crc standard probability and statistics tables and formulae - Free

Loading...
CRC

standard probability and

Statistics tables and formulae

c 2000 by Chapman & Hall/CRC 

CRC

standard probability and

Statistics tables and formulae DANIEL ZWILLINGER Rensselaer Polytechnic Institute Troy, New York

STEPHEN KOKOSKA Bloomsburg University Bloomsburg, Pennsylvania

CHAPMAN & HALL/CRC Boca Raton London New York Washington, D.C.

Library of Congress Cataloging-in-Publication Data Zwillinger, Daniel, 1957CRC standard probability and statistics tables and formulae / Daniel Zwillinger, Stephen Kokoska. p. cm. Includes bibliographical references and index. ISBN 1-58488-059-7 (alk. paper) 1. Probabilities—Tables. 2. Mathematical statistics—Tables. I. Kokoska, Stephen. II. Title. QA273.3 .Z95 1999 519.2′02′1—dc21 99-045786

This book contains information obtained from authentic and highly regarded sources. Reprinted material is quoted with permission, and sources are indicated. A wide variety of references are listed. Reasonable efforts have been made to publish reliable data and information, but the author and the publisher cannot assume responsibility for the validity of all materials or for the consequences of their use. Neither this book nor any part may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopying, microfilming, and recording, or by any information storage or retrieval system, without prior permission in writing from the publisher. The consent of CRC Press LLC does not extend to copying for general distribution, for promotion, for creating new works, or for resale. Specific permission must be obtained in writing from CRC Press LLC for such copying. Direct all inquiries to CRC Press LLC, 2000 N.W. Corporate Blvd., Boca Raton, Florida 33431. Trademark Notice: Product or corporate names may be trademarks or registered trademarks, and are used only for identification and explanation, without intent to infringe.

Visit the CRC Press Web site at www.crcpress.com © 2000 by Chapman & Hall/CRC No claim to original U.S. Government works International Standard Book Number 1-58488-059-7 Library of Congress Card Number 99-045786 Printed in the United States of America 2 3 4 5 6 7 8 9 0 Printed on acid-free paper

Preface It has long been the established policy of CRC Press to publish, in handbook form, the most up-to-date, authoritative, logically arranged, and readily usable reference material available. This book fills the need in probability and statistics. Prior to the preparation of this book the contents of similar books were considered. It is easy to fill a statistics reference book with many hundred pages of tables—indeed, some large books contain statistical tables for only a single test. The authors of this book focused on the basic principles of statistics. We have tried to ensure that each topic had an understandable textual introduction as well as easily understood examples. There are more than 80 examples; they usually follow the same format: start with a word problem, interpret the words as a statistical problem, find the solution, interpret the solution in words. We have organized this reference in an efficient and useful format. We believe both students and researchers will find this reference easy to read and understand. Material is presented in a multi-sectional format, with each section containing a valuable collection of fundamental reference material—tabular and expository. This Handbook serves as a guide for determining appropriate statistical procedures and interpretation of results. We have assembled the most important concepts in probability and statistics, as experienced through our own teaching, research, and work in industry. For most topics, concise yet useful tables were created. In most cases, the tables were re-generated and verified against existing tables. Even very modest statistical software can generate many of the tables in the book—often to more decimal places and for more values of the parameters. The values in this book are designed to illustrate the range of possible values and act as a handy reference for the most commonly needed values. This book also contains many useful topics from more advanced areas of statistics, but these topics have fewer examples. Also included are a large collection of short topics containing many classical results and puzzles. Finally, a section on notation used in the book and a comprehensive index are also included.

c 2000 by Chapman & Hall/CRC 

In line with the established policy of CRC Press, this Handbook will be kept as current and timely as is possible. Revisions and anticipated uses of newer materials and tables will be introduced as the need arises. Suggestions for the inclusion of new material in subsequent editions and comments concerning the accuracy of stated information are welcomed. If any errata are discovered for this book, they will be posted to http://vesta.bloomu.edu/~skokoska/prast/errata. Many people have helped in the preparation of this manuscript. The authors are especially grateful to our families who have remained lighthearted and cheerful throughout the process. A special thanks to Janet and Kent, and to Joan, Mark, and Jen. Daniel Zwillinger [email protected] Stephen Kokoska [email protected] ACKNOWLEDGMENTS Plans 6.1–6.6, 6A.1–6A.6, and 13.1–13.5 (appearing on pages 331–337) originally appeared on pages 234–237, 276–279, and 522–523 of W. G. Cochran and G. M. Cox, Experimental Designs, Second Edition, John Wiley & Sons, Inc, New York, 1957. Reprinted by permission of John Wiley & Sons, Inc. The tables of Bartlett’s critical values (in section 10.6.2) are from D. D. Dyer and J. P. Keating, “On the Determination of Critical Values for Bartlett’s Test”, JASA, Volume 75, 1980, pages 313–319. Reprinted with permission from the Journal of American Statistical Association. Copyright 1980 by the American Statistical Association. All rights reserved. The tables of Cochran’s critical values (in section 10.7.1) are from C. Eisenhart, M. W. Hastay, and W. A. Wallis, Techniques of Statistical Analysis, McGraw-Hill Book Company, 1947, Tables 15.1 and 15.2 (pages 390-391). Reprinted courtesy of The McGraw-Hill Companies. The tables of Dunnett’s critical values (in section 12.1.4.5) are from C. W. Dunnett, “A Multiple Comparison Procedure for Comparing Several Treatments with a Control”, JASA, Volume 50, 1955, pages 1096–1121. Reprinted with permission from the Journal of American Statistical Association. Copyright 1980 by the American Statistical Association. All rights reserved. The tables of Duncan’s critical values (in section 12.1.4.3) are from L. Hunter, “Critical Values for Duncan’s New Multiple Range Test”, Biometrics, 1960, Volume 16, pages 671– 685. Reprinted with permission from the Journal of American Statistical Association. Copyright 1960 by the American Statistical Association. All rights reserved. Table 15.1 is reproduced, by permission, from ASTM Manual on Quality Control of Materials, American Society for Testing and Materials, Philadelphia, PA, 1951. The table in section 15.1.2 and much of Chapter 18 originally appeared in D. Zwillinger, Standard Mathematical Tables and Formulae, 30th edition, CRC Press, Boca Raton, FL, 1995. Reprinted courtesy of CRC Press, LLC. Much of section 17.17 is taken from the URL http://members.aol.com/johnp71/javastat.html Permission courtesy of John C. Pezzullo.

c 2000 by Chapman & Hall/CRC 

Contents 1

Introduction 1.1 Background 1.2 Data sets 1.3

2

Summarizing Data 2.1 Tabular and graphical procedures 2.2

3

References

Numerical summary measures

Probability 3.1 Algebra of sets 3.2 Combinatorial methods 3.3 3.4

Probability Random variables

3.5 3.6 3.7

Mathematical expectation Multivariate distributions Inequalities

4 4.1 4.2 4.3 4.4 4.5 4.6 4.7

Functions of Random Variables Finding the probability distribution Sums of random variables Sampling distributions Finite population Theorems Order statistics Range and studentized range

© 2000 by Chapman & Hall/CRC

5

Discrete Probability Distributions 5.1 Bernoulli distribution 5.2 Beta binomial distribution 5.3 5.4 5.5 5.6 5.7 5.8 5.9 5.10

6

Beta Pascal distribution Binomial distribution Geometric distribution Hypergeometric distribution Multinomial distribution Negative binomial distribution Poisson distribution Rectangular (discrete uniform) distribution

Continuous Probability Distributions 6.1 Arcsin distribution 6.2 Beta distribution 6.3 Cauchy distribution 6.4 6.5 6.6 6.7

Chi–square distribution Erlang distribution Exponential distribution Extreme–value distribution

6.8 6.9 6.10

F distribution Gamma distribution Half–normal distribution

6.11 6.12 6.13

Inverse Gaussian (Wald) distribution Laplace distribution Logistic distribution

6.14 6.15 6.16

Lognormal distribution Noncentral chi–square distribution Noncentral F distribution

6.17 6.18 6.19

Noncentral t distribution Normal distribution Normal distribution: multivariate

6.20 6.21 6.22 6.23

Pareto distribution Power function distribution Rayleigh distribution t distribution

c 2000 by Chapman & Hall/CRC 

7

6.24 6.25 6.26

Triangular distribution Uniform distribution Weibull distribution

6.27

Relationships among distributions

Standard Normal Distribution 7.1 Density function and related functions 7.2 Critical values 7.3 7.4 7.5 7.6 7.7 7.8

8

Estimation 8.1 Definitions 8.2 Cram´er–Rao inequality 8.3 8.4 8.5

Theorems The method of moments The likelihood function

8.6 8.7 8.8

The method of maximum likelihood Invariance property of MLEs Different estimators

8.9 8.10

Estimators for small samples Estimators for large samples

9 9.1 9.2 9.3 9.4 9.5 9.6 9.7 10

Tolerance factors for normal distributions Operating characteristic curves Multivariate normal distribution Distribution of the correlation coefficient Circular normal probabilities Circular error probabilities

Confidence Intervals Definitions Common critical values Sample size calculations Summary of common confidence intervals Confidence intervals: one sample Confidence intervals: two samples Finite population correction factor Hypothesis Testing

c 2000 by Chapman & Hall/CRC 

10.1 10.2 10.3

Introduction The Neyman–Pearson lemma Likelihood ratio tests

10.4 10.5 10.6 10.7 10.8 10.9 10.10 10.11

Goodness of fit test Contingency tables Bartlett’s test Cochran’s test Number of observations required Critical values for testing outliers Significance test in 2 × 2 contingency tables Determining values in Bernoulli trials

11

Regression Analysis 11.1 Simple linear regression 11.2 Multiple linear regression 11.3 Orthogonal polynomials

12

Analysis of Variance

13

14

12.1 12.2 12.3

One-way anova Two-way anova Three-factor experiments

12.4 12.5 12.6

Manova Factor analysis Latin square design

Experimental Design 13.1 Latin squares 13.2 Graeco–Latin squares 13.3 Block designs 13.4 13.5

Factorial experimentation: 2 factors 2r Factorial experiments

13.6 13.7

Confounding in 2n factorial experiments Tables for design of experiments

13.8

References

Nonparametric Statistics 14.1 Friedman test for randomized block design

c 2000 by Chapman & Hall/CRC 

14.2 14.3 14.4

Kendall’s rank correlation coefficient Kolmogorov–Smirnoff tests Kruskal–Wallis test

14.5 14.6 14.7 14.8 14.9 14.10

The runs test The sign test Spearman’s rank correlation coefficient Wilcoxon matched-pairs signed-ranks test Wilcoxon rank–sum (Mann–Whitney) test Wilcoxon signed-rank test

15

Quality Control and Risk Analysis 15.1 Quality assurance 15.2 Acceptance sampling 15.3 Reliability 15.4 Risk analysis and decision rules

16

General Linear Models 16.1 Notation

17

16.2 16.3 16.4

The general linear model Summary of rules for matrix operations Quadratic forms

16.5 16.6

General linear hypothesis of full rank General linear model of less than full rank

Miscellaneous Topics 17.1 17.2 17.3 17.4

Geometric probability Information and communication theory Kalman filtering Large deviations (theory of rare events)

17.5 17.6 17.7 17.8 17.9 17.10 17.11 17.12

Markov chains Martingales Measure theoretical probability Monte Carlo integration techniques Queuing theory Random matrix eigenvalues Random number generation Resampling methods

c 2000 by Chapman & Hall/CRC 

17.13 Self-similar processes 17.14 Signal processing 17.15 Stochastic calculus 17.16 Classic and interesting problems 17.17 Electronic resources 17.18 Tables 18

Special Functions 18.1 18.2 18.3 18.4 18.5 18.6 18.7 18.8

Bessel functions Beta function Ceiling and floor functions Delta function Error functions Exponential function Factorials and Pochhammer’s symbol Gamma function

18.9 18.10 18.11 18.12

Hypergeometric functions Logarithmic functions Partitions Signum function

18.13 Stirling numbers 18.14 Sums of powers of integers 18.15 Tables of orthogonal polynomials 18.16 References Notation

c 2000 by Chapman & Hall/CRC 

CHAPTER 1

Introduction Contents 1.1 1.2 1.3

1.1

Background Data sets References

BACKGROUND

The purpose of this book is to provide a modern set of tables and a comprehensive list of definitions, concepts, theorems, and formulae in probability and statistics. While the numbers in these tables have not changed since they were first computed (in some cases, several hundred years ago), the presentation format here is modernized. In addition, nearly all table values have been re-computed to ensure accuracy. Almost every table is presented along with a textual description and at least one example using a value from the table. Most concepts are illustrated with examples and step-by-step solutions. Several data sets are described in this chapter; they are used in this book in order for users to be able to check algorithms. The emphasis of this book is on what is often called basic statistics. Most real-world statistics users will be able to refer to this book in order to quickly verify a formula, definition, or theorem. In addition, the set of tables here should make this a complete statistics reference tool. Some more advanced useful and current topics, such as Brownian motion and decision theory are also included. 1.2

DATA SETS

We have established a few data sets which we have used in examples throughout this book. With these, a user can check a local statistics program by verifying that it returns the same values as given in this book. For example, the correlation coefficient between the first 100 elements of the sequence of integers {1, 2, 3 . . . } and the first 100 elements of the sequence of squares {1, 4, 9 . . . } is 0.96885. Using this value is an easy way to check for correct computation of a computer program. These data sets may be obtained from http://vesta.bloomu.edu/~skokoska/prast/data. c 2000 by Chapman & Hall/CRC 

Ticket data: Forty random speeding tickets were selected from the courthouse records in Columbia County. The speed indicated on each ticket is given in the table below. 58 64 74 62

72 59 67 63

64 65 55 83

65 55 68 64

67 75 74 51

92 56 43 63

55 89 67 49

51 60 71 78

69 84 72 65

73 68 66 75

Swimming pool data: Water samples from 35 randomly selected pools in Beverly Hills were tested for acidity. The following table lists the PH for each sample. 6.4 7.0 7.0 5.9 6.4

6.6 5.9 7.0 7.2 6.3

6.2 5.7 6.0 7.3 6.2

7.2 7.0 6.3 7.7 7.5

6.2 7.4 5.6 6.8 6.7

8.1 6.5 6.3 5.2 6.4

7.0 6.8 5.8 5.2 7.8

Soda pop data: A new soda machine placed in the Mathematics Building on campus recorded the following sales data for one week in April. Soda Pepsi Wild Cherry Pepsi Diet Pepsi Seven Up Mountain Dew Lipton Ice Tea 1.3

Number of cans 72 60 85 54 32 64

REFERENCES

Gathered here are some of the books referenced in later sections; each has a broad coverage of the topics it addresses. 1. W. G. Cochran and G. M. Cox, Experimental Designs, Second Edition, John Wiley & Sons, Inc., New York, 1957. 2. C. J. Colbourn and J. H. Dinitz, CRC Handbook of Combinatorial Designs, CRC Press, Boca Raton, FL, 1996. 3. L. Devroye, Non-Uniform Random Variate Generation, Springer–Verlag, New York, 1986. 4. W. Feller, An Introduction to Probability Theory and Its Applications, Volumes 1 and 2, John Wiley & Sons, New York, 1968. 5. C. W. Gardiner, Handbook of Stochastic Methods, Second edition, Springer– Verlag, New York, 1985. 6. D. J. Sheskin, Handbook of Parametric and Nonparametric Statistical Procedures, CRC Press LLC, Boca Raton, FL, 1997. c 2000 by Chapman & Hall/CRC 

CHAPTER 2

Summarizing Data Contents 2.1

Tabular and graphical procedures 2.1.1 Stem-and-leaf plot 2.1.2 Frequency distribution 2.1.3 Histogram 2.1.4 Frequency polygons 2.1.5 Chernoff faces 2.2 Numerical summary measures 2.2.1 (Arithmetic) mean 2.2.2 Weighted (arithmetic) mean 2.2.3 Geometric mean 2.2.4 Harmonic mean 2.2.5 Mode 2.2.6 Median 2.2.7 p% trimmed mean 2.2.8 Quartiles 2.2.9 Deciles 2.2.10 Percentiles 2.2.11 Mean deviation 2.2.12 Variance 2.2.13 Standard deviation 2.2.14 Standard errors 2.2.15 Root mean square 2.2.16 Range 2.2.17 Interquartile range 2.2.18 Quartile deviation 2.2.19 Box plots 2.2.20 Coefficient of variation 2.2.21 Coefficient of quartile variation 2.2.22 Z score 2.2.23 Moments 2.2.24 Measures of skewness

c 2000 by Chapman & Hall/CRC 

2.2.25 Measures of kurtosis 2.2.26 Data transformations 2.2.27 Sheppard’s corrections for grouping

Numerical descriptive statistics and graphical techniques may be used to summarize information about central tendency and/or variability. 2.1

TABULAR AND GRAPHICAL PROCEDURES

2.1.1

Stem-and-leaf plot

A stem-and-leaf plot is a a graphical summary used to describe a set of observations (as symmetric, skewed, etc.). Each observation is displayed on the graph and should have at least two digits. Split each observation (at the same point) into a stem (one or more of the leading digit(s)) and a leaf (remaining digits). Select the split point so that there are 5–20 total stems. List the stems in a column to the left, and write each leaf in the corresponding stem row. Example 2.1 : Construct a stem-and-leaf plot for the Ticket Data (page 2). Solution:

Stem 4 5 6 7 8 9

Leaf 3 1 0 1 3 2

9 1 2 2 4

5 5 5 6 8 9 3 3 4 4 4 5 5 5 6 7 7 7 8 8 9 2 3 4 4 5 5 8 9

Stem = 10, Leaf = 1 Figure 2.1: Stem–and–leaf plot for Ticket Data. 2.1.2

Frequency distribution

A frequency distribution is a tabular method for summarizing continuous or discrete numerical data or categorical data. (1) Partition the measurement axis into 5–20 (usually equal) reasonable subintervals called classes, or class intervals. Thus, each observation falls into exactly one class. (2) Record, or tally, the number of observations in each class, called the frequency of each class. (3) Compute the proportion of observations in each class, called the relative frequency. (4) Compute the proportion of observations in each class and all preceding classes, called the cumulative relative frequency. c 2000 by Chapman & Hall/CRC 

Example 2.2 : Construct a frequency distribution for the Ticket Data (page 2). Solution: (S1) Determine the classes. It seems reasonable to use 40 to less than 50, 50 to less than 60, . . . , 90 to less than 100. Note: For continuous data, one end of each class must be open. This ensures that each observation will fall into only one class. The open end of each class may be either the left or right, but should be consistent. (S2) Record the number of observations in each class. (S3) Compute the relative frequency and cumulative relative frequency for each class. (S4) The resulting frequency distribution is in Figure 2.2.

Class

Frequency

Relative frequency

Cumulative relative frequency

2 8 17 9 3 1

0.050 0.200 0.425 0.225 0.075 0.025

0.050 0.250 0.625 0.900 0.975 1.000

[40, 50) [50, 60) [60, 70) [70, 80) [80, 90) [90, 100)

Figure 2.2: Frequency distribution for Ticket Data. 2.1.3

Histogram

A histogram is a graphical representation of a frequency distribution. A (relative) frequency histogram is a plot of (relative) frequency versus class interval. Rectangles are constructed over each class with height proportional (usually equal) to the class (relative) frequency. A frequency and relative frequency histogram have the same shape, but different scales on the vertical axis. Example 2.3 : Construct a frequency histogram for the Ticket Data (page 2). Solution: (S1) Using the frequency distribution in Figure 2.2, construct rectangles above each class, with height equal to class frequency. (S2) The resulting histogram is in Figure 2.3.

Note: A probability histogram is constructed so that the area of each rectangle equals the relative frequency. If the class widths are unequal, this histogram presents a more accurate description of the distribution. 2.1.4

Frequency polygons

A frequency polygon is a line plot of points with x coordinate being class midpoint and y coordinate being class frequency. Often the graph extends to

c 2000 by Chapman & Hall/CRC 

Figure 2.3: Frequency histogram for Ticket Data. an additional empty class on both ends. The relative frequency may be used in place of frequency. Example 2.4 : Construct a frequency polygon for the Ticket Data (page 2). Solution: (S1) Using the frequency distribution in Figure 2.2, plot each point and connect the graph. (S2) The resulting frequency polygon is in Figure 2.4.

Figure 2.4: Frequency polygon for Ticket Data. An ogive, or cumulative frequency polygon, is a plot of cumulative frequency versus the upper class limit. Figure 2.5 is an ogive for the Ticket Data (page 2). Another type of frequency polygon is a more-than cumulative frequency polygon. For each class this plots the number of observations in that class and every class above versus the lower class limit. c 2000 by Chapman & Hall/CRC 

Figure 2.5: Ogive for Ticket Data. A bar chart is often used to graphically summarize discrete or categorical data. A rectangle is drawn over each bin with height proportional to frequency. The chart may be drawn with horizontal rectangles, in three dimensions, and may be used to compare two or more sets of observations. Figure 2.6 is a bar chart for the Soda Pop Data (page 2).

Figure 2.6: Bar chart for Soda Pop Data. A pie chart is used to illustrate parts of the total. A circle is divided into slices proportional to the bin frequency. Figure 2.7 is a pie chart for the Soda Pop Data (page 2). 2.1.5

Chernoff faces

Chernoff faces are used to illustrate trends in multidimensional data. They are effective because people are used to differentiating between facial features. Chernoff faces have been used for cluster, discriminant, and time-series analyses. Facial features that might be controllable by the data include: (a) ear: level, radius (b) eyebrow: height, slope, length (c) eyes: height, size, separation, eccentricity, pupil position or size

c 2000 by Chapman & Hall/CRC 

Figure 2.7: Pie chart for Soda Pop Data. (d) face: width, half-face height, lower or upper eccentricity (e) mouth: position of center, curvature, length, openness (f) nose: width, length The Chernoff faces in Figure 2.8 come from data about this book. For the even chapters: (a) eye size is proportional to the approximate number of pages (b) mouth size is proportional to the approximate number of words (c) face shape is proportional to the approximate number of occurrences of the word “the” The data are as follows: Chapter

2

4

6

8

10

12

14

16

18

Number of pages 18 30 56 8 36 40 40 26 23 Number of words 4514 5426 12234 2392 9948 18418 8179 11739 5186 Occurrences of “the” 159 147 159 47 153 118 264 223 82

An interactive program for creating Chernoff faces is available at http:// www.hesketh.com/schampeo/projects/Faces/interactive.shtml. See H. Chernoff, “The use of faces to represent points in a K-dimensional space graphically,” Journal of the American Statistical Association, Vol. 68, No. 342, 1973, pages 361–368. 2.2

NUMERICAL SUMMARY MEASURES

The following conventions will be used in the definitions and formulas in this section. (C1) Ungrouped data: Let x1 , x2 , x3 , . . . , xn be a set of observations. (C2) Grouped data: Let x1 , x2 , x3 , . . . , xk be a set of class marks from a frequency distribution, or a representative set of observations, with corre-

c 2000 by Chapman & Hall/CRC 

Figure 2.8: Chernoff faces for chapter data. sponding frequencies f1 , f2 , f3 , . . . , fk . The total number of observations k  is n = fi . Let c denote the (constant) width of each bin and xo one i=1

of the class marks selected to be the computing origin. Each class mark, xi , may be coded by ui = (xi − xo )/c. Each ui will be an integer and the bin mark taken as the computing origin will be coded as a 0. 2.2.1

(Arithmetic) mean

The (arithmetic) mean of a set of observations is the sum of the observations divided by the total number of observations. (1) Ungrouped data: 1 x1 + x2 + x3 + · · · + xn x= xi = n i=1 n n

(2.1)

(2) Grouped data: 1 f1 x1 + f2 x2 + f3 x3 + · · · + fn xn fi xi = n i=1 n k

x=

c 2000 by Chapman & Hall/CRC 

(2.2)

(3) Coded data: k 

x = xo + c · 2.2.2

fi ui

i=1

n

(2.3)

Weighted (arithmetic) mean

Let wi ≥ 0 be the weight associated with observation xi . The total weight is n  given by wi and the weighted mean is i=1 n 

wi xi w1 x1 + w2 x2 + w3 x3 + · · · + wn xn xw = i=1 = . n  w1 + w2 + w3 + · · · + wn wi

(2.4)

i=1

2.2.3

Geometric mean

For ungrouped data such that xi > 0, the geometric mean is the nth root of the product of the observations: √ GM = n x1 · x2 · x3 · · · xn . (2.5) In logarithmic form: 1 log x1 + log x2 + log x3 + · · · + log xn . log xi = n i=1 n n

log(GM) =

For grouped data with each class mark xi > 0:  n GM = xf11 · xf22 · xf33 · · · xfkk .

(2.6)

(2.7)

In logarithmic form: 1 fi log(xi ) n i=1 k

log(GM) =

= 2.2.4

(2.8)

f1 log(x1 ) + f2 log(x2 ) + f3 log(x3 ) + · · · + fk log(xk ) . n

Harmonic mean

For ungrouped data the harmonic mean is given by n n HM =  . n 1 = 1 1 1 1 + + + ··· + x1 x2 x3 xn i=1 xi

c 2000 by Chapman & Hall/CRC 

(2.9)

For grouped data: HM =

n n = . k f f f f fk  1 2 3 i + + + ··· + x1 x2 x3 xk i=1 xi

(2.10)

Note: The equation involving the arithmetic, geometric, and harmonic mean is HM ≤ GM ≤ x .

(2.11)

Equality holds if all n observations are equal. 2.2.5

Mode

For ungrouped data, the mode, Mo , is the value that occurs most often, or with the greatest frequency. A mode may not exist, for example, if all observations occur with the same frequency. If the mode does exist, it may not be unique, for example, if two observations occur with the greatest frequency. For grouped data, select the class containing the largest frequency, called the modal class. Let L be the lower boundary of the modal class, dL the difference in frequencies between the modal class and the class immediately below, and dH the difference in frequencies between the modal class and the class immediately above. The mode may be approximated by Mo ≈ L + c · 2.2.6

dL . dL + dH

(2.12)

Median

The median, x ˜, is another measure of central tendency, resistant to outliers. For ungrouped data, arrange the observations in order from smallest to largest. If n is odd, the median is the middle value. If n is even, the median is the mean of the two middle values. For grouped data, select the class containing the median (median class). Let L be the lower boundary of the median class, fm the frequency of the median class, and CF the sum of frequencies for all classes below the median class (a cumulative frequency). The median may be approximated by n − CF x ˜≈L+c· 2 . (2.13) fm Note: If x > x ˜ the distribution is positively skewed. If x < x ˜ the distribution is negatively skewed. If x ≈ x ˜ the distribution is approximately symmetric. 2.2.7

p% trimmed mean

A trimmed mean is a measure of central tendency and a compromise between a mean and a median. The mean is more sensitive to outliers, and the median is less sensitive to outliers. Order the observations from smallest to largest. c 2000 by Chapman & Hall/CRC 

Delete the smallest p% and the largest p% of the observations. The p% trimmed mean, xtr(p) , is the arithmetic mean of the remaining observations. Note: If p% of n (observations) is not an integer, several (computer) algorithms exist for interpolating at each end of the distribution and for determining xtr(p) . Example 2.5 : Using the Swimming Pool data (page 2) find the mean, median, and mode. Compute the geometric mean and the harmonic mean, and verify the relationship between these three measures. Solution: (S1) x =

1 (6.4 + 6.6 + 6.2 + · · · + 7.8) = 6.5886 35

(S2) x ˜ = 6.5, the middle values when the observations are arranged in order from smallest to largest. (S3) Mo = 7.0, the observation that occurs most often.  (S4) GM = 35 (6.4)(6.6)(6.2) · · · (7.8) = 6.5513 (S5) HM =

35 = 6.5137 (1/6.4) + (1/6.6) + (1/6.2) + · · · + (1/7.8)

(S6) To verify the inequality: 6.5137   ≤ 6.5513  ≤ 6.5886  HM

2.2.8

GM

x

Quartiles

Quartiles split the data into four parts. For ungrouped data, arrange the observations in order from smallest to largest. (1) The second quartile is the median: Q2 = x ˜. (2) If n is even: The first quartile, Q1 , is the median of the smallest n/2 observations; and the third quartile, Q3 , is the median of the largest n/2 observations. (3) If n is odd: The first quartile, Q1 , is the median of the smallest (n + 1)/2 observations; and the third quartile, Q3 , is the median of the largest (n + 1)/2 observations. For grouped data, the quartiles are computed by applying equation (2.13) for the median. Compute the following: L1 = the lower boundary of the class containing Q1 . L3 = the lower boundary of the class containing Q3 . f1 = the frequency of the class containing the first quartile. f3 = the frequency of the class containing the third quartile. CF1 = cumulative frequency for classes below the one containing Q1 . CF3 = cumulative frequency for classes below the one containing Q3 . c 2000 by Chapman & Hall/CRC 

The (approximate) quartiles are given by n − CF1 Q1 = L1 + c · 4 f1 2.2.9

3n − CF3 Q3 = L3 + c · 4 . f3

(2.14)

Deciles

Deciles split the data into 10 parts. (1) For ungrouped data, arrange the observations in order from smallest to largest. The ith decile, Di (for i = 1, 2, . . . , 9), is the i(n + 1)/10th observation. It may be necessary to interpolate between successive values. (2) For grouped data, apply equation (2.13) (as in equation (2.14)) for the median to find the approximate deciles. Di is in the class containing the i n/10th largest observation. 2.2.10

Percentiles

Percentiles split the data into 100 parts. (1) For ungrouped data, arrange the observations in order from smallest to largest. The ith percentile, Pi (for i = 1, 2, . . . , 99), is the i(n + 1)/100th observation. It may be necessary to interpolate between successive values. (2) For grouped data, apply equation (2.13) (as in equation (2.14)) for the median to find the approximate percentiles. Pi is in the class containing the i n/100th largest observation. 2.2.11

Mean deviation

The mean deviation is a measure of variability based on the absolute value of the deviations about the mean or median. (1) For ungrouped data: 1 1 |xi − x| or MD = |xi − x ˜| . n i=1 n i=1 n

MD =

n

(2.15)

(2) For grouped data: 1 1 fi |xi − x| or MD = fi |xi − x ˜| . n i=1 n i=1 k

MD = 2.2.12

k

(2.16)

Variance

The variance is a measure of variability based on the squared deviations about the mean.

c 2000 by Chapman & Hall/CRC 

(1) For ungrouped data: 1  (xi − x)2 . n − 1 i=1 n

s2 =

(2.17)

The computational formula for s2 : 

2 

n n n    1 1  1 2 2 2 2 s = x − xi  = x − nx . (2.18) n − 1 i=1 i n i=1 n − 1 i=1 i (2) For grouped data: 1  fi (xi − x)2 . n − 1 i=1 k

s2 =

The computational formula for s2 : 

2  k k   1 1  s2 = fi x2i − fi xi  n − 1 i=1 n i=1 1 = n−1

k 

(2.19)

(2.20)

fi x2i − nx2 .

i=1

(3) For coded data:



2  k k   c 1  s2 = fi u2i − fi ui  . n − 1 i=1 n i=1

2.2.13

Standard deviation

The standard deviation is the positive square root of the variance: s =

(2.21)



s2 .

The probable error is 0.6745 times the standard deviation. 2.2.14

Standard errors

The standard error of a statistic is the standard deviation of the sampling distribution of that statistic. The standard error of a statistic is often designated by σ with a subscript indicating the statistic. 2.2.14.1

Standard error of the mean

The standard error of the mean is used in hypothesis testing and is an indication of the accuracy of the estimate x. √ SEM = s/ n . (2.22)

c 2000 by Chapman & Hall/CRC 

2.2.15

Root mean square

(1) For ungrouped data: RMS =

1 2 x n i=1 i n

1/2 .

(2.23)

(2) For grouped data: RMS = 2.2.16

1 fi x2i n i=1 k

1/2 .

(2.24)

Range

The range is the difference between the largest and smallest values. R = max{x1 , x2 , . . . , xn } − min{x1 , x2 , . . . , xn } = x(n) − x(1) . 2.2.17

(2.25)

Interquartile range

The interquartile range, or fourth spread, is the difference between the third and first quartile. IQR = Q3 − Q1 . 2.2.18

(2.26)

Quartile deviation

The quartile deviation, or semi-interquartile range, is half the interquartile range. QD = 2.2.19

Q3 − Q1 . 2

(2.27)

Box plots

Box plots, also known as quantile plots, are graphics which display the center portions of the data and some information about the range of the data. There are a number of variations and a box plot may be drawn with either a horizontal or vertical scale. The inner and outer fences are used in constructing a box plot and are markers used in identifying mild and extreme outliers. Inner Fences: Q1 − 1.5 · IQR, Q1 + 1.5 · IQR Outer Fences: Q3 − 3 · IQR,

c 2000 by Chapman & Hall/CRC 

Q3 + 3 · IQR

(2.28)

A general description:

Multiple box plots on the same measurement axis may be used to compare the center and spread of distributions. Figure 2.9 presents box plots for randomly selected August residential electricity bills for three different parts of the country.

Figure 2.9: Example of multiple box plots.

c 2000 by Chapman & Hall/CRC 

2.2.20

Coefficient of variation

The coefficient of variation is a measure of relative variability. Reported as percentage it is defined as: s CV = 100 . (2.29) x 2.2.21

Coefficient of quartile variation

The coefficient of quartile variation is a measure of variability. CQV = 100 2.2.22

Q3 − Q1 . Q3 + Q1

(2.30)

Z score

The z score, or standard score, associated with an observation is a measure of relative standing. z= 2.2.23

xi − x s

(2.31)

Moments

Moments are used to characterize a set of observations. (1) For ungrouped data: The rth moment about the origin: 1 r x . n i=1 i n

mr =

The rth moment about the mean x: n r    r 1 r mr = (−1)j mr−j xj . (xi − x) = j n i=1 j=0

(2.32)

(2.33)

(2) For grouped data: The rth moment about the origin: 1 fi xri . n i=1 k

mr =

The rth moment about the mean x:  1 mr = fi (xi − x ¯)r = n i=1 j=0 k

r

  r (−1)j mr−j x ¯j . j

(2.34)

(2.35)

(3) For coded data: mr =

c 2000 by Chapman & Hall/CRC 

n cr  fi uri . n i=1

(2.36)

2.2.24

Measures of skewness

The following descriptive statistics measure the lack of symmetry. Larger values (in magnitude) indicate more skewness in the distribution of observations. 2.2.24.1

Coefficient of skewness g1 =

2.2.24.2

m3

(2.37)

3/2

m2

Coefficient of momental skewness g1 m3 = 3/2 2 2m2

2.2.24.3

Pearson’s first coefficient of skewness Sk 1 =

2.2.24.4

3(x − Mo ) s

(2.39)

Pearson’s second moment of skewness Sk2 =

2.2.24.5

(2.38)

˜) 3(x − x s

(2.40)

Quartile coefficient of skewness Sk q =

Q3 − 2˜ x + Q1 Q3 − Q1

(2.41)

Example 2.6 : Using the Swimming Pool data (page 2) find the coefficient of skewness, coefficient of momental skewness, Pearson’s first coefficient of skewness, Pearson’s second moment of skewness, and the quartile coefficient of skewness. Solution: (S1) x ¯ = 6.589, x ˜ = 6.5, s = 0.708, Q1 = 6.2, Q3 = 7.0, Mo = 7.0 1 (xi − x ¯)2 = 0.4867 n i=1 35

(S2) m2 =

1 (xi − x ¯)3 = 0.0126 n i=1 35

m3 =

(S3) g1 = 0.0126/(0.4867)3/2 = 0.0371 , g1 /2 = 0.0372/2 = 0.0186 3(6.589 − 7) 3(6.589 − 6.5) (S4) Sk1 = = −1.7415 , Sk2 = = 0.3771 0.708 0.708 7.0 − 2(6.5) + 6.2 (S5) Skq = = 0.25 7.0 − 6.2 2.2.25

Measures of kurtosis

The following statistics describe the extent of the peak in a distribution. Smaller values (in magnitude) indicate a flatter, more uniform distribution. c 2000 by Chapman & Hall/CRC 

2.2.25.1

Coefficient of kurtosis g2 =

2.2.25.2

(2.42)

m4 −3 m22

(2.43)

Coefficient of excess kurtosis g2 − 3 =

2.2.26

m4 m22

Data transformations

Suppose yi = axi + b for i = 1, 2, . . . , n. The following summary statistics for the distribution of y’s are related to summary statistics for the distribution of x’s. s2y = a2 s2x ,

y = ax + b , 2.2.27

sy = |a|sx

(2.44)

Sheppard’s corrections for grouping

For grouped data, suppose every class interval has width c. If both tails of the distribution are very flat and close to the measurement axis, the grouped data approximation to the sample variance may be improved by using Sheppard’s correction, −c2 /12: c2 12 and mrc :

corrected variance = grouped data variance − There are similar corrected sample moments, denoted mrc m1c = m1

m1 c = m 1

c2 12 c2  = m3 − m 4 1 2 c 7c2 = m4 − m1 + 2 240

m2c = m2 −

m2 c = m 2 −

m3c

m3 c = m 3

m4c

(2.45)

m4 c = m 4 −

c2 12

(2.46)

c2 7c2 m2 + 2 240

Example 2.7 : Consider the grouped Ticket Data (page 2) as presented in the frequency distribution in Example 2.2 (on page 5). Find the corrected sample variance and corrected sample moments. Solution: (S1) x ¯ = 66.5 ,

s2 = 115.64 (for grouped data), c = 10

(S2) corrected variance = 115.64 − (102 /12) = 107.31 (S3) r 1 2 3 4

mr

mrc

mr

mrc

66.5 4535.0 316962.5 22692125.0

66.5 4526.7 315300.0 22688802.9

0.0 112.8 389.3 40637.3

0.0 104.4 389.3 35002.7

c 2000 by Chapman & Hall/CRC 

CHAPTER 3

Probability Contents 3.1 3.2

Algebra of sets Combinatorial methods 3.2.1 The product rule for ordered pairs 3.2.2 The generalized product rule for k-tuples 3.2.3 Permutations 3.2.4 Circular permutations 3.2.5 Combinations (binomial coefficients) 3.2.6 Sample selection 3.2.7 Balls into cells 3.2.8 Multinomial coefficients 3.2.9 Arrangements and derangements 3.3 Probability 3.3.1 Relative frequency concept of probability 3.3.2 Axioms of probability (discrete sample space) 3.3.3 The probability of an event 3.3.4 Probability theorems 3.3.5 Probability and odds 3.3.6 Conditional probability 3.3.7 The multiplication rule 3.3.8 The law of total probability 3.3.9 Bayes’ theorem 3.3.10 Independence 3.4 Random variables 3.4.1 Discrete random variables 3.4.2 Continuous random variables 3.4.3 Random functions 3.5 Mathematical expectation 3.5.1 Expected value 3.5.2 Variance 3.5.3 Moments 3.5.4 Generating functions c 2000 by Chapman & Hall/CRC 

3.6

Multivariate distributions 3.6.1 Discrete case 3.6.2 Continuous case 3.6.3 Expectation 3.6.4 Moments 3.6.5 Marginal distributions 3.6.6 Independent random variables 3.6.7 Conditional distributions 3.6.8 Variance and covariance 3.6.9 Correlation coefficient 3.6.10 Moment generating function 3.6.11 Linear combination of random variables 3.6.12 Bivariate distribution 3.7 Inequalities

3.1

ALGEBRA OF SETS

Properties of and operations on sets are important since events may be thought of as sets. Some set facts: (1) A set A is a collection of objects called the elements of the set. a ∈ A means a is an element of the set A. a ∈ A means a is not an element of the set A. A = {a, b, c} is used to denote the elements of the set A. (2) The null set, denoted by φ or { }, is the empty set; the set that contains no elements. (3) Two sets A and B are equal, written A = B, if 1) every element of A is an element of B, and 2) every element of B is an element of A. (4) The set A is a subset of the set B if every element of A is also in B; written A ⊂ B (or B ⊃ A). For every set A, φ ⊂ A. (5) If A ⊂ B and B ⊂ A then A = B and A is an improper subset of B. If A ⊂ B and there is at least one element of B not in A then A is a proper subset of B. The subset symbol ⊂ is often used to denote a proper subset while the symbol ⊆ indicates an improper subset. (6) Let S be the universal set, the set consisting of all elements of interest. For any set A, A ⊂ S. (7) The complement of the set A, denoted A , is the set consisting of all elements in S but not in A (Figure 3.1). (8) For any two sets A and B: The union of A and B, denoted A∪B, is the set consisting of all elements in A, or B, or both (Figure 3.2). The intersection of A and B, denoted A ∩ B, is the set consisting of all elements in both A and B (Figure 3.3). c 2000 by Chapman & Hall/CRC 

(9) A and B are disjoint or mutually exclusive if A ∩ B = φ (Figure 3.4).

Figure 3.1: Shaded region = A .

Figure 3.2: Shaded region = A ∪ B.

Figure 3.3: Shaded region = A ∩ B.

Figure 3.4: Mutually exclusive sets.

For the following properties, suppose A, B, and C are sets. It is necessary to assume these sets lie in a universal set S only in those properties that explicitly involve S. (1) Closure (a) There is a unique set A ∪ B. (b) There is a unique set A ∩ B. (2) Commutative laws (a) A ∪ B = B ∪ A (b) A ∩ B = B ∩ A (3) Associative laws (a) (A ∪ B) ∪ C = A ∪ (B ∪ C) (b) (A ∩ B) ∩ C = A ∩ (B ∩ C) (4) Distributive laws (a) A ∪ (B ∩ C) = (A ∪ B) ∩ (A ∪ C) (b) A ∩ (B ∪ C) = (A ∩ B) ∪ (A ∩ C) (5) Idempotent laws (a) A ∪ A = A (b) A ∩ A = A (6) Properties of S and φ

c 2000 by Chapman & Hall/CRC 

(a) A ∩ S = A (b) A ∪ φ = A (c) A ∩ φ = φ (d) A ∪ S = S (7) Properties of ⊂ (a) A ⊂ (A ∪ B) (b) (A ∩ B) ⊂ A (c) A ⊂ S (d) φ ⊂ A (e) If A ⊂ B, then A ∪ B = B and A ∩ B = A. (8) Properties of  (set complement) (a) For every set A, there is a unique set A . (b) A ∪ A = S (c) A ∩ A = φ (d)

(A ∪ B) = A ∩ B  (A ∩ B) = A ∪ B 

 DeMorgan’s laws

(9) Some generalizations Suppose A1 , A2 , A3 , . . . , An is a collection of sets. (a) The generalized union, A1 ∪ A2 ∪ · · · ∪ An , is the set consisting of all elements in at least one Ai . (b) The generalized intersection, A1 ∩A2 ∩· · ·∩An , is the set consisting of all elements in every Ai .   n n (c) ∪ Ai = (A1 ∪ · · · ∪ An ) = A1 ∩ · · · ∩ An = ∩ Ai  (d) 3.2

i=1 n

∩ Ai

i=1



i=1 n

= (A1 ∩ · · · ∩ An ) = A1 ∪ · · · ∪ An = ∪ Ai i=1

COMBINATORIAL METHODS

In an equally likely outcome experiment, computing the probability of an event involves counting. The following techniques are useful for determining the number of outcomes in an event and/or the sample space. 3.2.1

The product rule for ordered pairs

If the first element of an ordered pair can be selected in n1 ways, and for each of these n1 ways the second element of the pair can be selected in n2 ways, then the number of possible pairs is n1 n2 .

c 2000 by Chapman & Hall/CRC 

3.2.2

The generalized product rule for k-tuples

Suppose a sample space, or set, consists of ordered collections of k-tuples. If there are n1 choices for the first element, and for each choice of the first element there are n2 choices for the second element, . . . , and for each of the first k − 1 elements there are nk choices for the k th element, then there are n1 n2 · · · nk possible k-tuples. 3.2.3

Permutations

The number of permutations of n distinct objects taken k at a time is P (n, k) =

n! . (n − k)!

(3.1)

A table of values is on page 500. 3.2.4

Circular permutations

The number of permutations of n distinct objects arranged in a circle is (n − 1)!. 3.2.5

Combinations (binomial coefficients)   The binomial coefficient nk is the number of combinations of n distinct objects taken k at a time without regard to order:   n n! P (n, k) C(n, k) = = = . (3.2) k k!(n − k)! k! A table of values is on page 500. Other formulas involving binomial coefficients:    n n(n − 1) · · · (n − k + 1) n (a) = = k k! n−k       n n n (b) = = 1 and =n 0 n 1       n n−1 n−1 (c) = + k k k−1       n n n (d) + + ··· + = 2n 0 1 n       n n n =0 (e) − + · · · + (−1)n n 0 1   2n (2n−1)(2n−3)···3·1 (f) 2n = n! nn n+1 n+2    n+m+1 (g) n + n + n + · · · + n+m = n+1 n    n   n  n  n  n−1 or nn ) (h) 0 + 2 + 4 + · · · + = 2 (last term in sum is n−1        n    (i) n1 + n3 + n5 + · · · + = 2n−1 (last term in sum is n−1 or nn ) c 2000 by Chapman & Hall/CRC 

 n 2

 2  2   + n1 + · · · + nn = 2n 0 n m n m+n mn m n  0 p + 1 p−1 + · · · + p 0 = p n  n  n n−1 1 1 + 2 2 + · · · + n n = n2       1 n1 − 2 n2 + · · · + (−1)n+1 n nn = 0

(j) (k) (l) (m)

Example 3.8 : For the 5 element set {a, b, c, d, e} find the number of subsets containing exactly 3 elements. Solution:

5! 5 (S1) There are = = 10 subsets containing exactly 3 elements. 3!2! 3 (S2) The subsets are (a, b, c) (a, d, e)

3.2.6

(a, b, d) (b, c, d)

(a, b, e) (b, c, e)

(a, c, d) (b, d, e)

(a, c, e) (c, d, e)

Sample selection

There are 4 ways in which a sample of k elements can be obtained from a set of n distinguishable objects. Order counts? No Yes

Repetitions allowed? No No

No

Yes

Yes

Yes

where

The sample is called a k-combination k-permutation k-combination with replacement k-permutation with replacement

Number of ways to choose the sample C(n, k) P (n, k) C R (n, k) P R (n, k)

  n n! C(n, k) = = k k! (n − k)! n! P (n, k) = (n)k = nk = (n − k)! (n + k − 1)! C R (n, k) = C(n + k − 1, k) = k!(n − 1)!

(3.3)

P R (n, k) = nk Example 3.9 : There are 4 ways in which to choose a 2 element sample from the set {a, b}: combination permutation combination with replacement permutation with replacement c 2000 by Chapman & Hall/CRC 

C(2, 2) = 1 P (2, 2) = 2 C R (2, 2) = 3 P R (2, 2) = 4

ab ab and ba aa, ab, and bb aa, ab, ba, and bb

3.2.7

Balls into cells

There are 8 different ways in which n balls can be placed into k cells. Distinguish Distinguish Can cells Number of ways to balls? cells? be empty? place n balls into k cells Yes Yes Yes kn   Yes Yes No k! nk   No Yes Yes C(k + n − 1, n) =  k+n−1 n n−1 No Yes No C(n  n  − 1,  nk− 1) = k−1 n Yes No Yes  n1  + 2 + · · · + k Yes No No k No No Yes p1 (n) + p2 (n) + · · · + pk (n) No No No pk (n) n where k is the Stirling cycle number (see page 525) and pk (n) is the number of partitions of the number n into exactly k integer pieces (see page 523). Given n distinguishable balls and k distinguishable cells, the number of ways in which we can place n1 balls into cell 1, n2 balls  into cell2, . . . , nk balls into cell k, is given by the multinomial coefficient n1 ,n2n,...,nk . 3.2.8

Multinomial coefficients   The multinomial coefficient, n1 ,n2n,...,nk = C(n; n1 , n2 , . . . , nk ), is the number of ways of choosing n1 objects, then n2 objects, . . . , then nk objects from a collection of n distinct objects without regard to order. This requires that k n = n. j=1 j Other ways to interpret the multinomial coefficient: (1) Permutations (all objects not distinct): Given n1 objects of one kind, n2 objects of a second kind, . . . , nk objects of a k th kind, and n1 + n2 + n· · · +nk = n. The number of permutations of the n objects is n1 ,n2 ,...,nk . (2) Partitions: The number of ways of partitioning a set of n distinct objects into k subsets with n1 objects in the first subset, n2 objects in the second subset, . . . , and nk objects in the k th subset is n1 ,n2n,...,nk . The multinomial symbol is numerically evaluated as   n n! = n1 , n2 , . . . , nk n 1 ! n 2 ! · · · nk !

(3.4)

Example 3.10 : The number  4 of ways to choose 2 objects, then 1 object, then 1 object from the set {a, b, c, d} is 2,1,1 = 12; they are as follows (commas separate the ordered

c 2000 by Chapman & Hall/CRC 

selections): {ab, c, d} {ad, b, c} {bd, a, c}

3.2.9

{ab, d, c} {ad, c, b} {bd, c, a}

{ac, b, d} {bc, a, d} {cd, a, b}

{ac, d, b} {bc, d, a} {cd, b, a}

Arrangements and derangements

(a) The number of ways to arrange n distinct objects in a row is n!; this is the number of permutations of n objects. Example 3.11 : For the three objects {a, b, c} the number of arrangements is 3! = 6. These permutations are {abc, bac, cab, acb, bca, cba}.

(b) The number of ways to arrange n non-distinct objects (assuming that there are k types of objects, of each object of type i) is   and ni copies the multinomial coefficient n1 ,n2n,...,nk . Example 3.12 : For the set {a, a, b, c} the parameters are n = 4, k = 3, n1 = 2,  4 n2 = 1, and n3 = 1. Hence, there are 2,1,1 = 2! 4! = 12 arrangements, they 1! 1! are: aabc aacb abac abca acab acba baac baca bcaa caab caba cbaa (c) A derangement is a permutation of objects, in which object i is not in the ith location. Example 3.13 : All the derangements of {1, 2, 3, 4} are: 2143 3142 4123

2341 3412 4312

2413 3421 4321

The number of derangements of n elements, Dn , satisfies the recursion relation: Dn = (n − 1) (Dn−1 + Dn−2 ), with the initial values D1 = 0 and D2 = 1. Hence,   1 1 1 1 Dn = n! 1 − + − + · · · + (−1)n 1! 2! 3! n! The numbers Dn are also called subfactorials and rencontres numbers. For large values of n, Dn /n! ∼ e−1 ≈ 0.37. Hence, more than one of every three permutations is a derangement. n 1 2 3 4 5 6 7 8 9 10 Dn 0 1 2 9 44 265 1854 14833 133496 1334961 3.3

PROBABILITY

The sample space of an experiment, denoted S, is the set of all possible outcomes. Each outcome of the sample space is also called an element of the sample space or a sample point. An event is any collection of outcomes contained in the sample space. A simple event consists of exactly one outcome and a compound event consists of more than one outcome. c 2000 by Chapman & Hall/CRC 

3.3.1

Relative frequency concept of probability

Suppose an experiment is conducted n identical and independent times and n(A) is the number of times the event A occurs. The quotient n(A)/n is the relative frequency of occurrence of the event A. As n increases, the relative frequency converges to the limiting relative frequency of the event A. The probability of the event A, Prob [A], is this limiting relative frequency. 3.3.2

Axioms of probability (discrete sample space)

(1) For any event A, Prob [A] ≥ 0. (2) Prob [S] = 1. (3) If A1 , A2 , A3 , . . . , is a finite or infinite collection of pairwise mutually exclusive events of S, then Prob [A1 ∪ A2 ∪ A3 ∪ · · ·] = Prob [A1 ] + Prob [A2 ] + Prob [A3 ] + · · · (3.5) 3.3.3

The probability of an event

The probability of an event A is the sum of Prob [ai ] for all sample points ai in the event A:  Prob [A] = Prob [ai ] . (3.6) ai ∈A

If all of the outcomes in S are equally likely: Prob [A] = 3.3.4

n(A) number of outcomes in A = . n(S) number of outcomes in S

(3.7)

Probability theorems

(1) Prob [φ] = 0 for any sample space S. (2) If A and A are complementary events, Prob [A] + Prob [A ] = 1. (3) For any events A and B, if A ⊂ B then Prob [A] ≤ Prob [B]. (4) For any events A and B, Prob [A ∪ B] = Prob [A] + Prob [B] − Prob [A ∩ B] .

(3.8)

If A and B are mutually exclusive events, Prob [A ∩ B] = 0 and Prob [A ∪ B] = Prob [A] + Prob [B] .

(3.9)

(5) For any events A and B, Prob [A] = Prob [A ∩ B] + Prob [A ∩ B  ] .

c 2000 by Chapman & Hall/CRC 

(3.10)

(6) For any events A, B, and C, Prob [A ∪ B ∪ C] =Prob [A] + Prob [B] + Prob [C] − Prob [A ∩ B] − Prob [A ∩ C] − Prob [B ∩ C] + Prob [A ∩ B ∩ C]. (3.11) (7) For any events A1 , A2 , . . . , An ,    n n Prob ∪ Ai ≤ Prob [Ai ] . i=1

(3.12)

i=1

Equality holds if the events are pairwise mutually exclusive. 3.3.5

Probability and odds

If the probability of an event A is Prob [A] then odds for A = Prob [A]/Prob [A ],

Prob [A ] = 0

odds against A = Prob [A ]/Prob [A],

Prob [A] = 0.

(3.13)

If the odds for the event A are a:b, then Prob [A] = a/(a + b). Example 3.14 :

The odds of a fair coin coming up heads are 1:1; that it, is has a

probability of 1/2. The odds of a die showing a “1” are 5:1 against; that it, there is a probability of 5/6 that a “1” does not appear.

3.3.6

Conditional probability

The conditional probability of A given the event B has occurred is Prob [A | B] =

Prob [A ∩ B] , Prob [B]

Prob [B] > 0 .

(3.14)

(1) If Prob [A1 ∩ A2 ∩ · · · ∩ An−1 ] > 0 then Prob [A1 ∩ A2 ∩ · · · ∩ An ] =Prob [A1 ] · Prob [A2 | A1 ] · Prob [A3 | A1 ∩ A2 ]

(3.15)

· · · Prob [An | A1 ∩ A2 ∩ · · · ∩ An−1 ]. (2) If A ⊂ B, then Prob [A | B] = Prob [A]/Prob [B] and Prob [B | A] = 1. (3) Prob [A | B] = 1 − Prob [A | B]. Example 3.15 : A local bank offers loans for three purposes: home (H), automobile (A), and personal (P), and two different types: fixed rate (FR) and adjustable rate (ADJ). The joint probability table given below presents the proportions for the various

c 2000 by Chapman & Hall/CRC 

categories of loan and type: Loan Purpose

Type

H .27 .13

FR ADJ

A .19 .09

P .14 .18

Suppose a person who took out a loan at this bank is selected at random. (a) What is the probability the person has an automobile loan and it is fixed rate? (b) Given the person has an adjustable rate loan, what is the probability it is for a home? (c) Given the person does not have a personal loan, what is the probability it is adjustable rate? Solution: (S1) Prob [A ∩ FR] = .19 (S2) Prob [H | ADJ] = Prob [H ∩ ADJ]/Prob [ADJ] = .13/.4 = .325       (S3) Prob ADJ | P = Prob ADJ ∩ P /Prob P = .22/.68 = .3235

3.3.7

The multiplication rule Prob [A ∩ B] = Prob [A | B] · Prob [B] , = Prob [B | A] · Prob [A] ,

3.3.8

Prob [B] = 0 Prob [A] = 0

(3.16)

The law of total probability

Suppose A1 , A2 , . . . , An is a collection of mutually exclusive, exhaustive events, Prob [Ai ] = 0, i = 1, 2, . . . , n. For any event B: Prob [B] =

n 

Prob [B | Ai ] · Prob [Ai ] .

(3.17)

i=1

Example 3.16 : A ball drawing strategy. There are two urns. A marked ball may be in urn 1 (with probability p) or urn 2 (with probability 1 − p). The probability of drawing the marked ball from the urn it is in is r (with r < 1). After a ball is drawn from an urn, it is replaced. What is the best way to use n draws of balls from any urn so that the probability of drawing the marked ball is largest? Solution: (S1) Let the event of selecting the marked ball be A. (S2) Let Hi be the hypothesis that the marked ball is in urn i. (S3) By assumption, Prob [H1 ] = p and Prob [H2 ] = 1 − p. (S4) Choose m balls from urn 1, and n − m balls from urn 2. The conditional probabilities are then: Prob [A | H1 ] = 1 − (1 − r)m ,

c 2000 by Chapman & Hall/CRC 

Prob [A | H2 ] = 1 − (1 − r)n−m

(3.18)

so that (using the law of total probability) Prob [A] = Prob [H1 ] · Prob [A | H1 ] + Prob [H2 ] · Prob [A | H2 ]   = p [1 − (1 − r)m ] + (1 − p) 1 − (1 − r)n−m . (S5) Differentiating this with respect to m, and setting r)2m−n = (1 − p)/p or   ln 1−p p n . m= + 2 2 ln(1 − r)

3.3.9

dProb[A] dm

(3.19)

= 0 results in (1 −

(3.20)

Bayes’ theorem

Suppose A1 , A2 , . . . , An is a collection of mutually exclusive, exhaustive events, Prob [Ai ] = 0, i = 1, 2, . . . , n. For any event B such that Prob [B] = 0: Prob [Ak | B] =

Prob [B | Ak ] · Prob [Ak ] Prob [Ak ∩ B] =  , n Prob [B] Prob [B | Ai ] · Prob [Ai ]

(3.21)

i=1

for k = 1, 2, . . . , n. Example 3.17 : A large manufacturer uses three different trucking companies (A, B, and C) to deliver products. The probability a randomly selected shipment is delivered by each company is Prob [A] = .60,

Prob [B] = .25,

Prob [C] = .15

Occasionally, a shipment is damaged (D) in transit. Prob [D | A] = .01,

Prob [D | B] = .005,

Prob [D | C] = .015

Suppose a shipment is selected at random. (a) Find the probability the shipment is sent by trucking company B and is damaged. (b) Find the probability the shipment is damaged. (c) Suppose a shipment arrives damaged. What is the probability it was shipped by company B? Solution: (S1) Prob [B ∩ D] = Prob [B] · Prob [D | B] = (.25)(.005) = .00125 (S2) Prob [D] = Prob [A ∩ D] + Prob [B ∩ D] + Prob [C ∩ D] = Prob [A] · Prob [D | A] + Prob [B] · Prob [D | B] + Prob [B] · Prob [D | B] = (.60)(.01) + (.25)(.005) + (.15)(.015) = .0095 (S3) Prob [B | D] = Prob [B ∩ D]/Prob [D] = .00125/.0095 = .1316

3.3.10

Independence

(1) A and B are independent events if Prob [A | B] = Prob [A] or, equivalently, if Prob [B | A] = Prob [B]. c 2000 by Chapman & Hall/CRC 

(2) A and B are independent events if and only if Prob [A ∩ B] = Prob [A] · Prob [B]. (3) A1 , A2 , . . . , An are pairwise independent events if Prob [Ai ∩ Aj ] = Prob [Ai ] · Prob [Aj ]

for every pair i, j with i = j. (3.22)

(4) A1 , A2 , . . . , An are mutually independent events if for every k, k = 2, 3, . . . , n, and every subset of indices i1 , i2 , . . . , ik , Prob [Ai1 ∩ Ai2 ∩ · · · ∩ Aik ] = Prob [Ai1 ] · Prob [Ai2 ] · · · Prob [Aik ] . 3.4

(3.23)

RANDOM VARIABLES

Given a sample space S, a random variable is a function with domain S and range some subset of the real numbers. A random variable is discrete if it can assume only a finite or countably infinite number of values. A random variable is continuous if its set of possible values is an entire interval of numbers. Random variables are denoted by upper-case letters, for example X. 3.4.1

Discrete random variables

3.4.1.1

Probability mass function

The probability distribution or probability mass function (pmf), p(x), of a discrete random variable is a rule defined for every number x by p(x) = Prob [X = x] such that (1) p(x) ≥ 0; and  (2) p(x) = 1 x

3.4.1.2

Cumulative distribution function

The cumulative distribution function (cdf), F (x), for a discrete random variable X with pmf p(x) is defined for every number x:  F (x) = Prob [X ≤ x] = p(y) . (3.24) y|y≤x

(1)

lim F (x) = 0

x→−∞

(2) lim F (x) = 1 x→∞

(3) If a and b are real numbers such that a < b, then F (a) ≤ F (b). (4) Prob [a ≤ X ≤ b] = Prob [X ≤ b] − Prob [X < a] = F (b) − F (a− ) where a− is the first value X assumes less than a. Valid for a, b, ∈ R and a < b.

c 2000 by Chapman & Hall/CRC 

3.4.2

Continuous random variables

3.4.2.1

Probability density function

The probability distribution or probability density function (pdf) of a continuous random variable X is a real-valued function f (x) such that  b f (x) dx, a, b ∈ R, a ≤ b. (3.25) Prob [a ≤ X ≤ b] = a

(1) f (x) ≥ 0 for −∞ < x < ∞  ∞ (2) f (x) dx = 1 −∞

(3) Prob [X = c] = 0 3.4.2.2

for c ∈ R.

Cumulative distribution function

The cumulative distribution function (cdf), F (x), for a continuous random variable X is defined by  x F (x) = Prob [X ≤ x] = f (y) dy − ∞ < x < ∞. (3.26) −∞

(1)

lim F (x) = 0

x→−∞

(2) lim F (x) = 1 x→∞

(3) If a and b are real numbers such that a < b, then F (a) ≤ F (b). (4) Prob [a ≤ X ≤ b] = Prob [X ≤ b] − Prob [X < a] = F (b) − F (a), a, b, ∈ R and a < b. (5) The pdf f (x) may be found from the cdf: f (x) = 3.4.3

dF (x) dx

whenever the derivative exists.

(3.27)

Random functions

A random function of a real variable t is a function, denoted X(t), that is a random variable for each value of t. If the variable t can assume any value in an interval, then X(t) is called a stochastic process; if the variable t can only assume discrete values then X(t) is called a random sequence. 3.5 3.5.1

MATHEMATICAL EXPECTATION Expected value

(1) If X is a discrete random variable with pmf p(x): (a) The expected value of X is  E [X] = µ = xp(x) , x c 2000 by Chapman & Hall/CRC 

(3.28)

(b) The expected value of a function g(X) is  E [g(X)] = µg(X) = g(x)p(x) .

(3.29)

x

(2) If X is a continuous random variable with pdf f (x): (a) The expected value of X is  ∞ E [X] = µ = xf (x) dx ,

(3.30)

−∞

(b) The expected value of a function g(X) is  ∞ E [g(X)] = µg(X) = g(x)f (x) dx .

(3.31)

−∞

(3) Jensen’s inequality Let h(x) be a function such that

d2 [h(x)] ≥ 0, then dx2

E [h(X)] ≥ h(E [X]). (4) Theorems: (a) E [aX + bY ] = aE [X] + bE [Y ] (b) E [X · Y ] = E [X] · E [Y ] if X and Y are independent. 3.5.2

Variance

The variance of a random variable X is   (x − µ)2 p(x)     x σ 2 = E (X − µ)2 =  ∞    (x − µ)2 f (x) dx −∞

The standard deviation of X is σ = 3.5.2.1



if X is discrete (3.32) if X is continuous

σ2 .

Theorems

Suppose X is a random variable, and a, b are constants.   2 (1) σX = E X 2 − (E [X])2 . 2 2 (2) σaX = a2 · σX ,

(3)

2 σX+b

=

2 σX

σaX = |a| · σX .

.

2 2 (4) σaX+b = a2 · σX ,

3.5.3 3.5.3.1

σaX+b = |a| · σX .

Moments Moments about the origin

The moments about the origin completely characterize a probability distribution. The rth moment about the origin, r = 0, 1, 2, . . . , of a random variable c 2000 by Chapman & Hall/CRC 

X is

 r  x p(x)   x  r µr = E [X ] =  ∞    xr f (x) dx

if X is discrete (3.33) if X is continuous

−∞

The first moment about the origin is the mean of the random variable: µ1 = E [X] = µ. 3.5.3.2

Moments about the mean

The rth moment about the mean, r = 0, 1, 2, . . . , of a random variable X is  r  if X is discrete  (x − µ) p(x)  x r µr = E [(X − µ) ] =  ∞ (3.34)  r   (x − µ) f (x) dx if X is continuous −∞

The second moment about the mean is the variance of the random variable: µ2 = E [(X − µ)r ] = σ 2 = µ2 − µ2 . 3.5.3.3

Factorial moments

The rth factorial moment, r = 0, 1, 2, . . . , of a random variable is  [r]  x p(x) if X is discrete  #  " x [r] µ[r] = E X =  ∞    x[r] f (x) dx if X is continuous

(3.35)

−∞

[r]

where x

is the factorial expression x[r] = x(x − 1)(x − 2) · · · (x − r + 1) .

3.5.4 3.5.4.1

(3.36)

Generating functions Moment generating function

The moment generating function (mgf) of a random variable X, where it exists, is  tx  e p(x) if X is discrete   tX   x mX (t) = E e =  ∞ (3.37)  tx   e f (x) dx if X is continuous −∞

c 2000 by Chapman & Hall/CRC 

The moment generating function mX (t) is the expected value of etX and may be written as   mX (t) = E etX   (Xt)2 (Xt)3 = E 1 + Xt + + + ··· (3.38) 2! 3! t2 t3 + µ3 + · · · 2! 3! The moments µr are the coefficients of tr /r! in equation (3.38). Therefore, mX (t) generates the moments since the rth derivative of mX (t) evaluated at t = 0 yields µr : $ dr mX (t) $$ µr = mx(r) (0) = (3.39) dtr $t=0 = 1 + µ1 t + µ2

Theorems: Suppose mX (t) is the moment generating function for the random variable X and a, b are constants. (1) maX (t) = mX (at) (2) mX+b (t) = ebt · mX (t) (3) m(X+b)/a (t) = e(b/a)t · mX (t/a) (4) If X1 , X2 , . . . , Xn are independent random variables and Y = X1 +X2 + · · · + Xn , then mY (t) = [mX (t)]n . The moment generating function for X − µ is mX−µ (t) = e−µt · mX (t) .

(3.40)

Equation (3.40) may be used to generate the moments about the mean for the random variable X: $ dr (e−µt · mX (t)) $$ (r) µr = mX−µ (0) = (3.41) $ dtr t=0 3.5.4.2

Factorial moment generating functions

The factorial moment generating function of a random variable X is  x  t p(x) if X is discrete   X  x P (t) = E t =  ∞    tx f (x) dx if X is continuous

(3.42)

−∞

th

The r derivative of the function P (in equation (3.42)) with respect to t, evaluated at t = 1 is the rth factorial moment. Therefore, the function P

c 2000 by Chapman & Hall/CRC 

generates the factorial moments: µ[r] = P (r) (1) =

$ dr P (t) $$ . dtr $t=1

(3.43)

In particular: 1 = P (1) µ = P  (1) 

“conservation of probability” (3.44) 



2

σ = P (1) + P (1) − [P (1)] 2

3.5.4.3

Factorial moment generating function theorems

Theorems: Suppose PX (t) is the factorial moment generating function for the random variable X and a, b are constants. (1) PaX (t) = PX (ta ) (2) PX+b (t) = tb · PX (t) (3) P(X+b)/a (t) = tb/a · PX (t1/a ) (4) PX (t) = mX (ln t), where mx (t) is the moment generating function for X. (5) If X1 , X2 , . . . , Xn are independent random variables with factorial moment generating function PX (t) and Y = X1 + X2 + · · · + Xn , then PY (t) = [PX (t)]n . 3.5.4.4

Cumulant generating function

Let mX (t) be a moment generating function. If ln mX (t) can be expanded in the form t2 t3 tr + κ3 + · · · + κ r + · · · , (3.45) 2! 3! r! then c(t) is the cumulant generating function (or semi–invariant generating function). The constants κr are the cumulants (or semi–invariants) of the distribution. The rth derivative of c with respect to t, evaluated at 0 is the rth cumulant. The function c generates the cumulants: $ dr c(t) $$ (r) κr = c (0) = . (3.46) dtr $ c(t) = ln mX (t) = κ1 t + κ2

t=0

Marcienkiewicz’s theorem states that either all but the first two cumulants vanish (i.e., it is a normal distribution) or there are an infinite number of non-vanishing cumulants.

c 2000 by Chapman & Hall/CRC 

3.5.4.5

Characteristic function

The characteristic function exists for every random variable X and is defined by  itx  e p(x) if X is discrete   itX   x φ(t) = E e =  ∞ (3.47)  itx   e f (x) dx if X is continuous −∞

where t is a real number and i2 = −1. The rth derivative of φ with respect to t, evaluated at t = 0 is ir µr . Therefore, the characteristic function also generates the moments: $ dr φ(t) $$ r  (r) i µr = φ (0) = . (3.48) dtr $ t=0

3.6

MULTIVARIATE DISTRIBUTIONS

Note that the specialization to bivariate distributions is on page 45. 3.6.1

Discrete case

A n-dimensional random variable (X1 , X2 , . . . , Xn ) is n-dimensional discrete if it can assume only a finite or countably infinite number of values. The joint probability distribution, joint probability mass function, or joint density, for (X1 , X2 , . . . , Xn ) is p(x1 , x2 , . . . , xn ) = Prob [X1 = x1 , X2 = x2 , . . . , Xn = xn ] ∀(x1 , x2 , . . . , xn ) .

(3.49)

Suppose E is a subset of values the random variable may assume. The probability the event E occurs is Prob [E] = Prob [(X1 , X2 , . . . , Xn ) ∈ E]   = ··· p(x1 , x2 , . . . , xn ) .

(3.50)

(x1 ,x2 ,...,xn )∈E

The cumulative distribution function for (X1 , X2 , . . . , Xn ) is    F (x1 , x2 , . . . , xn ) = ··· p(x1 , x2 , . . . , xn ) . t1 |t1 ≤x1 t2 |t2 ≤x2

3.6.2

(3.51)

tn |tn ≤xn

Continuous case

The continuous random variables X1 , X2 , . . . , Xn are jointly distributed if there exists a function f such that f (x1 , x2 , . . . , xn ) ≥ 0 for −∞ < xi < ∞,

c 2000 by Chapman & Hall/CRC 

i = 1, 2, . . . , n, and for any event E Prob [E] = Prob [(X1 , X2 , . . . , Xn ) ∈ E]   = · · · f (x1 , x2 , . . . , xn ) dxn · · · dx1

(3.52)

E

where f is the joint distribution function or joint probability density function for the random variables X1 , X2 , . . . , Xn . The cumulative distribution function for X1 , X2 , . . . , Xn is  x1  x2  xn F (x1 , x2 , . . . , xn ) = ··· f (x1 , x2 , . . . , xn ) dxn · · · dx1 . (3.53) −∞

−∞

−∞

Given the cumulative distribution function, F , the probability density function may be found by f (x1 , x2 , . . . , xn ) =

∂n F (x1 , x2 , . . . , xn ) ∂x1 ∂x2 · · · ∂xn

(3.54)

wherever the partials exist. 3.6.3

Expectation

Let g(X1 , X2 , . . . , Xn ) be a function of the random variables X1 , . . . , Xn . The expected value of g(X1 , X2 , . . . , Xn ) is   E [g(X1 , X2 , . . . , Xn )] = ··· g(x1 , x2 , . . . , xn )p(x1 , x2 , . . . , xn ) x1

x2

xn

(3.55)

if X1 , X2 , . . . , Xn are discrete, and E [g(X1 , X2 , . . . , Xn )] =  ∞ ∞  ∞ ··· g(x1 , . . . , xn )f (x1 , . . . , xn ) dxn · · · dx1 −∞

−∞

−∞

if X1 , X2 , . . . , Xn are continuous. If c1 , c2 , . . . , cn are constants, then % n & n   E ci gi (X1 , X2 , . . . , Xn ) = ci E [gi (X1 , X2 , . . . , Xn )] . i=1

3.6.4

(3.56)

(3.57)

i=1

Moments

If X1 , X2 , . . . , Xn are jointly distributed, the rth moment of Xi is   E [Xir ] = ··· xri p(x1 , x2 , . . . , xn ) x1

x2

if X1 , X2 , . . . , Xn are discrete, and  ∞ ∞  ∞ r E [Xi ] = ··· xri f (x1 , x2 , . . . , xn ) dxn · · · dx1 −∞

−∞

c 2000 by Chapman & Hall/CRC 

−∞

(3.58)

xn

(3.59)

if X1 , X2 , . . . , Xn are continuous. The joint (product) moments about the origin are   E [X1r1 X2r2 · · · Xnrn ] = ··· xr11 xr22 · · · xrnn p(x1 , x2 , . . . , xn ) x1

x2

(3.60)

xn

if X1 , X2 , . . . , Xn are discrete, and E [X1r1 X2r2 · · · Xnrn ]  ∞ ∞  ∞ = ··· xr11 xr22 · · · xrnn f (x1 , x2 , . . . , xn ) dxn · · · dx1 −∞

−∞

(3.61)

−∞

if X1 , X2 , . . . , Xn are continuous. The value r = r1 + r2 + · · · + rn is the order of the moment. If E [Xi ] = µi , then the joint moments about the mean are E [(X1 − µ1 )r1 (X2 − µ2 )r2 · · · (Xn − µn )rn ] = (3.62)   ··· (x1 − µ1 )r1 (x2 − µ2 )r1 · · · (xn − µn )rn p(x1 , x2 , . . . , xn ) x1

x2

xn

if the X1 , X2 , . . . , Xn are discrete, and E [(X1 − µ1 )r1 (X2 − µ2 )r2 · · · (Xn − µn )rn ] = (3.63)  ∞ ∞  ∞ r1 rn ··· (x1 − µ1 ) · · · (xn − µn ) f (x1 , . . . , xn ) dxn · · · dx1 , −∞

−∞

−∞

if the X1 , X2 , . . . , Xn are continuous. 3.6.5

Marginal distributions

Let X1 , X2 , . . . , Xn be a collection of random variables. The marginal distribution of a subset of the random variables X1 , X2 , . . . , Xk (with (k < n)) is    g(x1 , x2 , . . . , xk ) = ··· p(x1 , x2 , . . . , xn ) (3.64) xk+1 xk+2

if X1 , X2 , . . . , Xn are discrete, and  ∞ ∞  g(x1 , x2 , . . . , xk ) = ··· −∞

−∞



−∞

xn

f (x1 , x2 , . . . , xn ) dxk+1 dxk+2 · · · dxn (3.65)

if X1 , X2 , . . . , Xn are continuous. Example 3.18 : The joint density functions g(x, y) = x+y and h(x, y) = (x+ 12 )(y+ 12 )

when 0 ≤ x ≤ 1 and 0 ≤ y ≤ 1 have the same marginal distributions. Using equation

c 2000 by Chapman & Hall/CRC 

(3.65): $y=1 y 2 $$ 1 =x+ 2 $y=0 2 0   2 $y=1  1 1 1 y $$ y h(x, y) dy = x + =x+ + hx (x) = 2 2 2 $y=0 2 0 

gx (x) =



1

g(x, y) dy =

xy +

(3.66)

and, by symmetry, gy (y) has the same form as gx (x) (likewise for hy (y) and hx (x)).

3.6.6

Independent random variables

Let X1 , X2 , . . . , Xn be a collection of discrete random variables with joint probability distribution function p(x1 , x2 , . . . , xn ). Let gXi (xi ) be the marginal distribution for Xi . The random variables X1 , X2 , . . . , Xn are independent if and only if p(x1 , x2 , . . . , xn ) = gX1 (x1 ) · gX2 (x2 ) · · · gXn (xn ) .

(3.67)

Let X1 , X2 , . . . , Xn be a collection of continuous random variables with joint probability distribution function f (x1 , x2 , . . . , xn ). Let gXi (xi ) be the marginal distribution for Xi . The random variables X1 , X2 , . . . , Xn are independent if and only if f (x1 , x2 , . . . , xn ) = gX1 (x1 ) · gX2 (x2 ) · · · gXn (xn ) .

(3.68)

Example 3.19 : Suppose X1 , X2 , and X3 are independent random variables with probability density functions given by ' −x1 e x1 > 0 gX1 (x1 ) = 0 elsewhere ' −3x2 ' −7x3 3e x2 > 0 7e x3 > 0 gX3 (x3 ) = gX2 (x2 ) = 0 elsewhere 0 elsewhere Using equation (3.68), the joint probability distribution for X1 , X2 , X3 is f (x1 , x2 , x3 ) = gX1 (x1 ) · gX2 (x2 ) · gX3 (x3 ) = (e−x1 ) · (3e−3x2 ) · (7e−7x3 ) = 21e−x1 −3x2 −7x3

3.6.7

[x1 > 0, x2 > 0, x3 > 0].

Conditional distributions

Let X1 , X2 , . . . , Xn be a collection of random variables. The conditional distribution of any subset of the random variables X1 , X2 , . . . , Xk given Xk+1 = xk+1 , Xk+2 = xk+2 , . . . , Xn = xn is p(x1 , x2 , . . . , xk | xk+1 , xk+2 . . . , xn ) =

c 2000 by Chapman & Hall/CRC 

p(x1 , x2 , . . . , xn ) g(xk+1 , xk+2 , . . . , xn )

(3.69)

if X1 , X2 , . . . , Xn are discrete with joint distribution function p(x1 , x2 , . . . , xn ) and Xk+1 , Xk+2 , . . . , Xn have marginal distribution g(xk+1 , xk+2 , . . . , xn ) = 0, and f (x1 , x2 , . . . , xk | xk+1 , xk+2 , . . . , xn ) =

f (x1 , x2 , . . . , xn ) g(xk+1 , xk+2 , . . . , xn )

(3.70)

if X1 , X2 , . . . , Xn are continuous with joint distribution function f (x1 , x2 , . . . , xn ) and Xk+1 , Xk+2 , . . . , Xn have marginal distribution g(xk+1 , xk+2 , . . . , xn ) = 0. Example 3.20 : Suppose X1 , X2 , X3 have a joint distribution function given by ( f (x1 , x2 , x3 ) =

(x1 + x2 )e−x3

when 0 < x1 < 1, 0 < x2 < 1, x3 > 0

0

elsewhere

The marginal distribution of X2 is  1 ∞ g(x2 ) = (x1 + x2 )e−x3 dx3 dx1 0



0 1

=

(x1 + x2 ) dx1 = 0

1 + x2 , 2

0 < x2 < 1.

The conditional distribution of X1 , X3 given X2 = x2 is f (x1 , x3 | x2 ) =

f (x1 , x2 , x3 ) (x1 + x2 )e−x3 . = 1 g(x2 ) + x2 2

If X2 = 3/4, then  f (x1 , x3 |

3.6.8

3/4)

=

x1 + 1 2

3 4



+

e−x3

3 4

=

4 5

 x1 +

3 4



e−x3 ,

0 < x1 < 1, x3 > 0.

Variance and covariance

Let X1 , X2 , . . . , Xn be a collection of random variables. The variance, σii , of Xi is   σii = σi2 = E (Xi − µi )2 (3.71) and the covariance, σij , of Xi and Xj is σij = ρij σi σj = E [(Xi − µi )(Xj − µj )]

(3.72)

where ρij is the correlation coefficient and σi and σj are the standard deviations of Xi and Xj , respectively. Theorems: (1) If X1 , X2 , . . . , Xn are independent, then E [X1 X2 · · · Xn ] = E [X1 ]E [X2 ] · · · E [Xn ] .

c 2000 by Chapman & Hall/CRC 

(3.73)

(2) For two random variables Xi and Xj : σij = E [Xi Xj ] − E [Xi ]E [Xj ] .

(3.74)

(3) If Xi and Xj are independent random variables, then σij = 0. (4) Two variables may be dependent and have zero covariance. For example, let X take the four values {−2, −1, 1, 2} with equal probability. If Y = X 2 then the covariance of X and Y is zero. The correlation function of a random function (see page 34) is KX (t1 , t2 ) = E [[X ∗ (t1 ) − µX ∗ (t1 )] [X(t2 ) − µX (t2 )]] where



denotes the complex conjugate. If X(t) is stationary then KX (t1 , t2 ) = KX (t1 − t2 ) and

3.6.9

(3.75)

µX (t) = constant.

(3.76)

Correlation coefficient

The correlation coefficient, defined by (see equation (3.72)) σij ρij = σi σj

(3.77)

is no greater than one in magnitude: |ρij | ≤ 1. Figure 3.5 contains 4 data sets of 100 points each; the correlation coefficients vary from −0.7 to 0.99. ···· ··· ·· · · ·· · · · · · ···· ······ ···· · ··· ······· · · · · · ·· · ·· · · ·· ·· ··· ·· ·· ··· · · · ·· · · ·· ·· · · ·· ··· · · · · ·· ·

· ·· · · · ··· · · · · ·· ·· · ··· · · ···· · ·· ··· ·· · ·· ·· ···· · · ·· ·· ·· · · ··· ·· · · · ·· · ·· · ·· · · · ·· ·· ·· · · · · · · ·· · · ·· ·

ρ=−0.71

· · ··· · · · · · · · ··· ··· · · · ·· ·· · · ···· · · · · ·· ·· ·· · ··· ·· ·· · ······· ··· · ···· · · ·· · · · ···· · · ·· ··· · ·· ·· ···· · · ·

ρ=0.09

····· ······ · · · · · ···· · · ···· · ········ ·· · · · ············ · ·········

ρ=0.28

ρ=0.99

Figure 3.5: Data sets illustrating different correlation coefficients. Example 3.21 : The correlation coefficient of the first 100 integers {1, 2, 3 . . . } and the first 100 squares {1, 4, 9 . . . } is 0.96885. 3.6.10

Moment generating function

Let X1 , X2 , . . . , Xn be a collection of random variables. The joint moment generating function is     (3.78) m(t1 , t2 , . . . , tn ) = m(t) = E et1 X1 +t2 X2 +···+tn Xn = E et·X c 2000 by Chapman & Hall/CRC 

if it exists for all values of ti such that |ti | < h2 (for some value h). The rth moment of Xi may be obtained (generated) by differentiating m(t1 , t2 , . . . , tn ) r times with respect to ti , and then evaluating the result with all t’s equal to zero: $ ∂ r m(t1 , t2 , . . . , tn ) $$ E [Xir ] = (3.79) $ ∂tri (t1 ,t2 ,...,tn )=(0,0,...,0) The rth joint moment, r = r1 +r2 +· · ·+rn , may be obtained by differentiating m(t1 , t2 , . . . , tn ) r1 times with respect to t1 , r2 times with respect to t2 , . . . , and rn times with respect to tn , and then evaluating the result with all t’s equal to zero: $ ∂ r m(t1 , t2 , . . . , tn ) $$ r1 r2 rn E [X1 X2 · · · Xn ] = (3.80) ∂tr11 ∂tr22 · · · ∂trnn $(t1 ,t2 ,...,tn )=(0,0,...,0) 3.6.11

Linear combination of random variables

Let X1 , X2 , . . . , Xm and Y1 , Y2 , . . . , Yn be random variables, let a1 , a2 , . . . , am and b1 , b2 , . . . , bn be constants, and let U and V be the linear combinations U=

m 

ai Xi ,

i=1

V =

n 

bj Yj .

(3.81)

j=1

Theorems: (1) E [U ] =

m 

ai E [Xi ] .

i=1

(2) σi2 =

m 

a2i σi2 + 2

i=1



ai aj σij ,

i
where the double sum extends over all pairs (i, j) with i < j. (3) If the random variables X1 , X2 , . . . , Xm are independent, then m  2 σU = a2i σi2 . i=1

(4) σU V =

m  n 

ai bj σij .

i=1 j=1

3.6.12 3.6.12.1

Bivariate distribution Joint probability distribution

(a) Discrete case Let X and Y be discrete random variables. The joint probability distribution for X and Y is p(x, y) = Prob [X = x, Y = y] c 2000 by Chapman & Hall/CRC 

∀(x, y) .

(3.82)

(1) For any subset E consisting of pairs (x, y),  Prob [(X, Y ) ∈ E] = p(x, y).

(3.83)

(x,y)∈E

(2) p(x, y) ≥ 0 ∀(x, y).  (3) p(x, y) = 1. x

y

(b) Continuous case Let X and Y be continuous random variables. The joint probability distribution for X and Y is a function f (x, y) such that for any twodimensional set E  Prob [(X, Y ) ∈ E] = f (x, y) dx dy . (3.84) E

(1) If E is a rectangle {(x, y) | a ≤ x ≤ b, c ≤ y ≤ d}, then Prob [(X, Y ) ∈ E] = Prob [a ≤ X ≤ b, c ≤ Y ≤ d]  b d = f (x, y) dy dx . a

(3.85)

c

(2) f (x, y) ≥ 0 ∀(x, y).  ∞ ∞ (3) f (x, y) dx dy = 1. −∞

3.6.12.2

−∞

Cumulative distribution function

For any two random variables X and Y the cumulative distribution function is F (x, y) = Prob [X ≤ x, Y ≤ y]:    p(x, y) if X and Y are discrete    x|x≤a y|y≤b F (a, b) =  a  b (3.86)    f (x, y) dy dx if X and Y are continuous  −∞

−∞

Properties: (1)

lim

F (x, y) = lim F (x, y) = lim F (x, y) = 0.

(x,y)→(−∞,−∞)

(2)

lim

x→−∞

y→−∞

F (x, y) = 1.

(x,y)→(∞,∞)

(3) If a ≤ b and c ≤ d, then Prob [a < X ≤ b, c < Y ≤ d] = F (b, d) − F (b, c) − F (a, d) + F (a, c) (3.87) ≥ 0. (4) Given the cumulative distribution function, F , for the continuous random variables X and Y , the probability density function may be found c 2000 by Chapman & Hall/CRC 

by f (x, y) =

∂2 F (x, y) ∂x∂y

(3.88)

wherever the partials exist. 3.6.12.3

Marginal distributions

Let X and Y be discrete random variables with joint distribution p(x, y). The marginal distributions for X and Y are   pX (x) = p(x, y) , pY (y) = p(x, y) . (3.89) y

x

Let X and Y be continuous random variables with joint distribution f (x, y). The marginal distributions for X and Y are  ∞ fX (x) = f (x, y) dy, −∞ < x < ∞  fY (y) = 3.6.12.4

−∞ ∞

(3.90) f (x, y) dx,

−∞

−∞ < y < ∞.

Conditional distributions

Let X and Y be discrete random variables with joint distribution p(x, y) and let pY (y) be the marginal distribution for Y . The conditional distribution for X given Y = y is p(x | y) =

p(x, y) , pY (y)

pY (y) = 0.

(3.91)

Let pX (x) be the conditional distribution for X. The conditional distribution for Y given X = x is p(y | x) =

p(x, y) , pX (x)

pX (x) = 0.

(3.92)

Let X and Y be continuous random variables with joint distribution f (x, y) and let fY (y) be the marginal distribution for Y . The conditional distribution for X given Y = y is f (x | y) =

f (x, y) , fY (y)

fY (y) = 0.

(3.93)

Let fX (x) be the conditional distribution for X. The conditional distribution for Y given X = x is f (y | x) =

c 2000 by Chapman & Hall/CRC 

f (x, y) , fX (x)

fX (x) = 0.

(3.94)

3.6.12.5

Conditional expectation

Let X and Y be random variables and let g(X) be a function of X. The conditional expectation of g(X) given Y = y is  g(x)p(x | y) if X and Y are discrete    x E [g(X) | y] =  ∞ (3.95)    g(x)f (x | y) dx if X and Y are continuous −∞

Properties: (1) The conditional mean, or conditional expectation, of X given Y = y is   xp(x | y) if X and Y are discrete    x µX|y = E [X | y] = (3.96)  ∞    xf (x | y) dx if X and Y are continuous −∞

(2) The conditional variance of X given Y = y is     2 σX|y = E (X − µX|y )2 | y = E X 2 | y − µ2X|y .

(3.97)

(3) E [X] = E [E [X | Y ]] 3.7

INEQUALITIES

1. Bienaym´e–Chebyshev’s inequality: If E [|X|r ] < ∞ for all r > 0 (r not necessarily an integer) then, for every a > 0 Prob [|X| ≥ a] ≤

E [|X|r ] ar

(3.98)

2. Bienaym´e–Chebyshev’s inequality (generalized ): Let g(x) be a nondecreasing nonnegative function defined on (0, ∞). Then, for a ≥ 0, Prob [|X| ≥ a] ≤

E [g(|X|)] g(a)

(3.99)

3. Cauchy–Schwartz inequality: Let X and Y be random variables in     which E Y 2 and E Z 2 exist, then     (E [Y Z])2 ≤ E Y 2 E Z 2 (3.100) 4. Chebyshev inequality:  Let c be any real number and let X be a random variable for which E (X − c)2 is finite. Then for every 3 > 0 the following holds  1  Prob [|X − c| ≥ 3] ≤ 2 E (X − c)2 (3.101) 3 c 2000 by Chapman & Hall/CRC 

5. Chebyshev inequality (one-sided): Let X be a random variable with zero mean (i.e., E [X] = 0) and variance σ 2 . Then for any positive a Prob [X > a] ≤

σ2 σ 2 + a2

(3.102)

6. Chernoff bound   n : This bound is useful for sums of random variables. Let Yn = i=1 Xi where each of the Xi is iid. Let mX (t) = E etX be the common moment generating function for the {Xi }, and define c(t) = log mX (t). Then 

Prob [Yn ≥ nc (t)] ≤ e−n[tc (t)−c(t)] 

Prob [Yn ≤ nc (t)] ≤ e−n[tc (t)−c(t)]

if t ≥ 0 if t ≤ 0

(3.103)

7. Jensen’s inequality: If E [X] exists, and if f (x) is a convex ∪ (“convex cup”) function, then E [f (X)] ≥ f (E [X])

(3.104)

8. Kolmogorov’s inequality: Let X1 , X2 , . . . , Xn be n independent random 2 is finite. Then, for all variables such that E [Xi ] = 0 and Var(Xi ) = σX i a > 0,    n σi2 Prob max |X1 + X2 + · · · + Xi | > a ≤ (3.105) i=1,...,n a2 i=1 9. Kolmogorov’s inequality: Let X1 , X2 , . . . , Xn be n mutually independent random variables with expectations µi = E [Xi ] and variances σk2 . Define the sums Sk = X1 + · · · + Xk so that mk = E [Sk ] = µ1 + · · · + µk and s2k = Var [Sk ] = σ12 + · · · + σk2 . For every t > 0, the probability of the simulteneous realization of the n inequalities |Sk − mk | < t sk is at least 1 − t

−2

(3.106)

. (When n = 1 this is Chebyshev’s inequality.)

10. Markov’s inequality: If X is random variable which takes only nonnegative values, then for any a > 0 Prob [X ≥ a] ≤

c 2000 by Chapman & Hall/CRC 

E [X] . a

(3.107)

CHAPTER 4

Functions of Random Variables Contents 4.1

4.2

4.3

4.4 4.5

4.6

4.7

Finding the probability distribution 4.1.1 Method of distribution functions 4.1.2 Method of transformations (one variable) 4.1.3 Method of transformations (many variables) 4.1.4 Method of moment generating functions Sums of random variables 4.2.1 Deterministic sums of random variables 4.2.2 Random sums of random variables Sampling distributions 4.3.1 Definitions 4.3.2 The sample mean 4.3.3 Central limit theorem 4.3.4 The law of large numbers 4.3.5 Laws of the iterated logarithm Finite population Theorems 4.5.1 Theorems: the chi–square distribution 4.5.2 Theorems: the t distribution 4.5.3 Theorems: the F distribution Order statistics 4.6.1 Definition 4.6.2 The first order statistic 4.6.3 The nth order statistic 4.6.4 The median 4.6.5 Joint distributions 4.6.6 Midrange and range 4.6.7 Uniform distribution: order statistics 4.6.8 Normal distribution: order statistics Range and studentized range 4.7.1 Probability integral of the range 4.7.2 Percentage points, studentized range

c 2000 by Chapman & Hall/CRC 

Let X1 , X2 , . . . , Xn be a collection of random variables with joint probability mass function p(x1 , x2 , . . . , xn ) (if the collection is discrete) or joint density function f (x1 , x2 , . . . , xn ) (if the collection is continuous). Suppose the random variable Y = Y (X1 , X2 , . . . , Xn ) is a function of X1 , X2 , . . . , Xn . Methods for finding the distribution of Y are presented below and sampling distributions are discussed in the following section. 4.1

FINDING THE PROBABILITY DISTRIBUTION

The following techniques may be used to determine the probability distribution for Y = Y (X1 , X2 , . . . , Xn ). 4.1.1

Method of distribution functions

Let X1 , X2 , . . . , Xn be a collection of continuous random variables. (1) Determine the region Y = y. (2) Determine the region Y ≤ y. (3) Compute F (y) = Prob [Y ≤ y] by integrating the joint density function f (x1 , x2 , . . . , xn ) over the region Y ≤ y. (4) Compute the probability density function for Y , f (y), by differentiating F (y): f (y) =

dF (y) . dy

(4.1)

Example 4.22 : Suppose the joint density function of X1 and X2 is given by ( f (x1 , x2 ) =

2

2

4x1 x2 e−(x1 +x2 ) 0

for x1 > 0, x2 > 0 elsewhere

 and Y = X12 + X22 . Find the cumulative distribution function for Y and the probability density function for Y . Solution: (S1) The region Y ≤ y is a quarter circle in quadrant I, shown shaded in Figure 4.1.

Figure 4.1: Integration region for example 4.22.

c 2000 by Chapman & Hall/CRC 

(S2) The cumulative distribution function for Y is given by   √ 2 2 y −x1

y

F (y) = 

y

=

2

2

4x1 x2 e−(x1 +x2 ) dx2 dx1

0

0

2

(4.2)

2

2x1 (e−x1 − e−y ) dx1

0

= 1 − (1 + y 2 )e−y

2

(S3) The probability density function for Y is given by 2

2

f (y) = F  (y) = −[(2y)e−y + (1 + y 2 )(−2y)e−y ] 2

= 2y 3 e−y ,

4.1.2

(4.3)

when y > 0

Method of transformations (one variable)

Let X be a continuous random variable with probability density function fX (x). If u(x) is differentiable and either increasing or decreasing, then Y = u(X) has probability density function fY (y) = fX (w(y)) · |w (y)|,

u (x) = 0

(4.4)

where x = w(y) = u−1 (y). Example 4.23 : Let X be a standard normal random variable, and let Y = X 2 . What is the distribution of Y ? Solution: (S1) Since X can be both positive and negative, two regions of X correspond to the same value of Y . (S2) The computation is  $ $ $ dy $ fy (y) = fx (x) + fx (−x) $$ $$ dx % & −x2 /2 −(−x)2 /2 e e 1 + √ + √ = √ 2 y 2π 2π

(4.5)

2 1 = √ e−x /2 2πy 1 e−y/2 = √ 2πy

which is the probability density function for a chi–square random variable with one degree of freedom.

Example 4.24 : Given two independent random variables X and Y with joint probability density f (x, y), let U = X/Y be the ratio distribution. The probability density is:  ∞ fU (u) = |x| f (x, ux) dx (4.6) −∞

c 2000 by Chapman & Hall/CRC 

If X and Y are normally distributed, then U has a Cauchy distribution. If X and Y are uniformly distributed on [0, 1], then   for u < 0 0 fU (u) = 1/2 (4.7) for 0 ≤ u ≤ 1   1 for u > 1 2u2

4.1.3

Method of transformations (two or more variables)

Let X1 and X2 be continuous random variables with joint density function f (x1 , x2 ). Let the functions y1 = u1 (x1 , x2 ) and y2 = u2 (x1 , x2 ) represent a one–to–one transformation from the x’s to the y’s and let the partial derivatives with respect to both x1 and x2 exist. The joint density function of Y1 = u1 (X1 , X2 ) and Y2 = u2 (X1 , X2 ) is g(y1 , y2 ) = f (w1 (y1 , y2 ), w2 (y1 , y2 )) · |J|

(4.8)

where y1 = u1 (x1 , x2 ) and y2 = u2 (x1 , x2 ) are uniquely solved for x1 = w1 (y1 , y2 ) and x2 = w2 (y1 , y2 ), and J is the determinant of the Jacobian $ $ $ ∂x1 ∂x1 $ $ $ $ ∂y1 ∂y2 $ $. (4.9) J = $$ $ $ ∂x2 ∂x2 $ $ $ ∂y1 ∂y2 This method of transformations may be extended to functions of n random variables. Let X1 , X2 , . . . , Xn be continuous random variables with joint density function f (x1 , x2 , . . . , xn ). Let the functions y1 = u1 (x1 , x2 , . . . , xn ), y2 = u2 (x1 , x2 , . . . , xn ), . . . , yn = un (x1 , x2 , . . . , xn ) represent a one–to–one transformation from the x’s to the y’s and let the partial derivatives with respect to x1 , x2 , . . . , xn exist. The joint density function of Y1 = u1 (X1 , X2 , . . . , Xn ), Y2 = u2 (X1 , X2 , . . . , Xn ), . . . , Yn = un (X1 , X2 , . . . , Xn ) is g(y1 , y2 , . . . , yn ) = f (w1 (y1 , . . . , yn ), . . . , wn (y1 , . . . , yn )) · |J|

(4.10)

where the functions y1 = u1 (x1 , x2 , . . . , xn ), y2 = u2 (x1 , x2 , . . . , xn ), . . . , yn = un (x1 , x2 , . . . , xn ) are uniquely solved for x1 = w1 (y1 , y2 , . . . , yn ), x2 = w2 (y1 , y2 , . . . , yn ), . . . , xn = wn (y1 , y2 , . . . , yn ) and J is the determinant of the Jacobian $ $ ∂x1 ∂x1 $ $ ∂x1 ··· $ $ $ ∂y1 ∂y2 ∂yn $ $ $ $ ∂x ∂x2 ∂x2 $$ $ 2 ··· $ $ ∂y2 ∂yn $ J = $$ ∂y1 (4.11) .. .. $$ .. $ .. . . . $ $ . $ $ $ ∂x ∂xn $$ $ n ∂xn ··· $ $ ∂y1 ∂y2 ∂yn c 2000 by Chapman & Hall/CRC 

Example 4.25 : Suppose the random variables X and Y are independent with probability density functions fX (x) and fY (y), then the probability density of their sum, Z = X + Y , is given by  ∞ fZ (z) = fX (t)fY (z − t) dt (4.12) −∞

Example 4.26 : Suppose the random variables X and Y are independent with probability density functions fX (x) and fY (y), then the probability density of their product, Z = XY , is given by  ∞ z 1 dt (4.13) fX (t)fY fZ (z) = t −∞ |t|

Example 4.27 : Two random variables X and Y have  a joint normal distribution. The probability density is f (x, y) =

1 exp 2πσ 2

x2 + y 2 . Find the probability density of 2σ 2

the system (R, Φ) if X = R cos Φ Y = R sin Φ

(4.14)

Solution:

$ $ $   $$ $ $ $ and $ ∂(x,y) = r. (S1) Use f (r, φ) = f x(r, φ), y(r, φ) $ ∂(x,y) ∂(r,φ) $ ∂(r,φ) $

(S2) Then f (r, φ) =

=

 2  r cos2 φ + r2 sin2 φ r exp − 2πσ 2 2σ 2 1 2π 

  r2 r exp − σ2 2σ 2  

fΦ (φ)

fR (r)

(4.15)

where fR (r) is a Rayleigh distribution and fΦ (φ) is a uniform distribution.

4.1.4

Method of moment generating functions

To determine the distribution of Y : (1) Determine the moment generating function for Y , mY (t). (2) Compare mY (t) with known moment generating functions. If mY (t) = mU (t) for all t, then Y and U have identical distributions. Theorems: (1) Let X and Y be random variables with moment generating functions mX (t) and mY (t), respectively. If mX (t) = mY (t) for all t, then X and Y have the same probability distributions.

c 2000 by Chapman & Hall/CRC 

(2) Let X1 , X2 , . . . , Xn be independent random variables and let Y = X1 + X2 + · · · + Xn , then mY (t) =

n )

mXi (t) .

(4.16)

i=1

4.2

SUMS OF RANDOM VARIABLES

4.2.1

Deterministic sums of random variables

If Y = X1 + X2 + · · · + Xn and (a) the X1 , X2 , . . . , Xn are independent random variables with factorial moment generating functions PXi (t), then PY (t) =

n )

PXi (t)

(4.17)

i=1

(b) the X1 , X2 , . . . , Xn are independent random variables with the same factorial moment generating function PX (t), then PY (t) = [PX (t)]n

(4.18)

(c) the X1 , X2 , . . . , Xn are independent random variables with characteristic functions φXi (t), then φY (t) =

n )

φXi (t)

(4.19)

i=1

(d) the X1 , X2 , . . . , Xn are independent random variables with the same characteristic function φX (t), then φY (t) = [φX (t)]n

(4.20)

Example 4.28 : What is the distribution of the sum of two normal random variables? Solution: (S1) Let X1 be N (µ1 , σ1 ) and let X2 be N (µ2 , σ2 ).

 (S2) The characteristic functions are (see page 148) φX1 (t) = exp µ1 it −   σ2 t φX2 (t) = exp µ2 it − 22 .

(S3) From equation (4.19) the characteristic function for Y = X1 + X2 is   (σ 2 + σ22 )t φY (t) = φX1 (t) · φX2 (t) = exp (µ1 + µ2 )it − 1 2

2 σ1 t 2

 and

(4.21)

(S4) This last expression is the characteristic function for a normal random variable with mean µY = µ1 + µ2 and variance of σY2 = σ12 + σ22 . (S5) Conclusion: the distribution of the sum of two normal random variables is normal; the means add and the variances add.

See section 3.6.11 for linear combinations of random variables. c 2000 by Chapman & Hall/CRC 

4.2.2

Random sums of random variables N If T = i=1 Xi where N is an integer valued random variable with factorial generating function PN (t), the {Xi } are discrete independent and identically distributed random variables with factorial generating function PX (t), and the {Xi } are independent of N , then the factorial generating function for T is PT (t) = PN (PX (t))

(4.22)

(If the {Xi } are continuous random variables, then φT (t) = PN (φX (t)).) Hence (using equation (3.44)) µT = µN µX

(4.23)

2 2 σT2 = µN σX + µX σN

Example 4.29 : A game is played as follows: There are two coins used to play the game. The probability of a head on the first coin is p1 and the probability of a head on the second coin is p2 . The first coin is tossed. If the resulting toss is a head, the game is over. If the outcome is a tail, then the second coin is tossed. If the second coin lands head up, a $1.00 payoff is made. There is no payoff for a tail. The first coin is tossed again and the game continues in this manner. What is the expected payoff for this game? Solution: (S1) In this game the number of rounds, N , has a geometric distribution, so that p1 t PN (t) = . 1 − (1 − p1 )t (S2) Let the random variable X be the payoff at each round. X has a Bernoulli distribution: PX (t) = (1 − p2 ) + p2 t. (S3) The generating function for the payoff is PT (t) = PN (PX (t)) =

p1 [(1 − p2 ) + p2 t] . 1 − (1 − p1 )[(1 − p2 ) + p2 t]

(4.24)

(S4) Using PT (t) in equation (3.44) or using equation (4.23) (with µN = 1/p1 and µX = p2 ) results in µT = p2 /p1 .

4.3 4.3.1

SAMPLING DISTRIBUTIONS Definitions

(1) The random variables X1 , X2 , . . . , Xn are a random sample of size n from an infinite population if X1 , X2 , . . . , Xn are independent and identically distributed (iid). (2) If X1 , X2 , . . . , Xn are a random sample, then the sample total and sample mean are T =

n  i=1

c 2000 by Chapman & Hall/CRC 

1 Xi , n i=1 n

Xi

and

X=

(4.25)

respectively. The sample variance is 1  (Xi − X)2 . n − 1 i=1 n

S2 = 4.3.2

(4.26)

The sample mean

Consider an infinite population with mean µ, variance σ 2 , skewness γ1 , and kurtosis γ2 . Using a sample of size n, the parameters describing the sample mean are: µx = µ σ2 n γ1 =√ n

σ σX = √ n γ2 γ2,X = n

2 σX =

γ1,X

(4.27)

When the population is finite and of size M , (M )

µx



2(M )

=

σx

σ2 M − N N M −1

(4.28)

If the underlying population is normal, then the sample mean X is normally distributed. 4.3.3

Central limit theorem

Let X1 , X2 , . . . , Xn be a random sample from an infinite population with mean µ and variance σ 2 . The limiting distribution of Z=

X −µ √ σ/ n

(4.29)

as n → ∞ is the standard normal distribution. The limiting distribution of T =

n 

Xi

(4.30)

i=1

as n → ∞ is normal with mean nµ and variance nσ 2 . 4.3.4

The law of large numbers

Let X1 , X2 , . . . , Xn be a random sample from an infinite population with mean µ and variance σ 2 . For any positive constant c, the probability the sample σ2 mean is within c units of µ is at least 1 − 2 : nc   σ2 Prob µ − c < X < µ + c ≥ 1 − 2 . (4.31) nc

c 2000 by Chapman & Hall/CRC 

As n → ∞ the probability approaches 1. (See Chebyshev inequality on page 48.) 4.3.5

Laws of the iterated logarithm

Laws of the iterated logarithm (the following hold “a.s.”, or “almost surely”): lim sup  t↓0

lim inf  t↓0

Wt

Wt =1 2t ln ln t Wt = −1 lim inf √ t→∞ 2t ln ln t

lim sup √

=1

2t ln ln(1/t) Wt

t→∞

= −1

2t ln ln(1/t)

where W is a Brownian motion. 4.4

FINITE POPULATION

Let {c1 , c2 , . . . , cN } be a collection of numbers representing a finite population of size N and assume the sampling from this population is done without replacement. Let the random variable Xi be the ith observation selected from the population. The collection X1 , X2 , . . . , Xn is a random sample of size n from the finite population if the joint probability mass function for X1 , X2 , . . . , Xn is p(x1 , x2 , . . . , xn ) =

1 . N (N − 1) · · · (N − n + 1)

(4.32)

(1) The marginal probability distribution for the random variable Xi , i = 1, 2, . . . , n is 1 for xi = c1 , c2 , . . . , cN . N (2) The mean and the variance of the finite population are pXi (xi ) =

µ=

N 

ci

i=1

1 N

and

σ2 =

N 

(ci − µ)2

i=1

(4.33)

1 . N

(4.34)

(3) The joint marginal probability mass function for any two random variables in the collection X1 , X2 , . . . , Xn is p(xi , xj ) =

1 . N (N − 1)

(4.35)

(4) The covariance between any two random variables in the collection X1 , X2 , . . . , Xn is Cov [Xi , Xj ] = −

c 2000 by Chapman & Hall/CRC 

σ2 . N −1

(4.36)

(5) Let X be the sample mean of the random sample of size n. The expected value and variance of X are     σ2 N − n E X =µ and Var X = · . (4.37) n N −1 The quantity (N − n)/(N − 1) is the finite population correction factor. 4.5 4.5.1

THEOREMS Theorems: the chi–square distribution

(1) Let Z be a standard normal random variable, then Z 2 has a chi–square distribution with 1 degree of freedom. (2) Let Z1 , Z2 , . . . , Zn be independent standard normal random variables. n  The random variable Y = Zi2 has a chi–square distribution with n i=1

degrees of freedom. (3) Let X1 , X2 , . . . , Xn be independent random variables such that Xi has a chi–square distribution with νi degrees of freedom. The random variable n  Y = Xi has a chi–square distribution with ν = ν1 + ν2 + · · · + νn i=1

degrees of freedom. (4) Let U have a chi–square distribution with ν1 degrees of freedom, U and V be independent, and U + V have a chi–square distribution with ν > ν1 degrees of freedom. The random variable V has a chi–sqaure distribution with ν − ν1 degrees of freedom. (5) Let X1 , X2 , . . . , Xn be a random sample from a normal population with mean µ and variance σ 2 . Then (a) The sample mean, X, and the sample variance, S 2 , are independent, and (n − 1)S 2 (b) The random variable has a chi–square distribution with σ2 n − 1 degrees of freedom. 4.5.2

Theorems: the t distribution

(1) Let Z have a standard normal distribution, X have a chi–square distribution with ν degrees of freedom, and X and Z be independent. The random variable Z T = (4.38) X/ν has a t distribution with ν degrees of freedom.

c 2000 by Chapman & Hall/CRC 

(2) Let X1 , X2 , . . . , Xn be a random sample from a normal population with mean µ and variance σ 2 . The random variable X −µ √ S/ n

T =

(4.39)

has a t distribution with n − 1 degrees of freedom. 4.5.3

Theorems: the F distribution

(1) Let U have a chi–square distribution with ν1 degrees of freedom, V have a chi–square distribution with ν2 degrees of freedom, and U and V be independent. The random variable F =

U/ν1 V /ν2

(4.40)

has an F distribution with ν1 and ν2 degrees of freedom. (2) Let X1 , X2 , . . . , Xm and Y1 , Y2 , . . . , Yn be random samples from nor2 mal populations with variances σX and σY2 , respectively. The random variable F =

2 2 SX /σX Sy2 /σY2

(4.41)

has an F distribution with m − 1 and n − 1 degrees of freedom. (3) Let Fα,ν1 ,ν2 be a critical value for the F distribution defined by Prob [F ≥ Fα,ν1 ,ν2 ] = α. Then F1−α,ν1 ,ν2 = 1/Fα,ν2 ,ν1 . 4.6 4.6.1

ORDER STATISTICS Definition

Let X1 , X2 , . . . , Xn be independent continuous random variables with probability density function f (x) and cumulative distribution function F (x). The order statistic, X(i) , i = 1, 2, . . . , n, is a random variable defined to be the ith largest of the set {X1 , X2 , . . . , Xn }. Therefore, X(1) ≤ X(2) ≤ · · · ≤ X(n)

(4.42)

X(1) = min{X1 , X2 , . . . , Xn } and X(n) = max{X1 , X2 , . . . , Xn }.

(4.43)

and in particular

The cumulative distribution function for the ith order statistic is   FX(i) (x) = Prob X(i) ≤ x = Prob [i or more observations are ≤ x] n    (4.44) n [F (x)]j [1 − F (x)]n−j = j j=i

c 2000 by Chapman & Hall/CRC 

and the probability density function is   n−1 fX(i) (x) = n [F (x)]i−1 [1 − F (x)]n−i f (x) i−1 n! = [F (x)]i−1 f (x)[1 − F (x)]n−i . (i − 1)!(n − i)! 4.6.2

(4.45)

The first order statistic

The probability density function, fX(1) (x), and the cumulative distribution function, FX(1) (x), for X(1) are fX(1) (x) = n[1 − F (x)]n−1 f (x) 4.6.3

FX(1) (x) = 1 − [1 − F (x)]n .

(4.46)

The nth order statistic

The probability density function, f(n) (x), and the cumulative distribution function, F(n) (x), for X(n) are fX(n) (x) = n[F (x)]n−1 f (x) 4.6.4

FX(n) (x) = [F (x)]n .

(4.47)

The median

If the number of observations is odd, the median is the middle observation when the observations are in numerical order. If the number of observations is even, the median is (arbitrarily) defined as the average of the middle two of the ordered observations. ( X(k) if n is odd and n = 2k − 1 median = 1 (4.48) 2 [X(k) + X(k+1) ] if n is even and n = 2k 4.6.5

Joint distributions

The joint density function for X(1) , X(2) , . . . , X(n) is g(x1 , x2 , . . . , xn ) = n!f (x1 )f (x2 ) · · · f (xn ).

(4.49)

The joint density function for the ith and j th (i < j) order statistics is fij (x, y) =

n! f (x)f (y) (i − 1)!(j − i − 1)!(n − j)!

(4.50)

×[F (x)]i−1 [1 − F (y)]n−j [F (y) − F (x)]j−i−1 . The joint distribution function for X(1) and X(n) is   F1n (x, y) = Prob X(1) ≤ x and X(n) ≤ y ( n [F (y)] − [F (y) − F (x)]n if x ≤ y = n if x > y [F (y)]

c 2000 by Chapman & Hall/CRC 

(4.51)

and the joint density function is ( n(n − 1)f (x)f (y)[F (y) − F (x)]n−2 f1n (x, y) = 0 4.6.6

if x ≤ y if x > y

(4.52)

Midrange and range

  The midrange is defined to be A = 12 X(1) + X(n ) . Using f1n (x, y) for the joint density function of X(1) and X(n ) results in  x fA (x) = 2 f1n (t, 2x − t) dt −∞ (4.53)  x n−2 = 2n(n − 1) f (t)f (2x − t) [F (2x − t) − F (t)] dt −∞

The range is the difference between the largest and smallest observations: R = X(n) − X(1) . The random variable R is used in the construction of tolerance intervals.  ∞ fR (r) = f1n (t, t + r) dt −∞

=

4.6.7

    n(n − 1)  



−∞

f (t)f (t + r)[F (t + r) − F (t)]n−2 dt

0

if r > 0 (4.54) if r ≤ 0

Uniform distribution: order statistics

If X is uniformly distributed on the interval [0, 1] then the density function for X(i) is   n − 1 i−1 fi (x) = n x (1 − x)n−i , 0 ≤ x ≤ 1 (4.55) i−1 which is a beta distribution with parameters i and n − i + 1.  1   i (1) E X(i) = . fk (t) dt = n+1 0

n . n+1 1 (3) The expected value of the smallest of n observations is . n+1 (4) The density function of the midrange is ( if 0 < x ≤ 12 n2n−1 xn−1 fA (x) = n2n−1 (1 − x)n−1 if 12 ≤ x < 1

(2) The expected value of the largest of n observations is

c 2000 by Chapman & Hall/CRC 

(4.56)

(5) The density function of the range is ( n(n − 1)(1 − r)rn−2 fR (r) = 0 4.6.7.1

if 0 < r < 1 otherwise

(4.57)

Tolerance intervals

In many applications, we need to estimate an interval in which a certain proportion of the population lies, with given probability. A tolerance interval may be constructed using the results relating to order statistics and the range. A table of required sample sizes for varying ranges and probabilities is in the following table. Example 4.30 : Assume a sample is drawn from a uniform population. Find a sample size n such that at least 99% of the sample population, with probability .95, lies between the smallest and largest observations. This problem may be written as a probability statement: 0.95 = Prob [F (Zn ) − F (Z1 ) > 0.99] = Prob [R > 0.99]  1 (1 − r)rn−2 dr = n(n − 1)

(4.58)

0.99

= 1 − (0.99)n−1 (0.01n + 0.99) Solving this results in the value n ≈ 473.

Tolerance intervals, uniform distribution Probability 0.500 0.750 0.900 0.950 0.975 0.990 0.995

This fraction of the total population is within the range 0.500 0.750 0.900 0.950 0.975 0.990 0.995 0.999 3 7 17 34 67 168 336 1679 5 10 26 53 107 269 538 2692 6 14 38 77 154 388 777 3889 8 17 46 93 188 473 947 4742 9 20 54 109 221 555 1112 5570 10 24 64 130 263 661 1325 6636 11 26 71 145 294 740 1483 7427

For tolerance intervals for normal samples, see section 7.3. 4.6.8

Normal distribution: order statistics

When the {Xi } come from a standard normal distribution, the {X(i) } are called standard order statistics. 4.6.8.1

Expected value of normal order statistics

The tables on pages 65–66 gives expected values of standard order statistics   ∞   n−1 E X(i) = n tf (t)[F (t)]i−1 [1 − F (t)]n−i dt (4.59) i−1 −∞ c 2000 by Chapman & Hall/CRC 

 x −t2 /2 2 e e−x /2 √ when f (x) = √ and F (x) = dt. Missing values (indicated 2π −∞  2π   by a dash) may be obtained from E X(i) = −E X(n−i+1) . Example 4.31 : If an average person takes five intelligence tests (each test having a normal distribution with a mean of 100 and a standard deviation of 20), what is the expected value of the largest score? Solution: (S1) We need to obtain the expected value of the largest normal order statistic when n = 5.  $ (S2) Using n = 5 and i = 5 in the table on page 65 yields (use j = 1) E X(5) $n=5 = 1.1630. (S3) The expected value of the largest score is 100 + (1.1630)(20) ≈ 123.

Expected value of the ith normal order statistic (use j = n − i + 1) j 1 2 3 4 5 j n = 10 1 1.5388 2 1.0014 3 0.6561 4 0.3757 5 0.1227 6 — 7 — 8 — 9 — 10 — j n = 20 1 1.8675 2 1.4076 3 1.1310 4 0.9210 5 0.7454 6 0.5903 7 0.4483 8 0.3149 9 0.1869 10 0.0620 11 — 12 — 13 — 14 — 15 —

11 1.5865 1.0619 0.7288 0.4619 0.2249 0.0000 — — — — 21 1.8892 1.4336 1.1605 0.9538 0.7816 0.6298 0.4915 0.3620 0.2384 0.1183 0.0000 — — — —

n=2 0.5642 — — — — 12 1.6292 1.1157 0.7929 0.5368 0.3122 0.1025 — — — — 22 1.9097 1.4581 1.1883 0.9846 0.8153 0.6667 0.5316 0.4056 0.2857 0.1699 0.0564 — — — —

c 2000 by Chapman & Hall/CRC 

3 0.8463 0.0000 — — — 13 1.6680 1.1641 0.8498 0.6028 0.3883 0.1905 0.0000 — — — 23 1.9292 1.4813 1.2145 1.0136 0.8470 0.7012 0.5690 0.4461 0.3296 0.2175 0.1081 0.0000 — — —

4 1.0294 0.2970 — — — 14 1.7034 1.2079 0.9011 0.6618 0.4556 0.2672 0.0882 — — — 24 1.9477 1.5034 1.2393 1.0409 0.8769 0.7336 0.6040 0.4839 0.3704 0.2616 0.1558 0.0518 — — —

5 1.1629 0.4950 0.0000 — — 15 1.7359 1.2479 0.9477 0.7149 0.5157 0.3353 0.1653 0.0000 — — 25 1.9653 1.5243 1.2628 1.0668 0.9051 0.7641 0.6369 0.5193 0.4086 0.3026 0.2000 0.0995 0.0000 — —

6 1.2672 0.6418 0.2015 — — 16 1.7660 1.2848 0.9903 0.7632 0.5700 0.3962 0.2337 0.0772 — — 26 1.9822 1.5442 1.2851 1.0914 0.9318 0.7929 0.6679 0.5527 0.4443 0.3410 0.2413 0.1439 0.0478 — —

7 1.3522 0.7574 0.3527 0.0000 — 17 1.7939 1.3188 1.0295 0.8074 0.6195 0.4513 0.2952 0.1459 0.0000 — 27 1.9983 1.5632 1.3064 1.1147 0.9571 0.8202 0.6973 0.5841 0.4780 0.3770 0.2798 0.1852 0.0922 0.0000 —

8 1.4236 0.8522 0.4728 0.1526 — 18 1.8200 1.3504 1.0657 0.8481 0.6648 0.5016 0.3508 0.2077 0.0688 — 28 2.0137 1.5814 1.3268 1.1370 0.9812 0.8462 0.7251 0.6138 0.5098 0.4109 0.3160 0.2239 0.1336 0.0444 —

9 1.4850 0.9323 0.5720 0.2745 0.0000 19 1.8445 1.3800 1.0995 0.8859 0.7066 0.5477 0.4016 0.2637 0.1307 0.0000 29 2.0285 1.5988 1.3462 1.1582 1.0042 0.8709 0.7515 0.6420 0.5398 0.4430 0.3501 0.2602 0.1724 0.0859 0.0000

Expected value of the ith normal order statistic (use j = n − i + 1) j n = 30 1 2.0427 2 1.6156 3 1.3648 4 1.1786 5 1.0262 6 0.8944 7 0.7767 8 0.6689 9 0.5683 10 0.4733 11 0.3823 12 0.2945 13 0.2088 14 0.1247 15 0.0415 16 — 17 — 18 — 19 — 20 —

4.6.8.2

31 2.0564 1.6316 1.3827 1.1980 1.0472 0.9169 0.8007 0.6944 0.5954 0.5020 0.4129 0.3268 0.2432 0.1613 0.0804 0.0000 — — — —

32 2.0696 1.6471 1.3999 1.2167 1.0673 0.9385 0.8236 0.7188 0.6213 0.5294 0.4418 0.3575 0.2757 0.1957 0.1170 0.0389 — — — —

33 2.0824 1.6620 1.4164 1.2347 1.0866 0.9591 0.8456 0.7420 0.6460 0.5555 0.4694 0.3867 0.3065 0.2283 0.1515 0.0755 0.0000 — — —

34 2.0947 1.6763 1.4323 1.2520 1.1052 0.9789 0.8666 0.7644 0.6695 0.5804 0.4957 0.4144 0.3358 0.2592 0.1842 0.1101 0.0367 — — —

35 2.1066 1.6902 1.4476 1.2686 1.1230 0.9979 0.8868 0.7857 0.6921 0.6043 0.5208 0.4409 0.3637 0.2886 0.2151 0.1428 0.0713 0.0000 — —

36 2.1181 1.7036 1.4624 1.2847 1.1402 1.0163 0.9063 0.8063 0.7138 0.6271 0.5449 0.4662 0.3903 0.3166 0.2446 0.1739 0.1040 0.0346 — —

37 2.1292 1.7165 1.4768 1.3002 1.1568 1.0339 0.9250 0.8261 0.7346 0.6490 0.5679 0.4904 0.4157 0.3433 0.2727 0.2034 0.1351 0.0674 0.0000 —

38 2.1401 1.7291 1.4906 1.3151 1.1729 1.0510 0.9430 0.8451 0.7547 0.6701 0.5900 0.5136 0.4401 0.3689 0.2995 0.2316 0.1647 0.0986 0.0328 —

39 2.1505 1.7413 1.5040 1.3296 1.1884 1.0674 0.9604 0.8634 0.7740 0.6904 0.6113 0.5359 0.4635 0.3934 0.3252 0.2585 0.1929 0.1282 0.0640 0.0000

Variances and covariances of order statistics

Given n observations of independent standard normal variables, arrange the sample in ascending order of magnitude X(1) , X(2) , . . . , X(n) . The variances and covariances for expected values and product moments may be found from   ∞   n−1 E X(i) = n tf (t)F i−1 (t)[1 − F (t)]n−i dt n−i −∞   ∞ " # n−1 2 E X(i) = n t2 f (t)F i−1 (t)[1 − F (t)]n−i dt n−i (4.60) −∞   ∞  y   n−1 E X(i) X(j) = n tyf (t)f (y) n−i −∞ −∞ × [F (t)]i−1 [1 − F (y)]n−j [F (y) − F (t)]j−i−1 dt dy  x −x2 /2 2 e−x /2 e √ where f (x) = √ and F (x) = dx. 2π 2π −∞ The following table gives the variances and covariances of order statistics in samples of sizes up to 10 from a standard Missing values   normal distribution.    may be obtained from E X(i) X(j) = E X(j) X(i) = E X(n−i+1) X(n−j+1) See G. L. Tietjen, D. K. Kahaner, and R. J. Beckman, “Variances and covariances of the normal order statistics for samples sizes 2 to 50”, Selected Tables in Mathematical Statistics, 5, American Mathematical Society, Providence, RI, 1977. c 2000 by Chapman & Hall/CRC 

Variances and covariances of normal order statistics  E X(i) X(j) is shown for samples of size n (use k = n − i + 1 and = = n − j + 1) n 2

k 1

3

2 1

4

2 1

2 5

1

2

6

3 1

2

3 7

1

2

. 1 2 2 1 2 3 2 1 2 3 4 2 3 1 2 3 4 5 2 3 4 3 1 2 3 4 5 6 2 3 4 5 3 4 1 2 3 4 5 6 7 2 3 4 5 6

value .6817 .3183 .6817 .5595 .2757 .1649 .4487 .4917 .2456 .1580 .1047 .3605 .2359 .4475 .2243 .1481 .1058 .0742 .3115 .2084 .1499 .2868 .4159 .2085 .1394 .1024 .0774 .0563 .2796 .1890 .1397 .1059 .2462 .1833 .3919 .1962 .1321 .0985 .0766 .0599 .0448 .2567 .1745 .1307 .1020 .0800

c 2000 by Chapman & Hall/CRC 

n 7

k 3

8

4 1

2

3

4 9

1

2

3

. 3 4 5 4 1 2 3 4 5 6 7 8 2 3 4 5 6 7 3 4 5 6 4 5 1 2 3 4 5 6 7 8 9 2 3 4 5 6 7 8 3 4 5 6 7

value .2197 .1656 .1296 .2104 .3729 .1863 .1260 .0947 .0748 .0602 .0483 .0368 .2394 .1632 .1233 .0976 .0787 .0632 .2008 .1524 .1210 .0978 .1872 .1492 .3574 .1781 .1207 .0913 .0727 .0595 .0491 .0401 .0311 .2257 .1541 .1170 .0934 .0765 .0632 .0517 .1864 .1421 .1138 .0934 .0772

n 9

k 4

10

5 1

2

3

4

5

. 4 5 6 5 1 2 3 4 5 6 7 8 9 10 2 3 4 5 6 7 8 9 3 4 5 6 7 8 4 5 6 7 5 6

value .1706 .1370 .1127 .1661 .3443 .1713 .1163 .0882 .0707 .0584 .0489 .0411 .0340 .0267 .2145 .1466 .1117 .0897 .0742 .0622 .0523 .0434 .1750 .1338 .1077 .0892 .0749 .0630 .1579 .1275 .1058 .0889 .1511 .1256

4.7 4.7.1

RANGE AND STUDENTIZED RANGE Probability integral of the range

Let {X1 , X2 , . . . , Xn } denote a random sample of size n from a population with standard deviation σ, density function f (x), and cumulative distribution function F (x). Let {X(1) , X(2) , . . . , X(n) } denote the same values in ascending order of magnitude. The sample range R is defined by R = X(n) − X(1)

(4.61)

In standardized form X(n) − X(1) R = (4.62) σ σ The probability that the range exceeds some value R, for a sample of size n, is (see equation (4.54))    ∞ range exceeds R for Prob = fR (r) dr a sample of size n R (4.63) n−1  ∞ =n F (t + R) − F (t) f (t) dt W =

−∞

The following tables provide values of this probability for the normal density 2 1 function f (x) = √ e−x /2 for various values of n and W . (Note that since 2π σ = 1 for this case R = W .)

c 2000 by Chapman & Hall/CRC 

Probability integral of the range W 0.00 0.05 0.10 0.15 0.20

n=2 0.0000 0.0282 0.0564 0.0845 0.1125

3 0.0000 0.0007 0.0028 0.0062 0.0110

4

0.0000 0.0001 0.0004 0.0000 0.0010 0.0001

0.25 0.30 0.35 0.40 0.45

0.1403 0.1680 0.1955 0.2227 0.2497

0.0171 0.0245 0.0332 0.0431 0.0543

0.0020 0.0034 0.0053 0.0079 0.0111

0.0002 0.0004 0.0008 0.0014 0.0022

0.0000 0.0001 0.0001 0.0002 0.0000 0.0004 0.0001

0.50 0.55 0.60 0.65 0.70

0.2763 0.3027 0.3286 0.3542 0.3794

0.0666 0.0800 0.0944 0.1099 0.1263

0.0152 0.0200 0.0257 0.0322 0.0398

0.0033 0.0048 0.0068 0.0092 0.0121

0.0007 0.0011 0.0017 0.0026 0.0036

0.0002 0.0003 0.0004 0.0007 0.0011

0.0000 0.0001 0.0001 0.0000 0.0002 0.0001 0.0003 0.0001

0.75 0.80 0.85 0.90 0.95

0.4041 0.4284 0.4522 0.4755 0.4983

0.1436 0.1616 0.1805 0.2000 0.2201

0.0483 0.0578 0.0682 0.0797 0.0922

0.0157 0.0200 0.0250 0.0308 0.0375

0.0050 0.0068 0.0090 0.0117 0.0150

0.0016 0.0023 0.0032 0.0044 0.0059

0.0005 0.0008 0.0011 0.0016 0.0023

0.0002 0.0002 0.0004 0.0006 0.0009

0.0000 0.0001 0.0001 0.0002 0.0003

1.00 1.05 1.10 1.15 1.20

0.5205 0.5422 0.5633 0.5839 0.6039

0.2407 0.2618 0.2833 0.3052 0.3272

0.1057 0.1201 0.1355 0.1517 0.1688

0.0450 0.0535 0.0629 0.0733 0.0847

0.0188 0.0234 0.0287 0.0348 0.0417

0.0078 0.0101 0.0129 0.0163 0.0203

0.0032 0.0043 0.0058 0.0076 0.0098

0.0013 0.0018 0.0025 0.0035 0.0047

0.0005 0.0008 0.0011 0.0016 0.0022

1.25 1.30 1.35 1.40 1.45

0.6232 0.6420 0.6602 0.6778 0.6948

0.3495 0.3719 0.3943 0.4168 0.4392

0.1867 0.2054 0.2248 0.2448 0.2654

0.0970 0.1104 0.1247 0.1400 0.1562

0.0495 0.0583 0.0680 0.0787 0.0904

0.0249 0.0304 0.0366 0.0437 0.0516

0.0125 0.0157 0.0195 0.0240 0.0292

0.0062 0.0080 0.0103 0.0131 0.0164

0.0030 0.0041 0.0054 0.0071 0.0092

1.50 1.55 1.60 1.65 1.70

0.7112 0.7269 0.7421 0.7567 0.7707

0.4614 0.4835 0.5053 0.5269 0.5481

0.2865 0.3080 0.3299 0.3521 0.3745

0.1733 0.1913 0.2101 0.2296 0.2498

0.1031 0.1168 0.1315 0.1473 0.1639

0.0606 0.0705 0.0814 0.0934 0.1064

0.0353 0.0421 0.0499 0.0587 0.0684

0.0204 0.0250 0.0304 0.0366 0.0437

0.0117 0.0148 0.0184 0.0227 0.0278

1.75 1.80 1.85 1.90 1.95

0.7841 0.7969 0.8092 0.8209 0.8321

0.5690 0.5894 0.6094 0.6290 0.6480

0.3970 0.4197 0.4423 0.4649 0.4874

0.2706 0.2920 0.3138 0.3361 0.3587

0.1815 0.2000 0.2193 0.2394 0.2602

0.1204 0.1355 0.1516 0.1686 0.1867

0.0792 0.0910 0.1039 0.1178 0.1329

0.0517 0.0607 0.0707 0.0818 0.0939

0.0336 0.0403 0.0479 0.0565 0.0661

2.00 2.05 2.10 2.15 2.20

0.8427 0.8528 0.8624 0.8716 0.8802

0.6665 0.6845 0.7019 0.7187 0.7349

0.5096 0.5317 0.5534 0.5748 0.5957

0.3816 0.4046 0.4277 0.4508 0.4739

0.2816 0.3035 0.3260 0.3489 0.3720

0.2056 0.2254 0.2460 0.2673 0.2893

0.1489 0.1661 0.1842 0.2032 0.2232

0.1072 0.1216 0.1371 0.1536 0.1712

0.0768 0.0886 0.1015 0.1155 0.1307

2.25

0.8884 0.7505 0.6163 0.4969 0.3955 0.3118 0.2440 0.1899 0.1470

c 2000 by Chapman & Hall/CRC 

5

6

7

8

9

10

Probability integral of the range W 0.00 0.05 0.10 0.15 0.20

n = 11

12

13

14

15

16

17

18

19

20

0.25 0.30 0.35 0.40 0.45 0.50 0.55 0.60 0.65 0.70 0.75 0.80 0.85 0.90 0.95

0.0000 0.0001 0.0001 0.0000

1.00 1.05 1.10 1.15 1.20

0.0002 0.0003 0.0005 0.0007 0.0010

0.0001 0.0001 0.0002 0.0003 0.0005

0.0000 0.0001 0.0001 0.0000 0.0001 0.0001 0.0000 0.0002 0.0001 0.0001

1.25 1.30 1.35 1.40 1.45

0.0015 0.0021 0.0028 0.0038 0.0051

0.0007 0.0010 0.0015 0.0021 0.0028

0.0004 0.0005 0.0008 0.0011 0.0016

0.0002 0.0003 0.0004 0.0006 0.0009

0.0001 0.0001 0.0002 0.0003 0.0005

0.0000 0.0001 0.0001 0.0002 0.0003

0.0000 0.0001 0.0001 0.0000 0.0001 0.0001 0.0000

1.50 1.55 1.60 1.65 1.70

0.0067 0.0087 0.0111 0.0140 0.0176

0.0038 0.0051 0.0067 0.0086 0.0111

0.0022 0.0030 0.0040 0.0053 0.0070

0.0012 0.0017 0.0024 0.0032 0.0044

0.0007 0.0010 0.0014 0.0020 0.0027

0.0004 0.0006 0.0008 0.0012 0.0017

0.0002 0.0003 0.0005 0.0007 0.0011

0.0001 0.0002 0.0003 0.0004 0.0007

0.0001 0.0001 0.0002 0.0003 0.0004

0.0000 0.0001 0.0001 0.0002 0.0003

1.75 1.80 1.85 1.90 1.95

0.0217 0.0266 0.0323 0.0388 0.0463

0.0140 0.0175 0.0217 0.0266 0.0323

0.0090 0.0115 0.0145 0.0182 0.0225

0.0058 0.0075 0.0097 0.0124 0.0156

0.0037 0.0049 0.0065 0.0084 0.0108

0.0023 0.0032 0.0043 0.0057 0.0075

0.0015 0.0021 0.0029 0.0039 0.0052

0.0010 0.0014 0.0019 0.0026 0.0036

0.0006 0.0009 0.0013 0.0018 0.0024

0.0004 0.0006 0.0008 0.0012 0.0017

2.00 2.05 2.10 2.15 2.20

0.0548 0.0643 0.0748 0.0866 0.0994

0.0389 0.0465 0.0550 0.0646 0.0753

0.0276 0.0335 0.0403 0.0481 0.0569

0.0195 0.0241 0.0295 0.0357 0.0429

0.0137 0.0173 0.0215 0.0265 0.0323

0.0097 0.0124 0.0156 0.0196 0.0242

0.0068 0.0088 0.0114 0.0144 0.0182

0.0048 0.0063 0.0082 0.0106 0.0136

0.0033 0.0045 0.0060 0.0078 0.0102

0.0023 0.0032 0.0043 0.0058 0.0076

2.25

0.1134 0.0872 0.0669 0.0511 0.0390 0.0297 0.0226 0.0172 0.0130 0.0099

c 2000 by Chapman & Hall/CRC 

Probability integral of the range W 2.25 2.30 2.35 2.40 2.45

n=2 0.8884 0.8961 0.9034 0.9103 0.9168

3 0.7505 0.7655 0.7799 0.7937 0.8069

4 0.6163 0.6363 0.6559 0.6748 0.6932

5 0.4969 0.5196 0.5421 0.5643 0.5861

6 0.3955 0.4190 0.4427 0.4663 0.4899

7 0.3118 0.3348 0.3582 0.3820 0.4059

8 0.2440 0.2656 0.2878 0.3107 0.3341

9 0.1899 0.2095 0.2300 0.2514 0.2735

10 0.1470 0.1645 0.1829 0.2025 0.2229

2.50 2.55 2.60 2.65 2.70

0.9229 0.9286 0.9340 0.9390 0.9438

0.8195 0.8315 0.8429 0.8537 0.8640

0.7110 0.7282 0.7448 0.7607 0.7759

0.6075 0.6283 0.6487 0.6685 0.6877

0.5132 0.5364 0.5592 0.5816 0.6036

0.4300 0.4541 0.4782 0.5022 0.5259

0.3579 0.3820 0.4064 0.4310 0.4555

0.2963 0.3198 0.3437 0.3680 0.3927

0.2443 0.2665 0.2894 0.3130 0.3372

2.75 2.80 2.85 2.90 2.95

0.9482 0.9523 0.9561 0.9597 0.9630

0.8737 0.8828 0.8915 0.8996 0.9073

0.7905 0.8045 0.8177 0.8304 0.8424

0.7063 0.7242 0.7415 0.7581 0.7739

0.6252 0.6461 0.6665 0.6863 0.7055

0.5494 0.5725 0.5952 0.6174 0.6391

0.4801 0.5045 0.5286 0.5525 0.5760

0.4175 0.4425 0.4675 0.4923 0.5171

0.3617 0.3867 0.4119 0.4372 0.4625

3.00 3.05 3.10 3.15 3.20

0.9661 0.9690 0.9716 0.9741 0.9763

0.9145 0.9212 0.9275 0.9334 0.9388

0.8537 0.8645 0.8746 0.8842 0.8931

0.7891 0.8036 0.8174 0.8305 0.8429

0.7239 0.7416 0.7587 0.7750 0.7905

0.6601 0.6806 0.7003 0.7194 0.7377

0.5991 0.6216 0.6436 0.6649 0.6856

0.5415 0.5656 0.5892 0.6124 0.6350

0.4878 0.5129 0.5378 0.5623 0.5864

3.25 3.30 3.35 3.40 3.45

0.9784 0.9804 0.9822 0.9838 0.9853

0.9439 0.9487 0.9531 0.9572 0.9610

0.9016 0.9095 0.9168 0.9237 0.9302

0.8546 0.8657 0.8761 0.8859 0.8951

0.8053 0.8194 0.8327 0.8454 0.8573

0.7553 0.7721 0.7881 0.8034 0.8179

0.7055 0.7248 0.7432 0.7609 0.7778

0.6569 0.6782 0.6988 0.7186 0.7376

0.6099 0.6329 0.6553 0.6769 0.6978

3.50 3.55 3.60 3.65 3.70

0.9867 0.9879 0.9891 0.9901 0.9911

0.9644 0.9677 0.9706 0.9734 0.9759

0.9361 0.9417 0.9468 0.9516 0.9560

0.9037 0.9117 0.9192 0.9261 0.9326

0.8685 0.8790 0.8889 0.8981 0.9067

0.8316 0.8446 0.8568 0.8683 0.8790

0.7938 0.8091 0.8236 0.8372 0.8501

0.7558 0.7732 0.7898 0.8055 0.8204

0.7180 0.7373 0.7558 0.7735 0.7903

3.75 3.80 3.85 3.90 3.95

0.9920 0.9928 0.9935 0.9942 0.9948

0.9782 0.9803 0.9822 0.9840 0.9856

0.9600 0.9637 0.9672 0.9703 0.9732

0.9386 0.9441 0.9493 0.9540 0.9583

0.9147 0.9222 0.9291 0.9355 0.9415

0.8891 0.8985 0.9073 0.9155 0.9230

0.8622 0.8736 0.8842 0.8941 0.9034

0.8345 0.8477 0.8602 0.8718 0.8827

0.8062 0.8212 0.8355 0.8488 0.8614

4.00 4.05 4.10 4.15 4.20

0.9953 0.9958 0.9963 0.9967 0.9970

0.9870 0.9883 0.9895 0.9906 0.9916

0.9758 0.9782 0.9804 0.9824 0.9842

0.9623 0.9660 0.9693 0.9724 0.9752

0.9469 0.9520 0.9566 0.9608 0.9647

0.9300 0.9365 0.9425 0.9480 0.9530

0.9120 0.9199 0.9273 0.9341 0.9404

0.8929 0.9024 0.9112 0.9193 0.9268

0.8731 0.8841 0.8943 0.9038 0.9126

4.25 4.30 4.35 4.40 4.45

0.9973 0.9976 0.9979 0.9981 0.9983

0.9925 0.9933 0.9941 0.9947 0.9953

0.9859 0.9874 0.9887 0.9899 0.9910

0.9777 0.9800 0.9821 0.9840 0.9857

0.9682 0.9715 0.9744 0.9771 0.9795

0.9576 0.9619 0.9657 0.9692 0.9724

0.9461 0.9514 0.9562 0.9607 0.9647

0.9338 0.9402 0.9460 0.9514 0.9563

0.9208 0.9283 0.9352 0.9416 0.9474

4.50

0.9985 0.9958 0.9920 0.9873 0.9817 0.9754 0.9684 0.9608 0.9527

c 2000 by Chapman & Hall/CRC 

Probability integral of the range W 2.25 2.30 2.35 2.40 2.45

n = 11 0.1134 0.1286 0.1450 0.1624 0.1810

12 0.0872 0.1003 0.1145 0.1299 0.1466

13 0.0669 0.0779 0.0902 0.1036 0.1183

14 0.0511 0.0604 0.0709 0.0825 0.0953

15 0.0390 0.0468 0.0556 0.0655 0.0766

16 0.0297 0.0361 0.0435 0.0519 0.0615

17 0.0226 0.0279 0.0340 0.0411 0.0493

18 0.0172 0.0214 0.0265 0.0325 0.0394

19 0.0130 0.0165 0.0207 0.0256 0.0315

20 0.0099 0.0127 0.0161 0.0202 0.0251

2.50 2.55 2.60 2.65 2.70

0.2007 0.2213 0.2429 0.2653 0.2885

0.1643 0.1833 0.2032 0.2243 0.2462

0.1342 0.1513 0.1696 0.1891 0.2096

0.1094 0.1247 0.1413 0.1590 0.1780

0.0890 0.1025 0.1174 0.1335 0.1509

0.0722 0.0842 0.0974 0.1119 0.1278

0.0585 0.0690 0.0807 0.0937 0.1080

0.0474 0.0565 0.0668 0.0783 0.0911

0.0383 0.0462 0.0552 0.0654 0.0768

0.0309 0.0377 0.0455 0.0545 0.0647

2.75 2.80 2.85 2.90 2.95

0.3124 0.3368 0.3618 0.3870 0.4125

0.2690 0.2926 0.3169 0.3417 0.3670

0.2311 0.2536 0.2770 0.3011 0.3258

0.1981 0.2194 0.2416 0.2647 0.2887

0.1696 0.1894 0.2103 0.2323 0.2553

0.1449 0.1632 0.1828 0.2036 0.2255

0.1236 0.1405 0.1587 0.1782 0.1989

0.1053 0.1208 0.1376 0.1557 0.1752

0.0896 0.1037 0.1191 0.1360 0.1541

0.0761 0.0889 0.1031 0.1186 0.1355

3.00 3.05 3.10 3.15 3.20

0.4382 0.4639 0.4895 0.5150 0.5401

0.3927 0.4186 0.4446 0.4706 0.4965

0.3511 0.3769 0.4029 0.4291 0.4554

0.3134 0.3387 0.3645 0.3907 0.4171

0.2792 0.3039 0.3292 0.3551 0.3814

0.2484 0.2723 0.2969 0.3223 0.3483

0.2207 0.2436 0.2675 0.2922 0.3177

0.1959 0.2177 0.2407 0.2646 0.2894

0.1736 0.1944 0.2163 0.2394 0.2634

0.1537 0.1733 0.1942 0.2163 0.2395

3.25 3.30 3.35 3.40 3.45

0.5649 0.5893 0.6131 0.6363 0.6589

0.5222 0.5475 0.5725 0.5970 0.6209

0.4817 0.5078 0.5337 0.5592 0.5842

0.4437 0.4703 0.4967 0.5230 0.5489

0.4080 0.4348 0.4617 0.4885 0.5150

0.3748 0.4016 0.4286 0.4557 0.4827

0.3438 0.3704 0.3974 0.4246 0.4519

0.3151 0.3413 0.3681 0.3953 0.4227

0.2884 0.3142 0.3407 0.3676 0.3950

0.2638 0.2890 0.3150 0.3416 0.3688

3.50 3.55 3.60 3.65 3.70

0.6807 0.7017 0.7220 0.7414 0.7600

0.6442 0.6668 0.6886 0.7096 0.7298

0.6087 0.6326 0.6558 0.6782 0.6999

0.5744 0.5994 0.6237 0.6474 0.6704

0.5413 0.5672 0.5926 0.6173 0.6414

0.5096 0.5362 0.5624 0.5881 0.6132

0.4792 0.5063 0.5332 0.5597 0.5856

0.4502 0.4777 0.5051 0.5322 0.5588

0.4226 0.4504 0.4781 0.5056 0.5329

0.3964 0.4242 0.4522 0.4801 0.5078

3.75 3.80 3.85 3.90 3.95

0.7776 0.7944 0.8103 0.8254 0.8395

0.7491 0.7675 0.7850 0.8016 0.8173

0.7206 0.7406 0.7596 0.7777 0.7948

0.6925 0.7138 0.7342 0.7537 0.7723

0.6648 0.6874 0.7090 0.7298 0.7497

0.6376 0.6613 0.6842 0.7062 0.7273

0.6110 0.6357 0.6596 0.6827 0.7050

0.5850 0.6106 0.6355 0.6596 0.6829

0.5598 0.5861 0.6118 0.6369 0.6611

0.5352 0.5622 0.5887 0.6145 0.6397

4.00 4.05 4.10 4.15 4.20

0.8528 0.8653 0.8769 0.8878 0.8978

0.8321 0.8460 0.8590 0.8712 0.8826

0.8111 0.8264 0.8408 0.8543 0.8669

0.7899 0.8066 0.8223 0.8371 0.8509

0.7686 0.7866 0.8036 0.8196 0.8347

0.7474 0.7666 0.7848 0.8021 0.8183

0.7263 0.7466 0.7660 0.7844 0.8018

0.7053 0.7268 0.7472 0.7667 0.7852

0.6845 0.7070 0.7285 0.7491 0.7686

0.6640 0.6874 0.7099 0.7315 0.7520

4.25 4.30 4.35 4.40 4.45

0.9072 0.9158 0.9238 0.9312 0.9379

0.8931 0.9029 0.9120 0.9204 0.9281

0.8787 0.8896 0.8998 0.9092 0.9178

0.8639 0.8760 0.8872 0.8976 0.9073

0.8488 0.8620 0.8744 0.8858 0.8964

0.8336 0.8479 0.8613 0.8737 0.8853

0.8182 0.8336 0.8480 0.8615 0.8740

0.8027 0.8191 0.8346 0.8490 0.8625

0.7871 0.8046 0.8210 0.8364 0.8508

0.7715 0.7899 0.8074 0.8237 0.8391

4.50

0.9441 0.9352 0.9258 0.9162 0.9062 0.8960 0.8856 0.8750 0.8643 0.8534

c 2000 by Chapman & Hall/CRC 

Probability integral of the range W 4.50 4.55 4.60 4.65 4.70

n=2 0.9985 0.9987 0.9989 0.9990 0.9991

3 0.9958 0.9963 0.9967 0.9971 0.9974

4 0.9920 0.9929 0.9937 0.9944 0.9951

5 0.9873 0.9887 0.9899 0.9911 0.9921

6 0.9817 0.9837 0.9855 0.9871 0.9885

7 0.9754 0.9780 0.9804 0.9825 0.9845

8 0.9684 0.9717 0.9747 0.9775 0.9799

9 0.9608 0.9649 0.9686 0.9719 0.9750

10 0.9527 0.9576 0.9620 0.9660 0.9696

4.75 4.80 4.85 4.90 4.95

0.9992 0.9993 0.9994 0.9995 0.9995

0.9977 0.9980 0.9982 0.9985 0.9986

0.9956 0.9962 0.9966 0.9970 0.9974

0.9930 0.9938 0.9945 0.9952 0.9958

0.9898 0.9910 0.9920 0.9930 0.9938

0.9862 0.9878 0.9892 0.9904 0.9916

0.9822 0.9842 0.9860 0.9876 0.9890

0.9777 0.9802 0.9824 0.9844 0.9862

0.9729 0.9759 0.9786 0.9810 0.9832

5.00 5.05 5.10 5.15 5.20

0.9996 0.9996 0.9997 0.9997 0.9998

0.9988 0.9990 0.9991 0.9992 0.9993

0.9977 0.9980 0.9982 0.9985 0.9987

0.9963 0.9967 0.9971 0.9975 0.9978

0.9945 0.9952 0.9958 0.9963 0.9968

0.9926 0.9935 0.9942 0.9950 0.9956

0.9903 0.9915 0.9925 0.9934 0.9942

0.9878 0.9893 0.9906 0.9917 0.9927

0.9851 0.9869 0.9884 0.9898 0.9911

5.25 5.30 5.35 5.40 5.45

0.9998 0.9998 0.9998 0.9999 0.9999

0.9994 0.9995 0.9995 0.9996 0.9997

0.9988 0.9990 0.9991 0.9992 0.9993

0.9981 0.9983 0.9985 0.9987 0.9989

0.9972 0.9975 0.9979 0.9981 0.9984

0.9961 0.9966 0.9971 0.9974 0.9978

0.9949 0.9956 0.9961 0.9966 0.9971

0.9936 0.9944 0.9951 0.9957 0.9963

0.9922 0.9931 0.9940 0.9948 0.9954

5.50 5.55 5.60 5.65 5.70

0.9999 0.9999 0.9999 0.9999 0.9999

0.9997 0.9997 0.9998 0.9998 0.9998

0.9994 0.9995 0.9996 0.9996 0.9997

0.9990 0.9992 0.9993 0.9994 0.9995

0.9986 0.9988 0.9989 0.9991 0.9992

0.9981 0.9983 0.9985 0.9987 0.9989

0.9974 0.9978 0.9981 0.9983 0.9986

0.9968 0.9972 0.9976 0.9979 0.9982

0.9960 0.9965 0.9970 0.9974 0.9977

5.75 5.80 5.85 5.90 5.95

1.0000 0.9999 0.9999 0.9999 0.9999 0.9999

0.9997 0.9998 0.9998 0.9998 0.9998

0.9995 0.9996 0.9997 0.9997 0.9997

0.9993 0.9994 0.9995 0.9996 0.9996

0.9991 0.9992 0.9993 0.9994 0.9995

0.9988 0.9989 0.9991 0.9992 0.9993

0.9984 0.9986 0.9988 0.9990 0.9991

0.9980 0.9983 0.9985 0.9988 0.9989

6.00

0.9999 0.9999 0.9998 0.9997 0.9996 0.9994 0.9993 0.9991

c 2000 by Chapman & Hall/CRC 

Probability integral of the range W 4.50 4.55 4.60 4.65 4.70

n = 11 0.9441 0.9498 0.9550 0.9597 0.9639

12 0.9352 0.9417 0.9476 0.9530 0.9579

13 0.9258 0.9332 0.9399 0.9460 0.9516

14 0.9162 0.9244 0.9319 0.9388 0.9451

15 0.9062 0.9153 0.9236 0.9313 0.9382

16 0.8960 0.9060 0.9151 0.9235 0.9312

17 0.8856 0.8964 0.9064 0.9155 0.9240

18 0.8750 0.8867 0.8975 0.9074 0.9165

19 0.8643 0.8768 0.8884 0.8991 0.9089

20 0.8534 0.8667 0.8791 0.8906 0.9012

4.75 4.80 4.85 4.90 4.95

0.9678 0.9713 0.9745 0.9774 0.9799

0.9624 0.9665 0.9702 0.9735 0.9765

0.9567 0.9614 0.9656 0.9694 0.9728

0.9508 0.9560 0.9608 0.9650 0.9689

0.9446 0.9505 0.9557 0.9605 0.9649

0.9383 0.9447 0.9505 0.9559 0.9607

0.9317 0.9387 0.9452 0.9510 0.9563

0.9249 0.9326 0.9396 0.9460 0.9518

0.9180 0.9263 0.9339 0.9409 0.9472

0.9110 0.9199 0.9281 0.9356 0.9424

5.00 5.05 5.10 5.15 5.20

0.9822 0.9843 0.9862 0.9878 0.9893

0.9791 0.9816 0.9837 0.9856 0.9874

0.9759 0.9786 0.9811 0.9833 0.9853

0.9724 0.9756 0.9784 0.9809 0.9832

0.9688 0.9723 0.9755 0.9783 0.9809

0.9650 0.9690 0.9725 0.9757 0.9785

0.9611 0.9655 0.9694 0.9729 0.9760

0.9571 0.9618 0.9661 0.9700 0.9735

0.9529 0.9581 0.9628 0.9670 0.9708

0.9486 0.9543 0.9593 0.9639 0.9681

5.25 5.30 5.35 5.40 5.45

0.9906 0.9917 0.9928 0.9937 0.9945

0.9889 0.9903 0.9915 0.9925 0.9935

0.9871 0.9887 0.9901 0.9913 0.9924

0.9852 0.9870 0.9886 0.9900 0.9913

0.9832 0.9852 0.9870 0.9886 0.9900

0.9811 0.9833 0.9854 0.9872 0.9888

0.9789 0.9814 0.9836 0.9856 0.9874

0.9766 0.9794 0.9819 0.9841 0.9860

0.9742 0.9773 0.9800 0.9824 0.9846

0.9718 0.9751 0.9781 0.9807 0.9831

5.50 5.55 5.60 5.65 5.70

0.9952 0.9958 0.9964 0.9969 0.9973

0.9943 0.9951 0.9957 0.9963 0.9968

0.9934 0.9942 0.9950 0.9956 0.9962

0.9924 0.9933 0.9942 0.9950 0.9956

0.9913 0.9924 0.9934 0.9943 0.9950

0.9902 0.9914 0.9925 0.9935 0.9944

0.9890 0.9904 0.9916 0.9927 0.9937

0.9878 0.9893 0.9907 0.9919 0.9930

0.9865 0.9882 0.9897 0.9910 0.9922

0.9852 0.9870 0.9887 0.9901 0.9914

5.75 5.80 5.85 5.90 5.95

0.9976 0.9980 0.9982 0.9985 0.9987

0.9972 0.9976 0.9979 0.9982 0.9985

0.9967 0.9972 0.9976 0.9979 0.9982

0.9962 0.9967 0.9972 0.9976 0.9979

0.9957 0.9963 0.9968 0.9972 0.9976

0.9951 0.9958 0.9963 0.9968 0.9973

0.9945 0.9952 0.9959 0.9964 0.9969

0.9939 0.9947 0.9954 0.9960 0.9966

0.9932 0.9941 0.9949 0.9956 0.9962

0.9925 0.9935 0.9944 0.9952 0.9958

6.00

0.9989 0.9987 0.9984 0.9982 0.9979 0.9977 0.9974 0.9971 0.9967 0.9964

c 2000 by Chapman & Hall/CRC 

Probability integral of the range W 6.00 6.05 6.10 6.15 6.20

n=2 3 4 5 6 1.0000 0.9999 0.9999 0.9998 0.9997 0.9999 0.9999 0.9998 0.9997 1.0000 0.9999 0.9998 0.9998 0.9999 0.9999 0.9998 0.9999 0.9999 0.9998

7 0.9996 0.9996 0.9997 0.9997 0.9998

8 0.9994 0.9995 0.9996 0.9996 0.9997

9 0.9993 0.9994 0.9995 0.9995 0.9996

10 0.9991 0.9992 0.9993 0.9994 0.9995

6.25 6.30 6.35 6.40 6.45

0.9999 0.9999 0.9999 1.0000 0.9999 0.9999 0.9999 0.9999 0.9999 0.9999 0.9999 0.9999

0.9998 0.9998 0.9999 0.9999 0.9999

0.9997 0.9998 0.9998 0.9998 0.9999

0.9997 0.9997 0.9998 0.9998 0.9998

0.9996 0.9996 0.9997 0.9997 0.9998

1.0000 0.9999 0.9999 0.9999 0.9999 0.9999 0.9999 0.9999 0.9999 1.0000 0.9999 0.9999 0.9999 0.9999 0.9999 0.9999 1.0000 0.9999 0.9999

0.9998 0.9998 0.9999 0.9999 0.9999

6.50 6.55 6.60 6.65 6.70 6.75 6.80 6.85 6.90 6.95 7.00 7.05 7.10 7.15 7.20 7.25

c 2000 by Chapman & Hall/CRC 

1.0000 0.9999 0.9999 0.9999 0.9999 1.0000 0.9999 1.0000

Probability integral of the range W 6.00 6.05 6.10 6.15 6.20

n = 11 0.9989 0.9990 0.9992 0.9993 0.9994

12 0.9987 0.9989 0.9990 0.9992 0.9993

13 0.9984 0.9987 0.9989 0.9990 0.9992

14 0.9982 0.9985 0.9987 0.9989 0.9990

15 0.9979 0.9982 0.9985 0.9987 0.9989

16 0.9977 0.9980 0.9983 0.9985 0.9987

17 0.9974 0.9977 0.9981 0.9983 0.9986

18 0.9971 0.9975 0.9978 0.9981 0.9984

19 0.9967 0.9972 0.9976 0.9979 0.9982

20 0.9964 0.9969 0.9973 0.9977 0.9980

6.25 6.30 6.35 6.40 6.45

0.9995 0.9996 0.9996 0.9997 0.9997

0.9994 0.9995 0.9996 0.9996 0.9997

0.9993 0.9994 0.9995 0.9996 0.9996

0.9992 0.9993 0.9994 0.9995 0.9996

0.9990 0.9992 0.9993 0.9994 0.9995

0.9989 0.9991 0.9992 0.9993 0.9994

0.9988 0.9990 0.9991 0.9992 0.9994

0.9986 0.9988 0.9990 0.9992 0.9993

0.9985 0.9987 0.9989 0.9991 0.9992

0.9983 0.9986 0.9988 0.9990 0.9991

6.50 6.55 6.60 6.65 6.70

0.9998 0.9998 0.9998 0.9999 0.9999

0.9997 0.9998 0.9998 0.9998 0.9999

0.9997 0.9997 0.9998 0.9998 0.9998

0.9996 0.9997 0.9997 0.9998 0.9998

0.9996 0.9996 0.9997 0.9997 0.9998

0.9995 0.9996 0.9997 0.9997 0.9998

0.9995 0.9995 0.9996 0.9997 0.9997

0.9994 0.9995 0.9996 0.9996 0.9997

0.9993 0.9994 0.9995 0.9996 0.9997

0.9993 0.9994 0.9995 0.9995 0.9996

6.75 6.80 6.85 6.90 6.95

0.9999 0.9999 0.9999 0.9999 1.0000

0.9999 0.9999 0.9999 0.9999 0.9999

0.9999 0.9999 0.9999 0.9999 0.9999

0.9998 0.9999 0.9999 0.9999 0.9999

0.9998 0.9998 0.9999 0.9999 0.9999

0.9998 0.9998 0.9999 0.9999 0.9999

0.9998 0.9998 0.9998 0.9999 0.9999

0.9997 0.9998 0.9998 0.9998 0.9999

0.9997 0.9998 0.9998 0.9998 0.9999

0.9997 0.9997 0.9998 0.9998 0.9998

1.0000 0.9999 0.9999 0.9999 1.0000 0.9999 0.9999 1.0000 0.9999 1.0000

0.9999 0.9999 0.9999 1.0000

0.9999 0.9999 0.9999 0.9999 1.0000

0.9999 0.9999 0.9999 0.9999 0.9999

0.9999 0.9999 0.9999 0.9999 0.9999

0.9999 0.9999 0.9999 0.9999 0.9999

7.00 7.05 7.10 7.15 7.20 7.25

4.7.2

1.0000 1.0000 0.9999

Percentage points, studentized range

The standardized range is W = R/σ as defined in the previous section. If the population standard deviation σ is replaced by the sample standard deviation s (computed from another sample from the same population), then the studentized range Q is given by Q = R/S. Here, R is the range of the sample of size n and S is the independent of R and has ν degrees of freedom. The probability integral for the studentized range is given by    ∞ 1−ν/2 ν/2 ν−1 −νs2 /2 R 2 ν s e f (qs) Prob [Q ≤ q] = Prob ≤q = ds (4.64) S Γ(ν/2) 0 where f is the probability integral of the range for samples of size n. The following tables provide values of the studentized range for the normal 2 1 density function f (x) = √ e−x /2 . 2π

c 2000 by Chapman & Hall/CRC 

c 2000 by Chapman & Hall/CRC 

4.76 4.72 4.69 4.66 4.63 4.52 4.45 4.37 4.29 4.20 4.13

4.08 4.05 4.03 4.01 3.99

16 17 18 19 20

25 3.92 30 3.87 40 3.82 60 3.76 1203w.71 1000 3.64

5.09 5.00 4.93 4.86 4.81

4.40 4.33 4.27 4.22 4.18

11 12 13 14 15

6.34 5.93 5.65 5.44 5.20

5.23 4.94 4.75 4.60 4.49

6 7 8 9 10

4.87 4.79 4.70 4.60 4.50 4.42

5.16 5.12 5.07 5.03 5.00

5.57 5.46 5.37 5.29 5.22

7.05 6.55 6.12 5.89 5.71

5.12 5.03 4.93 4.82 4.71 4.62

5.45 5.39 5.35 5.30 5.27

5.91 5.78 5.68 5.59 5.51

7.44 6.91 6.54 6.27 6.07

5.32 5.22 5.10 4.99 4.87 4.78

5.67 5.61 5.56 5.51 5.47

6.17 6.03 5.92 5.82 5.74

7.87 7.28 6.87 6.57 6.35

5.48 5.37 5.25 5.13 5.00 4.91

5.92 5.79 5.73 5.68 5.64

6.48 6.33 6.20 6.09 6.00

8.22 7.57 7.13 6.92 6.68

5.61 5.50 5.37 5.25 5.12 5.02

6.08 6.01 5.95 5.89 5.84

6.68 6.51 6.38 6.26 6.17

8.51 7.83 7.48 7.14 6.88

5.73 5.61 5.48 5.35 5.21 5.11

6.23 6.15 6.08 6.03 5.97

6.85 6.67 6.53 6.41 6.31

8.77 8.05 7.69 7.33 7.06

5.89 5.71 5.57 5.44 5.30 5.20

6.35 6.27 6.20 6.14 6.09

7.00 6.82 6.67 6.55 6.44

8.99 8.38 7.87 7.50 7.22

5.99 5.80 5.66 5.52 5.37 5.27

6.47 6.38 6.31 6.25 6.19

7.13 6.95 6.80 6.67 6.56

9.19 8.56 8.04 7.65 7.36

6.07 5.94 5.73 5.60 5.44 5.34

6.57 6.48 6.41 6.34 6.29

7.25 7.06 6.90 6.77 6.66

9.37 8.72 8.18 7.79 7.49

6.15 6.01 5.80 5.66 5.50 5.40

6.66 6.57 6.50 6.43 6.37

7.36 7.17 7.01 6.87 6.76

9.54 8.87 8.32 7.91 7.60

6.23 6.08 5.87 5.72 5.56 5.46

6.74 6.66 6.58 6.51 6.45

7.46 7.26 7.10 6.96 6.84

9.69 9.01 8.44 8.03 7.71

6.29 6.14 5.93 5.78 5.61 5.51

6.82 6.73 6.65 6.58 6.52

7.56 7.35 7.19 7.05 6.93

9.83 9.13 8.56 8.13 7.81

6.36 6.20 5.98 5.83 5.66 5.56

6.90 6.81 6.72 6.65 6.59

7.65 7.44 7.27 7.12 7.00

6.42 6.26 6.03 5.88 5.70 5.61

6.97 6.87 6.79 6.72 6.66

7.73 7.52 7.34 7.20 7.07

6.47 6.31 6.08 5.93 5.75 5.65

7.03 6.94 6.85 6.78 6.71

7.81 7.59 7.42 7.27 7.14

6.52 6.36 6.13 5.97 5.79 5.69

7.09 6.99 6.91 6.84 6.77

7.88 7.66 7.48 7.33 7.20

6.57 6.41 6.17 6.01 5.82 5.73

7.15 7.05 6.97 6.89 6.82

7.95 7.73 7.55 7.39 7.26

9.96 10.09 10.20 10.31 10.42 9.25 9.36 9.47 9.57 9.66 8.66 8.77 8.86 8.95 9.03 8.23 8.33 8.42 8.50 8.58 7.91 7.99 8.08 8.15 8.23

νn = 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 1 77.75129.44147.54170.27188.43202.60215.08225.53234.69242.85250.15255.43261.60238.57243.92248.69253.42257.88262.10 2 13.58 18.33 22.09 24.66 26.66 28.28 29.68 30.87 30.63 31.76 32.78 33.71 34.57 35.36 36.10 36.79 37.44 38.05 38.63 3 8.10 10.54 12.18 13.35 13.96 14.86 15.63 16.29 16.88 17.40 17.87 18.31 18.70 19.07 19.41 19.73 20.04 20.32 20.71 4 6.33 8.11 9.20 9.75 10.46 11.03 11.52 11.94 12.31 12.64 12.94 13.22 13.47 13.70 13.92 14.13 14.32 14.50 14.67 5 5.64 6.98 7.82 8.28 8.81 9.23 9.59 9.91 10.18 10.43 10.65 10.86 11.05 11.22 11.38 11.53 11.67 11.81 11.94

Upper 1% points of the studentized range The entries are q.01 where Prob [Q < q.01 ] = .99.

3.46 3.35 3.26 3.20 3.15

3.11 3.08 3.06 3.04 3.02

3.00 2.99 2.97 2.95 2.94

2.91 2.88 2.86 2.83 2.80 2.77

6 7 8 9 10

11 12 13 14 15

c 2000 by Chapman & Hall/CRC 

16 17 18 19 20

25 30 40 60 120 1000

3.55 3.51 3.47 3.43 3.36 3.35

3.67 3.65 3.63 3.62 3.60

3.83 3.78 3.75 3.72 3.69

4.34 4.15 4.03 3.95 3.88

3.92 3.88 3.83 3.78 3.68 3.68

4.07 4.05 4.02 4.00 3.99

4.27 4.21 4.17 4.13 4.10

4.88 4.67 4.53 4.42 4.33

4.19 4.14 4.09 4.03 3.91 3.92

4.35 4.33 4.30 4.28 4.26

4.58 4.52 4.46 4.42 4.38

5.28 5.05 4.88 4.75 4.65

4.39 4.34 4.28 4.22 4.10 4.11

4.57 4.54 4.52 4.49 4.47

4.82 4.75 4.69 4.65 4.61

5.63 5.36 5.15 5.01 4.90

4.56 4.51 4.44 4.38 4.25 4.25

4.75 4.72 4.69 4.67 4.65

5.02 4.94 4.88 4.83 4.79

5.90 5.61 5.40 5.25 5.11

4.71 4.65 4.58 4.52 4.38 4.37

4.90 4.87 4.84 4.81 4.79

5.20 5.11 5.04 4.99 4.94

6.12 5.82 5.60 5.43 5.31

4.83 4.77 4.70 4.63 4.49 4.48

5.03 5.00 4.97 4.94 4.92

5.35 5.27 5.18 5.12 5.08

6.32 6.00 5.77 5.60 5.46

4.94 4.88 4.80 4.73 4.59 4.58

5.15 5.11 5.08 5.05 5.03

5.49 5.40 5.32 5.24 5.19

6.49 6.16 5.92 5.74 5.60

5.03 4.97 4.90 4.81 4.67 4.67

5.26 5.22 5.18 5.15 5.13

5.61 5.51 5.43 5.35 5.30

6.65 6.30 6.06 5.87 5.72

5.12 5.06 4.98 4.89 4.75 4.75

5.35 5.31 5.28 5.24 5.22

5.71 5.62 5.53 5.46 5.39

6.79 6.43 6.18 5.98 5.83

5.20 5.14 5.06 4.97 4.83 4.82

5.44 5.39 5.36 5.33 5.30

5.81 5.71 5.63 5.56 5.48

6.92 6.55 6.29 6.09 5.94

5.28 5.21 5.13 5.04 4.89 4.88

5.51 5.47 5.44 5.40 5.38

5.90 5.80 5.71 5.64 5.56

7.04 6.66 6.39 6.19 6.03

5.35 5.28 5.20 5.10 4.95 4.95

5.59 5.55 5.51 5.48 5.45

5.99 5.88 5.79 5.71 5.64

7.14 6.76 6.48 6.28 6.12

5.41 5.34 5.26 5.16 5.01 5.00

5.66 5.61 5.57 5.54 5.52

6.06 5.95 5.86 5.79 5.70

7.24 6.85 6.57 6.36 6.19

5.47 5.40 5.30 5.22 5.07 5.06

5.72 5.68 5.64 5.60 5.58

6.14 6.02 5.93 5.85 5.79

7.34 6.94 6.65 6.44 6.27

5.53 5.46 5.35 5.27 5.12 5.11

5.78 5.73 5.70 5.66 5.63

6.20 6.09 6.00 5.92 5.85

7.43 7.02 6.73 6.51 6.34

5.58 5.51 5.40 5.32 5.16 5.15

5.84 5.79 5.75 5.72 5.69

6.27 6.15 6.06 5.97 5.90

7.51 7.10 6.80 6.58 6.41

5.63 5.56 5.45 5.36 5.20 5.20

5.89 5.84 5.80 5.77 5.74

6.33 6.21 6.11 6.03 5.96

7.59 7.17 6.87 6.64 6.47

ν n=2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 1 17.79 26.70 32.79 37.07 39.84 42.67 45.05 47.10 48.89 50.49 51.91 53.14 54.28 55.36 56.34 57.22 58.05 58.83 59.56 2 6.10 8.31 9.81 10.89 11.70 12.52 13.25 13.87 14.42 14.91 15.35 15.75 16.13 16.47 16.79 17.09 17.38 17.64 17.90 3 4.50 5.91 6.83 7.46 8.03 8.50 8.90 9.24 9.56 9.84 10.09 10.31 10.52 10.72 10.90 11.06 11.22 11.37 11.52 4 3.93 5.04 5.76 6.26 6.68 7.03 7.32 7.58 7.80 8.00 8.19 8.35 8.50 8.65 8.78 8.90 9.02 9.12 9.23 5 3.63 4.60 5.22 5.65 6.04 6.33 6.58 6.80 7.00 7.17 7.33 7.47 7.60 7.72 7.83 7.94 8.03 8.12 8.21

Upper 5% points of the studentized range The entries are q.05 where Prob [Q < q.05 ] = .95.

2.75 2.68 2.63 2.59 2.56

2.54 2.52 2.51 2.49 2.48

2.47 2.46 2.45 2.44 2.43

2.41 2.40 2.38 2.36 2.34 2.33

6 7 8 9 10

11 12 13 14 15

c 2000 by Chapman & Hall/CRC 

16 17 18 19 20

25 30 40 60 120 1000

3.04 3.02 2.99 2.96 2.93 2.90

3.12 3.11 3.10 3.09 3.08

3.23 3.20 3.18 3.16 3.14

3.56 3.45 3.37 3.32 3.27

3.42 3.39 3.35 3.31 3.28 3.24

3.52 3.50 3.49 3.47 3.46

3.66 3.62 3.59 3.56 3.54

4.07 3.93 3.83 3.76 3.70

3.68 3.65 3.61 3.56 3.52 3.48

3.81 3.78 3.77 3.75 3.74

3.97 3.92 3.89 3.85 3.83

4.44 4.28 4.17 4.08 4.02

3.89 3.85 3.80 3.76 3.71 3.66

4.03 4.00 3.98 3.97 3.95

4.20 4.16 4.12 4.08 4.05

4.73 4.55 4.43 4.34 4.26

4.06 4.02 3.96 3.91 3.86 3.81

4.21 4.18 4.16 4.14 4.12

4.40 4.35 4.30 4.27 4.24

4.97 4.78 4.65 4.54 4.47

4.20 4.16 4.10 4.04 3.99 3.93

4.36 4.33 4.31 4.29 4.27

4.57 4.51 4.46 4.42 4.39

5.17 4.97 4.83 4.72 4.64

4.32 4.28 4.22 4.16 4.10 4.04

4.49 4.46 4.44 4.42 4.40

4.71 4.65 4.60 4.56 4.52

5.34 5.14 4.99 4.87 4.78

4.43 4.38 4.32 4.25 4.19 4.13

4.61 4.58 4.55 4.53 4.51

4.84 4.78 4.72 4.68 4.64

5.50 5.28 5.12 5.01 4.91

4.53 4.47 4.41 4.34 4.28 4.21

4.71 4.68 4.66 4.63 4.61

4.95 4.89 4.83 4.79 4.75

5.64 5.41 5.25 5.13 5.03

4.62 4.56 4.49 4.42 4.35 4.28

4.81 4.77 4.75 4.72 4.70

5.05 4.99 4.93 4.88 4.84

5.76 5.53 5.36 5.23 5.13

4.69 4.64 4.56 4.49 4.42 4.35

4.89 4.86 4.83 4.80 4.78

5.15 5.08 5.02 4.97 4.93

5.87 5.64 5.46 5.33 5.23

4.77 4.71 4.63 4.56 4.49 4.41

4.97 4.93 4.91 4.88 4.86

5.23 5.16 5.10 5.05 5.01

5.98 5.73 5.56 5.42 5.32

4.83 4.77 4.69 4.62 4.54 4.48

5.04 5.01 4.98 4.95 4.92

5.31 5.24 5.18 5.12 5.08

6.07 5.82 5.64 5.50 5.40

4.89 4.83 4.75 4.68 4.60 4.53

5.11 5.07 5.04 5.01 4.99

5.38 5.31 5.25 5.19 5.15

6.16 5.91 5.72 5.58 5.47

4.95 4.89 4.81 4.73 4.65 4.59

5.17 5.13 5.10 5.07 5.05

5.45 5.37 5.31 5.26 5.21

6.25 5.99 5.80 5.65 5.54

5.00 4.94 4.86 4.78 4.69 4.64

5.23 5.19 5.16 5.13 5.10

5.51 5.44 5.37 5.32 5.27

6.32 6.06 5.87 5.72 5.61

5.06 4.99 4.91 4.82 4.74 4.68

5.28 5.24 5.21 5.18 5.16

5.57 5.49 5.43 5.37 5.32

6.40 6.13 5.93 5.79 5.67

5.10 5.04 4.95 4.87 4.78 4.73

5.33 5.30 5.26 5.23 5.21

5.63 5.55 5.48 5.43 5.38

6.46 6.19 6.00 5.84 5.73

ν n=2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 1 8.94 13.43 16.37 18.43 20.12 21.49 22.63 23.61 24.47 25.23 25.91 26.52 26.81 27.46 28.06 28.62 29.15 29.65 30.18 2 4.13 5.73 6.77 7.54 8.14 8.63 9.05 9.41 9.73 10.01 10.26 10.49 10.70 10.89 11.07 11.24 11.40 11.54 11.63 3 3.33 4.47 5.20 5.74 6.16 6.51 6.81 7.06 7.29 7.49 7.67 7.83 7.98 8.12 8.25 8.37 8.48 8.58 8.68 4 3.01 3.98 4.59 5.03 5.39 5.68 5.92 6.14 6.33 6.49 6.64 6.78 6.91 7.02 7.13 7.23 7.33 7.41 7.50 5 2.85 3.72 4.26 4.66 4.98 5.24 5.46 5.65 5.81 5.96 6.10 6.22 6.33 6.44 6.53 6.62 6.71 6.79 6.86

Upper 10% points of the studentized range The entries are q.10 where Prob [Q < q.10 ] = .90.

CHAPTER 5

Discrete Probability Distributions Contents 5.1

5.2

5.3 5.4

5.5

5.6

5.7

5.8

5.9

Bernoulli distribution 5.1.1 Properties 5.1.2 Variates Beta binomial distribution 5.2.1 Properties 5.2.2 Variates Beta Pascal distribution 5.3.1 Properties Binomial distribution 5.4.1 Properties 5.4.2 Variates 5.4.3 Tables Geometric distribution 5.5.1 Properties 5.5.2 Variates 5.5.3 Tables Hypergeometric distribution 5.6.1 Properties 5.6.2 Variates 5.6.3 Tables Multinomial distribution 5.7.1 Properties 5.7.2 Variates Negative binomial distribution 5.8.1 Properties 5.8.2 Variates 5.8.3 Tables Poisson distribution 5.9.1 Properties

c 2000 by Chapman & Hall/CRC 

5.9.2 Variates 5.9.3 Tables 5.10 Rectangular (discrete uniform) distribution 5.10.1 Properties

This chapter presents some common discrete probability distributions along with their properties. Relevant numerical tables are also included. Notation used throughout this chapter: Probability mass function (pmf)

p(x) = Prob [X = x]

Mean

µ = E [X]   σ 2 = E (X − µ)2   β1 = E (X − µ)3   β2 = E (X − µ)4   m(t) = E etX   φ(t) = E eitX   P (t) = E tX

Variance Coefficient of skewness Coefficient of kurtosis Moment generating function (mgf) Characteristic function (char function) Factorial moment generating function (fact mgf) 5.1

BERNOULLI DISTRIBUTION

A Bernoulli distribution is used to describe an experiment in which there are only two possible outcomes, typically a success or a failure. This type of experiment is called a Bernoulli trial, or simply a trial. The probability of a success is p and a sequence of Bernoulli trials is referred to as repeated trials. 5.1.1

Properties (

pmf

p(x) =

q p

x=0 x=1

(or px q 1−x for x = 0, 1)

0 ≤ p ≤ 1, q = 1 − p mean

µ=p

variance

σ 2 = pq

skewness

1 − 2p β1 = √ pq

kurtosis

β2 = 3 +

1 − 6pq pq

mgf m(t) = q + pet

c 2000 by Chapman & Hall/CRC 

char function

φ(t) = q + peit

fact mgf P (t) = q + pt 5.1.2

Variates

(1) Let X1 , X2 , . . . , Xn be independent, identically distributed (iid) Bernoulli random variables with probability of a success p. The random variable Y = X1 + X2 + · · · + Xn has a binomial distribution with parameters n and p. 5.2

BETA BINOMIAL DISTRIBUTION

The beta binomial distribution is also known as the negative hypergeometric distribution, inverse hypergeometric distribution, hypergeometric waiting–time distribution, and Markov–P´ olya distribution. 5.2.1

Properties  pmf

mean variance

skewness

p(x) =

  *  −a −b −a − b x = 0, 1, . . . , n n n−x n

a, b, n > 0, n an integer an µ= a+b abn(a + b + n) σ2 = (a + b)2 (a + b + 1) + (a + b + 1) (a − b)(a + b + 2n) β1 = abn(a + b + n) (a + b + 2)

(a + b)2 (a + b + 1) × abn(a + b + n)(a + b + 2)(a + b + 3)  (a + b)(a + b + 1 + 6n) + 3ab(n − 2) + 6n2  3abn(6 − n) 18abn2 +3 − − a+b (a + b)2 mgf m(t) = 2 F1 (a, −n; a + b; −et )

kurtosis

char function

β2 =

φ(t) = 2 F1 (a, −n; a + b; e−it )

fact mgf P (t) = 2 F1 (a, −n; a + b; −t) where p Fq is the generalized hypergeometric function defined in Chapter 18 (see page 520).

c 2000 by Chapman & Hall/CRC 

5.2.2

Variates

Let X be a beta binomial random variable with parameters a, b, and n. (1) If a = b = 1 and n is reduced by 1, then X is a rectangular (discrete uniform) random variable. (2) As n → ∞, X is approximately a binomial random variable with parameters n and p = a/b. 5.3

BETA PASCAL DISTRIBUTION

The beta Pascal distribution arises from a special case of the urn scheme. 5.3.1

Properties

pmf p(x) =

Γ(x)Γ(ν)Γ(ρ + ν)Γ(ν + x − (ρ + r)) Γ(r)Γ(x − r + 1)Γ(ρ)Γ(ν − ρ)Γ(ν + x) x = r, r + 1, . . . , r ∈ N , ν > ρ > 0

mean variance

µ=r

ν−1 , ρ>1 ρ−1

σ 2 = r(r + ρ − 1)

(ν − 1)(ν − ρ) , ρ>2 (ρ − 1)2 (ρ − 2)

where Γ(x) is the gamma function defined in Chapter 18 (see page 515). 5.4

BINOMIAL DISTRIBUTION

The binomial distribution is used to characterize the number of successes in n Bernoulli trials. It is used to model some very common experiments in which a sample of size n is taken from an infinite population such that each element is selected independently and has the same probability, p, of having a specified attribute. 5.4.1

Properties   n x n−x pmf p(x) = p q x = 0, 1, 2, . . . , n x 0 ≤ p ≤ 1, q = 1 − p

mean

µ = np

variance

σ 2 = npq

skewness

1 − 2p β1 = √ npq

kurtosis

β2 = 3 +

1 − 6pq npq

c 2000 by Chapman & Hall/CRC 

mgf m(t) = (q + pet )n char function

φ(t) = (q + peit )n

fact mgf P (t) = (q + pt)n 5.4.2

Variates

Let X be a binomial random variable with parameters n and p. (1) If n = 1, then X is a Bernoulli random variable with probability of success p. (2) As n → ∞ if np ≥ 5 and n(1 − p) ≥ 5, then X is approximately normal with parameters µ = np and σ 2 = np(1 − p). (3) As n → ∞ if p < 0.1 and np < 10, then X is approximately a Poisson random variable with parameter λ = np. (4) Let X1 , . . . , Xk be independent, binomial random variables with parameters ni and p, respectively. The random variable Y = X1 +X2 +· · ·+Xk has a binomial distribution with parameters n = n1 +n2 +· · ·+nk and p. 5.4.3

Tables

The following tables only contain values of p up to p = 1/2. By symmetry (replacing p with 1 − p and replacing x with n − x) values for p > 1/2 can be reduced to the present tables. Example 5.32 : A biased coin has a probability of heads of .75; what is the probability of obtaining 5 or more heads in 8 flips? Solution: (S1) The answer is given by looking in cumulative distribution tables with n = 8, x = 5, and p = 0.75. (S2) Making the substitutions mentioned above this is the same as n = 8, x = 3 and p = 0.25. (S3) This value is in the tables and is equal to 0.8862. Hence, 89% of the time 5 or more heads would be likely to occur. (S4) This result can be interpreted as the likelihood of flipping a biased coin that has a probability of tails equal to 25% and asking how likely it is to have 3 or fewer tails. (S5) Note: The following tables are for the cumulative distribution function, not the probability mass function. The probability of obtaining exactly 5 heads in 8 flips   is 85 (0.75)5 (.25)3 = .2076.

Example 5.33 : The probability a randomly selected home in Columbia County will lose power during a summer storm is .25. Suppose 14 homes in this county are selected at random. What is the probability exactly 4 homes will lose power, more than 6 will lose power, and between 2 and 7 (inclusive) will lose power?

c 2000 by Chapman & Hall/CRC 

Solution: (S1) Let X be the number of homes (out of 14) that will lose power. The random variable X has a binomial distribution with parameters n = 14 and p = 0.25. Use the cumulative terms for the binomial distribution to answer each probability question. (S2) Prob [X = 4] = Prob [X ≤ 4] − Prob [X ≤ 3] = 0.7415 − 0.5213 = 0.2202 (S3) Prob [X > 6] = 1 − Prob [X ≤ 6] = 1 − 0.9617 = 0.0383 (S4) Prob [2 ≤ X ≤ 7] = Prob [X ≤ 7] − Prob [X ≤ 1] = 0.9897 − 0.1010 = 0.8887

c 2000 by Chapman & Hall/CRC 

Cumulative probability, Binomial distribution n x p =0.05 0.10 0.15 0.20 0.25 0.30 0.40 0.50 2 0 0.9025 0.8100 0.7225 0.6400 0.5625 0.4900 0.3600 0.2500 1 0.9975 0.9900 0.9775 0.9600 0.9375 0.9100 0.8400 0.7500 3 0 0.8574 0.7290 0.6141 0.5120 0.4219 0.3430 0.2160 0.1250 1 0.9928 0.9720 0.9393 0.8960 0.8438 0.7840 0.6480 0.5000 2 0.9999 0.9990 0.9966 0.9920 0.9844 0.9730 0.9360 0.8750 4 0 0.8145 0.6561 0.5220 0.4096 0.3164 0.2401 0.1296 0.0625 1 0.9860 0.9477 0.8905 0.8192 0.7383 0.6517 0.4752 0.3125 2 0.9995 0.9963 0.9880 0.9728 0.9492 0.9163 0.8208 0.6875 3 1.0000 0.9999 0.9995 0.9984 0.9961 0.9919 0.9744 0.9375 5 0 0.7738 0.5905 0.4437 0.3277 0.2373 0.1681 0.0778 0.0313 1 0.9774 0.9185 0.8352 0.7373 0.6328 0.5282 0.3370 0.1875 2 0.9988 0.9914 0.9734 0.9421 0.8965 0.8369 0.6826 0.5000 3 1.0000 0.9995 0.9978 0.9933 0.9844 0.9692 0.9130 0.8125 4 1.0000 1.0000 0.9999 0.9997 0.9990 0.9976 0.9898 0.9688 6 0 0.7351 0.5314 0.3771 0.2621 0.1780 0.1177 0.0467 0.0156 1 0.9672 0.8857 0.7765 0.6554 0.5339 0.4202 0.2333 0.1094 2 0.9978 0.9841 0.9527 0.9011 0.8306 0.7443 0.5443 0.3438 3 0.9999 0.9987 0.9941 0.9830 0.9624 0.9295 0.8208 0.6563 4 1.0000 1.0000 0.9996 0.9984 0.9954 0.9891 0.9590 0.8906 5 1.0000 1.0000 1.0000 0.9999 0.9998 0.9993 0.9959 0.9844 7 0 0.6983 0.4783 0.3206 0.2097 0.1335 0.0824 0.0280 0.0078 1 0.9556 0.8503 0.7166 0.5767 0.4450 0.3294 0.1586 0.0625 2 0.9962 0.9743 0.9262 0.8520 0.7564 0.6471 0.4199 0.2266 3 0.9998 0.9973 0.9879 0.9667 0.9294 0.8740 0.7102 0.5000 4 1.0000 0.9998 0.9988 0.9953 0.9871 0.9712 0.9037 0.7734 5 1.0000 1.0000 0.9999 0.9996 0.9987 0.9962 0.9812 0.9375 6 1.0000 1.0000 1.0000 1.0000 0.9999 0.9998 0.9984 0.9922 8 0 0.6634 0.4305 0.2725 0.1678 0.1001 0.0576 0.0168 0.0039 1 0.9428 0.8131 0.6572 0.5033 0.3671 0.2553 0.1064 0.0352 2 0.9942 0.9619 0.8948 0.7969 0.6785 0.5518 0.3154 0.1445 3 0.9996 0.9950 0.9787 0.9437 0.8862 0.8059 0.5941 0.3633 4 1.0000 0.9996 0.9971 0.9896 0.9727 0.9420 0.8263 0.6367 5 1.0000 1.0000 0.9998 0.9988 0.9958 0.9887 0.9502 0.8555 6 1.0000 1.0000 1.0000 0.9999 0.9996 0.9987 0.9915 0.9648 7 1.0000 1.0000 1.0000 1.0000 1.0000 0.9999 0.9993 0.9961 9 0 0.6302 0.3874 0.2316 0.1342 0.0751 0.0403 0.0101 0.0019 1 0.9288 0.7748 0.5995 0.4362 0.3003 0.1960 0.0705 0.0195 2 0.9916 0.9470 0.8591 0.7382 0.6007 0.4628 0.2318 0.0898 3 0.9994 0.9917 0.9661 0.9144 0.8343 0.7297 0.4826 0.2539 4 1.0000 0.9991 0.9944 0.9804 0.9511 0.9012 0.7334 0.5000 5 1.0000 0.9999 0.9994 0.9969 0.9900 0.9747 0.9006 0.7461 6 1.0000 1.0000 1.0000 0.9997 0.9987 0.9957 0.9750 0.9102 7 1.0000 1.0000 1.0000 1.0000 0.9999 0.9996 0.9962 0.9805 8 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9997 0.9980

c 2000 by Chapman & Hall/CRC 

Cumulative probability, Binomial distribution n 10

x p =0.05 0.10 0.15 0.20 0.25 0.30 0.40 0.50 0 0.5987 0.3487 0.1969 0.1074 0.0563 0.0283 0.0060 0.0010 1 0.9139 0.7361 0.5443 0.3758 0.2440 0.1493 0.0464 0.0107 2 0.9885 0.9298 0.8202 0.6778 0.5256 0.3828 0.1673 0.0547 3 0.9990 0.9872 0.9500 0.8791 0.7759 0.6496 0.3823 0.1719 4 0.9999 0.9984 0.9901 0.9672 0.9219 0.8497 0.6331 0.3770 5 1.0000 0.9999 0.9986 0.9936 0.9803 0.9526 0.8338 0.6230 6 1.0000 1.0000 0.9999 0.9991 0.9965 0.9894 0.9452 0.8281 7 1.0000 1.0000 1.0000 0.9999 0.9996 0.9984 0.9877 0.9453 8 1.0000 1.0000 1.0000 1.0000 1.0000 0.9999 0.9983 0.9893 9 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9999 0.9990 11 0 0.5688 0.3138 0.1673 0.0859 0.0422 0.0198 0.0036 0.0005 1 0.8981 0.6974 0.4922 0.3221 0.1971 0.1130 0.0302 0.0059 2 0.9848 0.9104 0.7788 0.6174 0.4552 0.3127 0.1189 0.0327 3 0.9984 0.9815 0.9306 0.8389 0.7133 0.5696 0.2963 0.1133 4 0.9999 0.9972 0.9841 0.9496 0.8854 0.7897 0.5328 0.2744 5 1.0000 0.9997 0.9973 0.9883 0.9657 0.9218 0.7535 0.5000 6 1.0000 1.0000 0.9997 0.9980 0.9924 0.9784 0.9006 0.7256 7 1.0000 1.0000 1.0000 0.9998 0.9988 0.9957 0.9707 0.8867 8 1.0000 1.0000 1.0000 1.0000 0.9999 0.9994 0.9941 0.9673 9 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9993 0.9941 10 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9995 12 0 0.5404 0.2824 0.1422 0.0687 0.0317 0.0138 0.0022 0.0002 1 0.8816 0.6590 0.4435 0.2749 0.1584 0.0850 0.0196 0.0032 2 0.9804 0.8891 0.7358 0.5584 0.3907 0.2528 0.0834 0.0193 3 0.9978 0.9744 0.9078 0.7946 0.6488 0.4925 0.2253 0.0730 4 0.9998 0.9957 0.9761 0.9274 0.8424 0.7237 0.4382 0.1938 5 1.0000 0.9995 0.9954 0.9806 0.9456 0.8821 0.6652 0.3872 6 1.0000 1.0000 0.9993 0.9961 0.9858 0.9614 0.8418 0.6128 7 1.0000 1.0000 0.9999 0.9994 0.9972 0.9905 0.9427 0.8062 8 1.0000 1.0000 1.0000 0.9999 0.9996 0.9983 0.9847 0.9270 9 1.0000 1.0000 1.0000 1.0000 1.0000 0.9998 0.9972 0.9807 10 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9997 0.9968 11 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9998 13 0 0.5133 0.2542 0.1209 0.0550 0.0238 0.0097 0.0013 0.0001 1 0.8646 0.6213 0.3983 0.2336 0.1267 0.0637 0.0126 0.0017 2 0.9755 0.8661 0.6920 0.5017 0.3326 0.2025 0.0579 0.0112 3 0.9969 0.9658 0.8820 0.7473 0.5843 0.4206 0.1686 0.0461 4 0.9997 0.9935 0.9658 0.9009 0.7940 0.6543 0.3530 0.1334 5 1.0000 0.9991 0.9925 0.9700 0.9198 0.8346 0.5744 0.2905 6 1.0000 0.9999 0.9987 0.9930 0.9757 0.9376 0.7712 0.5000 7 1.0000 1.0000 0.9998 0.9988 0.9943 0.9818 0.9023 0.7095 8 1.0000 1.0000 1.0000 0.9998 0.9990 0.9960 0.9679 0.8666 9 1.0000 1.0000 1.0000 1.0000 0.9999 0.9993 0.9922 0.9539 10 1.0000 1.0000 1.0000 1.0000 1.0000 0.9999 0.9987 0.9888 11 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9999 0.9983 12 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9999 c 2000 by Chapman & Hall/CRC 

Cumulative probability, Binomial distribution n 14

x p =0.05 0.10 0.15 0.20 0.25 0.30 0.40 0.50 0 0.4877 0.2288 0.1028 0.0440 0.0178 0.0068 0.0008 0.0001 1 0.8470 0.5846 0.3567 0.1979 0.1010 0.0475 0.0081 0.0009 2 0.9699 0.8416 0.6479 0.4481 0.2811 0.1608 0.0398 0.0065 3 0.9958 0.9559 0.8535 0.6982 0.5213 0.3552 0.1243 0.0287 4 0.9996 0.9908 0.9533 0.8702 0.7415 0.5842 0.2793 0.0898 5 1.0000 0.9985 0.9885 0.9562 0.8883 0.7805 0.4859 0.2120 6 1.0000 0.9998 0.9978 0.9884 0.9617 0.9067 0.6925 0.3953 7 1.0000 1.0000 0.9997 0.9976 0.9897 0.9685 0.8499 0.6047 8 1.0000 1.0000 1.0000 0.9996 0.9979 0.9917 0.9417 0.7880 9 1.0000 1.0000 1.0000 1.0000 0.9997 0.9983 0.9825 0.9102 10 1.0000 1.0000 1.0000 1.0000 1.0000 0.9998 0.9961 0.9713 11 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9994 0.9935 12 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9999 0.9991 13 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9999 15 0 0.4633 0.2059 0.0873 0.0352 0.0134 0.0047 0.0005 0.0000 1 0.8290 0.5490 0.3186 0.1671 0.0802 0.0353 0.0052 0.0005 2 0.9638 0.8159 0.6042 0.3980 0.2361 0.1268 0.0271 0.0037 3 0.9945 0.9444 0.8227 0.6482 0.4613 0.2969 0.0905 0.0176 4 0.9994 0.9873 0.9383 0.8358 0.6865 0.5155 0.2173 0.0592 5 1.0000 0.9978 0.9832 0.9389 0.8516 0.7216 0.4032 0.1509 6 1.0000 0.9997 0.9964 0.9819 0.9434 0.8689 0.6098 0.3036 7 1.0000 1.0000 0.9994 0.9958 0.9827 0.9500 0.7869 0.5000 8 1.0000 1.0000 0.9999 0.9992 0.9958 0.9848 0.9050 0.6964 9 1.0000 1.0000 1.0000 0.9999 0.9992 0.9963 0.9662 0.8491 10 1.0000 1.0000 1.0000 1.0000 0.9999 0.9993 0.9907 0.9408 11 1.0000 1.0000 1.0000 1.0000 1.0000 0.9999 0.9981 0.9824 12 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9997 0.9963 13 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9995 14 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 16 0 0.4401 0.1853 0.0742 0.0282 0.0100 0.0033 0.0003 0.0000 1 0.8108 0.5147 0.2839 0.1407 0.0635 0.0261 0.0033 0.0003 2 0.9571 0.7893 0.5614 0.3518 0.1971 0.0994 0.0183 0.0021 3 0.9930 0.9316 0.7899 0.5981 0.4050 0.2459 0.0651 0.0106 4 0.9991 0.9830 0.9210 0.7983 0.6302 0.4499 0.1666 0.0384 5 0.9999 0.9967 0.9765 0.9183 0.8104 0.6598 0.3288 0.1051 6 1.0000 0.9995 0.9944 0.9733 0.9204 0.8247 0.5272 0.2273 7 1.0000 0.9999 0.9989 0.9930 0.9729 0.9256 0.7161 0.4018 8 1.0000 1.0000 0.9998 0.9985 0.9925 0.9743 0.8577 0.5982 9 1.0000 1.0000 1.0000 0.9998 0.9984 0.9929 0.9417 0.7728 10 1.0000 1.0000 1.0000 1.0000 0.9997 0.9984 0.9809 0.8949 11 1.0000 1.0000 1.0000 1.0000 1.0000 0.9997 0.9951 0.9616 12 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9991 0.9894 13 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9999 0.9979 14 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9997 15 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000

c 2000 by Chapman & Hall/CRC 

Cumulative probability, Binomial distribution n 17

x p =0.05 0.10 0.15 0.20 0.25 0.30 0.40 0.50 0 0.4181 0.1668 0.0631 0.0225 0.0075 0.0023 0.0002 0.0000 1 0.7922 0.4818 0.2525 0.1182 0.0501 0.0193 0.0021 0.0001 2 0.9497 0.7618 0.5198 0.3096 0.1637 0.0774 0.0123 0.0012 3 0.9912 0.9174 0.7556 0.5489 0.3530 0.2019 0.0464 0.0064 4 0.9988 0.9779 0.9013 0.7582 0.5739 0.3887 0.1260 0.0245 5 0.9999 0.9953 0.9681 0.8943 0.7653 0.5968 0.2639 0.0717 6 1.0000 0.9992 0.9917 0.9623 0.8929 0.7752 0.4478 0.1661 7 1.0000 0.9999 0.9983 0.9891 0.9598 0.8954 0.6405 0.3145 8 1.0000 1.0000 0.9997 0.9974 0.9876 0.9597 0.8011 0.5000 9 1.0000 1.0000 1.0000 0.9995 0.9969 0.9873 0.9081 0.6855 10 1.0000 1.0000 1.0000 0.9999 0.9994 0.9968 0.9652 0.8338 11 1.0000 1.0000 1.0000 1.0000 0.9999 0.9993 0.9894 0.9283 12 1.0000 1.0000 1.0000 1.0000 1.0000 0.9999 0.9975 0.9755 13 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9996 0.9936 14 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9999 0.9988 15 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9999 16 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 18 0 0.3972 0.1501 0.0537 0.0180 0.0056 0.0016 0.0001 0.0000 1 0.7735 0.4503 0.2240 0.0991 0.0395 0.0142 0.0013 0.0001 2 0.9419 0.7338 0.4797 0.2713 0.1353 0.0600 0.0082 0.0007 3 0.9891 0.9018 0.7202 0.5010 0.3057 0.1646 0.0328 0.0038 4 0.9984 0.9718 0.8794 0.7164 0.5187 0.3327 0.0942 0.0154 5 0.9998 0.9936 0.9581 0.8671 0.7175 0.5344 0.2088 0.0481 6 1.0000 0.9988 0.9882 0.9487 0.8610 0.7217 0.3743 0.1189 7 1.0000 0.9998 0.9973 0.9837 0.9431 0.8593 0.5634 0.2403 8 1.0000 1.0000 0.9995 0.9958 0.9807 0.9404 0.7368 0.4073 9 1.0000 1.0000 0.9999 0.9991 0.9946 0.9790 0.8653 0.5927 10 1.0000 1.0000 1.0000 0.9998 0.9988 0.9939 0.9424 0.7597 11 1.0000 1.0000 1.0000 1.0000 0.9998 0.9986 0.9797 0.8811 12 1.0000 1.0000 1.0000 1.0000 1.0000 0.9997 0.9942 0.9519 13 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9987 0.9846 14 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9998 0.9962 15 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9993 16 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9999 17 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000

c 2000 by Chapman & Hall/CRC 

Cumulative probability, Binomial distribution n 19

x p =0.05 0.10 0.15 0.20 0.25 0.30 0.40 0.50 0 0.3774 0.1351 0.0456 0.0144 0.0042 0.0011 0.0001 0.0000 1 0.7547 0.4203 0.1985 0.0829 0.0310 0.0104 0.0008 0.0000 2 0.9335 0.7054 0.4413 0.2369 0.1113 0.0462 0.0055 0.0004 3 0.9868 0.8850 0.6842 0.4551 0.2631 0.1332 0.0230 0.0022 4 0.9980 0.9648 0.8556 0.6733 0.4654 0.2822 0.0696 0.0096 5 0.9998 0.9914 0.9463 0.8369 0.6678 0.4739 0.1629 0.0318 6 1.0000 0.9983 0.9837 0.9324 0.8251 0.6655 0.3081 0.0835 7 1.0000 0.9997 0.9959 0.9767 0.9225 0.8180 0.4878 0.1796 8 1.0000 1.0000 0.9992 0.9933 0.9712 0.9161 0.6675 0.3238 9 1.0000 1.0000 0.9999 0.9984 0.9911 0.9675 0.8139 0.5000 10 1.0000 1.0000 1.0000 0.9997 0.9977 0.9895 0.9115 0.6762 11 1.0000 1.0000 1.0000 1.0000 0.9995 0.9972 0.9648 0.8204 12 1.0000 1.0000 1.0000 1.0000 0.9999 0.9994 0.9884 0.9165 13 1.0000 1.0000 1.0000 1.0000 1.0000 0.9999 0.9969 0.9682 14 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9994 0.9904 15 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9999 0.9978 16 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9996 17 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 18 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 20 0 0.3585 0.1216 0.0388 0.0115 0.0032 0.0008 0.0000 0.0000 1 0.7358 0.3917 0.1756 0.0692 0.0243 0.0076 0.0005 0.0000 2 0.9245 0.6769 0.4049 0.2061 0.0913 0.0355 0.0036 0.0002 3 0.9841 0.8670 0.6477 0.4114 0.2252 0.1071 0.0160 0.0013 4 0.9974 0.9568 0.8298 0.6297 0.4148 0.2375 0.0510 0.0059 5 0.9997 0.9888 0.9327 0.8042 0.6172 0.4164 0.1256 0.0207 6 1.0000 0.9976 0.9781 0.9133 0.7858 0.6080 0.2500 0.0577 7 1.0000 0.9996 0.9941 0.9679 0.8982 0.7723 0.4159 0.1316 8 1.0000 0.9999 0.9987 0.9900 0.9591 0.8867 0.5956 0.2517 9 1.0000 1.0000 0.9998 0.9974 0.9861 0.9520 0.7553 0.4119 10 1.0000 1.0000 1.0000 0.9994 0.9961 0.9829 0.8725 0.5881 11 1.0000 1.0000 1.0000 0.9999 0.9991 0.9949 0.9435 0.7483 12 1.0000 1.0000 1.0000 1.0000 0.9998 0.9987 0.9790 0.8684 13 1.0000 1.0000 1.0000 1.0000 1.0000 0.9997 0.9935 0.9423 14 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9984 0.9793 15 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9997 0.9941 16 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9987 17 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9998 18 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 19 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000

c 2000 by Chapman & Hall/CRC 

5.5

GEOMETRIC DISTRIBUTION

In a series of Bernoulli trials with probability of success p, a geometric random variable is the number of the trial on which the first success occurs. Hence, this is a waiting time distribution. The geometric distribution, sometimes called the Pascal distribution, is often thought of as the discrete version of an exponential distribution. 5.5.1

Properties pmf

p(x) = pq x−1 x = 1, 2, 3, . . . , 0 ≤ p ≤ 1, q = 1 − p

mean

µ = 1/p

variance

σ 2 = q/p2

skewness

2−p β1 = √ q β2 =

p2 + 6q q

mgf m(t) =

pet 1 − qet

kurtosis

char function

φ(t) =

fact mgf P (t) = 5.5.2

peit 1 − qeit pt 1 − qt

Variates

Let X1 , X2 , . . . , Xn be independent, identically distributed geometric random variables with parameter p. (1) The random variable Y = X1 + X2 + · · · + Xn has a negative binomial distribution with parameters n and p. (2) The random variable Y = min(X1 , X2 , . . . , Xn ) has a geometric distribution with parameter p. 5.5.3

Tables

Example 5.34 : When flipping a biased coin (so that heads occur only 30% of the time), what is the probability that the first head occurs on the 10th flip? Solution: (S1) Using the probability mass table below with x = 10 and p = 0.3 results in 0.0121. (S2) Hence, this is likely to occur only about 1% of the time.

c 2000 by Chapman & Hall/CRC 

Example 5.35 : The probability a randomly selected customer has the correct change when making a purchase at the local donut shop is 0.1. What is the probability the first person to have correct change will be the fifth customer? What is the probability the first person with correct change will be at least the sixth customer? Solution: (S1) Let X be the number of the first customer with correct change. The random variable X has a geometric distribution with parameter p = 0.1. Use the table for cumulative terms of the geometric probabilities to answer each question. (S2) Prob [X = 5] = 0.0656 (S3) Prob [X ≥ 6] = 1 − Prob [X ≤ 5] = 1 − (0.1000 + 0.0900 + 0.0810 + 0.0729 + 0.656) = 1 − 0.4095 = 0.5905

Probability mass, Geometric distribution x

p = 0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 20

0.1000 0.0900 0.0810 0.0729 0.0656 0.0590 0.0531 0.0478 0.0430 0.0387 0.0349 0.0314 0.0282 0.0254 0.0229 0.0135

0.2000 0.1600 0.1280 0.1024 0.0819 0.0655 0.0524 0.0419 0.0336 0.0268 0.0215 0.0172 0.0137 0.0110 0.0088 0.0029

0.3000 0.2100 0.1470 0.1029 0.0720 0.0504 0.0353 0.0247 0.0173 0.0121 0.0085 0.0059 0.0042 0.0029 0.0020 0.0003

0.4000 0.2400 0.1440 0.0864 0.0518 0.0311 0.0187 0.0112 0.0067 0.0040 0.0024 0.0015 0.0009 0.0005 0.0003

0.5000 0.2500 0.1250 0.0625 0.0313 0.0156 0.0078 0.0039 0.0020 0.0010 0.0005 0.0002 0.0001 0.0001

0.6000 0.2400 0.0960 0.0384 0.0154 0.0061 0.0025 0.0010 0.0004 0.0002 0.0001

0.7000 0.2100 0.0630 0.0189 0.0057 0.0017 0.0005 0.0002

0.8000 0.1600 0.0320 0.0064 0.0013 0.0003 0.0001

0.9000 0.0900 0.0090 0.0009 0.0001

Cumulative probability, Geometric distribution x

p = 0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 20

0.1000 0.1900 0.2710 0.3439 0.4095 0.4686 0.5217 0.5695 0.6126 0.6513 0.6862 0.7176 0.7458 0.7712 0.7941 0.8784

0.2000 0.3600 0.4880 0.5904 0.6723 0.7379 0.7903 0.8322 0.8658 0.8926 0.9141 0.9313 0.9450 0.9560 0.9648 0.9885

0.3000 0.5100 0.6570 0.7599 0.8319 0.8824 0.9176 0.9424 0.9596 0.9718 0.9802 0.9862 0.9903 0.9932 0.9953 0.9992

0.4000 0.6400 0.7840 0.8704 0.9222 0.9533 0.9720 0.9832 0.9899 0.9940 0.9964 0.9978 0.9987 0.9992 0.9995 1

0.5000 0.7500 0.8750 0.9375 0.9688 0.9844 0.9922 0.9961 0.9980 0.9990 0.9995 0.9998 0.9999 0.9999 1

0.6000 0.8400 0.9360 0.9744 0.9898 0.9959 0.9984 0.9993 0.9997 0.9999 1 1 1 1

0.7000 0.9100 0.9730 0.9919 0.9976 0.9993 0.9998 0.9999 1 1

0.8000 0.9600 0.9920 0.9984 0.9997 0.9999 1 1

0.9000 0.9900 0.9990 0.9999 1 1

c 2000 by Chapman & Hall/CRC 

5.6

HYPERGEOMETRIC DISTRIBUTION

In a finite population of size N suppose there are M successes (and N − M failures). The hypergeometric distribution is used to describe the number of successes, X, in n trials (n observations drawn from the population). Unlike a binomial distribution, the probability of a success does not remain constant from trial to trial. 5.6.1

Properties M N −M  pmf

p(x) =

x

Nn−x 

x = 0, 1, . . . , n

x≤M

n

n − x ≤ N − M, n, M, N ∈ N , 1 ≤ n ≤ N 1 ≤ M ≤ N, N = 1, 2, . . . M µ=n N     N −n M M 2 σ = n 1− N −1 N N √ (N − 2M )(N − 2n) N − 1  β1 = (N − 2) nM (N − M )(N − n)

mean variance skewness kurtosis

β2 =

N 2 (N − 1) × (N − 2)(N − 3)nM (N − m)(N − n) " M N (N + 1) − 6n(N − n) + 3 2 (N − M )× N # [N 2 (n − 2) − N n2 + 6n(N − n)]

mgf m(t) = 2 F1 (−n, −M ; −N ; 1 − et ) char function

φ(t) = 2 F1 (−n, −M ; −N ; 1 − eit )

fact mgf P (t) = 2 F1 (−n, −M ; −N ; 1 − t) where p Fq is the generalized hypergeometric function defined in Chapter 18 (see page 520). 5.6.2

Variates

Let X be a hypergeometric random variable with parameters n, m, and N . (1) As N → ∞ if n/N < 0.1 then X is approximately a binomial random variable with parameters n and p = M/N . (2) As n, M , and N all tend to infinity, if M/N is small then X has approximately a Poisson distribution with parameter λ = nM/N .

c 2000 by Chapman & Hall/CRC 

5.6.3

Tables

Let X be a hypergeometric random variable with parameters n, M , and N . The probability mass function p(x) = f (x; n, M, N ) is the probability of exactly x successes in a sample of n items. The cumulative distribution function    M N −M x x   r n−r   F (x; n, M, N ) = (5.1) f (r; n, M, N ) = N r=0 r=0 n is the probability of x or fewer successes in the sample of n items. The following table contains values for f (x; n, M, N ) and F (x; n, M, N ) for various values of x, n, M , and N . Example 5.36 : A New York City transportation company has 10 taxis, 3 of which have broken meters. Suppose 4 taxis are selected at random. What is the probability exactly 1 will have a broken meter, fewer than 2 will have a broken meter, all will have working meters? Solution: (S1) Let X be the number of taxis selected with broken meters. The random variable X has a hypergeometric distribution with N = 10, n = 4, and M = 3. M N −M  37 3 · 35  = 1103 = = 0.5 (S2) Prob [X = 1] = 1 Nn−1 210 4 n M N −M  37 1 · 35 = 0104 = = 0.16667 (S3) Prob [X = 0] = 0 N n 210 4 n (S4) Prob [X < 2] = Prob [X ≤ 1] = Prob [X = 0] + Prob [X = 1] = 0.66667

c 2000 by Chapman & Hall/CRC 

Hypergeometric probability and distribution functions N 2 2 3 3 3 3 3 3 4 4

n 1 1 1 1 2 2 2 2 1 1

M 1 1 1 1 1 1 2 2 1 1

x 0 1 0 1 0 1 1 2 0 1

F (x) 0.50000 1.00000 0.66667 1.00000 0.33333 1.00000 0.66667 1.00000 0.75000 1.00000

f (x) 0.50000 0.50000 0.66667 0.33333 0.33333 0.66667 0.66667 0.33333 0.75000 0.25000

N 6 6 6 6 6 6 6 6 6 6

n 2 3 3 3 3 3 3 3 3 3

M 2 1 1 2 2 2 3 3 3 3

x 2 0 1 0 1 2 0 1 2 3

F (x) 1.00000 0.50000 1.00000 0.20000 0.80000 1.00000 0.05000 0.50000 0.95000 1.00000

f (x) 0.06667 0.50000 0.50000 0.20000 0.60000 0.20000 0.05000 0.45000 0.45000 0.05000

4 4 4 4 4 4 4 4 4 4

2 2 2 2 2 3 3 3 3 3

1 1 2 2 2 1 1 2 2 3

0 1 0 1 2 0 1 1 2 2

0.50000 1.00000 0.16667 0.83333 1.00000 0.25000 1.00000 0.50000 1.00000 0.75000

0.50000 0.50000 0.16667 0.66667 0.16667 0.25000 0.75000 0.50000 0.50000 0.75000

6 6 6 6 6 6 6 6 6 6

4 4 4 4 4 4 4 4 4 4

1 1 2 2 2 3 3 3 4 4

0 1 0 1 2 1 2 3 2 3

0.33333 1.00000 0.06667 0.60000 1.00000 0.20000 0.80000 1.00000 0.40000 0.93333

0.33333 0.66667 0.06667 0.53333 0.40000 0.20000 0.60000 0.20000 0.40000 0.53333

4 5 5 5 5 5 5 5 5 5

3 1 1 2 2 2 2 2 3 3

3 1 1 1 1 2 2 2 1 1

3 0 1 0 1 0 1 2 0 1

1.00000 0.80000 1.00000 0.60000 1.00000 0.30000 0.90000 1.00000 0.40000 1.00000

0.25000 0.80000 0.20000 0.60000 0.40000 0.30000 0.60000 0.10000 0.40000 0.60000

6 6 6 6 6 6 6 6 6 6

4 5 5 5 5 5 5 5 5 5

4 1 1 2 2 3 3 4 4 5

4 0 1 1 2 2 3 3 4 4

1.00000 0.16667 1.00000 0.33333 1.00000 0.50000 1.00000 0.66667 1.00000 0.83333

0.06667 0.16667 0.83333 0.33333 0.66667 0.50000 0.50000 0.66667 0.33333 0.83333

5 5 5 5 5 5 5 5 5 5

3 3 3 3 3 3 4 4 4 4

2 2 2 3 3 3 1 1 2 2

0 1 2 1 2 3 0 1 1 2

0.10000 0.70000 1.00000 0.30000 0.90000 1.00000 0.20000 1.00000 0.40000 1.00000

0.10000 0.60000 0.30000 0.30000 0.60000 0.10000 0.20000 0.80000 0.40000 0.60000

6 7 7 7 7 7 7 7 7 7

5 1 1 2 2 2 2 2 3 3

5 1 1 1 1 2 2 2 1 1

5 0 1 0 1 0 1 2 0 1

1.00000 0.85714 1.00000 0.71429 1.00000 0.47619 0.95238 1.00000 0.57143 1.00000

0.16667 0.85714 0.14286 0.71429 0.28571 0.47619 0.47619 0.04762 0.57143 0.42857

5 5 5 5 6 6 6 6 6 6

4 4 4 4 1 1 2 2 2 2

3 3 4 4 1 1 1 1 2 2

2 3 3 4 0 1 0 1 0 1

0.60000 1.00000 0.80000 1.00000 0.83333 1.00000 0.66667 1.00000 0.40000 0.93333

0.60000 0.40000 0.80000 0.20000 0.83333 0.16667 0.66667 0.33333 0.40000 0.53333

7 7 7 7 7 7 7 7 7 7

3 3 3 3 3 3 3 4 4 4

2 2 2 3 3 3 3 1 1 2

0 1 2 0 1 2 3 0 1 0

0.28571 0.85714 1.00000 0.11429 0.62857 0.97143 1.00000 0.42857 1.00000 0.14286

0.28571 0.57143 0.14286 0.11429 0.51429 0.34286 0.02857 0.42857 0.57143 0.14286

c 2000 by Chapman & Hall/CRC 

Hypergeometric probability and distribution functions N 7 7 7 7 7 7 7 7 7 7

n 4 4 4 4 4 4 4 4 4 4

M 2 2 3 3 3 3 4 4 4 4

x 1 2 0 1 2 3 1 2 3 4

F (x) 0.71429 1.00000 0.02857 0.37143 0.88571 1.00000 0.11429 0.62857 0.97143 1.00000

f (x) 0.57143 0.28571 0.02857 0.34286 0.51429 0.11429 0.11429 0.51429 0.34286 0.02857

N 8 8 8 8 8 8 8 8 8 8

n 3 3 4 4 4 4 4 4 4 4

M 3 3 1 1 2 2 2 3 3 3

x 2 3 0 1 0 1 2 0 1 2

F (x) 0.98214 1.00000 0.50000 1.00000 0.21429 0.78571 1.00000 0.07143 0.50000 0.92857

f (x) 0.26786 0.01786 0.50000 0.50000 0.21429 0.57143 0.21429 0.07143 0.42857 0.42857

7 7 7 7 7 7 7 7 7 7

5 5 5 5 5 5 5 5 5 5

1 1 2 2 2 3 3 3 4 4

0 1 0 1 2 1 2 3 2 3

0.28571 1.00000 0.04762 0.52381 1.00000 0.14286 0.71429 1.00000 0.28571 0.85714

0.28571 0.71429 0.04762 0.47619 0.47619 0.14286 0.57143 0.28571 0.28571 0.57143

8 8 8 8 8 8 8 8 8 8

4 4 4 4 4 4 5 5 5 5

3 4 4 4 4 4 1 1 2 2

3 0 1 2 3 4 0 1 0 1

1.00000 0.01429 0.24286 0.75714 0.98571 1.00000 0.37500 1.00000 0.10714 0.64286

0.07143 0.01429 0.22857 0.51429 0.22857 0.01429 0.37500 0.62500 0.10714 0.53571

7 7 7 7 7 7 7 7 7 7

5 5 5 5 6 6 6 6 6 6

4 5 5 5 1 1 2 2 3 3

4 3 4 5 0 1 1 2 2 3

1.00000 0.47619 0.95238 1.00000 0.14286 1.00000 0.28571 1.00000 0.42857 1.00000

0.14286 0.47619 0.47619 0.04762 0.14286 0.85714 0.28571 0.71429 0.42857 0.57143

8 8 8 8 8 8 8 8 8 8

5 5 5 5 5 5 5 5 5 5

2 3 3 3 3 4 4 4 4 5

2 0 1 2 3 1 2 3 4 2

1.00000 0.01786 0.28571 0.82143 1.00000 0.07143 0.50000 0.92857 1.00000 0.17857

0.35714 0.01786 0.26786 0.53571 0.17857 0.07143 0.42857 0.42857 0.07143 0.17857

7 7 7 7 7 7 8 8 8 8

6 6 6 6 6 6 1 1 2 2

4 4 5 5 6 6 1 1 1 1

3 4 4 5 5 6 0 1 0 1

0.57143 1.00000 0.71429 1.00000 0.85714 1.00000 0.87500 1.00000 0.75000 1.00000

0.57143 0.42857 0.71429 0.28571 0.85714 0.14286 0.87500 0.12500 0.75000 0.25000

8 8 8 8 8 8 8 8 8 8

5 5 5 6 6 6 6 6 6 6

5 5 5 1 1 2 2 2 3 3

3 4 5 0 1 0 1 2 1 2

0.71429 0.98214 1.00000 0.25000 1.00000 0.03571 0.46429 1.00000 0.10714 0.64286

0.53571 0.26786 0.01786 0.25000 0.75000 0.03571 0.42857 0.53571 0.10714 0.53571

8 8 8 8 8 8 8 8 8 8

2 2 2 3 3 3 3 3 3 3

2 2 2 1 1 2 2 2 3 3

0 1 2 0 1 0 1 2 0 1

0.53571 0.96429 1.00000 0.62500 1.00000 0.35714 0.89286 1.00000 0.17857 0.71429

0.53571 0.42857 0.03571 0.62500 0.37500 0.35714 0.53571 0.10714 0.17857 0.53571

8 8 8 8 8 8 8 8 8 8

6 6 6 6 6 6 6 6 6 6

3 4 4 4 5 5 5 6 6 6

3 2 3 4 3 4 5 4 5 6

1.00000 0.21429 0.78571 1.00000 0.35714 0.89286 1.00000 0.53571 0.96429 1.00000

0.35714 0.21429 0.57143 0.21429 0.35714 0.53571 0.10714 0.53571 0.42857 0.03571

c 2000 by Chapman & Hall/CRC 

Hypergeometric probability and distribution functions N 8 8 8 8 8 8 8 8 8 8

n 7 7 7 7 7 7 7 7 7 7

M 1 1 2 2 3 3 4 4 5 5

x 0 1 1 2 2 3 3 4 4 5

F (x) 0.12500 1.00000 0.25000 1.00000 0.37500 1.00000 0.50000 1.00000 0.62500 1.00000

f (x) 0.12500 0.87500 0.25000 0.75000 0.37500 0.62500 0.50000 0.50000 0.62500 0.37500

N 9 9 9 9 9 9 9 9 9 9

n 5 5 5 5 5 5 5 5 5 5

M 3 3 3 4 4 4 4 4 5 5

x 1 2 3 0 1 2 3 4 1 2

F (x) 0.40476 0.88095 1.00000 0.00794 0.16667 0.64286 0.96032 1.00000 0.03968 0.35714

f (x) 0.35714 0.47619 0.11905 0.00794 0.15873 0.47619 0.31746 0.03968 0.03968 0.31746

8 8 8 8 9 9 9 9 9 9

7 7 7 7 1 1 2 2 2 2

6 6 7 7 1 1 1 1 2 2

5 6 6 7 0 1 0 1 0 1

0.75000 1.00000 0.87500 1.00000 0.88889 1.00000 0.77778 1.00000 0.58333 0.97222

0.75000 0.25000 0.87500 0.12500 0.88889 0.11111 0.77778 0.22222 0.58333 0.38889

9 9 9 9 9 9 9 9 9 9

5 5 5 6 6 6 6 6 6 6

5 5 5 1 1 2 2 2 3 3

3 4 5 0 1 0 1 2 0 1

0.83333 0.99206 1.00000 0.33333 1.00000 0.08333 0.58333 1.00000 0.01191 0.22619

0.47619 0.15873 0.00794 0.33333 0.66667 0.08333 0.50000 0.41667 0.01191 0.21429

9 9 9 9 9 9 9 9 9 9

2 3 3 3 3 3 3 3 3 3

2 1 1 2 2 2 3 3 3 3

2 0 1 0 1 2 0 1 2 3

1.00000 0.66667 1.00000 0.41667 0.91667 1.00000 0.23810 0.77381 0.98809 1.00000

0.02778 0.66667 0.33333 0.41667 0.50000 0.08333 0.23810 0.53571 0.21429 0.01191

9 9 9 9 9 9 9 9 9 9

6 6 6 6 6 6 6 6 6 6

3 3 4 4 4 4 5 5 5 5

2 3 1 2 3 4 2 3 4 5

0.76191 1.00000 0.04762 0.40476 0.88095 1.00000 0.11905 0.59524 0.95238 1.00000

0.53571 0.23810 0.04762 0.35714 0.47619 0.11905 0.11905 0.47619 0.35714 0.04762

9 9 9 9 9 9 9 9 9 9

4 4 4 4 4 4 4 4 4 4

1 1 2 2 2 3 3 3 3 4

0 1 0 1 2 0 1 2 3 0

0.55556 1.00000 0.27778 0.83333 1.00000 0.11905 0.59524 0.95238 1.00000 0.03968

0.55556 0.44444 0.27778 0.55556 0.16667 0.11905 0.47619 0.35714 0.04762 0.03968

9 9 9 9 9 9 9 9 9 9

6 6 6 6 7 7 7 7 7 7

6 6 6 6 1 1 2 2 2 3

3 4 5 6 0 1 0 1 2 1

0.23810 0.77381 0.98809 1.00000 0.22222 1.00000 0.02778 0.41667 1.00000 0.08333

0.23810 0.53571 0.21429 0.01191 0.22222 0.77778 0.02778 0.38889 0.58333 0.08333

9 9 9 9 9 9 9 9 9 9

4 4 4 4 5 5 5 5 5 5

4 4 4 4 1 1 2 2 2 3

1 2 3 4 0 1 0 1 2 0

0.35714 0.83333 0.99206 1.00000 0.44444 1.00000 0.16667 0.72222 1.00000 0.04762

0.31746 0.47619 0.15873 0.00794 0.44444 0.55556 0.16667 0.55556 0.27778 0.04762

9 9 9 9 9 9 9 9 9 9

7 7 7 7 7 7 7 7 7 7

3 3 4 4 4 5 5 5 6 6

2 3 2 3 4 3 4 5 4 5

0.58333 1.00000 0.16667 0.72222 1.00000 0.27778 0.83333 1.00000 0.41667 0.91667

0.50000 0.41667 0.16667 0.55556 0.27778 0.27778 0.55556 0.16667 0.41667 0.50000

c 2000 by Chapman & Hall/CRC 

Hypergeometric probability and distribution functions N 9 9 9 9 9 9 9 9 9 9

n 7 7 7 7 8 8 8 8 8 8

M 6 7 7 7 1 1 2 2 3 3

x 6 5 6 7 0 1 1 2 2 3

F (x) 1.00000 0.58333 0.97222 1.00000 0.11111 1.00000 0.22222 1.00000 0.33333 1.00000

f (x) 0.08333 0.58333 0.38889 0.02778 0.11111 0.88889 0.22222 0.77778 0.33333 0.66667

N 10 10 10 10 10 10 10 10 10 10

n 5 5 5 5 5 5 5 5 5 5

M 1 1 2 2 2 3 3 3 3 4

x 0 1 0 1 2 0 1 2 3 0

F (x) 0.50000 1.00000 0.22222 0.77778 1.00000 0.08333 0.50000 0.91667 1.00000 0.02381

f (x) 0.50000 0.50000 0.22222 0.55556 0.22222 0.08333 0.41667 0.41667 0.08333 0.02381

9 9 9 9 9 9 9 9 9 9

8 8 8 8 8 8 8 8 8 8

4 4 5 5 6 6 7 7 8 8

3 4 4 5 5 6 6 7 7 8

0.44444 1.00000 0.55556 1.00000 0.66667 1.00000 0.77778 1.00000 0.88889 1.00000

0.44444 0.55556 0.55556 0.44444 0.66667 0.33333 0.77778 0.22222 0.88889 0.11111

10 10 10 10 10 10 10 10 10 10

5 5 5 5 5 5 5 5 5 5

4 4 4 4 5 5 5 5 5 5

1 2 3 4 0 1 2 3 4 5

0.26190 0.73809 0.97619 1.00000 0.00397 0.10318 0.50000 0.89682 0.99603 1.00000

0.23810 0.47619 0.23810 0.02381 0.00397 0.09921 0.39682 0.39682 0.09921 0.00397

10 10 10 10 10 10 10 10 10 10

1 1 2 2 2 2 2 3 3 3

1 1 1 1 2 2 2 1 1 2

0 1 0 1 0 1 2 0 1 0

0.90000 1.00000 0.80000 1.00000 0.62222 0.97778 1.00000 0.70000 1.00000 0.46667

0.90000 0.10000 0.80000 0.20000 0.62222 0.35556 0.02222 0.70000 0.30000 0.46667

10 10 10 10 10 10 10 10 10 10

6 6 6 6 6 6 6 6 6 6

1 1 2 2 2 3 3 3 3 4

0 1 0 1 2 0 1 2 3 0

0.40000 1.00000 0.13333 0.66667 1.00000 0.03333 0.33333 0.83333 1.00000 0.00476

0.40000 0.60000 0.13333 0.53333 0.33333 0.03333 0.30000 0.50000 0.16667 0.00476

10 10 10 10 10 10 10 10 10 10

3 3 3 3 3 3 4 4 4 4

2 2 3 3 3 3 1 1 2 2

1 2 0 1 2 3 0 1 0 1

0.93333 1.00000 0.29167 0.81667 0.99167 1.00000 0.60000 1.00000 0.33333 0.86667

0.46667 0.06667 0.29167 0.52500 0.17500 0.00833 0.60000 0.40000 0.33333 0.53333

10 10 10 10 10 10 10 10 10 10

6 6 6 6 6 6 6 6 6 6

4 4 4 4 5 5 5 5 5 6

1 2 3 4 1 2 3 4 5 2

0.11905 0.54762 0.92857 1.00000 0.02381 0.26190 0.73809 0.97619 1.00000 0.07143

0.11429 0.42857 0.38095 0.07143 0.02381 0.23810 0.47619 0.23810 0.02381 0.07143

10 10 10 10 10 10 10 10 10 10

4 4 4 4 4 4 4 4 4 4

2 3 3 3 3 4 4 4 4 4

2 0 1 2 3 0 1 2 3 4

1.00000 0.16667 0.66667 0.96667 1.00000 0.07143 0.45238 0.88095 0.99524 1.00000

0.13333 0.16667 0.50000 0.30000 0.03333 0.07143 0.38095 0.42857 0.11429 0.00476

10 10 10 10 10 10 10 10 10 10

6 6 6 6 7 7 7 7 7 7

6 6 6 6 1 1 2 2 2 3

3 4 5 6 0 1 0 1 2 0

0.45238 0.88095 0.99524 1.00000 0.30000 1.00000 0.06667 0.53333 1.00000 0.00833

0.38095 0.42857 0.11429 0.00476 0.30000 0.70000 0.06667 0.46667 0.46667 0.00833

c 2000 by Chapman & Hall/CRC 

Hypergeometric probability and distribution functions N 10 10 10 10 10 10 10 10 10 10

n 7 7 7 7 7 7 7 7 7 7

M 3 3 3 4 4 4 4 5 5 5

x 1 2 3 1 2 3 4 2 3 4

F (x) 0.18333 0.70833 1.00000 0.03333 0.33333 0.83333 1.00000 0.08333 0.50000 0.91667

f (x) 0.17500 0.52500 0.29167 0.03333 0.30000 0.50000 0.16667 0.08333 0.41667 0.41667

N 10 10 10 10 10 10 10 10 10 10

n 9 9 9 9 9 9 9 9 9 9

M 5 5 6 6 7 7 8 8 9 9

x 4 5 5 6 6 7 7 8 8 9

F (x) 0.50000 1.00000 0.60000 1.00000 0.70000 1.00000 0.80000 1.00000 0.90000 1.00000

f (x) 0.50000 0.50000 0.60000 0.40000 0.70000 0.30000 0.80000 0.20000 0.90000 0.10000

10 10 10 10 10 10 10 10 10 10

7 7 7 7 7 7 7 7 7 8

5 6 6 6 6 7 7 7 7 1

5 3 4 5 6 4 5 6 7 0

1.00000 0.16667 0.66667 0.96667 1.00000 0.29167 0.81667 0.99167 1.00000 0.20000

0.08333 0.16667 0.50000 0.30000 0.03333 0.29167 0.52500 0.17500 0.00833 0.20000

11 11 11 11 11 11 11 11 11 11

1 1 2 2 2 2 2 3 3 3

1 1 1 1 2 2 2 1 1 2

0 1 0 1 0 1 2 0 1 0

0.90909 1.00000 0.81818 1.00000 0.65455 0.98182 1.00000 0.72727 1.00000 0.50909

0.90909 0.09091 0.81818 0.18182 0.65455 0.32727 0.01818 0.72727 0.27273 0.50909

10 10 10 10 10 10 10 10 10 10

8 8 8 8 8 8 8 8 8 8

1 2 2 2 3 3 3 4 4 4

1 0 1 2 1 2 3 2 3 4

1.00000 0.02222 0.37778 1.00000 0.06667 0.53333 1.00000 0.13333 0.66667 1.00000

0.80000 0.02222 0.35556 0.62222 0.06667 0.46667 0.46667 0.13333 0.53333 0.33333

11 11 11 11 11 11 11 11 11 11

3 3 3 3 3 3 4 4 4 4

2 2 3 3 3 3 1 1 2 2

1 2 0 1 2 3 0 1 0 1

0.94546 1.00000 0.33939 0.84849 0.99394 1.00000 0.63636 1.00000 0.38182 0.89091

0.43636 0.05455 0.33939 0.50909 0.14546 0.00606 0.63636 0.36364 0.38182 0.50909

10 10 10 10 10 10 10 10 10 10

8 8 8 8 8 8 8 8 8 8

5 5 5 6 6 6 7 7 7 8

3 4 5 4 5 6 5 6 7 6

0.22222 0.77778 1.00000 0.33333 0.86667 1.00000 0.46667 0.93333 1.00000 0.62222

0.22222 0.55556 0.22222 0.33333 0.53333 0.13333 0.46667 0.46667 0.06667 0.62222

11 11 11 11 11 11 11 11 11 11

4 4 4 4 4 4 4 4 4 4

2 3 3 3 3 4 4 4 4 4

2 0 1 2 3 0 1 2 3 4

1.00000 0.21212 0.72121 0.97576 1.00000 0.10606 0.53030 0.91212 0.99697 1.00000

0.10909 0.21212 0.50909 0.25455 0.02424 0.10606 0.42424 0.38182 0.08485 0.00303

10 10 10 10 10 10 10 10 10 10

8 8 9 9 9 9 9 9 9 9

8 8 1 1 2 2 3 3 4 4

7 8 0 1 1 2 2 3 3 4

0.97778 1.00000 0.10000 1.00000 0.20000 1.00000 0.30000 1.00000 0.40000 1.00000

0.35556 0.02222 0.10000 0.90000 0.20000 0.80000 0.30000 0.70000 0.40000 0.60000

11 11 11 11 11 11 11 11 11 11

5 5 5 5 5 5 5 5 5 5

1 1 2 2 2 3 3 3 3 4

0 1 0 1 2 0 1 2 3 0

0.54546 1.00000 0.27273 0.81818 1.00000 0.12121 0.57576 0.93939 1.00000 0.04546

0.54546 0.45454 0.27273 0.54546 0.18182 0.12121 0.45454 0.36364 0.06061 0.04546

c 2000 by Chapman & Hall/CRC 

5.7

MULTINOMIAL DISTRIBUTION

The multinomial distribution is a generalization of the binomial distribution. Suppose there are n independent trials, and each trial results in exactly one of k possible distinct outcomes. For i = 1, 2, . . . , k let pi be the probability k that outcome i occurs on any given trial (with i=1 pi = 1). The multinomial random variable is the random vector X = [X1 , X2 , . . . , Xk ]T where Xi is the number of times outcome i occurs. 5.7.1

Properties pmf p(x1 , x2 , . . . , xk ) = n!

k ) pxi i

i=1

mean of Xi variance of Xi

xi !

,

k 

xi = n

i=1

µi = npi σi2 = npi (1 − pi )

Cov[Xi , Xj ] σij = −npi pj ,

i = j

joint mgf m(t1 , t2 , . . . , tk ) = (p1 et1 + p2 et2 + · · · + pk etk )n joint char function φ(t1 , t2 , . . . , tk ) = (p1 eit1 + p2 eit2 + · · · + pk eitk )n joint fact mgf P (t1 , t2 , . . . , tk ) = (p1 t1 + p2 t2 + · · · + pk tk )n 5.7.2

Variates

Let X be a multinomial random variable with parameters n (number of trials) and p1 , p2 , . . . , pk . (1) The marginal distribution of Xi is binomial with parameters n and pi . (2) If k = 2 and p1 = p, then the multinomial random variable corresponds to the binomial random variable with parameters n and p. 5.8

NEGATIVE BINOMIAL DISTRIBUTION

Consider a sequence of Bernoulli trials with probability of success p. The negative binomial distribution is used to describe the number of failures, X, before the nth success. 5.8.1

Properties   x+n−1 n x pmf p(x) = p q x = 0, 1, 2, . . . , n = 1, 2, . . . n−1

mean

0 ≤ p ≤ 1, q = 1 − p nq µ= p

c 2000 by Chapman & Hall/CRC 

nq p2 2−p β1 = √ nq

σ2 =

variance skewness

p2 + 6q nq  n p mgf m(t) = 1 − qet n  p char function φ(t) = 1 − qeit  n p fact mgf P (t) = 1 − qt kurtosis

β2 = 3 +

Using p = k/(m + k) and n = k, there is the following alternative characterization. 5.8.1.1

Alternative characterization  k  x m Γ(k + x) k pmf p(x) = x!Γ(k) m+k m+k x = 0, 1, 2, . . . , m, k > 0 mean

µ=m

variance

σ 2 = m + m2 /k

skewness

β1 = 

kurtosis mgf char function fact mgf where Γ(x) is 5.8.2

2m + k mk(m + k)

6m2 + 6mk + k 2 mk(m + k)  −k m m(t) = 1 − (et − 1) k  −k m it φ(t) = 1 − (e − 1) k  −k m P (t) = 1 − (t − 1) k the gamma function defined in Chapter 18 (see page 515). β2 = 3 +

Variates

Let X be a negative binomial random variable with parameters n and p. (1) If n = 1 then X is a geometric random variable with probability of success p. c 2000 by Chapman & Hall/CRC 

(2) As n → ∞ and p → 1 with n(1 − p) held constant, X is approximately a Poisson random variable with λ = n(1 − p). (3) Let X1 , X2 , . . . , Xk be independent negative binomial random variables with parameters ni and p, respectively. The random variable Y = X1 + X2 + · · · + Xk has a negative binomial distribution with parameters n = n1 + n2 + · · · + nk and p. 5.8.3

Tables

Example 5.37 : Suppose a biased coin has probability of heads 0.3. What is the probability that the 5th head occurs after the 8th tail? Solution: (S1) Recognizing that n = 5 and x = 8 with p = 0.3 and q = 1 − p = 0.7, the probability is

5+8−1 (0.3)5 (0.7)8 = 495(0.3)5 (0.7)8 = 0.0693 Prob [X = 8] = 5−1 (S2) This value is in the table below with n = 5, x = 8, and p = 0.3.

Probability mass, Negative binomial distribution

5.9

(n, x) (1,2) (1,5) (1,8) (1,10)

p = 0.1 0.0810 0.0590 0.0430 0.0349

0.2 0.1280 0.0655 0.0336 0.0215

0.3 0.1470 0.0504 0.0173 0.0085

0.4 0.1440 0.0311 0.0067 0.0024

0.5 0.1250 0.0156 0.0020 0.0005

0.6 0.7 0.8 0.9 0.0960 0.0630 0.0320 0.00900 0.0061 0.0017 0.0003 0.0004 0.0001

(3,2) (3,5) (3,8) (3,10)

0.0049 0.0124 0.0194 0.0230

0.0307 0.0551 0.0604 0.0567

0.0794 0.0953 0.0700 0.0503

0.1382 0.1045 0.0484 0.0255

0.1875 0.0820 0.0220 0.0081

0.2074 0.0464 0.0064 0.0015

0.1852 0.1229 0.04370 0.0175 0.0034 0.00020 0.0010 0.0001 0.0001

(5,2) (5,5) (5,8) (5,10)

0.0001 0.0007 0.0021 0.0035

0.0031 0.0132 0.0266 0.0344

0.0179 0.0515 0.0693 0.0687

0.0553 0.1003 0.0851 0.0620

0.1172 0.1230 0.0604 0.0305

0.1866 0.1003 0.0252 0.0082

0.2269 0.1966 0.08860 0.0515 0.0132 0.00070 0.0055 0.0004 0.0010

(8,2) (8,5) (8,8) (8,10) 0.0001

0.0001 0.0007 0.0028 0.0053

0.0012 0.0087 0.0243 0.0360

0.0085 0.0404 0.0708 0.0771

0.0352 0.0967 0.0982 0.0742

0.0967 0.1362 0.0708 0.0343

0.1868 0.1109 0.0243 0.0066

0.2416 0.15500 0.0425 0.00340 0.0028 0.0003

POISSON DISTRIBUTION

The Poisson, or rare event, distribution is completely described by a single parameter, λ. This distribution is used to model the number of successes, X, in a specified time interval or given region. It is assumed the numbers of successes occurring in different time intervals or regions are independent, the probability of a success in a time interval or region is very small and proportional to the length of the time interval or the size of the region, and the probability of more than one success during any one time interval or region is negligible. c 2000 by Chapman & Hall/CRC 

5.9.1

Properties pmf

e−λ λx x = 0, 1, 2, . . . , λ > 0 x! µ=λ

p(x) =

mean

σ2 = λ

variance

√ β1 = 1/ λ

skewness kurtosis

β2 = 3 + (1/λ)

mgf m(t) = exp[λ(et − 1)] char function

φ(t) = exp[λ(eit − 1)]

fact mgf P (t) = exp[λ(t − 1)] Note that the waiting time between Poisson arrivals is exponentially distributed. 5.9.2

Variates

Let X be a Poisson random variable with parameter λ. (1) As λ → ∞, X is approximately normal with parameters µ = λ and σ 2 = λ. (2) Let X1 , X2 , . . . , Xn be independent Poisson random variables with parameters λi , respectively. The random variable Y = X1 + X2 + · · · + Xn has a Poisson distribution with parameter λ = λ1 + λ2 + · · · + λn . 5.9.3

Tables

Example 5.38 : The number of black bear sightings in Northeastern Pennsylvania during a given week has a Poisson distribution with λ = 3. For a randomly selected week, what is the probability of exactly 2 sightings, more than 5 sightings, between 4 and 7 sightings (inclusive)? Solution: (S1) Let X be the random variable representing the number of black bear sightings during any given week; X is Poisson with λ = 3. Use the table below to answer the probability questions. (S2) Prob [X = 2] = Prob [X ≤ 2] − Prob [X ≤ 1] = 0.423 − 0.199 = 0.224 (S3) Prob [X > 5] = 1 − Prob [X ≤ 4] = 1 − 0.815 = 0.185 (S4) Prob [4 ≤ X ≤ 7] = Prob [X ≤ 7] − Prob [X ≤ 3] = .988 − .647 = .341

c 2000 by Chapman & Hall/CRC 

Cumulative probability, Poisson distribution λ 0.02 0.04 0.06 0.08 0.10 0.15 0.20 0.25 0.30 0.35 0.40 0.45 0.50 0.55 0.60 0.65 0.70 0.75 0.80 0.85 0.90 0.95 1.00 1.1 1.2 1.3 1.4 1.5 1.6 1.7 1.8 1.9 2.0 2.2 2.4 2.6 2.8 3.0 3.2 3.4 3.6 3.8

x =0 0.980 0.961 0.942 0.923 0.905 0.861 0.819 0.779 0.741 0.705 0.670 0.638 0.607 0.577 0.549 0.522 0.497 0.472 0.449 0.427 0.407 0.387 0.368 0.333 0.301 0.273 0.247 0.223 0.202 0.183 0.165 0.150 0.135 0.111 0.091 0.074 0.061 0.050 0.041 0.033 0.027 0.022

1 1.000 0.999 0.998 0.997 0.995 0.990 0.983 0.974 0.963 0.951 0.938 0.925 0.910 0.894 0.878 0.861 0.844 0.827 0.809 0.791 0.772 0.754 0.736 0.699 0.663 0.627 0.592 0.558 0.525 0.493 0.463 0.434 0.406 0.355 0.308 0.267 0.231 0.199 0.171 0.147 0.126 0.107

2

3

4

5

1.000 1.000 1.000 1.000 1.000 0.999 0.998 0.996 0.995 0.992 0.989 0.986 0.982 0.977 0.972 0.966 0.960 0.953 0.945 0.937 0.929 0.920 0.900 0.879 0.857 0.834 0.809 0.783 0.757 0.731 0.704 0.677 0.623 0.570 0.518 0.469 0.423 0.380 0.340 0.303 0.269

1.000 1.000 1.000 1.000 1.000 0.999 0.999 0.998 0.998 0.997 0.996 0.994 0.993 0.991 0.989 0.987 0.984 0.981 0.974 0.966 0.957 0.946 0.934 0.921 0.907 0.891 0.875 0.857 0.819 0.779 0.736 0.692 0.647 0.603 0.558 0.515 0.473

1.000 1.000 1.000 1.000 1.000 0.999 0.999 0.999 0.999 0.998 0.998 0.997 0.996 0.995 0.992 0.989 0.986 0.981 0.976 0.970 0.964 0.956 0.947 0.927 0.904 0.877 0.848 0.815 0.781 0.744 0.706 0.668

1.000 1.000 1.000 1.000 1.000 1.000 1.000 0.999 0.999 0.999 0.998 0.997 0.996 0.994 0.992 0.990 0.987 0.983 0.975 0.964 0.951 0.935 0.916 0.895 0.871 0.844 0.816

c 2000 by Chapman & Hall/CRC 

6

7

1.000 1.000 1.000 1.000 0.999 1.000 0.999 1.000 0.999 1.000 0.998 1.000 0.997 0.999 0.997 0.999 0.996 0.999 0.993 0.998 0.988 0.997 0.983 0.995 0.976 0.992 0.967 0.988 0.955 0.983 0.942 0.977 0.927 0.969 0.909 0.960 continued

8

9

1.000 1.000 1.000 1.000 0.999 1.000 0.999 1.000 0.998 0.999 0.996 0.999 0.994 0.998 0.992 0.997 0.988 0.996 0.984 0.994 on next page

continued from previous page λ x =0 1 2 3 4.0 0.018 0.092 0.238 0.433 4.2 0.015 0.078 0.210 0.395 4.4 0.012 0.066 0.185 0.359 4.6 0.010 0.056 0.163 0.326 4.8 0.008 0.048 0.142 0.294 5.0 0.007 0.040 0.125 0.265 5.2 0.005 0.034 0.109 0.238 5.4 0.004 0.029 0.095 0.213 5.6 0.004 0.024 0.082 0.191 5.8 0.003 0.021 0.071 0.170 6.0 0.003 0.017 0.062 0.151 6.2 0.002 0.015 0.054 0.134 6.4 0.002 0.012 0.046 0.119 6.6 0.001 0.010 0.040 0.105 6.8 0.001 0.009 0.034 0.093 7.0 0.001 0.007 0.030 0.082 7.2 0.001 0.006 0.025 0.072 7.4 0.001 0.005 0.022 0.063 7.6 0.001 0.004 0.019 0.055 7.8 0.000 0.004 0.016 0.049 8.0 0.000 0.003 0.014 0.042 8.5 0.000 0.002 0.009 0.030 9.0 0.000 0.001 0.006 0.021 9.5 0.000 0.001 0.004 0.015 10.0 0.000 0.001 0.003 0.010 10.5 0.000 0.000 0.002 0.007 11.0 0.000 0.000 0.001 0.005 11.5 0.000 0.000 0.001 0.003 12.0 0.000 0.000 0.001 0.002 12.5 0.000 0.000 0.000 0.002 13.0 0.000 0.000 0.000 0.001 13.5 0.000 0.000 0.000 0.001 14.0 0.000 0.000 0.000 0.001 14.5 0.000 0.000 0.000 0.000 15.0 0.000 0.000 0.000 0.000

4 0.629 0.590 0.551 0.513 0.476 0.441 0.406 0.373 0.342 0.313 0.285 0.259 0.235 0.213 0.192 0.173 0.155 0.140 0.125 0.112 0.100 0.074 0.055 0.040 0.029 0.021 0.015 0.011 0.008 0.005 0.004 0.003 0.002 0.001 0.001

5 0.785 0.753 0.720 0.686 0.651 0.616 0.581 0.546 0.512 0.478 0.446 0.414 0.384 0.355 0.327 0.301 0.276 0.253 0.231 0.210 0.191 0.150 0.116 0.088 0.067 0.050 0.037 0.028 0.020 0.015 0.011 0.008 0.005 0.004 0.003

6 0.889 0.868 0.844 0.818 0.791 0.762 0.732 0.702 0.670 0.638 0.606 0.574 0.542 0.511 0.480 0.450 0.420 0.392 0.365 0.338 0.313 0.256 0.207 0.165 0.130 0.102 0.079 0.060 0.046 0.035 0.026 0.019 0.014 0.011 0.008

7 0.949 0.936 0.921 0.905 0.887 0.867 0.845 0.822 0.797 0.771 0.744 0.716 0.687 0.658 0.628 0.599 0.569 0.539 0.510 0.481 0.453 0.386 0.324 0.269 0.220 0.178 0.143 0.114 0.089 0.070 0.054 0.042 0.032 0.024 0.018

8 0.979 0.972 0.964 0.955 0.944 0.932 0.918 0.903 0.886 0.867 0.847 0.826 0.803 0.780 0.755 0.729 0.703 0.676 0.648 0.620 0.593 0.523 0.456 0.392 0.333 0.279 0.232 0.191 0.155 0.125 0.100 0.079 0.062 0.048 0.037

Cumulative probability, Poisson distribution λ x =10 2.8 1.000 3.0 1.000 3.2 1.000 3.4 0.999 3.6 0.999 3.8 0.998 4.0 0.997

11

12

13 14 15 16 17 18 19

1.000 1.000 0.999 1.000 0.999 1.000 continued on next page

c 2000 by Chapman & Hall/CRC 

9 0.992 0.989 0.985 0.981 0.975 0.968 0.960 0.951 0.941 0.929 0.916 0.902 0.886 0.869 0.850 0.831 0.810 0.788 0.765 0.741 0.717 0.653 0.587 0.522 0.458 0.397 0.341 0.289 0.242 0.201 0.166 0.135 0.109 0.088 0.070

continued from previous page λ x =10 11 12 13 4.2 0.996 0.999 1.000 4.4 0.994 0.998 0.999 1.000 4.6 0.992 0.997 0.999 1.000 4.8 0.990 0.996 0.999 1.000 5.0 0.986 0.995 0.998 0.999 5.2 0.982 0.993 0.997 0.999 5.4 0.978 0.990 0.996 0.999 5.6 0.972 0.988 0.995 0.998 5.8 0.965 0.984 0.993 0.997 6.0 0.957 0.980 0.991 0.996 6.2 0.949 0.975 0.989 0.995 6.4 0.939 0.969 0.986 0.994 6.6 0.927 0.963 0.982 0.992 6.8 0.915 0.955 0.978 0.990 7.0 0.901 0.947 0.973 0.987 7.2 0.887 0.937 0.967 0.984 7.4 0.871 0.926 0.961 0.981 7.6 0.854 0.915 0.954 0.976 7.8 0.835 0.902 0.945 0.971 8.0 0.816 0.888 0.936 0.966 8.5 0.763 0.849 0.909 0.949 9.0 0.706 0.803 0.876 0.926 9.5 0.645 0.752 0.836 0.898 10.0 0.583 0.697 0.792 0.865 10.5 0.521 0.639 0.742 0.825 11.0 0.460 0.579 0.689 0.781 11.5 0.402 0.520 0.633 0.733 12.0 0.347 0.462 0.576 0.681 12.5 0.297 0.406 0.519 0.628 13.0 0.252 0.353 0.463 0.573 13.5 0.211 0.304 0.409 0.518 14.0 0.176 0.260 0.358 0.464 14.5 0.145 0.220 0.311 0.412 15.0 0.118 0.185 0.268 0.363

14

15

16

17

18

19

1.000 1.000 1.000 0.999 0.999 0.999 0.998 0.997 0.997 0.996 0.994 0.993 0.991 0.989 0.986 0.983 0.973 0.959 0.940 0.916 0.888 0.854 0.815 0.772 0.725 0.675 0.623 0.570 0.518 0.466

1.000 1.000 1.000 0.999 0.999 0.999 0.998 0.998 0.997 0.996 0.995 0.993 0.992 0.986 0.978 0.967 0.951 0.932 0.907 0.878 0.844 0.806 0.764 0.718 0.669 0.619 0.568

1.000 1.000 1.000 1.000 0.999 0.999 0.999 0.998 0.998 0.997 0.996 0.993 0.989 0.982 0.973 0.960 0.944 0.924 0.899 0.869 0.836 0.797 0.756 0.711 0.664

1.000 1.000 1.000 1.000 0.999 0.999 0.999 0.998 0.997 0.995 0.991 0.986 0.978 0.968 0.954 0.937 0.916 0.890 0.861 0.827 0.790 0.749

1.000 1.000 1.000 0.999 0.999 0.998 0.996 0.993 0.989 0.982 0.974 0.963 0.948 0.930 0.908 0.883 0.853 0.820

1.000 1.000 0.999 0.998 0.997 0.994 0.991 0.986 0.979 0.969 0.957 0.942 0.923 0.901 0.875

Cumulative probability, Poisson distribution λ x =20 8.5 1.000 9.0 1.000 9.5 0.999 10.0 0.998 10.5 0.997 11.0 0.995 11.5 0.993 12.0 0.988

21

22

1.000 0.999 0.999 0.998 0.996 0.994

1.000 0.999 0.999 0.998 0.997

c 2000 by Chapman & Hall/CRC 

23

24

25

26 27 28 29

1.000 1.000 0.999 1.000 0.999 0.999 1.000 continued on next page

continued from previous page λ x =20 21 22 23 12.5 0.983 0.991 0.995 0.998 13.0 0.975 0.986 0.992 0.996 13.5 0.965 0.980 0.989 0.994 14.0 0.952 0.971 0.983 0.991 14.5 0.936 0.960 0.976 0.986 15.0 0.917 0.947 0.967 0.981

24 0.999 0.998 0.997 0.995 0.992 0.989

25 0.999 0.999 0.998 0.997 0.996 0.994

26 1.000 1.000 0.999 0.999 0.998 0.997

27

28

29

1.000 0.999 1.000 0.999 1.000 1.000 0.998 0.999 1.000

Cumulative probability, Poisson distribution λ 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30

x =5 0.001 0.001 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000

6 0.004 0.002 0.001 0.001 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000

7 0.010 0.005 0.003 0.002 0.001 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000

8 0.022 0.013 0.007 0.004 0.002 0.001 0.001 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000

9 0.043 0.026 0.015 0.009 0.005 0.003 0.002 0.001 0.000 0.000 0.000 0.000 0.000 0.000 0.000

10 0.077 0.049 0.030 0.018 0.011 0.006 0.004 0.002 0.001 0.001 0.000 0.000 0.000 0.000 0.000

11 0.127 0.085 0.055 0.035 0.021 0.013 0.008 0.004 0.003 0.001 0.001 0.000 0.000 0.000 0.000

12 0.193 0.135 0.092 0.061 0.039 0.025 0.015 0.009 0.005 0.003 0.002 0.001 0.001 0.000 0.000

13 0.275 0.201 0.143 0.098 0.066 0.043 0.028 0.017 0.011 0.006 0.004 0.002 0.001 0.001 0.000

14 0.367 0.281 0.208 0.150 0.105 0.072 0.048 0.031 0.020 0.012 0.008 0.005 0.003 0.002 0.001

Cumulative probability, Poisson distribution λ x =15 16 0.467 17 0.371 18 0.287 19 0.215 20 0.157 21 0.111 22 0.077 23 0.052 24 0.034 25 0.022 26 0.014 27 0.009 28 0.005 29 0.003 30 0.002

16 0.566 0.468 0.375 0.292 0.221 0.163 0.117 0.082 0.056 0.038 0.025 0.016 0.010 0.006 0.004

17 0.659 0.564 0.469 0.378 0.297 0.227 0.169 0.123 0.087 0.060 0.041 0.027 0.018 0.011 0.007

c 2000 by Chapman & Hall/CRC 

18 0.742 0.655 0.562 0.469 0.381 0.302 0.233 0.175 0.128 0.092 0.065 0.044 0.030 0.020 0.013

19 0.812 0.736 0.651 0.561 0.470 0.384 0.306 0.238 0.180 0.134 0.097 0.069 0.048 0.033 0.022

20 0.868 0.805 0.731 0.647 0.559 0.471 0.387 0.310 0.243 0.185 0.139 0.102 0.073 0.051 0.035

21 0.911 0.862 0.799 0.726 0.644 0.558 0.472 0.389 0.314 0.247 0.191 0.144 0.106 0.077 0.054

22 0.942 0.905 0.855 0.793 0.721 0.640 0.556 0.472 0.392 0.318 0.252 0.195 0.148 0.110 0.081

23 0.963 0.937 0.899 0.849 0.787 0.716 0.637 0.555 0.473 0.394 0.321 0.256 0.200 0.153 0.115

24 0.978 0.959 0.932 0.893 0.843 0.782 0.712 0.635 0.554 0.473 0.396 0.324 0.260 0.204 0.157

Cumulative probability, Poisson distribution λ x =25 16 0.987 17 0.975 18 0.955 19 0.927 20 0.888 21 0.838 22 0.777 23 0.708 24 0.632 25 0.553 26 0.474 27 0.398 28 0.327 29 0.264 30 0.208

26 0.993 0.985 0.972 0.951 0.922 0.883 0.832 0.772 0.704 0.629 0.552 0.474 0.400 0.330 0.267

27 0.996 0.991 0.983 0.969 0.948 0.917 0.877 0.827 0.768 0.700 0.627 0.551 0.475 0.401 0.333

28 0.998 0.995 0.990 0.981 0.966 0.944 0.913 0.873 0.823 0.763 0.697 0.625 0.550 0.475 0.403

29 0.999 0.997 0.994 0.988 0.978 0.963 0.940 0.908 0.868 0.818 0.759 0.694 0.623 0.549 0.476

30 0.999 0.999 0.997 0.993 0.987 0.976 0.960 0.936 0.904 0.863 0.813 0.755 0.690 0.621 0.548

31 1.000 0.999 0.998 0.996 0.992 0.985 0.974 0.956 0.932 0.900 0.859 0.809 0.751 0.687 0.619

32

33

34

1.000 0.999 0.998 0.995 0.991 0.983 0.971 0.953 0.928 0.896 0.855 0.805 0.748 0.684

1.000 0.999 0.997 0.995 0.990 0.981 0.969 0.950 0.925 0.892 0.851 0.801 0.744

0.999 0.999 0.997 0.994 0.988 0.979 0.966 0.947 0.921 0.888 0.847 0.797

Cumulative probability, Poisson distribution λ x =35 16 1.000 17 1.000 18 1.000 19 1.000 20 0.999 21 0.998 22 0.996 23 0.993 24 0.987 25 0.978 26 0.964 27 0.944 28 0.918 29 0.884 30 0.843

36

37

38

39

40

41

42

43

44

1.000 0.999 0.998 0.996 0.992 0.985 0.976 0.961 0.941 0.914 0.880

1.000 0.999 0.997 0.995 0.991 0.984 0.974 0.959 0.938 0.911

1.000 0.999 0.999 0.997 0.994 0.990 0.983 0.972 0.956 0.935

1.000 0.999 0.998 0.997 0.994 0.989 0.981 0.970 0.954

1.000 0.999 0.998 0.996 0.993 0.988 0.980 0.968

1.000 0.999 0.998 0.996 0.992 0.986 0.978

1.000 0.999 0.999 0.997 0.995 0.991 0.985

1.000 0.999 0.998 0.997 0.994 0.990

1.000 0.999 0.998 0.997 0.994

c 2000 by Chapman & Hall/CRC 

5.10

RECTANGULAR (DISCRETE UNIFORM) DISTRIBUTION

A general rectangular distribution is used to describe a random variable, X, that can assume n different values with equal probabilities. In the special case presented here, we assume the random variable can assume the first n positive integers. 5.10.1

Properties pmf mean

p(x) = 1/n, x = 1, 2, . . . , n, n ∈ N µ = (n + 1)/2

variance

σ 2 = (n2 − 1)/12

skewness

β1 = 0

kurtosis

3 β2 = 5

mgf m(t) = char function

φ(t) =

fact mgf P (t) =



4 3− 2 n −1



et (1 − ent ) n(1 − et ) eit (1 − enit ) n(1 − eit ) t(1 − tn ) n(1 − t)

Example 5.39 : A new family game has a special 12-sided numbered die, manufactured so that each side is equally likely to occur. Find the mean and variance of the number rolled, and the probability of rolling a 2, 3, or 12. Solution: (S1) Let X be the number on the side facing up; X has a discrete uniform distribution with n = 12. (S2) Using the properties given above: 13/2

= 6.5

σ = (n − 1)/12 = (12 − 1)/12 =

143/12

µ = (n + 1)/2 = (12 + 1)/2 = 2

(S3) Prob [X = 2, 3, 12] =

2

2

1 1 3 1 + + = = 0.25 12 12 12 12

c 2000 by Chapman & Hall/CRC 

= 11.9167

CHAPTER 6

Continuous Probability Distributions Contents 6.1

6.2

6.3

6.4

6.5

6.6

6.7

Arcsin distribution 6.1.1 Properties 6.1.2 Probability density function Beta distribution 6.2.1 Properties 6.2.2 Probability density function 6.2.3 Related distributions Cauchy distribution 6.3.1 Properties 6.3.2 Probability density function 6.3.3 Related distributions Chi–square distribution 6.4.1 Properties 6.4.2 Probability density function 6.4.3 Related distributions 6.4.4 Critical values for chi–square distribution 6.4.5 Percentage points, chi–square over dof Erlang distribution 6.5.1 Properties 6.5.2 Probability density function 6.5.3 Related distributions Exponential distribution 6.6.1 Properties 6.6.2 Probability density function 6.6.3 Related distributions Extreme–value distribution 6.7.1 Properties 6.7.2 Probability density function 6.7.3 Related distributions

c 2000 by Chapman & Hall/CRC 

6.8

F distribution 6.8.1 Properties 6.8.2 Probability density function 6.8.3 Related distributions 6.8.4 Critical values for the F distribution 6.9 Gamma distribution 6.9.1 Properties 6.9.2 Probability density function 6.9.3 Related distributions 6.10 Half–normal distribution 6.10.1 Properties 6.10.2 Probability density function 6.11 Inverse Gaussian (Wald) distribution 6.11.1 Properties 6.11.2 Probability density function 6.11.3 Related distributions 6.12 Laplace distribution 6.12.1 Properties 6.12.2 Probability density function 6.12.3 Related distributions 6.13 Logistic distribution 6.13.1 Properties 6.13.2 Probability density function 6.13.3 Related distributions 6.14 Lognormal distribution 6.14.1 Properties 6.14.2 Probability density function 6.14.3 Related distributions 6.15 Noncentral chi–square distribution 6.15.1 Properties 6.15.2 Probability density function 6.15.3 Related distributions 6.16 Noncentral F distribution 6.16.1 Properties 6.16.2 Probability density function 6.16.3 Related distributions 6.17 Noncentral t distribution 6.17.1 Properties 6.17.2 Probability density function 6.17.3 Related distributions 6.18 Normal distribution 6.18.1 Properties c 2000 by Chapman & Hall/CRC 

6.18.2 Probability density function 6.18.3 Related distributions 6.19 Normal distribution: multivariate 6.19.1 Properties 6.19.2 Probability density function 6.20 Pareto distribution 6.20.1 Properties 6.20.2 Probability density function 6.20.3 Related distributions 6.21 Power function distribution 6.21.1 Properties 6.21.2 Probability density function 6.21.3 Related distributions 6.22 Rayleigh distribution 6.22.1 Properties 6.22.2 Probability density function 6.22.3 Related distributions 6.23 t distribution 6.23.1 Properties 6.23.2 Probability density function 6.23.3 Related distributions 6.23.4 Critical values for the t distribution 6.24 Triangular distribution 6.24.1 Properties 6.24.2 Probability density function 6.25 Uniform distribution 6.25.1 Properties 6.25.2 Probability density function 6.25.3 Related distributions 6.26 Weibull distribution 6.26.1 Properties 6.26.2 Probability density function 6.26.3 Related distributions 6.27 Relationships among distributions 6.27.1 Other relationships among distributions

This chapter presents some common continuous probability distributions along with their properties. Relevant numerical tables are also included. Notation used throughout this chapter:

c 2000 by Chapman & Hall/CRC 

 Prob [a ≤ X ≤ b] =

Probability density function (pdf)

f (x)

Cumulative distrib function (cdf)

F (x) = Prob [X ≤ x] =

f (x) dx a



Mean

x

f (x) dx −∞

µ = E [X]   σ 2 = E (X − µ)2   β1 = E (X − µ)3 /σ 3   β2 = E (X − µ)4 /σ 4   m(t) = E etX

Variance Coefficient of skewness Coefficient of kurtosis Moment generating function (mgf)

  φ(t) = E eitX

Characteristic function (char function) 6.1

b

ARCSIN DISTRIBUTION

6.1.1

Properties pdf f (x) =

π



mean

µ = 1/2

variance

σ 2 = 1/8

skewness

β1 = 0

kurtosis

1 x(1 − x)

,

0
β2 = 3/2

mgf m(t) = et/2 I0 (t/2) char function

φ(t) = J0 (t/2) cos(t/2) + iJ0 (t/2) sin(t/2)

where Jn (x) is the Bessel function of the first kind and Ip (x) is the modified Bessel function of the first kind defined in Chapter 18 (see page 506). 6.1.2

Probability density function

The probability density function is “U” shaped. As x → 0+ and as x → 1− , f (x) → ∞.

c 2000 by Chapman & Hall/CRC 

Figure 6.1: Probability density function for an arcsin random variable. 6.2

BETA DISTRIBUTION

6.2.1

Properties pdf f (x) =

mean variance skewness kurtosis

Γ(α + β) α−1 xα−1 (1 − x)β−1 (1 − x)β−1 = x Γ(α)Γ(β) B(α, β)

0 ≤ x ≤ 1, α, β > 0 α µ= α+β αβ σ2 = (α + β)2 (α + β + 1) √ 2(β − α) α + β + 1 β1 = √ αβ(α + β + 2) β2 =

3(α + β + 1)[2(α + β)2 + αβ(α + β − 6)] αβ(α + β + 2)(α + β + 3)

mgf m(t) = 1 F1 (α; β; t) char function φ(t) = '  , ' ,  1 1 α α 3 1 α β α β t2 iat 2 F3 + ,1 + ; , + + ,1 + + ;− α+β 2 2 2 2 2 2 2 2 2 4 ' , ' ,  2 1 α α 1 α β 1 α β t + 2 F3 + , ; , + , + + ;− 2 2 2 2 2 2 2 2 2 4 where Γ(x) is the gamma function, B(a, b) is the beta function, and p Fq is the generalized hypergeometric function defined in Chapter 18 (see pages 515, c 2000 by Chapman & Hall/CRC 

511, and 520). The rth moment about the origin is µr = 6.2.2

Γ(α + β)Γ(α + r) Γ(α)Γ(α + β + r)

(6.1)

Probability density function

If α < 1 and β < 1 the probability density function is “U” shaped. If the product (α − 1)(β − 1) < 0 the probability density function is “J” shaped. Let f (x; α, β) denote the probability density function for a beta random variable with parameters α and β. If both α > 1 and β > 1 then f (x; α, β) and f (x; β, α) are symmetric with respect to the line x = .5.

Figure 6.2: Probability density functions for a beta random variable, various shape parameters. 6.2.3

Related distributions

Let X be a beta random variable with parameters α and β. (1) If α = β = 1/2, then X is an arcsin random variable. (2) If α = β = 1, then X is a uniform random variable with parameters a = 0 and b = 1. (3) If β = 1, then X is a power function random variable with parameters b = 1 and c = α. (4) As α and β tend to infinity such that α/β is constant, X tends to a standard normal random variable.

c 2000 by Chapman & Hall/CRC 

Figure 6.3: Probability density functions for a beta random variable, example of symmetry. 6.3

CAUCHY DISTRIBUTION

6.3.1

Properties pdf f (x) =

 bπ 1 +

1  x−a 2  ,

x ∈ R, a ∈ R, b > 0

b

mean

µ = does not exist

variance

σ 2 = does not exist

skewness

β1 = does not exist

kurtosis

β2 = does not exist

mgf m(t) = does not exist char function 6.3.2

φ(t) = eait−b|t|

Probability density function

The probability density function for a Cauchy random variable is unimodal and symmetric about the parameter a. The tails are heavier than those of a normal random variable. 6.3.3

Related distributions

Let X be a Cauchy random variable with parameters a and b. (1) If a = 0 and b = 1 then X is a standard Cauchy random variable.

c 2000 by Chapman & Hall/CRC 

Figure 6.4: Probability density functions for a Cauchy random variable. (2) The random variable 1/X is also a Cauchy random variable with parameters a/(a2 + b2 ) and b/(a2 + b2 ). (3) Let Xi (for i = 1, 2, . . . , n) be independent, Cauchy random variables with parameters ai and bi , respectively. The random variable Y = X1 + X2 + · · · + Xn has a Cauchy distribution with parameters a = a1 + a2 + · · · + an and b = b1 + b2 + · · · + bn . 6.4

CHI–SQUARE DISTRIBUTION

6.4.1

Properties pdf f (x) = mean

variance skewness

e−x/2 x(ν/2)−1 , 2ν/2 Γ(ν/2)

µ=ν σ 2 = 2ν  β1 = 2 2/ν

12 ν mgf m(t) = (1 − 2t)−ν/2 ,

kurtosis

char function

x ≥ 0, ν ∈ N

β2 = 3 +

t < 1/2

−ν/2

φ(t) = (1 − 2it)

where Γ(x) is the gamma function (see page 515). A chi–square(χ2 ) distribution is completely characterized by the parameter ν, the degrees of freedom.

c 2000 by Chapman & Hall/CRC 

6.4.2

Probability density function

The probability density function for a chi–square random variable is positively skewed. As ν tends to infinity, the density function becomes more bell–shaped and symmetric.

Figure 6.5: Probability density functions for a chi–square random variable. 6.4.3

Related distributions

(1) If X is a chi–square random variable with ν = 2, then X is an exponential random variable with λ = 1/2. (2) If X1 and X2 are independent chi–square random variables with parameters ν1 and ν2 , then the random variable (X1 /ν1 )/(X2 /ν2 ) has an F distribution with ν1 and ν2 degrees of freedom. (3) If X1 and X2 are independent chi–square random variables with parameters ν1 = ν2 = ν, the random variable √ ν X1 − X2 √ Y = (6.2) 2 X1 X2 has a t distribution with ν degrees of freedom. (4) Let Xi (for i = 1, 2, . . . , n) be independent chi–square random variables with parameters νi . The random variable Y = X1 + X2 + · · · + Xn has a chi–square distribution with ν = ν1 + ν2 + · · · + νn degrees of freedom. (5) If X is a chi–square √ random variable with ν degrees of freedom, the random variable X has a chi distribution with parameter ν. Properties of a chi random variable:

c 2000 by Chapman & Hall/CRC 

xn−1 e−x /2 , x ≥ 0, n ∈ N (n/2)−1 2 Γ(n/2)   Γ n+1 2   µ= Γ n2  %  n+1  &2  Γ 2 Γ n+2 2 2 n −   σ = Γ 2 Γ n2 2

pdf f (x) = mean

variance

where Γ(x) is the gamma function (see page 515). If X is a chi random variable with parameter n = 2, then X is a Rayleigh random variable with σ = 1. 6.4.4

Critical values for chi–square distribution

The following tables give values of χ2α,ν such that  χ2α,ν   1 1 − α = F χ2α,ν = x(ν−2)/2 e−x/2 dx ν/2 Γ(ν/2) 2 0

(6.3)

where ν, the number of degrees of freedom, varies from 1 to 10,000 and α varies from 0.0001 to 0.9999.  √ (a) For ν > 30, the expression 2χ2 − 2ν − 1 is approximately a standard normal distribution. Hence, χ2α,ν is approximately √ 2 1 zα + 2ν − 1 for ν  1 2   (b) For even values of ν, F χ2α,ν can be written as χ2α,ν ≈

1−F



χ2α,ν



=

 x −1

x=0

e−λ λx x!

(6.4)

(6.5)

with λ = χ2α,ν /2 and x = ν/2. Hence, the cumulative chi–square distribution is related to the cumulative Poisson distribution. Example 6.40 : Use the table on page 121 to find the values χ2.99,36 and χ2.05,20 . Solution: (S1) The left–hand column of the table on page 121 contains entries for the number of degrees of freedom and the top row lists values for α. The intersection of the ν degrees of freedom row and the α column contains χ2α,ν such that Prob χ2 ≥ χ2α,ν = α.   (S2) χ2.99,36 = 19.2327 =⇒ Prob χ2 ≥ 19.2327 = .99   χ2.05,20 = 31.4104 =⇒ Prob χ2 ≥ 31.4104 = .05

c 2000 by Chapman & Hall/CRC 

Critical values for the chi–square distribution χ2α,ν . α ν 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40

.9999 7

.0 157 .0002 .0052 .0284 .0822 .1724 .3000 .4636 .6608 .8889 1.1453 1.4275 1.7333 2.0608 2.4082 2.7739 3.1567 3.5552 3.9683 4.3952 4.8348 5.2865 5.7494 6.2230 6.7066 7.1998 7.7019 8.2126 8.7315 9.2581 9.7921 10.3331 10.8810 11.4352 11.9957 12.5622 13.1343 13.7120 14.2950 14.8831

.9995 6

.0 393 .0010 .0153 .0639 .1581 .2994 .4849 .7104 .9717 1.2650 1.5868 1.9344 2.3051 2.6967 3.1075 3.5358 3.9802 4.4394 4.9123 5.3981 5.8957 6.4045 6.9237 7.4527 7.9910 8.5379 9.0932 9.6563 10.2268 10.8044 11.3887 11.9794 12.5763 13.1791 13.7875 14.4012 15.0202 15.6441 16.2729 16.9062

.999

.995

5

4

.0 157 .0020 .0243 .0908 .2102 .3811 .5985 .8571 1.1519 1.4787 1.8339 2.2142 2.6172 3.0407 3.4827 3.9416 4.4161 4.9048 5.4068 5.9210 6.4467 6.9830 7.5292 8.0849 8.6493 9.2221 9.8028 10.3909 10.9861 11.5880 12.1963 12.8107 13.4309 14.0567 14.6878 15.3241 15.9653 16.6112 17.2616 17.9164

c 2000 by Chapman & Hall/CRC 

.0 393 .0100 .0717 .2070 .4117 .6757 .9893 1.3444 1.7349 2.1559 2.6032 3.0738 3.5650 4.0747 4.6009 5.1422 5.6972 6.2648 6.8440 7.4338 8.0337 8.6427 9.2604 9.8862 10.5197 11.1602 11.8076 12.4613 13.1211 13.7867 14.4578 15.1340 15.8153 16.5013 17.1918 17.8867 18.5858 19.2889 19.9959 20.7065

.99

.975

.95

.90

.0002 .0201 .1148 .2971 .5543 .8721 1.2390 1.6465 2.0879 2.5582 3.0535 3.5706 4.1069 4.6604 5.2293 5.8122 6.4078 7.0149 7.6327 8.2604 8.8972 9.5425 10.1957 10.8564 11.5240 12.1981 12.8785 13.5647 14.2565 14.9535 15.6555 16.3622 17.0735 17.7891 18.5089 19.2327 19.9602 20.6914 21.4262 22.1643

.0010 .0506 .2158 .4844 .8312 1.2373 1.6899 2.1797 2.7004 3.2470 3.8157 4.4038 5.0088 5.6287 6.2621 6.9077 7.5642 8.2307 8.9065 9.5908 10.2829 10.9823 11.6886 12.4012 13.1197 13.8439 14.5734 15.3079 16.0471 16.7908 17.5387 18.2908 19.0467 19.8063 20.5694 21.3359 22.1056 22.8785 23.6543 24.4330

.0039 .1026 .3518 .7107 1.1455 1.6354 2.1673 2.7326 3.3251 3.9403 4.5748 5.2260 5.8919 6.5706 7.2609 7.9616 8.6718 9.3905 10.1170 10.8508 11.5913 12.3380 13.0905 13.8484 14.6114 15.3792 16.1514 16.9279 17.7084 18.4927 19.2806 20.0719 20.8665 21.6643 22.4650 23.2686 24.0749 24.8839 25.6954 26.5093

.0158 .2107 .5844 1.0636 1.6103 2.2041 2.8331 3.4895 4.1682 4.8652 5.5778 6.3038 7.0415 7.7895 8.5468 9.3122 10.0852 10.8649 11.6509 12.4426 13.2396 14.0415 14.8480 15.6587 16.4734 17.2919 18.1139 18.9392 19.7677 20.5992 21.4336 22.2706 23.1102 23.9523 24.7967 25.6433 26.4921 27.3430 28.1958 29.0505

Critical values for the chi–square distribution χ2α,ν . α ν

.9999

.9995

.999

.995

.99

.975

.95

.90

41 42 43 44 45 46 47 48 49 50 60 70 80 90 100 200 300 400 500 600 700 800 900 1000 1500 2000 2500 3000 3500 4000 4500 5000 5500 6000 6500 7000 7500 8000 8500 9000 9500 10000

15.48 16.07 16.68 17.28 17.89 18.51 19.13 19.75 20.38 21.01 27.50 34.26 41.24 48.41 55.72 134.02 217.33 303.26 390.85 479.64 569.32 659.72 750.70 842.17 1304.80 1773.30 2245.54 2720.44 3197.36 3675.88 4155.71 4636.62 5118.47 5601.13 6084.50 6568.49 7053.05 7538.11 8023.63 8509.57 8995.90 9482.59

17.54 18.19 18.83 19.48 20.14 20.79 21.46 22.12 22.79 23.46 30.34 37.47 44.79 52.28 59.90 140.66 225.89 313.43 402.45 492.52 583.39 674.89 766.91 859.36 1326.30 1798.42 2273.86 2751.65 3231.23 3712.22 4194.37 4677.48 5161.42 5646.08 6131.36 6617.20 7103.53 7590.32 8077.51 8565.07 9052.97 9541.19

18.58 19.24 19.91 20.58 21.25 21.93 22.61 23.29 23.98 24.67 31.74 39.04 46.52 54.16 61.92 143.84 229.96 318.26 407.95 498.62 590.05 682.07 774.57 867.48 1336.42 1810.24 2287.17 2766.32 3247.14 3729.29 4212.52 4696.67 5181.58 5667.17 6153.35 6640.05 7127.22 7614.81 8102.78 8591.09 9079.73 9568.67

21.42 22.14 22.86 23.58 24.31 25.04 25.77 26.51 27.25 27.99 35.53 43.28 51.17 59.20 67.33 152.24 240.66 330.90 422.30 514.53 607.38 700.73 794.47 888.56 1362.67 1840.85 2321.62 2804.23 3288.25 3773.37 4259.39 4746.17 5233.60 5721.59 6210.07 6698.98 7188.28 7677.94 8167.91 8658.17 9148.70 9639.48

22.91 23.65 24.40 25.15 25.90 26.66 27.42 28.18 28.94 29.71 37.48 45.44 53.54 61.75 70.06 156.43 245.97 337.16 429.39 522.37 615.91 709.90 804.25 898.91 1375.53 1855.82 2338.45 2822.75 3308.31 3794.87 4282.25 4770.31 5258.96 5748.11 6237.70 6727.69 7218.03 7708.68 8199.63 8690.83 9182.28 9673.95

25.21 26.00 26.79 27.57 28.37 29.16 29.96 30.75 31.55 32.36 40.48 48.76 57.15 65.65 74.22 162.73 253.91 346.48 439.94 534.02 628.58 723.51 818.76 914.26 1394.56 1877.95 2363.31 2850.08 3337.92 3826.60 4315.96 4805.90 5296.34 5787.20 6278.43 6769.99 7261.85 7753.98 8246.35 8738.94 9231.74 9724.72

27.33 28.14 28.96 29.79 30.61 31.44 32.27 33.10 33.93 34.76 43.19 51.74 60.39 69.13 77.93 168.28 260.88 354.64 449.15 544.18 639.61 735.36 831.37 927.59 1411.06 1897.12 2384.84 2873.74 3363.53 3854.03 4345.10 4836.66 5328.63 5820.96 6313.60 6806.52 7299.69 7793.08 8286.68 8780.46 9274.42 9768.53

29.91 30.77 31.63 32.49 33.35 34.22 35.08 35.95 36.82 37.69 46.46 55.33 64.28 73.29 82.36 174.84 269.07 364.21 459.93 556.06 652.50 749.19 846.07 943.13 1430.25 1919.39 2409.82 2901.17 3393.22 3885.81 4378.86 4872.28 5366.03 5860.05 6354.32 6848.80 7343.48 7838.33 8333.34 8828.50 9323.78 9819.19

c 2000 by Chapman & Hall/CRC 

Critical values for the chi–square distribution χ2α,ν . α ν

.10

.05

.025

.01

.005

.001

.0005

.0001

1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40

2.7055 4.6052 6.2514 7.7794 9.2364 10.6446 12.0170 13.3616 14.6837 15.9872 17.2750 18.5493 19.8119 21.0641 22.3071 23.5418 24.7690 25.9894 27.2036 28.4120 29.6151 30.8133 32.0069 33.1962 34.3816 35.5632 36.7412 37.9159 39.0875 40.2560 41.4217 42.5847 43.7452 44.9032 46.0588 47.2122 48.3634 49.5126 50.6598 51.8051

3.8415 5.9915 7.8147 9.4877 11.0705 12.5916 14.0671 15.5073 16.9190 18.3070 19.6751 21.0261 22.3620 23.6848 24.9958 26.2962 27.5871 28.8693 30.1435 31.4104 32.6706 33.9244 35.1725 36.4150 37.6525 38.8851 40.1133 41.3371 42.5570 43.7730 44.9853 46.1943 47.3999 48.6024 49.8018 50.9985 52.1923 53.3835 54.5722 55.7585

5.0239 7.3778 9.3484 11.1433 12.8325 14.4494 16.0128 17.5345 19.0228 20.4832 21.9200 23.3367 24.7356 26.1189 27.4884 28.8454 30.1910 31.5264 32.8523 34.1696 35.4789 36.7807 38.0756 39.3641 40.6465 41.9232 43.1945 44.4608 45.7223 46.9792 48.2319 49.4804 50.7251 51.9660 53.2033 54.4373 55.6680 56.8955 58.1201 59.3417

6.6349 9.2103 11.3449 13.2767 15.0863 16.8119 18.4753 20.0902 21.6660 23.2093 24.7250 26.2170 27.6882 29.1412 30.5779 31.9999 33.4087 34.8053 36.1909 37.5662 38.9322 40.2894 41.6384 42.9798 44.3141 45.6417 46.9629 48.2782 49.5879 50.8922 52.1914 53.4858 54.7755 56.0609 57.3421 58.6192 59.8925 61.1621 62.4281 63.6907

7.8794 10.5966 12.8382 14.8603 16.7496 18.5476 20.2777 21.9550 23.5894 25.1882 26.7568 28.2995 29.8195 31.3193 32.8013 34.2672 35.7185 37.1565 38.5823 39.9968 41.4011 42.7957 44.1813 45.5585 46.9279 48.2899 49.6449 50.9934 52.3356 53.6720 55.0027 56.3281 57.6484 58.9639 60.2748 61.5812 62.8833 64.1814 65.4756 66.7660

10.8276 13.8155 16.2662 18.4668 20.5150 22.4577 24.3219 26.1245 27.8772 29.5883 31.2641 32.9095 34.5282 36.1233 37.6973 39.2524 40.7902 42.3124 43.8202 45.3147 46.7970 48.2679 49.7282 51.1786 52.6197 54.0520 55.4760 56.8923 58.3012 59.7031 61.0983 62.4872 63.8701 65.2472 66.6188 67.9852 69.3465 70.7029 72.0547 73.4020

12.1157 15.2018 17.7300 19.9974 22.1053 24.1028 26.0178 27.8680 29.6658 31.4198 33.1366 34.8213 36.4778 38.1094 39.7188 41.3081 42.8792 44.4338 45.9731 47.4985 49.0108 50.5111 52.0002 53.4788 54.9475 56.4069 57.8576 59.3000 60.7346 62.1619 63.5820 64.9955 66.4025 67.8035 69.1986 70.5881 71.9722 73.3512 74.7253 76.0946

15.1367 18.4207 21.1075 23.5127 25.7448 27.8563 29.8775 31.8276 33.7199 35.5640 37.3670 39.1344 40.8707 42.5793 44.2632 45.9249 47.5664 49.1894 50.7955 52.3860 53.9620 55.5246 57.0746 58.6130 60.1403 61.6573 63.1645 64.6624 66.1517 67.6326 69.1057 70.5712 72.0296 73.4812 74.9262 76.3650 77.7977 79.2247 80.6462 82.0623

c 2000 by Chapman & Hall/CRC 

Critical values for the chi–square distribution χ2α,ν . ν 41 42 43 44 45 46 47 48 49 50 60 70 80 90 100 200 300 400 500 600 700 800 900 1000 1500 2000 2500 3000 3500 4000 4500 5000 5500 6000 6500 7000 7500 8000 8500 9000 9500 10000

.10

.05

.025

α .01

.005

52.95 56.94 60.56 64.95 68.05 54.09 58.12 61.78 66.21 69.34 55.23 59.30 62.99 67.46 70.62 56.37 60.48 64.20 68.71 71.89 57.51 61.66 65.41 69.96 73.17 58.64 62.83 66.62 71.20 74.44 59.77 64.00 67.82 72.44 75.70 60.91 65.17 69.02 73.68 76.97 62.04 66.34 70.22 74.92 78.23 63.17 67.50 71.42 76.15 79.49 74.40 79.08 83.30 88.38 91.95 85.53 90.53 95.02 100.43 104.21 96.58 101.88 106.63 112.33 116.32 107.57 113.15 118.14 124.12 128.30 118.50 124.34 129.56 135.81 140.17 226.02 233.99 241.06 249.45 255.26 331.79 341.40 349.87 359.91 366.84 436.65 447.63 457.31 468.72 476.61 540.93 553.13 563.85 576.49 585.21 644.80 658.09 669.77 683.52 692.98 748.36 762.66 775.21 789.97 800.13 851.67 866.91 880.28 895.98 906.79 954.78 970.90 985.03 1001.63 1013.04 1057.72 1074.68 1089.53 1106.97 1118.95 1570.61 1591.21 1609.23 1630.35 1644.84 2081.47 2105.15 2125.84 2150.07 2166.66 2591.04 2617.43 2640.47 2667.43 2685.89 3099.69 3128.54 3153.70 3183.13 3203.28 3607.64 3638.75 3665.87 3697.57 3719.26 4115.05 4148.25 4177.19 4211.01 4234.14 4622.00 4657.17 4687.83 4723.63 4748.12 5128.58 5165.61 5197.88 5235.57 5261.34 5634.83 5673.64 5707.45 5746.93 5773.91 6140.81 6181.31 6216.59 6257.78 6285.92 6646.54 6688.67 6725.36 6768.18 6797.45 7152.06 7195.75 7233.79 7278.19 7308.53 7657.38 7702.58 7741.93 7787.86 7819.23 8162.53 8209.19 8249.81 8297.20 8329.58 8667.52 8715.59 8757.44 8806.26 8839.60 9172.36 9221.81 9264.85 9315.05 9349.34 9677.07 9727.86 9772.05 9823.60 9858.81 10181.66 10233.75 10279.07 10331.93 10368.03

c 2000 by Chapman & Hall/CRC 

.001

.0005

.0001

74.74 77.46 83.47 76.08 78.82 84.88 77.42 80.18 86.28 78.75 81.53 87.68 80.08 82.88 89.07 81.40 84.22 90.46 82.72 85.56 91.84 84.04 86.90 93.22 85.35 88.23 94.60 86.66 89.56 95.97 99.61 102.69 109.50 112.32 115.58 122.75 124.84 128.26 135.78 137.21 140.78 148.63 149.45 153.17 161.32 267.54 272.42 283.06 381.43 387.20 399.76 493.13 499.67 513.84 603.45 610.65 626.24 712.77 720.58 737.46 821.35 829.71 847.78 929.33 938.21 957.38 1036.83 1046.19 1066.40 1143.92 1153.74 1174.93 1674.97 1686.81 1712.30 2201.16 2214.68 2243.81 2724.22 2739.25 2771.57 3245.08 3261.45 3296.66 3764.26 3781.87 3819.74 4282.11 4300.88 4341.22 4798.87 4818.73 4861.40 5314.73 5335.62 5380.48 5829.81 5851.68 5898.63 6344.23 6367.02 6415.98 6858.05 6881.74 6932.61 7371.35 7395.90 7448.62 7884.18 7909.57 7964.06 8396.59 8422.78 8479.00 8908.62 8935.59 8993.48 9420.30 9448.03 9507.53 9931.67 9960.13 10021.21 10442.73 10471.91 10534.52

6.4.5 Percentage points, chi–square over degrees of freedom distribution The following table gives the percentage points of the sampling distribution of s2 /σ 2 , referred to as the percentage points of the χ2 /d.f. distribution (read “chi–square over degrees of freedom”). These percentage points are a function of the sample size. ν 1 2 3 4 5

0.05 0.000 0.001 0.005 0.016 0.032

Probability in percent 0.1 0.5 1.0 2.5 0.000 0.000 0.000 0.001 0.001 0.005 0.010 0.025 0.008 0.024 0.038 0.072 0.023 0.052 0.074 0.121 0.042 0.082 0.111 0.166

5.0 0.004 0.051 0.117 0.178 0.229

95 3.841 2.996 2.605 2.372 2.214

Probability in percent 97.5 99 99.5 99.9 99.95 5.024 6.635 7.879 10.828 12.116 3.689 4.605 5.298 6.908 7.601 3.116 3.782 4.279 5.422 5.910 2.786 3.319 3.715 4.617 4.999 2.567 3.017 3.350 4.103 4.421

6 7 8 9 10

0.050 0.069 0.089 0.108 0.126

0.064 0.085 0.107 0.128 0.148

0.113 0.141 0.168 0.193 0.216

0.145 0.177 0.206 0.232 0.256

0.206 0.241 0.272 0.300 0.325

0.273 0.310 0.342 0.369 0.394

2.099 2.010 1.938 1.880 1.831

2.408 2.288 2.192 2.114 2.048

2.802 2.639 2.511 2.407 2.321

3.091 2.897 2.744 2.621 2.519

3.743 3.475 3.266 3.097 2.959

4.017 3.717 3.484 3.296 3.142

15 20 25 30 40 50

0.207 0.270 0.320 0.360 0.423 0.469

0.232 0.296 0.346 0.386 0.448 0.493

0.307 0.372 0.421 0.460 0.518 0.560

0.349 0.413 0.461 0.498 0.554 0.594

0.417 0.480 0.525 0.560 0.611 0.647

0.484 0.543 0.584 0.616 0.663 0.695

1.666 1.571 1.506 1.459 1.394 1.350

1.833 1.708 1.626 1.566 1.484 1.428

2.039 1.878 1.773 1.696 1.592 1.523

2.187 2.000 1.877 1.789 1.669 1.590

2.513 2.266 2.105 1.990 1.835 1.733

2.648 2.375 2.198 2.072 1.902 1.791

75 100 150 200 500 1000

0.548 0.599 0.663 0.703 0.810 0.859

0.570 0.619 0.681 0.719 0.816 0.868

0.629 0.673 0.728 0.761 0.845 0.889

0.660 0.701 0.751 0.782 0.859 0.899

0.706 0.742 0.787 0.814 0.880 0.914

0.747 0.779 0.818 0.841 0.898 0.928

1.283 1.243 1.197 1.170 1.111 1.075

1.345 1.296 1.239 1.205 1.128 1.090

1.419 1.358 1.288 1.247 1.153 1.107

1.470 1.402 1.322 1.276 1.170 1.119

1.581 1.494 1.395 1.338 1.207 1.144

1.626 1.532 1.424 1.362 1.221 1.154

6.5

ERLANG DISTRIBUTION

6.5.1

Properties pdf f (x) = mean

variance skewness

xn−1 e−x/β , β n (n − 1)!

µ = nβ σ 2 = nβ 2 √ β1 = 2/ n

6 n mgf m(t) = (1 − βt)−n

kurtosis

char function

β2 = 3 +

φ(t) = (1 − βit)−n

c 2000 by Chapman & Hall/CRC 

x ≥ 0, β > 0, n ∈ N

6.5.2

Probability density function

The probability density function is skewed to the right with n as the shape parameter.

Figure 6.6: Probability density functions for an Erlang random variable. 6.5.3

Related distributions

If X is an Erlang random variable with parameters β and n = 1, then X is an exponential random variable with parameter λ = 1/β. 6.6

EXPONENTIAL DISTRIBUTION

6.6.1

Properties pdf f (x) = λe−λx , mean

x ≥ 0, λ > 0

µ = 1/λ

variance

σ 2 = 1/λ2

skewness

β1 = 2

kurtosis

β2 = 9 λ λ−t λ φ(t) = λ − it

mgf m(t) = char function 6.6.2

Probability density function

The probability density function is skewed to the right. The tail of the distribution is heavier for larger values of λ. c 2000 by Chapman & Hall/CRC 

Figure 6.7: Probability density functions for an exponential random variable. 6.6.3

Related distributions

Let X be an exponential random variable with parameter λ. (1) If λ = 1/2, then X is a chi–square random variable with ν = 2. √ (2) The  random variable X has a Rayleigh distribution with parameter σ = 1/(2λ). (3) The random variable Y = X 1/α has a Weibull distribution with parameters α and λ−1/α . (4) The random variable Y = e−X has a power function distribution with parameters b = 1 and c = λ. (5) The random variable Y = aeX has a Pareto distribution with parameters a and θ = λ. (6) The random variable Y = α − ln X has an extreme–value distribution with parameters α and β = 1/λ. (7) Let X1 , X2 , . . . , Xn be independent exponential random variables each with parameter λ. (a) The random variable Y = min(X1 , X2 , . . . , Xn ) has an exponential distribution with parameter nλ. (b) The random variable Y = X1 + X2 + · · · + Xn has an Erlang distribution with parameters β = 1/λ and n. (8) Let X1 and X2 be independent exponential random variables each with parameter λ. The random variable Y = X1 − X2 has a Laplace distribution with parameters 0 and 1/λ. (9) Let X be an exponential random variable with parameter λ = 1. The random variable Y = − ln[e−X /(1 + e−X )] has a (standard) logistic c 2000 by Chapman & Hall/CRC 

distribution with parameters α = 0 and β = 1. (10) Let X1 and X2 be independent exponential random variables with parameter λ = 1. (a) The random variable Y = X1 /(X1 + X2 ) has a (standard) uniform distribution with parameters a = 0 and b = 1. (b) The random variable W = − ln(X1 /X2 ) has a (standard) logistic distribution with parameters α = 0 and β = 1. 6.7

EXTREME–VALUE DISTRIBUTION

6.7.1

Properties −(x−α)/β ] pdf f (x) = (1/β)e−(x−α)/β e[−e

mean variance skewness kurtosis

µ = α + γβ,

γ = 0.5772156649 . . . (Euler’s constant)

2 2

π β 6 √ 6 6 ϕ (1) β1 = − π3 β2 = 27/5 = 5.4

σ2 =

mgf m(t) = eαt Γ(1 − βt), char function

x, α ∈ R, β > 0

t < 1/β

φ(t) = eαit Γ(1 − βit)

where Γ(x) is the gamma function and ϕ(x) is the digamma function (see pages 515 and 518). 6.7.2

Probability density function

The probability density function is skewed slightly to the right with location parameter α. 6.7.3

Related distributions

(1) The standard extreme–value distribution has α = 0 and β = 1. (2) If X is an extreme–value random variable with parameters α and β, then the random variable Y = (X − α)/β has a (standard) extreme– value distribution with parameters 0 and 1. (3) If X is a (standard) extreme–value random variable with parameters −X/c ) has a power α = 0 and β = 1, then the random variable Y = e(−e function distribution with parameters b = 0 and c. (4) If X is a extreme–value random variable with parameters α = 0 and " #1/θ −X β = 1, then the random variable Y = a 1 − e(−e ) has a Pareto distribution with parameters a and θ. c 2000 by Chapman & Hall/CRC 

Figure 6.8: Probability density functions for an extreme–value random variable. (5) Let X1 and X2 be independent extreme–value random variables with parameters α and β. The random variable Y = X1 − X2 has a logistic distribution with parameters 0 and β. 6.8

F DISTRIBUTION

6.8.1

Properties  ν21 ν22  2 ν1 ν2 Γ ν1 +ν 2 pdf f (x) = x(ν1 /2)−1 (ν2 + ν1 x)−(ν1 +ν2 )/2 Γ(ν1 /2)Γ(ν2 /2)

mean variance skewness kurtosis

x > 0, ν1 , ν2 > 0 ν2 µ= , ν2 ≥ 3 ν2 − 2 2ν22 (ν1 + ν2 − 2) , ν2 ≥ 5 ν1 (ν2 − 2)2 (ν2 − 4)  (2ν1 + ν2 − 2) 8(ν2 − 4) √ β1 = √ , ν2 ≥ 7 ν1 (ν2 − 6) ν1 + ν2 − 2

σ2 =

β2 = 3 + 12[(ν2 − 2)2 (ν2 − 4)+ν1 (ν1 + ν2 − 2)(5ν2 − 22)] ν1 (ν2 − 6)(ν2 − 8)(ν1 + ν2 − 2) ν2 ≥ 9

c 2000 by Chapman & Hall/CRC 

mgf m(t) = does not exist       ν1 + ν2 ν2 ν1 ν2 itν2 char function φ(t) = Γ Γ ψ ,1 − ; 2 2 2 2 ν1 where Γ(x) is the gamma function and ψ is the confluent hypergeometric function of the second kind (see pages 515 and 521). 6.8.2

Probability density function

The probability density function is skewed to the right with shape parameters ν1 and ν2 . For fixed ν2 , the tail becomes lighter as ν1 increases.

Figure 6.9: Probability density functions for an F random variable. 6.8.3

Related distributions

(1) If X has an F distribution with ν1 and ν2 degrees of freedom, then the random variable Y = 1/X has an F distribution with ν2 and ν1 degrees of freedom. (2) If X has an F distribution with ν1 and ν2 degrees of freedom, the random variable ν1 X tends to a chi–square distribution with ν1 degrees of freedom as ν2 → ∞. (3) Let X1 and X2 be independent F random variables with ν1 = ν2 = ν degrees of freedom. The random variable √    ν Y = (6.6) X1 − X2 2 has a t distribution with ν degrees of freedom.

c 2000 by Chapman & Hall/CRC 

(4) If X has an F distribution with parameters ν1 and ν2 , the random variable ν1 X/ν2 Y = (6.7) ν1 X 1+ ν2 has a beta distribution with parameters α = ν2 /2 and β = ν1 /2. 6.8.4

Critical values for the F distribution

Given values of ν1 , ν2 , and α, the tables on pages 132–137 contain values of Fα,ν1 ,ν2 such that  Fα,ν1 ,ν2 1−α= f (x) dx 0

 =

0

Fα,ν1 ,ν2

(6.8)  ν21 ν22  2 ν Γ ν1 +ν ν 1 2 2 x(ν1 /2)−1 (ν2 + ν1 x)−(ν1 +ν2 )/2 dx Γ(ν1 /2)Γ(ν2 /2)

Note that F1−α for ν1 and ν2 degrees of freedom is the reciprocal of Fα for ν2 and ν1 degrees of freedom. For example, F.05,4,7 =

1 1 = = .164 F.95,7,4 6.09

(6.9)

Example 6.41 : Use the following tables to find the values F.1,4,9 and F.95,12,15 . Solution: (S1) The top rows of the tables on pages 132–137 contain entries for the numerator degrees of freedom and the left–hand column contains the denominator degrees of freedom. The intersection of the ν1 degrees of freedom column and the ν2 row may be used to find critical values of the form Fα,ν1 ,ν2 such that Prob [F ≥ Fα,ν1 ,ν2 ] = α. (S2) F.1,4,9 = 2.69 =⇒ Prob [F ≥ 2.69] = .1 1 1 = = .3817 =⇒ Prob [F ≥ .3817] = .95 F.95,12,15 = F.05,15,12 2.62 (S3) Illustrations:

c 2000 by Chapman & Hall/CRC 

ν1 =1 39.86 8.53 5.54 4.54

4.06 3.78 3.59 3.46 3.36

3.29 3.23 3.18 3.14 3.10

3.07 3.05 3.03 3.01 2.99

2.97 2.92 2.81 2.76 2.71

ν2 1 2 3 4

5 6 7 8 9

10 11 12 13 14

15 16 17 18 19

25 50 100 ∞

20 25 50 100 ∞

2.92 2.81 2.76 2.71

2.53 2.41 2.36 2.30 2.59 2.53 2.41 2.36 2.30

2.70 2.67 2.64 2.62 2.61

2.92 2.86 2.81 2.76 2.73

3.78 3.46 3.26 3.11 3.01

2 49.50 9.00 5.46 4.32

2.38 2.32 2.20 2.14 2.08

2.49 2.46 2.44 2.42 2.40

2.73 2.66 2.61 2.56 2.52

3.62 3.29 3.07 2.92 2.81

3 53.59 9.16 5.39 4.19

2.25 2.18 2.06 2.00 1.94

2.36 2.33 2.31 2.29 2.27

2.61 2.54 2.48 2.43 2.39

3.52 3.18 2.96 2.81 2.69

4 55.83 9.24 5.34 4.11

2.16 2.09 1.97 1.91 1.85

2.27 2.24 2.22 2.20 2.18

2.52 2.45 2.39 2.35 2.31

3.45 3.11 2.88 2.73 2.61

5 57.24 9.29 5.31 4.05

2.09 2.02 1.90 1.83 1.77

2.21 2.18 2.15 2.13 2.11

2.46 2.39 2.33 2.28 2.24

3.40 3.05 2.83 2.67 2.55

6 58.20 9.33 5.28 4.01

2.04 1.97 1.84 1.78 1.72

2.16 2.13 2.10 2.08 2.06

2.41 2.34 2.28 2.23 2.19

3.37 3.01 2.78 2.62 2.51

7 58.91 9.35 5.27 3.98

2.00 1.93 1.80 1.73 1.67

2.12 2.09 2.06 2.04 2.02

2.38 2.30 2.24 2.20 2.15

3.34 2.98 2.75 2.59 2.47

8 59.44 9.37 5.25 3.95

1.96 1.89 1.76 1.69 1.63

2.09 2.06 2.03 2.00 1.98

2.35 2.27 2.21 2.16 2.12

3.32 2.96 2.72 2.56 2.44

9 59.86 9.38 5.24 3.94

1.94 1.87 1.73 1.66 1.60

2.06 2.03 2.00 1.98 1.96

2.32 2.25 2.19 2.14 2.10

3.30 2.94 2.70 2.54 2.42

10 60.19 9.39 5.23 3.92

1.69 1.61 1.44 1.35 1.24

1.83 1.79 1.76 1.74 1.71

2.12 2.04 1.97 1.92 1.87

3.15 2.77 2.52 2.35 2.22

50 62.69 9.47 5.15 3.80

1.65 1.56 1.39 1.29 1.17

1.79 1.76 1.73 1.70 1.67

2.09 2.01 1.94 1.88 1.83

3.13 2.75 2.50 2.32 2.19

100 63.01 9.48 5.14 3.78

1.61 1.52 1.34 1.20 1.00

1.76 1.72 1.69 1.66 1.63

2.06 1.97 1.90 1.85 1.80

3.10 2.72 2.47 2.29 2.16

∞ 63.33 9.49 5.13 3.76

Critical values for the F distribution

For given values of ν1 and ν2 , the following table contains values of F0.1,ν1 ,ν2 ; defined by Prob [F ≥ F0.1,ν1 ,ν2 ] = α = 0.1.

c 2000 by Chapman & Hall/CRC 

2.32 2.20 2.14 2.08

2.18 2.06 2.00 1.94

2.09 1.97 1.91 1.85

2.02 1.90 1.83 1.77

1.97 1.84 1.78 1.72

1.93 1.80 1.73 1.67

ν1 =1 161.4 18.51 10.13 7.71

6.61 5.99 5.59 5.32 5.12

4.96 4.84 4.75 4.67 4.60

4.54 4.49 4.45 4.41 4.38

4.35 4.24 4.03 3.94 3.84

ν2 1 2 3 4

5 6 7 8 9

10 11 12 13 14

c 2000 by Chapman & Hall/CRC 

15 16 17 18 19

20 25 50 100 ∞

3.49 3.39 3.18 3.09 3.00

3.68 3.63 3.59 3.55 3.52

4.10 3.98 3.89 3.81 3.74

5.79 5.14 4.74 4.46 4.26

2 199.5 19.00 9.55 6.94

3.10 2.99 2.79 2.70 2.60

3.29 3.24 3.20 3.16 3.13

3.71 3.59 3.49 3.41 3.34

5.41 4.76 4.35 4.07 3.86

3 215.7 19.16 9.28 6.59

2.87 2.76 2.56 2.46 2.37

3.06 3.01 2.96 2.93 2.90

3.48 3.36 3.26 3.18 3.11

5.19 4.53 4.12 3.84 3.63

4 224.6 19.25 9.12 6.39

2.71 2.60 2.40 2.31 2.21

2.90 2.85 2.81 2.77 2.74

3.33 3.20 3.11 3.03 2.96

5.05 4.39 3.97 3.69 3.48

5 230.2 19.30 9.01 6.26

2.60 2.49 2.29 2.19 2.10

2.79 2.74 2.70 2.66 2.63

3.22 3.09 3.00 2.92 2.85

4.95 4.28 3.87 3.58 3.37

6 234.0 19.33 8.94 6.16

2.51 2.40 2.20 2.10 2.01

2.71 2.66 2.61 2.58 2.54

3.14 3.01 2.91 2.83 2.76

4.88 4.21 3.79 3.50 3.29

7 236.8 19.35 8.89 6.09

2.45 2.34 2.13 2.03 1.94

2.64 2.59 2.55 2.51 2.48

3.07 2.95 2.85 2.77 2.70

4.82 4.15 3.73 3.44 3.23

8 238.9 19.37 8.85 6.04

2.39 2.28 2.07 1.97 1.88

2.59 2.54 2.49 2.46 2.42

3.02 2.90 2.80 2.71 2.65

4.77 4.10 3.68 3.39 3.18

9 240.5 19.38 8.81 6.00

2.35 2.24 2.03 1.93 1.83

2.54 2.49 2.45 2.41 2.38

2.98 2.85 2.75 2.67 2.60

4.74 4.06 3.64 3.35 3.14

10 241.9 19.40 8.79 5.96

1.97 1.84 1.60 1.48 1.35

2.18 2.12 2.08 2.04 2.00

2.64 2.51 2.40 2.31 2.24

4.44 3.75 3.32 3.02 2.80

50 251.8 19.48 8.58 5.70

1.91 1.78 1.52 1.39 1.25

2.12 2.07 2.02 1.98 1.94

2.59 2.46 2.35 2.26 2.19

4.41 3.71 3.27 2.97 2.76

100 253.0 19.49 8.55 5.66

1.84 1.71 1.45 1.28 1.00

2.07 2.01 1.96 1.92 1.88

2.54 2.40 2.30 2.21 2.13

4.36 3.67 3.23 2.93 2.71

∞ 254.3 19.50 8.53 5.63

Critical values for the F distribution

For given values of ν1 and ν2 , the following table contains values of F0.05,ν1 ,ν2 ; defined by Prob [F ≥ F0.05,ν1 ,ν2 ] = α = 0.05.

ν1 =1 647.8 38.51 17.44 12.22

10.01 8.81 8.07 7.57 7.21

6.94 6.72 6.55 6.41 6.30

6.20 6.12 6.04 5.98 5.92

5.87 5.69 5.34 5.18 5.02

ν2 1 2 3 4

5 6 7 8 9

10 11 12 13 14

c 2000 by Chapman & Hall/CRC 

15 16 17 18 19

20 25 50 100 ∞

4.46 4.29 3.97 3.83 3.69

4.77 4.69 4.62 4.56 4.51

5.46 5.26 5.10 4.97 4.86

8.43 7.26 6.54 6.06 5.71

2 799.5 39.00 16.04 10.65

3.86 3.69 3.39 3.25 3.12

4.15 4.08 4.01 3.95 3.90

4.83 4.63 4.47 4.35 4.24

7.76 6.60 5.89 5.42 5.08

3 864.2 39.17 15.44 9.98

3.51 3.35 3.05 2.92 2.79

3.80 3.73 3.66 3.61 3.56

4.47 4.28 4.12 4.00 3.89

7.39 6.23 5.52 5.05 4.72

4 899.6 39.25 15.10 9.60

3.29 3.13 2.83 2.70 2.57

3.58 3.50 3.44 3.38 3.33

4.24 4.04 3.89 3.77 3.66

7.15 5.99 5.29 4.82 4.48

5 921.8 39.30 14.88 9.36

3.13 2.97 2.67 2.54 2.41

3.41 3.34 3.28 3.22 3.17

4.07 3.88 3.73 3.60 3.50

6.98 5.82 5.12 4.65 4.32

6 937.1 39.33 14.73 9.20

3.01 2.85 2.55 2.42 2.29

3.29 3.22 3.16 3.10 3.05

3.95 3.76 3.61 3.48 3.38

6.85 5.70 4.99 4.53 4.20

7 948.2 39.36 14.62 9.07

2.91 2.75 2.46 2.32 2.19

3.20 3.12 3.06 3.01 2.96

3.85 3.66 3.51 3.39 3.29

6.76 5.60 4.90 4.43 4.10

8 956.7 39.37 14.54 8.98

2.84 2.68 2.38 2.24 2.11

3.12 3.05 2.98 2.93 2.88

3.78 3.59 3.44 3.31 3.21

6.68 5.52 4.82 4.36 4.03

9 963.3 39.39 14.47 8.90

2.77 2.61 2.32 2.18 2.05

3.06 2.99 2.92 2.87 2.82

3.72 3.53 3.37 3.25 3.15

6.62 5.46 4.76 4.30 3.96

10 968.6 39.40 14.42 8.84

2.25 2.08 1.75 1.59 1.43

2.55 2.47 2.41 2.35 2.30

3.22 3.03 2.87 2.74 2.64

6.14 4.98 4.28 3.81 3.47

50 1008 39.48 14.01 8.38

2.17 2.00 1.66 1.48 1.27

2.47 2.40 2.33 2.27 2.22

3.15 2.96 2.80 2.67 2.56

6.08 4.92 4.21 3.74 3.40

100 1013 39.49 13.96 8.32

2.09 1.91 1.54 1.37 1.00

2.40 2.32 2.25 2.19 2.13

3.08 2.88 2.72 2.60 2.49

6.02 4.85 4.14 3.67 3.33

∞ 1018 39.50 13.90 8.26

Critical values for the F distribution

For given values of ν1 and ν2 , the following table contains values of F0.025,ν1 ,ν2 ; defined by Prob [F ≥ F0.025,ν1 ,ν2 ] = α = 0.025.

ν1 =1 4052 98.50 34.12 21.20

16.26 13.75 12.25 11.26 10.56

10.04 9.65 9.33 9.07 8.86

8.68 8.53 8.40 8.29 8.18

8.10 7.77 7.17 6.90 6.63

ν2 1 2 3 4

5 6 7 8 9

10 11 12 13 14

c 2000 by Chapman & Hall/CRC 

15 16 17 18 19

20 25 50 100 ∞

5.85 5.57 5.06 4.82 4.61

6.36 6.23 6.11 6.01 5.93

7.56 7.21 6.93 6.70 6.51

13.27 10.92 9.55 8.65 8.02

2 5000 99.00 30.82 18.00

4.94 4.68 4.20 3.98 3.78

5.42 5.29 5.18 5.09 5.01

6.55 6.22 5.95 5.74 5.56

12.06 9.78 8.45 7.59 6.99

3 5403 99.17 29.46 16.69

4.43 4.18 3.72 3.51 3.32

4.89 4.77 4.67 4.58 4.50

5.99 5.67 5.41 5.21 5.04

11.39 9.15 7.85 7.01 6.42

4 5625 99.25 28.71 15.98

4.10 3.85 3.41 3.21 3.02

4.56 4.44 4.34 4.25 4.17

5.64 5.32 5.06 4.86 4.69

10.97 8.75 7.46 6.63 6.06

5 5764 99.30 28.24 15.52

3.87 3.63 3.19 2.99 2.80

4.32 4.20 4.10 4.01 3.94

5.39 5.07 4.82 4.62 4.46

10.67 8.47 7.19 6.37 5.80

6 5859 99.33 27.91 15.21

3.70 3.46 3.02 2.82 2.64

4.14 4.03 3.93 3.84 3.77

5.20 4.89 4.64 4.44 4.28

10.46 8.26 6.99 6.18 5.61

7 5928 99.36 27.67 14.98

3.56 3.32 2.89 2.69 2.51

4.00 3.89 3.79 3.71 3.63

5.06 4.74 4.50 4.30 4.14

10.29 8.10 6.84 6.03 5.47

8 5981 99.37 27.49 14.80

3.46 3.22 2.78 2.59 2.41

3.89 3.78 3.68 3.60 3.52

4.94 4.63 4.39 4.19 4.03

10.16 7.98 6.72 5.91 5.35

9 6022 99.39 27.35 14.66

3.37 3.13 2.70 2.50 2.32

3.80 3.69 3.59 3.51 3.43

4.85 4.54 4.30 4.10 3.94

10.05 7.87 6.62 5.81 5.26

10 6056 99.40 27.23 14.55

2.64 2.40 1.95 1.74 1.53

3.08 2.97 2.87 2.78 2.71

4.12 3.81 3.57 3.38 3.22

9.24 7.09 5.86 5.07 4.52

50 6303 99.48 26.35 13.69

2.54 2.29 1.82 1.60 1.32

2.98 2.86 2.76 2.68 2.60

4.01 3.71 3.47 3.27 3.11

9.13 6.99 5.75 4.96 4.41

100 6334 99.49 26.24 13.58

2.42 2.17 1.70 1.45 1.00

2.87 2.75 2.65 2.57 2.49

3.91 3.60 3.36 3.17 3.00

9.02 6.88 5.65 4.86 4.31

∞ 6336 99.50 26.13 13.46

Critical values for the F distribution

For given values of ν1 and ν2 , the following table contains values of F0.01,ν1 ,ν2 ; defined by Prob [F ≥ F0.01,ν1 ,ν2 ] = α = 0.01.

ν1 =1 16211 198.5 55.55 31.33

22.78 18.63 16.24 14.69 13.61

12.83 12.23 11.75 11.37 11.06

10.80 10.58 10.38 10.22 10.07

9.94 9.48 8.63 8.24 7.88

ν2 1 2 3 4

5 6 7 8 9

10 11 12 13 14

c 2000 by Chapman & Hall/CRC 

15 16 17 18 19

20 25 50 100 ∞

6.99 6.60 5.90 5.59 5.30

7.70 7.51 7.35 7.21 7.09

9.43 8.91 8.51 8.19 7.92

18.31 14.54 12.40 11.04 10.11

2 20000 199.0 49.80 26.28

5.82 5.46 4.83 4.54 4.28

6.48 6.30 6.16 6.03 5.92

8.08 7.60 7.23 6.93 6.68

16.53 12.92 10.88 9.60 8.72

3 21615 199.2 47.47 24.26

5.17 4.84 4.23 3.96 3.72

5.80 5.64 5.50 5.37 5.27

7.34 6.88 6.52 6.23 6.00

15.56 12.03 10.05 8.81 7.96

4 22500 199.2 46.19 23.15

4.76 4.43 3.85 3.59 3.35

5.37 5.21 5.07 4.96 4.85

6.87 6.42 6.07 5.79 5.56

14.94 11.46 9.52 8.30 7.47

5 23056 199.3 45.39 22.46

4.47 4.15 3.58 3.33 3.09

5.07 4.91 4.78 4.66 4.56

6.54 6.10 5.76 5.48 5.26

14.51 11.07 9.16 7.95 7.13

6 23437 199.3 44.84 21.97

4.26 3.94 3.38 3.13 2.90

4.85 4.69 4.56 4.44 4.34

6.30 5.86 5.52 5.25 5.03

14.20 10.79 8.89 7.69 6.88

7 23715 199.4 44.43 21.62

4.09 3.78 3.22 2.97 2.74

4.67 4.52 4.39 4.28 4.18

6.12 5.68 5.35 5.08 4.86

13.96 10.57 8.68 7.50 6.69

8 23925 199.4 44.13 21.35

3.96 3.64 3.09 2.85 2.62

4.54 4.38 4.25 4.14 4.04

5.97 5.54 5.20 4.94 4.72

13.77 10.39 8.51 7.34 6.54

9 24091 199.4 43.88 21.14

3.85 3.54 2.99 2.74 2.52

4.42 4.27 4.14 4.03 3.93

5.85 5.42 5.09 4.82 4.60

13.62 10.25 8.38 7.21 6.42

10 24224 199.4 43.69 20.97

2.96 2.65 2.10 1.84 1.60

3.52 3.37 3.25 3.14 3.04

4.90 4.49 4.17 3.91 3.70

12.45 9.17 7.35 6.22 5.45

50 25211 199.5 42.21 19.67

2.83 2.52 1.95 1.68 1.36

3.39 3.25 3.12 3.01 2.91

4.77 4.36 4.04 3.78 3.57

12.30 9.03 7.22 6.09 5.32

100 25337 199.5 42.02 19.50

2.69 2.38 1.81 1.51 1.00

3.26 3.11 2.98 2.87 2.78

4.64 4.23 3.90 3.65 3.44

12.14 8.88 7.08 5.95 5.19

∞ 25465 199.5 41.83 19.32

Critical values for the F distribution

For given values of ν1 and ν2 , the following table contains values of F0.005,ν1 ,ν2 ; defined by Prob [F ≥ F0.005,ν1 ,ν2 ] = α = 0.005.

ν1 =1 998.5 167.0 74.14

47.18 35.51 29.25 25.41 22.86

21.04 19.69 18.64 17.82 17.14

16.59 16.12 15.72 15.38 15.08

14.82 13.88 12.22 11.50 10.83

ν2 2 3 4

5 6 7 8 9

10 11 12 13 14

c 2000 by Chapman & Hall/CRC 

15 16 17 18 19

20 25 50 100 ∞

9.95 9.22 7.96 7.41 6.91

11.34 10.97 10.66 10.39 10.16

14.91 13.81 12.97 12.31 11.78

37.12 27.00 21.69 18.49 16.39

2 999.0 148.5 61.25

8.10 7.45 6.34 5.86 5.42

9.34 9.01 8.73 8.49 8.28

12.55 11.56 10.80 10.21 9.73

33.20 23.70 18.77 15.83 13.90

3 999.2 141.1 56.18

7.10 6.49 5.46 5.02 4.62

8.25 7.94 7.68 7.46 7.27

11.28 10.35 9.63 9.07 8.62

31.09 21.92 17.20 14.39 12.56

4 999.2 137.1 53.44

6.46 5.89 4.90 4.48 4.10

7.57 7.27 7.02 6.81 6.62

10.48 9.58 8.89 8.35 7.92

29.75 20.80 16.21 13.48 11.71

5 999.3 134.6 51.71

6.02 5.46 4.51 4.11 3.74

7.09 6.80 6.56 6.35 6.18

9.93 9.05 8.38 7.86 7.44

28.83 20.03 15.52 12.86 11.13

6 999.3 132.8 50.53

5.69 5.15 4.22 3.83 3.47

6.74 6.46 6.22 6.02 5.85

9.52 8.66 8.00 7.49 7.08

28.16 19.46 15.02 12.40 10.70

7 999.4 131.6 49.66

5.44 4.91 4.00 3.61 3.27

6.47 6.19 5.96 5.76 5.59

9.20 8.35 7.71 7.21 6.80

27.65 19.03 14.63 12.05 10.37

8 999.4 130.6 49.00

5.24 4.71 3.82 3.44 3.10

6.26 5.98 5.75 5.56 5.39

8.96 8.12 7.48 6.98 6.58

27.24 18.69 14.33 11.77 10.11

9 999.4 129.9 48.47

5.08 4.56 3.67 3.30 2.96

6.08 5.81 5.58 5.39 5.22

8.75 7.92 7.29 6.80 6.40

26.92 18.41 14.08 11.54 9.89

10 999.4 129.2 48.05

3.77 3.28 2.44 2.08 1.75

4.70 4.45 4.24 4.06 3.90

7.19 6.42 5.83 5.37 5.00

24.44 16.31 12.20 9.80 8.26

50 999.5 124.7 44.88

3.58 3.09 2.25 1.87 1.45

4.51 4.26 4.05 3.87 3.71

6.98 6.21 5.63 5.17 4.81

24.12 16.03 11.95 9.57 8.04

100 999.5 124.1 44.47

3.38 2.89 2.06 1.65 1.00

4.31 4.06 3.85 3.67 3.51

6.76 6.00 5.42 4.97 4.60

23.79 15.75 11.70 9.33 7.81

∞ 999.5 123.5 44.05

Critical values for the F distribution

For given values of ν1 and ν2 , the following table contains values of F0.001,ν1 ,ν2 ; defined by Prob [F ≥ F0.001,ν1 ,ν2 ] = α = 0.001.

6.9

GAMMA DISTRIBUTION

6.9.1

Properties pdf f (x) = mean

variance skewness kurtosis

xα−1 e−x/β β α Γ(α)

µ = αβ σ 2 = αβ 2 √ β1 = 2/ α   2 β2 = 3 1 + α

mgf m(t) = (1 − βt)−α char function

φ(t) = (1 − iβt)−α

where Γ(x) is the gamma function (see page 515). 6.9.2

Probability density function

The probability density function is skewed to the right. For fixed β the tail becomes heavier as α increases.

Figure 6.10: Probability density functions for a gamma random variable. 6.9.3

Related distributions

Let X be a gamma random variable with parameters α and β. (1) The random variable X has a standard gamma distribution if α = 1. (2) If α = 1 and β = 1/λ, then X has an exponential distribution with parameter λ.

c 2000 by Chapman & Hall/CRC 

(3) If α = ν/2 and β = 2, then X has a chi–square distribution with ν degrees of freedom. (4) If α = n is an integer, then X has an Erlang distribution with parameters β and n. (5) If α = ν/2 and β = 1, then the random variable Y = 2X has a chi– square distribution with ν degrees of freedom. (6) As α → ∞, X tends to a normal distribution with parameters µ = αβ and σ 2 = αβ 2 . (7) Suppose X1 is a gamma random variable with parameters α = 1 and β = β1 , X2 is a gamma random variable with parameters α = 1 and β = β2 , and X1 and X2 are independent. The random variable Y = X1 /(X1 + X2 ) has a beta distribution with parameters β1 and β2 . (8) Let X1 , X2 , . . . , Xn be independent gamma random variables with parameters α and βi for i = 1, 2, . . . , n. The random variable Y = X1 + X2 + · · · + Xn has a gamma distribution with parameters α and β = β1 + β2 + · · · + βn . 6.10

HALF–NORMAL DISTRIBUTION

6.10.1

Properties  2 2 2θ θ x , pdf f (x) = exp − 2 π π mean

variance skewness

x ≥ 0, θ > 0

µ = 1/θ π−2 2θ2 √ 2(4 − π) β1 = (π − 2)3/2

σ2 =

3π 2 − 4π − 12 (π − 2)2  √   2 πt πt 1 + erf mgf m(t) = exp 2 4θ 2θ    √  2 πt πit char function φ(t) = exp − 2 1 + erf 4θ 2θ kurtosis

β2 =

where erf(x) is the error function (see page 512). 6.10.2

Probability density function

The probability density function is skewed to the right. As θ increases the tail becomes lighter.

c 2000 by Chapman & Hall/CRC 

Figure 6.11: Probability density functions for a half–normal random variable. 6.11

INVERSE GAUSSIAN (WALD) DISTRIBUTION

6.11.1

Properties pdf f (x) = mean

variance skewness

λ exp 2πx3



−λ(x − µ)2 2µ2 x

 x, µ, λ > 0

µ=µ σ 2 = µ3 /λ  β1 = 3 µ/λ

15µ λ % & λ 2µ2 t mgf m(t) = exp 1− 1− µ λ % & λ 2µ2 it char function φ(t) = exp 1− 1− µ λ kurtosis

6.11.2

β2 = 3 +

Probability density function

The probability density function is skewed right. For fixed µ the probability density function becomes more bell–shaped as λ increases. 6.11.3

Related distributions

If X is an inverse Gaussian random variable with parameters µ and λ, the λ(X − µ)2 has a chi–square distribution with 1 degree random variable Y = µ2 X c 2000 by Chapman & Hall/CRC 

Figure 6.12: Probability density functions for an inverse Gaussian random variable. of freedom. 6.12

LAPLACE DISTRIBUTION

6.12.1

Properties pdf f (x) = mean

  1 |x − α| exp − , 2β β

µ=α

variance

σ 2 = 2β 2

skewness

β1 = 0

kurtosis

β2 = 6

mgf m(t) = char function 6.12.2

x ∈ R, α ∈ R, β > 0

φ(t) =

eαt 1 − β 2 t2 eαit 1 + β 2 t2

Probability density function

The probability density function is symmetric about the parameter α. For fixed α the tails become heavier as β increases. 6.12.3

Related distributions

(1) Let X be a Laplace random variable with parameters α and β. The random variable Y = |X − α| has an exponential distribution with pac 2000 by Chapman & Hall/CRC 

Figure 6.13: Probability density functions for a Laplace random variable. rameter λ = β. The random variable W = |X −α|/β has an exponential distribution with parameter λ = 1. (2) Let X1 and X2 be independent Laplace random variables with parameters α = 0, and β1 and β2 , respectively. The random variable Y = |X1 /X2 | has an F distribution with parameters ν1 = ν2 = 2. 6.13

LOGISTIC DISTRIBUTION

6.13.1

Properties pdf f (x) = mean

e−(x−α)/β , β(1 + e−(x−α)/β )2

µ=α

variance

σ 2 = β 2 π 2 /3

skewness

β1 = 0

kurtosis

x ∈ R, α ∈ R, β ∈ R

β2 = 21/5

mgf m(t) = πβteαt / sin(πβt) char function 6.13.2

φ(t) = πβteiαt / sinh(πβt)

Probability density function

The probability density function is symmetric about the parameter α. For fixed α the tails become heavier as β increases.

c 2000 by Chapman & Hall/CRC 

Figure 6.14: Probability density functions for a logistic random variable. 6.13.3

Related distributions

(1) The random variable X has a standard logistic distribution if α = 0 and β = 1. (2) If X is a logistic random variable with parameters α and β, then the random variable Y = (X − α)/β has a (standard) logistic distribution with parameters 0 and 1. 6.14

LOGNORMAL DISTRIBUTION

6.14.1

Properties pdf f (x) = √

  1 1 exp − 2 (ln x − µ)2 2σ 2π σx

x > 0, µ ∈ R, σ > 0 mean

µ = eµ+σ

2

/2 2

2

skewness

σ 2 = e2µ+σ (eσ − 1)  2 β1 = (eσ + 2) eσ2 − 1

kurtosis

β2 = e4σ + 2e3σ + 3e2σ

variance

2

2

mgf m(t) = does not exist char function

φ(t) = does not exist

c 2000 by Chapman & Hall/CRC 

2

6.14.2

Probability density function

The probability density function is skewed to the right. The scale parameter is µ and the shape parameter is σ.

Figure 6.15: Probability density functions for a lognormal random variable. 6.14.3

Related distributions

(1) If X is a lognormal random variable with parameters µ and σ, then the random variable Y = ln X has a normal distribution with mean µ and variance σ 2 . (2) If X is a lognormal random variable with parameters µ and σ and a and b are constants, then the random variable Y = ea X b has a lognormal distribution with parameters a + bµ and bσ. (3) Let X1 and X2 be independent lognormal random variables with parameters µ1 , σ1 and µ2 , σ2 , respectively. The random variable Y = X1 /X2 has a lognormal distribution with parameters µ1 − µ2 and σ1 + σ2 . (4) Let X1 , X2 , . . . , Xn be independent lognormal random variables with parameters µi and σi for i = 1, 2, . . . , n. The random variable Y = X1 · X2 · · · Xn has a lognormal distribution with parameters µ = µ1 + µ2 + · · · + µn and σ = σ1 + σ2 + · · · + σn . (5) Let X1 , X2 , . . . , Xn be independent lognormal random √ variables with parameters µ and σ. The random variable Y = n X1 · · · · Xn has a lognormal distribution with parameters µ and σ/n.

c 2000 by Chapman & Hall/CRC 

6.15

NONCENTRAL CHI–SQUARE DISTRIBUTION

6.15.1

Properties 1 ∞ e[ 2 (x+λ)]  x(ν/2)+j−1 λj   pdf f (x) = 2ν/2 j=1 Γ ν2 + j 22j j!

mean variance skewness kurtosis

x, λ > 0, ν ∈ N

µ = ν+λ σ 2 = 2ν + 4λ √ 2 2(ν + 3λ) β1 = (ν + 2λ)3/2 β2 = 3 +

12(ν + 4λ) (ν + 2λ)2



 λt mgf m(t) = (1 − 2t) exp 1 − 2t   λit char function φ(t) = (1 − 2it)−ν/2 exp 1 − 2it −ν/2

where Γ(x) is the gamma function (see page 515). 6.15.2

Probability density function

The probability density function is skewed to the right. For fixed ν the tail becomes heavier as the noncentrality parameter λ increases.

Figure 6.16: Probability density functions for a noncentral chi–square random variable.

c 2000 by Chapman & Hall/CRC 

6.15.3

Related distributions

(1) If X is a noncentral chi–square random variable with parameters ν and λ = 0, then X is a chi–square random variable with ν degrees of freedom. (2) If X1 is a noncentral chi–square random variable with parameters ν1 and λ, X2 is a chi–square random variable with parameter ν2 , and X1 and X2 are independent, then the random variable Y = (X1 /ν1 )/(X2 /ν2 ) has a noncentral F distribution with parameters ν1 , ν2 , and λ. (3) Let X1 , X2 , . . . , Xn be independent noncentral chi–square random variables with parameters νi and λi (for i = 1, 2, . . . , n). The random variable Y = X1 + X2 + · · · + Xn has a noncentral chi–square distribution with parameters ν = ν1 + ν2 + · · · + νn and λ = λ1 + λ2 + · · · + λn . 6.16

NONCENTRAL F DISTRIBUTION

6.16.1

Properties e−λ/2 ν1 1 ν2 2 x 2 (ν1 −2) (ν1 x + ν2 )− 2 (ν1 +ν2 ) × B(ν1 /2, ν2 /2)   ν1 + ν 2 ν 1 ν1 λx , , 1 F1 2 2 2(ν1 x + ν2 ) ν /2 ν /2

pdf f (x) =

1

1

x > 0, ν1 , ν2 ∈ N , λ > 0 ν2 (ν1 + λ) , ν1 (ν2 − 2)

mean

µ=

variance

σ2 =

skewness

β1 = does not exist

kurtosis

β2 = does not exist

ν2 > 2

2ν22 ((ν1 + λ) + (ν2 − 2)(ν1 + 2λ)) , ν12 (ν2 − 4)(ν2 − 2)2

ν2 > 4

mgf m(t) = does not exist char function

φ(t) = does not exist

where B(a, b) is the beta function and p Fq is the generalized hypergeometric function defined in Chapter 18 (see pages 511 and 520). 6.16.2

Probability density function

The probability density function is skewed to the right. The parameters ν1 and ν2 are the shape parameters and λ is the noncentrality parameter. 6.16.3

Related distributions

If X has a noncentral F distribution with parameters ν1 , ν2 , and λ, then the random variable X tends to an F distribution as λ → 0. c 2000 by Chapman & Hall/CRC 

Figure 6.17: Probability density functions for a noncentral F random variable. 6.17 6.17.1

NONCENTRAL t DISTRIBUTION Properties ν ν/2 e−λ /2 × π Γ(ν/2)(ν + x2 )(ν+1)/2 j/2   j ∞  ν+j+1 λ 2x2 Γ 2 j! ν + x2 j=0 2

pdf f (x) = √

x, λ ∈ R, ν ∈ N

where Γ(x) is the gamma function (see page 515). The moments about the origin are µr = cr

ν r/2 Γ[(ν − r)/2] , 2r/2 Γ(ν/2)

ν>r

(6.10)

where c2r−1 = c2r =

r  j=1 r  j=0

6.17.2

(2r − 1)!λ2r−1 , (2j − 1)(r − j)!2r−j (2r)!λ2j , (2j)!(r − j)!2r−j

r = 1, 2, 3, . . . (6.11)

r = 1, 2, 3, . . .

Probability density function

The probability density function is skewed to the right. The shape parameter is ν and the noncentrality parameter is λ. For fixed ν the tail becomes heavier as λ increases. For large values of ν, the probability density function is approximately symmetric.

c 2000 by Chapman & Hall/CRC 

Figure 6.18: Probability density functions for a noncentral t random variable. 6.17.3

Related distributions

If X has a noncentral t distribution with parameters ν and λ = 0, then X has a t distribution with ν degrees of freedom. 6.18

NORMAL DISTRIBUTION

6.18.1

Properties pdf f (x) = mean

2 2 1 √ e−(x−µ) /2σ , σ 2π

x ∈ R, µ ∈ R, σ > 0

µ=µ

variance

σ2 = σ2

skewness

β1 = 0

kurtosis

β2 = 3

  σ 2 t2 mgf m(t) = exp µt + 2   σ 2 t2 char function φ(t) = exp µit − 2 See Chapter 7 for more details. 6.18.2

Probability density function

The probability density function is symmetric and bell–shaped about the location parameter µ. For small values of the scale parameter σ the probability density function is more compact. c 2000 by Chapman & Hall/CRC 

Figure 6.19: Probability density functions for a normal random variable. 6.18.3

Related distributions

(1) The random variable X has a standard normal distribution if µ = 0 and σ = 1. (2) If X is a normal random variable with parameters µ and σ, the random variable Y = (X − µ)/σ has a (standard) normal distribution with parameters 0 and 1. (3) If X is a normal random variable with parameters µ and σ, the random variable Y = eX has a lognormal distribution with parameters µ and σ. (4) If X is a normal random variable with parameters µ = 0 and σ = 1, then the random variable Y = eµ+σX has a lognormal distribution with parameters µ and σ. (5) If X is a normal random variable with parameters µ and σ, and a and b are constants, then the random variable Y = a + bX has a normal distribution with parameters a + bµ and bσ. (6) If X1 and X2 are independent standard normal random variables, the random variable Y = X1 /X2 has a Cauchy distribution with parameters a = 0 and b = 1. (7) If X1 and X2 are independent normal randomvariables with parameters µ = 0 and σ, then the random variable Y = X12 + X22 has a Rayleigh distribution with parameter σ. (8) Let Xi (for i = 1, 2, . . . , n) be independent, normal random variables with parameters µi and σi , and let ci be any constants. The random n  variable Y = ci Xi has a normal distribution with parameters µ = n  i=1

i=1

ci µi and σ 2 =

n  i=1

c2i σi2 .

c 2000 by Chapman & Hall/CRC 

(9) Let Xi (for i = 1, 2, . . . , n) be independent, normal random variables with parameters µ and σ, then the random variable Y = X1 + X2 + · · · + Xn has a normal distribution with mean nµ and variance nσ 2 . (10) Let Xi (for i = 1, 2, . . . , n) be independent standard normal random n  variables. The random variable Y = Xi2 has a chi–square distribui=1

tion with ν = n degrees of freedom. If µi = λi > 0 (σi = 1), then the random variable Y has a noncentral chi–square distribution with n  parameters ν = n and noncentrality parameter λ = λ2i . i=1

6.19 6.19.1

NORMAL DISTRIBUTION: MULTIVARIATE Properties pdf f (x) =

(2π)n/2

mean

µ

covariance matrix

Σ

1 

  (x − µ)T Σ−1 (x − µ) exp − 2 det(Σ)

  1 φ(t) = exp itT µ − tT Σt 2 T  where x = x1 , x2 , . . . , xn (with xi ∈ R) and Σ is a positive semi-definite matrix. Section 7.6 discusses the bivariate normal. char function

6.19.2

Probability density function

The probability density function is smooth and unimodal. Figure  6.20 shows  T 12 . two views of a bivariate normal with µ = 1 0 and Σ = 04

Figure 6.20: Two views of the probability density for a bivariate normal. c 2000 by Chapman & Hall/CRC 

6.20

PARETO DISTRIBUTION

6.20.1

Properties θaθ , x ≥ a, θ > 0, a > 0 xθ+1 aθ µ= , θ>1 θ−1 a2 θ σ2 = , θ>2 (θ − 1)2 (θ − 2) √ 2(θ + 1) θ − 2 √ β1 = , θ>3 (θ − 3) θ

pdf f (x) = mean variance skewness kurtosis

β2 =

3(θ − 2)(3θ2 + θ + 2) , θ(θ − 3)(θ − 4)

θ>4

mgf m(t) = does not exist char function

φ(t) = −aθ tθ cos(πθ/2)Γ(1 − θ) +  θ   1   − 2 , 2 , 1 − θ2 , − 14 a2 t2 − 1 F2   1 θ   3 3 θ    1 1 2 2 1−θ atiθ 1 F2 2 − 2 , 2 , 2 − 2 , − 4 a t sgn(t) + iaθ tθ Γ(1 − θ) sgn(t) sin(πθ/2)

where p Fq is the generalized hypergeometric function and sgn(t) is the signum function defined in Chapter 18 (see pages 520 and 523). 6.20.2

Probability density function

The probability density function is skewed to the right. The shape parameter is θ and the location parameter is a. 6.20.3

Related distributions

(1) Let X be a Pareto random variable with parameters a and θ. (a) The random variable Y = ln(X/a) has an exponential distribution with parameter λ = 1/θ. (b) The random variable Y = 1/X has a power function distribution with parameters 1/a and θ.   (c) The random variable Y = − ln (X/a)θ − 1 has a logistic distribution with parameters α = 0 and β = 1. (2) Let Xi (for i = 1, 2, . . . , n) be independent Pareto random variables n  with parameters a and θ. The random variable Y = 2a ln(Xi /θ) has i=1

a chi–square distribution with ν = 2n.

c 2000 by Chapman & Hall/CRC 

Figure 6.21: Probability density functions for a Pareto random variable. 6.21

POWER FUNCTION DISTRIBUTION

6.21.1

Properties cxc−1 , 0 ≤ x ≤ b, b > 0, c > 0 bc bc µ= c+1 b2 c σ2 = (c + 1)2 (c + 2) √ 2(1 − c) c + 2 √ β1 = (c + 3) c

pdf f (x) = mean variance skewness kurtosis

β2 =

3(c + 2)(3c2 − c + 2) c(c + 3)(c + 4)

mgf m(t) = does not exist      char function φ(t) = 1 F2 2c , 12 , 1 + 2c , − 14 b2 t2 +   1 c   3 3 c    1 1 2 2 c+1 ibct1 F2 2 + 2 , 2 , 2 + 2 , − 4 b t sgn(t) where p Fq is the generalized hypergeometric function and sgn(t) is the signum function defined in Chapter 18 (see pages 520 and 523). 6.21.2

Probability density function

The probability density function is “J” shaped for c < 1 and is skewed left for c > 1.

c 2000 by Chapman & Hall/CRC 

Figure 6.22: Probability density functions for a power function random variable. 6.21.3

Related distributions

(1) Let X be a power function random variable with parameters b and c. (a) The random variable Y = 1/X has a power function distribution with parameters 1/b and c. (b) If b = 1: (1) The random variable X has a beta distribution with parameters α = c and β = 1. (2) The random variable Y = − ln X has an exponential distribution with parameter λ = c. (3) The random variable Y = 1/X has a Pareto distribution with parameters a = 0 and θ = c. (4) The random variable Y = − ln(X −c − 1) has a logistic distribution with parameters α = 0 and β = 1. (5) The random variable Y = (− ln X c )1/k has a Weibull distribution with parameters α = k and β = 1. (6) The random variable Y = − ln(−c ln X) has an extreme–value distribution with parameters α = 0 and β = 1. (c) If c = 1 then X has a uniform distribution with parameters a = 0 and b. (7) Let X1 , X2 be independent power function random variables with parameters b = 1 and c. The random variable Y = −c ln(X1 /X2 ) has a Laplace distribution with parameters α = 0 and β = 1.

c 2000 by Chapman & Hall/CRC 

6.22

RAYLEIGH DISTRIBUTION

6.22.1

Properties   x x2 , exp − σ2 2σ 2  µ = σ π/2  π σ2 = σ2 2 − 2  (π − 3) π/2 β1 =  3/2 2 − π2

pdf f (x) = mean variance skewness

x ≥ 0, σ > 0

32 − 3π 2 (4 − π)2     √ σt 1 σ 2 t2 /2 1 + erf √ mgf m(t) = 2 + 2π σ t e 2 2    π iσt −σ 2 t2 /2 √ char function φ(t) = 1 + ie σ t 1 − erf − 2 2 kurtosis

β2 =

where erf(x) is the error function (see page 512). 6.22.2

Probability density function

The probability density function is skewed to the right. For large values of σ the tail is heavier.

Figure 6.23: Probability density functions for a Rayleigh random variable.

c 2000 by Chapman & Hall/CRC 

6.22.3

Related distributions

(1) If X is a Rayleigh random variable with parameter σ = 1, then X is a chi random variable with parameter n = 2. (2) If X is a Rayleigh random variable with parameter σ, then the random variable Y = X 2 has an exponential distribution with parameter λ = 1/(2σ 2 ). 6.23

t DISTRIBUTION

6.23.1

Properties    −(ν+1)/2 x2 1 Γ ν+1  ν2  1+ pdf f (x) = √ x ∈ R, ν ∈ N ν πν Γ 2 mean

variance skewness kurtosis

ν≥2 ν σ2 = , ν≥3 ν−2 µ = 0,

β1 = 0,

ν≥4

β2 = 3 +

6 , ν−4

ν≥5

mgf m(t) = does not exist char function

√ ν 21− 2 ν ν/4 |t|ν/2 Kν/2 ( ν|t|) φ(t) = Γ(ν/2)

where Kn (x) is a modified Bessel function and Γ(x) is the gamma function defined in Chapter 18 (see pages 506 and 18.8). 6.23.2

Probability density function

The probability density function is symmetric and bell–shaped centered about 0. As the degrees of freedom, ν, increases the distribution becomes more compact. 6.23.3

Related distributions

(1) If X is a t random variable with parameter ν, then the random variable Y = X 2 has an F distribution with 1 and ν degrees of freedom. (2) If X is a t random variable with parameter ν = 1, then X has a Cauchy distribution with parameters a = 0 and b = 1. (3) If X is a t random variable with parameter ν, as ν tends to infinity X tends to a standard normal distribution. The approximation is reasonable for ν ≥ 30.

c 2000 by Chapman & Hall/CRC 

Figure 6.24: Probability density functions for a t random variable. 6.23.4

Critical values for the t distribution

For a given value of ν, the number of degrees of freedom, the table on page 157 contains values of tα,ν such that Prob [t ≥ tα,ν ] = α

(6.12)

Example 6.42 : Use the table on page 157 to find the values t.05,11 and −t.01,24 . Solution: (S1) The top row of the following table contains cumulative probability and the left– hand column contains the degrees of freedom. The values in the body of the table may be used to find critical values. (S2) t.05,11 = 1.7959 since F (1.7959; 11) = .95 =⇒ Prob [t ≥ 1.7959] = .05 (S3) −t.01,24 = −2.4922 since F (2.4922; 24) = .99 =⇒ Prob [t ≤ −2.4922] = .01 (S4) Illustrations:

c 2000 by Chapman & Hall/CRC 

Critical values for the t distribution. ν 1 2 3 4 5

α = 0.1 3.078 1.886 1.638 1.533 1.476

0.05 6.314 2.920 2.353 2.132 2.015

0.025 12.706 4.303 3.182 2.776 2.571

0.01 31.821 6.965 4.541 3.747 3.365

0.005 63.657 9.925 5.841 4.604 4.032

0.0025 318.309 22.327 10.215 7.173 5.893

0.001 636.619 31.599 12.924 8.610 6.869

6 7 8 9 10

1.440 1.415 1.397 1.383 1.372

1.943 1.895 1.860 1.833 1.812

2.447 2.365 2.306 2.262 2.228

3.143 2.998 2.896 2.821 2.764

3.707 3.499 3.355 3.250 3.169

5.208 4.785 4.501 4.297 4.144

5.959 5.408 5.041 4.781 4.587

11 12 13 14 15

1.363 1.356 1.350 1.345 1.341

1.796 1.782 1.771 1.761 1.753

2.201 2.179 2.160 2.145 2.131

2.718 2.681 2.650 2.624 2.602

3.106 3.055 3.012 2.977 2.947

4.025 3.930 3.852 3.787 3.733

4.437 4.318 4.221 4.140 4.073

16 17 18 19 20

1.337 1.333 1.330 1.328 1.325

1.746 1.740 1.734 1.729 1.725

2.120 2.110 2.101 2.093 2.086

2.583 2.567 2.552 2.539 2.528

2.921 2.898 2.878 2.861 2.845

3.686 3.646 3.610 3.579 3.552

4.015 3.965 3.922 3.883 3.850

21 22 23 24 25

1.323 1.321 1.319 1.318 1.316

1.721 1.717 1.714 1.711 1.708

2.080 2.074 2.069 2.064 2.060

2.518 2.508 2.500 2.492 2.485

2.831 2.819 2.807 2.797 2.787

3.527 3.505 3.485 3.467 3.450

3.819 3.792 3.768 3.745 3.725

26 27 28 29 30

1.315 1.314 1.313 1.311 1.310

1.706 1.703 1.701 1.699 1.697

2.056 2.052 2.048 2.045 2.042

2.479 2.473 2.467 2.462 2.457

2.779 2.771 2.763 2.756 2.750

3.435 3.421 3.408 3.396 3.385

3.707 3.69o 3.674 3.659 3.646

35 40 45 50 100 ∞

1.306 1.303 1.301 1.299 0.290 1.282

1.690 1.684 1.679 1.676 1.660 1.645

2.030 2.021 2.014 2.009 1.984 1.960

2.438 2.423 2.412 2.403 2.364 2.326

2.724 2.704 2.690 2.678 2.626 2.576

3.340 3.307 3.281 3.261 3.174 3.091

3.591 3.551 3.520 3.496 3.390 3.291

c 2000 by Chapman & Hall/CRC 

6.24

TRIANGULAR DISTRIBUTION

6.24.1

Properties  0     4(x − a)/(b − a)2 pdf f (x) =  4(b − x)/(b − a)2    0

x≤a a < x ≤ (a + b)/2 (a + b)/2 < x < b x≥b

a
a+b 2 (b − a)2 σ2 = 24 β1 = 0 µ=

β2 = 12/5

mgf m(t) = char function 6.24.2

4(eat/2 − ebt/2 )2 (b − a)2 t2

φ(t) = −

4(eait/2 − ebit/2 )2 (b − a)2 t2

Probability density function

The probability density function is symmetric about the mean and consists of two line segments.

Figure 6.25: Probability density function for a triangular random variable.

c 2000 by Chapman & Hall/CRC 

6.25

UNIFORM DISTRIBUTION

6.25.1

Properties 1 , a ≤ x ≤ b, a < b ∈ R b−a a+b µ= 2 (b − a)2 σ2 = 12 β1 = 0

pdf f (x) = mean variance skewness kurtosis

β2 = 9/5

mgf m(t) = char function 6.25.2

φ(t) =

ebt − eat (b − a)t ebit − eait (b − a)it

Probability density function

The probability density function is a horizontal line segment between a and b at 1/(b − a).

Figure 6.26: Probability density functions for a uniform random variable. 6.25.3

Related distributions

(1) The random variable X has a standard uniform distribution if a = 0 and b = 1. (2) If X is a uniform random variable with parameters a = 0 and b = 1, the random variable Y = −(ln X)/λ has an exponential distribution with parameter λ.

c 2000 by Chapman & Hall/CRC 

(3) Let X1 and X2 be independent uniform random variables with parameters a = 0 and b = 1. The random variable Y = (X1 + X2 )/2 has a triangular distribution with parameters 0 and 1. (4) If X is a uniform random variable with parameters a = −π/2 and b = π/2, then the random variable Y = tan X has a Cauchy distribution with parameters a = 0 and b = 1. 6.26

WEIBULL DISTRIBUTION

6.26.1

Properties α α−1 −(x/β)α x e βα   1 µ = βΓ 1 + α      2 1 σ2 = β 2 Γ 1 + − Γ2 1 + α α         1 1 2Γ3 1 + α − 3Γ 1 + α Γ 1 + α2 + Γ 1 + α3 β1 =     3/2 Γ 1 + α2 − Γ2 1 + α1

pdf f (x) = mean variance skewness

kurtosis β2 =             −3Γ4 1 + α1 + 6Γ2 1 + α1 Γ 1 + α2 − 4Γ 1 + α1 Γ 1 + α3 + Γ 1 + α4     2 Γ 1 + α2 − Γ2 1 + α1 mgf m(t) = does not exist char function

φ(t) = does not exist

where Γ(x) is the gamma function (see page 515). 6.26.2

Probability density function

The probability density function is skewed to the right. For fixed β the tail becomes lighter and the distribution becomes more bell–shaped as α increases. 6.26.3

Related distributions

Suppose X is a Weibull random variable with parameters α and β. (1) The random variable X has a standard Weibull distribution if β = 1. (2) If α = 1 then X has an exponential distribution with parameter λ = 1/β. (3) The random variable Y = X α has an exponential distribution with parameter λ = β. √ (4) If α = 2 then X has a Rayleigh distribution with parameter σ = β/ 2. (5) The random variable Y = −α ln(X/β) has a (standard) extreme–value distribution with parameters α = 0 and β = 1. c 2000 by Chapman & Hall/CRC 

Figure 6.27: Probability density functions for a Weibull random variable. 6.27

RELATIONSHIPS AMONG DISTRIBUTIONS

Figure 6.27 presents some of the relationships among common univariate distributions. The first line of each box is the name of the distribution and the second line lists the parameters that characterize the distribution. The random variable X is used to represent each distribution. The three types of relationships presented in the figure are transformations (independent random variables are assumed) and special cases (both indicated with a solid arrow), and limiting distributions (indicated with a dashed arrow). 6.27.1

Other relationships among distributions

(1) If X1 has a standard normal distribution, X2 has a chi–square distribution with ν degrees of freedom, and X1 and X2 are independent, then the random variable X1 Y = (6.13) X2 /ν has a t distribution with ν degrees of freedom. (2) Let X1 , X2 , . . . , Xn be independent normal random variables with parameters µ and σ, and define 1 Xi n i=1 n

X=

1 (Xi − X)2 . n i=1 n

and S 2 =

(6.14)

(a) The random variable Y = nS 2 /σ 2 has a chi–square distribution with n − 1 degrees of freedom.

c 2000 by Chapman & Hall/CRC 

Figure 6.28: Relationships among distributions (see page 161).

c 2000 by Chapman & Hall/CRC 

(b) The random variable W =

X −µ √ S/ n − 1

(6.15)

has a t distribution with n − 1 degrees of freedom. (3) Let X1 , X2 , . . . , Xn be independent normal random variables with parameters µ and σ, and define 1 Xi n i=1 n

X=

1  (Xi − X)2 . n − 1 i=1 n

and S 2 =

(6.16)

The random variable Y =

X −µ √ S/ n

(6.17)

has a t distribution with n − 1 degrees of freedom. (4) Let X1 , X2 , . . . , Xn1 be independent normal random variables with parameters µ1 and σ, and Y1 , Y2 , . . . , Yn2 be independent normal random variables with parameters µ2 and σ. Define X=

n1 1  Xi n1 i=1

n2 1  Y = Yi n2 i=1

S12 = S22

n1 1  (Xi − X)2 n1 i=1

n2 1  = (Yi − Y )2 n2 i=1

(6.18)

(a) The random variable Y = (n1 S12 + n2 S22 )/σ 2 has a chi–square distribution with n1 + n2 − 2 degrees of freedom. (b) The random variable (X − Y ) − (µ1 − µ2 )  W = n1 S12 +n2 S22 1 1 + n1 n2 n1 +n2 −2

(6.19)

has a t distribution with n1 + n2 − 2 degrees of freedom. (5) Let X1 , X2 , . . . , Xn1 be independent normal random variables with parameters µ1 and σ1 , and Y1 , Y2 , . . . , Yn2 be independent normal random variables with parameters µ2 and σ2 . Define X=

n1 1  Xi n1 i=1

n2 1  Y = Yi n2 i=1

The random variable Y =

c 2000 by Chapman & Hall/CRC 

n1 S12 (n1 − 1)σ12

S12 = S22 *

n1 1  (Xi − X)2 n1 i=1

n2 1  = (Yi − Y )2 n2 i=1

n2 S22 (n2 − 1)σ22

(6.20)

(6.21)

has an F distribution with n1 and n2 degrees of freedom. (6) Let X1 be a normal random variable with parameters µ = λ and σ = 1, X2 a chi–square random variable with parameter  ν, and X1 and X2 be independent. The random variable Y = X1 / X2 /ν has a noncentral t distribution with parameters ν and λ. (7) Let X be a continuous random variable with cumulative distribution function F (x). (a) The random variable Y = F (X) has a (standard) uniform distribution with parameters a = 0 and b = 1. (b) The random variable Y = − ln[1 − F (X)] has a (standard) exponential distribution with parameter λ = 1. Let X be a continuous random variable with probability density function f (x). The random variable Y = |X| has probability density function g(y) given by ( f (y) + f (−y) if y > 0 g(y) = (6.22) 0 elsewhere If X has a standard normal distribution (µ = 0, σ = 1) then g(y) = 2f (y).

c 2000 by Chapman & Hall/CRC 

CHAPTER 7

Standard Normal Distribution Contents 7.1 7.2 7.3 7.4

7.5 7.6

7.7 7.8

Density function and related functions Critical values Tolerance factors for normal distributions 7.3.1 Tables of tolerance intervals Operating characteristic curves 7.4.1 One-sample Z test 7.4.2 Two-sample Z test Multivariate normal distribution Distribution of the correlation coefficient 7.6.1 Normal approximation 7.6.2 Zero coefficient for bivariate normal Circular normal probabilities Circular error probabilities

7.1 THE PROBABILITY DENSITY FUNCTION AND RELATED FUNCTIONS Let Z be a standard normal random variable (µ = 0, σ = 1). The probability density function is given by 2 1 f (z) = √ e−z /2 . 2π

The following tables contain values for: (1) f (z)



z

2 1 √ e−t /2 dt. 2π −∞ = the cumulative distribution function

(2) F (z) = Prob [Z ≤ z] =

c 2000 by Chapman & Hall/CRC 

(7.1)

Figure 7.1: Cumulative distribution function for a standard normal random variable. Note: (1) For all z, f (−z) = f (z). (2) For all z, F (−z) = 1 − F (z) (3) For all z, Prob [|Z| ≤ z] = F (z) − F (−z) (4) For all z, Prob [|Z| ≥ z] = 1 − F (z) + F (−z) (5) The function Φ(z) = F (z) is often used to represent the normal distribution function. (6) f  (x) = − √x2π e−x /2 = −x f (x)   (7) f  (x) = x2 − 1 f (x)   (8) f  (x) = 3x − x3 f (x)   (9) f (4) (x) = x4 − 6x2 + 3 f (x) 2

(10) For large values of x: % % &   & 2 2 e−x /2 1 1 e−x /2 1 √ − < 1 − Φ(x) < √ x x3 x 2π 2π

(7.2)

(11) Order statistics for the normal distribution may be found in section 4.6.8 on page 64.

c 2000 by Chapman & Hall/CRC 

Normal distribution z

f (z)

F (z)

1 − F (z)

z

f (z)

F (z)

1 − F (z)

−4.00 −3.99 −3.98 −3.97 −3.96 −3.95 −3.94 −3.93 −3.92 −3.91

0.0001 0.0001 0.0001 0.0001 0.0002 0.0002 0.0002 0.0002 0.0002 0.0002

0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0001

1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000

−3.50 −3.49 −3.48 −3.47 −3.46 −3.45 −3.44 −3.43 −3.42 −3.41

0.0009 0.0009 0.0009 0.0010 0.0010 0.0010 0.0011 0.0011 0.0011 0.0012

0.0002 0.0002 0.0003 0.0003 0.0003 0.0003 0.0003 0.0003 0.0003 0.0003

0.9998 0.9998 0.9998 0.9997 0.9997 0.9997 0.9997 0.9997 0.9997 0.9997

−3.90 −3.89 −3.88 −3.87 −3.86 −3.85 −3.84 −3.83 −3.82 −3.81

0.0002 0.0002 0.0002 0.0002 0.0002 0.0002 0.0003 0.0003 0.0003 0.0003

0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001

1.0000 1.0000 1.0000 1.0000 0.9999 0.9999 0.9999 0.9999 0.9999 0.9999

−3.40 −3.39 −3.38 −3.37 −3.36 −3.35 −3.34 −3.33 −3.32 −3.31

0.0012 0.0013 0.0013 0.0014 0.0014 0.0015 0.0015 0.0016 0.0016 0.0017

0.0003 0.0003 0.0004 0.0004 0.0004 0.0004 0.0004 0.0004 0.0004 0.0005

0.9997 0.9997 0.9996 0.9996 0.9996 0.9996 0.9996 0.9996 0.9996 0.9995

−3.80 −3.79 −3.78 −3.77 −3.76 −3.75 −3.74 −3.73 −3.72 −3.71

0.0003 0.0003 0.0003 0.0003 0.0003 0.0003 0.0004 0.0004 0.0004 0.0004

0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001

0.9999 0.9999 0.9999 0.9999 0.9999 0.9999 0.9999 0.9999 0.9999 0.9999

−3.30 −3.29 −3.28 −3.27 −3.26 −3.25 −3.24 −3.23 −3.22 −3.21

0.0017 0.0018 0.0018 0.0019 0.0020 0.0020 0.0021 0.0022 0.0022 0.0023

0.0005 0.0005 0.0005 0.0005 0.0006 0.0006 0.0006 0.0006 0.0006 0.0007

0.9995 0.9995 0.9995 0.9995 0.9994 0.9994 0.9994 0.9994 0.9994 0.9993

−3.70 −3.69 −3.68 −3.67 −3.66 −3.65 −3.64 −3.63 −3.62 −3.61

0.0004 0.0004 0.0005 0.0005 0.0005 0.0005 0.0005 0.0006 0.0006 0.0006

0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001

0.9999 0.9999 0.9999 0.9999 0.9999 0.9999 0.9999 0.9999 0.9999 0.9999

−3.20 −3.19 −3.18 −3.17 −3.16 −3.15 −3.14 −3.13 −3.12 −3.11

0.0024 0.0025 0.0025 0.0026 0.0027 0.0028 0.0029 0.0030 0.0031 0.0032

0.0007 0.0007 0.0007 0.0008 0.0008 0.0008 0.0008 0.0009 0.0009 0.0009

0.9993 0.9993 0.9993 0.9992 0.9992 0.9992 0.9992 0.9991 0.9991 0.9991

−3.60 −3.59 −3.58 −3.57 −3.56 −3.55 −3.54 −3.53 −3.52 −3.51

0.0006 0.0006 0.0007 0.0007 0.0007 0.0007 0.0008 0.0008 0.0008 0.0008

0.0002 0.0002 0.0002 0.0002 0.0002 0.0002 0.0002 0.0002 0.0002 0.0002

0.9998 0.9998 0.9998 0.9998 0.9998 0.9998 0.9998 0.9998 0.9998 0.9998

−3.10 −3.09 −3.08 −3.07 −3.06 −3.05 −3.04 −3.03 −3.02 −3.01

0.0033 0.0034 0.0035 0.0036 0.0037 0.0038 0.0039 0.0040 0.0042 0.0043

0.0010 0.0010 0.0010 0.0011 0.0011 0.0011 0.0012 0.0012 0.0013 0.0013

0.9990 0.9990 0.9990 0.9989 0.9989 0.9989 0.9988 0.9988 0.9987 0.9987

−3.50

0.0009

0.0002

0.9998

−3.00

0.0044

0.0014

0.9987

c 2000 by Chapman & Hall/CRC 

Normal distribution z

f (z)

F (z)

1 − F (z)

z

f (z)

F (z)

1 − F (z)

−3.00 −2.99 −2.98 −2.97 −2.96 −2.95 −2.94 −2.93 −2.92 −2.91

0.0044 0.0046 0.0047 0.0049 0.0050 0.0051 0.0053 0.0054 0.0056 0.0058

0.0014 0.0014 0.0014 0.0015 0.0015 0.0016 0.0016 0.0017 0.0018 0.0018

0.9987 0.9986 0.9986 0.9985 0.9985 0.9984 0.9984 0.9983 0.9982 0.9982

−2.50 −2.49 −2.48 −2.47 −2.46 −2.45 −2.44 −2.43 −2.42 −2.41

0.0175 0.0180 0.0184 0.0189 0.0194 0.0198 0.0203 0.0208 0.0213 0.0219

0.0062 0.0064 0.0066 0.0068 0.0069 0.0071 0.0073 0.0076 0.0078 0.0080

0.9938 0.9936 0.9934 0.9932 0.9930 0.9929 0.9927 0.9925 0.9922 0.9920

−2.90 −2.89 −2.88 −2.87 −2.86 −2.85 −2.84 −2.83 −2.82 −2.81

0.0060 0.0061 0.0063 0.0065 0.0067 0.0069 0.0071 0.0073 0.0075 0.0077

0.0019 0.0019 0.0020 0.0021 0.0021 0.0022 0.0023 0.0023 0.0024 0.0025

0.9981 0.9981 0.9980 0.9980 0.9979 0.9978 0.9977 0.9977 0.9976 0.9975

−2.40 −2.39 −2.38 −2.37 −2.36 −2.35 −2.34 −2.33 −2.32 −2.31

0.0224 0.0229 0.0235 0.0241 0.0246 0.0252 0.0258 0.0264 0.0271 0.0277

0.0082 0.0084 0.0087 0.0089 0.0091 0.0094 0.0096 0.0099 0.0102 0.0104

0.9918 0.9916 0.9913 0.9911 0.9909 0.9906 0.9904 0.9901 0.9898 0.9896

−2.80 −2.79 −2.78 −2.77 −2.76 −2.75 −2.74 −2.73 −2.72 −2.71

0.0079 0.0081 0.0084 0.0086 0.0089 0.0091 0.0094 0.0096 0.0099 0.0101

0.0026 0.0026 0.0027 0.0028 0.0029 0.0030 0.0031 0.0032 0.0033 0.0034

0.9974 0.9974 0.9973 0.9972 0.9971 0.9970 0.9969 0.9968 0.9967 0.9966

−2.30 −2.29 −2.28 −2.27 −2.26 −2.25 −2.24 −2.23 −2.22 −2.21

0.0283 0.0290 0.0296 0.0303 0.0310 0.0317 0.0325 0.0332 0.0339 0.0347

0.0107 0.0110 0.0113 0.0116 0.0119 0.0122 0.0126 0.0129 0.0132 0.0135

0.9893 0.9890 0.9887 0.9884 0.9881 0.9878 0.9875 0.9871 0.9868 0.9865

−2.70 −2.69 −2.68 −2.67 −2.66 −2.65 −2.64 −2.63 −2.62 −2.61

0.0104 0.0107 0.0110 0.0113 0.0116 0.0119 0.0122 0.0126 0.0129 0.0132

0.0035 0.0036 0.0037 0.0038 0.0039 0.0040 0.0042 0.0043 0.0044 0.0045

0.9965 0.9964 0.9963 0.9962 0.9961 0.9960 0.9959 0.9957 0.9956 0.9955

−2.20 −2.19 −2.18 −2.17 −2.16 −2.15 −2.14 −2.13 −2.12 −2.11

0.0355 0.0363 0.0371 0.0379 0.0387 0.0396 0.0404 0.0413 0.0422 0.0431

0.0139 0.0143 0.0146 0.0150 0.0154 0.0158 0.0162 0.0166 0.0170 0.0174

0.9861 0.9857 0.9854 0.9850 0.9846 0.9842 0.9838 0.9834 0.9830 0.9826

−2.60 −2.59 −2.58 −2.57 −2.56 −2.55 −2.54 −2.53 −2.52 −2.51

0.0136 0.0139 0.0143 0.0147 0.0151 0.0155 0.0158 0.0163 0.0167 0.0171

0.0047 0.0048 0.0049 0.0051 0.0052 0.0054 0.0055 0.0057 0.0059 0.0060

0.9953 0.9952 0.9951 0.9949 0.9948 0.9946 0.9945 0.9943 0.9941 0.9940

−2.10 −2.09 −2.08 −2.07 −2.06 −2.05 −2.04 −2.03 −2.02 −2.01

0.0440 0.0449 0.0459 0.0468 0.0478 0.0488 0.0498 0.0508 0.0519 0.0529

0.0179 0.0183 0.0188 0.0192 0.0197 0.0202 0.0207 0.0212 0.0217 0.0222

0.9821 0.9817 0.9812 0.9808 0.9803 0.9798 0.9793 0.9788 0.9783 0.9778

−2.50

0.0175

0.0062

0.9938

−2.00

0.0540

0.0227

0.9772

c 2000 by Chapman & Hall/CRC 

Normal distribution z

f (z)

F (z)

1 − F (z)

z

f (z)

F (z)

1 − F (z)

−2.00 −1.99 −1.98 −1.97 −1.96 −1.95 −1.94 −1.93 −1.92 −1.91

0.0540 0.0551 0.0562 0.0573 0.0584 0.0596 0.0608 0.0619 0.0632 0.0644

0.0227 0.0233 0.0238 0.0244 0.0250 0.0256 0.0262 0.0268 0.0274 0.0281

0.9772 0.9767 0.9761 0.9756 0.9750 0.9744 0.9738 0.9732 0.9726 0.9719

−1.50 −1.49 −1.48 −1.47 −1.46 −1.45 −1.44 −1.43 −1.42 −1.41

0.1295 0.1315 0.1334 0.1354 0.1374 0.1394 0.1415 0.1435 0.1456 0.1476

0.0668 0.0681 0.0694 0.0708 0.0722 0.0735 0.0749 0.0764 0.0778 0.0793

0.9332 0.9319 0.9306 0.9292 0.9278 0.9265 0.9251 0.9236 0.9222 0.9207

−1.90 −1.89 −1.88 −1.87 −1.86 −1.85 −1.84 −1.83 −1.82 −1.81

0.0656 0.0669 0.0681 0.0694 0.0707 0.0721 0.0734 0.0748 0.0761 0.0775

0.0287 0.0294 0.0301 0.0307 0.0314 0.0322 0.0329 0.0336 0.0344 0.0352

0.9713 0.9706 0.9699 0.9693 0.9686 0.9678 0.9671 0.9664 0.9656 0.9648

−1.40 −1.39 −1.38 −1.37 −1.36 −1.35 −1.34 −1.33 −1.32 −1.31

0.1497 0.1518 0.1540 0.1561 0.1582 0.1604 0.1626 0.1647 0.1669 0.1691

0.0808 0.0823 0.0838 0.0853 0.0869 0.0885 0.0901 0.0918 0.0934 0.0951

0.9192 0.9177 0.9162 0.9147 0.9131 0.9115 0.9099 0.9082 0.9066 0.9049

−1.80 −1.79 −1.78 −1.77 −1.76 −1.75 −1.74 −1.73 −1.72 −1.71

0.0790 0.0804 0.0818 0.0833 0.0848 0.0863 0.0878 0.0893 0.0909 0.0925

0.0359 0.0367 0.0375 0.0384 0.0392 0.0401 0.0409 0.0418 0.0427 0.0436

0.9641 0.9633 0.9625 0.9616 0.9608 0.9599 0.9591 0.9582 0.9573 0.9564

−1.30 −1.29 −1.28 −1.27 −1.26 −1.25 −1.24 −1.23 −1.22 −1.21

0.1714 0.1736 0.1759 0.1781 0.1804 0.1827 0.1849 0.1872 0.1895 0.1919

0.0968 0.0985 0.1003 0.1020 0.1038 0.1056 0.1075 0.1094 0.1112 0.1131

0.9032 0.9015 0.8997 0.8980 0.8962 0.8943 0.8925 0.8907 0.8888 0.8869

−1.70 −1.69 −1.68 −1.67 −1.66 −1.65 −1.64 −1.63 −1.62 −1.61

0.0940 0.0957 0.0973 0.0989 0.1006 0.1023 0.1040 0.1057 0.1074 0.1091

0.0446 0.0455 0.0465 0.0475 0.0485 0.0495 0.0505 0.0515 0.0526 0.0537

0.9554 0.9545 0.9535 0.9525 0.9515 0.9505 0.9495 0.9485 0.9474 0.9463

−1.20 −1.19 −1.18 −1.17 −1.16 −1.15 −1.14 −1.13 −1.12 −1.11

0.1942 0.1965 0.1989 0.2012 0.2036 0.2059 0.2083 0.2107 0.2131 0.2155

0.1151 0.1170 0.1190 0.1210 0.1230 0.1251 0.1271 0.1292 0.1314 0.1335

0.8849 0.8830 0.8810 0.8790 0.8770 0.8749 0.8729 0.8708 0.8686 0.8665

−1.60 −1.59 −1.58 −1.57 −1.56 −1.55 −1.54 −1.53 −1.52 −1.51

0.1109 0.1127 0.1145 0.1163 0.1182 0.1200 0.1219 0.1238 0.1257 0.1276

0.0548 0.0559 0.0570 0.0582 0.0594 0.0606 0.0618 0.0630 0.0643 0.0655

0.9452 0.9441 0.9429 0.9418 0.9406 0.9394 0.9382 0.9370 0.9357 0.9345

−1.10 −1.09 −1.08 −1.07 −1.06 −1.05 −1.04 −1.03 −1.02 −1.01

0.2178 0.2203 0.2226 0.2251 0.2275 0.2299 0.2323 0.2347 0.2371 0.2396

0.1357 0.1379 0.1401 0.1423 0.1446 0.1469 0.1492 0.1515 0.1539 0.1563

0.8643 0.8621 0.8599 0.8577 0.8554 0.8531 0.8508 0.8485 0.8461 0.8438

−1.50

0.1295

0.0668

0.9332

−1.00

0.2420

0.1587

0.8413

c 2000 by Chapman & Hall/CRC 

Normal distribution z

f (z)

F (z)

1 − F (z)

z

f (z)

F (z)

1 − F (z)

−1.00 −0.99 −0.98 −0.97 −0.96 −0.95 −0.94 −0.93 −0.92 −0.91

0.2420 0.2444 0.2468 0.2492 0.2516 0.2541 0.2565 0.2589 0.2613 0.2637

0.1587 0.1611 0.1635 0.1660 0.1685 0.1711 0.1736 0.1762 0.1788 0.1814

0.8413 0.8389 0.8365 0.8340 0.8315 0.8289 0.8264 0.8238 0.8212 0.8186

−0.50 −0.49 −0.48 −0.47 −0.46 −0.45 −0.44 −0.43 −0.42 −0.41

0.3521 0.3538 0.3555 0.3572 0.3589 0.3605 0.3621 0.3637 0.3653 0.3668

0.3085 0.3121 0.3156 0.3192 0.3228 0.3264 0.3300 0.3336 0.3372 0.3409

0.6915 0.6879 0.6844 0.6808 0.6772 0.6736 0.6700 0.6664 0.6628 0.6591

−0.90 −0.89 −0.88 −0.87 −0.86 −0.85 −0.84 −0.83 −0.82 −0.81

0.2661 0.2685 0.2709 0.2732 0.2756 0.2780 0.2803 0.2827 0.2850 0.2874

0.1841 0.1867 0.1894 0.1921 0.1949 0.1977 0.2004 0.2033 0.2061 0.2090

0.8159 0.8133 0.8106 0.8078 0.8051 0.8023 0.7995 0.7967 0.7939 0.7910

−0.40 −0.39 −0.38 −0.37 −0.36 −0.35 −0.34 −0.33 −0.32 −0.31

0.3683 0.3697 0.3711 0.3725 0.3739 0.3752 0.3765 0.3778 0.3790 0.3802

0.3446 0.3483 0.3520 0.3557 0.3594 0.3632 0.3669 0.3707 0.3745 0.3783

0.6554 0.6517 0.6480 0.6443 0.6406 0.6368 0.6331 0.6293 0.6255 0.6217

−0.80 −0.79 −0.78 −0.77 −0.76 −0.75 −0.74 −0.73 −0.72 −0.71

0.2897 0.2920 0.2943 0.2966 0.2989 0.3011 0.3034 0.3056 0.3079 0.3101

0.2119 0.2148 0.2177 0.2207 0.2236 0.2266 0.2296 0.2327 0.2358 0.2389

0.7881 0.7852 0.7823 0.7793 0.7764 0.7734 0.7703 0.7673 0.7642 0.7611

−0.30 −0.29 −0.28 −0.27 −0.26 −0.25 −0.24 −0.23 −0.22 −0.21

0.3814 0.3825 0.3836 0.3847 0.3857 0.3867 0.3876 0.3885 0.3894 0.3902

0.3821 0.3859 0.3897 0.3936 0.3974 0.4013 0.4052 0.4091 0.4129 0.4168

0.6179 0.6141 0.6103 0.6064 0.6026 0.5987 0.5948 0.5909 0.5871 0.5832

−0.70 −0.69 −0.68 −0.67 −0.66 −0.65 −0.64 −0.63 −0.62 −0.61

0.3123 0.3144 0.3166 0.3187 0.3209 0.3230 0.3251 0.3271 0.3292 0.3312

0.2420 0.2451 0.2482 0.2514 0.2546 0.2579 0.2611 0.2643 0.2676 0.2709

0.7580 0.7549 0.7518 0.7486 0.7454 0.7421 0.7389 0.7357 0.7324 0.7291

−0.20 −0.19 −0.18 −0.17 −0.16 −0.15 −0.14 −0.13 −0.12 −0.11

0.3910 0.3918 0.3925 0.3932 0.3939 0.3945 0.3951 0.3956 0.3961 0.3965

0.4207 0.4247 0.4286 0.4325 0.4364 0.4404 0.4443 0.4483 0.4522 0.4562

0.5793 0.5754 0.5714 0.5675 0.5636 0.5596 0.5557 0.5517 0.5478 0.5438

−0.60 −0.59 −0.58 −0.57 −0.56 −0.55 −0.54 −0.53 −0.52 −0.51

0.3332 0.3352 0.3372 0.3391 0.3411 0.3429 0.3448 0.3467 0.3485 0.3503

0.2742 0.2776 0.2810 0.2843 0.2877 0.2912 0.2946 0.2981 0.3015 0.3050

0.7258 0.7224 0.7190 0.7157 0.7123 0.7088 0.7054 0.7019 0.6985 0.6950

−0.10 −0.09 −0.08 −0.07 −0.06 −0.05 −0.04 −0.03 −0.02 −0.01

0.3970 0.3973 0.3977 0.3980 0.3982 0.3984 0.3986 0.3988 0.3989 0.3989

0.4602 0.4641 0.4681 0.4721 0.4761 0.4801 0.4840 0.4880 0.4920 0.4960

0.5398 0.5359 0.5319 0.5279 0.5239 0.5199 0.5160 0.5120 0.5080 0.5040

−0.50

0.3521

0.3085

0.6915

0.00

0.3989

0.5000

0.5000

c 2000 by Chapman & Hall/CRC 

Normal distribution z

f (z)

F (z)

1 − F (z)

z

f (z)

F (z)

1 − F (z)

0.00 0.01 0.02 0.03 0.04 0.05 0.06 0.07 0.08 0.09

0.3989 0.3989 0.3989 0.3988 0.3986 0.3984 0.3982 0.3980 0.3977 0.3973

0.5000 0.5040 0.5080 0.5120 0.5160 0.5199 0.5239 0.5279 0.5319 0.5359

0.5000 0.4960 0.4920 0.4880 0.4840 0.4801 0.4761 0.4721 0.4681 0.4641

0.50 0.51 0.52 0.53 0.54 0.55 0.56 0.57 0.58 0.59

0.3521 0.3503 0.3485 0.3467 0.3448 0.3429 0.3411 0.3391 0.3372 0.3352

0.6915 0.6950 0.6985 0.7019 0.7054 0.7088 0.7123 0.7157 0.7190 0.7224

0.3085 0.3050 0.3015 0.2981 0.2946 0.2912 0.2877 0.2843 0.2810 0.2776

0.10 0.11 0.12 0.13 0.14 0.15 0.16 0.17 0.18 0.19

0.3970 0.3965 0.3961 0.3956 0.3951 0.3945 0.3939 0.3932 0.3925 0.3918

0.5398 0.5438 0.5478 0.5517 0.5557 0.5596 0.5636 0.5675 0.5714 0.5754

0.4602 0.4562 0.4522 0.4483 0.4443 0.4404 0.4364 0.4325 0.4286 0.4247

0.60 0.61 0.62 0.63 0.64 0.65 0.66 0.67 0.68 0.69

0.3332 0.3312 0.3292 0.3271 0.3251 0.3230 0.3209 0.3187 0.3166 0.3144

0.7258 0.7291 0.7324 0.7357 0.7389 0.7421 0.7454 0.7486 0.7518 0.7549

0.2742 0.2709 0.2676 0.2643 0.2611 0.2579 0.2546 0.2514 0.2482 0.2451

0.20 0.21 0.22 0.23 0.24 0.25 0.26 0.27 0.28 0.29

0.3910 0.3902 0.3894 0.3885 0.3876 0.3867 0.3857 0.3847 0.3836 0.3825

0.5793 0.5832 0.5871 0.5909 0.5948 0.5987 0.6026 0.6064 0.6103 0.6141

0.4207 0.4168 0.4129 0.4091 0.4052 0.4013 0.3974 0.3936 0.3897 0.3859

0.70 0.71 0.72 0.73 0.74 0.75 0.76 0.77 0.78 0.79

0.3123 0.3101 0.3079 0.3056 0.3034 0.3011 0.2989 0.2966 0.2943 0.2920

0.7580 0.7611 0.7642 0.7673 0.7703 0.7734 0.7764 0.7793 0.7823 0.7852

0.2420 0.2389 0.2358 0.2327 0.2296 0.2266 0.2236 0.2207 0.2177 0.2148

0.30 0.31 0.32 0.33 0.34 0.35 0.36 0.37 0.38 0.39

0.3814 0.3802 0.3790 0.3778 0.3765 0.3752 0.3739 0.3725 0.3711 0.3697

0.6179 0.6217 0.6255 0.6293 0.6331 0.6368 0.6406 0.6443 0.6480 0.6517

0.3821 0.3783 0.3745 0.3707 0.3669 0.3632 0.3594 0.3557 0.3520 0.3483

0.80 0.81 0.82 0.83 0.84 0.85 0.86 0.87 0.88 0.89

0.2897 0.2874 0.2850 0.2827 0.2803 0.2780 0.2756 0.2732 0.2709 0.2685

0.7881 0.7910 0.7939 0.7967 0.7995 0.8023 0.8051 0.8078 0.8106 0.8133

0.2119 0.2090 0.2061 0.2033 0.2004 0.1977 0.1949 0.1921 0.1894 0.1867

0.40 0.41 0.42 0.43 0.44 0.45 0.46 0.47 0.48 0.49

0.3683 0.3668 0.3653 0.3637 0.3621 0.3605 0.3589 0.3572 0.3555 0.3538

0.6554 0.6591 0.6628 0.6664 0.6700 0.6736 0.6772 0.6808 0.6844 0.6879

0.3446 0.3409 0.3372 0.3336 0.3300 0.3264 0.3228 0.3192 0.3156 0.3121

0.90 0.91 0.92 0.93 0.94 0.95 0.96 0.97 0.98 0.99

0.2661 0.2637 0.2613 0.2589 0.2565 0.2541 0.2516 0.2492 0.2468 0.2444

0.8159 0.8186 0.8212 0.8238 0.8264 0.8289 0.8315 0.8340 0.8365 0.8389

0.1841 0.1814 0.1788 0.1762 0.1736 0.1711 0.1685 0.1660 0.1635 0.1611

0.50

0.3521

0.6915

0.3085

1.00

0.2420

0.8413

0.1587

c 2000 by Chapman & Hall/CRC 

Normal distribution z

f (z)

F (z)

1 − F (z)

z

f (z)

F (z)

1 − F (z)

1.00 1.01 1.02 1.03 1.04 1.05 1.06 1.07 1.08 1.09

0.2420 0.2396 0.2371 0.2347 0.2323 0.2299 0.2275 0.2251 0.2226 0.2203

0.8413 0.8438 0.8461 0.8485 0.8508 0.8531 0.8554 0.8577 0.8599 0.8621

0.1587 0.1563 0.1539 0.1515 0.1492 0.1469 0.1446 0.1423 0.1401 0.1379

1.50 1.51 1.52 1.53 1.54 1.55 1.56 1.57 1.58 1.59

0.1295 0.1276 0.1257 0.1238 0.1219 0.1200 0.1182 0.1163 0.1145 0.1127

0.9332 0.9345 0.9357 0.9370 0.9382 0.9394 0.9406 0.9418 0.9429 0.9441

0.0668 0.0655 0.0643 0.0630 0.0618 0.0606 0.0594 0.0582 0.0570 0.0559

1.10 1.11 1.12 1.13 1.14 1.15 1.16 1.17 1.18 1.19

0.2178 0.2155 0.2131 0.2107 0.2083 0.2059 0.2036 0.2012 0.1989 0.1965

0.8643 0.8665 0.8686 0.8708 0.8729 0.8749 0.8770 0.8790 0.8810 0.8830

0.1357 0.1335 0.1314 0.1292 0.1271 0.1251 0.1230 0.1210 0.1190 0.1170

1.60 1.61 1.62 1.63 1.64 1.65 1.66 1.67 1.68 1.69

0.1109 0.1091 0.1074 0.1057 0.1040 0.1023 0.1006 0.0989 0.0973 0.0957

0.9452 0.9463 0.9474 0.9485 0.9495 0.9505 0.9515 0.9525 0.9535 0.9545

0.0548 0.0537 0.0526 0.0515 0.0505 0.0495 0.0485 0.0475 0.0465 0.0455

1.20 1.21 1.22 1.23 1.24 1.25 1.26 1.27 1.28 1.29

0.1942 0.1919 0.1895 0.1872 0.1849 0.1827 0.1804 0.1781 0.1759 0.1736

0.8849 0.8869 0.8888 0.8907 0.8925 0.8943 0.8962 0.8980 0.8997 0.9015

0.1151 0.1131 0.1112 0.1094 0.1075 0.1056 0.1038 0.1020 0.1003 0.0985

1.70 1.71 1.72 1.73 1.74 1.75 1.76 1.77 1.78 1.79

0.0940 0.0925 0.0909 0.0893 0.0878 0.0863 0.0848 0.0833 0.0818 0.0804

0.9554 0.9564 0.9573 0.9582 0.9591 0.9599 0.9608 0.9616 0.9625 0.9633

0.0446 0.0436 0.0427 0.0418 0.0409 0.0401 0.0392 0.0384 0.0375 0.0367

1.30 1.31 1.32 1.33 1.34 1.35 1.36 1.37 1.38 1.39

0.1714 0.1691 0.1669 0.1647 0.1626 0.1604 0.1582 0.1561 0.1540 0.1518

0.9032 0.9049 0.9066 0.9082 0.9099 0.9115 0.9131 0.9147 0.9162 0.9177

0.0968 0.0951 0.0934 0.0918 0.0901 0.0885 0.0869 0.0853 0.0838 0.0823

1.80 1.81 1.82 1.83 1.84 1.85 1.86 1.87 1.88 1.89

0.0790 0.0775 0.0761 0.0748 0.0734 0.0721 0.0707 0.0694 0.0681 0.0669

0.9641 0.9648 0.9656 0.9664 0.9671 0.9678 0.9686 0.9693 0.9699 0.9706

0.0359 0.0352 0.0344 0.0336 0.0329 0.0322 0.0314 0.0307 0.0301 0.0294

1.40 1.41 1.42 1.43 1.44 1.45 1.46 1.47 1.48 1.49

0.1497 0.1476 0.1456 0.1435 0.1415 0.1394 0.1374 0.1354 0.1334 0.1315

0.9192 0.9207 0.9222 0.9236 0.9251 0.9265 0.9278 0.9292 0.9306 0.9319

0.0808 0.0793 0.0778 0.0764 0.0749 0.0735 0.0722 0.0708 0.0694 0.0681

1.90 1.91 1.92 1.93 1.94 1.95 1.96 1.97 1.98 1.99

0.0656 0.0644 0.0632 0.0619 0.0608 0.0596 0.0584 0.0573 0.0562 0.0551

0.9713 0.9719 0.9726 0.9732 0.9738 0.9744 0.9750 0.9756 0.9761 0.9767

0.0287 0.0281 0.0274 0.0268 0.0262 0.0256 0.0250 0.0244 0.0238 0.0233

1.50

0.1295

0.9332

0.0668

2.00

0.0540

0.9772

0.0227

c 2000 by Chapman & Hall/CRC 

Normal distribution z

f (z)

F (z)

1 − F (z)

z

f (z)

F (z)

1 − F (z)

2.00 2.01 2.02 2.03 2.04 2.05 2.06 2.07 2.08 2.09

0.0540 0.0529 0.0519 0.0508 0.0498 0.0488 0.0478 0.0468 0.0459 0.0449

0.9772 0.9778 0.9783 0.9788 0.9793 0.9798 0.9803 0.9808 0.9812 0.9817

0.0227 0.0222 0.0217 0.0212 0.0207 0.0202 0.0197 0.0192 0.0188 0.0183

2.50 2.51 2.52 2.53 2.54 2.55 2.56 2.57 2.58 2.59

0.0175 0.0171 0.0167 0.0163 0.0158 0.0155 0.0151 0.0147 0.0143 0.0139

0.9938 0.9940 0.9941 0.9943 0.9945 0.9946 0.9948 0.9949 0.9951 0.9952

0.0062 0.0060 0.0059 0.0057 0.0055 0.0054 0.0052 0.0051 0.0049 0.0048

2.10 2.11 2.12 2.13 2.14 2.15 2.16 2.17 2.18 2.19

0.0440 0.0431 0.0422 0.0413 0.0404 0.0396 0.0387 0.0379 0.0371 0.0363

0.9821 0.9826 0.9830 0.9834 0.9838 0.9842 0.9846 0.9850 0.9854 0.9857

0.0179 0.0174 0.0170 0.0166 0.0162 0.0158 0.0154 0.0150 0.0146 0.0143

2.60 2.61 2.62 2.63 2.64 2.65 2.66 2.67 2.68 2.69

0.0136 0.0132 0.0129 0.0126 0.0122 0.0119 0.0116 0.0113 0.0110 0.0107

0.9953 0.9955 0.9956 0.9957 0.9959 0.9960 0.9961 0.9962 0.9963 0.9964

0.0047 0.0045 0.0044 0.0043 0.0042 0.0040 0.0039 0.0038 0.0037 0.0036

2.20 2.21 2.22 2.23 2.24 2.25 2.26 2.27 2.28 2.29

0.0355 0.0347 0.0339 0.0332 0.0325 0.0317 0.0310 0.0303 0.0296 0.0290

0.9861 0.9865 0.9868 0.9871 0.9875 0.9878 0.9881 0.9884 0.9887 0.9890

0.0139 0.0135 0.0132 0.0129 0.0126 0.0122 0.0119 0.0116 0.0113 0.0110

2.70 2.71 2.72 2.73 2.74 2.75 2.76 2.77 2.78 2.79

0.0104 0.0101 0.0099 0.0096 0.0094 0.0091 0.0089 0.0086 0.0084 0.0081

0.9965 0.9966 0.9967 0.9968 0.9969 0.9970 0.9971 0.9972 0.9973 0.9974

0.0035 0.0034 0.0033 0.0032 0.0031 0.0030 0.0029 0.0028 0.0027 0.0026

2.30 2.31 2.32 2.33 2.34 2.35 2.36 2.37 2.38 2.39

0.0283 0.0277 0.0271 0.0264 0.0258 0.0252 0.0246 0.0241 0.0235 0.0229

0.9893 0.9896 0.9898 0.9901 0.9904 0.9906 0.9909 0.9911 0.9913 0.9916

0.0107 0.0104 0.0102 0.0099 0.0096 0.0094 0.0091 0.0089 0.0087 0.0084

2.80 2.81 2.82 2.83 2.84 2.85 2.86 2.87 2.88 2.89

0.0079 0.0077 0.0075 0.0073 0.0071 0.0069 0.0067 0.0065 0.0063 0.0061

0.9974 0.9975 0.9976 0.9977 0.9977 0.9978 0.9979 0.9980 0.9980 0.9981

0.0026 0.0025 0.0024 0.0023 0.0023 0.0022 0.0021 0.0021 0.0020 0.0019

2.40 2.41 2.42 2.43 2.44 2.45 2.46 2.47 2.48 2.49

0.0224 0.0219 0.0213 0.0208 0.0203 0.0198 0.0194 0.0189 0.0184 0.0180

0.9918 0.9920 0.9922 0.9925 0.9927 0.9929 0.9930 0.9932 0.9934 0.9936

0.0082 0.0080 0.0078 0.0076 0.0073 0.0071 0.0069 0.0068 0.0066 0.0064

2.90 2.91 2.92 2.93 2.94 2.95 2.96 2.97 2.98 2.99

0.0060 0.0058 0.0056 0.0054 0.0053 0.0051 0.0050 0.0049 0.0047 0.0046

0.9981 0.9982 0.9982 0.9983 0.9984 0.9984 0.9985 0.9985 0.9986 0.9986

0.0019 0.0018 0.0018 0.0017 0.0016 0.0016 0.0015 0.0015 0.0014 0.0014

2.50

0.0175

0.9938

0.0062

3.00

0.0044

0.9987

0.0014

c 2000 by Chapman & Hall/CRC 

Normal distribution z

f (z)

F (z)

1 − F (z)

z

f (z)

F (z)

1 − F (z)

3.00 3.01 3.02 3.03 3.04 3.05 3.06 3.07 3.08 3.09

0.0044 0.0043 0.0042 0.0040 0.0039 0.0038 0.0037 0.0036 0.0035 0.0034

0.9987 0.9987 0.9987 0.9988 0.9988 0.9989 0.9989 0.9989 0.9990 0.9990

0.0014 0.0013 0.0013 0.0012 0.0012 0.0011 0.0011 0.0011 0.0010 0.0010

3.50 3.51 3.52 3.53 3.54 3.55 3.56 3.57 3.58 3.59

0.0009 0.0008 0.0008 0.0008 0.0008 0.0007 0.0007 0.0007 0.0007 0.0006

0.9998 0.9998 0.9998 0.9998 0.9998 0.9998 0.9998 0.9998 0.9998 0.9998

0.0002 0.0002 0.0002 0.0002 0.0002 0.0002 0.0002 0.0002 0.0002 0.0002

3.10 3.11 3.12 3.13 3.14 3.15 3.16 3.17 3.18 3.19

0.0033 0.0032 0.0031 0.0030 0.0029 0.0028 0.0027 0.0026 0.0025 0.0025

0.9990 0.9991 0.9991 0.9991 0.9992 0.9992 0.9992 0.9992 0.9993 0.9993

0.0010 0.0009 0.0009 0.0009 0.0008 0.0008 0.0008 0.0008 0.0007 0.0007

3.60 3.61 3.62 3.63 3.64 3.65 3.66 3.67 3.68 3.69

0.0006 0.0006 0.0006 0.0006 0.0005 0.0005 0.0005 0.0005 0.0005 0.0004

0.9998 0.9999 0.9999 0.9999 0.9999 0.9999 0.9999 0.9999 0.9999 0.9999

0.0002 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001

3.20 3.21 3.22 3.23 3.24 3.25 3.26 3.27 3.28 3.29

0.0024 0.0023 0.0022 0.0022 0.0021 0.0020 0.0020 0.0019 0.0018 0.0018

0.9993 0.9993 0.9994 0.9994 0.9994 0.9994 0.9994 0.9995 0.9995 0.9995

0.0007 0.0007 0.0006 0.0006 0.0006 0.0006 0.0006 0.0005 0.0005 0.0005

3.70 3.71 3.72 3.73 3.74 3.75 3.76 3.77 3.78 3.79

0.0004 0.0004 0.0004 0.0004 0.0004 0.0003 0.0003 0.0003 0.0003 0.0003

0.9999 0.9999 0.9999 0.9999 0.9999 0.9999 0.9999 0.9999 0.9999 0.9999

0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001

3.30 3.31 3.32 3.33 3.34 3.35 3.36 3.37 3.38 3.39

0.0017 0.0017 0.0016 0.0016 0.0015 0.0015 0.0014 0.0014 0.0013 0.0013

0.9995 0.9995 0.9996 0.9996 0.9996 0.9996 0.9996 0.9996 0.9996 0.9997

0.0005 0.0005 0.0004 0.0004 0.0004 0.0004 0.0004 0.0004 0.0004 0.0003

3.80 3.81 3.82 3.83 3.84 3.85 3.86 3.87 3.88 3.89

0.0003 0.0003 0.0003 0.0003 0.0003 0.0002 0.0002 0.0002 0.0002 0.0002

0.9999 0.9999 0.9999 0.9999 0.9999 0.9999 0.9999 1.0000 1.0000 1.0000

0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001 0.0001

3.40 3.41 3.42 3.43 3.44 3.45 3.46 3.47 3.48 3.49

0.0012 0.0012 0.0011 0.0011 0.0011 0.0010 0.0010 0.0010 0.0009 0.0009

0.9997 0.9997 0.9997 0.9997 0.9997 0.9997 0.9997 0.9997 0.9998 0.9998

0.0003 0.0003 0.0003 0.0003 0.0003 0.0003 0.0003 0.0003 0.0003 0.0002

3.90 3.91 3.92 3.93 3.94 3.95 3.96 3.97 3.98 3.99

0.0002 0.0002 0.0002 0.0002 0.0002 0.0002 0.0002 0.0001 0.0001 0.0001

1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000

0.0001 0.0001 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000 0.0000

3.50

0.0009

0.9998

0.0002

4.00

0.0001

1.0000

0.0000

c 2000 by Chapman & Hall/CRC 

7.2

CRITICAL VALUES

Table 7.1 lists common critical values for a standard normal random variable, zα , defined by (see Figure 7.2): Prob [Z ≥ zα ] = α.

(7.3)

Figure 7.2: Critical values for a normal random variable. α .10 .05 .025 .01 .005

zα 1.2816 1.6449 1.9600 2.3263 2.5758

α .00009 .00008 .00007 .00006 .00005

zα 3.7455 3.7750 3.8082 3.8461 3.8906

.0025 .001 .0005 .0001

2.8070 3.0902 3.2905 3.7190

.00004 .00003 .00002 .00001

3.9444 4.0128 4.1075 4.2649

α .000001 .0000001 .00000001 .000000001 .0000000001

zα 4.75 5.20 5.61 6.00 6.36

Table 7.1: Common critical values 7.3

TOLERANCE FACTORS FOR NORMAL DISTRIBUTIONS

Suppose X1 , X2 , . . . , Xn is a random sample of size n from a normal population with mean µ and standard deviation σ. Using the summary statistics x and s, a tolerance interval [L, U ] may be constructed to capture 100P % of the population with probability 1 − α. The following procedures may be used. (1) Two-sided tolerance interval: A 100(1 − α)% tolerance interval that captures 100P % of the population has as endpoints [L, U ] = x ± Kα,n,P · s

c 2000 by Chapman & Hall/CRC 

(7.4)

(2) One-sided tolerance interval, upper tailed: A 100(1 − α)% tolerance interval bounded below has L = x − kα,n,P · s

U =∞

(7.5)

(3) One-sided tolerance interval, lower tailed: A 100(1 − α)% tolerance interval bounded above has L = −∞

U = x + kα,n,P · s

(7.6)

where Kα,n,P is the tolerance factor given in section 7.3.1 and kα,n,P is computed using the formula below. Values of Kα,n,P are given in section 7.3.1 for P = 0.75, 0.90, 0.95, 0.99, 0.999, α = 0.75, 0.90, 0.95, 0.99, and various values of n. The value of kα,n,P is given by  z1−P + (z1−P )2 − ab kα,n,P = a (zα )2 (7.7) a=1− 2(n − 1) z2 2 b = z1−P − α n where z1−P and zα are critical values for a standard normal random variable (see page 175). Example 7.43 : Suppose a sample of size n = 30 from a normal distribution has x = 10.02 and s = 0.13. Find tolerance intervals with a confidence level 95% (α = .05) and P = .90. Solution: (S1) Two-sided interval: 1. From the tables in section 7.3.1 we find K.05,30,.90 = 2.413. 2. The interval is x ± K · s = 10.02 ± 0.31; or I = [9.71, 10.33]. 3. We conclude: in each sample of size 30, at least 90% of the normal population being sampled will be in the interval I, with probability 95%. (S2) One-sided intervals: 1. The critical values used in equation (7.7) are z1−P = z.10 = 1.282 and 2 = 0.9533, zα = z.05 = 1.645. Using this equation: a = 1 − (1.645) 2(29) 2

= 1.553, and k.05,30,.90 = 1.768 b = (1.282)2 − (1.645) 30 2. The lower bound is L = x − k · s = 9.79. 3. The upper bound is U = x + k · s = 10.25. 4. We conclude: (a) In each sample of size 30, at least 90% of the normal population being sampled will be greater than L, with probability 95%. (b) In each sample of size 30, at least 90% of the normal population being sampled will be smaller than U , with probability 95%.

c 2000 by Chapman & Hall/CRC 

7.3.1

n 2 3 4 5 6 7 8 9 10 15

n 2 3 4 5 6 7 8 9 10 15

Tables of tolerance intervals for normal distributions Tolerance factors for normal distributions P = .90 α = .10 .05 .01 .001 n α = .10 .05 .01 15.978 18.800 24.167 30.227 20 2.152 2.564 3.368 5.847 6.919 8.974 11.309 25 2.077 2.474 3.251 4.166 4.943 6.440 8.149 30 2.025 2.413 3.170 3.494 4.152 5.423 6.879 40 1.959 2.334 3.066 3.131 3.723 4.870 6.188 50 1.916 2.284 3.001 2.902 2.743 2.626 2.535 2.278

2.906 2.854 2.689 2.654 2.576

3.712 3.646 3.434 3.390 3.291

Tolerance factors for normal distributions P = .95 α = .10 .05 .01 .001 n α = .10 .05 .01 32.019 37.674 48.430 60.573 20 2.310 2.752 3.615 8.380 9.916 12.861 16.208 25 2.208 2.631 3.457 5.369 6.370 8.299 10.502 30 2.140 2.549 3.350 4.275 5.079 6.634 8.415 40 2.052 2.445 3.213 3.712 4.414 5.775 7.337 50 1.996 2.379 3.126

.001 4.614 4.413 4.278 4.104 3.993

3.369 3.136 2.967 2.839 2.480

n α = .10 2 160.193 3 18.930 4 9.398 5 6.612 6 5.337 7 8 9 10 15

4.613 4.147 3.822 3.582 2.945

3.452 3.264 3.125 3.018 2.713

4.007 3.732 3.532 3.379 2.954

4.521 4.278 4.098 3.959 3.562

5.248 4.891 4.631 4.433 3.878

5.750 75 5.446 100 5.220 500 5.046 1000 4.545 ∞

6.676 75 6.226 100 5.899 500 5.649 1000 4.949 ∞

1.856 1.822 1.717 1.695 1.645

1.917 1.874 1.737 1.709 1.645

2.211 2.172 2.046 2.019 1.960

.001 4.300 4.151 4.049 3.917 3.833

2.285 2.233 2.070 2.036 1.960

3.002 2.934 2.721 2.676 2.576

Tolerance factors for normal distributions P = .99 .05 .01 .001 n α = .10 .05 .01 188.491 242.300 303.054 20 2.659 3.168 4.161 22.401 29.055 36.616 25 2.494 2.972 3.904 11.150 14.527 18.383 30 2.385 2.841 3.733 7.855 10.260 13.015 40 2.247 2.677 3.518 6.345 8.301 10.548 50 2.162 2.576 3.385 5.488 4.936 4.550 4.265 3.507

7.187 6.468 5.966 5.594 4.605

c 2000 by Chapman & Hall/CRC 

9.142 75 8.234 100 7.600 500 7.129 1000 5.876 ∞

2.042 1.977 1.777 1.736 1.645

2.433 2.355 2.117 2.068 1.960

3.197 3.096 2.783 2.718 2.576

3.835 3.748 3.475 3.418 3.291

.001 5.312 4.985 4.768 4.493 4.323 4.084 3.954 3.555 3.472 3.291

7.4 7.4.1

OPERATING CHARACTERISTIC CURVES One-sample Z test

Consider a one-sample hypothesis test on a population mean of a normal distribution with known standard deviation σ (see section 10.2). The general form of the hypothesis test (for each possible alternative hypothesis) is: H0 : µ = µ0 Ha : µ > µ0 ,

µ < µ0 ,

µ = µ0

¯ − µ0 X √ σ/ n RR: Z ≥ zα , Z ≤ −zα , TS: Z =

|Z| ≥ zα/2

Let α be the probability of a Type I error, β the probability of a Type II error, and µa an alternative mean. For ∆ = |µa − µ0 |/σ the operating characteristic curve returns the probability of not rejecting the null hypothesis given µ = µa . The curves may be used to determine the appropriate sample size for given values of α, β, and ∆. 7.4.2

Two-sample Z test

Consider a two-sample hypothesis test for comparing population means from normal distributions with known standard deviations σ1 and σ2 (see section 10.3). The general form of the hypothesis test for testing the equality of means (for each possible alternative hypothesis) is: H0 : µ1 − µ2 = 0 Ha : µ1 − µ2 > 0,

µ1 − µ2 < 0,

µ1 − µ2 = 0

¯1 − X ¯2 X TS: Z =  2 σ1 σ22 n1 + n2 RR: Z ≥ zα ,

Z ≤ −zα ,

|Z| ≥ zα/2

Let α be the probability of a Type I error and β the probability of a Type II |µ1 − µ2 | error. For given values of α, and ∆ =  2 the operating characteristic σ1 + σ22 curve returns the probability of not rejecting the null hypothesis. The curves may be used to determine an appropriate sample size (n = n1 = n2 ) for desired levels of α, β, and ∆.

c 2000 by Chapman & Hall/CRC 

Figure 7.3: Operating characteristic curves, various values of n, Z test, twosided alternative, α = .05.

Figure 7.4: Operating characteristic curves, various values of n, Z test, twosided alternative, α = .01.

c 2000 by Chapman & Hall/CRC 

Figure 7.5: Operating characteristic curves, various values of n, Z test, onesided alternative, α = .05.

Figure 7.6: Operating characteristic curves, various values of n, Z test, onesided alternative, α = .01.

c 2000 by Chapman & Hall/CRC 

7.5

MULTIVARIATE NORMAL DISTRIBUTION

Let each {Xi } (for i = 1, . . . , n) be a normal random variable with mean µi and variance σii . If the covariance of Xi and Xj is σij , then the joint probability density of the {Xi } is:   1 1  f (x) = exp − (x − µ)T Σ−1 (x − µ) (7.8) 2 (2π)n/2 det(Σ) where  T (a) x = x1 x2 . . . xn  T (b) µ = µ1 µ2 . . . µn (c) Σ is an n × n matrix with elements σij The corresponding characteristic function is   1 T T φ(t) = exp iµ t − t Σt 2

(7.9)

The form of the characteristic function implies that all cumulants of higher order than 2 vanish (see Marcienkiewicz’s theorem). Therefore, all moments of order higher than 2 may be expressed in terms of those of order 1 and 2. If µ = 0 then the odd moments vanish and the (2n)th moment satisfies   (2n)! E Xi Xj Xk Xl · · · = {σij σkl · · ·}sym (7.10) n!2n   2n terms

where the subscript “sym” means the symmetrized form of the product of the σ’s. Example 7.44 : For n = 2 we can compute fourth moments E [X1 X2 X3 X4 ] = 

4

E X1

4! 2! · 22



 1 (σ12 σ34 + σ41 σ23 + σ13 σ24 ) 3

= σ12 σ34 + σ41 σ23 + σ13 σ24 4!  2  2 = σ11 = 3σ11 2! · 22

(7.11)

See C. W. Gardiner Handbook of Stochastic Methods, Springer–Verlag, New York, 1985, pages 36–37.

c 2000 by Chapman & Hall/CRC 

7.6 DISTRIBUTION OF THE CORRELATION COEFFICIENT FOR A BIVARIATE NORMAL The bivariate normal probability function is given by 1 1  (7.12) f (x, y) = exp − 2 2(1 − ρ2 ) 2πσx σy 1 − ρ % 2    2 &  y − µy y − µy x − µx x − µx + × − 2ρ σx σx σy σy where µx = mean of x µy = mean of y σx = standard deviation of x σy = standard deviation of y ρ = correlation coefficient between x and y Given a sample {(x1 , y1 ), . . . , (xn , yn )} of size n, the sample correlation coefficient, an estimate of ρ, is n  (xi − x) (yi − y) i=1 r = + (7.13)  n  n   2 2 (xi − x) (yi − y) i=1

i=1

n n where x = ( i=1 xi ) /n and y = ( i=1 yi ) /n. For given n, the distribution of r is independent of {µx , µy , σx , σy }, but depends on ρ. For −1 ≤ ρ ≤ 1, the density function for r is fn (r; ρ):  ∞   dz 1 2 (n−4)/2 2 (n−1)/2 fn (r; ρ) = (n − 2)(1 − r ) 1−ρ π (cosh z − ρr)n−1 0  (n−1)/2 π Γ(n − 1) 1 = (n − 2)(1 − r2 )(n−4)/2 1 − ρ2 π 2 Γ(n − 1/2)   1 1 2n − 1 ρr + 1 −(n−3/2) × (1 − ρr) , ; ; 2 F1 2 2 2 2 (7.14)

c 2000 by Chapman & Hall/CRC 

where Γ(x) is the gamma function and 2 F1 is the hypergeometric function defined in Chapter 18 (see pages 515 and 520). The moments are given by ρ(1 − ρ2 ) (n + 1)   (1 − ρ2 )2 11ρ2 2 σr = 1+ + ... n+1 2(n + 1)   6ρ 77ρ2 − 30 γ1 = √ + ... 1+ 12(n + 1) n+1  6  γ2 = 12ρ2 − 1 + . . . n+1 µr = ρ −

7.6.1

(7.15)

Normal approximation

If r is the sample correlation coefficient (defined in equation (7.13)), the random variable 1 1+r Z = tanh−1 r = ln (7.16) 2 1−r is approximately normally distributed with parameters   1 1 1+ρ 2 µZ = ln = = tanh−1 ρ and σZ (7.17) 2 1−ρ n−3 7.6.2

Zero correlation coefficient for bivariate normal

In the special case where ρ = 0, the density function of r becomes (n−4)/2 1 Γ ((n − 1)/2)  fn (r; 0) = √ 1 − r2 π Γ ((n − 2)/2)

(7.18)

Under the transformation r2 =

t2 t2 + ν

(7.19)

fn (r; 0), as given by equation (7.18), has a t-distribution with ν = n−1 degrees of freedom. The following table gives percentage points of the distribution of the correlation coefficient when ρ = 0.

c 2000 by Chapman & Hall/CRC 

Percentage points of the correlation coefficient, when ρ = 0 Prob [r ≤ tabulated value] = 1 − α α= 2α = ν=1 2 3 4 5

0.05 0.1 0.988 0.900 0.805 0.729 0.669

0.025 0.05 0.997 0.950 0.878 0.811 0.754

0.01 0.02 0.93 507 0.980 0.934 0.882 0.833

0.005 0.01 0.93 877 0.990 0.959 0.917 0.875

0.0025 0.005 0.94 69 0.995 0.974 0.942 0.906

0.0005 0.001 0.96 0.999 0.991 0.974 0.951

6 7 8 9 10

0.621 0.582 0.549 0.521 0.497

0.707 0.666 0.632 0.602 0.576

0.789 0.750 0.715 0.685 0.658

0.834 0.798 0.765 0.735 0.708

0.870 0.836 0.805 0.776 0.750

0.925 0.898 0.872 0.847 0.823

11 12 13 14 15

0.476 0.458 0.441 0.426 0.412

0.553 0.532 0.514 0.497 0.482

0.634 0.612 0.592 0.574 0.558

0.684 0.661 0.641 0.623 0.606

0.726 0.703 0.683 0.664 0.647

0.801 0.780 0.760 0.742 0.725

16 17 18 19 20

0.400 0.389 0.378 0.369 0.360

0.468 0.456 0.444 0.433 0.423

0.543 0.529 0.516 0.503 0.492

0.590 0.575 0.561 0.549 0.537

0.631 0.616 0.602 0.589 0.576

0.708 0.693 0.679 0.665 0.652

25 30 35 40 45 50

0.323 0.296 0.275 0.257 0.243 0.231

0.381 0.349 0.325 0.304 0.288 0.273

0.445 0.409 0.381 0.358 0.338 0.322

0.487 0.449 0.418 0.393 0.372 0.354

0.524 0.484 0.452 0.425 0.403 0.384

0.597 0.554 0.519 0.490 0.465 0.443

60 70 80 90 100

0.211 0.195 0.183 0.173 0.164

0.250 0.232 0.217 0.205 0.195

0.295 0.274 0.257 0.242 0.230

0.325 0.302 0.283 0.267 0.254

0.352 0.327 0.307 0.290 0.276

0.408 0.380 0.357 0.338 0.321

Use the α value for a single-tail test. For a two-tail test, use the 2α value. If r is computed from n paired observations, enter the table with ν = n − 2. For partial correlations, enter the table with ν = n − 2 − k, where k is the number of variables held constant.

c 2000 by Chapman & Hall/CRC 

7.7

CIRCULAR NORMAL PROBABILITIES

The joint probability density of two independent and normally distributed random variables X and Y (each of mean zero and variance σ 2 ) is    1 1  2 2 x (7.20) f (x, y) = exp − + y 2πσ 2 2σ 2 The following table gives the probability that a sample value of X and Y is obtained inside a circle C of radius R at a distance d from the origin:       d R 1 1  2 2 F , = exp − 2 x + y dx dy (7.21) σ σ 2πσ 2 2σ C Circular normal probabilities R/σ d/σ 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 1.2 1.4 1.6 1.8 2.0 2.2 2.4 2.6 2.8 3.0 3.2 3.4 3.6 3.8 4.0 4.2 4.4 4.6 4.8 5.0 5.2 5.4 5.6 5.8

0.2 0.020 0.019 0.019 0.018 0.017 0.017 0.016 0.014 0.013 0.012 0.0097 0.0075 0.0056 0.0040 0.0027 0.0018 0.0011 0.0007 0.0004 0.0002 0.0001 0.0001

0.4 0.077 0.075 0.074 0.071 0.068 0.065 0.061 0.057 0.052 0.048 0.038 0.030 0.022 0.016 0.011 0.0075 0.0048 0.0030 0.0018 0.0010 0.0006 0.0003 0.0002 0.0001

0.6 0.164 0.162 0.158 0.153 0.147 0.140 0.132 0.123 0.114 0.104 0.085 0.067 0.051 0.037 0.026 0.018 0.012 0.0074 0.0045 0.0027 0.0015 0.0008 0.0004 0.0002

c 2000 by Chapman & Hall/CRC 

0.8 0.273 0.269 0.264 0.256 0.246 0.235 0.222 0.209 0.194 0.179 0.148 0.119 0.092 0.069 0.050 0.034 0.023 0.015 0.0094 0.0057 0.0033 0.0018 0.0010 0.0005 0.0001 0.0001

1.0 0.392 0.387 0.380 0.370 0.357 0.342 0.326 0.307 0.288 0.267 0.225 0.184 0.145 0.111 0.082 0.059 0.040 0.027 0.017 0.011 0.0065 0.0038 0.0021 0.0011 0.0003 0.0001 0.0001

1.5 0.674 0.668 0.659 0.647 0.631 0.612 0.591 0.566 0.540 0.512 0.451 0.388 0.325 0.264 0.209 0.161 0.120 0.086 0.060 0.041 0.027 0.017 0.010 0.006 0.0019 0.0010 0.0005 0.0003 0.0001 0.0001

2.0 0.863 0.859 0.852 0.843 0.831 0.816 0.799 0.779 0.756 0.731 0.674 0.610 0.541 0.469 0.396 0.327 0.262 0.204 0.154 0.113 0.080 0.055 0.037 0.024 0.0088 0.0051 0.0029 0.0015 0.0008 0.0004 0.0002 0.0001

2.5 0.955 0.953 0.950 0.945 0.938 0.930 0.920 0.909 0.895 0.879 0.841 0.795 0.739 0.676 0.606 0.532 0.456 0.381 0.311 0.246 0.189 0.141 0.102 0.072 0.0320 0.0200 0.0120 0.0073 0.0041 0.0023 0.0012 0.0006 0.0003

3.0 0.989 0.988 0.987 0.985 0.982 0.979 0.975 0.970 0.964 0.956 0.937 0.912 0.878 0.837 0.786 0.726 0.659 0.586 0.510 0.433 0.358 0.288 0.225 0.171 0.090 0.062 0.041 0.027 0.017 0.0100 0.0058 0.0033 0.0018

7.8

CIRCULAR ERROR PROBABILITIES

The joint probability density of two independent and normally distributed random variables X and Y , each of mean zero and having variances σx2 and σy2 , is % (   2 & 2 1 y 1 x exp − + f (x, y) = (7.22) 2πσx σy 2 σx σy The probability that a sample value of X and Y will lie within a circle with center at the origin and radius Kσx is given by  P (K, σx , σy ) = f (x, y) dx dy (7.23) R

 where R is the region x2 + y 2 < Kσx . For various values of K and c = σx /σy (for convenience we assume that σy ≤ σx ) the following table gives the value of P . Circular error probabilities K 0.1 0.2 0.3 0.4 0.5

c = 0.0 0.0797 0.1585 0.2358 0.3108 0.3829

0.2 0.0242 0.0885 0.1739 0.2635 0.3482

0.4 0.0124 0.0482 0.1039 0.1742 0.2533

0.6 0.0083 0.0327 0.0719 0.1238 0.1857

0.8 0.0062 0.0247 0.0547 0.0951 0.1444

1.0 0.0050 0.0198 0.0440 0.0769 0.1175

0.6 0.7 0.8 0.9 1.0

0.4515 0.5161 0.5763 0.6319 0.6827

0.4256 0.4961 0.5604 0.6191 0.6724

0.3357 0.4171 0.4942 0.5652 0.6291

0.2548 0.3280 0.4026 0.4759 0.5461

0.2010 0.2629 0.3283 0.3953 0.4621

0.1647 0.2173 0.2739 0.3330 0.3935

1.2 1.4 1.6 1.8 2.0

0.7699 0.8385 0.8904 0.9281 0.9545

0.7630 0.8340 0.8875 0.9263 0.9534

0.7359 0.8170 0.8769 0.9197 0.9494

0.6714 0.7721 0.8478 0.9019 0.9388

0.5893 0.7008 0.7917 0.8613 0.9116

0.5132 0.6247 0.7220 0.8021 0.8647

2.2 2.4 2.6 2.8 3.0

0.9722 0.9836 0.9907 0.9949 0.9973

0.9715 0.9832 0.9905 0.9948 0.9972

0.9692 0.9819 0.9897 0.9944 0.9970

0.9631 0.9785 0.9879 0.9934 0.9965

0.9459 0.9683 0.9821 0.9903 0.9949

0.9111 0.9439 0.9660 0.9802 0.9889

3.2 3.4 3.6 3.8 4.0

0.9986 0.9993 0.9997 0.9999 0.9999

0.9986 0.9993 0.9997 0.9999 0.9999

0.9985 0.9993 0.9997 0.9998 0.9999

0.9982 0.9991 0.9996 0.9998 0.9999

0.9974 0.9988 0.9994 0.9997 0.9999

0.9940 0.9969 0.9985 0.9993 0.9997

c 2000 by Chapman & Hall/CRC 

CHAPTER 8

Estimation Contents 8.1 8.2 8.3 8.4 8.5 8.6 8.7 8.8 8.9 8.10

Definitions Cram´ er–Rao inequality Theorems The method of moments The likelihood function The method of maximum likelihood Invariance property of MLEs Different estimators Estimators for small samples Estimators for large samples

A nonconstant function of a set of random variables is a statistic. It is a function of observable random variables, which does not contain any unknown parameters. A statistic is itself an observable random variable. Let θ be a parameter appearing in the density function for the random variable X. Let g be a function that returns an approximate value θ. of θ from a given sample {x1 , . . . , xn }. Then θ. = g(x1 , x2 , . . . , xn ) may be considered . = g(X1 , X2 , . . . , Xn ). The a single observation of the random variable Θ . is an estimator for the parameter θ. random variable Θ 8.1

DEFINITIONS . is an unbiased estimator for θ if E [Θ] . = θ. (1) Θ . is B [Θ] . = E [Θ] . − θ. (2) The bias of the estimator Θ . is (3) The mean square error of Θ   . = E (Θ . − θ)2 = Var [Θ] . + B [Θ] . 2. MSE [Θ]

. − θ|. (4) The error of estimation is 3 = |Θ . 1 and Θ . 2 be unbiased estimators for θ. (5) Let Θ . 1 ] < Var [Θ . 2 ] then the estimator Θ . 1 is relatively more (a) If Var [Θ . efficient than the estimator Θ2 . c 2000 by Chapman & Hall/CRC 

. 2 relative to Θ . 1 is (b) The efficiency of Θ . 1] Var [Θ Efficiency = . . Var [Θ2 ] . is a consistent estimator for θ if for every 3 > 0, (6) Θ . − θ| ≤ 3] = 1 or, equivalently lim Prob [|Θ n→∞

. − θ| > 3] = 0. lim Prob [|Θ

n→∞

. is a sufficient estimator for θ if for each value of Θ . the conditional (7) Θ . distribution of X1 , X2 , . . . , Xn given Θ = θ0 is independent of θ. . be an estimator for the parameter θ and suppose Θ . has sampling (8) Let Θ . Then Θ . is a complete statistic if for all θ, E [h (Θ)] . = distribution g(Θ). . = 0 for all functions h (Θ). . 0 implies h (Θ) . for θ is a minimum variance unbiased estimator (9) An estimator Θ (MVUE) if it is unbiased and has the smallest possible variance. 2

To determine if X is an unbiased estimator of µ2 , consider the following expected value: %

2 & n " 2# 1 E X =E Xi n i=1   n n  1  2 = 2E Xi + Xi Xj  n i=1 i=1,j=1,i=j (8.1)   1 = 2 n(σ 2 + µ2 ) + (n2 − n)µ2 n

Example 8.45 :

= µ2 +

σ2 n

> µ2 2

This shows X is a biased estimator of µ2 .

8.2

´ CRAMER–RAO INEQUALITY

Let {X1 , X2 , . . . , Xn } be a random sample from a population with probability . be an unbiased estimator for θ. Under very density function f (x). Let Θ general conditions it can be shown that 1 1 . ≥ #.  " 2 Var [Θ] (8.2) 2  = ∂ ln f (X) −n · E 2 n · E ∂ ln f (X) ∂θ ∂θ

. is a minimum variance unbiased estimator (MVUE) If equality holds then Θ for θ. c 2000 by Chapman & Hall/CRC 

Example 8.46 :

The probability density function for a normalrandom variable with  2 (x − θ) 1 unknown mean θ and known variance σ 2 is f (x; θ) = √ exp − . Use the 2σ 2 2πσ . for Cram´er–Rao inequality to show the minimum variance of any unbiased estimator, Θ, θ is at least σ 2 /n. Solution: (X − θ) ∂ ln f (X; θ) = (S1) ∂θ σ2  2 (X − θ)2 ∂ ln f (X; θ) (S2) = ∂θ σ4      ∞ (X − θ)2 (x − θ)2 1 1 −(x−θ)2 /2σ 2 √ e (S3) E = dx = 2 σ4 σ4 σ 2πσ −∞ 2 . ≥ 1 = σ (S4) Var [Θ] 1 n n σ2

8.3

THEOREMS . (1) Θ is a consistent estimator for θ if . is unbiased, and (a) Θ . = 0. (b) lim Var [Θ] n→∞

. is a sufficient estimator for the parameter θ if the joint distribution (2) Θ of {X1 , X2 , . . . , Xn } can be factored into . θ) · h(x1 , x2 , . . . , xn ) f (x1 , x2 , . . . , xn ; θ) = g(Θ,

(8.3)

. θ) depends only on the estimate θ. and the parameter θ, and where g(Θ, h(x1 , x2 , . . . , xn ) does not depend on the parameter θ. (3) Unbiased estimators: (a) An unbiased estimator may not exist. (b) An unbiased estimator is not unique. (c) An unbiased estimator may be meaningless. (d) An unbiased estimator is not necessarily consistent. (4) Consistent estimators: (a) A consistent estimator is not unique. (b) A consistent estimator may be meaningless. (c) A consistent estimator is not necessarily unbiased. (5) Maximum likelihood estimators (MLE): (a) A MLE need not be consistent. (b) A MLE may not be unbiased. (c) A MLE is not unique. c 2000 by Chapman & Hall/CRC 

(d) If a single sufficient statistic T exists for the parameter θ, the MLE of θ must be a function of T . . be a MLE of θ. If τ (·) is a function with a single-valued (e) Let Θ . inverse, then a MLE of τ (θ) is τ (Θ). (6) Method of moments (MOM) estimators: (a) MOM estimators are not uniquely defined. (b) MOM estimators may not be functions of sufficient or complete statistics. (7) A single sufficient estimator may not exist. 8.4

THE METHOD OF MOMENTS

The moment estimators are the solutions to the systems of equations 1 r x = mr , n i=1 i n

µr = E [X r ] =

r = 1, 2, . . . , k

where k is the number of parameters. Example 8.47 : Suppose X1 , X2 , . . . , Xn is a random sample from a population having a gamma distribution with parameters α and β. Use the method of moments to obtain estimators for the parameters α and β. Solution: (S1) The system of equations to solve: µ1 = m1 ; µ2 = m2   (S2) µ1 = E [X] = αβ ; µ2 = E X 2 = α(α + 1)β 2 (S3) Solve αβ = m1 and α(α + 1)β 2 = m2 for α and β. m − (m1 )2 β. = 2 m1 n n 1 1 2 (S5) Given m1 = x = xi and m2 = xi , then n i=1 n i=1 n  (xi − x)2 2 nx i=1 α .=  and β. = n nx (xi − x)2

.= (S4) α

(m1 )2 ; − (m1 )2

m2

i=1

8.5

THE LIKELIHOOD FUNCTION

Let x1 , x2 , . . . , xn be the values of a random sample from a population characterized by the parameters θ1 , θ2 , . . . , θr . The likelihood function of the sample is (1) the joint probability mass function evaluated at (x1 , x2 , . . . , xn ) if (X1 , X2 , . . . , Xn ) are discrete, L(θ1 , θ2 , . . . , θr ) = p(x1 , x2 , . . . , xn ; θ1 , θ2 , . . . , θr ) c 2000 by Chapman & Hall/CRC 

(8.4)

(2) the joint probability density function evaluated at (x1 , x2 , . . . , xn ) if (X1 , X2 , . . . , Xn ) are continuous. L(θ1 , θ2 , . . . , θr ) = f (x1 , x2 , . . . , xn ; θ1 , θ2 , . . . , θr ) 8.6

(8.5)

THE METHOD OF MAXIMUM LIKELIHOOD

The maximum likelihood estimators (MLEs) are those values of the parameters that maximize the likelihood function of the sample: L(θ1 , . . . , θr ). In practice, it is often easier to maximize ln L(θ1 , . . . , θr ). This is equivalent to maximizing the likelihood function, L(θ1 , . . . , θr ), since ln L(θ1 , . . . , θr ) is a monotonic function of L(θ1 , . . . , θr ). Example 8.48 : Suppose X1 , X2 , . . . , Xn is a random sample from a population having a Poisson distribution with parameter λ. Find the maximum likelihood estimator for the parameter λ. Solution: (S1) The probability mass function for a Poisson random variable is e−λ λx f (x; λ) = x! (S2) We compute  −λ x1   −λ x2   −λ xn  e λ e λ e λ ··· L(θ) = x1 ! x2 ! xn ! e−nλ λx1 +x2 +···+xn x1 !x2 ! · · · xn ! ln L(θ) = −nλ + (x1 + x2 + · · · + xn ) ln λ + ln (x1 !x2 ! · · · xn !) =

(8.6)

∂ ln L(λ) x1 + x2 + · · · + xn = −n + =0 ∂λ λ . = x1 + x2 + · · · + xn = x is the MLE for λ. (S4) Solving for λ: λ n

(S3)

8.7 INVARIANCE PROPERTY OF MAXIMUM LIKELIHOOD ESTIMATORS . 1, Θ . 2, . . . , Θ . r be the maximum likelihood estimators for θ1 , θ2 , . . . , θr and Let Θ let h(θ1 , θ2 , . . . , θr ) be a function of θ1 , θ2 , . . . , θr . The maximum likelihood es. 1, Θ . 2, . . . , Θ . r ). h(θ1 , θ2 , . . . , θr ) = h(Θ timator of the parameter h(θ1 , θ2 , . . . , θr ) is . 8.8

DIFFERENT ESTIMATORS

Assume {x1 , x2 , . . . , xn } is a set of observations. Let UMV stand for uniformly minimum variance unbiased and let MLE stand for maximum likelihood estimator. (1) Normal distribution: N (µ, σ 2 ) (a) When σ is known:  1. xi is necessary, sufficient, and complete. c 2000 by Chapman & Hall/CRC 

2. Point estimate for µ: µ .=x=

1 xi is UMV, MLE. n

(b) When µ is known:  1. (xi − µ)2 is necessary, sufficient, and complete.  − µ)2 (x i /2 = 2. Point estimate for σ 2 : σ is UMV, MLE. n (c) When µ and σ are unknown:   1. { xi , (xi − x)2 } are necessary, sufficient, and complete. 1 2. Point estimate for µ: µ .=x= xi is UMV, MLE. n 2 /2 = (xi − x) is MLE. 3. Point estimate for σ 2 : σ  n 2 /2 = (xi − x) is UMV. 4. Point estimate for σ 2 : σ n − 1 - Γ [(n − 1)/2] (xi − x)2 5. Point estimate for σ: σ .= √ is UMV. n−1 2Γ(n/2) (2) Poisson distribution with parameter λ:  (a) xi is necessary, sufficient, and complete.  .= 1 (b) Point estimate for λ: λ xi is UMV, MLE. n (3) Uniform distribution on an interval: (a) Interval is [0, θ] 1. max(xi ) is necessary, sufficient, and complete. . = max(xi ) is MLE. 2. Point estimate for θ: Θ . = n + 1 max(xi ) is UMV. 3. Point estimate for θ: Θ n (b) Interval is [α, β] 1. {min(xi ), max(xi )} are necessary, sufficient, and complete. n min(xi ) − max(xi ) 2. Point estimate for α: α .= is UMV. n−1 3. Point estimate for α: α . = min(xi ) is MLE. α0 +β min(xi ) + max(xi ) 4. Point estimate for α+β : = is UMV. 2 2 2   (c) Interval is θ − 12 , θ + 12 1. {min(xi ), max(xi )} are necessary and sufficient. . = min(xi ) + max(xi ) is MLE. 2. Point estimate for θ: Θ 2

c 2000 by Chapman & Hall/CRC 

8.9 ESTIMATORS FOR MEAN AND STANDARD DEVIATION IN SMALL SAMPLES In all cases below, the variance of an estimate must be multiplied by the the true variance of the sample, σ 2 . Different estimators for the mean Median n Var

Eff

Midrange

Average of best two

Var

Statistic

Eff

2 .500 1.000 .500 1.000 3 .449

.743 .362

.920

4 .298

.838 .298

.838

5 .287

.697 .261

.767

Eff



n−1 i=2

 xi

Var

Eff

1 2 (x1 1 2 (x1

+ x2 )

.500 1.000

+ x3 )

.362

.920

.449

.743

1 2 (x2 1 2 (x2

+ x3 )

.298

.838

.298

.838

+ x4 )

.231

.867

.227

.881

.119

.840

.105

.949

.081

.825

.069

.969

+ x15 ) .061

.824

.051

.978

1.24 n

.808

10 .138

.723 .186

.539

15 .102

.656 .158

.422

1 2 (x3 + x8 ) 1 2 (x4 + x12 )

20 .073

.681 .143

.350

1 2 (x6



.637

.000

1 2 (P25

1.57 n

Var

1 n−2

+ P75 )

1.000

Estimating standard deviation σ from the sample range w n Estimator Variance Efficiency 2 .886w .571 1.000 3 .591w .275 .992 4 .486w .183 .975 5 .430w .138 .955 6 .395w .112 .933 7 .370w .095 .911 8 .351w .083 .890 9 .337w .074 .869 10 .325w .067 .850 15 .288w .047 .766 20 .268w .038 .700

n 2 3 4 5 6 7

Best linear estimate of the standard deviation σ Estimator Efficiency .8862(x2 − x1 ) 1.000 .5908(x3 − x1 ) .992 .4539(x4 − x1 ) + .1102(x3 − x2 ) .989 .3724(x5 − x1 ) + .1352(x3 − x2 ) .988 .3175(x6 − x1 ) + .1386(x5 − x2 ) + .0432(x4 − x3 ) .988 .2778(x7 − x1 ) + .1351(x6 − x2 ) + .0625(x5 − x3 ) .989

c 2000 by Chapman & Hall/CRC 

8.10 ESTIMATORS FOR MEAN AND STANDARD DEVIATION IN LARGE SAMPLES Percentile estimates may be used to estimate both the mean and the standard deviation. Estimators for the mean Number of terms 1 2 3 4 5

Estimator using percentiles P50 1/2 (P 25 + P75 ) 1/3 (P 17 + P50 + P83 ) 1/4 (P + P37.5 + P62.5 + P87.5 ) 12.5 1/5 (P + P 10 30 + P50 + P70 + P90 )

Efficiency .64 .81 .88 .91 .93

Estimators for the standard deviation Number of terms Estimator using percentiles Efficiency 2 .3388 (P93 − P7 ) .65 4 .1714 (P97 + P85 − P15 − P3 ) .80 6 .1180 (P98 + P91 + P80 − P20 − P9 − P2 ) .87

c 2000 by Chapman & Hall/CRC 

CHAPTER 9

Confidence Intervals Contents 9.1 9.2 9.3 9.4 9.5

Definitions Common critical values Sample size calculations Summary of common confidence intervals Confidence intervals: one sample 9.5.1 Mean of normal population, known variance 9.5.2 Mean of normal population, unknown var 9.5.3 Variance of normal population 9.5.4 Success in binomial experiments 9.5.5 Confidence interval for percentiles 9.5.6 Confidence interval for medians 9.5.7 Confidence interval for Poisson distribution 9.5.8 Confidence interval for binomial distribution 9.6 Confidence intervals: two samples 9.6.1 Difference in means, known variances 9.6.2 Difference in means, equal unknown var 9.6.3 Difference in means, unequal unknown var 9.6.4 Difference in means, paired observations 9.6.5 Ratio of variances 9.6.6 Difference in success probabilities 9.6.7 Difference in medians 9.7 Finite population correction factor

9.1

DEFINITIONS

A simple point estimate θ. of a parameter θ serves as a best guess for the value of θ, but conveys no sense of confidence in the estimate. A confidence . is used to make statements about θ when the sample interval I, based on θ, size, the underlying distribution of θ, and the confidence coefficient 1 − α are known. We make statements of the form: The probability that θ is in a specified interval is 1 − α. c 2000 by Chapman & Hall/CRC 

(9.1)

To construct a confidence interval for a parameter θ, the confidence coefficient must be specified. The usual procedure is to specify the confidence coefficient 1 − α and then determine the confidence interval. A typical value is α = 0.05 (also written as α = 5%); so that 1 − α = 0.95 (or 95% confidence). The confidence interval may be denoted (θlow , θhigh ). There are many ways to specify θlow and θhigh , depending on the parameter θ and the underlying distribution. The bounds on the confidence interval are usually defined to satisfy The probability that θ < θlow is α/2: Prob [θ < θlow ] = α/2 (9.2) The probability that θ > θhigh is α/2: Prob [θ > θhigh ] = α/2 so that Prob [θlow ≤ θ ≤ θhigh ] = 1 − α. It is also possible to construct one-sided confidence intervals. For these, (1) θlow = −∞ and Prob [θ > θhigh ] = α or Prob [θ ≤ θhigh ] = 1 − α (one-sided, lower-tailed confidence interval), or (2) θhigh = ∞ and Prob [θ < θlow ] = α or Prob [θ ≥ θlow ] = 1 − α (one-sided, upper-tailed confidence interval). Notes: (1) When the sample size is at least 5% of the total population, a finite population correction factor may be used to modify a confidence interval. See section (9.7). (2) If the test statistic is significant in an ANOVA, confidence intervals may then be used to determine which pairs of means differ. 9.2

COMMON CRITICAL VALUES

The formulas for common confidence intervals usually involve critical values from the normal distribution, the t distribution, or the chi–square distribution; see Tables 9.3 and 9.4. More extensive critical value tables for the normal distribution are given on page 175, for the t distribution on page 156, for the chi–square distribution on page 6.4.4, and for the F distribution on page 131.

9.3

SAMPLE SIZE CALCULATIONS

In order to construct a confidence interval of specified width, a priori parameter estimates and a bound on the error of estimation may be used to determine the necessary sample size. For a 100(1 − α)% confidence interval, let E = error of estimation (half the width of the confidence interval). Table 9.2 presents some common sample size calculations. Example 9.49 : A researcher would like to estimate the probability of a success, p, in a binomial experiment. How large a sample is necessary in order to estimate c 2000 by Chapman & Hall/CRC 

α Distribution

0.10

0.05

0.01

0.001

0.0001

tα/2,10

1.8125

2.2281

3.1693

4.5869

6.2111

tα/2,100

1.6602

1.9840

2.6259

3.3905

4.0533

tα/2,1000

1.6464

1.9623

2.5808

3.3003

3.9063

1.6449

1.9600

2.5758

3.2905

3.8906

3.9403

3.2470

2.1559

1.2650

0.7660

18.3070

20.4832

25.1882

31.4198

37.3107

77.9295

74.2219

67.3276

59.8957

54.1129

124.3421

129.5612

140.1695

153.1670

164.6591

927.5944

914.2572

888.5635

859.3615

835.3493

t distribution

Normal distribution zα/2 χ2 distribution χ21−α/2,10 χ2α/2,10 χ21−α/2,100 χ2α/2,100 χ21−α/2,1000 χ2α/2,1000

1074.6790 1089.5310 1118.9480 1153.7380 1183.4920 0.90

0.95

0.99

0.999

0.9999

1−α

Table 9.1: Common critical values used with confidence intervals Parameter

Estimate

Sample size z 2 α/2 · σ n= E

µ

x

p

p.

µ2 − µ2

x1 − x2

n1 = n 2 =

(zα/2 )2 (σ12 + σ22 ) E2

(3)

p1 − p2

p.1 − p.2

n1 = n 2 =

(zα/2 )2 (p1 q1 + p2 q2 ) E2

(4)

n=

(zα/2 )2 · pq E2

(1)

(2)

Table 9.2: Common sample size calculations this proportion to within .05 with 99% confidence, i.e., find a value of n such that Prob [|ˆ p − p| ≤ 0.05] ≥ 0.99. Solution: (S1) Since no a priori estimate of p is available, use p = .5. The bound on the error of estimation is E = .05 and 1 − α = .99. c 2000 by Chapman & Hall/CRC 

2 (2.5758)(.5)(.5) z.005 · pq = = 663.47 E2 .052 (S3) This formula produces a conservative value for the necessary sample size (since no a priori estimate of p is known). A sample size of at least 664 should be used.

(S2) From Table 9.2, n =

9.4

SUMMARY OF COMMON CONFIDENCE INTERVALS

Table 9.3 presents a summary of common confidence intervals for one sample, Table 9.4 is for two samples. For each population parameter, the assumptions and formula for a 100(1 − α)% confidence interval are given. Assumptions (reference)

100(1 − α)% Confidence interval

µ

n large, σ 2 known, or normality, σ 2 known (§9.5.1)

σ x ± zα/2 · √ n

µ

normality, σ 2 unknown (§9.5.2)

s x ± tα/2,n−1 · √ n

Parameter

σ

2

normality (§9.5.3)

(n − 1)s2 (n − 1)s2 , χ2α/2,n−1 χ21−α/2,n−1 -

p

binomial experiment, n large (§9.5.4)

p. ± zα/2 ·

(1)

(2)

p.(1 − p.) n

(3)

(4)

Table 9.3: Summary of common confidence intervals: one sample 9.5

CONFIDENCE INTERVALS: ONE SAMPLE

Let x1 , x2 , . . . , xn be a random sample of size n. 9.5.1 Confidence interval for mean of normal population, known variance Find a 100(1 − α)% confidence interval for the mean µ of a normal population with known variance σ 2 , or Find a 100(1 − α)% confidence interval for the mean µ of a population with known variance σ 2 where n is large. (a) Compute the sample mean x. (b) Determine the critical value zα/2 such that Φ(zα/2 ) = 1 − α/2, where Φ(z) is the standard normalcumulative  distribution function. That is, zα/2 is defined so that Prob Z ≥ zα/2 = α/2. √ (c) Compute the constant √ k = σ zα/2 / n. (Table 9.5 on page 200 has common values of zα/2 / n.) (d) A 100(1 − α)% confidence interval for µ is given by (x − k, x + k). c 2000 by Chapman & Hall/CRC 

Parameter

Assumptions (reference)

µ1 − µ2

normality, independence, σ12 , σ22 known or n1 , n2 large, independence, σ12 , σ22 known, (§9.6.1)

µ1 − µ2

normality, independence, σ12 = σ22 unknown (§9.6.2)

100(1 − α)% Confidence interval + (x1 − x2 ) ± zα/2 ·

(1)

σ2 σ12 + 2 n1 n2

(x1 − x2 ) ± 1 1 α t 2 ,n1 +n2 −2 · sp + n1 n2 (n1 − 1)s21 + (n2 − 1)s22 n1 + n2 − 2 + s2 s21 (x1 − x2 ) ± tα/2,ν · + 2 n1 n2  2  2 s1 s2 + n22 n1 ν ≈ (s2 /n )2 (s2 /n2 )2 1 1 + n2 2 −1 n1 −1 s2p =

µ1 − µ2

µ1 − µ2

σ12 /σ22

normality, independence, σ12 = σ22 unknown (§9.6.3)

normality, n pairs, dependence (§9.6.4) normality, independence (§9.6.5)

sd d ± tα/2,n−1 · √ n 1 s21 · , s22 F α2 ,n1 −1,n2 −1 1 s21 · s22 F1− α2 ,n1 −1,n2 −1

p1 − p2

binomial experiments, n1 , n2 large, independence (§9.6.6)

(2)

(3)

(4)

(. p1 − p.2 ) ± p.1 (1 − p.1 ) p.2 (1 − p.2 ) zα/2 · + n1 n2

(5)

(6)

Table 9.4: Summary of common confidence intervals: two samples 9.5.2 Confidence interval for mean of normal population, unknown variance Find a 100(1 − α)% confidence interval for the mean µ of a normal population with unknown variance σ 2 . (a) Compute the sample mean x and the sample standard deviation s. (b) Determine the critical value tα/2,n−1 such that F (tα/2,n−1 ) = 1 − α/2, where F (t) is the cumulative distribution function for a t distribution with n − 1 degrees of freedom. That is, tα/2,n−1 is defined so that Prob T ≥ tα/2,n−1 = α/2. c 2000 by Chapman & Hall/CRC 

√ zα/2 / n when n

α = 0.05

α = 0.01

2 3 4 5 6 7 8 9 10

8.99 2.48 1.59 1.24 1.05 0.925 0.836 0.769 0.715

45.0 5.73 2.92 2.06 1.65 1.40 1.24 1.12 1.03

√ zα/2 / n when n

α = 0.05

α = 0.01

0.554 0.468 0.413 0.373 0.320 0.284 0.198 0.139 0.088

0.769 0.640 0.559 0.503 0.428 0.379 0.263 0.184 0.116

15 20 25 30 40 50 100 200 500

√ Table 9.5: Common values of zα/2 / n √ (c) Compute the constant k = tα/2,n−1 s/ n. (d) A 100(1 − α)% confidence interval for µ is given by (x − k, x + k). Example 9.50 : A software company conducted a survey on the size of a typical word processing file. For n = 23 randomly selected files, x = 4822 kb and s = 127. Find a 95% confidence interval for the true mean size of word processing files. Solution: (S1) The underlying population, the size of word processing files, is assumed to be normal. The confidence interval for µ is based on a t distribution. (S2) 1 − α = .95 ; α = .05 ; α/2 = .025 ; tα/2,n−1 = t.025,22 = 2.0739 √ (S3) k = (2.0739)(127)/ 23 = 54.92 (S4) A 99% confidence interval for µ: (x − k, x + k) = (4767.08, 4876.92)

9.5.3

Confidence interval for variance of normal population

Find a 100(1 − α)% confidence interval for the variance σ 2 of a normal population. (a) Compute the sample variance s2 . (b) Determine the critical values χ2α/2,n−1 and χ21−α/2,n−1 such that # " # " Prob χ2 ≥ χ2α/2,n−1 = Prob χ2 ≤ χ21−α/2,n−1 = α/2. (c) Compute the constants k1 =

(n − 1)s2 (n − 1)s2 and k = . 2 χ2α/2,n−1 χ21−α/2,n−1

(d) A 100(1 − α)% confidence interval for σ 2 is given by (k1 , k2 ). √ √ (e) A 100(1 − α)% confidence interval for σ is given by ( k1 , k2 ).

c 2000 by Chapman & Hall/CRC 

9.5.4 Confidence interval for the probability of a success in a binomial experiment Find a 100(1 − α)% confidence interval for the probability of a success p in a binomial experiment where n is large. (a) Compute the sample proportion of successes p..   (b) Determine the critical value zα/2 such that Prob Z ≥ zα/2 = α/2. p.(1 − p.) (c) Compute the constant k = zα/2 . n (d) A 100(1 − α)% confidence interval for p is given by (. p − k, p. + k). 9.5.5

Confidence interval for percentiles

Find an approximate 100(1 − α)% confidence interval for the pth percentile, ξp , where n is large. (1) Compute the order statistics {x(1) , x(2) , . . . , x(n) }.   (2) Determine the critical value zα/2 such that Prob Z ≥ zα/2 = α/2. 1 2  (3) Compute the constants k1 = np − zα/2 np(1 − p) and 3 4  k2 = np + zα/2 np(1 − p) . (4) A 100(1 − α)% confidence interval for ξp is given by (x(k1 ) , x(k2 ) ). 9.5.6

Confidence interval for medians

Find an approximate 100(1 − α)% confidence interval for the median µ ˜ where n is large (based on the Wilcoxon one-sample statistic).   (1) Compute the order statistics {w(1) , w(2) , . . . , w(N ) } of the N = n2 = n(n−1) averages (xi + xj )/2, for 1 ≤ i < j ≤ n. 2   (2) Determine the critical value zα/2 such that Prob Z ≥ zα/2 = α/2. 5 6 zα/2 N N (3) Compute the constants k1 = − √ and 2 3n 7 8 zα/2 N N k2 = + √ . 2 3n (4) A 100(1 − α)% confidence interval for µ ˜ is given by (w(k1 ) , w(k2 ) ). (See Table 9.6.) 9.5.6.1

Table of confidence interval for medians

If the n observations {x1 , x2 , . . . , xn } are arranged in ascending order {x(1) , x(2) , . . . , x(n) }, a 100(1 − α)% confidence interval for the median of the population can be found. Table 9.7 lists values of l and u such that the probability the median is between x(l) and x(u) is (1 − α).

c 2000 by Chapman & Hall/CRC 

n 7 8 9 10

α = .05 k1 k2 1 20 2 26 4 32 6 39

α = .01 k1 k2

1

44

11 12 13 14 15

8 11 14 17 21

47 55 64 74 84

2 4 6 9 12

53 62 72 82 93

16 17 18 19 20

26 30 35 41 46

94 106 118 130 144

15 18 22 27 31

105 118 131 144 159

Table 9.6: Confidence interval for median (see section 9.5.6)

9.5.7

Confidence interval for parameter in a Poisson distribution

The probability distribution function for a Poisson random variable is given by (see page 103) e−λ λx for x = 0, 1, 2, . . . (9.3) x! For any value of x and α < 1, lower and upper values of λ (λlower and λupper ) may be determined such that λlower < λupper and f (x; λ) =



x  e−λlower λx

lower

x=0

x!

=

α 2

and

∞  e−λupper λxupper α = x! 2 

(9.4)

x=x

Table 9.8 lists λlower and λupper for α = 0.01 and α = 0.05. For x > 50, λupper and λlower may be approximated by χ21−α,n where 1 − F (χ21−α,n ) = α, and n = 2(x + 1) 2 χ2α,n = where F (χ2α,n ) = α, and n = 2x 2

λupper = λlower

(9.5)

where F (χ2 ) is the cumulative distribution function for a chi–square random variable with n degrees of freedom. Example 9.51 : In a Poisson process, 5 outcomes are observed during a specified time interval. Find a 95% and a 99% confidence interval for the parameter λ in this Poisson process.

c 2000 by Chapman & Hall/CRC 

n 6 7 8 9 10

l 1 1 1 2 2

u 6 7 8 8 9

actual α ≤ 0.05 0.031 0.016 0.008 0.039 0.021

l

u

actual α ≤ 0.01

1 1 1

8 9 10

0.008 0.004 0.002

11 12 13 14 15 16 17 18 19 20

2 3 3 3 4 4 5 5 5 6

10 10 11 12 12 13 13 14 15 15

0.012 0.039 0.022 0.013 0.035 0.021 0.049 0.031 0.019 0.041

1 2 2 2 3 3 3 4 4 4

11 11 12 13 13 14 15 15 16 17

0.001 0.006 0.003 0.002 0.007 0.004 0.002 0.008 0.004 0.003

21 22 23 24 25 26 27 28 29 30

6 6 7 7 8 8 8 9 9 10

16 17 17 18 18 19 20 20 21 21

0.027 0.017 0.035 0.023 0.043 0.029 0.019 0.036 0.024 0.043

5 5 5 6 6 7 7 7 8 8

17 18 19 19 20 20 21 22 22 23

0.007 0.004 0.003 0.007 0.004 0.009 0.006 0.004 0.008 0.005

35 40 50 60 70 75 80 90 100 110 120

12 14 18 22 27 29 31 36 40 45 49

24 27 33 39 44 47 50 55 61 66 72

0.041 0.038 0.033 0.027 0.041 0.037 0.033 0.045 0.035 0.045 0.035

10 12 16 20 24 26 29 33 37 42 46

26 29 35 41 47 50 52 58 64 69 75

0.006 0.006 0.007 0.006 0.006 0.005 0.010 0.008 0.007 0.010 0.008

Table 9.7: Confidence intervals for medians Solution: (S1) Using Table 9.8 with an observed count of 5 and a 95% significance level (α = 0.05), the confidence interval bounds are λlower = 1.6 and λupper = 11.7. (S2) Hence, the probability is .95 that the interval (1.6, 11.7) contains the true value of λ.

c 2000 by Chapman & Hall/CRC 

Significance level Observed α = 0.01 α = 0.05 count λlower λhigher λlower λhigher 0 0.0 5.3 0.0 3.7 1 0.0 7.4 0.0 5.6 2 0.1 9.3 0.2 7.2 3 0.3 11.0 0.6 8.8 4 0.7 12.6 1.1 10.2 5 1.1 14.1 1.6 11.7 6 1.5 15.7 2.2 13.1 7 2.0 17.1 2.8 14.4 8 2.6 18.6 3.5 15.8 9 3.1 20.0 4.1 17.1 10 3.7 21.4 4.8 18.4 11 4.3 22.8 5.5 19.7 12 4.9 24.1 6.2 21.0 13 5.6 25.5 6.9 22.2 14 6.2 26.8 7.7 23.5 15 6.9 28.2 8.4 24.7 16 7.6 29.5 9.1 26.0 17 8.3 30.8 9.9 27.2 18 8.9 32.1 10.7 28.4 19 9.6 33.4 11.4 29.7 20 10.4 34.7 12.2 30.9 21 11.1 35.9 13.0 32.1 22 11.8 37.2 13.8 33.3 23 12.5 38.5 14.6 34.5 24 13.3 39.7 15.4 35.7 25 14.0 41.0 16.2 36.9

Significance level Observed α = 0.01 α = 0.05 count λlower λhigher λlower λhigher 26 14.7 42.3 17.0 38.1 27 15.5 43.5 17.8 39.3 28 16.2 44.7 18.6 40.5 29 17.0 46.0 19.4 41.6 30 17.8 47.2 20.2 42.8 31 18.5 48.4 21.1 44.0 32 19.3 49.7 21.9 45.2 33 20.1 50.9 22.7 46.3 34 20.9 52.1 23.5 47.5 35 21.6 53.3 24.4 48.7 36 22.4 54.5 25.2 49.8 37 23.2 55.7 26.1 51.0 38 24.0 57.0 26.9 52.2 39 24.8 58.2 27.7 53.3 40 25.6 59.4 28.6 54.5 41 26.4 60.6 29.4 55.6 42 27.2 61.8 30.3 56.8 43 28.0 63.0 31.1 57.9 44 28.8 64.1 32.0 59.1 45 29.6 65.3 32.8 60.2 46 30.4 66.5 33.7 61.4 47 31.2 67.7 34.5 62.5 48 32.0 68.9 35.4 63.6 49 32.8 70.1 36.3 64.8 50 33.7 71.3 37.1 65.9

Table 9.8: Confidence limits for the parameter in a Poisson distribution

(S3) Using Table 9.8 with an observed count of 5 and a 99% significance level (α = 0.01), the confidence interval bounds are λlower = 1.1 and λupper = 14.1. (S4) Hence, the probability is .99 that the interval (1.1, 14.1) contains the true value of λ.

c 2000 by Chapman & Hall/CRC 

9.5.8

Confidence interval for parameter in a binomial distribution

The probability distribution function of a binomial random variable is given by (see page 84)   n x f (x; n, p) = p (1 − p)n−x for x = 0, 1, 2, . . . , n (9.6) x For known n, any value of x less than n, and α < 1, lower and upper values of p (plower and pupper ) may be determined such that plower < pupper and 

x 

f (x; n, plower ) =

x=0

α 2

and

n  x=x

f (x; n, pupper ) =

α 2

(9.7)

The tables on pages 206–221 list plower and pupper for α = 0.01 and α = 0.05. Example 9.52 : In a binomial experiment with n = 30, x = 8 successes are observed. Determine a 95% and a 99% confidence interval for the probability of a success p. Solution: (S1) The table on page 210 may be used to construct a 95% confidence interval (α = 0.05). Using this Table with n − x = 22 and x = 8 the bounds on the confidence interval are plower = 0.123 and pupper = 0.459. (S2) Hence, the probability is .95 that the interval (0.123, 0.459) contains the true value of p. (S3) Using the Table on page 218 for a 99% confidence level (α = 0.01), the bounds on the confidence interval are plower = 0.093 and pupper = 0.516. (S4) Hence, the probability is .99 that the interval (0.093, 0.516) contains the true value of p.

c 2000 by Chapman & Hall/CRC 

Confidence limits of proportions (confidence coefficient .95)

x 0

Denominator minus numerator: n − x (Lower limit in italic type, upper limit in roman type) 1 2 3 4 5 6 7 8 9 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.975 0.842 0.708 0.602 0.522 0.459 0.410 0.369 0.336

1

0.013 0.987

0.008 0.906

0.006 0.806

0.005 0.716

0.004 0.641

0.004 0.579

0.003 0.527

0.003 0.482

0.003 0.445

2

0.094 0.992

0.068 0.932

0.053 0.853

0.043 0.777

0.037 0.710

0.032 0.651

0.028 0.600

0.025 0.556

0.023 0.518

3

0.194 0.994

0.147 0.947

0.118 0.882

0.099 0.816

0.085 0.755

0.075 0.701

0.067 0.652

0.060 0.610

0.055 0.572

4

0.284 0.995

0.223 0.957

0.184 0.901

0.157 0.843

0.137 0.788

0.122 0.738

0.109 0.692

0.099 0.651

0.091 0.614

5

0.359 0.996

0.290 0.963

0.245 0.915

0.212 0.863

0.187 0.813

0.168 0.766

0.152 0.723

0.139 0.684

0.128 0.649

6

0.421 0.996

0.349 0.968

0.299 0.925

0.262 0.878

0.234 0.832

0.211 0.789

0.192 0.749

0.177 0.711

0.163 0.677

7

0.473 0.997

0.400 0.972

0.348 0.933

0.308 0.891

0.277 0.848

0.251 0.808

0.230 0.770

0.213 0.734

0.198 0.701

8

0.518 0.997

0.444 0.975

0.390 0.940

0.349 0.901

0.316 0.861

0.289 0.823

0.266 0.787

0.247 0.753

0.230 0.722

9

0.555 0.997

0.482 0.977

0.428 0.945

0.386 0.909

0.351 0.872

0.323 0.837

0.299 0.802

0.278 0.770

0.260 0.740

10

0.587 0.998

0.516 0.979

0.462 0.950

0.419 0.916

0.384 0.882

0.354 0.848

0.329 0.816

0.308 0.785

0.289 0.756

11

0.615 0.998

0.545 0.981

0.492 0.953

0.449 0.922

0.413 0.890

0.383 0.858

0.357 0.827

0.335 0.797

0.315 0.769

12

0.640 0.998

0.572 0.982

0.519 0.957

0.476 0.927

0.440 0.897

0.410 0.867

0.384 0.837

0.361 0.809

0.340 0.782

13

0.661 0.998

0.595 0.983

0.544 0.960

0.501 0.932

0.465 0.903

0.435 0.874

0.408 0.846

0.384 0.819

0.364 0.793

14

0.680 0.998

0.617 0.984

0.566 0.962

0.524 0.936

0.488 0.909

0.457 0.881

0.430 0.854

0.407 0.828

0.385 0.803

15

0.698 0.998

0.636 0.985

0.586 0.964

0.544 0.940

0.509 0.913

0.478 0.887

0.451 0.861

0.427 0.836

0.406 0.812

c 2000 by Chapman & Hall/CRC 

Confidence limits of proportions (confidence coefficient .95)

x 16

Denominator minus numerator: n − x (Lower limit in italic type, upper limit in roman type) 1 2 3 4 5 6 7 8 9 0.713 0.653 0.604 0.563 0.528 0.498 0.471 0.447 0.425 0.998 0.986 0.966 0.943 0.918 0.893 0.868 0.844 0.820

17

0.727 0.999

0.669 0.987

0.621 0.968

0.581 0.945

0.546 0.922

0.516 0.898

0.489 0.874

0.465 0.851

0.443 0.828

18

0.740 0.999

0.683 0.988

0.637 0.970

0.597 0.948

0.563 0.925

0.533 0.902

0.506 0.879

0.482 0.857

0.460 0.835

19

0.751 0.999

0.696 0.988

0.651 0.971

0.612 0.951

0.578 0.929

0.549 0.906

0.522 0.884

0.498 0.862

0.477 0.841

20

0.762 0.999

0.708 0.989

0.664 0.972

0.626 0.953

0.593 0.932

0.564 0.910

0.537 0.889

0.513 0.868

0.492 0.847

22

0.780 0.999

0.730 0.990

0.688 0.975

0.651 0.956

0.619 0.937

0.591 0.917

0.565 0.897

0.541 0.877

0.520 0.858

24

0.796 0.999

0.749 0.991

0.708 0.977

0.673 0.960

0.642 0.942

0.614 0.923

0.589 0.904

0.566 0.885

0.545 0.867

26

0.810 0.999

0.765 0.991

0.727 0.978

0.693 0.962

0.663 0.945

0.636 0.928

0.611 0.910

0.588 0.893

0.567 0.875

28

0.822 0.999

0.779 0.992

0.742 0.980

0.710 0.965

0.681 0.949

0.655 0.932

0.631 0.916

0.608 0.899

0.588 0.882

30

0.833 0.999

0.792 0.992

0.757 0.981

0.726 0.967

0.697 0.952

0.672 0.936

0.648 0.920

0.627 0.904

0.607 0.889

35

0.855 0.999

0.818 0.993

0.786 0.983

0.758 0.971

0.732 0.958

0.708 0.944

0.686 0.930

0.666 0.916

0.647 0.902

40

0.871 0.999

0.838 0.994

0.809 0.985

0.783 0.975

0.759 0.963

0.737 0.951

0.717 0.938

0.698 0.925

0.680 0.912

45

0.885 0.999

0.855 0.995

0.828 0.987

0.804 0.977

0.782 0.967

0.761 0.956

0.742 0.944

0.724 0.933

0.707 0.921

50

0.896 0.999

0.868 0.995

0.843 0.988

0.821 0.979

0.800 0.970

0.781 0.960

0.763 0.949

0.746 0.939

0.730 0.928

60

0.912 1.000

0.888 0.996

0.867 0.990

0.848 0.983

0.830 0.975

0.813 0.966

0.797 0.957

0.781 0.948

0.767 0.939

80

0.933 1.000

0.915 0.997

0.898 0.992

0.883 0.987

0.868 0.981

0.854 0.974

0.841 0.967

0.829 0.960

0.817 0.953

100

0.946 1.000

0.931 0.998

0.917 0.994

0.904 0.989

0.892 0.984

0.881 0.979

0.870 0.973

0.859 0.967

0.849 0.962



1.000 1.000

1.000 1.000

1.000 1.000

1.000 1.000

1.000 1.000

1.000 1.000

1.000 1.000

1.000 1.000

1.000 1.000

c 2000 by Chapman & Hall/CRC 

Confidence limits of proportions (confidence coefficient .95)

x 0

Denominator minus numerator: n − x (Lower limit in italic type, upper limit in roman type) 10 11 12 13 14 15 16 17 18 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.309 0.285 0.265 0.247 0.232 0.218 0.206 0.195 0.185

1

0.002 0.413

0.002 0.385

0.002 0.360

0.002 0.339

0.002 0.320

0.002 0.302

0.002 0.287

0.001 0.273

0.001 0.260

2

0.021 0.484

0.019 0.455

0.018 0.428

0.017 0.405

0.016 0.383

0.015 0.364

0.014 0.347

0.013 0.331

0.012 0.317

3

0.050 0.538

0.047 0.508

0.043 0.481

0.040 0.456

0.038 0.434

0.036 0.414

0.034 0.396

0.032 0.379

0.030 0.363

4

0.084 0.581

0.078 0.551

0.073 0.524

0.068 0.499

0.064 0.476

0.060 0.456

0.057 0.437

0.055 0.419

0.052 0.403

5

0.118 0.616

0.110 0.587

0.103 0.560

0.097 0.535

0.091 0.512

0.087 0.491

0.082 0.472

0.078 0.454

0.075 0.437

6

0.152 0.646

0.142 0.617

0.133 0.590

0.126 0.565

0.119 0.543

0.113 0.522

0.107 0.502

0.102 0.484

0.098 0.467

7

0.184 0.671

0.173 0.643

0.163 0.616

0.154 0.592

0.146 0.570

0.139 0.549

0.132 0.529

0.126 0.511

0.121 0.494

8

0.215 0.692

0.203 0.665

0.191 0.639

0.181 0.616

0.172 0.593

0.164 0.573

0.156 0.553

0.149 0.535

0.143 0.518

9

0.244 0.711

0.231 0.685

0.218 0.660

0.207 0.636

0.197 0.615

0.188 0.594

0.180 0.575

0.172 0.557

0.165 0.540

10

0.272 0.728

0.257 0.702

0.244 0.678

0.232 0.655

0.221 0.634

0.211 0.613

0.202 0.594

0.194 0.576

0.186 0.559

11

0.298 0.743

0.282 0.718

0.268 0.694

0.256 0.672

0.244 0.651

0.234 0.631

0.224 0.612

0.215 0.594

0.207 0.577

12

0.322 0.756

0.306 0.732

0.291 0.709

0.278 0.687

0.266 0.666

0.255 0.647

0.245 0.628

0.235 0.611

0.227 0.594

13

0.345 0.768

0.328 0.744

0.313 0.722

0.299 0.701

0.287 0.680

0.275 0.661

0.264 0.643

0.255 0.626

0.245 0.609

14

0.366 0.779

0.349 0.756

0.334 0.734

0.320 0.713

0.306 0.694

0.294 0.675

0.283 0.657

0.273 0.640

0.264 0.623

15

0.387 0.789

0.369 0.766

0.353 0.745

0.339 0.725

0.325 0.706

0.313 0.687

0.302 0.669

0.291 0.653

0.281 0.637

c 2000 by Chapman & Hall/CRC 

Confidence limits of proportions (confidence coefficient .95)

x 16

Denominator minus numerator: n − x (Lower limit in italic type, upper limit in roman type) 10 11 12 13 14 15 16 17 18 0.406 0.388 0.372 0.357 0.343 0.331 0.319 0.308 0.298 0.798 0.776 0.755 0.736 0.717 0.698 0.681 0.665 0.649

17

0.424 0.806

0.406 0.785

0.389 0.765

0.374 0.745

0.360 0.727

0.347 0.709

0.335 0.692

0.324 0.676

0.314 0.660

18

0.441 0.814

0.423 0.793

0.406 0.773

0.391 0.755

0.377 0.736

0.363 0.719

0.351 0.702

0.340 0.686

0.329 0.671

19

0.457 0.821

0.439 0.801

0.422 0.782

0.406 0.763

0.392 0.745

0.379 0.728

0.367 0.712

0.355 0.696

0.344 0.681

20

0.472 0.827

0.454 0.808

0.437 0.789

0.421 0.771

0.407 0.753

0.393 0.737

0.381 0.721

0.369 0.705

0.358 0.690

22

0.500 0.839

0.482 0.820

0.465 0.803

0.449 0.785

0.435 0.769

0.421 0.753

0.408 0.737

0.396 0.722

0.385 0.707

24

0.525 0.849

0.507 0.831

0.490 0.814

0.475 0.798

0.460 0.782

0.446 0.766

0.433 0.751

0.421 0.737

0.410 0.723

26

0.548 0.858

0.530 0.841

0.513 0.825

0.498 0.809

0.483 0.794

0.469 0.779

0.456 0.764

0.444 0.750

0.433 0.737

28

0.569 0.866

0.551 0.850

0.535 0.834

0.519 0.819

0.504 0.804

0.491 0.790

0.478 0.776

0.465 0.762

0.454 0.749

30

0.588 0.873

0.570 0.858

0.554 0.843

0.539 0.828

0.524 0.814

0.510 0.800

0.497 0.786

0.485 0.773

0.473 0.760

35

0.629 0.888

0.612 0.874

0.596 0.861

0.582 0.847

0.567 0.834

0.554 0.821

0.541 0.809

0.529 0.797

0.517 0.785

40

0.663 0.900

0.647 0.887

0.632 0.875

0.617 0.862

0.603 0.850

0.590 0.839

0.578 0.827

0.566 0.816

0.555 0.805

45

0.691 0.909

0.676 0.898

0.661 0.886

0.647 0.875

0.634 0.864

0.621 0.853

0.609 0.842

0.598 0.831

0.586 0.821

50

0.715 0.917

0.700 0.906

0.686 0.896

0.673 0.885

0.660 0.875

0.648 0.865

0.636 0.855

0.625 0.845

0.614 0.835

60

0.753 0.929

0.740 0.920

0.727 0.911

0.715 0.902

0.703 0.893

0.692 0.883

0.681 0.875

0.670 0.866

0.660 0.857

80

0.805 0.945

0.794 0.938

0.783 0.931

0.773 0.923

0.763 0.916

0.753 0.909

0.743 0.902

0.734 0.894

0.725 0.887

100

0.839 0.956

0.830 0.950

0.820 0.943

0.811 0.937

0.803 0.931

0.794 0.925

0.786 0.919

0.778 0.913

0.770 0.907



1.000 1.000

1.000 1.000

1.000 1.000

1.000 1.000

1.000 1.000

1.000 1.000

1.000 1.000

1.000 1.000

1.000 1.000

c 2000 by Chapman & Hall/CRC 

Confidence limits of proportions (confidence coefficient .95)

x 0

Denominator minus numerator: n − x (Lower limit in italic type, upper limit in roman type) 19 20 22 24 26 28 30 35 40 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.176 0.168 0.154 0.143 0.132 0.123 0.116 0.100 0.088

1

0.001 0.249

0.001 0.238

0.001 0.220

0.001 0.204

0.001 0.190

0.001 0.178

0.001 0.167

0.001 0.145

0.001 0.129

2

0.012 0.304

0.011 0.292

0.010 0.270

0.009 0.251

0.009 0.235

0.008 0.221

0.008 0.208

0.007 0.182

0.006 0.162

3

0.029 0.349

0.028 0.336

0.025 0.312

0.023 0.292

0.022 0.273

0.020 0.258

0.019 0.243

0.017 0.214

0.015 0.191

4

0.049 0.388

0.047 0.374

0.044 0.349

0.040 0.327

0.038 0.307

0.035 0.290

0.033 0.274

0.029 0.242

0.025 0.217

5

0.071 0.422

0.068 0.407

0.063 0.381

0.058 0.358

0.055 0.337

0.051 0.319

0.048 0.303

0.042 0.268

0.037 0.241

6

0.094 0.451

0.090 0.436

0.083 0.409

0.077 0.386

0.072 0.364

0.068 0.345

0.064 0.328

0.056 0.292

0.049 0.263

7

0.116 0.478

0.111 0.463

0.103 0.435

0.096 0.411

0.090 0.389

0.084 0.369

0.080 0.352

0.070 0.314

0.062 0.283

8

0.138 0.502

0.132 0.487

0.123 0.459

0.115 0.434

0.107 0.412

0.101 0.392

0.096 0.373

0.084 0.334

0.075 0.302

9

0.159 0.523

0.153 0.508

0.142 0.480

0.133 0.455

0.125 0.433

0.118 0.412

0.111 0.393

0.098 0.353

0.088 0.320

10

0.179 0.543

0.173 0.528

0.161 0.500

0.151 0.475

0.142 0.452

0.134 0.431

0.127 0.412

0.112 0.371

0.100 0.337

11

0.199 0.561

0.192 0.546

0.180 0.518

0.169 0.493

0.159 0.470

0.150 0.449

0.142 0.430

0.126 0.388

0.113 0.353

12

0.218 0.578

0.211 0.563

0.197 0.535

0.186 0.510

0.175 0.487

0.166 0.465

0.157 0.446

0.139 0.404

0.125 0.368

13

0.237 0.594

0.229 0.579

0.215 0.551

0.202 0.525

0.191 0.502

0.181 0.481

0.172 0.461

0.153 0.418

0.138 0.383

14

0.255 0.608

0.247 0.593

0.231 0.565

0.218 0.540

0.206 0.517

0.196 0.496

0.186 0.476

0.166 0.433

0.150 0.397

15

0.272 0.621

0.263 0.607

0.247 0.579

0.234 0.554

0.221 0.531

0.210 0.509

0.200 0.490

0.179 0.446

0.161 0.410

c 2000 by Chapman & Hall/CRC 

Confidence limits of proportions (confidence coefficient .95)

x 16

Denominator minus numerator: n − x (Lower limit in italic type, upper limit in roman type) 19 20 22 24 26 28 30 35 40 0.288 0.279 0.263 0.249 0.236 0.224 0.214 0.191 0.173 0.633 0.619 0.592 0.567 0.544 0.522 0.503 0.459 0.422

17

0.304 0.645

0.295 0.631

0.278 0.604

0.263 0.579

0.250 0.556

0.238 0.535

0.227 0.515

0.203 0.471

0.184 0.434

18

0.319 0.656

0.310 0.642

0.293 0.615

0.277 0.590

0.263 0.567

0.251 0.546

0.240 0.527

0.215 0.483

0.195 0.445

19

0.334 0.666

0.324 0.652

0.307 0.626

0.291 0.601

0.277 0.579

0.264 0.557

0.252 0.538

0.227 0.494

0.206 0.456

20

0.348 0.676

0.338 0.662

0.320 0.636

0.304 0.612

0.289 0.589

0.276 0.568

0.264 0.548

0.238 0.504

0.217 0.467

22

0.374 0.693

0.364 0.680

0.346 0.654

0.329 0.631

0.314 0.608

0.300 0.587

0.287 0.568

0.260 0.524

0.237 0.487

24

0.399 0.709

0.388 0.696

0.369 0.671

0.352 0.648

0.337 0.626

0.322 0.605

0.309 0.586

0.281 0.543

0.257 0.505

26

0.421 0.723

0.411 0.711

0.392 0.686

0.374 0.663

0.358 0.642

0.343 0.622

0.330 0.603

0.300 0.560

0.276 0.522

28

0.443 0.736

0.432 0.724

0.413 0.700

0.395 0.678

0.378 0.657

0.363 0.637

0.350 0.618

0.319 0.575

0.294 0.538

30

0.462 0.748

0.452 0.736

0.432 0.713

0.414 0.691

0.397 0.670

0.382 0.650

0.368 0.632

0.337 0.590

0.311 0.553

35

0.506 0.773

0.496 0.762

0.476 0.740

0.457 0.719

0.440 0.700

0.425 0.681

0.410 0.663

0.378 0.622

0.351 0.586

40

0.544 0.794

0.533 0.783

0.513 0.763

0.495 0.743

0.478 0.724

0.462 0.706

0.447 0.689

0.414 0.649

0.386 0.614

45

0.576 0.811

0.565 0.801

0.546 0.782

0.528 0.763

0.511 0.745

0.495 0.728

0.480 0.711

0.447 0.673

0.418 0.639

50

0.604 0.825

0.594 0.816

0.575 0.798

0.557 0.780

0.540 0.763

0.524 0.747

0.510 0.731

0.476 0.694

0.447 0.660

60

0.650 0.849

0.641 0.840

0.622 0.824

0.605 0.808

0.589 0.792

0.574 0.777

0.560 0.763

0.526 0.728

0.497 0.697

80

0.717 0.880

0.708 0.873

0.692 0.860

0.676 0.846

0.662 0.833

0.648 0.820

0.634 0.808

0.603 0.778

0.575 0.750

100

0.762 0.901

0.754 0.895

0.740 0.883

0.726 0.872

0.712 0.861

0.700 0.849

0.687 0.839

0.658 0.812

0.632 0.787



1.000 1.000

1.000 1.000

1.000 1.000

1.000 1.000

1.000 1.000

1.000 1.000

1.000 1.000

1.000 1.000

1.000 1.000

c 2000 by Chapman & Hall/CRC 

Confidence limits of proportions (confidence coefficient .95)

x 0

Denominator minus numerator: n − x (Lower limit in italics, upper limit in roman) 45 50 60 80 100 ∞ 0.000 0.000 0.000 0.000 0.000 0.000 0.079 0.071 0.060 0.045 0.036 0.000

1

0.001 0.115

0.001 0.104

0.000 0.088

0.000 0.067

0.000 0.054

0.000 0.000

2

0.005 0.145

0.005 0.132

0.004 0.112

0.003 0.085

0.002 0.069

0.000 0.000

3

0.013 0.172

0.012 0.157

0.010 0.133

0.008 0.102

0.006 0.083

0.000 0.000

4

0.023 0.196

0.021 0.179

0.017 0.152

0.013 0.117

0.011 0.096

0.000 0.000

5

0.033 0.218

0.030 0.200

0.025 0.170

0.019 0.132

0.016 0.108

0.000 0.000

6

0.044 0.239

0.040 0.219

0.034 0.187

0.026 0.146

0.021 0.119

0.000 0.000

7

0.056 0.258

0.051 0.237

0.043 0.203

0.033 0.159

0.027 0.130

0.000 0.000

8

0.067 0.276

0.061 0.254

0.052 0.219

0.040 0.171

0.033 0.141

0.000 0.000

9

0.079 0.293

0.072 0.270

0.061 0.233

0.047 0.183

0.038 0.151

0.000 0.000

10

0.091 0.309

0.083 0.285

0.071 0.247

0.055 0.195

0.044 0.161

0.000 0.000

11

0.102 0.324

0.094 0.300

0.080 0.260

0.062 0.206

0.050 0.170

0.000 0.000

12

0.114 0.339

0.104 0.314

0.089 0.273

0.069 0.217

0.057 0.180

0.000 0.000

13

0.125 0.353

0.115 0.327

0.098 0.285

0.077 0.227

0.063 0.189

0.000 0.000

14

0.136 0.366

0.125 0.340

0.107 0.297

0.084 0.237

0.069 0.197

0.000 0.000

15

0.147 0.379

0.135 0.352

0.117 0.308

0.091 0.247

0.075 0.206

0.000 0.000

c 2000 by Chapman & Hall/CRC 

Confidence limits of proportions (confidence coefficient .95)

x 16

Denominator minus numerator: n − x (Lower limit in italics, upper limit in roman) 45 50 60 80 100 ∞ 0.158 0.145 0.125 0.098 0.081 0.000 0.391 0.364 0.319 0.257 0.214 0.000

17

0.169 0.402

0.155 0.375

0.134 0.330

0.106 0.266

0.087 0.222

0.000 0.000

18

0.179 0.414

0.165 0.386

0.143 0.340

0.113 0.275

0.093 0.230

0.000 0.000

19

0.189 0.424

0.175 0.396

0.151 0.350

0.120 0.283

0.099 0.238

0.000 0.000

20

0.199 0.435

0.184 0.406

0.160 0.359

0.127 0.292

0.105 0.246

0.000 0.000

22

0.218 0.454

0.202 0.425

0.176 0.378

0.140 0.308

0.117 0.260

0.000 0.000

24

0.237 0.472

0.220 0.443

0.192 0.395

0.154 0.324

0.128 0.274

0.000 0.000

26

0.255 0.489

0.237 0.460

0.208 0.411

0.167 0.338

0.139 0.288

0.000 0.000

28

0.272 0.505

0.253 0.476

0.223 0.426

0.180 0.352

0.151 0.300

0.000 0.000

30

0.289 0.520

0.269 0.490

0.237 0.440

0.192 0.366

0.161 0.313

0.000 0.000

35

0.327 0.553

0.306 0.524

0.272 0.474

0.222 0.397

0.188 0.342

0.000 0.000

40

0.361 0.582

0.340 0.553

0.303 0.503

0.250 0.425

0.213 0.368

0.000 0.000

45

0.393 0.607

0.370 0.579

0.332 0.529

0.276 0.451

0.236 0.392

0.000 0.000

50

0.421 0.630

0.398 0.602

0.359 0.552

0.301 0.474

0.259 0.415

0.000 0.000

60

0.471 0.668

0.448 0.641

0.407 0.593

0.345 0.515

0.300 0.455

0.000 0.000

80

0.549 0.724

0.526 0.699

0.485 0.655

0.420 0.580

0.371 0.520

0.000 0.000

100

0.608 0.764

0.585 0.741

0.545 0.700

0.480 0.629

0.429 0.571

0.000 0.000



1.000 1.000

1.000 1.000

1.000 1.000

1.000 1.000

1.000 1.000

1.000 1.000

c 2000 by Chapman & Hall/CRC 

Confidence limits of proportions (confidence coefficient .99)

x 0

Denominator minus numerator: n − x (Lower limit in italic type, upper limit in roman type) 1 2 3 4 5 6 7 8 9 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.995 0.929 0.829 0.734 0.653 0.586 0.531 0.484 0.445

1

0.003 0.997

0.002 0.959

0.001 0.889

0.001 0.815

0.001 0.746

0.001 0.685

0.001 0.632

0.001 0.585

0.001 0.544

2

0.041 0.998

0.029 0.971

0.023 0.917

0.019 0.856

0.016 0.797

0.014 0.742

0.012 0.693

0.011 0.648

0.010 0.608

3

0.111 0.999

0.083 0.977

0.066 0.934

0.055 0.882

0.047 0.830

0.042 0.781

0.037 0.735

0.033 0.693

0.030 0.655

4

0.185 0.999

0.144 0.981

0.118 0.945

0.100 0.900

0.087 0.854

0.077 0.809

0.069 0.767

0.062 0.727

0.057 0.691

5

0.254 0.999

0.203 0.984

0.170 0.953

0.146 0.913

0.128 0.872

0.114 0.831

0.103 0.791

0.094 0.755

0.087 0.720

6

0.315 0.999

0.258 0.986

0.219 0.958

0.191 0.923

0.169 0.886

0.152 0.848

0.138 0.811

0.127 0.777

0.117 0.744

7

0.368 0.999

0.307 0.988

0.265 0.963

0.233 0.931

0.209 0.897

0.189 0.862

0.172 0.828

0.159 0.795

0.147 0.764

8

0.415 0.999

0.352 0.989

0.307 0.967

0.273 0.938

0.245 0.906

0.223 0.873

0.205 0.841

0.190 0.810

0.176 0.781

9

0.456 0.999

0.392 0.990

0.345 0.970

0.309 0.943

0.280 0.913

0.256 0.883

0.236 0.853

0.219 0.824

0.205 0.795

10

0.491 1.000

0.427 0.991

0.379 0.972

0.342 0.947

0.312 0.920

0.287 0.891

0.266 0.863

0.247 0.835

0.232 0.808

11

0.523 1.000

0.459 0.992

0.411 0.974

0.373 0.951

0.341 0.925

0.315 0.899

0.293 0.872

0.274 0.845

0.257 0.819

12

0.551 1.000

0.488 0.992

0.440 0.976

0.401 0.955

0.369 0.930

0.342 0.905

0.319 0.879

0.299 0.854

0.282 0.829

13

0.576 1.000

0.514 0.993

0.466 0.978

0.427 0.957

0.394 0.935

0.367 0.910

0.343 0.886

0.323 0.862

0.305 0.838

14

0.598 1.000

0.537 0.993

0.490 0.979

0.451 0.960

0.418 0.938

0.390 0.915

0.366 0.892

0.345 0.869

0.326 0.846

15

0.619 1.000

0.559 0.994

0.512 0.980

0.473 0.962

0.440 0.942

0.412 0.920

0.388 0.898

0.366 0.875

0.347 0.854

c 2000 by Chapman & Hall/CRC 

Confidence limits of proportions (confidence coefficient .99)

x 16

Denominator minus numerator: n − x (Lower limit in italic type, upper limit in roman type) 1 2 3 4 5 6 7 8 9 0.637 0.578 0.532 0.493 0.461 0.433 0.408 0.386 0.367 1.000 0.994 0.981 0.964 0.945 0.924 0.903 0.881 0.860

17

0.654 1.000

0.596 0.994

0.550 0.982

0.512 0.966

0.480 0.947

0.452 0.927

0.427 0.907

0.405 0.887

0.385 0.866

18

0.669 1.000

0.613 0.995

0.568 0.983

0.530 0.968

0.498 0.950

0.470 0.931

0.445 0.911

0.422 0.891

0.403 0.872

19

0.683 1.000

0.628 0.995

0.584 0.984

0.547 0.969

0.515 0.952

0.486 0.934

0.461 0.915

0.439 0.896

0.419 0.877

20

0.696 1.000

0.642 0.995

0.599 0.985

0.562 0.971

0.530 0.954

0.502 0.936

0.477 0.918

0.455 0.900

0.435 0.881

22

0.719 1.000

0.668 0.996

0.626 0.986

0.590 0.973

0.559 0.958

0.531 0.941

0.507 0.924

0.484 0.907

0.464 0.890

24

0.738 1.000

0.690 0.996

0.649 0.987

0.615 0.975

0.584 0.961

0.557 0.945

0.533 0.930

0.511 0.913

0.491 0.897

26

0.755 1.000

0.709 0.996

0.670 0.988

0.637 0.977

0.607 0.963

0.581 0.949

0.557 0.934

0.535 0.919

0.515 0.904

28

0.770 1.000

0.726 0.997

0.689 0.989

0.656 0.978

0.627 0.966

0.602 0.952

0.578 0.938

0.557 0.924

0.537 0.909

30

0.784 1.000

0.741 0.997

0.705 0.989

0.674 0.980

0.646 0.968

0.621 0.955

0.597 0.942

0.576 0.928

0.557 0.914

35

0.811 1.000

0.773 0.997

0.740 0.991

0.711 0.982

0.685 0.972

0.661 0.961

0.639 0.949

0.619 0.937

0.600 0.924

40

0.832 1.000

0.797 0.997

0.767 0.992

0.741 0.984

0.716 0.975

0.694 0.965

0.673 0.955

0.654 0.944

0.636 0.933

45

0.849 1.000

0.817 0.998

0.789 0.993

0.765 0.986

0.742 0.978

0.721 0.969

0.702 0.959

0.683 0.949

0.666 0.939

50

0.863 1.000

0.834 0.998

0.808 0.993

0.785 0.987

0.763 0.980

0.744 0.972

0.725 0.963

0.708 0.954

0.692 0.945

60

0.884 1.000

0.859 0.998

0.836 0.995

0.816 0.989

0.797 0.983

0.780 0.976

0.763 0.969

0.747 0.961

0.733 0.953

80

0.912 1.000

0.892 0.999

0.874 0.996

0.858 0.992

0.842 0.987

0.828 0.982

0.814 0.976

0.801 0.970

0.789 0.964

100

0.929 1.000

0.912 0.999

0.898 0.997

0.884 0.993

0.871 0.990

0.859 0.985

0.847 0.981

0.836 0.976

0.826 0.971



1.000 1.000

1.000 1.000

1.000 1.000

1.000 1.000

1.000 1.000

1.000 1.000

1.000 1.000

1.000 1.000

1.000 1.000

c 2000 by Chapman & Hall/CRC 

Confidence limits of proportions (confidence coefficient .99)

x 0

Denominator minus numerator: n − x (Lower limit in italic type, upper limit in roman type) 10 11 12 13 14 15 16 17 18 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.411 0.382 0.357 0.335 0.315 0.298 0.282 0.268 0.255

1

0.000 0.509

0.000 0.477

0.000 0.449

0.000 0.424

0.000 0.402

0.000 0.381

0.000 0.363

0.000 0.346

0.000 0.331

2

0.009 0.573

0.008 0.541

0.008 0.512

0.007 0.486

0.007 0.463

0.006 0.441

0.006 0.422

0.006 0.404

0.005 0.387

3

0.028 0.621

0.026 0.589

0.024 0.560

0.022 0.534

0.021 0.510

0.020 0.488

0.019 0.468

0.018 0.450

0.017 0.432

4

0.053 0.658

0.049 0.627

0.045 0.599

0.043 0.573

0.040 0.549

0.038 0.527

0.036 0.507

0.034 0.488

0.032 0.470

5

0.080 0.688

0.075 0.659

0.070 0.631

0.065 0.606

0.062 0.582

0.058 0.560

0.055 0.539

0.053 0.520

0.050 0.502

6

0.109 0.713

0.101 0.685

0.095 0.658

0.090 0.633

0.085 0.610

0.080 0.588

0.076 0.567

0.073 0.548

0.069 0.530

7

0.137 0.734

0.128 0.707

0.121 0.681

0.114 0.657

0.108 0.634

0.102 0.612

0.097 0.592

0.093 0.573

0.089 0.555

8

0.165 0.753

0.155 0.726

0.146 0.701

0.138 0.677

0.131 0.655

0.125 0.634

0.119 0.614

0.113 0.595

0.109 0.578

9

0.192 0.768

0.181 0.743

0.171 0.718

0.162 0.695

0.154 0.674

0.146 0.653

0.140 0.633

0.134 0.615

0.128 0.597

10

0.218 0.782

0.206 0.758

0.195 0.734

0.185 0.712

0.176 0.690

0.168 0.670

0.161 0.651

0.154 0.633

0.148 0.616

11

0.242 0.794

0.229 0.771

0.218 0.748

0.207 0.726

0.197 0.706

0.189 0.686

0.181 0.667

0.173 0.649

0.167 0.632

12

0.266 0.805

0.252 0.782

0.240 0.760

0.228 0.739

0.218 0.719

0.209 0.700

0.200 0.681

0.192 0.664

0.185 0.647

13

0.288 0.815

0.274 0.793

0.261 0.772

0.249 0.751

0.238 0.732

0.228 0.713

0.219 0.695

0.211 0.677

0.203 0.661

14

0.310 0.824

0.294 0.803

0.281 0.782

0.268 0.762

0.257 0.743

0.247 0.724

0.237 0.707

0.228 0.690

0.220 0.674

15

0.330 0.832

0.314 0.811

0.300 0.791

0.287 0.772

0.276 0.753

0.265 0.735

0.255 0.718

0.246 0.701

0.237 0.685

c 2000 by Chapman & Hall/CRC 

Confidence limits of proportions (confidence coefficient .99)

x 16

Denominator minus numerator: n − x (Lower limit in italic type, upper limit in roman type) 10 11 12 13 14 15 16 17 18 0.349 0.333 0.319 0.305 0.293 0.282 0.272 0.262 0.253 0.839 0.819 0.800 0.781 0.763 0.745 0.728 0.712 0.696

17

0.367 0.846

0.351 0.827

0.336 0.808

0.323 0.789

0.310 0.772

0.299 0.754

0.288 0.738

0.278 0.722

0.269 0.706

18

0.384 0.852

0.368 0.833

0.353 0.815

0.339 0.797

0.326 0.780

0.315 0.763

0.304 0.747

0.294 0.731

0.284 0.716

19

0.401 0.858

0.384 0.840

0.369 0.822

0.355 0.804

0.342 0.787

0.330 0.771

0.319 0.755

0.308 0.740

0.299 0.725

20

0.417 0.863

0.400 0.845

0.384 0.828

0.370 0.811

0.357 0.794

0.345 0.778

0.333 0.763

0.323 0.748

0.313 0.733

22

0.446 0.873

0.429 0.856

0.413 0.839

0.399 0.823

0.385 0.807

0.372 0.792

0.361 0.777

0.350 0.763

0.340 0.748

24

0.472 0.881

0.455 0.865

0.439 0.849

0.425 0.834

0.411 0.819

0.398 0.804

0.386 0.789

0.375 0.776

0.364 0.762

26

0.496 0.888

0.479 0.873

0.463 0.858

0.449 0.843

0.435 0.829

0.422 0.815

0.410 0.801

0.398 0.787

0.388 0.774

28

0.518 0.894

0.502 0.880

0.486 0.866

0.471 0.852

0.457 0.838

0.444 0.824

0.432 0.811

0.420 0.798

0.409 0.785

30

0.539 0.900

0.522 0.886

0.506 0.873

0.492 0.859

0.478 0.846

0.465 0.833

0.452 0.820

0.441 0.807

0.430 0.795

35

0.583 0.912

0.567 0.900

0.551 0.887

0.537 0.875

0.523 0.863

0.510 0.851

0.498 0.839

0.486 0.828

0.475 0.816

40

0.620 0.921

0.604 0.910

0.589 0.899

0.575 0.888

0.561 0.877

0.549 0.866

0.536 0.855

0.525 0.844

0.514 0.833

45

0.650 0.929

0.635 0.919

0.621 0.908

0.607 0.898

0.594 0.888

0.582 0.878

0.570 0.867

0.558 0.857

0.548 0.848

50

0.676 0.935

0.662 0.926

0.648 0.916

0.635 0.906

0.622 0.897

0.610 0.888

0.599 0.878

0.587 0.869

0.577 0.860

60

0.719 0.945

0.705 0.937

0.692 0.928

0.680 0.920

0.668 0.912

0.657 0.903

0.646 0.895

0.636 0.887

0.626 0.879

80

0.777 0.957

0.766 0.951

0.755 0.944

0.744 0.938

0.734 0.931

0.724 0.924

0.714 0.918

0.705 0.911

0.696 0.905

100

0.815 0.965

0.806 0.960

0.796 0.955

0.787 0.949

0.778 0.944

0.769 0.938

0.761 0.933

0.752 0.927

0.744 0.921



1.000 1.000

1.000 1.000

1.000 1.000

1.000 1.000

1.000 1.000

1.000 1.000

1.000 1.000

1.000 1.000

1.000 1.000

c 2000 by Chapman & Hall/CRC 

Confidence limits of proportions (confidence coefficient .99)

x 0

Denominator minus numerator: n − x (Lower limit in italic type, upper limit in roman type) 19 20 22 24 26 28 30 35 40 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.000 0.243 0.233 0.214 0.198 0.184 0.172 0.162 0.140 0.124

1

0.000 0.317

0.000 0.304

0.000 0.281

0.000 0.262

0.000 0.245

0.000 0.230

0.000 0.216

0.000 0.189

0.000 0.168

2

0.005 0.372

0.005 0.358

0.004 0.332

0.004 0.310

0.004 0.291

0.003 0.274

0.003 0.259

0.003 0.227

0.003 0.203

3

0.016 0.416

0.015 0.401

0.014 0.374

0.013 0.351

0.012 0.330

0.011 0.311

0.011 0.295

0.009 0.260

0.008 0.233

4

0.031 0.453

0.029 0.438

0.027 0.410

0.025 0.385

0.023 0.363

0.022 0.344

0.020 0.326

0.018 0.289

0.016 0.259

5

0.048 0.485

0.046 0.470

0.042 0.441

0.039 0.416

0.037 0.393

0.034 0.373

0.032 0.354

0.028 0.315

0.025 0.284

6

0.066 0.514

0.064 0.498

0.059 0.469

0.055 0.443

0.051 0.419

0.048 0.398

0.045 0.379

0.039 0.339

0.035 0.306

7

0.085 0.539

0.082 0.523

0.076 0.493

0.070 0.467

0.066 0.443

0.062 0.422

0.058 0.403

0.051 0.361

0.045 0.327

8

0.104 0.561

0.100 0.545

0.093 0.516

0.087 0.489

0.081 0.465

0.076 0.443

0.072 0.424

0.063 0.381

0.056 0.346

9

0.123 0.581

0.119 0.565

0.110 0.536

0.103 0.509

0.096 0.485

0.091 0.463

0.086 0.443

0.076 0.400

0.067 0.364

10

0.142 0.599

0.137 0.583

0.127 0.554

0.119 0.528

0.112 0.504

0.106 0.482

0.100 0.461

0.088 0.417

0.079 0.380

11

0.160 0.616

0.155 0.600

0.144 0.571

0.135 0.545

0.127 0.521

0.120 0.498

0.114 0.478

0.100 0.433

0.090 0.396

12

0.178 0.631

0.172 0.616

0.161 0.587

0.151 0.561

0.142 0.537

0.134 0.514

0.127 0.494

0.113 0.449

0.101 0.411

13

0.196 0.645

0.189 0.630

0.177 0.601

0.166 0.575

0.157 0.551

0.148 0.529

0.141 0.508

0.125 0.463

0.112 0.425

14

0.213 0.658

0.206 0.643

0.193 0.615

0.181 0.589

0.171 0.565

0.162 0.543

0.154 0.522

0.137 0.477

0.123 0.439

15

0.229 0.670

0.222 0.655

0.208 0.628

0.196 0.602

0.185 0.578

0.176 0.556

0.167 0.535

0.149 0.490

0.134 0.451

c 2000 by Chapman & Hall/CRC 

Confidence limits of proportions (confidence coefficient .99)

x 16

Denominator minus numerator: n − x (Lower limit in italic type, upper limit in roman type) 19 20 22 24 26 28 30 35 40 0.245 0.237 0.223 0.211 0.199 0.189 0.180 0.161 0.145 0.681 0.667 0.639 0.614 0.590 0.568 0.548 0.502 0.464

17

0.260 0.692

0.252 0.677

0.237 0.650

0.224 0.625

0.213 0.602

0.202 0.580

0.193 0.559

0.172 0.514

0.156 0.475

18

0.275 0.701

0.267 0.687

0.252 0.660

0.238 0.636

0.226 0.612

0.215 0.591

0.205 0.570

0.184 0.525

0.167 0.486

19

0.289 0.711

0.281 0.697

0.265 0.670

0.251 0.646

0.239 0.623

0.227 0.601

0.217 0.581

0.195 0.536

0.177 0.497

20

0.303 0.719

0.295 0.705

0.279 0.679

0.264 0.655

0.251 0.632

0.239 0.611

0.229 0.591

0.206 0.546

0.187 0.507

22

0.330 0.735

0.321 0.721

0.304 0.696

0.289 0.672

0.275 0.650

0.263 0.629

0.251 0.609

0.227 0.565

0.207 0.526

24

0.354 0.749

0.345 0.736

0.328 0.711

0.312 0.688

0.298 0.666

0.285 0.646

0.273 0.626

0.247 0.582

0.226 0.543

26

0.377 0.761

0.368 0.749

0.350 0.725

0.334 0.702

0.319 0.681

0.306 0.661

0.293 0.642

0.266 0.598

0.244 0.560

28

0.399 0.773

0.389 0.761

0.371 0.737

0.354 0.715

0.339 0.694

0.325 0.675

0.313 0.656

0.285 0.613

0.262 0.575

30

0.419 0.783

0.409 0.771

0.391 0.749

0.374 0.727

0.358 0.707

0.344 0.687

0.331 0.669

0.303 0.626

0.279 0.589

35

0.464 0.805

0.454 0.794

0.435 0.773

0.418 0.753

0.402 0.734

0.387 0.715

0.374 0.697

0.343 0.657

0.318 0.620

40

0.503 0.823

0.493 0.813

0.474 0.793

0.457 0.774

0.440 0.756

0.425 0.738

0.411 0.721

0.380 0.682

0.353 0.647

45

0.537 0.838

0.527 0.829

0.508 0.810

0.491 0.792

0.474 0.775

0.459 0.758

0.445 0.742

0.413 0.704

0.386 0.670

50

0.567 0.851

0.557 0.842

0.538 0.824

0.521 0.807

0.505 0.791

0.489 0.775

0.475 0.759

0.443 0.723

0.415 0.690

60

0.616 0.871

0.607 0.863

0.589 0.847

0.572 0.832

0.556 0.817

0.541 0.802

0.527 0.788

0.495 0.755

0.466 0.724

80

0.687 0.898

0.679 0.892

0.662 0.879

0.647 0.866

0.632 0.853

0.618 0.841

0.605 0.829

0.574 0.800

0.547 0.773

100

0.736 0.916

0.729 0.910

0.714 0.899

0.700 0.888

0.686 0.878

0.674 0.867

0.661 0.857

0.632 0.832

0.606 0.807



1.000 1.000

1.000 1.000

1.000 1.000

1.000 1.000

1.000 1.000

1.000 1.000

1.000 1.000

1.000 1.000

1.000 1.000

c 2000 by Chapman & Hall/CRC 

Confidence limits of proportions (confidence coefficient .99)

x 0

Denominator minus numerator: n − x (Lower limit in italics, upper limit in roman) 45 50 60 80 100 ∞ 0.000 0.000 0.000 0.000 0.000 0.000 0.111 0.101 0.085 0.064 0.052 0.000

1

0.000 0.151

0.000 0.137

0.000 0.116

0.000 0.088

0.000 0.071

0.000 0.000

2

0.002 0.183

0.002 0.166

0.002 0.141

0.001 0.108

0.001 0.088

0.000 0.000

3

0.007 0.211

0.007 0.192

0.005 0.164

0.004 0.126

0.003 0.102

0.000 0.000

4

0.014 0.235

0.013 0.215

0.011 0.184

0.008 0.142

0.007 0.116

0.000 0.000

5

0.022 0.258

0.020 0.237

0.017 0.203

0.013 0.158

0.010 0.129

0.000 0.000

6

0.031 0.279

0.028 0.256

0.024 0.220

0.018 0.172

0.015 0.141

0.000 0.000

7

0.041 0.298

0.037 0.275

0.031 0.237

0.024 0.186

0.019 0.153

0.000 0.000

8

0.051 0.317

0.046 0.292

0.039 0.253

0.030 0.199

0.024 0.164

0.000 0.000

9

0.061 0.334

0.055 0.308

0.047 0.267

0.036 0.211

0.029 0.174

0.000 0.000

10

0.071 0.350

0.065 0.324

0.055 0.281

0.043 0.223

0.035 0.185

0.000 0.000

11

0.081 0.365

0.074 0.338

0.063 0.295

0.049 0.234

0.040 0.194

0.000 0.000

12

0.092 0.379

0.084 0.352

0.072 0.308

0.056 0.245

0.045 0.204

0.000 0.000

13

0.102 0.393

0.094 0.365

0.080 0.320

0.062 0.256

0.051 0.213

0.000 0.000

14

0.112 0.406

0.103 0.378

0.088 0.332

0.069 0.266

0.056 0.222

0.000 0.000

15

0.122 0.418

0.112 0.390

0.097 0.343

0.076 0.276

0.062 0.231

0.000 0.000

16

0.133 0.430

0.122 0.401

0.105 0.354

0.082 0.286

0.067 0.239

0.000 0.000

17

0.143 0.442

0.131 0.413

0.113 0.364

0.089 0.295

0.073 0.248

0.000 0.000

c 2000 by Chapman & Hall/CRC 

Confidence limits of proportions (confidence coefficient .99)

x 18

Denominator minus numerator: n − x (Lower limit in italics, upper limit in roman) 45 50 60 80 100 ∞ 0.152 0.140 0.121 0.095 0.079 0.000 0.452 0.423 0.374 0.304 0.256 0.000

19

0.162 0.463

0.149 0.433

0.129 0.384

0.102 0.313

0.084 0.264

0.000 0.000

20

0.171 0.473

0.158 0.443

0.137 0.393

0.108 0.321

0.090 0.271

0.000 0.000

22

0.190 0.492

0.176 0.462

0.153 0.411

0.121 0.338

0.101 0.286

0.000 0.000

24

0.208 0.509

0.193 0.479

0.168 0.428

0.134 0.353

0.112 0.300

0.000 0.000

26

0.225 0.526

0.209 0.495

0.183 0.444

0.147 0.368

0.122 0.314

0.000 0.000

28

0.242 0.541

0.225 0.511

0.198 0.459

0.159 0.382

0.133 0.326

0.000 0.000

30

0.258 0.555

0.241 0.525

0.212 0.473

0.171 0.395

0.143 0.339

0.000 0.000

35

0.296 0.587

0.277 0.557

0.245 0.505

0.200 0.426

0.168 0.368

0.000 0.000

40

0.330 0.614

0.310 0.585

0.276 0.534

0.227 0.453

0.193 0.394

0.000 0.000

45

0.362 0.638

0.341 0.609

0.305 0.559

0.253 0.478

0.216 0.418

0.000 0.000

50

0.391 0.659

0.369 0.631

0.332 0.581

0.277 0.501

0.238 0.440

0.000 0.000

60

0.441 0.695

0.419 0.668

0.380 0.620

0.321 0.541

0.278 0.479

0.000 0.000

80

0.522 0.747

0.499 0.723

0.459 0.679

0.396 0.604

0.349 0.543

0.000 0.000

100

0.582 0.784

0.560 0.762

0.521 0.722

0.457 0.651

0.407 0.593

0.000 0.000



1.000 1.000

1.000 1.000

1.000 1.000

1.000 1.000

1.000 1.000

1.000 1.000

c 2000 by Chapman & Hall/CRC 

9.6

CONFIDENCE INTERVALS: TWO SAMPLES

Let x1 , x2 , . . . , xn1 be a random sample of size n1 from population 1 and y1 , y2 , . . . , yn2 a random sample of size n2 from population 2. 9.6.1

Confidence interval for difference in means, known variances

Find a 100(1 − α)% confidence interval for the difference in means µ1 − µ2 if the populations are normal, σ12 and σ22 are known, and the samples are independent, or Find a 100(1 − α)% confidence interval for the difference in means µ1 − µ2 if n1 and n2 are large, σ12 and σ22 are known, and the samples are independent. (a) Compute the sample means x1 and x2 .   (b) Determine the critical value zα/2 such that Prob Z ≥ zα/2 = α/2. + σ12 σ2 (c) Compute the constant k = zα/2 + 2. n1 n2 (d) A 100(1 − α)% confidence interval for µ1 − µ2 is given by ((x1 − x2 ) − k, (x1 − x2 ) + k). 9.6.2 Confidence interval for difference in means, equal unknown variances Find a 100(1 − α)% confidence interval for the difference in means µ1 − µ2 if the populations are normal, the samples are independent, and the variances are unknown but assumed equal (σ12 = σ22 = σ 2 ). (a) Compute the sample means x1 and x2 , the sample variances s21 and s22 , and the pooled estimate of the common variance σ 2 s2p =

(n1 − 1)s21 + (n2 − 1)s22 . n1 + n2 − 2

(9.8)

(b) Determine the critical value tα/2,n1 +n2 −2 such that   Prob T ≥ tα/2,n1 +n2 −2 = α/2. 1 1 (c) Compute the constant k = tα/2,n1 +n2 −2 · sp + , n1 n2 (d) A 100(1 − α)% confidence interval for µ1 − µ2 is given by ((x1 − x2 ) − k, (x1 − x2 ) + k). 9.6.3 Confidence interval for difference in means, unequal unknown variances Find an approximate 100(1 − α)% confidence interval for the difference in means µ1 − µ2 if the populations are normal, the samples are independent, and the variances are unknown and unequal.

c 2000 by Chapman & Hall/CRC 

(a) Compute the sample means x1 and x2 , the sample variances s21 and s22 , and the approximate degrees of freedom  2 2 s1 s22 + n1 n2 ν = (s2 /n )2 . (9.9) (s22 /n2 )2 1 1 + n1 −1 n2 −1 Round ν to the nearest integer.   (b) Determine the critical value tα/2,ν such that Prob T ≥ tα/2,ν = α/2. + s21 s2 (c) Compute the constant k = tα/2,ν · + 2. n1 n2 (d) An approximate 100(1 − α)% confidence interval for µ1 − µ2 is given by ((x1 − x2 ) − k, (x1 − x2 ) + k). 9.6.4 Confidence interval for difference in means, paired observations Find a 100(1 − α)% confidence interval for the difference in means µ1 − µ2 if the populations are normal and the observations are paired (dependent). (a) Compute the paired differences x1 − y1 , x2 − y2 , . . . , xn − yn , the sample mean for the differences d, and the sample variance for the differences s2d .   (b) Determine the critical value tα/2,n−1 such that Prob T ≥ tα/2,n−1 = α/2. √ (c) Compute the constant k = tα/2,n−1 sd / n. (d) A 100(1 − α)% confidence interval for µ1 − µ2 is given by (d − k, d + k). 9.6.5

Ratio of variances

Find a 100(1 − α)% confidence interval for the ratio of variances σ12 /σ22 if the populations are normal and the samples are independent. (a) Compute the sample variances s21 and s22 . (b) Determine the critical values Fα/2,n  1 −1,n2 −1 and  F1−α/2,n such that Prob F ≥ Fα/2,n1 −1,n2 −1 = 1 − α/2 and  1 −2,n2 −1  Prob F ≤ F1−α/2,n1 −2,n2 −1 = α/2. (c) Compute the constants k1 = 1/Fα/2,n1 −1,n2 −1 and k2 = 1/F1−α/2,n1 −1,n2 −1 .   2 s1 s21 σ12 (d) A 100(1 − α)% confidence interval for σ2 is k1 , 2 k2 . 2 s22 s2 9.6.6

Difference in success probabilities

Find a 100(1 − α)% confidence interval for the difference in success probabilities p1 − p2 if the samples are from a binomial experiment, the sample sizes are large, and the samples are independent. c 2000 by Chapman & Hall/CRC 

(a) Compute the proportion of success for each sample p.1 and p.2 .   (b) Determine the critical value zα/2 such that Prob Z ≥ zα/2 = α/2. + p.1 (1 − p.1 ) p.2 (1 − p.2 ) (c) Compute the constant k = zα/2 + . n1 n2 (d) A 100(1 − α)% confidence interval for p.1 − p.2 is given by p1 − p.2 ) + k). ((. p1 − p.2 ) − k, (. Example 9.53 : A researcher would like to compare the quality of incoming students at a public college and a private college. One measure of the strength of a class is the proportion of students who took an Advanced Placement test in High School. Random samples were selected at each institution. For the public college, p.1 = 190/500 = .38 and for the private college, p.2 = 215/500 = .43. Find a 95% confidence interval for the difference in proportions of students who took an AP test in high school. Solution: (S1) The samples are assumed to be from binomial experiments and independent, and the samples are large. (S2) 1 − α = .95 ; α = .05 ; α/2 = .025 ; zα/2 = z.025 = 1.96 (.43)(.57) (.38)(.62) + = .0608 (S3) k = (1.96) 500 500 p1 −. p1 −. p2 )−k, (. p2 )+k) = (−.1108, .0108) (S4) A 95% confidence interval for p1 −p2 : ((.

9.6.7

Difference in medians

The following technique, based on the Mann–Whitney–Wilcoxon procedure, may be used to find an approximate 100(1 − α)% confidence interval for the difference in medians, µ ˜1 − µ ˜2 . Assume the sample sizes are large and the samples are independent. (1) Compute the order statistics {w(1) , w(2) , . . . , w(N ) } for the N = n1 n2 differences xi − yj , for 1 ≤ i ≤ n1 and 1 ≤ j ≤ n2 .   (2) Determine the critical value zα/2 such that Prob Z ≥ zα/2 = α/2. (3) Compute 9 the constants : n1 n2 (n1 + n2 + 1) n 1 n2 k1 = + 0.5 − zα/2 and 2 12 ; < n 1 n2 n1 n2 (n1 + n2 + 1) k2 = + 0.5 + zα/2 . 2 12 (4) An approximate 100(1 − α)% confidence interval for µ ˜1 − µ ˜2 is given by (w(k1 ) , w(k2 ) ).

c 2000 by Chapman & Hall/CRC 

9.7

FINITE POPULATION CORRECTION FACTOR

Suppose a sample of size n is taken without replacement from a (finite) population of size N . If n is large or a significant portion of the population then, intuitively, a point estimate based on this sample should be more accurate than if the population were infinite. In such cases, therefore, the standard deviation of the sample mean and the standard deviation of the sample proportion are corrected (multiplied) by the finite population correction factor: N −n . (9.10) N −1 When constructing a confidence interval, the critical distance is multiplied by this function of n and N to yield a more accurate interval estimate. If the sample size is less than 5% of the total population, the finite population correction factor is usually not applied. Confidence intervals constructed using the finite population correction factor: (1) Suppose a random sample of size n is taken from a population of size N . If the population is assumed normal, the endpoints for a 100(1 − α)% confidence interval for the population mean µ are N −n s x ± zα/2 · √ · . (9.11) N −1 n (2) In a binomial experiment, suppose a random sample of size n is taken from a population of size N . The endpoints for a 100(1−α)% confidence interval for the population proportion p are p.(1 − p.) N −n p. ± zα/2 · · . (9.12) n N −1

c 2000 by Chapman & Hall/CRC 

CHAPTER 10

Hypothesis Testing Contents 10.1

Introduction 10.1.1 Tables 10.2 The Neyman–Pearson lemma 10.3 Likelihood ratio tests 10.4 Goodness of fit test 10.5 Contingency tables 10.6 Bartlett’s test 10.6.1 Approximate test procedure 10.6.2 Tables for Bartlett’s test 10.7 Cochran’s test 10.7.1 Tables for Cochran’s test 10.8 Number of observations required 10.9 Critical values for testing outliers 10.10 Significance test in 2 × 2 contingency tables 10.11 Determining values in Bernoulli trials

10.1

INTRODUCTION

A hypothesis test is a formal procedure used to investigate a claim about one or more population parameters. Using the information in the sample the claim is either rejected or not rejected. There are four parts to every hypothesis test: (1) The null hypothesis, H0 , is a claim about the value of one or more population parameters; assumed to be true. (2) The alternative, or (research), hypothesis, Ha , is an opposing statement; believed to be true if the null hypothesis is false. (3) The test statistic, TS, is a quantity computed from the sample and used to decide whether or not to reject the null hypothesis. (4) The rejection region, RR, is a set or interval of numbers selected in such a way that if the value of the test statistic lies in the rejection region the null hypothesis is rejected. One or more critical values separate the rejection region from the remaining values of the test statistic. c 2000 by Chapman & Hall/CRC 

Decision

Nature

Do not reject H0

Reject H0

H0 True

Correct decision

Type I error: α

H0 False

Type II error: β

Correct decision

Table 10.1: Hypothesis test errors There are two error probabilities associated with hypothesis testing; they are illustrated in Table 10.1 and described below. (1) A type I error occurs if the null hypothesis is rejected when it is really true. The probability of a type I error is usually denoted by α, so that Prob [type I error] = α. Common values of α include 0.05, 0.01, and 0.001. (2) A type II error occurs if the null hypothesis is accepted when it is really false. The probability of a type II error depends upon the true value of the population parameter(s) and is usually denoted by β (or β(θ)), so that Prob [type II error] = β. The power of the hypothesis test is 1 − α. Note: (1) α is the significance level of the hypothesis test. The test statistic is significant if it lies in the rejection region. (2) The values α and β are inversely related, that is, when α increases then β decreases, and conversely. (3) To decrease both α and β, increase the sample size. The p-value is the smallest value of α (the smallest significance level) that would result in rejecting the null hypothesis. A p-value for a hypothesis test is often reported rather than whether or not the value of the test statistic lies in the rejection region. 10.1.1

Tables

Tables 10.2 and 10.3 contain hypothesis tests for one and two samples. The small numbers on the right-hand side of each table are for referencing these tests. Example 10.54 : A breakfast cereal manufacturer claims each box is filled with 24 ounces of cereal. To check this claim, a consumer group randomly selected 17 boxes and carefully weighed the contents. The summary statistics: x = 23.55 and s = 1.5. Is there any evidence to suggest the cereal boxes are underfilled? Use α = .05. Solution: (S1) This is a question about a population mean µ. The distribution of cereal box weights is assumed normal and the population variance is unknown. A one-sample c 2000 by Chapman & Hall/CRC 

Null hypothesis, assumptions

Alternative Test hypotheses statistic

Rejection regions

µ = µ0 ,

µ > µ0

Z ≥ zα

(1)

Z ≤ −zα

(2)

|Z| ≥ zα/2

(3)

T ≥ tα,n−1

(4)

T ≤ −tα,n−1

(5)

|T | ≥ tα/2,n−1

(6)

n large,

σ2

known, or µ < µ0

normality, σ 2 known

µ = µ0

µ = µ0 ,

µ > µ0

normality,

µ < µ0

σ 2 unknown

µ = µ0

σ2

=

2

σ02 ,

σ > 2

normality

σ < σ 2 =

σ02 σ02 σ02

X − µ0 Z= √ σ/ n

X − µ0 T = √ S/ n

2

(n − 1)S 2 χ = σ02 2

χ ≥ 2

χ ≤ χ2 ≤ χ2 ≥

p = p0 ,

p > p0

binomial experiment,

p < p0

n large

p = p0

Z= 

p. − p0 p0 (1 − p0 )/n

χ2α,n−1 χ21−α,n−1 χ21−α/2,n−1 , χ2α/2,n−1

(7) (8)

or (9)

Z ≥ zα

(10)

Z ≤ −zα

(11)

|Z| ≥ zα/2

(12)

Table 10.2: Hypothesis tests: one sample t test is appropriate (Table 10.2, number (5)). (S2) The four parts to the hypothesis test are: H0 : µ = 24 = µ0 Ha : µ < 24 X − µ0 √ S/ n RR: T ≤ −tα,n−1 = −t.05,16 = −1.7459 23.55 − 24 √ (S3) T = = −1.2369 1.5/ 17 (S4) Conclusion: The value of the test statistic does not lie in the rejection region (equivalently, p = .1170 > .05). There is no evidence to suggest the population mean is less than 24 ounces. TS: T =

Example 10.55 : A newspaper article claimed the proportion of local residents in favor of a property tax increase to fund new educational programs is .45. A school board member selected 192 random residents and found 65 were in favor of the tax increase. Is there any evidence to suggest the proportion reported in the newspaper article is wrong? Use α = 0.1. Solution: (S1) This is a question about a population proportion p. A binomial experiment is assumed and n is large. A one-sample test based on a Z statistic is appropriate (Table 10.2, number (12)). (S2) The four parts to the hypothesis test are: c 2000 by Chapman & Hall/CRC 

Assumptions Null hypothesis

Alternative hypotheses

Test statistic

Rejection regions

n1 , n2 large, independence, σ12 , σ22 known, or normality, independence, σ12 , σ22 known µ1 − µ2 = ∆ 0 µ 1 − µ 2 > ∆ 0

(X 1 − X 2 ) − ∆0 -

µ1 − µ2 < ∆ 0 Z =

2 σ1 n1

µ1 − µ2 = ∆0

2 σ2 n2

+

Z ≥ zα

(1)

Z ≤ −zα

(2)

|Z| ≥ zα/2

(3)

T ≥ tα,n1 +n2 −2

(4)

T ≤ −tα,n1 +n2 −2

(5)

|T | ≥ t α ,n1 +n2 −2

(6)

T  ≥ tα,ν

(7)

normality, independence, σ12 = σ22 unknown µ1 − µ2 = ∆ 0 µ 1 − µ 2 > ∆ 0 µ1 − µ2 < ∆ 0

(X 1 − X 2 ) − ∆0  Sp n1 + n1

T =

1

µ1 − µ2 = ∆0

2

2

(n1 −1)S12 + (n2 −1)S22 Sp = n1 + n2 − 2 normality, independence, σ12 , σ22 unknown, σ12 = σ22 µ1 − µ2 = ∆ 0 µ 1 − µ 2 > ∆ 0 µ1 − µ2 < ∆ 0

T =

(X 1 − X 2 ) − ∆0 2 S1 n1

µ1 − µ2 = ∆0

 ν≈

s2 1 n1

2 (s2 1 /n1 ) n1 −1

+

+

2 S2 n2

s2 2 n2

+

2



T ≤ −tα,ν

(8)

|T  | ≥ tα/2,ν

(9)

T ≥ tα,n−1

(10)

T ≤ −tα,n−1

(11)

|T | ≥ tα/2,n−1

(12)

F ≥ Fα,n1−1,n2−1

(13)

F ≤ F1−α,n1−1,n2−1

(14)

2 (s2 2 /n2 ) n2 −1

normality, n pairs, dependence µD = ∆ 0

µD > ∆ 0 µD < ∆ 0

T =

µD = ∆0

D − ∆0 √ SD / n

normality, independence σ12 = σ22

σ12 > σ22 σ12 σ12

< =

σ22 σ22

F =

S12 /S22

F ≤ F1−α ,n1−1,n2−1 2 or F ≥ F α ,n1−1,n2−1 2

binomial experiments, n1 , n2 large, independence p.1 − p.2 p1 = p2 = 0 p1 − p2 > 0 Z=  p . q . (1/n 1 + 1/n2 ) p1 − p2 < 0 p1 − p2 = 0

p. =

X1 + X2 , n1 + n2

q. = 1 − p.

(15)

Z ≥ zα

(16)

Z ≤ −zα

(17)

|Z| ≥ zα/2

(18)

Z ≥ zα

(19)

Z ≤ −zα

(20)

|Z| ≥ zα/2

(21)

binomial experiments, n1 , n2 large, independence p1 − p2 = ∆0

p1 − p2 > ∆0 p1 − p2 < ∆0 p1 − p2 = ∆0

Z= 

(. p1 − p.2 ) − ∆0 p .1 (1−. p1 ) n1

+

p .2 (1−. p2 ) n2

Table 10.3: Hypothesis tests: two samples c 2000 by Chapman & Hall/CRC 

H0 : p = .45 = p0 Ha : p = .45 p. − p0 p0 (1 − p0 )/n RR: |Z| ≥ zα/2 = z.005 = 2.5758 65 .3385 − .45 = .3385 ; Z =  (S3) p. = = −3.1044 192 (.3385)(.6615)/192 (S4) Conclusion: The value of the test statistic lies in the rejection region (equivalently, p = .0019 < .005). There is evidence to suggest the true proportion of residents in favor of the property tax increase is different from .45. TS: Z = 

Example 10.56 :

An automobile parts seller claims a new product when attached to an engine’s air filter will significantly improve gas mileage. To test this claim, a consumer group randomly selected 10 cars and drivers. The miles per gallon for each automobile was recorded without the product and then using the new product. The summary statistics for the differences (before − after) were: d = −1.2 and sD = 3.5. Is there any evidence to suggest the new product improves gas mileage? Use α = .01. Solution: (S1) This is a question about a difference in population means, µD . The data are assumed to be from a normal distribution and the observations are dependent. A paired t test is appropriate (Table 10.3, number (5)). (S2) The four parts to the hypothesis test are: H0 : µD = 0 = ∆ 0 Ha : µ D < 0 D − ∆0 √ Sd / n RR: T ≤ −tα,n−1 = −t.01,9 = −2.8214 −1.2 − 0 √ = −1.0842 (S3) T = 3.5/ 10 (S4) Conclusion: The value of the test statistic does not lie in the rejection region (equivalently, p = .1532 > .01). There is no evidence to suggest the new product improves gas mileage. TS: T =

10.2

THE NEYMAN–PEARSON LEMMA

Given the null hypothesis H0 : θ = θ0 versus the alternative hypothesis Ha : θ = θa , let L(θ) be the likelihood function evaluated at θ. For a given α, the test that maximizes the power at θa has a rejection region determined by L(θ0 )
c 2000 by Chapman & Hall/CRC 

(10.1)

10.3

LIKELIHOOD RATIO TESTS

Given the null hypothesis H0 : θ ∈ Ω0 versus the alternative hypothesis Ha : . 0 ) be the likelihood θ ∈ Ωa with Ω0 ∩ Ωa = φ and Ω = Ω0 ∪ Ωa . Let L(Ω function with all unknown parameters replaced by their maximum likelihood . be defined similarly estimators subject to the constraint θ ∈ Ω0 , and let L(Ω) so that θ ∈ Ω. Define . 0) L(Ω λ= . (10.2) . L(Ω) A likelihood ratio test of H0 versus Ha uses λ as a test statistic and has a rejection region given by λ ≤ k (for 0 < k < 1). Under very general conditions and for large n, −2 ln λ has approximately a chi–square distribution with degrees of freedom equal to the number of parameters or functions of parameters assigned specific values under H0 . 10.4

GOODNESS OF FIT TEST

Let ni be the number of observations falling into the ith category (for i = 1, 2, . . . , k) and let n = n1 + n2 + · · · + nk . H0 : p1 = p10 , p2 = p20 , . . . , pk = pk0 Ha : pi = pi0 for at least one i TS: χ2 =

k k   (observed − estimated expected)2 (ni − npi0 )2 = estimated expected npi0 i=1 i=1

Under the null hypothesis χ2 has approximately a chi–square distribution with k − 1 degrees of freedom. The approximation is satisfactory if n pi0 ≥ 5 for all i. RR: χ2 ≥ χ2α,k−1 Example 10.57 :

The bookstore at a large university stocks four brands of graphing calculators. Recent sales figures indicated 55% of all graphing calculator sales were Texas Instruments (TI), 25% were Hewlett Packard (HP), 15% were Casio, and 5% were Sharp. This semester 200 graphing calculators were sold according to the table given below. Is there any evidence to suggest the sales proportions have changed? Use α = .05. Calculator Sales TI

HP

Casio

Sharp

120

47

21

12

c 2000 by Chapman & Hall/CRC 

Solution: (S1) There are k = 4 categories (of calculators) with unequal expected frequencies. The bookstore would like to determine if sales are consistent with previous results. This problem involves a goodness of fit test based on a chi–square distribution. (S2) The four parts to the hypothesis test are: H0 : p1 = .55, p2 = .25, p3 = .15, p4 = .05. Ha : pi = pi0 for at least one i 4 4   (observed − estimated expected)2 (ni − npi0 )2 = TS: χ2 = estimated expected npi0 i=1 i=1 RR: χ2 ≥ χ2α,k−1 = χ2.05,3 = 7.8147 (120 − 110)2 (50 − 47)2 (30 − 21)2 10 − 12)2 + + + = 4.1891 110 50 30 10 (S4) Conclusion: The value of the test statistic does not lie in the rejection region (equivalently, p = .2418 > .05). There is no evidence to suggest the proportions of graphing calculator sales have changed. (S3) χ2 =

If k = 2, this test is equivalent to a one proportion Z test, Table 10.2, number (3). This result follows from section 6.18.3 (page 149): If Z is a standard normal random variable, then Z 2 has a chi–square distribution with 1 degree of freedom. 10.5

CONTINGENCY TABLES

The general I × J contingency table has the form: Sample 1 Sample 2 .. .

Treatment 1 n11 n21 .. .

Sample I Totals

nI1 n.1

Treatment 1 n12 n22 .. .

... ... ... .. .

Treatment J n1J n2J .. .

Totals n1. n2. .. .

nI2 ... nIJ nI. n.2 ... n.J n J I where nk. = j=1 nkj and n.k = i=1 nik . If complete independence is assumed, then the probability of any specific configuration, given the row and column totals {n.k , nk. }, is Prob [n11 , . . . , nIJ | n1. , . . . , n.J ] =

(ΠIi ni. !)(ΠJj n.j !) n! ΠIi ΠJj nij !

(10.3)

Let a contingency table contain I rows and J columns, let nij be the count in the (i, j)th cell, and let . 3ij be the estimated expected count in that cell. The test statistic is χ2 =

I  J   (observed − estimated expected)2 (nij − 3ˆij )2 (10.4) = estimated expected 3ˆij i=1 j=1

all cells

c 2000 by Chapman & Hall/CRC 

where 3ˆij =

ni. n.j (ith row total)(j th column total) = grand total n

(10.5)

Under the null hypothesis χ2 has approximately a chi–square distribution with (I − 1)(J − 1) degrees of freedom. The approximation is satisfactory if 3ˆij ≥ 5 for all i and j. Example 10.58 : Recent reports indicate meals served during flights are rated similar regardless of airline. A survey given to randomly selected passengers asked each to rate the quality of in-flight meals. The results are given in the table below.

Poor Acceptable Good

A

Airline B C

D

42 50 10

35 75 17

23 28 18

22 33 21

Is there any evidence to suggest the quality of meals differs by airline? Use α = .01. Solution: (S1) The contingency table has I = 3 rows and J = 4 columns. To determine if the meal ratings differ by airline, a contingency table analysis is appropriate. The test statistic is based on a chi–square distribution. (S2) The four parts to the hypothesis test are: H0 : Airline and meal ratings are independent Ha : Airline and meal ratings are dependent 3  4  (nij − Aˆij )2 TS: χ2 = Aˆij i=1 j=1 RR: χ2 ≥ χ2.01,6 = 18.5476 (42 − 33.27)2 (35 − 41.43)2 (22 − 24.79)2 (23 − 22.51)2 + + + 33.27 41.43 24.79 22.51 (75 − 63.16)2 (33 − 37.80)2 (28 − 34.32)2 (50 − 50.73)2 + + + + 50.73 63.16 37.80 34.32 (17 − 22.41)2 (21 − 13.41)2 (18 − 12.18)2 (10 − 18.00)2 + + + + 18.00 22.41 13.41 12.18 = 19.553

(S3) χ2 =

(S4) The value of the test statistic lies in the rejection region (equivalently, p = .003 < .01). There is evidence to suggest the meal rating proportions differ by airline.

10.6

BARTLETT’S TEST

Let there be k independent samples with ni (for i = 1, 2, . . . , k) observations in each sample, N = n1 + n2 + · · · + nk , and let Si2 be the ith sample variance.

c 2000 by Chapman & Hall/CRC 

H0 : σ12 = σ22 = · · · = σk2 Ha : the variances are not all equal  2 n −1 2 n −1 1/(N −k) (S1 ) 1 (S2 ) 2 · · · (Sk2 )nk −1 TS: B = Sp2 k  (ni − 1)Si2 i=1 2 Sp = N −k RR: B ≤ bα,k,n

(n1 = n2 = · · · = nk = n)

B ≤ bα,k,n1 ,n2 ,...,nk

(when sample sizes are unequal)

n1 bα,k,n1 + n2 bα,k,n2 + · · · + nk bα,k,nk N Here bα,k,n is a critical value for Bartlett’s test with α being the significance level, k is the number of populations, and n is the sample size from each population. A table of values is in section 10.6.2. where bα,k,n1 ,n2 ,...,nk ≈

10.6.1

Approximate test procedure

Let νi = ni − 1 TS: χ2 = M/C where

k k   2 M= νi ln S − νi ln Si2 , i=1

i=1

2

S =

k  i=1

& % k = k  1  1 C =1+ − 1 νi 3(k − 1) i=1 νi i=1

νi Si2

= k 

νi

i=1

Under the null hypothesis χ2 has approximately a chi–square distribution with k − 1 degrees of freedom. RR: χ2 ≥ χ2α,k−1 10.6.2

Tables for Bartlett’s test

These tables contain critical values, bα,k,n , for Bartlett’s test where α is the significance level, k is the number of populations, and n is the sample size from each population. These tables are from D. D. Dyer and J. P. Keating, “On the Determination of Critical Values for Bartlett’s Test”, JASA, Volume 75, 1980, pages 313–319. Reprinted with permission from the Journal of American Statistical Association. Copyright 1980 by the American Statistical Association. All rights reserved.

c 2000 by Chapman & Hall/CRC 

Critical values for Bartlett’s test, bα,k,n α = .05 n 2 3 .3123 4 .4780 5 .5845 6 .6563 7 .7075 8 .7456 9 .7751

k 3 .3058 .4699 .5762 .6483 .7000 .7387 .7686

4 .3173 .4803 .5850 .6559 .7065 .7444 .7737

5 .3299 .4921 .5952 .6646 .7142 .7512 .7798

6 ∗ .5028 .6045 .6727 .7213 .7574 .7854

7 ∗ .5122 .6126 .6798 .7275 .7629 .7903

8 ∗ .5204 .6197 .6860 .7329 .7677 .7946

9 ∗ .5277 .6260 .6914 .7376 .7719 .7984

10 ∗ .5341 .6315 .6961 .7418 .7757 .8017

10 11 12 13 14

.7984 .8175 .8332 .8465 .8578

.7924 .8118 .8280 .8415 .8532

.7970 .8160 .8317 .8450 .8564

.8025 .8210 .8364 .8493 .8604

.8076 .8257 .8407 .8533 .8641

.8121 .8298 .8444 .8568 .8673

.8160 .8333 .8477 .8598 .8701

.8194 .8365 .8506 .8625 .8726

.8224 .8392 .8531 .8648 .8748

15 16 17 18 19

.8676 .8761 .8836 .8902 .8961

.8632 .8719 .8796 .8865 .8926

.8662 .8747 .8823 .8890 .8949

.8699 .8782 .8856 .8921 .8979

.8734 .8815 .8886 .8949 .9006

.8764 .8843 .8913 .8975 .9030

.8790 .8868 .8936 .8997 .9051

.8814 .8890 .8957 .9016 .9069

.8834 .8909 .8975 .9033 .9086

20 21 22 23 24

.9015 .9063 .9106 .9146 .9182

.8980 .9030 .9075 .9116 .9153

.9003 .9051 .9095 .9135 .9172

.9031 .9078 .9120 .9159 .9195

.9057 .9103 .9144 .9182 .9217

.9080 .9124 .9165 .9202 .9236

.9100 .9143 .9183 .9219 .9253

.9117 .9160 .9199 .9235 .9267

.9132 .9175 .9213 .9248 .9280

25 26 27 28 29

.9216 .9246 .9275 .9301 .9326

.9187 .9219 .9249 .9276 .9301

.9205 .9236 .9265 .9292 .9316

.9228 .9258 .9286 .9312 .9336

.9249 .9278 .9305 .9330 .9354

.9267 .9296 .9322 .9347 .9370

.9283 .9311 .9337 .9361 .9383

.9297 .9325 .9350 .9374 .9396

.9309 .9336 .9361 .9385 .9406

30 40 50 60 80 100

.9348 .9513 .9612 .9677 .9758 .9807

.9325 .9495 .9597 .9665 .9749 .9799

.9340 .9506 .9606 .9672 .9754 .9804

.9358 .9520 .9617 .9681 .9761 .9809

.9376 .9533 .9628 .9690 .9768 .9815

.9391 .9545 .9637 .9698 .9774 .9819

.9404 .9555 .9645 .9705 .9779 .9823

.9416 .9564 .9652 .9710 .9783 .9827

.9426 .9572 .9658 .9716 .9787 .9830

c 2000 by Chapman & Hall/CRC 

Critical values for Bartlett’s test, bα,k,n α = .01 n 2 3 .1411 4 .2843 5 .3984 6 .4850 7 .5512 8 .6031 9 .6445

k 3 .1672 .3165 .4304 .5149 .5787 .6282 .6676

4 ∗ .3475 .4607 .5430 .6045 .6518 .6892

5 ∗ .3729 .4850 .5653 .6248 .6704 .7062

6 ∗ .3937 .5046 .5832 .6410 .6851 .7197

7 ∗ .4110 .5207 .5978 .6542 .6970 .7305

8 ∗ ∗ .5343 .6100 .6652 .7069 .7395

9 ∗ ∗ .5458 .6204 .6744 .7153 .7471

10 ∗ ∗ .5558 .6293 .6824 .7225 .7536

10 11 12 13 14

.6783 .7063 .7299 .7501 .7674

.6996 .7260 .7483 .7672 .7835

.7195 .7445 .7654 .7832 .7985

.7352 .7590 .7789 .7958 .8103

.7475 .7703 .7894 .8056 .8195

.7575 .7795 .7980 .8135 .8269

.7657 .7871 .8050 .8201 .8330

.7726 .7935 .8109 .8256 .8382

.7786 .7990 .8160 .8303 .8426

15 16 17 18 19

.7825 .7958 .8076 .8181 .8275

.7977 .8101 .8211 .8309 .8397

.8118 .8235 .8338 .8429 .8512

.8229 .8339 .8436 .8523 .8601

.8315 .8421 .8514 .8596 .8670

.8385 .8486 .8576 .8655 .8727

.8443 .8541 .8627 .8704 .8773

.8491 .8586 .8670 .8745 .8811

.8532 .8625 .8707 .8780 .8845

20 21 22 23 24

.8360 .8437 .8507 .8571 .8630

.8476 .8548 .8614 .8673 .8728

.8586 .8653 .8714 .8769 .8820

.8671 .8734 .8791 .8844 .8892

.8737 .8797 .8852 .8902 .8948

.8791 .8848 .8901 .8949 .8993

.8835 .8890 .8941 .8988 .9030

.8871 .8926 .8975 .9020 .9061

.8903 .8956 .9004 .9047 .9087

25 26 27 28 29

.8684 .8734 .8781 .8824 .8864

.8779 .8825 .8869 .8909 .8946

.8867 .8911 .8951 .8988 .9023

.8936 .8977 .9015 .9050 .9083

.8990 .9029 .9065 .9099 .9130

.9034 .9071 .9105 .9138 .9167

.9069 .9105 .9138 .9169 .9198

.9099 .9134 .9166 .9196 .9224

.9124 .9158 .9190 .9219 .9246

30 40 50 60 80 100

.8902 .9175 .9339 .9449 .9586 .9669

.8981 .9235 .9387 .9489 .9617 .9693

.9056 .9291 .9433 .9527 .9646 .9716

.9114 .9335 .9468 .9557 .9668 .9734

.9159 .9370 .9496 .9580 .9685 .9748

.9195 .9397 .9518 .9599 .9699 .9759

.9225 .9420 .9536 .9614 .9711 .9769

.9250 .9439 .9551 .9626 .9720 .9776

.9271 .9455 .9564 .9637 .9728 .9783

c 2000 by Chapman & Hall/CRC 

10.7

COCHRAN’S TEST

Let there be k independent samples with n observations in each sample, and let Si2 be the ith sample variance (for i = 1, 2, . . . , k). H0 : σ12 = σ22 = · · · = σk2 Ha : the variances are not all equal = k  2 TS: G = largest Si Si2 i=1

RR: G ≥ gα,k,n Here gα,k,n is a critical value for Cochran’s test with α being the significance level, k is the number of populations, and n is the sample size from each population. A table of values is in section 10.7.1. 10.7.1

Tables for Cochran’s test

These tables contain critical values, gα,k,n , for Cochran’s test where α is the significance level, k is the number of independent estimates of variance, each of which is based on n degrees of freedom. These tables are from C. Eisenhart, M. W. Hastay, and W. A. Wallis, Techniques of Statistical Analysis, McGrawHill Book Company, 1947, Tables 15.1 and 15.2 (pages 390-391). Reprinted courtesy of The McGraw-Hill Companies.

c 2000 by Chapman & Hall/CRC 

.8412 .7808 .7271 .6798 .6385 .6020 .5410 .4709 .3894 .3434 .2929 .2370 .1737 .0998 0

5 6 7 8 9 10 12 15 20 24

c 2000 by Chapman & Hall/CRC 

30 40 60 120 ∞

.1980 .1576 .1131 .0632 0

.4450 .3924 .3346 .2705 .2354

.6838 .6161 .5612 .5157 .4775

.1593 .1259 .0895 .0495 0

.3733 .3264 .2758 .2205 .1907

.5981 .5321 .4800 .4377 .4027

.1377 .1082 .0765 .0419 0

.3311 .2880 .2419 .1921 .1656

.5441 .4803 .4307 .3910 .3584

.1237 .0968 .0682 .0371 0

.3029 .2624 .2195 .1735 .1493

.5065 .4447 .3974 .3595 .3286

.1137 .0887 .0623 .0337 0

.2823 .2439 .2034 .1602 .1374

.4783 .4184 .3726 .3362 .3067

.1061 .0827 .0583 .0312 0

.2666 .2299 .1911 .1501 .1286

.4564 .3980 .3535 .3185 .2901

.1002 .0780 .0552 .0292 0

.2541 .2187 .1815 .1422 .1216

.4387 .3817 .3384 .3043 .2768

.0958 .0745 .0520 .0279 0

.2439 .2098 .1736 .1357 .1160

.4241 .3682 .3259 .2926 .2659

.0921 .0713 .0497 .0266 0

.2353 .2020 .1671 .1303 .1113

.4118 .3568 .3154 .2829 .2568

.0771 .0595 .0411 .0218 0

.2032 .1737 .1429 .1108 .0942

.3645 .3135 .2756 .2462 .2226

.0604 .0462 .0316 .0165 0

.1655 .1403 .1144 .0879 .0743

.3066 .2612 .2278 .2022 .1820

.0457 .0347 .0234 .0120 0

.1308 .1100 .0889 .0675 .0567

.2513 .2119 .1833 .1616 .1446

.0333 .0250 .0167 .0083 0

.1000 .0833 .0667 .0500 .0417

.2000 .1667 .1429 .1250 .1111

α = .05 n k 1 2 3 4 5 6 7 8 9 10 16 36 144 ∞ 2 .9985 .9750 .9392 .9057 .8772 .8534 .8332 .8159 .8010 .7880 .7341 .6602 .5813 .5000 3 .9669 .8709 .7977 .7457 .7071 .6771 .6530 .6333 .6167 .6025 .5466 .4748 .4031 .3333 4 .9065 .7679 .6841 .6287 .5895 .5598 .5365 .5175 .5017 .4884 .4366 .3720 .3093 .2500

Critical values for Cochran’s test, gα,k,n

.9279 .8828 .8376 .7945 .7544 .7175 .6528 .5747 .4799 .4247 .3632 .2940 .2151 .1225 0

5 6 7 8 9 10 12 15 20 24

c 2000 by Chapman & Hall/CRC 

30 40 60 120 ∞

.2412 .1915 .1371 .0759 0

.5358 .4751 .4069 .3297 .2871

.7885 .7218 .6644 .6152 .5727

.1913 .1508 .1069 .0585 0

.4469 .3919 .3317 .2654 .2295

.6957 .6258 .5685 .5209 .4810

.1635 .1281 .0902 .0489 0

.3934 .3428 .2882 .2288 .1970

.6329 .5635 .5080 .4627 .4251

.1454 .1135 .0796 .0429 0

.3572 .3099 .2593 .2048 .1759

.5875 .5195 .4659 .4226 .3870

.1327 .1033 .0722 .0387 0

.3308 .2861 .2386 .1877 .1608

.5531 .4866 .4347 .3932 .3592

.1232 .0957 .0668 .0357 0

.3106 .2680 .2228 .1748 .1495

.5259 .4608 .4105 .3704 .3378

.1157 .0898 .0625 .0334 0

.2945 .2535 .2104 .1646 .1406

.5037 .4401 .3911 .3522 .3207

.1100 .0853 .0594 .0316 0

.2813 .2419 .2002 .1567 .1338

.4854 .4229 .3751 .3373 .3067

.1054 .0816 .0567 .0302 0

.2704 .2320 .1918 .1501 .1283

.4697 .4084 .3616 .3248 .2950

.0867 .0668 .0461 .0242 0

.2297 .1961 .1612 .1248 .1060

.4094 .3529 .3105 .2779 .2514

.0658 .0503 .0344 .0178 0

.1811 .1535 .1251 .0960 .0810

.3351 .2858 .2494 .2214 .1992

.0480 .0363 .0245 .0125 0

.1376 .1157 .0934 .0709 .0595

.2644 .2229 .1929 .1700 .1521

.0333 .0250 .0167 .0083 0

.1000 .0833 .0667 .0500 .0417

.2000 .1667 .1429 .1250 .1111

α = .05 n k 1 2 3 4 5 6 7 8 9 10 16 36 144 ∞ 2 .9999 .9950 .9794 .9586 .9373 .9172 .8988 .8823 .8674 .8539 .7949 .7067 .6062 .5000 3 .9933 .9423 .8831 .8335 .7933 .7606 .7335 .7107 .6912 .6743 .6059 .5153 .4230 .3333 4 .9676 .8643 .7814 .7212 .6761 .6410 .6129 .5897 .5702 .5536 .4884 .4057 .3251 .2500

Critical values for Cochran’s test, gα,k,n

10.8 NUMBER OF OBSERVATIONS REQUIRED FOR THE COMPARISON OF A POPULATION VARIANCE WITH A STANDARD VALUE USING THE CHI–SQUARE TEST Suppose x1 , x2 , . . . , xn+1 is a random sample from a population with variance sigma21 . The sample variance, s21 has n degrees of freedom, and may be used to test the hypothesis that σ12 = σ02 . Let R be the ratio of the variances σ02 and σ12 . The table below shows the value of the ratio R for which a chi-square test, with significance level α, will not be able to detect the difference in the variances with probability β. Note that when R is far from one few samples will be required to distinguish σ02 from σ12 , while for R near one large samples will be required. Example 10.59 : Testing for an increase in variance. Let α = 0.05, β = 0.01, and R = 4. Using the table below with these values the value R = 4 occurs between the rows corresponding to n = 15 and n = 20. Using rough, linear, interpolation, the table indicates that the estimate of variance should be based on 19 degrees of freedom. Example 10.60 :

Testing for an decrease in variance. Let α = 0.05, β = 0.01, and R = 0.33. Using the table below with α = β = 0.01, β  = α = 0.05 and R = 1/R = 3, the value R = 3 occurs between the rows corresponding to n = 24 and n = 30. Using rough, linear, interpolation, the table indicates that the estimate of variance should be based on 26 degrees of freedom.

Values of R given n, α, and β α = 0.01 α = 0.05 n β = 0.01 β = 0.05 β = 0.1 β = 0.5 β = 0.01 β = 0.05 β = 0.1 β = 0.5 1 42236.852 1687.350 420.176 14.584 24454.206 976.938 243.272 8.444 2 458.211 89.781 43.709 6.644 298.073 58.404 28.433 4.322 3 98.796 32.244 19.414 4.795 68.054 22.211 13.373 3.303 4 44.686 18.681 12.483 3.955 31.933 13.349 8.920 2.827 5 27.217 13.170 9.369 3.467 19.972 9.665 6.875 2.544 6 7 8 9 10

19.278 14.911 12.202 10.377 9.072

10.280 8.524 7.352 6.516 5.890

7.627 6.521 5.757 5.198 4.770

3.144 2.911 2.736 2.597 2.484

14.438 11.353 9.418 8.103 7.156

7.699 6.490 5.675 5.088 4.646

5.713 4.965 4.444 4.059 3.763

2.354 2.217 2.112 2.028 1.960

15 20 25 30 40 50

5.847 4.548 3.845 3.403 2.874 2.564

4.211 3.462 3.033 2.752 2.403 2.191

3.578 3.019 2.690 2.471 2.192 2.021

2.133 1.943 1.821 1.735 1.619 1.544

4.780 3.803 3.267 2.927 2.516 2.272

3.442 2.895 2.577 2.367 2.103 1.942

2.925 2.524 2.286 2.125 1.919 1.791

1.743 1.624 1.547 1.492 1.418 1.368

75 100 150 ∞

2.150 1.938 1.715 1.000

1.898 1.743 1.575 1.000

1.779 1.649 1.506 1.000

1.431 1.367 1.297 1.000

1.945 1.775 1.594 1.000

1.716 1.596 1.464 1.000

1.609 1.510 1.400 1.000

1.294 1.252 1.206 1.000

c 2000 by Chapman & Hall/CRC 

10.9

CRITICAL VALUES FOR TESTING OUTLIERS

Tests for outliers may be based on the largest deviation max (xi − x) of the i=1,2,...

observations from their mean (which has to be normalized by the standard deviation or an estimate of the standard deviation). An alternative technique is to look at ratios of approximations to the range. (a) To determine if the smallest element in a sample, x(1) , is an outlier compute x(2) − x(1) r10 = (10.6) x(n) − x(1) Equivalently, to determine if the largest element in a sample, x(n) , is an outlier compute x(n) − x(n−1) r10 = (10.7) x(n) − x(1) (b) To determine if the smallest element in a sample, x(1) , is an outlier, and the value x(n) is not to be used, then compute r11 =

x(2) − x(1) x(n−1) − x(1)

(10.8)

Equivalently, to determine if the largest element in a sample, x(n) , is an outlier, without using the value x(1) , compute r11 =

x(n) − x(n−1) x(n) − x(2)

(10.9)

(c) To determine if the smallest element in a sample, x(1) , is an outlier, and the value x(2) is not to be used, then compute r20 =

x(3) − x(1) x(n) − x(1)

(10.10)

Equivalently, to determine if the largest element in a sample, x(n) , is an outlier, without using the value x(n−1) , compute r20 =

x(n) − x(n−2) x(n) − x(2)

(10.11)

The following tables contain critical values for r10 , r11 , and r20 . See W. J. Dixon, Annals of Mathematical Statistics, 22, 1951, pages 68–78.

c 2000 by Chapman & Hall/CRC 

Percentage values for r10

(Prob [r10 > R] = α)

n 3 4 5 6 7 8 9

α = .005 .994 .926 .821 .740 .680 .634 .598

.01 .988 .889 .780 .698 .637 .590 .555

.02 .976 .846 .729 .644 .586 .543 .510

.05 .941 .745 .642 .560 .507 .468 .437

.10 .886 .679 .557 .482 .434 .399 .370

.50 .500 .324 .250 .210 .184 .166 .152

.90 .114 .065 .048 .038 .032 .029 .026

.95 .059 .033 .023 .018 .016 .014 .013

10 15 20 25 30

.568 .475 .425 .393 .372

.527 .438 .391 .362 .341

.483 .399 .356 .329 .309

.412 .338 .300 .277 .260

.349 .285 .252 .230 .215

.142 .111 .096 .088 .082

.025 .019 .017 .015 .014

.012 .010 .008 .008 .007

Percentage values for r11

(Prob [r11 > R] = α)

n 4 5 6 7 8 9

α = .005 .995 .937 .839 .782 .725 .677

.01 .991 .916 .805 .740 .683 .635

.02 .981 .876 .763 .689 .631 .587

.05 .955 .807 .689 .610 .554 .512

.10 .910 .728 .609 .530 .479 .441

.50 .554 .369 .288 .241 .210 .189

.90 .131 .078 .056 .045 .037 .033

.95 .069 .039 .028 .022 .019 .016

10 15 20 25 30

.639 .522 .464 .426 .399

.597 .486 .430 .394 .369

.551 .445 .392 .359 .336

.477 .381 .334 .394 .283

.409 .323 .282 .255 .236

.173 .129 .110 .098 .090

.030 .023 .019 .017 .016

.014 .011 .010 .009 .008

Percentage values for r20

(Prob [r20 > R] = α)

n 4 5 6 7 8 9 10

α = .005 .996 .950 .865 .814 .746 .700 .664

.01 .992 .929 .836 .778 .719 .667 .632

.02 .987 .901 .800 .732 .670 .627 .592

.05 .967 .845 .736 .661 .607 .565 .531

.10 .935 .782 .670 .596 .545 .505 .474

.50 .676 .500 .411 .355 .317 .288 .268

.90 .321 .218 .172 .144 .125 .114 .104

.95 .235 .155 .126 .099 .085 .077 .070

15 20 25 30

.554 .494 .456 .428

.522 .464 .428 .402

.486 .430 .395 .372

.430 .372 .343 .322

.382 .333 .304 .285

.209 .179 .161 .149

.079 .067 .060 .056

.052 .046 .041 .039

c 2000 by Chapman & Hall/CRC 

10.10 TEST OF SIGNIFICANCE IN 2 × 2 CONTINGENCY TABLES A 2 × 2 contingency table (see section 10.5) is a special case that occurs often. Suppose n elements are simultaneously classified as having either property 1 or 2 and as having property I or II. The 2 × 2 contingency table may be written as: 1 2 Totals

I a b r

II A−a B−b n−r

Totals A B n

If the marginal totals are fixed, the probability of a given configuration may be written as AB  A! B! r! (n − r)! f (a | r, A, B) = anb = (10.12) n! a! b! (A − a)! (B − b)! r The following tables are designed to be used in conducting a hypothesis test concerning the difference between observed and expected frequencies in a 2×2 contingency table. For given values of a, A, and B, table entries show the largest value of b (in bold type, with b < a) for which there is a significant difference (between observed and expected frequencies, or equivalently, between a/A and b/B). Critical values of b (probability levels) are presented for α = .05, .025, .01, and .005. The tables also satisfy the following conditions: (1) Categories 1 and 2 are determined so that A ≥ B. a b (2) ≥ or, aB ≥ bA. A B (3) If b is less than or equal to the integer in bold type, then a/A is significantly greater than b/B (for a one tailed–test) at the probability level (α) indicated by the column heading. For a two-tailed test the significance level is 2α. (4) A dash in the body of the table indicates no 2 × 2 table may show a significant effect at that probability level and combination of a, A, and B. (5) For a given r, the probability b is less than the integer in bold type is shown in small type following an entry. Note that as A and B get large, this test may be approximated by a twosample Z test of proportions.

c 2000 by Chapman & Hall/CRC 

Example 10.61 : In order to compare the probability of a success in two populations, the following 2 × 2 contingency table was obtained.

Sample from population 1 Sample from population 2 Totals

Success

Failure

Totals

7 3

2 3

9 6

10

5

15

Is there any evidence to suggest the two population proportions are different? Use α = .05. Solution: (S1) In this 2 × 2 contingency table, a = 7, A = 9, and B = 6. For α = .05 the table entry is 1.035. (S2) The critical value for b is 1. If b ≤ 1 then the null hypothesis H0 : p1 = p2 is rejected. (S3) Conclusion: The value of the test statistic does not lie in the rejection region, b = 3. There is no evidence to suggest the population proportions are different. (S4) Note there are six 2 × 2 tables with the same marginal totals as the table in this example (that is, A = 9, B = 6, and r = 10): 9 1

0 5

8 2

1 4

7 3

2 3

6 4

3 2

5 5

4 1

4 6

5 0

Assuming independence, the probability of obtaining each of these six tables (using equation (10.12), rounded) is {.002, .045, .24, .42, .25, .042}. That is, the first configuration is the least likely, and the fourth configuration is the most likely.

c 2000 by Chapman & Hall/CRC 

Contingency tables: 2 × 2 a A=3B=3 A=4B=4 B=3 A=5B=5 B=4 B=3 B=2 A=6B=6

B=5

B=4 B=3 B=2 A=7B=7

B=6

B=5

B=4

B=3 B=2 A=8B=8

B=7

B=6

B=5

3 4 4 5 4 5 4 5 5 6 5 4 6 5 4 6 5 6 5 6 7 6 5 4 7 6 5 4 7 6 5 7 6 5 7 6 7 8 7 6 5 4 8 7 6 5 8 7 6 5 8 7

0.05 0.050 0.014 0.029 1.024 0.024 1.048 0.040 0.018 0.048 2.030 1.039 0.030 1.015 0.013 0.045 1.033 0.024 0.012 0.048 0.036 3.035 2.049 1.049 0.035 2.021 1.024 0.016 0.049 2.045 1.044 0.027 1.024 0.015 0.045 0.008 0.033 0.028 4.038 2.020 1.020 0.013 0.038 3.026 2.034 1.030 0.019 2.015 1.016 1.049 0.028 2.035 1.031

Probability 0.025 0.01 − − 0.014 − − − 1.024 0.004 0.024 − 0.008 0.008 − − 0.018 − − − 1.008 1.008 0.008 0.008 − − 1.015 0.002 0.013 − − − 0.005 0.005 0.024 − 0.012 − − − − − 2.010 1.002 1.014 0.002 0.010 − − − 2.021 1.005 1.024 0.004 0.016 − − − 1.010 0.001 0.008 0.008 − − 1.024 0.003 0.015 − − − 0.008 0.008 − − − − 3.013 2.003 2.020 1.005 1.020 0.003 0.013 − − − 2.007 2.007 1.009 1.009 0.006 0.006 0.019 − 2.015 1.003 1.016 0.002 0.009 0.009 − − 1.007 1.007 0.005 0.005

c 2000 by Chapman & Hall/CRC 

a 0.005 − A=8 B − B − 0.004 − B − − B − A=9 B − 0.001 − − 0.002 − B − 0.005 − − − B − 1.002 0.002 − − B 1.005 0.004 − − 0.001 B − − 0.003 − B − − − − B 2.003 1.005 0.003 B − A = 10 B − 1.001 0.001 − − 1.003 0.002 B − − 0.001 0.005

=5 =4

=3 =2 =9

=8

=7

=6

=5

=4

=3

=2 = 10

=9

6 5 8 7 6 8 7 8 9 8 7 6 5 4 9 8 7 6 5 9 8 7 6 5 9 8 7 6 5 9 8 7 6 9 8 7 6 9 8 7 9 10 9 8 7 6 5 4 10 9 8 7 6 5

0.05 0.016 0.044 1.018 0.010 0.030 0.006 0.024 0.022 5.041 3.024 2.027 1.024 0.015 0.041 4.029 3.041 2.041 1.035 0.020 3.019 2.024 1.020 0.010 0.029 3.044 2.045 1.034 0.017 0.042 2.027 1.022 0.010 0.028 1.014 0.007 0.021 0.049 1.045 0.018 0.045 0.018 6.043 4.027 3.032 2.032 1.027 0.016 0.043 5.033 4.046 2.018 2.047 1.038 0.022

Probability 0.025 0.01 0.016 − − − 1.018 0.002 0.010 − − − 0.006 0.006 0.024 − 0.022 − 4.015 3.005 3.024 2.007 1.007 1.007 1.024 0.005 0.015 − − − 3.009 3.009 2.013 1.003 1.012 0.002 0.007 0.007 0.020 − 3.019 2.005 2.024 1.006 1.020 0.003 0.010 − − − 2.011 1.002 1.011 0.001 0.006 0.006 0.017 − − − 1.005 1.005 1.022 0.003 0.010 − − − 1.014 0.001 0.007 0.007 0.021 − − − 0.005 0.005 0.018 − − − 0.018 − 5.016 4.005 3.010 3.010 2.011 1.003 1.010 1.010 0.005 0.005 0.016 − − − 4.011 3.003 3.017 2.005 2.018 1.004 1.014 0.002 0.008 0.008 0.022 −

0.005 − − 0.002 − − − − − 3.005 1.002 0.001 0.005 − − 2.002 1.003 0.002 − − 2.005 0.001 0.003 − − 1.002 0.001 − − − 1.005 0.003 − − 0.001 − − − 0.005 − − − 3.002 2.003 1.003 0.002 − − − 3.003 2.005 1.004 0.002 − −

Contingency tables: 2 × 2 a A = 10 B = 8 10 9 8 7 6 5 B = 7 10 9 8 7 6 5 B = 6 10 9 8 7 6 B = 5 10 9 8 7 6 B = 4 10 9 8 7 B = 3 10 9 8 B = 2 10 9 A = 11 B = 11 11 10 9 8 7 6 5 4 B = 10 11 10 9 8 7 6 5 B = 9 11 10 9 8 7 6 5 B = 8 11 10

0.05 4.023 3.030 2.029 1.022 0.011 0.029 3.015 2.017 2.049 1.035 0.017 0.041 3.036 2.034 1.024 0.010 0.026 2.022 1.017 1.045 0.019 0.042 1.011 1.040 0.015 0.035 1.038 0.014 0.035 0.015 0.045 7.045 5.030 4.036 3.039 2.036 1.030 0.018 0.045 6.035 4.020 3.022 2.021 1.016 1.040 0.023 5.026 4.036 3.037 2.032 1.024 0.012 0.030 4.018 3.023

Probability 0.025 0.01 4.023 3.007 2.009 2.009 1.007 1.007 1.022 0.004 0.011 − − − 3.015 2.003 2.017 1.004 1.013 0.002 0.006 0.006 0.017 − − − 2.008 2.008 1.007 1.007 1.024 0.003 0.010 − − − 2.022 1.004 1.017 0.002 0.007 0.007 0.019 − − − 1.011 0.001 0.005 0.005 0.015 − − − 0.003 0.003 0.014 − − − 0.015 − − − 6.018 5.006 4.011 3.004 3.014 2.004 2.014 1.004 1.011 0.002 0.006 0.006 0.018 − − − 5.012 4.004 4.020 3.006 3.022 2.007 2.021 1.006 1.016 0.003 0.009 0.009 0.023 − 4.008 4.008 3.012 2.003 2.012 1.003 1.009 1.009 1.024 0.004 0.012 − − − 4.018 3.005 3.023 2.006

c 2000 by Chapman & Hall/CRC 

a 0.005 2.002 A = 11 1.002 0.001 0.004 − − 2.003 1.004 0.002 − − − 1.001 0.001 0.003 − − 1.004 0.002 − − − 0.001 0.005 − − 0.003 − − − − 4.002 A = 12 3.004 2.004 1.004 0.002 − − − 4.004 2.002 1.002 0.001 0.003 − − 3.002 2.003 1.003 0.001 0.004 − − 3.005 1.001

B=8

B=7

B=6

B=5

B=4

B=3

B=2 B = 12

B = 11

B = 10

9 8 7 6 5 11 10 9 8 7 6 11 10 9 8 7 6 11 10 9 8 7 11 10 9 8 11 10 9 11 10 12 11 10 9 8 7 6 5 4 12 11 10 9 8 7 6 5 12 11 10 9 8 7 6

0.05 2.020 1.014 1.035 0.017 0.040 4.043 3.045 2.036 1.024 0.010 0.025 3.029 2.027 1.017 1.041 0.017 0.037 2.018 1.013 1.034 0.013 0.029 1.009 1.032 0.011 0.026 1.033 0.011 0.027 0.013 0.038 8.047 6.032 5.040 4.044 3.044 2.040 1.032 0.019 0.047 7.037 5.023 4.027 3.027 2.024 1.018 1.041 0.024 6.029 5.041 4.043 3.041 2.034 1.025 0.012

Probability 0.025 0.01 2.020 1.005 1.014 0.002 0.007 0.007 0.017 − − − 3.011 2.002 2.012 1.002 1.009 1.009 1.024 0.004 0.010 − 0.025 − 2.006 2.006 1.005 1.005 1.017 0.002 0.007 0.007 0.017 − − − 2.018 1.003 1.013 0.001 0.005 0.005 0.013 − − − 1.009 1.009 0.004 0.004 0.011 − − − 0.003 0.003 0.011 − − − 0.013 − − − 7.019 6.007 5.013 4.005 4.017 3.006 3.018 2.006 2.017 1.005 1.013 0.002 0.007 0.007 0.019 − − − 6.014 5.005 5.023 4.008 3.010 3.010 2.009 2.009 2.024 1.007 1.018 0.003 0.009 0.009 0.024 − 5.010 5.010 4.015 3.005 3.016 2.005 2.014 1.003 1.010 1.010 1.025 0.005 0.012 −

0.005 1.005 0.002 − − − 2.002 1.002 0.001 0.004 − − 1.001 0.001 0.002 − − − 1.003 0.001 0.005 − − 0.001 0.004 − − 0.003 − − − − 5.002 4.005 2.002 1.001 1.005 0.002 − − − 5.005 3.002 2.003 1.002 0.001 0.003 − − 4.003 3.005 2.005 1.003 0.002 0.005 −

Contingency tables: 2 × 2 a A = 12 B = 10 5 B = 9 12 11 10 9 8 7 6 5 B = 8 12 11 10 9 8 7 6 B = 7 12 11 10 9 8 7 6 B = 6 12 11 10 9 8 7 6 B = 5 12 11 10 9 8 7 B = 4 12 11 10 9 8 B = 3 12 11 10 9 B = 2 12 11 A = 13 B = 13 13 12 11 10 9 8 7 6

0.05 0.030 5.021 4.028 3.027 2.022 1.015 1.035 0.017 0.039 5.049 3.017 3.048 2.037 1.024 0.010 0.024 4.036 3.036 2.028 1.017 1.038 0.016 0.034 3.025 2.021 1.012 1.030 0.011 0.025 0.050 2.015 1.010 1.027 0.009 0.020 0.041 2.050 1.026 0.008 0.019 0.038 1.029 0.009 0.022 0.044 0.011 0.033 9.048 7.034 6.043 5.048 4.049 3.048 2.043 1.034

Probability 0.025 0.01 − − 5.021 4.006 3.009 3.009 2.008 2.008 2.022 1.006 1.015 0.002 0.007 0.007 0.017 − − − 4.014 3.004 3.017 2.004 2.015 1.003 1.010 1.010 1.024 0.004 0.010 − 0.024 − 3.009 3.009 2.009 2.009 1.006 1.006 1.017 0.002 0.007 0.007 0.016 − − − 3.025 2.005 2.021 1.004 1.012 0.002 0.005 0.005 0.011 − 0.025 − − − 2.015 1.002 1.010 1.010 0.003 0.003 0.009 0.009 0.020 − − − 1.007 1.007 0.003 0.003 0.008 0.008 0.019 − − − 0.002 0.002 0.009 0.009 0.022 − − − 0.011 − − − 8.020 7.007 6.014 5.005 5.019 4.007 4.021 3.008 3.021 2.007 2.019 1.005 1.014 0.003 0.007 0.007

c 2000 by Chapman & Hall/CRC 

a 0.005 − A = 13 3.002 2.002 1.002 0.001 0.002 − − − 3.004 2.004 1.003 0.001 0.004 − − 2.002 1.002 0.001 0.002 − − − 2.005 1.004 0.002 0.005 − − − 1.002 0.001 0.003 − − − 0.001 0.003 − − − 0.002 − − − − − 6.003 4.002 3.002 2.002 1.002 0.001 0.003 −

B = 13 5 4 B = 12 13 12 11 10 9 8 7 6 5 B = 11 13 12 11 10 9 8 7 6 5 B = 10 13 12 11 10 9 8 7 6 5 B = 9 13 12 11 10 9 8 7 6 5 B = 8 13 12 11 10 9 8 7 6 B = 7 13 12 11 10 9 8 7 6 B = 6 13

0.05 0.020 0.048 8.039 6.025 5.030 4.032 3.030 2.026 1.019 1.043 0.024 7.031 6.045 5.049 4.048 3.044 2.036 1.026 0.013 0.030 6.024 5.032 4.033 3.030 2.024 1.016 1.035 0.017 0.038 5.017 4.022 3.020 3.048 2.036 1.023 1.048 0.023 0.049 5.042 4.045 3.038 2.027 1.016 1.035 0.015 0.032 4.031 3.029 2.021 2.048 1.027 0.010 0.022 0.044 3.021

Probability 0.025 0.01 0.020 − − − 7.015 6.005 6.025 5.010 4.012 3.004 3.012 2.004 2.011 1.003 1.008 1.008 1.019 0.004 0.010 0.010 0.024 − 6.011 5.003 5.017 4.006 4.020 3.007 3.019 2.006 2.016 1.004 1.011 0.002 0.005 0.005 0.013 − − − 6.024 5.007 4.011 3.003 3.011 2.003 2.010 2.010 2.024 1.006 1.016 0.003 0.007 0.007 0.017 − − − 5.017 4.005 4.022 3.006 3.020 2.006 2.016 1.004 1.010 1.010 1.023 0.004 0.010 − 0.023 − − − 4.012 3.003 3.013 2.003 2.011 1.002 1.006 1.006 1.016 0.002 0.006 0.006 0.015 − − − 3.007 3.007 2.007 2.007 2.021 1.004 1.012 0.002 0.004 0.004 0.010 − 0.022 − − − 3.021 2.004

0.005 − − 5.002 4.003 3.004 2.004 1.003 0.001 0.004 − − 5.003 3.002 2.002 1.001 1.004 0.002 0.005 − − 4.002 3.003 2.003 1.002 0.001 0.003 − − − 4.005 2.001 1.001 1.004 0.001 0.004 − − − 3.003 2.003 1.002 0.001 0.002 − − − 2.001 1.001 1.004 0.002 0.004 − − − 2.004

Contingency tables: 2 × 2 a A = 13 B = 6 12 11 10 9 8 7 B = 5 13 12 11 10 9 8 B = 4 13 12 11 10 9 B = 3 13 12 11 10 B = 2 13 12 A = 14 B = 14 14 13 12 11 10 9 8 7 6 5 4 B = 13 14 13 12 11 10 9 8 7 6 5 B = 12 14 13 12 11 10 9 8 7 6 5 B = 11 14

0.05 2.017 2.043 1.023 1.046 0.017 0.034 2.012 2.042 1.021 1.045 0.015 0.029 2.044 1.022 0.006 0.015 0.029 1.025 0.007 0.018 0.036 0.010 0.029 10.049 8.036 7.045 5.024 4.025 3.024 2.021 2.045 1.036 0.020 0.049 9.041 7.027 6.033 5.036 4.036 3.033 2.028 1.020 1.044 0.025 8.033 7.048 5.023 4.023 3.021 3.046 2.037 1.026 0.013 0.030 7.026

Probability 0.025 0.01 2.017 1.003 1.009 1.009 1.023 0.003 0.008 0.008 0.017 − − − 2.012 1.002 1.008 1.008 1.021 0.002 0.007 0.007 0.015 − − − 1.006 1.006 1.022 0.002 0.006 0.006 0.015 − − − 1.025 0.002 0.007 0.007 0.018 − − − 0.010 0.010 − − 9.020 8.008 7.015 6.006 6.021 5.008 5.024 4.010 4.025 3.010 3.024 2.008 2.021 1.006 1.015 0.003 0.008 0.008 0.020 − − − 8.016 7.006 6.011 5.004 5.014 4.005 4.015 3.005 3.014 2.004 2.012 1.003 1.008 1.008 1.020 0.004 0.010 − 0.025 − 7.012 6.004 6.020 5.007 5.023 4.008 4.023 3.008 3.021 2.007 2.017 1.005 1.012 0.002 0.005 0.005 0.013 − − − 6.009 6.009

c 2000 by Chapman & Hall/CRC 

a 0.005 1.003 A = 14 0.001 0.003 − − − 1.002 0.001 0.002 − − − 0.000 0.002 − − − 0.002 − − − − − 7.003 5.002 4.003 3.003 2.003 1.002 0.001 0.003 − − − 6.002 5.004 4.005 2.001 2.004 1.003 0.001 0.004 − − 6.004 4.002 3.003 2.002 1.002 1.005 0.002 − − − 5.003

B = 11 13 12 11 10 9 8 7 6 5 B = 10 14 13 12 11 10 9 8 7 6 5 B = 9 14 13 12 11 10 9 8 7 6 B = 8 14 13 12 11 10 9 8 7 6 B = 7 14 13 12 11 10 9 8 7 B = 6 14 13 12 11 10 9 8 7 B = 5 14 13

0.05 6.037 5.039 4.037 3.032 2.025 1.016 1.035 0.017 0.038 6.020 5.026 4.026 3.022 3.048 2.036 1.023 1.047 0.022 0.047 6.047 4.017 4.047 3.037 2.027 1.016 1.033 0.014 0.030 5.036 4.037 3.030 2.020 2.043 1.025 1.048 0.020 0.040 4.026 3.024 2.016 2.038 1.020 1.040 0.015 0.030 3.018 2.014 2.035 1.017 1.036 0.012 0.024 0.044 2.010 2.036

Probability 0.025 0.01 5.013 4.004 4.015 3.005 3.013 2.004 2.011 1.002 2.025 1.007 1.016 0.003 0.007 0.007 0.017 − − − 6.020 5.006 4.008 4.008 3.008 3.008 3.022 2.007 2.017 1.004 1.010 0.002 1.023 0.004 0.010 0.010 0.022 − − − 5.014 4.004 4.017 3.005 3.016 2.004 2.011 1.002 1.007 1.007 1.016 0.002 0.006 0.006 0.014 − − − 4.010 4.010 3.011 2.002 2.008 2.008 2.020 1.005 1.011 0.002 1.025 0.004 0.009 0.009 0.020 − − − 3.006 3.006 3.024 2.005 2.016 1.003 1.009 1.009 1.020 0.003 0.007 0.007 0.015 − − − 3.018 2.003 2.014 1.002 1.007 1.007 1.017 0.002 0.005 0.005 0.012 − 0.024 − − − 2.010 1.001 1.006 1.006

0.005 4.004 3.005 2.004 1.002 0.001 0.003 − − − 4.002 3.002 2.002 1.001 1.004 0.002 0.004 − − − 4.004 3.005 2.004 1.002 0.001 0.002 − − − 3.002 2.002 1.001 1.005 0.002 0.004 − − − 2.001 1.001 1.003 0.001 0.003 − − − 2.003 1.002 0.001 0.002 − − − − 1.001 0.001

Contingency tables: 2 × 2 a A = 14 B = 5 12 11 10 9 8 B = 4 14 13 12 11 10 9 B = 3 14 13 12 11 B = 2 14 13 12 A = 15 B = 15 15 14 13 12 11 10 9 8 7 6 5 4 B = 14 15 14 13 12 11 10 9 8 7 6 5 B = 13 15 14 13 12 11 10 9 8 7 6 5 B = 12 15 14 13

0.05 1.017 1.036 0.011 0.022 0.040 2.039 1.018 1.042 0.011 0.023 0.041 1.022 0.006 0.015 0.029 0.008 0.025 0.050 11.050 9.037 8.047 6.026 5.028 4.028 3.026 2.022 2.047 1.037 0.021 0.050 10.042 8.029 7.036 6.039 5.040 4.039 3.035 2.029 1.021 1.045 0.025 9.035 7.022 6.026 5.027 4.026 3.023 3.047 2.038 1.027 0.013 0.031 8.028 7.040 6.044

Probability 0.025 0.01 1.017 0.002 0.005 0.005 0.011 − 0.022 − − − 1.005 1.005 1.018 0.002 0.005 0.005 0.011 − 0.023 − − − 1.022 0.001 0.006 0.006 0.015 − − − 0.008 0.008 0.025 − − − 10.021 9.008 8.016 7.007 7.022 6.010 5.011 4.004 4.012 3.004 3.011 2.004 2.010 2.010 2.022 1.007 1.016 0.003 0.008 0.008 0.021 − − − 9.017 8.006 7.012 6.004 6.016 5.006 5.018 4.007 4.018 3.006 3.016 2.005 2.013 1.003 1.009 1.009 1.021 0.004 0.011 − − − 8.013 7.005 7.022 6.008 5.010 4.003 4.011 3.003 3.010 3.010 3.023 2.008 2.018 1.005 1.012 0.002 0.005 0.005 0.013 − − − 7.010 7.010 6.016 5.005 5.018 4.006

c 2000 by Chapman & Hall/CRC 

a 0.005 0.002 A = 15 0.005 − − − 1.005 0.002 0.005 − − − 0.001 − − − − − − 8.003 6.002 5.004 4.004 3.004 2.004 1.002 0.001 0.003 − − − 7.002 6.004 4.002 3.002 2.002 1.001 1.003 0.001 0.004 − − 7.005 5.003 4.003 3.003 2.003 1.002 1.005 0.002 − − − 6.003 4.002 3.002

B = 12 12 11 10 9 8 7 6 5 B = 11 15 14 13 12 11 10 9 8 7 6 5 B = 10 15 14 13 12 11 10 9 8 7 6 B = 9 15 14 13 12 11 10 9 8 7 6 B = 8 15 14 13 12 11 10 9 8 7 6 B = 7 15 14 13 12 11 10

0.05 5.043 4.039 3.033 2.025 1.016 1.035 0.017 0.037 7.022 6.030 5.031 4.028 3.023 3.048 2.036 1.023 1.045 0.022 0.046 6.017 5.021 4.020 4.047 3.037 2.026 1.015 1.031 0.013 0.028 6.042 5.044 4.038 3.029 2.020 2.040 1.023 1.044 0.019 0.037 5.032 4.031 3.024 2.016 2.033 1.018 1.035 0.013 0.026 0.050 4.023 3.020 3.049 2.030 1.015 1.030

Probability 0.025 0.01 4.017 3.006 3.015 2.004 2.011 1.003 1.007 1.007 1.016 0.003 0.007 0.007 0.017 − − − 7.022 6.007 5.011 4.003 4.011 3.003 3.010 3.010 3.023 2.007 2.017 1.004 1.010 0.002 1.023 0.004 0.010 0.010 0.022 − − − 6.017 5.005 5.021 4.007 4.020 3.006 3.017 2.005 2.012 1.003 1.007 1.007 1.015 0.002 0.006 0.006 0.013 − − − 5.012 4.003 4.014 3.004 3.012 2.003 2.008 2.008 2.020 1.005 1.011 0.002 1.023 0.004 0.009 0.009 0.019 − − − 4.008 4.008 3.008 3.008 3.024 2.006 2.016 1.003 1.008 1.008 1.018 0.003 0.006 0.006 0.013 − − − − − 4.023 3.005 3.020 2.004 2.013 1.002 1.006 1.006 1.015 0.002 0.005 0.005

0.005 2.001 2.004 1.003 0.001 0.003 − − − 5.002 4.003 3.003 2.003 1.002 1.004 0.002 0.004 − − − 5.005 3.002 2.001 2.005 1.003 0.001 0.002 − − − 4.003 3.004 2.003 1.002 1.005 0.002 0.004 − − − 3.002 2.002 1.001 1.003 0.001 0.003 − − − − 3.005 2.004 1.002 0.001 0.002 0.005

Contingency tables: 2 × 2 a A = 15 B = 7

B=6

B=5

B=4

B=3

B=2

A = 16 B = 16

B = 15

9 8 7 15 14 13 12 11 10 9 8 15 14 13 12 11 10 9 15 14 13 12 11 10 15 14 13 12 11 15 14 13 16 15 14 13 12 11 10 9 8 7 6 5 16 15 14 13 12 11 10 9 8 7 6

0.05 0.010 0.020 0.038 3.015 2.011 2.029 1.013 1.028 0.009 0.017 0.032 2.009 2.031 1.014 1.029 0.008 0.016 0.030 2.035 1.015 1.036 0.009 0.018 0.033 1.020 0.005 0.012 0.025 0.043 0.007 0.022 0.044 11.022 10.038 9.049 7.028 6.031 5.032 4.031 3.028 2.024 2.049 1.038 0.022 11.043 9.030 8.038 7.043 6.044 5.044 4.041 3.037 2.030 1.022 1.046

Probability 0.025 0.01 0.010 − 0.020 − − − 3.015 2.003 2.011 1.002 1.005 1.005 1.013 0.002 0.004 0.004 0.009 0.009 0.017 − − − 2.009 2.009 1.005 1.005 1.014 0.001 0.004 0.004 0.008 0.008 0.016 − − − 1.004 1.004 1.015 0.001 0.004 0.004 0.009 0.009 0.018 − − − 1.020 0.001 0.005 0.005 0.012 − 0.025 − − − 0.007 0.007 0.022 − − − 11.022 10.009 9.017 8.007 8.024 6.004 6.013 5.005 5.014 4.006 4.014 3.005 3.013 2.004 2.011 1.003 2.024 1.007 1.017 0.003 0.009 0.009 0.022 − 10.018 9.007 8.013 7.005 7.017 6.007 6.020 5.008 5.021 4.008 4.020 3.007 3.018 2.006 2.014 1.004 1.010 1.010 1.022 0.004 0.011 −

c 2000 by Chapman & Hall/CRC 

a 0.005 − A = 16 − − 2.003 1.002 0.001 0.002 0.004 − − − 1.001 1.005 0.001 0.004 − − − 1.004 0.001 0.004 − − − 0.001 0.005 − − − − − − 9.003 7.003 6.004 4.002 3.002 2.002 2.004 1.003 0.001 0.003 − − 8.002 6.002 5.003 4.003 3.003 2.002 1.001 1.004 0.002 0.004 −

B = 15 5 B = 14 16 15 14 13 12 11 10 9 8 7 6 5 B = 13 16 15 14 13 12 11 10 9 8 7 6 5 B = 12 16 15 14 13 12 11 10 9 8 7 6 5 B = 11 16 15 14 13 12 11 10 9 8 7 6 B = 10 16 15 14 13 12 11 10

0.05 0.026 10.037 8.024 7.029 6.031 5.030 4.028 3.024 3.048 2.039 1.027 0.013 0.031 9.030 8.043 7.048 6.048 5.045 4.040 3.034 2.026 1.017 1.035 0.017 0.037 8.024 7.034 6.036 5.034 4.030 3.024 3.047 2.035 1.022 1.044 0.021 0.044 7.019 6.025 5.025 4.022 4.046 3.036 2.025 2.048 1.030 0.013 0.027 7.046 5.018 5.046 4.038 3.028 2.019 2.037

Probability 0.025 0.01 − − 9.014 8.005 8.024 7.009 6.012 5.004 5.013 4.005 4.013 3.004 3.011 2.003 3.024 2.008 2.019 1.005 1.013 0.002 0.006 0.006 0.013 − − − 8.011 7.004 7.018 6.006 6.021 5.008 5.021 4.008 4.019 3.007 3.016 2.005 2.012 1.003 1.007 1.007 1.017 0.003 0.007 0.007 0.017 − − − 8.024 7.008 6.012 5.004 5.014 4.005 4.013 3.004 3.011 2.003 3.024 2.008 2.017 1.004 1.010 0.002 1.022 0.004 0.010 0.010 0.021 − − − 7.019 6.006 6.025 5.008 5.025 4.008 4.022 3.007 3.017 2.005 2.012 1.003 1.007 1.007 1.015 0.002 0.006 0.006 0.013 − − − 6.014 5.004 5.018 4.005 4.016 3.005 3.013 2.003 2.008 2.008 2.019 1.005 1.010 0.002

0.005 − 7.002 6.003 5.004 4.005 3.004 2.003 1.002 0.001 0.002 − − − 7.004 5.002 4.002 3.002 2.002 1.001 1.003 0.001 0.003 − − − 6.002 5.004 4.005 3.004 2.003 1.002 1.004 0.002 0.004 − − − 5.002 4.002 3.002 2.002 2.005 1.003 0.001 0.002 − − − 5.004 3.001 3.005 2.003 1.002 1.005 0.002

Contingency tables: 2 × 2 a A = 16 B = 10 9 8 7 6 B = 9 16 15 14 13 12 11 10 9 8 7 6 B = 8 16 15 14 13 12 11 10 9 8 7 B = 7 16 15 14 13 12 11 10 9 8 7 B = 6 16 15 14 13 12 11 10 9 8 B = 5 16 15 14 13 12 11 10 9 B = 4 16 15 14

0.05 1.022 1.041 0.017 0.035 6.037 5.038 4.031 3.023 3.047 2.030 1.016 1.031 0.012 0.024 0.045 5.028 4.026 3.019 3.043 2.026 2.049 1.026 1.047 0.017 0.033 4.020 3.017 3.042 2.024 2.047 1.023 1.041 0.014 0.026 0.047 3.013 3.044 2.024 2.049 1.022 1.041 0.012 0.023 0.040 3.048 2.027 1.011 1.024 1.045 0.012 0.023 0.039 2.032 1.013 1.031

Probability 0.025 0.01 1.022 0.004 0.008 0.008 0.017 − − − 5.010 5.010 4.011 3.003 3.009 3.009 3.023 2.006 2.015 1.003 1.008 1.008 1.016 0.002 0.006 0.006 0.012 − 0.024 − − − 4.007 4.007 3.007 3.007 3.019 2.005 2.012 1.002 1.006 1.006 1.013 0.002 0.004 0.004 0.009 0.009 0.017 − − − 4.020 3.004 3.017 2.003 2.010 1.002 2.024 1.005 1.011 0.001 1.023 0.003 0.007 0.007 0.014 − − − − − 3.013 2.002 2.009 2.009 2.024 1.004 1.011 0.001 1.022 0.003 0.006 0.006 0.012 − 0.023 − − − 2.008 2.008 1.004 1.004 1.011 0.001 1.024 0.003 0.006 0.006 0.012 − 0.023 − − − 1.004 1.004 1.013 0.001 0.003 0.003

c 2000 by Chapman & Hall/CRC 

a 0.005 0.004 A = 16 − − − 4.002 3.003 2.002 1.001 1.003 0.001 0.002 − − A = 17 − − 3.001 2.001 2.005 1.002 0.001 0.002 0.004 − − − 3.004 2.003 1.002 1.005 0.001 0.003 − − − − 2.002 1.001 1.004 0.001 0.003 − − − − 1.001 1.004 0.001 0.003 − − − − 1.004 0.001 0.003

B = 4 13 12 11 10 B = 3 16 15 14 13 12 B = 2 16 15 14 B = 17 17 16 15 14 13 12 11 10 9 8 7 6 5 B = 16 17 16 15 14 13 12 11 10 9 8 7 6 5 B = 15 17 16 15 14 13 12 11 10 9 8 7 6 5 B = 14 17 16 15 14

0.05 0.007 0.014 0.026 0.043 1.018 1.050 0.010 0.021 0.036 0.007 0.020 0.039 12.022 11.039 9.025 8.030 7.033 6.035 5.035 4.033 3.030 2.025 1.018 1.039 0.022 12.044 10.032 9.040 8.045 7.048 6.048 5.046 4.043 3.038 2.032 1.022 1.046 0.026 11.038 9.025 8.031 7.034 6.034 5.033 4.030 3.025 3.049 2.040 1.028 0.014 0.031 10.032 9.046 7.023 6.024

Probability 0.025 0.01 0.007 0.007 0.014 − − − − − 1.018 0.001 0.004 0.004 0.010 − 0.021 − − − 0.007 0.007 0.020 − − − 12.022 11.009 10.018 9.008 8.012 7.005 7.014 6.006 6.016 5.007 5.016 4.007 4.016 3.006 3.014 2.005 2.012 1.003 1.008 1.008 1.018 0.004 0.009 0.009 0.022 − 11.018 10.007 9.014 8.006 8.019 7.008 7.022 6.010 6.023 4.004 5.023 4.010 4.022 3.008 3.019 2.007 2.015 1.004 1.010 0.002 1.022 0.005 0.011 − − − 10.015 9.006 8.010 7.004 7.014 6.005 6.015 5.006 5.015 4.006 4.014 3.005 3.012 2.004 2.009 2.009 2.020 1.006 1.013 0.002 0.006 0.006 0.014 − − − 9.012 8.004 8.019 7.007 7.023 6.009 6.024 5.010

0.005 − − − − 0.001 0.004 − − − − − − 10.004 8.003 7.005 5.002 4.002 3.002 2.002 2.005 1.003 0.001 0.004 − − 9.003 7.002 6.003 5.004 4.004 3.003 2.003 1.002 1.004 0.002 0.005 − − 8.002 7.004 5.002 4.002 3.002 3.005 2.004 1.002 0.001 0.002 − − − 8.004 6.003 5.003 4.003

Contingency tables: 2 × 2 a A = 17 B = 14 13 12 11 10 9 8 7 6 5 B = 13 17 16 15 14 13 12 11 10 9 8 7 6 5 B = 12 17 16 15 14 13 12 11 10 9 8 7 6 B = 11 17 16 15 14 13 12 11 10 9 8 7 6 B = 10 17 16 15 14 13 12 11 10 9

0.05 5.023 5.047 4.041 3.034 2.026 2.050 1.035 0.017 0.036 9.026 8.037 7.040 6.039 5.035 4.030 3.024 3.046 2.035 1.022 1.043 0.021 0.043 8.021 7.028 6.029 5.027 4.023 4.045 3.035 2.025 2.046 1.029 0.012 0.026 7.016 6.021 5.020 5.045 4.037 3.027 2.018 2.035 1.020 1.039 0.016 0.033 7.041 6.044 5.039 4.030 3.022 3.043 2.028 1.015 1.029

Probability 0.025 0.01 5.023 4.009 4.021 3.007 3.017 2.005 2.013 1.003 1.008 1.008 1.017 0.003 0.007 0.007 0.017 − − − 8.009 8.009 7.014 6.005 6.016 5.006 5.016 4.006 4.014 3.005 3.011 2.003 3.024 2.008 2.018 1.005 1.011 0.002 1.022 0.004 0.010 0.010 0.021 − − − 8.021 7.007 6.010 5.003 5.011 4.003 4.010 4.010 4.023 3.008 3.018 2.005 2.012 1.003 2.025 1.007 1.015 0.002 0.006 0.006 0.012 − − − 7.016 6.005 6.021 5.007 5.020 4.006 4.017 3.005 3.013 2.003 2.008 2.008 2.018 1.004 1.010 1.010 1.020 0.004 0.008 0.008 0.016 − − − 6.012 5.003 5.014 4.004 4.013 3.003 3.010 3.010 3.022 2.006 2.014 1.003 1.007 1.007 1.015 0.002 0.005 0.005

c 2000 by Chapman & Hall/CRC 

a 0.005 3.003 A = 17 2.002 1.001 1.003 0.001 0.003 − − − 7.003 6.005 4.002 3.002 3.005 2.003 1.002 1.005 0.002 0.004 − − − 6.002 5.003 4.003 3.003 2.002 1.001 1.003 0.001 0.002 − − − 6.005 4.002 3.002 2.001 2.003 1.002 1.004 0.001 0.004 − − − 5.003 4.004 3.003 2.002 1.001 1.003 0.001 0.002 −

B = 10 8 7 6 B = 9 17 16 15 14 13 12 11 10 9 8 7 B = 8 17 16 15 14 13 12 11 10 9 8 7 B = 7 17 16 15 14 13 12 11 10 9 8 B = 6 17 16 15 14 13 12 11 10 9 8 B = 5 17 16 15 14 13 12 11 10 9 B = 4 17

0.05 0.011 0.022 0.042 6.032 5.033 4.026 3.018 3.038 2.023 2.043 1.023 1.041 0.016 0.030 5.024 4.022 3.016 3.035 2.020 2.039 1.019 1.035 0.012 0.022 0.040 4.017 3.014 3.035 2.019 2.038 1.017 1.032 0.010 0.019 0.033 3.011 3.038 2.020 2.042 1.017 1.032 0.009 0.017 0.030 0.050 3.043 2.023 1.009 1.020 1.037 0.010 0.018 0.030 0.049 2.029

Probability 0.025 0.01 0.011 − 0.022 − − − 5.008 5.008 4.009 4.009 3.007 3.007 3.018 2.005 2.011 1.002 2.023 1.005 1.012 0.002 1.023 0.004 0.008 0.008 0.016 − − − 5.024 4.006 4.022 3.005 3.016 2.004 2.009 2.009 2.020 1.004 1.010 1.010 1.019 0.003 0.006 0.006 0.012 − 0.022 − − − 4.017 3.003 3.014 2.003 2.008 2.008 2.019 1.004 1.008 1.008 1.017 0.002 0.005 0.005 0.010 0.010 0.019 − − − 3.011 2.002 2.008 2.008 2.020 1.003 1.008 1.008 1.017 0.002 0.005 0.005 0.009 0.009 0.017 − − − − − 2.006 2.006 2.023 1.003 1.009 1.009 1.020 0.002 0.005 0.005 0.010 0.010 0.018 − − − − − 1.003 1.003

0.005 − − − 4.002 3.002 2.002 2.005 1.002 0.001 0.002 0.004 − − − 3.001 2.001 2.004 1.002 1.004 0.001 0.003 − − − − 3.003 2.003 1.001 1.004 0.001 0.002 0.005 − − − 2.002 1.001 1.003 0.001 0.002 0.005 − − − − 1.001 1.003 0.001 0.002 0.005 − − − − 1.003

Contingency tables: 2 × 2 a A = 17 B = 4 16 15 14 13 12 11 B = 3 17 16 15 14 13 12 B = 2 17 16 15 A = 18 B = 18 18 17 16 15 14 13 12 11 10 9 8 7 6 5 B = 17 18 17 16 15 14 13 12 11 10 9 8 7 6 5 B = 16 18 17 16 15 14 13 12 11 10 9 8 7

0.05 1.011 1.027 0.006 0.012 0.021 0.035 1.016 1.045 0.009 0.018 0.031 0.049 0.006 0.018 0.035 13.023 12.040 10.026 9.032 8.035 7.037 6.038 5.037 4.035 3.032 2.026 1.019 1.040 0.023 13.045 11.033 10.042 9.048 7.026 6.026 5.025 5.049 4.045 3.040 2.032 1.023 1.047 0.026 12.039 10.027 9.033 8.037 7.038 6.037 5.035 4.031 3.026 3.050 2.040 1.028

Probability 0.025 0.01 1.011 0.001 0.003 0.003 0.006 0.006 0.012 − 0.021 − − − 1.016 0.001 0.004 0.004 0.009 0.009 0.018 − − − − − 0.006 0.006 0.018 − − − 13.023 12.010 11.019 10.008 9.012 8.005 8.015 7.007 7.017 6.008 6.019 5.008 5.019 4.008 4.017 3.007 3.015 2.005 2.012 1.003 1.008 1.008 1.019 0.004 0.010 0.010 0.023 − 12.019 11.008 10.015 9.006 9.020 8.009 8.024 6.004 6.012 5.005 5.012 4.004 4.011 3.004 4.023 3.009 3.020 2.007 2.016 1.005 1.011 0.002 1.023 0.005 0.011 − − − 11.016 10.006 9.011 8.004 8.015 7.006 7.017 6.007 6.018 5.007 5.017 4.007 4.015 3.006 3.013 2.004 2.010 2.010 2.020 1.006 1.013 0.002 0.006 0.006

c 2000 by Chapman & Hall/CRC 

a 0.005 0.001 A = 18 0.003 − − − − 0.001 0.004 − − − − − − − 11.004 9.003 7.002 6.003 5.003 4.003 3.003 2.002 1.001 1.003 0.001 0.004 − − 10.003 8.002 7.004 6.004 5.005 4.004 3.004 2.003 1.002 1.005 0.002 0.005 − − 9.002 8.004 6.002 5.003 4.003 3.002 2.002 2.004 1.002 0.001 0.002 −

B = 16 6 5 B = 15 18 17 16 15 14 13 12 11 10 9 8 7 6 5 B = 14 18 17 16 15 14 13 12 11 10 9 8 7 6 5 B = 13 18 17 16 15 14 13 12 11 10 9 8 7 6 B = 12 18 17 16 15 14 13 12 11 10 9 8 7

0.05 0.014 0.031 11.033 10.049 8.026 7.027 6.027 5.025 5.048 4.042 3.035 2.026 2.050 1.034 0.017 0.036 10.028 9.040 8.044 7.043 6.041 5.036 4.031 3.025 3.046 2.034 1.022 1.042 0.020 0.043 9.023 8.031 7.033 6.032 5.028 4.023 4.044 3.034 2.024 2.045 1.028 0.012 0.025 8.018 7.024 6.024 5.022 5.044 4.035 3.026 3.048 2.033 1.019 1.037 0.016

Probability 0.025 0.01 0.014 − − − 10.013 9.005 9.021 8.008 7.011 6.004 6.012 5.004 5.011 4.004 5.025 3.003 4.022 3.008 3.018 2.006 2.013 1.003 1.008 1.008 1.017 0.003 0.007 0.007 0.017 − − − 9.010 9.010 8.016 7.006 7.019 6.007 6.019 5.007 5.018 4.006 4.015 3.005 3.012 2.004 3.025 2.008 2.018 1.005 1.011 0.002 1.022 0.004 0.009 0.009 0.020 − − − 9.023 8.008 7.012 6.004 6.013 5.004 5.012 4.004 4.011 3.003 4.023 3.008 3.018 2.005 2.012 1.003 2.024 1.007 1.014 0.002 0.006 0.006 0.012 − − − 8.018 7.006 7.024 6.008 6.024 5.008 5.022 4.007 4.018 3.006 3.013 2.004 2.008 2.008 2.018 1.004 1.010 1.010 1.019 0.003 0.007 0.007 0.016 −

0.005 − − 9.005 7.003 6.004 5.004 4.004 3.003 2.002 1.001 1.003 0.001 0.003 − − − 8.003 6.002 5.002 4.002 3.002 2.001 2.004 1.002 1.005 0.002 0.004 − − − 7.002 6.004 5.004 4.004 3.003 2.002 1.001 1.003 0.001 0.002 − − − 6.002 5.002 4.003 3.002 2.001 2.004 1.002 1.004 0.001 0.003 − −

Contingency tables: 2 × 2 a A = 18 B = 12 6 B = 11 18 17 16 15 14 13 12 11 10 9 8 7 6 B = 10 18 17 16 15 14 13 12 11 10 9 8 7 6 B = 9 18 17 16 15 14 13 12 11 10 9 8 7 B = 8 18 17 16 15 14 13 12 11 10 9 8 7 B = 7 18 17 16 15

0.05 0.031 8.045 6.018 6.045 5.038 4.029 3.021 3.039 2.026 2.046 1.027 1.048 0.020 0.039 7.037 6.038 5.033 4.025 4.049 3.034 2.021 2.038 1.020 1.037 0.014 0.027 0.049 6.029 5.028 4.022 4.046 3.030 2.018 2.033 1.016 1.030 0.010 0.020 0.036 5.022 4.019 4.047 3.029 2.016 2.031 1.014 1.026 1.045 0.016 0.028 0.048 4.015 4.050 3.030 2.016

Probability 0.025 0.01 − − 7.014 6.004 6.018 5.005 5.016 4.005 4.013 3.004 3.010 3.010 3.021 2.006 2.013 1.003 1.007 1.007 1.014 0.002 0.005 0.005 0.010 − 0.020 − − − 6.010 5.003 5.012 4.003 4.010 3.003 4.025 3.007 3.017 2.005 2.010 1.002 2.021 1.005 1.010 0.001 1.020 0.003 0.007 0.007 0.014 − − − − − 5.007 5.007 4.008 4.008 4.022 3.006 3.015 2.003 2.008 2.008 2.018 1.004 1.008 1.008 1.016 0.002 0.005 0.005 0.010 − 0.020 − − − 5.022 4.005 4.019 3.004 3.013 2.003 2.007 2.007 2.016 1.003 1.007 1.007 1.014 0.002 0.004 0.004 0.008 0.008 0.016 − − − − − 4.015 3.003 3.012 2.002 2.007 2.007 2.016 1.003

c 2000 by Chapman & Hall/CRC 

a 0.005 − A = 18 6.004 4.001 3.001 3.004 2.002 1.001 1.003 0.001 0.002 0.005 − − − 5.003 4.003 3.003 2.002 2.005 1.002 1.005 0.001 0.003 − − − − 4.002 3.002 2.001 2.003 1.002 1.004 0.001 0.002 − − − − 4.005 3.004 2.003 1.001 1.003 A = 19 0.001 0.002 0.004 − − − − 3.003 2.002 1.001 1.003

B = 7 14 13 12 11 10 9 8 B = 6 18 17 16 15 14 13 12 11 10 9 B = 5 18 17 16 15 14 13 12 11 10 B = 4 18 17 16 15 14 13 12 11 B = 3 18 17 16 15 14 13 B = 2 18 17 16 B = 19 19 18 17 16 15 14 13 12 11 10 9 8

0.05 2.031 1.013 1.025 1.043 0.013 0.024 0.040 3.010 3.034 2.017 2.035 1.014 1.026 1.045 0.013 0.022 0.037 3.040 2.020 2.045 1.017 1.031 0.007 0.014 0.024 0.038 2.026 1.010 1.023 1.044 0.010 0.017 0.029 0.045 1.014 1.041 0.008 0.015 0.026 0.042 0.005 0.016 0.032 14.023 13.041 11.027 10.033 9.037 8.040 7.041 6.041 5.040 4.037 3.033 2.027

Probability 0.025 0.01 1.007 1.007 1.013 0.002 1.025 0.004 0.007 0.007 0.013 − 0.024 − − − 3.010 3.010 2.006 2.006 2.017 1.003 1.007 1.007 1.014 0.002 0.003 0.003 0.007 0.007 0.013 − 0.022 − − − 2.006 2.006 2.020 1.003 1.008 1.008 1.017 0.002 0.004 0.004 0.007 0.007 0.014 − 0.024 − − − 1.003 1.003 1.010 1.010 1.023 0.002 0.005 0.005 0.010 0.010 0.017 − − − − − 1.014 0.001 0.003 0.003 0.008 0.008 0.015 − − − − − 0.005 0.005 0.016 − − − 14.023 13.010 12.020 11.009 10.013 9.006 9.017 8.008 8.019 7.009 7.020 6.009 6.021 5.009 5.020 4.009 4.019 3.008 3.017 2.006 2.013 1.004 1.009 1.009

0.005 0.001 0.002 0.004 − − − − 2.001 1.001 1.003 0.001 0.002 0.003 − − − − 1.001 1.003 0.001 0.002 0.004 − − − − 1.003 0.001 0.002 0.005 − − − − 0.001 0.003 − − − − − − − 12.004 10.004 8.002 7.003 6.004 5.004 4.004 3.003 2.002 1.001 1.004 0.002

Contingency tables: 2 × 2 a A = 19 B = 19 7 6 5 B = 18 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 B = 17 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 B = 16 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 B = 15 19 18 17 16 15 14 13

0.05 1.020 1.041 0.023 14.046 12.034 11.044 10.050 8.028 7.029 6.029 5.027 4.025 4.046 3.041 2.033 1.023 1.047 0.027 13.040 11.028 10.035 9.039 8.041 7.041 6.039 5.036 4.032 3.027 2.021 2.040 1.028 0.014 0.031 12.035 10.023 9.028 8.030 7.030 6.029 5.026 5.049 4.042 3.035 2.027 2.049 1.034 0.017 0.036 11.029 10.042 9.047 8.048 7.045 6.042 5.037

Probability 0.025 0.01 1.020 0.004 0.010 0.010 0.023 − 13.020 12.008 11.016 10.007 10.021 9.010 8.012 7.005 7.013 6.006 6.014 5.006 5.013 4.005 4.012 3.004 4.025 2.003 3.021 2.008 2.017 1.005 1.011 0.002 1.023 0.005 0.012 − − − 12.016 11.006 10.012 9.005 9.016 8.007 8.019 7.008 7.020 6.009 6.020 5.008 5.019 4.008 4.016 3.006 3.014 2.004 2.010 1.003 2.021 1.006 1.014 0.002 0.006 0.006 0.014 − − − 11.013 10.005 10.023 9.009 8.012 7.005 7.013 6.005 6.013 5.005 5.013 4.005 4.011 3.004 4.023 3.009 3.018 2.006 2.013 1.004 1.008 1.008 1.017 0.003 0.007 0.007 0.017 − − − 10.011 9.004 9.018 8.007 8.021 7.008 7.022 6.009 6.021 5.008 5.019 4.007 4.016 3.006

c 2000 by Chapman & Hall/CRC 

a 0.005 0.004 A = 19 − − 11.003 9.003 8.004 6.002 5.002 4.002 3.002 3.004 2.003 1.002 1.005 0.002 0.005 − − 10.002 9.005 7.003 6.003 5.003 4.003 3.003 2.002 2.004 1.003 0.001 0.002 − − − 10.005 8.003 7.005 5.002 4.002 4.005 3.004 2.003 1.001 1.004 0.001 0.003 − − − 9.004 7.002 6.003 5.003 4.003 3.002 2.002

B = 15 12 11 10 9 8 7 6 5 B = 14 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 B = 13 19 18 17 16 15 14 13 12 11 10 9 8 7 6 B = 12 19 18 17 16 15 14 13 12 11 10 9 8 7 6 B = 11 19 18 17 16

0.05 4.031 3.025 3.045 2.034 1.022 1.042 0.020 0.042 10.024 9.034 8.037 7.036 6.033 5.028 4.023 4.043 3.034 2.024 2.043 1.027 0.012 0.024 0.049 9.020 8.027 7.028 6.026 5.022 5.043 4.034 3.025 3.046 2.032 1.019 1.035 0.015 0.030 9.049 7.020 6.020 6.044 5.036 4.028 3.020 3.037 2.024 2.043 1.025 1.045 0.019 0.037 8.041 7.044 6.039 5.031

Probability 0.025 0.01 3.012 2.004 3.025 2.009 2.018 1.005 1.011 0.002 1.022 0.004 0.009 0.009 0.020 − − − 10.024 9.008 8.013 7.005 7.015 6.006 6.015 5.005 5.014 4.005 4.011 3.004 4.023 3.008 3.018 2.006 2.012 1.003 2.024 1.007 1.014 0.002 0.005 0.005 0.012 − 0.024 − − − 9.020 8.006 7.010 7.010 6.010 5.003 5.010 5.010 5.022 4.008 4.018 3.006 3.013 2.004 2.008 2.008 2.017 1.004 1.009 1.009 1.019 0.003 0.007 0.007 0.015 − − − 8.016 7.005 7.020 6.007 6.020 5.007 5.017 4.006 4.014 3.004 3.010 3.010 3.020 2.006 2.013 1.003 2.024 1.006 1.013 0.002 1.025 0.005 0.010 0.010 0.019 − − − 7.012 6.003 6.015 5.004 5.013 4.004 4.011 3.003

0.005 2.004 1.002 1.005 0.002 0.004 − − − 8.003 7.005 5.002 4.002 4.005 3.004 2.002 1.001 1.003 0.001 0.002 − − − − 7.002 6.003 5.003 4.003 3.002 2.002 2.004 1.002 1.004 0.001 0.003 − − − 7.005 5.002 4.002 3.002 3.004 2.003 1.001 1.003 0.001 0.002 0.005 − − − 6.003 5.004 4.004 3.003

Contingency tables: 2 × 2 a A = 19 B = 11 15 14 13 12 11 10 9 8 7 6 B = 10 19 18 17 16 15 14 13 12 11 10 9 8 7 B = 9 19 18 17 16 15 14 13 12 11 10 9 8 7 B = 8 19 18 17 16 15 14 13 12 11 10 9 8 B = 7 19 18 17 16 15 14 13

0.05 4.023 4.044 3.031 2.019 2.035 1.019 1.034 0.013 0.025 0.046 7.033 6.034 5.028 4.020 4.041 3.027 3.048 2.029 1.015 1.027 1.046 0.018 0.032 6.026 5.024 4.018 4.039 3.025 3.045 2.026 2.045 1.022 1.039 0.013 0.024 0.043 5.019 4.016 4.040 3.024 3.046 2.025 2.044 1.020 1.035 0.011 0.020 0.034 4.013 4.044 3.026 2.013 2.026 2.046 1.020

Probability 0.025 0.01 4.023 3.007 3.016 2.004 2.010 2.010 2.019 1.005 1.010 1.010 1.019 0.003 0.006 0.006 0.013 − 0.025 − − − 6.009 6.009 5.010 4.003 4.008 4.008 4.020 3.006 3.013 2.003 2.008 2.008 2.016 1.003 1.007 1.007 1.015 0.002 0.005 0.005 0.009 0.009 0.018 − − − 5.006 5.006 5.024 4.006 4.018 3.005 3.012 2.003 3.025 2.006 2.014 1.003 1.006 1.006 1.012 0.002 1.022 0.004 0.007 0.007 0.013 − 0.024 − − − 5.019 4.004 4.016 3.004 3.011 2.002 3.024 2.006 2.013 1.002 2.025 1.005 1.011 0.001 1.020 0.003 0.006 0.006 0.011 − 0.020 − − − 4.013 3.002 3.010 2.002 2.005 2.005 2.013 1.002 1.005 1.005 1.011 0.001 1.020 0.003

c 2000 by Chapman & Hall/CRC 

a 0.005 2.002 A = 19 2.004 1.002 1.005 0.001 0.003 − − − − 5.002 4.003 3.002 2.001 2.003 1.001 1.003 0.001 0.002 0.005 − − − 4.001 3.001 3.005 2.003 1.001 1.003 0.001 0.002 0.004 − − − − 4.004 3.004 2.002 1.001 1.002 0.001 0.001 0.003 − A = 20 − − − 3.002 2.002 1.001 1.002 0.001 0.001 0.003

B = 7 12 11 10 9 8 B = 6 19 18 17 16 15 14 13 12 11 10 9 B = 5 19 18 17 16 15 14 13 12 11 10 B = 4 19 18 17 16 15 14 13 12 B = 3 19 18 17 16 15 14 B = 2 19 18 17 16 B = 20 20 19 18 17 16 15 14 13 12 11 10

0.05 1.034 0.010 0.017 0.030 0.048 4.050 3.030 2.014 2.030 1.011 1.021 1.037 0.010 0.017 0.028 0.045 3.036 2.018 2.040 1.014 1.026 1.044 0.011 0.019 0.030 0.047 2.024 1.009 1.020 1.038 0.008 0.014 0.024 0.037 1.013 1.037 0.006 0.013 0.023 0.036 0.005 0.014 0.029 0.048 15.024 14.042 12.028 11.034 10.039 9.041 8.043 7.044 6.043 5.041 4.039

Probability 0.025 0.01 0.005 0.005 0.010 0.010 0.017 − − − − − 3.009 3.009 2.005 2.005 2.014 1.002 1.005 1.005 1.011 0.001 1.021 0.003 0.005 0.005 0.010 0.010 0.017 − − − − − 2.005 2.005 2.018 1.002 1.006 1.006 1.014 0.001 0.003 0.003 0.006 0.006 0.011 − 0.019 − − − − − 2.024 1.002 1.009 1.009 1.020 0.002 0.004 0.004 0.008 0.008 0.014 − 0.024 − − − 1.013 0.001 0.003 0.003 0.006 0.006 0.013 − 0.023 − − − 0.005 0.005 0.014 − − − − − 15.024 13.004 13.020 12.009 11.014 10.006 10.018 9.008 9.020 8.010 8.022 6.005 7.023 5.005 6.023 4.004 5.022 4.010 4.020 3.008 3.018 2.006

0.005 − − − − − 2.001 1.001 1.002 0.000 0.001 0.003 − − − − − 2.005 1.002 0.000 0.001 0.003 − − − − − 1.002 0.001 0.002 0.004 − − − − 0.001 0.003 − − − − 0.005 − − − 13.004 11.004 9.003 8.004 7.004 6.005 5.005 4.004 3.004 2.003 1.002

Contingency tables: 2 × 2 a A = 20 B = 20 9 8 7 6 5 B = 19 20 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 B = 18 20 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 B = 17 20 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 B = 16 20 19

0.05 3.034 2.028 1.020 1.042 0.024 15.047 13.035 12.045 10.027 9.030 8.031 7.031 6.031 5.029 4.026 4.047 3.042 2.034 1.024 1.048 0.027 14.041 12.029 11.037 10.041 9.044 8.044 7.043 6.041 5.038 4.033 3.028 2.021 2.041 1.029 0.014 0.031 13.036 11.024 10.030 9.032 8.033 7.032 6.030 5.027 5.049 4.043 3.035 2.027 2.049 1.034 0.017 0.036 12.031 11.045

Probability 0.025 0.01 2.014 1.004 1.009 1.009 1.020 0.004 0.010 − 0.024 − 14.020 13.008 12.016 11.007 11.023 9.004 9.013 8.006 8.015 7.006 7.015 6.007 6.015 5.007 5.014 4.006 4.013 3.005 3.011 2.004 3.022 2.008 2.017 1.005 1.011 0.002 1.024 0.005 0.012 − − − 13.017 12.007 11.013 10.005 10.018 9.008 9.020 8.009 8.022 7.010 7.022 5.004 6.021 5.009 5.020 4.008 4.017 3.007 3.014 2.005 2.010 1.003 2.021 1.006 1.014 0.003 0.006 0.006 0.014 − − − 12.014 11.005 11.024 9.004 9.013 8.005 8.015 7.006 7.015 6.006 6.015 5.006 5.014 4.005 4.012 3.004 4.023 3.009 3.019 2.006 2.014 1.004 1.008 1.008 1.017 0.003 0.008 0.008 0.017 − − − 11.012 10.004 10.019 9.008

c 2000 by Chapman & Hall/CRC 

a 0.005 1.004 A = 20 0.002 0.004 − − 12.003 10.003 9.004 7.002 6.003 5.003 4.002 3.002 3.005 2.004 1.002 0.001 0.002 − − − 11.003 9.002 8.003 7.004 6.004 5.004 4.004 3.003 2.002 2.005 1.003 0.001 0.003 − − − 10.002 9.004 7.002 6.002 5.002 4.002 3.002 3.004 2.003 1.002 1.004 0.001 0.003 − − − 10.004 8.003

B = 16 18 17 16 15 14 13 12 11 10 9 8 7 6 5 B = 15 20 19 18 17 16 15 14 13 12 11 10 9 8 7 6 5 B = 14 20 19 18 17 16 15 14 13 12 11 10 9 8 7 6 B = 13 20 19 18 17 16 15 14 13 12 11

0.05 9.023 8.024 8.050 7.047 6.042 5.037 4.031 3.025 3.045 2.034 1.021 1.041 0.020 0.041 11.026 10.037 9.040 8.040 7.037 6.033 5.029 4.023 4.042 3.033 2.023 2.042 1.027 1.049 0.024 0.048 10.022 9.030 8.031 7.030 6.027 5.022 5.042 4.033 3.025 3.044 2.031 1.018 1.034 0.014 0.029 9.017 8.023 7.023 6.021 6.043 5.035 4.027 4.048 3.035 2.023

Probability 0.025 0.01 9.023 8.010 8.024 6.004 7.024 5.004 6.022 5.009 5.020 4.008 4.016 3.006 3.013 2.004 3.025 2.009 2.018 1.005 1.011 0.002 1.021 0.004 0.009 0.009 0.020 − − − 10.009 10.009 9.015 8.005 8.017 7.007 7.018 6.007 6.016 5.006 5.014 4.005 4.012 3.004 4.023 3.009 3.018 2.006 2.012 1.003 2.023 1.007 1.014 0.002 0.005 0.005 0.012 − 0.024 − − − 10.022 9.007 8.011 7.004 7.012 6.004 6.012 5.004 5.010 4.003 5.022 4.008 4.018 3.006 3.013 2.004 3.025 2.008 2.016 1.004 1.009 1.009 1.018 0.003 0.007 0.007 0.014 − − − 9.017 8.005 8.023 7.008 7.023 6.008 6.021 5.008 5.018 4.006 4.014 3.004 3.010 3.010 3.019 2.006 2.012 1.003 2.023 1.006

0.005 7.004 6.004 5.004 4.003 3.003 2.002 2.004 1.002 1.005 0.002 0.004 − − − 9.003 7.002 6.002 5.002 4.002 3.002 3.004 2.003 1.001 1.003 0.001 0.002 − − − − 8.002 7.004 6.004 5.004 4.003 3.003 2.002 2.004 1.002 1.004 0.001 0.003 − − − 7.002 6.002 5.003 4.002 3.002 3.004 2.003 1.001 1.003 0.001

Contingency tables: 2 × 2 a A = 20 B = 13 10 9 8 7 6 B = 12 20 19 18 17 16 15 14 13 12 11 10 9 8 7 6 B = 11 20 19 18 17 16 15 14 13 12 11 10 9 8 7 B = 10 20 19 18 17 16 15 14 13 12 11 10 9 8 7 B = 9 20 19 18 17 16 15 14

0.05 2.041 1.024 1.042 0.018 0.035 9.044 8.049 7.045 6.038 5.030 4.022 4.041 3.028 3.049 2.032 1.017 1.031 0.012 0.023 0.043 8.037 7.039 6.033 5.026 4.019 4.036 3.024 3.043 2.026 2.045 1.024 1.042 0.016 0.029 7.030 6.029 5.024 5.049 4.034 3.022 3.039 2.022 2.039 1.019 1.034 0.012 0.022 0.038 6.023 5.021 4.015 4.033 3.020 3.038 2.021

Probability 0.025 0.01 1.012 0.002 1.024 0.004 0.009 0.009 0.018 − − − 8.014 7.004 7.017 6.005 6.017 5.005 5.014 4.004 4.011 3.003 4.022 3.007 3.015 2.004 2.009 2.009 2.018 1.004 1.009 1.009 1.017 0.003 0.006 0.006 0.012 − 0.023 − − − 7.010 6.003 6.013 5.004 5.011 4.003 4.008 4.008 4.019 3.006 3.012 2.003 3.024 2.007 2.014 1.003 1.007 1.007 1.013 0.002 1.024 0.004 0.008 0.008 0.016 − − − 6.008 6.008 5.008 5.008 5.024 4.007 4.017 3.005 3.011 2.003 3.022 2.006 2.012 1.002 2.022 1.005 1.011 0.001 1.019 0.003 0.006 0.006 0.012 − 0.022 − − − 6.023 5.005 5.021 4.005 4.015 3.004 3.010 3.010 3.020 2.005 2.011 1.002 2.021 1.004

c 2000 by Chapman & Hall/CRC 

a 0.005 0.002 A = 20 0.004 − − − 7.004 5.002 4.001 4.004 3.003 2.002 2.004 1.002 1.004 0.001 0.003 − − − − 6.003 5.004 4.003 3.002 2.001 2.003 1.001 1.003 0.001 0.002 0.004 − − − 5.002 4.002 3.002 3.005 2.003 1.001 1.002 0.001 0.001 0.003 − − − − 4.001 3.001 3.004 2.002 1.001 1.002 1.004

B = 9 13 12 11 10 9 8 7 B = 8 20 19 18 17 16 15 14 13 12 11 10 9 8 B = 7 20 19 18 17 16 15 14 13 12 11 10 9 B = 6 20 19 18 17 16 15 14 13 12 11 10 B = 5 20 19 18 17 16 15 14 13 12 11 B = 4 20 19

0.05 2.036 1.017 1.029 1.048 0.017 0.029 0.050 5.017 4.014 4.035 3.021 3.039 2.020 2.036 1.015 1.027 1.044 0.014 0.024 0.041 4.012 4.040 3.022 3.045 2.022 2.039 1.016 1.027 1.044 0.013 0.022 0.036 4.046 3.027 2.012 2.026 2.047 1.018 1.030 1.048 0.013 0.022 0.035 3.033 2.016 2.036 1.012 1.022 1.038 0.009 0.015 0.024 0.038 2.022 1.008

Probability 0.025 0.01 1.009 1.009 1.017 0.002 0.005 0.005 0.009 0.009 0.017 − − − − − 5.017 4.003 4.014 3.003 3.009 3.009 3.021 2.005 2.010 1.002 2.020 1.004 1.008 1.008 1.015 0.002 0.004 0.004 0.008 0.008 0.014 − 0.024 − − − 4.012 3.002 3.009 3.009 3.022 2.004 2.011 1.002 2.022 1.004 1.008 1.008 1.016 0.002 0.004 0.004 0.007 0.007 0.013 − 0.022 − − − 3.008 3.008 2.005 2.005 2.012 1.002 1.004 1.004 1.009 1.009 1.018 0.002 0.004 0.004 0.007 0.007 0.013 − 0.022 − − − 2.004 2.004 2.016 1.002 1.005 1.005 1.012 0.001 1.022 0.002 0.005 0.005 0.009 0.009 0.015 − 0.024 − − − 2.022 1.002 1.008 1.008

0.005 0.001 0.002 0.005 − − − − 4.003 3.003 2.002 2.005 1.002 1.004 0.001 0.002 0.004 − − − − 3.002 2.001 2.004 1.002 1.004 0.001 0.002 0.004 − − − − 2.001 2.005 1.002 1.004 0.001 0.002 0.004 − − − − 2.004 1.002 0.000 0.001 0.002 0.005 − − − − 1.002 0.000

10.11

DETERMINING VALUES IN BERNOULLI TRIALS

Suppose the probability of heads for a biased coin is either (H0 ) p = α or (H1 ) p = β (with α < β). Assume the values α and β are known and the purpose of an experiment is to determine the true value of p. Toss the coin repeatedly and let the random variable Y be the number of tosses until the rth head appears. Let 1 − θ be the probability the identification is correct under either hypothesis and define N by 1 − θ ≈ Prob [Y ≤ N | p = β] ≈ Prob [Y > N | p = α]

(10.13)

This hypothesis test has significance level θ and power 1 − θ. The random variable Y has a negative binomial distribution with mean r/p and variance r(1 − p)/p2 but may be approximated by a normal distribution. If ξ is defined by Φ(ξ) = θ, then  ?2 √ ξ >  r≈ α 1−β+β 1−α β−α (10.14)  r − ξ r(1 − α) N≈ α Example 10.62 : If θ = 0.05, equation (10.14) yields the values below. In this table, E [T ] is the expected number of tosses required to reach a decision. α 0.1 0.3 0.5 0.7 0.1 0.4

β 0.2 0.4 0.6 0.8 0.6 0.9

r 21 87 148 153 4 7

N 139 247 268 203 10 9

E [T ] 122 232 257 197 8 8

See G. J. Manas and D. H. Meyer, “On a problem of coin identification,” SIAM Review, 31, Number 1, March 1989, SIAM, Philadelphia, pages 114–117.

c 2000 by Chapman & Hall/CRC 

CHAPTER 11

Regression Analysis Contents 11.1

Simple linear regression 11.1.1 Least squares estimates 11.1.2 Sum of squares 11.1.3 Inferences regarding regression coefficients 11.1.4 The mean response 11.1.5 Prediction interval 11.1.6 Analysis of variance table 11.1.7 Test for linearity of regression 11.1.8 Sample correlation coefficient 11.1.9 Example 11.2 Multiple linear regression 11.2.1 Least squares estimates 11.2.2 Sum of squares 11.2.3 Inferences regarding regression coefficients 11.2.4 The mean response 11.2.5 Prediction interval 11.2.6 Analysis of variance table 11.2.7 Sequential sum of squares 11.2.8 Partial F test 11.2.9 Residual analysis 11.2.10 Example 11.3 Orthogonal polynomials 11.3.1 Tables for orthogonal polynomials

11.1

SIMPLE LINEAR REGRESSION

Let (x1 , y1 ), (x2 , y2 ), . . . , (xn , yn ) be n pairs of observations such that yi is an observed value of the random variable Yi . Assume there exist constants β0 and β1 such that Yi = β0 + β1 xi + 3i

c 2000 by Chapman & Hall/CRC 

(11.1)

where 31 , 32 , . . . , 3n are independent, normal random variables having mean 0 and variance σ 2 . Assumptions In terms of Yi ’s

In terms of 3i ’s

3i ’s are normally distributed

Yi ’s are normally distributed

E [3i ] = 0

E [Yi ] = β0 + β1 xi

Var [3i ] = σ

2

Var [Yi ] = σ 2

Cov [3i , 3j ] = 0, i = j

Cov [Yi , Yj ] = 0, i = j

Principle of least squares: The sum of squared deviations about the true regression line is S(β0 , β1 ) =

n 

[yi − (β0 + β1 xi )]2 .

(11.2)

i=1

The point estimates of β0 and β1 , denoted by β.0 and β.1 , are those values that minimize S(β0 , β1 ). The estimates β.0 and β.1 are called the least squares estimates. The estimated regression line or least squares line is y. = β.0 + β.1 x. The normal equations for β.0 and β.1 are n

n   yi = β.0 n + β.1 xi i=1

n 

xi yi

= β.0

i=1

Sxx =

(xi − x) = 2

i=1

Syy =

n 

(yi − y) =

i=1

Sxy

n 

x2i

i=1 2

i=1

+ β.1

xi

i=1

Notation: n 

n 

n 

yi2

i=1

n 

n 

1 − n 1 − n

x2i

i=1

2 xi

i=1

n 2  yi

n 

i=1

1 = (xi − x)(yi − y)2 = xi yi − n i=1 i=1

c 2000 by Chapman & Hall/CRC 

n 

n  i=1

n  xi yi i=1

(11.3)

11.1.1

Least squares estimates  Sxy β.1 = = Sxx

n



n 

xi yi

i=1



n

n  i=1

n 

β.0 =

n  yi − β.1 xi

i=1

i=1

n

 −

xi

i=1

 x2i



n 

n 

 yi

i=1



n 



2 xi

(11.4)

i=1

= y − β.1 x

The ith predicted (fitted) value: y.i = β.0 + β.1 xi (for i = 1, 2, . . . , n). The ith residual: ei = yi − y.i (for i = 1, 2, . . . , n). Properties: (1) E [β.1 ] = β1 ,

σ2 σ2 Var [β.1 ] =  = n Sxx (xi − x)2 i=1

(2) E [β.0 ] = β0 ,

n 

σ2

Var [β.0 ] = n

n 

σ2

xi

i=1

n 

xi

i=1

=

nSxx

(xi − x)2

=

σ2 x Sxx

i=1

(3) β.0 and β.1 are normally distributed. 11.1.2 n  i=1



Sum of squares

(yi − y)2 = 



n  i=1



SST

(. yi − y)2 + 



n 

(yi − y.i )2

i=1



SSR





SSE

SST = total sum of squares = Syy SSR = sum of squares due to regression = β.2 Sxy SSE = sum of squares due to error n n n n     = [yi − (β.0 + β.1 xi )]2 = yi2 − β.0 yi − β.1 xi yi i=1

i=1

i=1

i=1

= Syy − 2β.1 Sxy + β.12 Sxx = Syy − β.12 Sxx = Syy − β.1 Sxy σ .2 = s2 =

SSE , n−2

  E S 2 = σ2

Sample coefficient of determination: r2 =

c 2000 by Chapman & Hall/CRC 

SSR SSE =1− SST SST

11.1.3

Inferences concerning the regression coefficients

The parameter β.1 (1) T =

β.1 − β1 β.1 − β1 √ = Sβ.1 S/ Sxx

has a t distribution with n − 2 degrees of freedom, where √ Sβ.1 = S/ Sxx is an estimate for the standard deviation of β.1 . (2) A 100(1 − α)% confidence interval for β1 has as endpoints β.1 ± tα/2,n−2 · s ˆ β1

(3) Hypothesis test: Null hypothesis

Alternative hypotheses

Test statistic

Rejection regions

β1 = β10

β1 > β10 β1 < β10 β1 = β10

β.1 − β10 T = Sβ.1

T ≥ tα,n−2 T ≤ −tα,n−2 |T | ≥ tα/2,n−2

(1) (2) (3)

The parameter β.0 +

(1) T = S

β.0 − β0 n  i=1

=

x2i /nSxx

β.0 − β0 Sβ.0

has a t distribution with n − 2 degrees of freedom, where Sβ.0 denotes the estimate for the standard deviation of β.0 . (2) A 100(1 − α)% confidence interval for β1 has as endpoints β.1 ± tα/2,n−2 · s ˆ β0

(3) Hypothesis test: Null hypothesis

Alternative hypotheses

Test statistic

β0 = β00

β0 > β00 β0 < β00 β0 = β00

T =

11.1.4

β.0 − β00 Sβ.0

Rejection regions T ≥ tα,n−2 T ≤ −tα,n−2 |T | ≥ tα/2,n−2

(1) (2) (3)

The mean response

The mean response of Y given x = x0 is µY |x0 = β0 + β1 x0 . The random variable Y.0 = β.0 + β.1 x0 is used to estimate µY |x0 .

c 2000 by Chapman & Hall/CRC 

(1) E [Y.0 ] = β0 + β1 x0   (x0 − x)2 1 + (2) Var [Y.0 ] = σ 2 n Sxx (3) Y.0 has a normal distribution. (4) T =



Y.0 − µY |x0

=

Y.0 − µY |x0 SY.0

S (1/n) + [(x0 − x)2 /Sxx ] has a t distribution with n − 2 degrees of freedom, where SY.0 denotes the estimate for the standard deviation of Y.0 .

(5) A 100(1 − α)% confidence interval for µY |x0 has as endpoints y.0 ± tα/2,n−2 · sY.0 . (6) Hypothesis test: Null hypothesis

Alternative hypotheses

Test statistic

β0 + β1 x0 = y0 = µ0

y0 > µ0 y0 < µ0 y0 = µ0

T =

11.1.5

Y.0 − µ0 SY.0

Rejection regions T ≥ tα,n−2 T ≤ −tα,n−2 |T | ≥ tα/2,n−2

(1) (2) (3)

Prediction interval

A prediction interval for a value y0 of the random variable Y0 = β0 + β1 x0 + 30 is obtained by considering the random variable Y.0 − Y0 . (1) E [Y.0 − Y0 ] = 0

  1 (x0 − x)2 (2) Var [Y.0 − Y0 ] = σ 2 1 + + n Sxx (3) Y.0 − Y0 has a normal distribution. (4) T =



Y.0 − Y0

=

Y.0 − Y0 SY.0 −Y0

S 1 + (1/n) + [(x0 − has a t distribution with n − 2 degrees of freedom. x)2 /Sxx ]

(5) A 100(1 − α)% prediction interval for y0 has as endpoints y.0 ± tα/2,n−2 · sY.0 −Y0

c 2000 by Chapman & Hall/CRC 

11.1.6

Analysis of variance table

Source of variation

Sum of Degrees of Mean squares freedom square

Regression

SSR

1

Error

SSE

n−2

Total

SST

n−1

Computed F

SSR 1 SSE MSE = n−2

MSR =

MSR/MSE

Hypothesis test of significant regression: Null hypothesis

Alternative Test hypothesis statistic

β1 = 0

β1 = 0

11.1.7

Rejection region

F = MSR/MSE F ≥ Fα,1,n−2

Test for linearity of regression

Suppose there are k distinct values of x, {x1 , x2 , . . . , xk }, ni observations for xi , and n = n1 + n2 + · · · + nk . Definitions: (1) yij = the j th observation on the random variable Yi . ni  (2) Ti = yij , y i. = Ti /ni j=1

(3) SSPE = sum of squares due to pure error =

ni k  

(yij − y i. )2 =

i=1 j=1

ni k  

2 yij −

i=1 j=1

k  T2 i

i=1

ni

(4) SSLF = Sum of squares due to lack of fit = SSE − SSPE Hypothesis test: Null hypothesis

Alternative Test hypothesis statistic

Linear regression Lack of fit 11.1.8

F =

SSLF/(k − 2) SSPE/(n − k)

Rejection region F ≥ Fα,k−2,n−k

Sample correlation coefficient

The sample correlation coefficient is a measure of linear association and is defined by + Sxx Sxy . r = β1 = . (11.5) Syy Sxx Syy

c 2000 by Chapman & Hall/CRC 

Hypothesis tests: Null hypothesis

Alternative hypothesis

Test statistic

Rejection region

ρ=0

ρ>0 ρ<0 ρ = 0

√ R n−2 β.1 T = √ = Sβ.1 1 − R2

T ≥ tα,n−2 T ≤ −tα,n−2 |T | ≥ tα/2,n−2

If X and Y have a bivariate normal distribution: √ ρ = ρ0 ρ > ρ0 n−3 Z= ρ < ρ0 2   ρ = ρ0 (1 + R)(1 − ρ0 ) × ln (1 − R)(1 + ρ0 ) 11.1.9

Z ≥ zα Z ≤ −zα |Z| ≥ zα/2

(1) (2) (3)

(4) (5) (6)

Example

Example 11.63 : A recent study at a manufacturing facility examined the relationship between the noon temperature (Fahrenheit) inside the plant and the number of defective items produced during the day shift. The data are given in the following table. Noon temperature (x)

Number defective (y)

Noon temperature (x)

Number defective (y)

68 78 71 69 66 75

27 52 39 22 21 66

74 65 72 73 67 77

48 33 45 51 29 65

(1) Find the regression equation using temperature as the independent variable and construct the anova table. (2) Test for a significant regression. Use α = .05. (3) Find a 95% confidence interval for the mean number of defective items produced when the temperature is 68◦ F. (4) Find a 99% prediction interval for a temperature of 75◦ F. Solution: Sxy 640.5 = = 3.1359 (S1) β.1 = Sxx 204.25 β.0 = y − β.1 x = 41.5 − (3.1359)(71.25) = −181.93 Regression line: y. = −181.93 + 3.1359x (S2) Source of variation Regression Error Total

Sum of squares

Degrees of freedom

Mean square

2008.5 644.5 2653.0

1 10 11

2008.5 64.4

c 2000 by Chapman & Hall/CRC 

Computed F 31.16

(S3) Hypothesis test of significant regression: H0 : β 1 = 0 Ha : β1 = 0 TS: F = MSR/MSE RR: F ≥ F.05,1,10 = 4.96 Conclusion: The value of the test statistic (F = 31.16) lies in the rejection region. There is evidence to suggest a significant regression. Note: this test is equivalent to the t test in section 11.1.3 with β10 = 0. (S4) A 95% confidence interval for µY |68 : y.0 ± t.025,10 · sY.0 = 31.31 ± (2.228)(2.9502) = (24.74, 37.88) (S5) A 99% prediction interval for y0 = −181.93 + (3.1359)(75) = 53.26 y.0 ± t.005,10 · sY.0 −Y0 = 53.26 ± (3.169)(8.6172) = (25.95, 80.56)

11.2

MULTIPLE LINEAR REGRESSION

Let there be n observations of the form (x1i , x2i , . . . , xki , yi ) such that yi is an observed value of the random variable Yi . Assume there exist constants β0 , β1 , . . . , βk such that Yi = β0 + β1 x1i + · · · + βk xki + 3i

(11.6)

where 31 , 32 , . . . , 3n are independent, normal random variables having mean 0 and variance σ 2 .

In terms of 3i ’s

Assumptions In terms of Yi ’s

3i ’s are normally distributed

Yi ’s are normally distributed

E [3i ] = 0

E [Yi ] = β0 + β1 x1i + · · · + βk xki

Var [3i ] = σ

2

Cov [3i , 3j ] = 0, i = j

Var [Yi ] = σ 2 Cov [Yi , Yj ] = 0, i = j

Notation: Let Y be the random vector of responses, y be the vector of observed responses, β be the vector of regression coefficients, be the vector of random errors, and let X be the design matrix: 

        Y1 y1 β0 31 1 x11 x21  Y2   y2   β1   32   1 x12 x22          Y= .  y= .  β= .  = .  X= . . ..  ..   ..   ..   ..   .. .. . Yn yn βk 3n 1 x1n x2n

c 2000 by Chapman & Hall/CRC 

 · · · xk1 · · · xk2   . . ..  . .  · · · xkn (11.7)

The model can now be written as Y = Xβ + where ∼ Nn (0, σ 2 In ) or equivalently Y ∼ Nn (Xβ, σ 2 In ). Principle of least squares: The sum of squared deviations about the true regression line is S(β) =

n 

[yi − (β0 + β1 x1i + · · · + βk xki )]2 = y − Xβ2 .

(11.8)

i=1

. = [β.0 , β.1 , . . . , β.k ]T that minimizes S(β) is the vector of least The vector β squares estimates. The estimated regression line or least squares line is y = β.0 + β.1 x1 + · · · + β.k xk .   . = XT y. The normal equations may be written as XT X β 11.2.1

Least squares estimates

. = (XT X)−1 XT y. If the matrix XT X is non–singular, then β The ith predicted (fitted) value: y.i = β.0 + β.1 x1i + · · · + β.k xki . . = Xβ. (for i = 1, 2, . . . , n), y .. The ith residual: ei = yi − y.i , i = 1, 2, . . . , n, e = y − y Properties: For i = 0, 1, 2, , . . . , k and j = 0, 1, 2, . . . , k: (1) E [β.i ] = βi . (2) Var [β.i ] = cii σ 2 , where cij is the value in the ith row and j th column of the matrix (XT X)−1 . (3) β.i is normally distributed. (4) Cov[β.i , β.j ] = cij σ 2 , i = j. 11.2.2 n  i=1



Sum of squares

(yi − y)2 = 



n  i=1



SST

(. yi − y)2 + 



SSR

n  i=1



(yi − y.i )2 



SSE

SST = total sum of squares = ||y − y1||2 = yT y − ny 2 . − y1||2 = β . T XT y − ny 2 SSR = sum of squares due to regression = ||Xβ . 2 = yT y − βX . Ty SSE = sum of squares due to error = ||y − Xβ|| where 1 = [1, 1, . . . , 1]T is a column vector of all 1’s.   SSE , E S 2 = σ2 n−k−1 (n − k − 1)S 2 (2) has a chi–square distribution with n − k − 1 degrees of σ2 freedom, and S 2 and β.i are independent. (1) σ .2 = s2 =

c 2000 by Chapman & Hall/CRC 

(3) The coefficient of multiple determination: SSE SSR =1− SST SST (4) Adjusted coefficient of multiple determination:     n−1 n−1 SSE Ra2 = 1 − = 1 − (1 − R2 ) n − k − 1 SST n−k−1 R2 =

11.2.3

Inferences concerning the regression coefficients

β.i − βi has a t distribution with n−k−1 degrees of freedom. √ S cii (2) A 100(1 − α)% confidence interval for βi has as endpoints √ β.i ± tα/2,n−k−1 · s cii (1) T =

(3) Hypothesis test for βi : Null hypothesis

Alternative hypotheses

Test statistic

Rejection regions

βi = βi0

βi > βi0 βi < βi0 βi = βi0

β.i − βi0 T = √ S cii

T ≥ tα,n−k−1 T ≤ −tα,n−k−1 |T | ≥ tα/2,n−k−1

11.2.4

(1) (2) (3)

The mean response

The mean response of Y given x = x0 = [1, x10 , x20 , . . . , xk0 ]T is µY |x10 ,x20 ,...,xk0 = β0 + β1 x10 + · · · + βk xk0 . The random variable . T x0 = β.0 + β.1 x10 + · · · + β.k xk0 is used to estimate µY |x ,x Y.0 = β 10

20 ,...,xk0

.

(1) E [Y.0 ] = β0 + β1 x10 + · · · + βk xk0 T −1 (2) Var [Y.0 ] = σ 2 xT x0 0 (X X)

(3) Y.0 has a normal distribution. (4) T =

Y.0 − µY |x10 ,x20 ,...,xk0  T −1 x S xT 0 0 (X X)

has a t distribution with n − k − 1 degrees of freedom. (5) A 100(1 − α)% confidence interval for µY |x10 ,x20 ,...,xk0 has as endpoints  T −1 x . ” y.0 ± tα/2,n−k−1 · s xT 0 0 (X X) (6) Hypothesis test:

c 2000 by Chapman & Hall/CRC 

Null Alternative Test hypothesis hypotheses statistic

Rejection regions

β0 + β1 x10 + · · · + βk xk0 = y0 = µ0 T ≥ tα,n−k−1 Y.0 − µ0  T= T ≤ −tα,n−k−1 T −1 x S xT 0 0 (X X) |T | ≥ tα/2,n−k−1

y0 > µ0 y0 < µ0 y0 = µ0 11.2.5

(1) (2) (3)

Prediction interval

A prediction interval for a value y0 of the random variable Y0 = β0 + β1 x10 + · · · + βk xk0 + 30 is obtained by considering the random variable Y.0 − Y0 . (1) E [Y.0 − Y0 ] = 0

  T −1 (2) Var [Y.0 − Y0 ] = σ 2 1 + xT x0 0 (X X) (3) Y.0 − Y0 has a normal distribution. 

(4) T =

Y.0 − Y0

T −1 x S 1 + xT 0 0 (X X) has a t distribution with n − k − 1 degrees of freedom.

(5) A 100(1 − α)% prediction interval for y0 has as endpoints  T −1 X y.0 ± tα/2,n−2 · s 1 + xT 0 0 (X X) 11.2.6

Analysis of variance table

Source of variation

Sum of Degrees of Mean squares freedom square

Regression SSR

k

Error

SSE

n−k−1

Total

SST

n−1

SSR k SSE MSE = n−k−1 MSR =

Hypothesis test of significant regression: H0 : β1 = β2 = · · · = βk = 0 Ha : βi =

0 for some i TS: F = MSR/MSE RR: F ≥ Fα,k,n−k−1

c 2000 by Chapman & Hall/CRC 

Computed F MSR/MSE

11.2.7

Sequential sum of squares

Define

 n

 yi

    i=1     n g0   g1   x y 1i i      g =  .  = XT y =  i=1    ..  ..    .   gk   n   xki yi i=1

SSR =

k 

β.j gj − ny 2

j=0

SS(β1 , β2 , . . . , βr ) = the sum of squares due to β1 , β2 , . . . , βr r  = β.j gj − ny 2 j=1

SS(β1 ) = the regression sum of squares due to x1 1  = β.j gj − ny 2 j=0

SS(β2 |β1 ) = the regression sum of squares due to x2 given x1 is in the model = SS(β1 , β2 ) − SS(β1 ) = β.2 g2 SS(β3 |β1 , β2 ) = the regression sum of squares due to x3 given x1 , x2 are in the model = SS(β1 , β2 , β3 ) − SS(β1 , β2 ) = β.3 g3 .. . SS(βr |β1 , . . . , βr−1 ) = the regression sum of squares due to xr given x1 , . . . , xr−1 are in the model = SS(β1 , . . . , βr ) − SS(β1 , . . . , βr−1 ) = β.r gr 11.2.8

Partial F test

Definitions: (1) Full model: yi = β0 + β1 x1i + · · · + βr xri + βr+1 x(r+1)i + · · · + βk xki + 3i (2) SSE(F) = sum of squares due to error in the full model. n  = (yi − y.i )2 where i=1 c 2000 by Chapman & Hall/CRC 

(11.9)

y.i = βˆ0 + βˆ1 x1i + · · · + βˆr xri + βˆr+1 x(r+1)i + · · · + βˆk xki (3) Reduced model: yi = β0 + β1 x1i + · · · + βr xri + 3i (4) SSE(R) = sum of squares due to error in the reduced model. n  = (yi − y.i )2 where y.i = βˆ0 + βˆ1 x1i + · · · + βˆr xri i=1

(5) SS(βr+1 , . . . , |β1 , . . . , βr ) = regression sum of squares due to xr+1 , . . . , xk given x1 , . . . , xr are in the model. It is given by: k  SS(β1 , . . . , βr , βr+1 , . . . , βk ) − SS(β1 , . . . , βr ) = β.j gj j=r+1

Hypothesis test: H0 : βr+1 = βr+2 = · · · = βk = 0 Ha : βm = 0 for some m = r + 1, r + 2, . . . , k TS: F = =

[SSE(R) − SSE(F)]/(k − r) SSE(F)/(n − k − 1) SS(βr+1 , . . . , βk |β1 , . . . , βr )/(k − r) SSE(R)/(n − k − 1)

RR: F ≥ Fα,k−r,n−k−1 11.2.9

Residual analysis

Let hii be the diagonal entries of the HAT matrix defined by H = X(XT X)−1 XT . ei ei Standardized residuals: √ = , i = 1, 2, . . . , n s MSE ei Studentized residuals: e∗i = √ , i = 1, 2, . . . , n s 1 − hii Deleted studentized residuals: + n−k−2 ∗ di = ei , i = 1, 2, . . . , n s2 (1 − hii ) − e2i Cook’s distance: Di =

  hii e2i , i = 1, 2, . . . , n (k + 1)s2 (1 − hii )2

Press residuals: δi = yi − y.i,−i =

ei , i = 1, 2, . . . , n 1 − hii

where y.i,−i is the ith predicted value by the model without using the ith observation in calculating the regression coefficients.

c 2000 by Chapman & Hall/CRC 

Prediction sum of squares = PRESS =

n 

δi2

i=1 n 

|δi |: is used for cross validation, it is less sensitive to large press residuals.

i=1

11.2.10

Example

Example 11.64 : A university foundation office recently investigated factors that might contribute to alumni donations. Fifteen years were randomly selected and the total donations (in millions of dollars), United States savings rate, the unemployment rate, and the number of games won by the school basketball team are given in the table below. Donations (y)

Savings rate (%) (x1 )

Unemployment rate (%) (x2 )

Games won

27.80 17.41 18.51 30.09 34.22 20.28 26.98 27.32 18.74 35.52 15.52 17.75 22.94 41.47 22.95

15.1 11.3 13.6 16.1 17.8 13.3 15.5 16.7 13.8 17.4 10.9 12.5 12.9 18.8 16.1

4.7 5.5 5.6 4.9 5.1 5.6 4.6 4.7 5.8 4.5 6.1 5.7 5.0 4.2 5.4

26 14 24 15 17 18 20 9 18 17 19 16 21 19 18

(1) Find the regression equation using donation as the dependent variable and construct the anova table. (2) Test for a significant regression. Use α = .05. (3) Is there any evidence to suggest the number of games won by the school basketball team affects alumni donations. Use α = .10. Solution: (S1) Regression line: y. = 23.89 − 5.54x1 + 1.95x2 + .057x3 (S2) Source of variation Regression Error Total

Sum of squares

Degrees of freedom

Mean square

755.64 71.28 826.92

3 11 14

251.88 6.48

(S3) Hypothesis test of significant regression: c 2000 by Chapman & Hall/CRC 

Computed F 38.87

H0 : β 1 = β 2 = β 3 = 0 Ha : βi = 0 for some i TS: F = MSR/MSE RR: F ≥ F.05,3,11 = 3.58 Conclusion: The value of the test statistic (F = 38.87) lies in the rejection region. There is evidence to suggest a significant regression. (S4) Hypothesis test for β3 : H0 : β 3 = 0 Ha : β3 = 0 β.3 − 0 TS: T = √ S cii RR: |T | ≥ t.05,11 = 1.796 t = .057/.1719 = .33 Conclusion: The value of the test statistic does not lie in the rejection region. There is no evidence to suggest the number of games won by the basketball team affects alumni donations.

11.3

ORTHOGONAL POLYNOMIALS

Polynomial regression models may contain several independent variables and each independent variable may appear in the model to various powers. Let (x1 , y1 ), (x2 , y2 ), . . . , (xn , yn ) be n pairs of observations such that yi is an observed value of the random variable Yi . Assume there exist constants β0 , β1 , . . . , βp such that Yi = β0 + β1 xi + β2 x2i + · · · + βp xpi + 3i

(11.10)

where 31 , 32 , . . . , 3n are independent, normal random variables having mean 0 and variance σ 2 . In order to determine the best polynomial model, the regression coefficients {βj } must be recalculated for various values of p. If the values {xi } are evenly spaced, then orthogonal polynomials are often used to determine the best model. This technique presents certain computational advantages and quickly isolates significant effects. The orthogonal polynomial model is Yi = α0 ξ0 (x) + α1 ξ1 (x) + · · · + αp ξp (x) + 3i

(11.11)

where ξi (x) are orthogonal polynomials in x of degree i. The orthogonal polynomials have the property n 

ξh (xi ) · ξj (xi ) = 0 , when h = j and xi = x0 + iδ

i=1

c 2000 by Chapman & Hall/CRC 

(11.12)

The least square estimator for αj is given by n 

α .j =

yi ξj (xi )

i=1 n 

(11.13)

[ξj (xi )]2

i=1

The estimator α .j (for j = 1, 2, . . . , n − 1) is a normal random = n variable with  mean αj if j ≤ p and mean 0 if j > p, and variance σ 2 [ξj (xi )]2 . An i=1

estimate of σ 2 is given by n  i=1

s2 = (αj − α .j ) Each ratio

-



p 

yi2 −

(. αj )2

j=0

n 

 2

[ξj (xi )]

i=1

n−p−1

.

(11.14)

2

[ξj (xi )]

i

(with α .j = 0 for j > p) has a t distribution.

s

The contribution of each term may be tested; if it is not significant then the term may be discarded without recalculating the previously obtained coefficients (i.e., the tests for significance effects are isolated). In order to tabulate the values of the orthogonal polynomials for repeated use, the values of xi are assumed to be one unit apart, and ξj (x) is defined to be a multiple, λj , of ξj (x) so that ξj (x) has a leading coefficient of unity. This adjustment makes all the tabulated values ξj (xi ) integers. Thus, in particular: ξ1 (x) = λ1 ξ1 (x) = λ1 (x − x)   n2 − 1 ξ2 (x) = λ2 ξ2 (x) = λ2 (x − x)2 − 12    2 3n − 7 ξ3 (x) = λ3 ξ3 (x) = λ3 (x − x)3 − (x − x) (11.15) 20 ' ξ4 (x) = λ4 ξ4 (x) = λ4 (x − x)4  2  , 3n − 13 3(n2 − 1)(n2 − 9) −(x − x)2 + 14 560 and, in general, ξr (x) = λr ξr with the recursion relation ξ0 = 1, ξr+1 = ξ1 ξr −

ξ1 = (x − x) r2 (n2 − r2 ) ξr−1 4(4r2 − 1)

(11.16)

The tables below provide values of {ξj (xi )} for various values of n and j. When n > 10 only half of the table is shown (the other half can be found c 2000 by Chapman & Hall/CRC 

by symmetry). The two values at the bottom of each column are the values n  2 Dj = [ξj (xi )] and λj . i=1

Example 11.65 : Consider n = 11 observations of data from a quadratic with noise: specifically the values are {1, 4, 9, . . . , 121} + 0.1{n1 , n2 , . . . , n11 } where each {ni } is a standard normal random variable. In order to fit a polynomial regression we must assume some maximum polynomial power of interest; we presume that terms up to degree three might be of interest. The {ξj (xi )} table for n = 11 (see page 279) gives the values for x ≤ 0; the values for x > 0 are given by symmetry. From the data and the {ξj (xi )} table we compute

Di =

xi 1 2 3 4 5 6 7 8 9 10 11

yi 1.53 4.27 8.94 15.92 24.99 35.94 48.77 64.13 80.80 99.89 120.99

ξ1 −5 −4 −3 −2 −1 0 1 2 3 4 5

ξ2 15 6 −1 −6 −9 −10 −9 −6 −1 6 15

ξ3 −30 6 22 23 14 0 −14 −23 −22 −6 30

ξi2

...

110

858

4290

λi

...

1

1

5/6

yi ξi

...

1315

869.5

−12.26





The values for α .i are found by: x=6

y = 46.01

α .2 =

869.5 = 1.013 858

1315 = 11.96 110 −12.26 α .3 = = −.00029 4290 α .1 =

(11.17)

Using the value of n and {λi } in equation (11.15), the orthogonal polynomials are ξ2 (x) = x2 − 12x + 26 ξ1 (x) = x − 8 1 ξ3 (x) = (5x3 − 90x2 + 451x − 546) 6

(11.18)

The regression equation is y. = y + α .1 ξ1 (x)+ α .2 ξ2 (x)+ α .3 ξ3 (x) = .85−.42x+1.06x2 − 3 .002x . The predicted values are {1.49, 4.22, 9.04, . . . , 120.92}.

c 2000 by Chapman & Hall/CRC 

The significance of the coefficients may be determined by computing: Mean square

Error mean square

Computed F statistic

1

15733

881.4

17.9

881.2

1

881.2

0.15

5958

(3) cubic regression  2 *  yi ξ3 ξ32

0.035

1

0.035

0.11

0.31

residual sum of squares: (1) − (2) − (3) − (4)

0.018

7

0.0026

0.095

0.027

Sum of squares

Degrees freedom

16615

10

(2) linear regression  2 *  yi ξ1 ξ12

15733

(3) quadratic regression  2 *  yi ξ2 ξ22

Quantity   n yi2 − ( yi )2 (1) n

where the computed F statistic is given by F = (mean square)/(error mean square). Conclusion: The cubic term should probably not be included in the model. The regresα1 ξ1 (x)+. α2 ξ2 (x)+0·ξ3 (x) = .59−.20x+1.0134x2 . sion equation then becomes y. = y+. The predicted values are then {1.40, 4.24, 9.11, . . . , 121.00}.

11.3.1

Tables for orthogonal polynomials

n = 3 points ξ1 ξ2 −1 0 1 D2 λ 1

1 −2 1 6 3

n = 4 points ξ1 ξ2 ξ3 −3 1 −1 −1 −1 3 1 −1 −3 3 1 1 D 20 4 20 λ 2 1 10/3

n = 7 points ξ1 ξ2 ξ3 ξ4 ξ5 −3 5 −1 −2 0 1 −1 −3 1 0 −4 0 1 −3 −1 2 0 −1 3 5 1 D 28 84 6 λ 1 1 1/6

3 −7 1 6 1 −7 3 154 7/12

−1 4 −5 0 5 −4 1 84 7/20

n = 5 points ξ1 ξ2 ξ3 ξ4 −2 2 −1 −1 −1 2 0 −2 0 1 −1 −2 2 2 1 D 10 14 10 λ 1 1 5/6

n = 8 points ξ1 ξ2 ξ3 ξ4

1 −4 6 −4 1 70 35/12

ξ5

−7 7 −7 7 −7 −5 1 5 −13 23 −3 −3 7 −3 −17 −1 −5 3 9 −15 1 −5 −3 9 15 3 −3 −7 −3 17 5 1 −5 −13 −23 7 7 7 7 7 D 168 168 264 616 2184 λ 2 1 2/3 7/12 7/10

c 2000 by Chapman & Hall/CRC 

n = 6 points ξ1 ξ2 ξ3 ξ4 ξ5 −5 5 −5 −3 −1 7 −1 −4 4 1 −4 −4 3 −1 −7 5 5 5 D 70 84 180 λ 2 3/2 5/3 ξ1

1 −1 −3 5 2 −10 2 10 −3 −5 1 1 28 252 7/12 21/10

n = 9 points ξ2 ξ3 ξ4 ξ5

−4 28 −14 14 −4 −3 7 7 −21 11 −2 −8 13 −11 −4 −1 −17 9 9 −9 0 −20 0 18 0 1 −17 −9 9 9 2 −8 −13 −11 4 3 7 −7 −21 −11 4 28 14 14 4 D 60 2772 990 2002 468 λ 1 3 5/6 7/12 3/20

n = 10 points ξ1 ξ2 ξ3 ξ4 ξ5

n = 11 points ξ1 ξ2 ξ3 ξ4 ξ5

−9 6 −42 18 −6 −7 2 14 −22 14 −5 −1 35 −17 −1 −3 −3 31 3 −11 −1 −4 12 18 −6 1 −4 −12 18 6 3 −3 −31 3 11 5 −1 −35 −17 1 7 2 −14 −22 −14 9 6 42 18 6 D 330 132 8580 2860 780 λ 2 1/2 5/3 5/12 1/10 ξ1

n = 13 points ξ2 ξ3 ξ4

−6 22 −11 99 −5 11 0 −66 −4 2 6 −96 −3 −5 8 −54 −2 −10 7 11 −1 −13 4 64 0 −14 0 84 D 182 2002 572 68068 λ 1 1 1/6 7/12 ξ1

−5 15 −30 −4 6 6 −3 −1 22 −2 −6 23 −1 −9 14 0 −10 0 D 110 858 4290 λ 1 1 5/6

n = 15 points ξ2 ξ3 ξ4

−8 −7 −6 −5 −4 −3 −2 −1 0 D 408 λ 1

n = 17 points ξ2 ξ3 ξ4

ξ5

ξ5

40 −28 52 −104 25 −7 −13 91 12 7 −39 104 1 15 −39 39 −8 18 −24 −36 −15 17 −3 −83 −20 13 17 −88 −23 7 31 −55 −24 0 36 0 7752 3876 16796 100776 1/20 1 1/6 1/12

n = 12 points ξ2 ξ3 ξ4

ξ1

ξ1

ξ5

n = 16 points ξ2 ξ3 ξ4

ξ5

−17 68 −68 68 −884 −15 44 −20 −12 676 −13 23 13 −47 871 −11 5 33 −51 429 −9 −10 42 −36 −156 −7 −22 42 −12 −588 −5 −31 35 13 −733 −3 −37 23 33 −583 −1 −40 8 44 −220 D 1938 23256 23256 28424 6953544 1/3 1/12 3/10 3/2 λ 2

c 2000 by Chapman & Hall/CRC 

ξ5

−15 35 −455 273 −143 −13 21 −91 −91 143 −11 9 143 −221 143 −9 −1 267 −201 33 −7 −9 301 −101 −77 −5 −15 265 23 −131 −3 −19 179 129 −115 −1 −21 63 189 −45 D 1360 5712 1007760 470288 201552 7/12 1/10 10/3 λ 2 1 n = 18 points ξ2 ξ3 ξ4

ξ5

−11 55 −33 33 −33 −9 25 3 −27 57 −7 1 21 −33 21 −5 −17 25 −13 −29 −3 −29 19 12 −44 −1 −35 7 28 −20 D 572 12012 5148 8008 15912 λ 2 3 2/3 7/24 3/20

−13 13 −143 143 −143 −11 7 −11 −77 187 −9 2 66 −132 132 −7 −2 98 −92 −28 −5 −5 95 −13 −139 −3 −7 67 63 −145 −1 −8 24 108 −60 D 910 728 97240 136136 235144 5/3 7/12 7/30 λ 2 1/2

−7 91 −91 1001 −1001 −6 52 −13 −429 1144 −5 19 35 −869 979 −4 −8 58 −704 44 −3 −29 61 −249 −751 −2 −44 49 251 −1000 −1 −53 27 621 −675 0 −56 0 756 0 D 280 37128 39780 6466460 10581480 35/12 21/20 5/6 λ 1 3 ξ1

−3 6 1 −4 −4 0 156 1/40

n = 14 points ξ1 ξ2 ξ3 ξ4

ξ5 −22 33 18 −11 −26 −20 0 6188 7/120

6 −6 −6 −1 4 6 286 1/12

ξ1

ξ1

n = 19 points ξ2 ξ3 ξ4

ξ5

ξ1

−9 51 −204 612 −102 −8 34 −68 −68 68 −7 19 28 −388 98 −6 6 89 −453 58 −5 −5 120 −354 −3 −4 −14 126 −168 −54 −3 −21 112 42 −79 −2 −26 83 227 −74 −1 −29 44 352 −44 0 −30 0 396 0 D 570 13566 213180 2288132 89148 5 7 1 λ 1 1 /6 /12 /40 ξ1

ξ2

n = 21 points ξ3 ξ4

n = 23 points ξ2 ξ3 ξ4

ξ5

ξ5

−11 77 −77 1463 −209 −10 56 −35 133 76 −9 37 −3 −627 171 −8 20 20 −950 152 −7 5 35 −955 77 −6 −8 43 −747 −12 −5 −19 45 −417 −87 −4 −28 42 −42 −132 −3 −35 35 315 −141 −2 −40 25 605 −116 −1 −43 13 793 −65 0 −44 0 858 0 D 1012 35420 32890 13123110 340860 1 7 1 λ 1 1 /6 /12 /60

c 2000 by Chapman & Hall/CRC 

n = 20 points ξ3 ξ4

ξ5

−19 57 −969 1938 −1938 −17 39 −357 −102 1122 −15 23 85 −1122 1802 −13 9 377 −1402 1222 −11 −3 539 −1187 187 −9 −13 591 −687 −771 −7 −21 553 −77 −1351 −5 −27 445 503 −1441 −3 −31 287 948 −1076 −1 −33 99 1188 −396 D 2660 17556 4903140 22881320 31201800 10 35 7/20 λ 2 1 /3 /24 ξ1

−10 190 −285 969 −3876 −9 133 −114 0 1938 −8 82 12 −510 3468 −7 37 98 −680 2618 −6 −2 149 −615 788 −5 −35 170 −406 −1063 −4 −62 166 −130 −2354 −3 −83 142 150 −2819 −2 −98 103 385 −2444 −1 −107 54 540 −1404 0 −110 0 594 0 D 770 201894 432630 5720330 121687020 7 21 5 λ 1 3 /6 /12 /40 ξ1

ξ2

−21 −19 −17 −15 −13 −11 −9 −7 −5 −3 −1 D 3542 λ 2 ξ1

n = 22 points ξ2 ξ3 ξ4

ξ5

35 −133 1197 −2261 25 −57 57 969 16 0 −570 1938 8 40 −810 1598 1 65 −775 663 −5 77 −563 −363 −10 78 −258 −1158 −14 70 70 −1554 −17 55 365 −1509 −19 35 585 −1079 −20 12 702 −390 7084 96140 8748740 40562340 1/3 7/12 7/30 1/2 ξ2

n = 24 points ξ3 ξ4

ξ5

−23 253 −1771 253 −4807 −21 187 −847 33 1463 −19 127 −133 −97 3743 −17 73 391 −157 3553 −15 25 745 −165 2071 −13 −17 949 −137 169 −11 −53 1023 −87 −1551 −9 −83 987 −27 −2721 −7 −107 861 33 −3171 −5 −125 665 85 −2893 −3 −137 419 123 −2005 −1 −143 143 143 −715 D 4600 394680 17760600 394680 177928920 10 1 3/10 λ 2 3 /3 /12

CHAPTER 12

Analysis of Variance Contents 12.1

One-way anova 12.1.1 Sum of squares 12.1.2 Properties 12.1.3 Analysis of variance table 12.1.4 Multiple comparison procedures 12.1.5 Contrasts 12.1.6 Example 12.2 Two-way anova 12.2.1 One observation per cell 12.2.2 Analysis of variance table 12.2.3 Nested classifications with equal samples 12.2.4 Nested classifications with unequal samples 12.2.5 Two-factor experiments 12.2.6 Example 12.3 Three-factor experiments 12.3.1 Models and assumptions 12.3.2 Sum of squares 12.3.3 Mean squares and properties 12.3.4 Analysis of variance table 12.4 Manova 12.5 Factor analysis 12.6 Latin square design 12.6.1 Models and assumptions 12.6.2 Sum of squares 12.6.3 Mean squares and properties 12.6.4 Analysis of variance table

c 2000 by Chapman & Hall/CRC 

12.1

ONE-WAY ANOVA

Let there be k treatments, or populations, and independent random samples of size ni (for i = 1, 2, . . . , k) from each population, and let N = n1 + n2 + · · · + nk . Let Yij be the j th random observation in the ith treatment group. Assume each population is normally distributed with mean µi and common variance σ 2 . In a fixed effects, or model I, experiment the treatment levels are predetermined. In a random effects, or model II, experiment, the treatment levels are selected at random. Notation: (1) Dot notation is used to indicate a sum over all values of the selected subscript. (2) The random error term is 3ij and an observed value is eij . Fixed effects experiment: Model:

Yij = µi + 3ij = µ + αi + 3ij i = 1, 2, . . . , k, j = 1, 2, . . . , ni

Assumptions:

The 3ij ’s are independent, normally distributed with k  ind mean 0 and variance σ 2 (3ij ∼ N(0, σ 2 )), αi = 0 i=1

Random effects experiment: Model:

Yij = µi + 3ij = µ + Ai + 3ij i = 1, 2, . . . , k, j = 1, 2, . . . , ni ind

ind

Assumptions: 3ij ∼ N(0, σ 2 ), Ai ∼ N(0, σα2 ) The 3ij ’s are independent of the Ai ’s. 12.1.1

Sum of squares

ni k   i=1 j=1



(yij − y¯.. )2 = 



k  i=1



SST

ni (y i. − y .. )2 +  SSA



ni k   i=1 j=1



(yij − y i. )2 



SSE

Notation: yij = observed value of Yij ni 1  y i. = yij = mean of the observations in the ith sample ni j=1 k ni 1  y .. = yij = mean of all observations N i=1 j=1

c 2000 by Chapman & Hall/CRC 

(12.1)

Ti. =

ni 

= sum of the observations in the ith sample

yij

j=1

T.. =

ni k  

yij = sum of all observations

i=1 j=1

SST = total sum of squares =

ni k  

(yij − y .. )2 =

i=1 j=1

ni k  

2 yij −

i=1 j=1

T..2 N

SSA = sum of squares due to treatment =

k 

ni (y i. − y .. )2 =

i=1

k  T2 i.

ni

i=1



T..2 N

SSE = sum of squares due to error =

ni k  

(yij − y i. )2 = SST − SSA

i=1 j=1

12.1.2

Properties Expected value

Mean square

Fixed model k 

Random model 

ni αi2

SSA σ 2 + i=1 k−1 k−1 SSE MSE = S 2 = σ2 N −k 2 MSA = SA =

σ2 +

1  n− k−1

k  i=1

 n2i 

n

σα2

σ2

2 F = SA /S 2 has an F distribution with k − 1 and N − k degrees of freedom.

12.1.3

Analysis of variance table

Source of variation

Sum of Degrees of squares freedom

Mean square

Computed F MSA/MSE

Treatments SSA (between groups) Error SSE (within groups)

k−1

MSA

N −k

MSE

Total

N −1

SST

Hypothesis test of significant regression:

c 2000 by Chapman & Hall/CRC 

H0 : µ1 = µ2 = · · · = µk (Fixed effects model: α1 = α2 = · · · = αk = 0) (Random effects model: σα2 = 0) Ha : at least two of the means are unequal (Fixed effects model: αi = 0 for some i) (Random effects model: σα2 = 0) 2 TS: F = SA /S 2

RR: F ≥ Fα,k−1,N −k 12.1.4 12.1.4.1

Multiple comparison procedures Tukey’s procedure

Equal sample sizes: Let n = n1 = n2 = · · · = nk and let Qα,ν1 ,ν2 be a critical value of the Studentized Range distribution (see page 76). The set of intervals with endpoints (y i. − y j. ) ± Qα,k,k(n−1) ·



s2 /n for all i and j, i = j

(12.2)

is a collection of simultaneous 100(1 − α)% confidence intervals for the differences between the true treatment means, µi − µj . Each confidence interval that does not include zero suggests µi = µj at the α significance level. Unequal sample sizes: The set of confidence intervals with endpoints (N = n1 + n2 + · · · + nk ) + 1 1 1 (y i. − y j. ) ± √ Qα,k,N −k · s + for all i and j, i = j (12.3) n n 2 i j is a collection of simultaneous 100(1 − α)% confidence intervals for the differences between the true treatment means, µi − µj . 12.1.4.2

Duncan’s multiple range test

Let n = n1 = n2 = · · · = nk and let rα,ν1 ,ν2 be a critical value for Duncan’s multiple range test (see page 285). Duncan’s procedure for determining significant differences between each treatment group at the joint significance level α is: s2 (1) Define Rp = rα,p,k(n−1) · for p = 2, 3, . . . , k. n (2) List the sample means in increasing order. (3) Compare the range of every subset of p sample means (for p = 2, 3, . . . , k) in the ordered list with Rp . (4) If the range of a p–subset is less than Rp then that subset of ordered means in not significantly different.

c 2000 by Chapman & Hall/CRC 

12.1.4.3

Duncan’s multiple range test

These tables contain critical values for the least significant studentized ranges, rα,p,ν , for Duncan’s multiple range test where α is the significance level, p is the number of successive values from an ordered list of k means of equal sample sizes (p = 2, 3, . . . , k), and n is the degrees of freedom for the independent estimate s2 . These tables are from L. Hunter, “Critical Values for Duncan’s New Multiple Range Test”, Biometrics, 1960, Volume 16, pages 671–685. Reprinted with permission from the Journal of American Statistical Association. Copyright 1960 by the American Statistical Association. All rights reserved.

c 2000 by Chapman & Hall/CRC 

p=2 17.97 6.085 4.501 3.927 3.635 3.461 3.344 3.261 3.199 3.151 3.113 3.082 3.055 3.033 3.014 2.998 2.984 2.971 2.960 2.950 2.919 2.888 2.858 2.829 2.800 2.772

n 1 2 3 4

5 6 7 8 9

10 11 12 13 14

15 16 17 18 19

20 24 30 40 60 120 ∞

20 24 30 40 60 120 ∞

2.950 2.919 2.888 2.858 2.829 2.800 2.772

3.097 3.066 3.035 3.006 2.976 2.947 2.918 3.097 3.066 3.035 3.006 2.976 2.947 2.918

3.160 3.144 3.130 3.118 3.107

3.293 3.256 3.225 3.200 3.178

3.749 3.587 3.477 3.399 3.339

3 17.97 6.085 4.516 4.013

3.190 3.160 3.131 3.102 3.073 3.045 3.017

3.250 3.235 3.222 3.210 3.199

3.376 3.342 3.313 3.289 3.268

3.797 3.649 3.548 3.475 3.420

4 17.97 6.085 4.516 4.033

3.255 3.226 3.199 3.171 3.143 3.116 3.089

3.312 3.298 3.285 3.274 3.264

3.430 3.397 3.370 3.348 3.329

3.814 3.680 3.588 3.521 3.470

5 17.97 6.085 4.516 4.033

3.303 3.276 3.250 3.224 3.198 3.172 3.146

3.356 3.343 3.331 3.321 3.311

3.465 3.435 3.410 3.389 3.372

3.814 3.694 3.611 3.549 3.502

6 17.97 6.085 4.516 4.033

3.339 3.315 3.290 3.266 3.241 3.217 3.193

3.389 3.376 3.366 3.356 3.347

3.489 3.462 3.439 3.419 3.403

3.814 3.697 3.622 3.566 3.523

7 17.97 6.085 4.516 4.033

3.368 3.345 3.322 3.300 3.277 3.254 3.232

3.413 3.402 3.392 3.383 3.375

3.505 3.480 3.459 3.442 3.426

3.814 3.697 3.626 3.575 3.536

8 17.97 6.085 4.516 4.033

3.391 3.370 3.349 3.328 3.307 3.287 3.265

3.432 3.422 3.412 3.405 3.397

3.516 3.493 3.474 3.458 3.444

3.814 3.697 3.626 3.579 3.544

9 17.97 6.085 4.516 4.033

3.409 3.390 3.371 3.352 3.333 3.314 3.294

3.446 3.437 3.429 3.421 3.415

3.522 3.501 3.484 3.470 3.457

3.814 3.697 3.626 3.579 3.547

10 17.97 6.085 4.516 4.033

3.424 3.406 3.389 3.373 3.355 3.337 3.320

3.457 3.449 3.441 3.435 3.429

3.525 3.506 3.491 3.478 3.467

3.814 3.697 3.626 3.579 3.547

11 17.97 6.085 4.516 4.033

3.436 3.420 3.405 3.390 3.374 3.359 3.343

3.465 3.458 3.451 3.445 3.440

3.526 3.509 3.496 3.484 3.474

3.814 3.697 3.626 3.579 3.547

12 17.97 6.085 4.516 4.033

3.445 3.432 3.418 3.405 3.391 3.377 3.363

3.471 3.465 3.459 3.454 3.449

3.526 3.510 3.498 3.488 3.479

3.814 3.697 3.626 3.579 3.547

13 17.97 6.085 4.516 4.033

3.453 3.441 3.430 3.418 3.406 3.394 3.382

3.476 3.470 3.465 3.460 3.456

3.526 3.510 3.499 3.490 3.482

3.814 3.697 3.626 3.579 3.547

14 17.97 6.085 4.516 4.033

3.459 3.449 3.439 3.429 3.419 3.409 3.399

3.478 3.473 3.469 3.465 3.462

3.526 3.510 3.499 3.490 3.484

3.814 3.697 3.626 3.579 3.547

15 17.97 6.085 4.516 4.033

3.464 3.456 3.447 3.439 3.431 3.423 3.414

3.480 3.477 3.473 3.470 3.467

3.526 3.510 3.499 3.490 3.484

3.814 3.697 3.626 3.579 3.547

16 17.97 6.085 4.516 4.033

3.467 3.461 3.454 3.448 3.442 3.435 3.428

3.481 3.478 3.475 3.472 3.470

3.526 3.510 3.499 3.490 3.485

3.814 3.697 3.626 3.579 3.547

17 17.97 6.085 4.516 4.033

3.470 3.465 3.460 3.456 3.451 3.446 3.442

3.481 3.478 3.476 3.474 3.472

3.526 3.510 3.499 3.490 3.485

3.814 3.697 3.626 3.579 3.547

18 17.97 6.085 4.516 4.033

3.472 3.469 3.466 3.463 3.460 3.457 3.454

3.481 3.478 3.476 3.474 3.473

3.526 3.510 3.499 3.490 3.485

3.814 3.697 3.626 3.579 3.547

19 17.97 6.085 4.516 4.033

3.473 3.471 3.470 3.469 3.467 3.466 3.466

3.481 3.478 3.476 3.474 3.474

3.526 3.510 3.499 3.490 3.485

3.814 3.697 3.626 3.579 3.547

20 17.97 6.085 4.516 4.033

Critical values for Duncan’s test, rα,p,n , for α = .05

c 2000 by Chapman & Hall/CRC 

3.190 3.160 3.131 3.102 3.073 3.045 3.017

3.255 3.226 3.199 3.171 3.143 3.116 3.089

3.303 3.276 3.250 3.224 3.198 3.172 3.146

3.339 3.315 3.290 3.266 3.241 3.217 3.193

3.368 3.345 3.322 3.300 3.277 3.254 3.232

3.391 3.370 3.349 3.328 3.307 3.287 3.265

3.409 3.390 3.371 3.352 3.333 3.314 3.294

3.424 3.406 3.389 3.373 3.355 3.337 3.320

3 3 3 3 3 3 3

5.702 5.243 4.949 4.746 4.596 4.482 4.392 4.320 4.260 4.210 4.168 4.131 4.099 4.071 4.046 4.024 3.956 3.889 3.825 3.762 3.702 3.643

5 6 7 8 9

10 11 12 13 14

15 16 17 18 19

20 24 30 40 60 120 ∞

20 24 30 40 60 120 ∞

4.024 3.956 3.889 3.825 3.762 3.702 3.643

4.197 4.126 4.056 3.988 3.922 3.858 3.796 4.197 4.126 4.056 3.988 3.922 3.858 3.796

4.347 4.309 4.275 4.246 4.220

4.671 4.579 4.504 4.442 4.391

5.893 5.439 5.145 4.939 4.787

4.312 4.239 4.168 4.098 4.031 3.965 3.900

4.463 4.425 4.391 4.362 4.335

4.790 4.697 4.622 4.560 4.508

5.989 5.549 5.260 5.057 4.906

4.395 4.322 4.250 4.180 4.111 4.044 3.978

4.547 4.509 4.475 4.445 4.419

4.871 4.780 4.706 4.644 4.591

6.040 5.614 5.334 5.135 4.986

4.459 4.386 4.314 4.244 4.174 4.107 4.040

4.610 4.572 4.539 4.509 4.483

4.931 4.841 4.767 4.706 4.654

6.065 5.655 5.383 5.189 5.043

4.510 4.437 4.366 4.296 4.226 4.158 4.091

4.660 4.622 4.589 4.560 4.534

4.975 4.887 4.815 4.755 4.704

6.074 5.680 5.416 5.227 5.086

4.552 4.480 4.409 4.339 4.270 4.202 4.135

4.700 4.663 4.630 4.601 4.575

5.010 4.924 4.852 4.793 4.743

6.074 5.694 5.439 5.256 5.118

4.587 4.516 4.445 4.376 4.307 4.239 4.172

4.733 4.696 4.664 4.635 4.610

5.037 4.952 4.883 4.824 4.775

6.074 5.701 5.454 5.276 5.142

4.617 4.546 4.477 4.408 4.340 4.272 4.205

4.760 4.724 4.693 4.664 4.639

5.058 4.975 4.907 4.850 4.802

6.074 5.703 5.464 5.291 5.160

4.642 4.573 4.504 4.436 4.368 4.301 4.235

4.783 4.748 4.717 4.689 4.665

5.074 4.994 4.927 4.872 4.824

6.074 5.703 5.470 5.302 5.174

4.664 4.596 4.528 4.461 4.394 4.327 4.261

4.803 4.768 4.738 4.711 4.686

5.088 5.009 4.944 4.889 4.843

6.074 5.703 5.472 5.309 5.185

4.684 4.616 4.550 4.483 4.417 4.351 4.285

4.820 4.786 4.756 4.729 4.705

5.098 5.021 4.958 4.904 4.859

6.074 5.703 5.472 5.314 5.193

4.701 4.634 4.569 4.503 4.438 4.372 4.307

4.834 4.800 4.771 4.745 4.722

5.106 5.031 4.969 4.917 4.872

6.074 5.703 5.472 5.316 5.199

4.716 4.651 4.586 4.521 4.456 4.392 4.327

4.846 4.813 4.785 4.759 4.736

5.112 5.039 4.978 4.928 4.884

6.074 5.703 5.472 5.317 5.203

4.729 4.665 4.601 4.537 4.474 4.410 4.345

4.857 4.825 4.797 4.772 4.749

5.117 5.045 4.986 4.937 4.894

6.074 5.703 5.472 5.317 5.205

4.741 4.678 4.615 4.553 4.490 4.426 4.363

4.866 4.835 4.807 4.783 4.761

5.120 5.050 4.993 4.944 4.902

6.074 5.703 5.472 5.317 5.206

4.751 4.690 4.628 4.566 4.504 4.442 4.379

4.874 4.844 4.816 4.792 4.771

5.122 5.054 4.998 4.950 4.910

6.074 5.703 5.472 5.317 5.206

4.761 4.700 4.640 4.579 4.518 4.456 4.394

4.881 4.851 4.824 4.801 4.780

5.124 5.057 5.002 4.956 4.916

6.074 5.703 5.472 5.317 5.206

4.769 4.710 4.650 4.591 4.530 4.469 4.408

4.887 4.858 4.832 4.808 4.788

5.124 5.059 5.006 4.960 4.921

6.074 5.703 5.472 5.317 5.206

2 14.04 14.04 14.04 14.04 14.04 14.04 14.04 14.04 14.04 14.04 14.04 14.04 14.04 14.04 14.04 14.04 14.04 14.04 14.04 3 8.261 8.321 8.321 8.321 8.321 8.321 8.321 8.321 8.321 8.321 8.321 8.321 8.321 8.321 8.321 8.321 8.321 8.321 8.321 4 6.512 6.677 6.740 6.756 6.756 6.756 6.756 6.756 6.756 6.756 6.756 6.756 6.756 6.756 6.756 6.756 6.756 6.756 6.756

n p=2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 1 90.03 90.03 90.03 90.03 90.03 90.03 90.03 90.03 90.03 90.03 90.03 90.03 90.03 90.03 90.03 90.03 90.03 90.03 90.03

Critical values for Duncan’s test, rα,p,n , for α = .01

c 2000 by Chapman & Hall/CRC 

4.312 4.239 4.168 4.098 4.031 3.965 3.900

4.395 4.322 4.250 4.180 4.111 4.044 3.978

4.459 4.386 4.314 4.244 4.174 4.107 4.040

4.510 4.437 4.366 4.296 4.226 4.158 4.091

4.552 4.480 4.409 4.339 4.270 4.202 4.135

4.587 4.516 4.445 4.376 4.307 4.239 4.172

4.617 4.546 4.477 4.408 4.340 4.272 4.205

4.642 4.573 4.504 4.436 4.368 4.301 4.235

4 4 4 4 4 4 4

5.444 5.297 5.156 5.022 4.894 4.771 4.654

20 24 30 40 60 120 ∞

6.487 6.275 6.106 5.970 5.856

10 11 12 13 14 5.760 5.678 5.608 5.546 5.492

9.714 8.427 7.648 7.130 6.762

5 6 7 8 9

15 16 17 18 19

p=2 900.3 44.69 18.28 12.18

n 1 2 3 4

20 24 30 40 60 120 ∞

5.444 5.297 5.156 5.022 4.894 4.771 4.654 5.640 5.484 5.335 5.191 5.055 4.924 4.798

5.974 5.888 5.813 5.748 5.691

6.738 6.516 6.340 6.195 6.075

10.05 8.743 7.943 7.407 7.024

3 900.3 44.69 18.45 12.52

5.774 5.612 5.457 5.308 5.166 5.029 4.898

6.119 6.030 5.953 5.886 5.826

6.902 6.676 6.494 6.346 6.223

10.24 8.932 8.127 7.584 7.195

4 900.3 44.69 18.45 12.67

5.873 5.708 5.549 5.396 5.249 5.109 4.974

6.225 6.135 6.056 5.988 5.927

7.021 6.791 6.607 6.457 6.332

10.35 9.055 8.252 7.708 7.316

5 900.3 44.69 18.45 12.73

5.952 5.784 5.622 5.466 5.317 5.173 5.034

6.309 6.217 6.138 6.068 6.007

7.111 6.880 6.695 6.543 6.416

10.42 9.139 8.342 7.799 7.407

6 900.3 44.69 18.45 12.75

6.017 5.846 5.682 5.524 5.372 5.226 5.085

6.377 6.284 6.204 6.134 6.072

7.182 6.950 6.765 6.612 6.485

10.46 9.198 8.409 7.869 7.478

7 900.3 44.69 18.45 12.75

6.071 5.899 5.734 5.574 5.420 5.271 5.128

6.433 6.340 6.260 6.189 6.127

7.240 7.008 6.822 6.670 6.542

10.48 9.241 8.460 7.924 7.535

8 900.3 44.69 18.45 12.75

6.117 5.945 5.778 5.617 5.461 5.311 5.166

6.481 6.388 6.307 6.236 6.174

7.287 7.056 6.870 6.718 6.590

10.49 9.272 8.500 7.968 7.582

9 900.3 44.69 18.45 12.75

6.158 5.984 5.817 5.654 5.498 5.346 5.199

6.522 6.429 6.348 6.277 6.214

7.327 7.097 6.911 6.759 6.631

10.49 9.294 8.530 8.004 7.619

10 900.3 44.69 18.45 12.75

6.193 6.020 5.851 5.688 5.530 5.377 5.229

6.558 6.465 6.384 6.313 6.250

7.361 7.132 6.947 6.795 6.667

10.49 9.309 8.555 8.033 7.652

11 900.3 44.69 18.45 12.75

6.225 6.051 5.882 5.718 5.559 5.405 5.256

6.590 6.497 6.416 6.345 6.281

7.390 7.162 6.978 6.826 6.699

10.49 9.319 8.574 8.057 7.679

12 900.3 44.69 18.45 12.75

6.254 6.079 5.910 5.745 5.586 5.431 5.280

6.619 6.525 6.444 6.373 6.310

7.415 7.188 7.005 6.854 6.727

10.49 9.325 8.589 8.078 7.702

13 900.3 44.69 18.45 12.75

6.279 6.105 5.935 5.770 5.610 5.454 5.303

6.644 6.551 6.470 6.399 6.336

7.437 7.211 7.029 6.878 6.752

10.49 9.328 8.600 8.094 7.722

14 900.3 44.69 18.45 12.75

6.303 6.129 5.958 5.793 5.632 5.476 5.324

6.666 6.574 6.493 6.422 6.359

7.456 7.231 7.050 6.900 6.774

10.49 9.329 8.609 8.108 7.739

15 900.3 44.69 18.45 12.75

6.324 6.150 5.980 5.814 5.653 5.496 5.343

6.687 6.595 6.514 6.443 6.380

7.472 7.250 7.069 6.920 6.794

10.49 9.329 8.616 8.119 7.753

16 900.3 44.69 18.45 12.75

6.344 6.170 6.000 5.834 5.672 5.515 5.361

6.706 6.614 6.533 6.462 6.400

7.487 7.266 7.086 6.937 6.812

10.49 9.329 8.621 8.129 7.766

17 900.3 44.69 18.45 12.75

6.362 6.188 6.018 5.852 5.690 5.532 5.378

6.723 6.631 6.551 6.480 6.418

7.500 7.280 7.102 6.954 6.829

10.49 9.329 8.624 8.137 7.777

18 900.3 44.69 18.45 12.75

6.379 6.205 6.036 5.869 5.707 5.549 5.394

6.739 6.647 6.567 6.497 6.434

7.511 7.293 7.116 6.968 6.844

10.49 9.329 8.626 8.143 7.786

19 900.3 44.69 18.45 12.75

6.394 6.221 6.051 5.885 5.723 5.565 5.409

6.753 6.661 6.582 6.512 6.450

7.522 7.304 7.128 6.982 6.858

10.49 9.329 8.627 8.149 7.794

20 900.3 44.69 18.45 12.75

Critical values for Duncan’s test, rα,p,n , for α = .001

c 2000 by Chapman & Hall/CRC 

5.640 5.484 5.335 5.191 5.055 4.924 4.798

5.774 5.612 5.457 5.308 5.166 5.029 4.898

5.873 5.708 5.549 5.396 5.249 5.109 4.974

5.952 5.784 5.622 5.466 5.317 5.173 5.034

6.017 5.846 5.682 5.524 5.372 5.226 5.085

6.071 5.899 5.734 5.574 5.420 5.271 5.128

6.117 5.945 5.778 5.617 5.461 5.311 5.166

6.158 5.984 5.817 5.654 5.498 5.346 5.199

6.193 6.020 5.851 5.688 5.530 5.377 5.229

6 6 5 5 5 5 5

12.1.4.4

Dunnett’s procedure

Let n = n0 = n1 = n2 = · · · = nk where treatment 0 is the control group and let dα,ν1 ,ν2 be a critical value for Dunnett’s procedure (see page 289). Dunnett’s procedure for determining significant differences between each treatment and the control at the joint significance level α is given in the following table. For i = 1, 2, . . . , k: Null hypothesis

Alternative hypotheses

Test statistic

Rejection regions

µ0 = µi

µ0 > µi µ0 < µi µ0 = µi

y − y 0. Di = i. 2S 2 /n

Di ≥ dα,k,k(n−1) Di ≤ −dα,k,k(n−1) |Di | ≥ dα,k,k(n−1)

12.1.4.5

(1) (2) (3)

Tables for Dunnett’s procedure

This table contains critical values dα/2,k,ν and dα,k,n for simultaneous comparisons of each treatment group with a control group; α is the significance level, k is the number of treatment groups, and n is the degrees of freedom of the independent estimate s2 . These tables are from C. W. Dunnett, “A Multiple Comparison Procedure for Comparing Several Treatments with a Control”, JASA, Volume 50, 1955, pages 1096–1121. Reprinted with permission from the Journal of American Statistical Association. Copyright 1980 by the American Statistical Association. All rights reserved.

3.07 3.01 2.96 2.90 2.85 2.80 2.75 3.02 2.96 2.91 2.86 2.81 2.76 2.71 2.96 2.91 2.86 2.81 2.76 2.71 2.67 2.89 2.84 2.79 2.75 2.70 2.66 2.62 2.81 2.76 2.72 2.67 2.63 2.59 2.55 2.70 2.66 2.62 2.58 2.55 2.51 2.47 2.57 2.53 2.50 2.47 2.43 2.40 2.37 2.09 2.06 2.04 2.02 2.00 1.98 1.96 20 24 30 40 60 120 ∞

2.38 2.35 2.32 2.29 2.27 2.24 2.21

3.19 3.16 3.13 3.11 3.09 3.13 3.10 3.08 3.05 3.04 3.07 3.04 3.01 2.99 2.97 2.99 2.96 2.94 2.92 2.90 2.90 2.88 2.85 2.84 2.82 2.79 2.77 2.75 2.73 2.72 2.64 2.63 2.61 2.59 2.58 2.13 2.12 2.11 2.10 2.09 15 16 17 18 19

2.44 2.42 2.41 2.40 2.39

3.46 3.38 3.32 3.27 3.23 3.39 3.31 3.25 3.21 3.17 3.31 3.24 3.18 3.14 3.10 3.21 3.15 3.10 3.06 3.02 3.11 3.05 3.00 2.96 2.93 2.97 2.92 2.88 2.84 2.81 2.81 2.76 2.72 2.69 2.67 2.23 2.20 2.18 2.16 2.14 10 11 12 13 14

2.57 2.53 2.50 2.48 2.46

k=1 2.57 2.45 2.36 2.31 2.26 n 5 6 7 8 9

2 3.03 2.86 2.75 2.67 2.61

3 3.39 3.18 3.04 2.94 2.86

4 3.66 3.41 3.24 3.13 3.04

5 3.88 3.60 3.41 3.28 3.18

6 4.06 3.75 3.54 3.40 3.29

7 4.22 3.88 3.66 3.51 3.39

8 4.36 4.00 3.76 3.60 3.48

9 4.49 4.11 3.86 3.68 3.55

Values of dα/2,k,n for two–sided comparisons (α = .05)

c 2000 by Chapman & Hall/CRC 

30 40 60 120 ∞

2.04 2.02 2.00 1.98 1.96

2.32 2.29 2.27 2.24 2.21

2.50 2.47 2.43 2.40 2.37

2.62 2.58 2.55 2.51 2.47

2.72 2.67 2.63 2.59 2.55

2.79 2.75 2.70 2.66 2.62

2.86 2.81 2.76 2.71 2.67

2.91 2.86 2.81 2.76 2.71

2.96 2.90 2.85 2.80 2.75

c 2000 by Chapman & Hall/CRC 

1.81 1.80 1.78 1.77 1.76 1.75 1.75 1.74 1.73 1.73 1.72 1.71 1.70 1.68 1.67

10 11 12 13 14

15 16 17 18 19

20 24 30 40 60

2.03 2.01 1.99 1.97 1.95

2.07 2.06 2.05 2.04 2.03

2.15 2.13 2.11 2.09 2.08

2.34 2.27 2.22 2.18

2.19 2.17 2.15 2.13 2.10

2.24 2.23 2.22 2.21 2.20

2.34 2.31 2.29 2.27 2.25

2.56 2.48 2.42 2.37

2.30 2.28 2.25 2.23 2.21

2.36 2.34 2.33 2.32 2.31

2.47 2.44 2.41 2.39 2.37

2.71 2.62 2.55 2.50

2.39 2.36 2.33 2.31 2.28

2.44 2.43 2.42 2.41 2.40

2.56 2.53 2.50 2.48 2.46

2.83 2.73 2.66 2.60

2.46 2.43 2.40 2.37 2.35

2.51 2.50 2.49 2.48 2.47

2.64 2.60 2.58 2.55 2.53

2.92 2.82 2.74 2.68

2.51 2.48 2.45 2.42 2.39

2.57 2.56 2.54 2.53 2.52

2.70 2.67 2.64 2.61 2.59

3.00 2.89 2.81 2.75

2.56 2.53 2.50 2.47 2.44

2.62 2.61 2.59 2.58 2.57

2.76 2.72 2.69 2.66 2.64

3.07 2.95 2.87 2.81

2.60 2.57 2.54 2.51 2.48

2.67 2.65 2.64 2.62 2.61

2.81 2.77 2.74 2.71 2.69

3.12 3.01 2.92 2.86

120 1.66 1.93 2.08 2.18 2.26 2.32 2.37 2.41 2.45 ∞ 1.64 1.92 2.06 2.16 2.23 2.29 2.34 2.38 2.42

1.94 1.89 1.86 1.83

6 7 8 9

n k=1 2 3 4 5 6 7 8 9 5 2.02 2.44 2.68 2.85 2.98 3.08 3.16 3.24 3.30

3.17 3.11 3.05 3.01 2.98 2.95 2.92 2.90 2.88 2.86 2.85 2.80 2.75 2.70 2.66

10 11 12 13 14 15 16 17 18 19 20 24 30 40 60

3.13 3.07 3.01 2.95 2.90

3.25 3.22 3.19 3.17 3.15

3.53 3.45 3.39 3.33 3.29

2 4.63 4.22 3.95 3.77 3.63

3.31 3.24 3.17 3.10 3.04

3.45 3.41 3.38 3.35 3.33

3.78 3.68 3.61 3.54 3.49

3 5.09 4.60 4.28 4.06 3.90

3.43 3.36 3.28 3.21 3.14

3.59 3.55 3.51 3.48 3.46

3.95 3.85 3.76 3.69 3.64

4 5.44 4.88 4.52 4.27 4.09

3.53 3.45 3.37 3.29 3.22

3.70 3.65 3.62 3.58 3.55

4.10 3.98 3.89 3.81 3.75

5 5.73 5.11 4.71 4.44 4.24

3.61 3.52 3.44 3.36 3.28

3.79 3.74 3.70 3.67 3.64

4.21 4.09 3.99 3.91 3.84

6 5.97 5.30 4.87 4.58 4.37

3.67 3.58 3.50 3.41 3.33

3.86 3.82 3.77 3.74 3.70

4.31 4.18 4.08 3.99 3.92

7 6.18 5.47 5.01 4.70 4.48

3.73 3.64 3.55 3.46 3.38

3.93 3.88 3.83 3.80 3.76

4.40 4.26 4.15 4.06 3.99

8 6.36 5.61 5.13 4.81 4.57

3.78 3.69 3.59 3.50 3.42

3.99 3.93 3.89 3.85 3.81

4.47 4.33 4.22 4.13 4.05

9 6.53 5.74 5.24 4.90 4.65

120 2.62 2.84 2.98 3.08 3.15 3.21 3.25 3.30 3.33 ∞ 2.58 2.79 2.92 3.01 3.08 3.14 3.18 3.22 3.25

k=1 4.03 3.71 3.50 3.36 3.25

n 5 6 7 8 9

Values of dα/2,k,n for two–sided comparisons (α = .01)

Values of dα/2,k,n for one–sided comparisons (α = .05)

2

3

4

5

6

7

8

9

12.1.5

.= Let L k 

L= k 

i=1

c 2000 by Chapman & Hall/CRC 

ci µi where k 

i=1

ci = 0.

(2) A 100(1 − α)% confidence interval for L has as endpoints F G k G . l ± tα/2,N −k · sH c2i /ni . Contrasts

A contrast L is a linear combination of the means µi such that the coefficients ci sum to zero:

i=1

(12.4)

i=1

ci y i. , then

k k " #  " #  c2i . has a normal distribution, E L . = . = σ2 (1) L ci µi , Var L . n i=1 i=1 i

(12.5)

∞ 2.33 2.56 2.68 2.77 2.84 2.89 2.93 2.97 3.00

120 2.36 2.60 2.73 2.82 2.89 2.94 2.99 3.03 3.06

60 2.39 2.64 2.78 2.87 2.94 3.00 3.04 3.08 3.12

40 2.42 2.68 2.82 2.92 2.99 3.05 3.10 3.14 3.18

30 2.46 2.72 2.87 2.97 3.05 3.11 3.16 3.21 3.24

24 2.49 2.77 2.92 3.03 3.11 3.17 3.22 3.27 3.31

20 2.53 2.81 2.97 3.08 3.17 3.23 3.29 3.34 3.38

19 2.54 2.83 2.99 3.10 3.18 3.25 3.31 3.36 3.40

18 2.55 2.84 3.01 3.12 3.21 3.27 3.33 3.38 3.42

17 2.57 2.86 3.03 3.14 3.23 3.30 3.36 3.41 3.45

16 2.58 2.88 3.05 3.17 3.26 3.33 3.39 3.44 3.48

15 2.60 2.91 3.08 3.20 3.29 3.36 3.42 3.47 3.52

14 2.62 2.94 3.11 3.23 3.32 3.40 3.46 3.51 3.56

13 2.65 2.97 3.15 3.27 3.37 3.44 3.51 3.56 3.61

12 2.68 3.01 3.19 3.32 3.42 3.50 3.56 3.62 3.67

11 2.72 3.06 3.25 3.38 3.48 3.56 3.63 3.69 3.74

10 2.76 3.11 3.31 3.45 3.56 3.64 3.71 3.78 3.83

9 2.82 3.19 3.40 3.55 3.66 3.75 3.82 3.89 3.94

8 2.90 3.29 3.51 3.67 3.79 3.88 3.96 4.03 4.09

7 3.00 3.42 3.66 3.83 3.96 4.07 4.15 4.23 4.30

6 3.14 3.61 3.88 4.07 4.21 4.33 4.43 4.51 4.59

5 3.37 3.90 4.21 4.43 4.60 4.73 4.85 4.94 5.03

n k=1

Values of dα/2,k,n for one–sided comparisons (α = .01)

(3) Single degree of freedom test: ith sample H0 :

k 

ci µi = c

i=1

Ha :

k 

ci µi > c,

i=1

k 

ci µi < c,

i=1

k 

ci µi = c

i=1

L−c (L − c)2 TS: T = + or F = T 2 = k  k  s2 c2i /ni s c2i /ni i=1

i=1

RR: T ≥ tα,N −k , T ≤ −tα,N −k , |T | ≥ tα/2,N −k or F ≥ Fα,1,N −k (4) The set of confidence intervals with endpoints F G k  G . l ± (k − 1)Fα,k−1,N −k · sH c2i /ni

(12.6)

i=1

is the collection of simultaneous 100(1 − α)% confidence intervals for all possible contrasts. (5) Let n = ni , i = 1, 2, . . . , k, then the contrast sum of squares, SSL, is given by  k 2  ci Ti. SSL = i=1 k . (12.7)  2 n ci i=1

(6) Two contrasts L1 =

k  i=1

k  bi ci i=1

12.1.6

ni

bi µi and L2 =

k 

ci µi are orthogonal if

i=1

= 0.

Example

Example 12.66 : A telephone company recently surveyed the length of long distance calls originating in four different parts of the country. The length of each randomly

c 2000 by Chapman & Hall/CRC 

selected call (in minutes) is given in the table below. North:

11.0 11.1

9.5 10.7

10.3 8.4

8.7 10.8

10.6

7.9

South:

13.6 13.0 13.6

12.2 15.6 17.2

12.5 12.1 13.1

17.5 10.2 13.3

9.7 14.6 11.9

16.4 16.2 12.0

Midwest:

15.0 10.6 13.4

14.2 13.8 10.9

11.9 16.8 11.1

14.5 11.4 11.0

12.7 15.7 10.3

17.1 11.0 10.4

West:

12.0 12.9

12.7 15.2

13.0 14.9

13.3 11.4

11.6 11.3

11.4

Is there any evidence to suggest the mean lengths of long distance calls from these four parts of the country are different? Use α = .05. Solution: (S1) There are k = 4 treatments and each sample is assumed to be independent and randomly selected. Each population is assumed to be normally distributed with a mean of µi (for i = 1, 2, 3, 4) and common variance σ 2 . (S2) Summary statistics: T1. = 99,

T2. = 244.7,

T3. = 231.8,

T4. = 139.7,

T.. = 715.2

(S3) Sum of squares: SST =

ni 4  

2 yij −

i=1 j=1

SSA =

T..2 = 9269.7 − 715.22 /57 = 295.82 N

 2  4  T2 244.72 231.82 139.72 Ti.2 99 − .. = − 715.22 /57 + + + ni N 10 18 18 11 i=1

= 9065.92 − 8973.88 = 92.04 SSE = SST − SSA = 295.82 − 92.04 = 203.78 (S4) The analysis of variance table: Source of variation

Sum of squares

Degrees of freedom

Mean square

Treatments Error Total

92.04 203.78 295.82

3 53 56

30.68 3.84

Computed F 7.98

(S5) Hypothesis test for significant regression (see section 12.1.3): H0 : µ 1 = µ 2 = µ3 = µ 4 = 0 Ha : at least two of the means are unequal 2 /S 2 TS: F = SA

RR: F ≥ F.05,3,53 = 2.78 Conclusion: The value of the test statistic lies in the rejection region. There is evidence to suggest at least two of the mean lengths are different. c 2000 by Chapman & Hall/CRC 

12.2

TWO-WAY ANOVA

12.2.1

One observation per cell

12.2.1.1

Models and assumptions

Let Yij be the random observation in the ith row and the j th column for i = 1, 2, . . . , r and j = 1, 2, . . . , c. Fixed effects experiment: Model:

Yij = µ + αi + βj + 3ij ind

Assumptions: 3ij ∼ N(0, σ 2 ),

r 

αi =

i=1

c 

βj = 0

j=1

Random effects experiment: Model:

Yij = µ + Ai + Bj + 3ij ind

ind

ind

Assumptions: 3ij ∼ N(0, σ 2 ), Ai ∼ N(0, σα2 ), Bj ∼ N(0, σβ2 ) The 3ij ’s, Ai ’s, and Bj ’s are independent. Mixed Effects Experiment: Model:

Yij = µ + Ai + βj + 3ij ind

ind

Assumptions: 3ij ∼ N(0, σ 2 ), Ai ∼ N(0, σα2 ),

c 

βj = 0

j=1

The 3ij ’s and Ai ’s are independent. 12.2.1.2

Sum of squares

Dots in the subscript of y and T indicate the mean and sum of yij , respectively, over the appropriate subscript(s). SST =

r  c 

(yij − y .. )2 =

i=1 j=1

SSR = c

r 

r  c  i=1 j=1

r 

(y i. − y .. )2 =

i=1

Ti.2

c

i=1

c 

SSC = r

c 

(y .j − y .. )2 =

j=1

j=1

SSE =

r  c 

2 yij −



T..2 rc



T..2 rc

T.j2

r

T..2 rc

(yij − y i. − y .j + y .. )2 = SST − SSR − SSC

i=1 j=1

c 2000 by Chapman & Hall/CRC 

12.2.1.3

Mean squares and properties

SSR 2 = SR = mean square due to rows r−1 SSC 2 MSC = = SC = mean square due to columns c−1 SSE MSE = = S 2 = mean square due to error (r − 1)(c − 1)

MSR =

Mean

Expected value

Square

Fixed model  r   2 α i   σ 2 + c i=1 r−1   c βi2 j=1  σ2 + r c−1

MSR

MSC MSE

σ

2

Random model

Mixed model

σ 2 + cσα2

σ 2 + cσα2  c

σ 2 + rσβ2 σ

σ2 + r

2

σ

j=1

βi2

 

c−1

2

(1) F = has an F distribution with r − 1 and (r − 1)(c − 1) degrees of freedom. 2 (2) F = SC /S 2 has an F distribution with c − 1 and (r − 1)(c − 1) degrees of freedom. 2 SR /S 2

12.2.2

Analysis of variance table

Source of Sum of variation squares

Degrees of freedom

Mean square

Rows Columns Error

SSR SSC SSE

r−1 MSR c−1 MSC (r − 1)(c − 1) MSE

Total

SST

rc − 1

Computed F MSR/MSE MSC/MSE

Hypothesis tests: (1) Test for significant row effect H0 : There is no effect due to rows (Fixed effects model: α1 = α2 = · · · = αr = 0) (Random effects model: σα2 = 0) (Mixed effects model: σα2 = 0)

c 2000 by Chapman & Hall/CRC 

Ha : There is an effect due to rows (Fixed effects model: αi = 0 for some i) (Random effects model: σα2 = 0) (Mixed effects model: σα2 = 0) 2 TS: F = SR /S 2

RR: F ≥ Fα,r−1,(r−1)(c−1) (2) Test for significant column effect H0 : There is no effect due to columns (Fixed effects model: β1 = β2 = · · · = βc = 0) (Random effects model: σβ2 = 0) (Mixed effects model: β1 = β2 = · · · = βc = 0) Ha : There is an effect due to columns (Fixed effects model: βj = 0 for some j) (Random effects model: σβ2 = 0) (Mixed effects model: βj = 0 for some j) 2 TS: F = SC /S 2

RR: F ≥ Fα,c−1,(r−1)(c−1) 12.2.3 12.2.3.1

Nested classifications with equal samples Models and assumptions

Let Yijk be the k th random observation for the ith level of factor A and the j th level of factor B. There are n observations for each factor combination: i = 1, 2, . . . , a, j = 1, 2, . . . , b, and k = 1, 2, . . . , n. Fixed effects experiment: Model:

Yijk = µ + αi + βj(i) + 3k(ij) ind

Assumptions: 3k(ij) ∼ N(0, σ 2 ),

a 

αi = 0,

i=1

b 

βj(i) = 0 for all i

j=1

Random effects experiment: Model:

Yijk = µ + Ai + Bj(i) + 3k(ij) ind

ind

ind

Assumptions: 3k(ij) ∼ N(0, σ 2 ), Ai ∼ N(0, σα2 ), Bj(i) ∼ N(0, σβ2 ) The 3k(ij) ’s, Ai ’s, and Bj(i) ’s are independent.

c 2000 by Chapman & Hall/CRC 

Mixed effects experiment (α): Model:

Yijk = µ + αi + Bj(i) + 3k(ij) a  ind ind Assumptions: 3k(ij) ∼ N(0, σ 2 ), αi = 0, Bj(i) ∼ N(0, σβ2 ) i=1

The 3k(ij) ’s and Bj(i) ’s are independent. Mixed effects experiment (β): Model:

Yijk = µ + Ai + βj(i) + 3k(ij) ind

ind

Assumptions: 3k(ij) ∼ N(0, σ 2 ), Ai ∼ N(0, σα2 ),

b 

βj(i) = 0 for all i

j=1

The 3k(ij) ’s and Ai ’s are independent. 12.2.3.2

Sum of squares

Dots in the subscript of y and T indicate the mean and sum of yijk , respectively, over the appropriate subscript(s). SST =

a  b  n 

a  b  n 

(yijk − y ... )2 =

i=1 j=1 k=1

SSA =

a 

i=1 j=1 k=1 a 

ni. (y i.. − y ... ) = 2

i=1

SSB(A) =

a  b 

i=1

2 Ti..

ni.

nij (y ij. − y i.. )2 =

i=1 j=1

SSE =

a  b  n 

2 yijk −



2 T... abn

2 T... abn

a  b 2  Tij. i=1 j=1

nij



a  T2

i..

i=1

ni.

(yijk − y ij. )2 = SST − SSA − SSB(A)

i=1 j=1 k=1

12.2.3.3

Mean squares and properties

SSA a−1 SSB(A) MSB(A) = a(b − 1) MSA =

MSE =

2 = SA

= mean square due to factor A

2 = SB(A) = mean square due to factor B

SSE = S2 ab(n − 1)

c 2000 by Chapman & Hall/CRC 

= mean square due to error

Expected value Fixed model Random model Mixed model (α) Mixed model (β) a  2 αi σ 2 + bn i=1 σ 2 + bnσα2 + nσβ2 a−1  a   2 α i   σ 2 + nσβ2 + bn i=1 σ 2 + bnσα2 a−1

Mean square

MSA

a  b 

σ2 + n

MSB(A)

i=1 j=1

2 βj(i)

σ 2 + nσβ2

a(b − 1)

a  b  2

σ + MSE

σ

nσβ2

2

σ +n

2

σ

σ2

i=1 j=1

2 βj(i)

a(b − 1)

2

σ2

2 (1) F = SA /S 2 has an F distribution with a − 1 and ab(n − 1) degrees of freedom. 2 (2) F = SB(A) /S 2 has an F distribution with a(b − 1) and ab(n − 1) degrees of freedom.

12.2.3.4

Analysis of variance table

Source of variation

Sum of squares

Degrees of freedom

Mean square

Between main groups

SSA

a−1

MSA

MSA MSE

Subgroups within main groups Error

SSB(A)

a(b − 1)

MSB(A)

MSB(A) MSE

SSE

ab(n − 1)

MSE

Total

SST

abn − 1

Hypothesis tests: (1) Test for significant factor A main effect

c 2000 by Chapman & Hall/CRC 

Computed F

H0 : There is no effect due to factor A (Fixed effects model: α1 = α2 = · · · = αa = 0) (Random effects model: σα2 = 0) (Mixed effects model (α): α1 = α2 = · · · = αa = 0) (Mixed effects model (β): σα2 = 0) Ha : There is an effect due to factor A (Fixed effects model: αi = 0 for some i) (Random effects model: σα2 = 0) (Mixed effects model (α): αi = 0 for some i) (Mixed effects model (β): σα2 = 0) 2 TS: F = SA /S 2

RR: F ≥ Fα,a−1,ab(n−1) (2) Test for significant factor B specific effect H0 : There is no effect due to factor B (Fixed effects model: all βj(i) = 0) (Random effects model: σβ2 = 0) (Mixed effects model (α): σβ2 = 0) (Mixed effects model (β): all βj(i) = 0) Ha : There is an effect due to factor B (Fixed effects model: not all βj(i) = 0) (Random effects model: σβ2 = 0) (Mixed effects model (α): σβ2 = 0) (Mixed effects model (β): not all βj(i) = 0) 2 TS: F = SB(A) /S 2

RR: F ≥ Fα,a(b−1),ab(n−1) 12.2.4 12.2.4.1

Nested classifications with unequal samples Models and assumptions

Let Yijk be the k th random observation for the ith level of factor A and the j th level of factor B. There are nij observations for each factor combination: mi a   i = 1, 2, . . . , a, j = 1, 2, . . . , mi , k = 1, 2, . . . , nij , and nij = n. i=1 j=1

c 2000 by Chapman & Hall/CRC 

Fixed effects experiment: Model:

Yijk = µ + αi + βj(i) + 3k(ij) ind

Assumptions: 3k(ij) ∼ N(0, σ 2 ),

a 

αi = 0,

i=1

mi 

βj(i) = 0 for all i

j=1

Random effects experiment: Model:

Yijk = µ + Ai + Bj(i) + 3k(ij) ind

ind

ind

Assumptions: 3k(ij) ∼ N(0, σ 2 ), Ai ∼ N(0, σα2 ), Bj(i) ∼ N(0, σβ2 ) The 3k(ij) ’s, Ai ’s, and Bj(i) ’s are independent. Mixed effects experiment (α): Model:

Yijk = µ + αi + Bj(i) + 3k(ij) a  ind ind Assumptions: 3k(ij) ∼ N(0, σ 2 ), αi = 0, Bj(i) ∼ N(0, σβ2 ) i=1

The 3k(ij) ’s and Bj(i) ’s are independent. Mixed effects experiment (β): Model:

Yijk = µ + Ai + βj(i) + 3k(ij) ind

ind

Assumptions: 3k(ij) ∼ N(0, σ 2 ), Ai ∼ N(0, σα2 ),

mi 

βj(i) = 0 for all i

j=1

The 3k(ij) ’s and Ai ’s are independent. 12.2.4.2

Sum of squares

Dots in the subscript of y and T indicate the mean and sum of yijk , respectively, over the appropriate subscript(s). SST =

nij mi  a  

nij mi  a  

(yijk − y ... )2 =

i=1 j=1 k=1

SSA =

a 

i=1 j=1 k=1 a 

ni. (y i.. − y ... ) = 2

i=1

SSB(A) =

mi a   i=1 j=1

SSE =

i=1

2 Ti..

ni.

nij (y ij. − y i.. )2 =

nij mi  a  

2 yijk −



2 T... n

2 T... n

mi a  2  Tij. i=1 j=1

nij



a  T2

i..

i=1

ni.

(yijk − y ij. )2 = SST − SSA − SSB(A)

i=1 j=1 k=1

c 2000 by Chapman & Hall/CRC 

12.2.4.3

Mean squares and properties

SSA a−1 SSB(A) MSB(A) =  a mi

2 = SA

MSA =

= mean square due to factor A

2 = SB(A) = mean square due to factor B

i=1

MSE =

SSE = S2 a  n− mi

= mean square due to error

i=1

Expected value Fixed model Random model Mixed model (α) Mixed model (β) a  mi αi2 σ 2 + i=1 σ 2 + c1 σα2 + c2 σβ2 a−1 a  mi αi2 i=1 2 2 σ + c2 σβ + σ 2 + c1 σα2 a−1 mi a   2 nij βj(i)

Mean square

MSA

σ2 +

MSB(A)

i=1 j=1 a 

σ 2 + c3 σβ2

mi − a

i=1 mi a  

σ 2 + c3 σβ2

σ2 +

i=1 j=1 a 

2 nij βj(i)

mi − a

i=1

MSE

σ2

σ2

σ2

σ2

where a 

c1 =

m i j=1

i=1

n2ij

mi



a m  i i=1 j=1

a 

n2ij

n

a−1

,

c2 =

m2i

n − i=1n a−1

n− ,

c3 =

a 

i=1 a 

m i j=1

n2ij

mi

.

mi − k

i=1 2 (1) F = SA /S 2 has an F distribution with a − 1 and n −

freedom.

c 2000 by Chapman & Hall/CRC 

a  i=1

mi degrees of

2 (2) F = SB(A) /S 2 has an F distribution with

a 

mi − a and n −

i=1

i=1

degrees of freedom. 12.2.4.4

Analysis of variance table

Source of variation

Sum of squares

Degrees of freedom

Mean square

Between main groups

SSA

a−1

MSA

MSA MSE

Subgroups within main groups

SSB(A)

MSB(A)

MSB(A) MSE

Error

SSE

a 

mi − a

i=1

n−

a 

mi

Computed F

MSE

i=1

Total

n−1

SST

Hypothesis tests: (1) Test for significant factor A main effect H0 : There is no effect due to factor A (Fixed effects model: α1 = α2 = · · · = αa = 0) (Random effects model: σα2 = 0) (Mixed effects model (α): α1 = α2 = · · · = αa = 0) (Mixed effects model (β): σα2 = 0) Ha : There is an effect due to factor A (Fixed effects model: αi = 0 for some i) (Random effects model: σα2 = 0) (Mixed effects model (α): αi = 0 for some i) (Mixed effects model (β): σα2 = 0) 2 TS: F = SA /S 2

RR: F ≥ F

α,a−1,n−

a 

mi

i=1

(2) Test for significant factor B specific effect

c 2000 by Chapman & Hall/CRC 

a 

mi

H0 : There is no effect due to factor B (Fixed effects model: all βj(i) = 0) (Random effects model: σβ2 = 0) (Mixed effects model (α): σβ2 = 0) (Mixed effects model (β): all βj(i) = 0) Ha : There is an effect due to factor B (Fixed effects model: not all βj(i) = 0) (Random effects model: σβ2 = 0) (Mixed effects model (α): σβ2 = 0) (Mixed effects model (β): not all βj(i) = 0) 2 TS: F = SB(A) /S 2

RR: F ≥ F

α,

a 

mi −a,n−

i=1

a 

mi

i=1

12.2.5

Two-factor experiments

12.2.5.1

Models and assumptions

Let Yijk be the k th random observation for the ith level of factor A and the j th level of factor B. There are n observations for each factor combination: i = 1, 2, . . . , a, j = 1, 2, . . . , b, and k = 1, 2, . . . , n. Fixed effects experiment: Model:

Yijk = µ + αi + βj + (αβ)ij + 3ijk ind

Assumptions: 3ijk ∼ N(0, σ ), 2

a 

αi = 0,

i=1 a 

(αβ)ij =

i=1

b 

b 

βj = 0

j=1

(αβ)ij = 0

j=1

Random effects experiment: Model:

Yijk = µ + Ai + Bj + (AB)ij + 3ijk ind

ind

ind

Assumptions: 3ijk ∼ N(0, σ 2 ), Ai ∼ N(0, σα2 ), Bj ∼ N(0, σβ2 ) ind

2 ) (AB)ij ∼ N(0, σαβ

The 3ijk ’s, Ai ’s, Bj ’s, and (AB)ij ’s are independent. Mixed effects experiment (α): Model:

Yijk = µ + αi + Bj + (αB)ij + 3ijk a a   ind 2 Assumptions: 3ijk ∼ N(0, σ ), αi = 0, (αB)ij = 0 for all j i=1

ind

Bj ∼

N(0, σβ2 ),

i=1

ind

2 (αB)ij ∼ N(0, a−1 a σαβ )

The 3ijk ’s, Bj ’s, and (αB)ij ’s are independent. c 2000 by Chapman & Hall/CRC 

12.2.5.2

Sum of squares

Dots in the subscript of y and T indicate the mean and sum of yijk , respectively, over the appropriate subscript(s). SST =

a  b  n 

(yijk − y ... )2 =

i=1 j=1 k=1

SSA = bn

a 

a  b  n  i=1 j=1 k=1

a 

(y i.. − y ... ) = 2

i=1

2 Ti..



bn

i=1

b 

SSB = an

b 

(y .j. − y ... ) = 2

j=1

a  b 

2 T... abn

2 T... abn

2 T.j.



an

j=1

SS(AB) = n

2 yijk −

2 T... abn

(y ij. − y i.. − y .j. + y ... )2

i=1 j=1 a 

i=1 j=1

= SSE =

b 

a 

2 Tij.

n a  b  n 



i=1

b 

2 Ti..

bn



j=1

2 T.j.

an

+

2 T... abn

(yijk − y ij. )2 = SST − SSA − SSB − SS(AB)

i=1 j=1 k=1

12.2.5.3

Mean squares and properties

SSA 2 = SA = mean square due to factor A a−1 SSB 2 MSB = = mean square due to factor B = SB b−1 SS(AB) 2 MS(AB) = = mean square due to interaction = SAB (a − 1)(b − 1) MSA =

MSE =

SSE ab(n − 1)

= S2

c 2000 by Chapman & Hall/CRC 

= mean square due to error

Mean square Expected value Random model

Fixed model

Mixed model (α)

MSA a 

σ 2 + nb

i=1

a 

αi2

a−1

i=1

2 σ 2 + nbσα2 + σαβ

σ 2 + nb

2 σ 2 + naσβ2 + nσαβ

σ 2 + naσβ2

2 σ 2 + nσαβ

2 σ 2 + nσαβ

σ2

σ2

αi2

a−1

2 + nσαβ

MSB b  2

σ + na

j=1

βj2

b−1

MS(AB) a  b  2

σ +n

i=1 j=1

(αβ)2ij

(a − 1)(b − 1)

MSE σ2

2 (1) F = SA /S 2 has an F distribution with a − 1 and ab(n − 1) degrees of freedom. 2 (2) F = SB /S 2 has an F distribution with b − 1 and ab(n − 1) degrees of freedom. 2 (3) F = SAB /S 2 has an F distribution with (a − 1)(b − 1) and ab(n − 1) degrees of freedom.

12.2.5.4

Analysis of variance table

Source of variation

Sum of squares

Degrees of freedom

Mean square

Factor A

SSA

a−1

MSA

MSA

SSB

b−1

MSB

MSB MS(AB)

Factor B

Computed F MSE MSE

Interaction AB

SS(AB)

(a−1)(b−1)

MS(AB)

Error

SSE

ab(n − 1)

MSE

Total

SST

abn − 1

MSE

Hypothesis tests: (1) Test for significant factor A main effect

c 2000 by Chapman & Hall/CRC 

H0 : There is no effect due to factor A (Fixed effects model: α1 = α2 = · · · = αa = 0) (Random effects model: σα2 = 0) (Mixed effects model (α): α1 = α2 = · · · = αa = 0) Ha : There is an effect due to factor A (Fixed effects model: αi = 0 for some i) (Random effects model: σα2 = 0) (Mixed effects model (α): αi = 0 for some i) 2 TS: F = SA /S 2

RR: F ≥ Fα,a−1,ab(n−1) (2) Test for significant factor B main effect H0 : There is no effect due to factor B (Fixed effects model: β1 = β2 = · · · = βb = 0) (Random effects model: σβ2 = 0) (Mixed effects model (α): σβ2 = 0) Ha : There is an effect due to factor B (Fixed effects model: βj = 0 for some j) (Random effects model: σβ2 = 0) (Mixed effects model (α): σβ2 = 0) 2 TS: F = SB /S 2

RR: F ≥ Fα,b−1,ab(n−1) (3) Test for significant AB interaction effect H0 : There is no effect due to interaction (Fixed effects model: (αβ)11 = (αβ)12 = · · · = (αβ)ab = 0 ) 2 =0 (Random effects model: σαβ 2 (Mixed effects model (α): σαβ =0 Ha : There is an effect due to interaction (Fixed effects model: (αβ)ij = 0 for some ij 2

= 0 (Random effects model: σαβ 2 (Mixed effects model (α): σαβ

= 0 2 TS: F = SAB /S 2

RR: F ≥ Fα,(a−1)(b−1),ab(n−1) 12.2.6

Example

Example 12.67 : An electrical engineer believes the brand of battery and the style of music played most often may have an effect on the lifetime of batteries in a portable CD player. Random samples were selected and the lifetime of each battery (in hours)

c 2000 by Chapman & Hall/CRC 

was recorded. The data are given in the table below. Battery Brand B C

A

D

Easy listening 61.1 58.3 60.3 68.8 58.0 59.9 55.7 48.9 61.3 69.2 69.5 61.9 66.4 60.3 64.1 60.5 Style Country

62.2 55.9 64.0 53.4 64.5 59.2 61.8 59.0 63.1 67.0 64.3 64.8 66.0 56.4 54.1 50.5 64.2 57.0 58.7 61.1 62.8 63.0 58.7 57.3 59.1 48.7 62.3 70.8 65.1 64.1 60.9 63.0

Rock

Is there any evidence to suggest a difference in battery life due to brand or music style? Solution: (S1) A fixed effects experiment is assumed. There are i = 3 styles of music, j = 4 battery brands, and k = 4 observations for each factor combination. (S2) The analysis of variance table: Source of variation

Sum of Degrees of Mean squares freedom square Computed F

Style 10.2 Brand 199.7 Interaction 125.8 Error 822.5 Total 1158.1

2 3 6 36 47

5.1 66.6 21.0 22.8

0.22 2.91 0.92

(S3) Test for significant interaction effect: H0 : There is no effect due to interaction. Ha : There is an effect due to interaction. 2 TS: F = SAB /S 2

RR: F ≥ F.05,6,36 = 2.36 Conclusion: The value of the test statistic (F = 0.92) does not lie in the rejection region. There is no evidence to suggest an interaction effect. Tests for main effects may be analyzed as though there were no interaction. (S4) Test for significant effect due to style: H0 : There is no effect due to style. Ha : There is an effect due to style. 2 /S 2 TS: F = SA

RR: F ≥ F.05,2,36 = 3.26 Conclusion: The value of the test statistic (F = 0.22) does not lie in the rejection region. There is no evidence to suggest a difference in battery life due to style of music. (S5) Test for significant effect due to brand: H0 : There is no effect due to brand. Ha : There is an effect due to brand. c 2000 by Chapman & Hall/CRC 

2 TS: F = SB /S 2

RR: F ≥ F.05,3,36 = 2.86 Conclusion: The value of the test statistic (F = 2.91) lies in the rejection region. There is some evidence to suggest a difference in battery life due to brand.

12.3

THREE-FACTOR EXPERIMENTS

12.3.1

Models and assumptions

Let Yijkl be the lth random observation for the ith level of factor A, the j th level of factor B, and the k th level of factor C. There are n observations for each factor combination: i = 1, 2, . . . , a, j = 1, 2, . . . , b, and k = 1, 2, . . . , c, l = 1, 2, . . . , n. Fixed effects experiment: Model: Yijkl = µ + αi + βj + γk + (αβ)ij + (αγ)ik + (βγ)jk + (αβγ)ijk + 3ijkl Assumptions: ind

3ijkl ∼ N(0, σ 2 ),

a 

αi = 0,

i=1 a 

βj = 0,

j=1

c 

γk = 0,

k=1

b c c b c      (αβ)ij = (αβ)ij = (αγ)ik = (αγ)ik = (βγ)jk = (βγ)jk = 0,

i=1 a 

b 

j=1

(αβγ)ijk =

i=1

i=1 b 

(αβγ)ijk =

j=1

j=1

k=1 c 

k=1

(αβγ)ijk = 0

k=1

Random effects experiment: Model: Yijkl = µ+Ai +Bj +Ck +(AB)ij +(AC)ik +(BC)jk +(ABC)ijk +3ijkl Assumptions: ind

ind

ind

ind

3ijkl ∼ N(0, σ 2 ), Ai ∼ N(0, σα2 ), Bj ∼ N(0, σβ2 ), Ck ∼ N(0, σγ2 ) ind

ind

ind

2 2 2 (AB)ij ∼ N(0, σαβ ), (AC)ik ∼ N(0, σαγ ), (BC)jk ∼ N(0, σβγ ),

ind

2 (ABC)ijk ∼ N(0, σαβγ ).

The 3ijkl ’s are independent of the other random components.

c 2000 by Chapman & Hall/CRC 

Mixed effects experiment (α): Model: Yijkl = µ+αi +Bj +Ck +(αB)ij +(αC)ik +(BC)jk +(αBC)ijk +3ijkl Assumptions: ind

3ijk ∼ N(0, σ 2 ),

a 

αi =

i=1

ind

a 

(αB)ij =

i=1

a 

(αC)ik =

i=1

ind

a 

(αBC)ijk = 0 ,

i=1

ind

2 Bj ∼ N(0, σβ2 ), Ck ∼ N(0, σγ2 ), (αB)ij ∼ N(0, σαβ )

ind

ind

ind

2 2 2 ), (BC)jk ∼ N(0, σβγ ), (αBC)ijk ∼ N(0, σαβγ ) (αC)ik ∼ N(0, σαγ

The 3ijkl ’s are independent of the other random components. Mixed effects experiment (α, β): Model: Yijkl = µ+αi +βj +Ck +(αβ)ij +(αC)ik +(βC)jk +(αβC)ijk +3ijkl Assumptions: ind

3ijkl ∼ N(0, σ ), 2

a 

αi =

i=1 a 

(αβ)ij =

i=1

b 

(αβ)ij =

j=1

ind

b  j=1

a  i=1

ind

βj =

a 

(αC)ik =

i=1

(αβC)ijk =

b 

(βC)jk = 0,

j=1 b 

(αβC)ijk = 0,

j=1

ind

2 2 Ck ∼ N(0, σγ2 ), (αC)ik ∼ N(0, σαγ ), (βC)jk ∼ N(0, σβγ ),

ind

2 (αβC)ijk ∼ N(0, σαβγ )

The 3ijkl ’s are independent of the other random components. 12.3.2

Sum of squares

Dots in the subscript of y and T indicate the mean and sum of yijkl , respectively, over the appropriate subscript(s).

c 2000 by Chapman & Hall/CRC 

SST =

a  b  c  n 

a  b  c  n 

(yijkl − y .... )2 =

i=1 j=1 k=1 l=1

SSA = bcn

a 

i=1 j=1 k=1 l=1 a  i=1

(y i... − y .... ) = 2

2 Ti...

b 

SSB = acn

j=1

(y .j.. − y .... ) = 2

SSC = abn

c 

c  k=1

(y ..k. − y .... ) = 2

a  b 

2 T.... abcn



2 T..k.



abn

k=1

SS(AB) = cn

2 T.j..

acn

j=1

2 T.... abcn

2 T.... abcn



bn

i=1 b 

2 yijkl −

2 T.... abcn

(y ij.. − y i... − y .j.. + y .... )2

i=1 j=1 b a  

=

i=1 j=1

SS(AC) = bn

a 

2 Tij..

cn a  c 



i=1

b 

2 Ti...

bcn



2 T.j..

j=1

+

acn

2 T.... abcn

(y i.k. − y i... − y ..k. + y .... )2

i=1 k=1 a 

=

c 

i=1 k=1

SS(BC) = an

a 

2 Ti.k.

bn b  c 



i=1

c 

2 Ti...

bcn



2 T..k.

k=1

+

abn

2 T.... abcn

(y .jk. − y .j.. − y ..k. + y .... )2

j=1 k=1 b 

= SS(ABC) =

c 

b 

2 T.jk.

j=1 k=1



an a  b  c 

j=1

c 

2 T.j..

k=1



acn

2 T..k.

+

abn

2 T.... abcn

(y ijk. −y ij.. −y i.k. −y .jk. +y i... +y .j.. +y ..k. −y .... )2

i=1 j=1 k=1 a  b  c 

=

i=1 j=1 k=1



n a 

+

i=1

i=1 j=1

b 

2 Ti...

bcn

a  b 

2 Tijk.

+

c 2000 by Chapman & Hall/CRC 

j=1



cn c 

2 T.j..

acn

a  b 

2 Tij..

+

k=1

i=1 j=1

bn

2 T..k.

abn

+

b  c 

2 Ti.k.

2 T.... abcn



j=1 k=1

an

2 T.jk.

SSE =

a  b  c  n 

(yijkl − y ijk. )2

i=1 j=1 k=1 l=1

= SST − SSA − SSB − SSC − SS(AB) − SS(AC) − SS(BC) − SS(ABC) 12.3.3

Mean squares and properties

SSA 2 = SA = mean square due to factor A a−1 SSB 2 MSB = = mean square due to factor B = SB b−1 SSC 2 MSC = = SC = mean square due to factor C c−1 SS(AB) MS(AB) = (a − 1)(b − 1) MSA =

2 = SAB = mean square due to AB interaction

MS(AC) =

SS(AC) (a − 1)(c − 1)

2 = SAC = mean square due to AC interaction

MS(BC) =

SS(BC) (b − 1)(c − 1)

2 = SBC = mean square due to BC interaction

MS(ABC) =

SS(ABC) (a − 1)(b − 1)(c − 1)

2 = SABC = mean square due to ABC interaction

MSE =

SSE = S 2 = mean square due to error abc(n − 1)

c 2000 by Chapman & Hall/CRC 

Mean square Expected value Random model

Fixed model MSA a  2

σ + bcn

i=1

αi2 2 2 2 σ 2 + bcnσα2 + cnσαβ + bnσαγ + nσαβγ

a−1

MSB b 

σ 2 + acn

j=1

βj2 2 2 2 σ 2 + acnσβ2 + cnσαβ + anσβγ + nσαβγ

b−1

MSC c  2

σ + abn

k=1

γk2 2 2 2 σ 2 + abnσγ2 + bnσαγ + anσβγ + nσαβγ

c−1

MS(AB) a  b  i=1 j=1

σ 2 + cn

(αβ)2ij

(a − 1)(b − 1)

MS(AC) a  c  i=1 k=1

σ 2 + bn

2 2 σ 2 + cnσαβ + nσαβγ

(αγ)2ik

(a − 1)(c − 1)

2 2 σ 2 + bnσαγ + nσαβγ

MS(BC) b  c 

σ 2 + an

j=1 k=1

(βγ)2jk

(b − 1)(c − 1)

MS(ABC) a  b  c  2

σ +n

i=1 j=1 k=1

2 2 σ 2 + anσβγ + nσαβγ

(αβγ)2ijk

(a − 1)(b − 1)(c − 1)

2 σ 2 + nσαβγ

MSE σ2

c 2000 by Chapman & Hall/CRC 

σ2

Mean square Expected value Mixed model (α, β)

Mixed model (α) MSA a 

σ 2 + bcn

i=1

a 

αi2

a−1

2 + cnσαβ

σ 2 + bcn

i=1

αi2 2 + bnσαγ

a−1

2 2 + bnσαγ + nσαβγ

MSB b  2 σ 2 + acnσβ2 + anσβγ

σ 2 + acn

j=1

βj2

b−1

2 + anσβγ

MSC 2 σ 2 + abnσγ2 + anσβγ

σ 2 + abnσγ2

MS(AB) a  b  2 2 σ 2 + cnσαβ + nσαβγ

σ 2 + cn

i=1 j=1

(αβ)2ij

(a − 1)(b − 1)

2 + nσαβγ

MS(AC) 2 2 σ 2 + bnσαγ + nσαβγ

2 σ 2 + bnσαγ

MS(BC) 2 σ 2 + anσβγ

2 σ 2 + anσβγ

MS(ABC) 2 σ 2 + nσαβγ

2 σ 2 + nσαβγ

MSE σ2

σ2

The following statistics have F distributions with the stated degrees of freedom. Statistic

Numerator df

Denominator df

2 SA /S 2 2 SB /S 2 2 SC /S 2 2 SAB /S 2 2 SAC /S 2 2 SBC /S 2 2 SABC /S 2

a−1

abc(n − 1)

b−1

abc(n − 1)

c−1

abc(n − 1)

(a − 1)(b − 1)

abc(n − 1)

(a − 1)(c − 1)

abc(n − 1)

(b − 1)(c − 1)

abc(n − 1)

(a − 1)(b − 1)(c − 1) abc(n − 1)

c 2000 by Chapman & Hall/CRC 

12.3.4

Analysis of variance table

Source of variation

Sum of squares

Degrees of freedom

Mean square

Factor A

SSA

a−1

MSA

Factor B

SSB

b−1

MSB

Factor C

SSC

c−1

MSC

A×B

SS(AB)

(a−1)(b−1)

MS(AB)

A×C

SS(AC)

(a−1)(c−1)

MS(AC)

B×C

SS(BC)

(b−1)(c−1)

MS(BC)

A × B × C SS(ABC) (a−1)(b−1)(c−1) MS(ABC) Error

SSE

abc(n − 1)

Total

SST

abcn − 1

Computed F MSA MSE MSB MSE MSC MSE MS(AB) MSE MS(AC) MSE MS(BC) MSE MS(ABC) MSE

MSE

Hypothesis tests: (1) Test for significant factor A main effect H0 : There is no effect due to factor A (Fixed effects model: α1 = α2 = · · · = αa = 0) (Random effects model: σα2 = 0) (Mixed effects model (α): α1 = α2 = · · · = αa = 0) (Mixed effects model (α, β): α1 = α2 = · · · = αa = 0) Ha : There is an effect due to factor A (Fixed effects model: αi = 0 for some i) (Random effects model: σα2 = 0) (Mixed effects model (α): αi = 0 for some i) (Mixed effects model (α, β): αi = 0 for some i) 2 TS: F = SA /S 2

RR: F ≥ Fα,a−1,abc(n−1) (2) Test for significant factor B main effect

c 2000 by Chapman & Hall/CRC 

H0 : There is no effect due to factor B (Fixed effects model: β1 = β2 = · · · = βb = 0) (Random effects model: σβ2 = 0) (Mixed effects model (α): σβ2 = 0) (Mixed effects model (α, β): β1 = β2 = · · · = βb = 0) Ha : There is an effect due to factor B (Fixed effects model: βj = 0 for some j) (Random effects model: σβ2 = 0) (Mixed effects model (α): σβ2 = 0) (Mixed effects model (α, β): βj = 0 for some j) 2 TS: F = SB /S 2

RR: F ≥ Fα,b−1,abc(n−1) (3) Test for significant factor C main effect H0 : There is no effect due to factor C (Fixed effects model: γ1 = γ2 = · · · = γc = 0) (Random effects model: σγ2 = 0) (Mixed effects model (α): σγ2 = 0) (Mixed effects model (α, β): σγ2 = 0) Ha : There is an effect due to factor C (Fixed effects model: γk = 0 for some k) (Random effects model: σγ2 = 0) (Mixed effects model (α): σγ2 = 0) (Mixed effects model (α, β): σγ2 = 0) 2 TS: F = SC /S 2

RR: F ≥ Fα,c−1,abc(n−1) (4) Test for significant AB interaction effect H0 : There is no effect due to interaction (Fixed effects model: (αβ)11 = · · · = (αβ)ab = 0) 2 = 0) (Random effects model: σαβ 2 (Mixed effects model (α): σαβ = 0) (Mixed effects model (α, β): (αβ)11 = · · · = (αβ)ab = 0) Ha : There is an effect due to AB interaction (Fixed effects model: (αβ)ij = 0 for some ij) 2

= 0) (Random effects model: σαβ 2 (Mixed effects model (α): σαβ

= 0) (Mixed effects model (α, β): (αβ)ij = 0 for some ij) 2 TS: F = SAB /S 2

RR: F ≥ Fα,(a−1)(b−1),abc(n−1) (5) Test for significant AC interaction effect

c 2000 by Chapman & Hall/CRC 

H0 : There is no effect due to interaction (Fixed effects model: (αγ)11 = · · · = (αγ)ac = 0) 2 (Random effects model: σαγ = 0) 2 = 0) (Mixed effects model (α): σαγ 2 (Mixed effects model (α, β): σαγ = 0) Ha : There is an effect due to AC interaction (Fixed effects model: (αγ)ik = 0 for some ik) 2 (Random effects model: σαγ

= 0) 2 (Mixed effects model (α): σαγ

= 0) 2

= 0) (Mixed effects model (α, β): σαγ 2 TS: F = SAC /S 2

RR: F ≥ Fα,(a−1)(c−1),abc(n−1) (6) Test for significant BC interaction effect H0 : There is no effect due to interaction (Fixed effects model: (βγ)11 = · · · = (βγ)bc = 0) 2 (Random effects model: σβγ = 0) 2 (Mixed effects model (α): σβγ = 0) 2 (Mixed effects model (α, β): σβγ = 0) Ha : There is an effect due to BC interaction (Fixed effects model: (βγ)jk = 0 for some jk 2

= 0) (Random effects model: σβγ 2 (Mixed effects model (α): σβγ

= 0) 2 (Mixed effects model (α, β): σβγ

= 0) 2 TS: F = SBC /S 2

RR: F ≥ Fα,(b−1)(c−1),abc(n−1) (7) Test for significant ABC interaction effect H0 : There is no effect due to interaction (Fixed effects model: (αβγ)111 = · · · = (αβγ)abc = 0) 2 = 0) (Random effects model: σαβγ 2 (Mixed effects model (α): σαβγ = 0) 2 (Mixed effects model (α, β): σαβγ = 0) Ha : There is an effect due to BC interaction (Fixed effects model: (αβγ)ijk = 0 for some ijk) 2

= 0) (Random effects model: σαβγ 2 (Mixed effects model (α): σαβγ

= 0) 2 (Mixed effects model (α, β): σαβγ

= 0) 2 TS: F = SABC /S 2

RR: F ≥ Fα,(a−1)(b−1)(c−1),abc(n−1)

c 2000 by Chapman & Hall/CRC 

12.4

MANOVA

Manova means multiple anova, used if there are multiple dependent variables to be analyzed simultaneously. The use of repeated measurement is a subset of manova. Using multiple oneway anovas to do this will raise the probability of a Type I error. Manova controls the experiment-wide error rate. (While it may seem that several simultaneous anovas raise power, the Type I error rate increases also.) Manova assumptions: (a) Usual anova assumptions (normality, independence, HOV). (b) Linearity or multicollinearity of dependent variables. (c) Manova does not have the compound symmetry requirement that the one-factor repeated measures anova model requires. Manova advantages: (a) Manova is a “gateway” test. If the multivariate F test is significant, then individual univariate analyses may be considered. (b) Manova may be used with assorted dependent variables, or with repeated measures. This is an important feature of the model if the factors cannot be collapsed because they are all essentially different. (c) Manova may detect combined differences not found by univariate analyses if there is multicollinearity (a linear combination of the dependent variables). Manova limitations: (a) Manova may be very sensitive to outliers, for small sample sizes. (b) Manova assumes a linear relationship between dependent variables. (c) Manova cannot give the interaction effects between the main effect and a repeated factor. 12.5

FACTOR ANALYSIS

The purpose of factor analysis is to examine the covariance, or correlation, relationships among all of the variables. This technique is used to group variables that tend to move together into an unobservable, random quantity called a factor. Suppose highly correlated observable variables are grouped together, i.e., grouped by correlations. Variables within a group tend to move together and have very little correlation with variables outside their group. It is possible that each group of variables may be represented by, and depend on, a single, unobserved factor. Factor analysis attempts to discover this model structure so that each factor has a large correlation with a few variables and little correlation with the remaining variables.

c 2000 by Chapman & Hall/CRC 

Let X be a p×1 random vector with mean µ and variance-covariance matrix Σ. Assume X is a linear function of a set of unobservable factors, F1 , F2 , . . . , Fm , and p error terms, 31 , 32 , . . . , 3p . The factor analysis model may be written as X1 − µ1 = =11 F1 + =12 F2 + · · · + =1m Fm + 31 X2 − µ2 = =21 F1 + =22 F2 + · · · + =2m Fm + 32 .. .. . . Xp − µp = =p1 F1 + =p2 F2 + · · · + =pm Fm + 3p

(12.8)

In matrix notation, the factor analysis model may be written as X−µ = (p × 1)

L

F

(p × m) (m × 1)

+

(p × 1)

(12.9)

where L is the matrix of factor loadings and =ij is the loading of the ith variable on the j th factor. There are additional model assumptions involving the unobservable random vectors F and : (1) F and are independent. (2) E [F] = 0, Cov [F] = I. (3) E [ ] = 0, Cov [ ] = Ψ, where Ψ is a diagonal matrix. The orthogonal factor analysis model with m common factors (equation (12.8) and these assumptions) implies the following covariance structure for the random vector X: (1) Cov [X] = LLT + Ψ or Var [Xi ] = =2i1 + · · · + =2im + Ψi Cov [Xi , Xk ] = =i1 =k1 + · · · + =im =km

(12.10) (12.11)

(2) Cov [X, F] = L, or Cov [Xi , Fj ] = =ij The variance of the ith variable, σii , is the sum of two terms: the ith communality and the specific variance. σii = =2i1 + =2i2 + · · · + =2im +    Var[Xi ]

communality

Ψi 

(12.12)

specific variance

Factor analysis is most useful when the number of unobserved factors, m, is small relative to the number of observed random variables, p. The objective of the factor analysis model is to provide a simpler explanation for the relationships in X rather than referring to the complete variance-covariance Σ. A problem with this procedure is that most variance-covariance matrices cannot be written as in equation (12.11) with m much less than p. If m > 1, there are additional conditions necessary in order to obtain unique estimates of L and Ψ. The estimate of the loading matrix L is determined only up to an orthogonal (rotation) matrix. The rotation matrix is usually constructed so that the model may be realistically interpreted. c 2000 by Chapman & Hall/CRC 

There are two common procedures used to estimate the parameters =ij and Ψi : the method of principal components and the method of maximum likelihood. Each of these solutions may be rotated in order to more appropriately interpret the model. See, for example, R. A. Johnson and D. W. Wichern, Applied Multivariate Statistical Analysis, Fourth Edition, Prentice-Hall, Inc., Upper Saddle River, NJ, 1998. 12.6

LATIN SQUARE DESIGN

12.6.1

Models and assumptions

Let Yij(k) be the random observation corresponding to the ith row, the j th column, and the k th treatment. The parentheses in the subscripts are used to denote the one value k assumes for each ij combination: i, j, k = 1, 2, . . . , r. It is assumed there are no interactions among these three factors. Fixed effects experiment: Model:

Yij(k) = µ + αi + βj + γk + 3ij(k) r r r    ind Assumptions: 3ij(k) ∼ N(0, σ 2 ), αi = βj = γk = 0 i=1

j=1

k=1

Random effects experiment: Model:

Yij(k) = µ + Ai + Bj + Ck + 3ij(k)

Assumptions: ind

ind

ind

ind

3ij(k) ∼ N(0, σ 2 ), Ai ∼ N(0, σα2 ), Bj ∼ N(0, σβ2 ), Ck ∼ N(0, σγ2 ) The 3ij(k) ’s are independent of the other random components. Mixed effects experiment (γ): Model:

Yij(k) = µ + Ai + Bj + γk + 3ij(k)

Assumptions: ind

ind

ind

3ij(k) ∼ N(0, σ 2 ), Ai ∼ N(0, σα2 ), Bj ∼ N(0, σβ2 ),

r 

γk = 0

k=1

The 3ij(k) ’s are independent of the other random components. Mixed effects experiment (α, γ): Model:

Yij(k) = µ + αi + Bj + γk + 3ij(k)

Assumptions: ind

ind

3ij(k) ∼ N(0, σ 2 ), Bj ∼ N(0, σβ2 ),

r 

αi =

i=1

The 3ij(k) ’s and the Bj ’s are independent.

c 2000 by Chapman & Hall/CRC 

r  k=1

γk = 0

12.6.2

Sum of squares

Dots in the subscript of y and T indicate the mean and sum of yij(k) , respectively, over the appropriate subscript(s). SST =

r  r 

r  r 

(yij(k) − y ... )2 =

i=1 j=1

SSR = r

r 

i=1 j=1 r 

(y i.. − y ... )2 =

i=1

2 Ti..

i=1

r r 

SSC = r

r 

(y .j. − y ... ) = 2

j=1

SSTr = r

r 

j=1

(y ..k − y ... ) = 2

r  r 



2 T... r2



2 T... 2 r

2 T.j.

r r 

k=1

SSE =

2 yij(k) −

k=1

2 T..k

r



2 T... r2

2 T... 2 r

(yij(k) − y i.. − y .j. − y ..k + 2y ... )2

i=1 j=1

= SST − SSR − SSC − SSTr 12.6.3

Mean squares and properties

SSR r−1 SSC MSC = r−1 SSTr MSTr = r−1 SSE MSE = (r − 1)(r − 2) MSR =

2 = SR = mean square due to rows 2 = SC = mean square due to columns 2 = mean square due to treatments = STr

= S 2 = mean square due to error

c 2000 by Chapman & Hall/CRC 

Mean square Expected value Random Mixed model model (γ)

Fixed model MSR r  2

σ +r

i=1

Mixed model (α, γ) r 

αi2

r−1

2

σ +

rσα2

2

σ +

rσα2

2

σ +r

i=1

αi2

r−1

MSC r 

σ2 + r

j=1

βj2

r−1

σ 2 + rσβ2

σ 2 + rσβ2

σ 2 + rσβ2

MSTr r  2

σ +r

k=1

r 

γk2

r−1

2

σ +

rσγ2

2

σ +r

k=1

r 

γk2

r−1

2

σ +r

k=1

γk2

r−1

MSE σ2

σ2

σ2

σ2

2 (1) F = SR /S 2 has an F distribution with r − 1 and (r − 1)(r − 2) degrees of freedom. 2 (2) F = SC /S 2 has an F distribution with r − 1 and (r − 1)(r − 2) degrees of freedom. 2 (3) F = STr /S 2 has an F distribution with r − 1 and (r − 1)(r − 2) degrees of freedom.

12.6.4

Analysis of variance table

Source of variation

Sum of Degrees of squares freedom

Rows

SSR

r−1

MSA

Columns

SSC

r−1

MSB

Treatments SSTr

r−1

MSC

Error

SSE

(r − 1)(r − 2) MSE

Total

SST

r2 − 1

Hypothesis tests: (1) Test for significant row effect

c 2000 by Chapman & Hall/CRC 

Mean square Computed F MSR MSE MSC MSE MSTr MSE

H0 : There is no effect due to rows (Fixed effects model: α1 = α2 = · · · = αr = 0) (Random effects model: σα2 = 0) (Mixed effects model (γ): σα2 = 0) (Mixed effects model (α, γ): α1 = α2 = · · · = αr = 0) Ha : There is an effect due to rows (Fixed effects model: αi = 0 for some i) (Random effects model: σα2 = 0) (Mixed effects model (γ): σα2 = 0) (Mixed effects model (α, γ): αi = 0 for some i) 2 TS: F = SR /S 2

RR: F ≥ Fα,r−1,(r−1)(r−2) (2) Test for significant column effect H0 : There is no effect due to columns (Fixed effects model: β1 = β2 = · · · = βr = 0) (Random effects model: σβ2 = 0) (Mixed effects model (γ): σβ2 = 0) (Mixed effects model (α, γ): σβ2 = 0) Ha : There is an effect due to columns (Fixed effects model: βj = 0 for some j) (Random effects model: σβ2 = 0) (Mixed effects model (γ): σβ2 = 0) (Mixed effects model (α, γ): σβ2 = 0) 2 TS: F = SC /S 2

RR: F ≥ Fα,r−1,(r−1)(r−2) (3) Test for significant treatment effect H0 : There is no effect due to treatments (Fixed effects model: γ1 = γ2 = · · · = γr = 0) (Random effects model: σγ2 = 0) (Mixed effects model (γ): γ1 = γ2 = · · · = γr = 0) (Mixed effects model (α, γ): γ1 = γ2 = · · · = γr = 0) Ha : There is an effect due to treatments (Fixed effects model: γk = 0 for some k) (Random effects model: σγ2 = 0) (Mixed effects model (γ): γk = 0 for some k) (Mixed effects model (α, γ): γk = 0 for some k) 2 TS: F = STr /S 2

RR: F ≥ Fα,r−1,(r−1)(r−2)

c 2000 by Chapman & Hall/CRC 

CHAPTER 13

Experimental Design Contents 13.1 13.2 13.3 13.4 13.5 13.6 13.7

Latin squares Graeco–Latin squares Block designs Factorial experimentation: 2 factors 2r Factorial experiments Confounding in 2n factorial experiments Tables for design of experiments 13.7.1 Plans of factorial experiments confounded 13.7.2 Plans of 2n factorials in fractional replication 13.7.3 Plans of incomplete block designs 13.7.4 Interactions in factorial designs 13.8 References

13.1

LATIN SQUARES

A Latin square of order n is an n × n array in which each cell contains a single element from an n-set, such that each element occurs exactly once in each row and exactly once in each column. A Latin square is in standard form if in the first row and column the elements occur in natural order. The number of Latin squares in standard form are: n number

1 1

2 1

3 1

4 4

5 56

6 9,408

7 16,942,080

The unique Latin squares of order 1 and 2 are

8 535,281,401,856

A and

A B . There are 4 B A

Latin squares of order 4. 3×3 A B C B C A C A B

4×4 a A B C D

B A D C

c 2000 by Chapman & Hall/CRC 

b C D B A

D C A B

A B C D

B C D A

c C D A B

D A B C

A B C D

B D A C

d C A D B

D C B A

A B C D

B A D C

C D A B

D C B A

A B C D E

A B C D E F G H

A B C D E F G H I J

5×5 B C D A E C D A E E B A C D B

B C D E F G H A

B C D E F G H I J A

13.2

C D E F G H A B

C D E F G H I J A B

8×8 D E E F F G G H H A A B B C C D

D E F G H I J A B C

A B C D E F

E D B C A

F G H A B C D E

G H A B C D E F

10 × 10 E F G F G H G H I H I J I J A J A B A B C B C D C D E D E F

B F D A C E

6×6 C D D C E F F E A B B A

E A B C F D

A B C D E F G

F E A B D C

H A B C D E F G

H I J A B C D E F G

I J A B C D E F G H

J A B C D E F G H I

A B C D E F G H I J K

B C D E F G H I J K A

B C D E F G A D E F G H I A B C

7×7 C D E D E F E F G F G A G A B A B C B C D

9×9 E F F G G H H I I A A B B C C D D E

A B C D E F G H I

B C D E F G H I A

C D E F G H I A B

C D E F G H I J K A B

D E F G H I J K A B C

11 × 11 E F G F G H G H I H I J I J K J K A K A B A B C B C D C D E D E F

H I J K A B C D E F G

F G A B C D E

G A B C D E F

G H I A B C D E F

H I A B C D E F G

I A B C D E F G H

I J K A B C D E F G H

J K A B C D E F G H I

K A B C D E F G H I J

GRAECO–LATIN SQUARES

Two Latin squares K and L of order n are orthogonal if K(a, b) = K(c, d) and L(a, b) = L(c, d) implies a = c and b = d. Equivalently, all of the n2 pairs (Ki,j , Li,j ) are distinct. A pair of orthogonal Latin squares are called Graeco–Latin squares.There is a pair of orthogonal Latin squares of order n for all n > 1 except n = 2 or 6. A set of Latin squares L1 , . . . , Lm are mutually orthogonal if for every 1 ≤ i < j ≤ m, the Latin squares Li and Lj are orthogonal. Three mutually orthogonal Latin squares of size 4 are       A B C D 1 3 4 2 a d b c B A D C  2 4 3 1  b c a d       (13.1)  C D A B  3 1 2 4  c b d a D C B A 4 2 1 3 d a c b

c 2000 by Chapman & Hall/CRC 

3×3 A1 B3 C2 B2 C1 A3 C3 A2 B1

A1 B2 C3 D4

4×4 B3 C4 A4 D3 D1 A2 C2 B1

A1 B2 C3 D4 E5

D2 C1 B4 A3

B3 C4 D5 E1 A2

5×5 C5 D2 D1 E3 E2 A4 A3 B5 B4 C1

E4 A5 B1 C2 D3

There are no 6 × 6 Graeco–Latin squares A1 B2 C3 D4 E5 F6 G7

B5 C6 D7 E1 F2 G3 A4

7×7 C2 D6 E3 D3 E7 F4 E4 F1 G5 F5 G2 A6 G6 A3 B7 A7 B4 C1 B1 C5 D2

A1 G7 F6 E5 H10 J9 C8 I2 D3 B4

13.3

F7 G1 A2 B3 C4 D5 E6

G4 A5 B6 C7 D1 E2 F3

A1 B2 C3 D4 E5 F6 G7 H8 I9

B3 C1 A2 E6 F4 D5 H9 I7 G8

C2 A3 B1 F5 D6 E4 I8 G9 H7

B8 H2 G1 F7 E6 I10 D9 J3 C4 A5

C9 A8 I3 G2 F1 E7 J10 D4 B5 H6

D10 B9 H8 J4 G3 F2 E1 C5 A6 I7

D7 E8 F8 G1 H2 I3 A4 B5 C6

9×9 E 9 F8 F7 D9 D8 E7 H3 I 2 I1 G3 G2 H1 B6 A5 C4 A6 A5 B4

10 × 10 E 2 F4 C10 E3 A9 B10 I8 H9 D5 J8 G4 C6 F3 G5 B6 A7 H7 I 1 J1 D2

A1 B2 C3 D4 E5 F6 G7 H8

B5 A8 G4 F3 H1 D7 C6 E2

C2 G1 A7 E6 D8 H4 B3 F5

G4 H5 I6 A7 B8 C9 D1 E2 F3

H6 I4 G5 B9 C7 A8 E3 F1 D2

I5 G6 H4 C8 A9 B7 F2 D3 E1

H3 I4 J5 D6 C7 B1 A2 G8 F10 E9

I5 J6 D7 C1 B2 A3 H4 F9 E8 G10

G6 F5 E4 A10 I9 D8 B7 H1 J2 C3

8×8 D3 E7 F7 H3 E1 D2 A5 C8 C4 A6 B8 G5 H2 F1 G6 B4

F4 D6 H5 B1 G3 A2 E8 C7

G8 C5 B6 H7 F2 E3 A4 D1

H6 E4 F8 G2 B7 C1 D5 A3

J7 D1 C2 B3 A4 H5 I6 E10 G9 F8

BLOCK DESIGNS

A balanced incomplete block design (BIBD) is a pair (V, B) where V is a v-set and B is a collection of b subsets of V (each subset containing k elements) such that each element of V is contained in exactly r blocks and any 2-subset of V is contained in exactly λ blocks. The numbers v, b, r, k, λ are parameters of the BIBD.

c 2000 by Chapman & Hall/CRC 

The parameters are necessarily related by vr = bk and r(k − 1) = λ(v − 1). BIBDs are usually described by specifying (v, k, λ); this is a (v, k, λ)-design. vλ(v−1) vr From these values r = λ(v−1) k−1 and b = k = k(k−1) . The complement of a design for (V, B) is a design for (V, B) where B = (V \B | B ∈ B). The complement of a design with parameters (v, b, r, k, λ) is a design with parameters (v, b, b − r, v − k, b − 2r + λ). For this reason, tables are usually given for v ≥ 2k (the designs for v < 2k are then obtained by taking complements). The example designs given below are from Colbourn and Dinitz. To conserve space, designs are displayed in a k × b array in which each column contains the elements forming a block. 0 0 0 0 0 1 1 1 2 2

(a) The unique (6,3,2) design is 1 1 2 3 4 2 3 4 3 3 2 3 4 5 5 5 4 5 4 5

Here there are v = 6 elements (numbered 0, 1, . . . , 5), k = 3 elements per block, and each pair is in λ = 2 blocks. The other parameters are r = 5 and b = 10. As an illustration of how to interpret this design, note that the pair (0,1) appears in columns 1 and 2, and in no other columns. Note that the pair (0,2) appears in columns 1 and 3, and in no other columns, etc. (b) One of the 4 nonisomorphic (7,3,2) designs is 0 0 0 0 0 0 1 1 1 1 2 2 2 2 1 1 3 3 5 5 3 3 4 4 3 3 4 4 2 2 4 4 6 6 5 5 6 6 6 6 5 5

(c) One of the 10 nonisomorphic (7,3,3) designs is 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 2 2 2 2 2 2 1 1 1 3 3 3 5 5 5 3 3 3 4 4 4 3 3 3 4 4 4 2 2 2 4 4 4 6 6 6 5 5 5 6 6 6 6 6 6 5 5 5

(d) One of the 4 nonisomorphic (8,4,3) designs is 0 1 2 3

0 1 2 4

0 1 5 6

0 2 5 7

0 3 4 5

0 3 6 7

0 4 6 7

1 2 6 7

1 3 4 6

1 3 5 7

1 4 5 7

2 3 4 7

2 3 5 6

2 4 5 6 0 0 0 0 1 1 1 2 2 2 3 6

(e) The unique (9,3,1) design is 1 3 4 5 3 4 5 3 4 5 4 7 2 6 8 7 8 7 6 7 6 8 5 8

(f) One of the 36 nonisomorphic (9,3,2) designs is 0 0 0 0 0 0 0 0 1 1 1 1 1 1 2 2 2 2 2 2 3 3 4 4 1 1 3 3 5 5 7 7 3 3 4 4 6 6 3 3 4 4 5 5 6 6 5 5 2 2 4 4 6 6 8 8 5 5 7 7 8 8 8 8 6 6 7 7 7 7 8 8

c 2000 by Chapman & Hall/CRC 

(g) One of the 11 nonisomorphic (9,4,3) designs is 0 1 2 3

0 1 2 4

0 1 5 6

0 2 5 6

0 3 4 7

0 3 4 8

0 5 7 8

0 6 7 8

1 2 7 8

1 3 5 7

1 3 5 8

1 4 6 7

1 4 6 8

2 3 6 7

2 3 6 8

2 4 5 7

2 4 5 8

3 4 5 6

(h) One of the 3 nonisomorphic (10,4,2) designs is 0 1 2 3

0 1 4 5

0 2 4 6

0 3 7 8

0 5 7 9

0 6 8 9

1 2 7 8

1 3 6 9

1 4 7 9

1 5 6 8

2 3 5 9

2 4 8 9

2 5 6 7

3 4 5 8

3 4 6 7

0 1 (i) The unique (11,5,2) design is 2 3 7

0 1 4 5 6

0 2 5 8 9

0 3 6 8 a

0 4 7 9 a

1 1 4 8 a

1 3 5 9 a

1 6 7 8 9

2 3 4 6 9

2 5 6 7 a

3 4 5 7 8

0 1 (j) The unique (13,4,1) design is 3 9

0 2 8 c

0 4 5 7

0 6 a b

1 2 4 a

1 5 6 8

1 7 b c

2 3 5 b

2 6 7 9

3 4 6 c

3 7 8 a

13.4

4 8 9 b

5 9 a c

FACTORIAL EXPERIMENTATION: 2 FACTORS

If two factors A and B are to be investigated at a levels and b levels, respectively, and if there are ab experimental conditions (treatments) corresponding to all possible combinations of the levels of the two factors, the resulting experiment is called a complete a × b factorial experiment. Assume the entire set of ab experimental conditions are repeated r times. Let yijk be the observation in the k th replicate taken at the ith factor of A and the j th factor of B. The model has the form yijk = µ + αi + βj + (αβ)ij + ρk + 3ijk

(13.2)

for i = 1, 2, . . . , a, j = 1, 2, . . . , b, and k = 1, 2, . . . , r. Here • • • •

µ is the grand mean αi is the effect of the ith level of factor A βj is the effect of the j th level of factor B (αβ)ij is the interaction effect or joint effect of the ith level of factor A and the j th level of factor B • ρk is the effect of the k th replicate • 3ijk are independent normally distributed random variables with mean zero and variance σ 2 The following conditions are also imposed a  i=1

αi =

b 

βj =

j=1

c 2000 by Chapman & Hall/CRC 

a  i=1

(αβ)ij =

b  j=1

(αβ)ij =

r  k=1

ρk = 0

(13.3)

In the usual way SS(Tr)

SST

      a  a  b  r b     2 2 (yijk − y ... ) = r y ij. − y ... i=1 j=1 k=1

i=1 j=1 r 

+ ab

2

(y ..k − y ... )

(13.4)

k=1

+

a  b  r  

yijk − y ij. − y ..k + y ...

i=1 j=1 k=1





2

SSE

SST is the total sum of squares, SS(Tr) is the treatment sum of squares, SSE is the error sum of squares. The distinguishing feature of a factorial experiment is that the treatment sum of squares can be further subdivided into components corresponding to the various factorial effects. Here:  r

SS(Tr) a  b 



 2



(yij. − y ... ) = rb

i=1 j=1

SS(A) a 





 2

(y i.. − y ... ) + ra

i=1

+r 

b a    i=1 j=1

SS(B) b  



y .j. − y ...

 2

j=1

y ij. − y i.. − y .j. + y ... 

2

(13.5)



SS(AB)

SS(A) is the factor A sum of squares, SS(B) is the factor B sum of squares, SS(AB) is the interaction sum of squares. 13.5

2r FACTORIAL EXPERIMENTS

A factorial experiment in which there are r factors, each at only two levels, is a 2r factorial experiment. The two levels are often denoted as high and low, or 0 and 1. A complete 2r factorial experiment includes observations for every combination of factor and level, for a total of 2r observations. A 23 factorial experiment has 8 treatment combinations, and the model is given by Yijkl = µ + αi + βj + γk + (αβ)ij + (αγ)ik + (βγ)jk + (αβγ)ijk + 3ijkl

(13.6)

where i = 0, 1, j = 0, 1, k = 0, 1, and l = 1, 2, . . . , n. The assumptions are ind

3ij ∼ N(0, σ 2 ),

α1 = −α0 ,

β1 = −β0 ,

(αβ)10 = (αβ)01 = −(αβ)11 = −(αβ)00 , . . . c 2000 by Chapman & Hall/CRC 

γ1 = −γ0 , (13.7)

A 2n factorial experiment requires 2n experimental conditions. These conditions are listed in a standard order using a special notation. Factor A at the low level, or level 0, is denoted by “1”, at the high level, or level 1, by “a”. The levels of factor B are represented by “1” and “b”, etc. In a 23 factorial experiment, the treatment combination that consists of high levels of factors A and C, and a low level of factor B, is denoted by ac. The treatment combination of all low levels is denoted simply by 1. The treatment combinations are given by a binary expansion of the factor levels. For n = 2 the standard order of combinations is {1, a, b, ab}. For n = 3 the standard form is: Experimental condition 1 a b ab c ac bc abc

Level of factor A B C 0 0 0 1 0 0 0 1 0 1 1 0 0 0 1 1 0 1 0 1 1 1 1 1

The symbols for the first four experimental conditions are like those for a two–factor experiment, and the second four are obtained by multiplying each of the first four symbols by c. Define (1), (a), (b), (ab), (c), . . . , to be the treatment totals corresponding to the experimental conditions 1, a, b, ab, c, . . . . For example, in a 23 factorial experiment, (1) =

r 

y000A

A=1

(a) =

r 

y100A

A=1

... (bc) =

(13.8) r  A=1

y011A

(abc) =

r 

y111A

A=1

Certain linear combinations of these totals result in the sum of squares for the main effects and interaction effects. Define the effect total for factor A: [A] = −(1) + (a) − (b) + (ab) − (c) + (ac) − (bc) + (abc).

(13.9)

The sum of squares due to each factor may be obtained from its effect total. [1] is the total effect. The linear combination for each experimental condition effect total may be presented in a table of signs (a larger table is on page 338):

c 2000 by Chapman & Hall/CRC 

Experimental condition 1 a b ab c ac bc abc

[1] + + + + + + + +

[A] − + − + − + − +

[B] − − + + − − + +

Effect total [AB] [C] [AC] + − + − − − − − + + − − + + − − + + − + − + + +

[BC] + + − − − − + +

[ABC] − + + − + − − +

To compute the sum of squares: SSA = [A]2 /(8r), SSB = [B]2 /(8r), SSC = [C]2 /(8r), SS(AB) = [AB]2 /(8r), . . . . Note: The expression for a single effect total may be found by expanding an algebraic expression. Consider the effect total [AB] in a 23 factorial experiment. Take the expression (a ± 1)(b ± 1)(c ± 1) and use a “−” if the corresponding letter appears in the symbol for the main effect, and use a “+” if the letter does not appear. Expand the expression and add parentheses. For [AB] the calculation is given by (a − 1)(b − 1)(c + 1) = abc − ac − bc + c + ab − a − b + 1 = (1) − (a) − (b) + (ab) + (c) − (ac) − (bc) + (abc) 13.6

(13.10)

CONFOUNDING IN 2N FACTORIAL EXPERIMENTS

Sometimes it is impossible to run all the required experiments in a single block. When experimental conditions are distributed over several blocks, one or more of the effects may become confounded (i.e., inseparable) with possible block effects, that is, between-block differences. For example, in a 23 factorial experiment, let 1, b, c, and bc be in one block, and let a, ab, ac, and abc be in another block. The “block effect”, the difference between the two block totals, is given by [(a) + (ab) + (ac) + (abc)] − [(1) + (b) + (c) + (bc)] This happens to be equal to [A]. Hence, the main effect of A is confounded with blocks. Using instead a block of a, b, c, abc and a block of 1, ab, ac, bc would result in ABC being confounded with blocks. If the number of blocks is 2p then a total of 2p − 1 effects are confounded by blocks. 13.7

TABLES FOR DESIGN OF EXPERIMENTS

The following tables present combinatorial patterns that may be used as experimental designs. The plans and plan numbers are from a more numerous c 2000 by Chapman & Hall/CRC 

set of patterns in W. G. Cochran and G. M. Cox, Experimental Designs, Second Edition, John Wiley & Sons, Inc, New York, 1957. Reprinted by permission of John Wiley & Sons, Inc. Plans 6.1–6.6 are plans of factorial experiments confounded in randomized incomplete blocks. Plans 6A.1–6A.6 are plans of 2n factorials in fractional replication. Plans 13.1–13.5 are plans of incomplete block designs. 13.7.1 Plans of factorial experiments confounded in randomized incomplete blocks Plan 6.1: 23 factorial, 4 unit blocks Rep. I, ABC confounded abc a b c

ab ac bc (1)

Plan 6.2: 24 factorial, 8 unit blocks Rep. I, ABCD confounded a b c d abc abd acd bcd

(1) ab ac bc ad bd cd abcd

Plan 6.3: 26 factorial, 16 unit blocks Rep. I, ABCD, ABEF, CDEF confounded a b acd bcd ce de abce abde cf df abcf abdf aef bef acdef bcdef

c 2000 by Chapman & Hall/CRC 

c d abc abd ae be acde bcde af bf acdf bcdf cef def abcef abdef

ab cd (1) abcd ace ade bce bde acf adf bcf bdf abef cdef ef abcdef

ac ad bc bd abe cde e abcde abf cdf f abcdf acef adef bcef bdef

Plan 6.4: Balanced group of sets for 24 factorial, 4 unit blocks Two–factor interactions are confounded in 1 replication and three–factor interactions are confounded in 3 replications. The columns are the blocks. Rep. I, AB, ACD, BCD confounded (1) abc abd cd

ab c d abcd

a bc bd acd

Rep. II, AC, ABD, BCD (1) abc acd bd

b ac ad bcd

Rep. IV, BC, ABD, ACD (1) abc bcd ad

bc a d abcd

b ac cd abd

ac b d abcd

a bc cd abd

Rep. III, AD, ABC, BCD (1) abd acd bc

c ab ad bcd

Rep. V, BD, ABC, ACD (1) abd bcd ac

c ab bd acd

bd a c abcd

b ad cd abc

ad b c abcd

a bd cd abc

d ab ac bcd

Rep. VI, CD, ABC, ABD (1) acd bcd ab

d ab bc acd

cd a b abcd

c ad bd abc

d ac bc abd

Plan 6.5: Balanced group of sets for 25 factorial, 8 unit blocks Three– and four–factor interactions are confounded in 1 replication. Rep. I, ABC, ADE, BCDE confounded (1) bc abd acd abe ace de bcde

ab ac d bcd e bce abde acde

a abc bd be ce ade abcde cd

ac ce b abe d ade abcd bcde

a e bc abce cd acde abd bde

(1) ad abc bcd abe bde ce acde

b c ad abcd ae abce bde cde

Rep. III, ACE, BCD, ABDE (1) ae abc bce acd cde bd abde

Rep. II, ABD, BCE, ACDE ab bd c acd e ade abce bcde

(1) ac abd bcd ade cde be abce

ad cd b abc e ace abde bcde

Rep. V, ABE, CDE, ABCD

c 2000 by Chapman & Hall/CRC 

ae be c abc d abd acde bcde

b abd ac cd ae de bce abcde

Rep. IV, ACD, BDE, ABCE

c ace ab be ad de bcd abcde

(1) ab ace bce ade bde cd abcd

a d bc abcd be abde ace cde

a b ce abce de abde acd bcd

e abe ac bc ad bd cde abcde

a c bd abcd de acde abe bce

d acd ab bc ae ce bde abcde

Plan 6.6: Balanced group of sets for 26 factorial, 8 unit blocks All three– and four–factor interactions are confounded in 2 replications. Rep. I, ABC, CDE, ADF, BEF, ABDE, BCDF, ACEF confounded abc bd ae cde cf adf bef abcdef

a cd abce bde bf abcdf cef adef

b abcd ce ade af cdf abcef bdef

(1) acd bce abde abf bcdf acef def

bc abd e acde acf df abef bcdef

ac d abe bcde bcf abdf ef acdef

c ad be abcde abcf bdf aef cdef

ab bcd ace de f acdf bcef abdef

Rep. II, ABD, DEF, BCF, ACE, ABEF, ACDF, BCDE abd cd be ace af bcf def abcdef

b ac abde cde df abcdf aef bcef

a bc de abcde abdf cdf bef acef

(1) abc ade bcde bdf acdf abef cef

ad bcd e abce abf cf bdef acdef

bd acd abe ce f abcf adef bcdef

d abcd ae bce bf acf abdef cdef

ab c bde acde adf bcdf ef abcef

Rep. III, ABE, BDF, ACD, CEF, ADEF, BCDE, ABCF bc acd abe de af bdf cef abcdef

a bd cd abcde bcf acdf abef def

ac bcd d abde bf adf abcef cdef

(1) abd ace bcde abcf cdf bef adef

abc cd be ade f abdf acef bcdef

ab d bce acde cf abcdf aef bdef

b ad abce cde acf bcdf ef abdef

c abcd ae bde abf df bcef acdef

Rep. IV, ABF, CDF, ADE, BCE, ABCD, BDEF, ACEF ac bd bce ade abf cdf ef abcdef

a bcd be acde abcf df cef abdef

b acd ae bcde cf abdf abcef def

(1) abcd abe cde bcf adf acef bdef

c abd abce de bf acdf aef bcdef

abc d ce abde af bcdf bef acdef

bc ad ace bde f abcdf abef cdef

ab cd e abcde acf bdf bcef adef

Rep. V, ACF, BCD, ADE, BEF, ABDF, CDEF, ABCE ab bcd ce ade acf df bef abcdef

a cd bce abde abcf bdf ef acdef

bc abd ae cde f acdf abcef bdef

c 2000 by Chapman & Hall/CRC 

(1) acd abce bde bcf abdf aef cdef

b abcd ace de cf adf abef bcdef

ac d be abcde abf bcdf cef adef

c ad abe bcde bf abcdf acef def

abc bd e acde af cdf bcef abdef

Note: (1) Replication VI, ABC, BDE, ADF, CEF, ACDE, BCDF, ABEF . Interchange B and C in replication I. (2) Replication VII, ABF, DEF, BCD, ACE, ABDE, ACDF, BCEF . Interchange F and D in replication II. (3) Replication VIII, ABE, BDF, CDE, ACF, ADEF, ABCD, BCEF . Interchange A and E in replication III. (4) Replication IX, ABD, CDF, AEF, BCE, ABCF, BDEF, ACDE. Interchange F and D in replication IV. (5) Replication X, AEF, BDE, ACD, BCF, ABDF, CDEF, ABCE. Interchange E and C in replication V.

13.7.2

Plans of 2n factorials in fractional replication

Plan 6A.1: 24 factorial in 8 units ( 1/2 replicate) Defining contrast: ABCD Estimable 2-factor interactions: AB = CD, AC = BD, AD = BC (1) ab ac ad bc bd cd abcd

Effect Main 2-factor Total

df 4 3 7

Plan 6A.2: 25 factorial in 8 units ( 1/4 replicate) Defining contrast: ABE, CDE, ABCD Main effects have 2-factors as aliases. The only estimatable 2-factors are AC = BD and AD = BC. (1) ab cd ace bce ade bde abcd

Effect Main 2-factor Total

df 5 2 7

Plan 6A.3: 25 factorial in 16 units ( 1/2 replicate) Defining contrast: ABCDE 1. Blocks of 4 units Estimatable 2-factors: All except CD, CE, DE (confounded with blocks) Blocks

(1) (2) (1) ac ab bc acde de bcde abde CD, CE, DE

c 2000 by Chapman & Hall/CRC 

(3) (4) ae ad be bd cd ce abcd abce confounded

Effect Block Main 2-factor Total

df 3 5 7 15

2. Blocks of 8 units (a) Estimatable 2-factors: All except DE (b) Combine blocks 1 and 2; and blocks 3 and 4. DE confounded.

Effect Block Main 2-factor Total

df 1 5 9 15

Effect Main 2-factor Total

df 5 10 15

3. Blocks of 16 units (a) Estimatable 2-factors: All (b) Combine blocks 1–4

Plan 6A.4: 26 factorial in 8 units ( 1/8 replicate) Defining contrasts: ACE, ADF , BCF , BDE, ABCD, ABEF , CDEF Main effects have 2-factors as aliases. The only estimable 2-factor is the set AB = CD = EF (1) acf ade bce bdf abcd abef cdef

Effect Main 2-factor (AB = CD = EF ) Total

df 6 1 7

Plan 6A.5: 26 factorial in 16 units ( 1/4 replicate) Defining contrasts: ABCE, ABDF , CDEF 1. Blocks of 4 units Estimatable 2-factors: The alias sets AC = BD, AD = BF , AE = BC, AF = BD, CD = EF , CF = DE Blocks

(1) (2) (3) (4) (1) acd ab acf abce aef ce ade abdf bcf df bcd cdef bde abcdef bef AB, ACF , BCF confounded

Effect Block Main 2-factor Total

2. Blocks of 8 units (a) Estimatable 2-factors: Same as in blocks of 4 units, plus the set AB = CE = DF . (b) Combine blocks 1 and 2; and blocks 3 and 4. ACF confounded.

Effect Block Main 2-factor 3-factor Total

df 1 6 7 1 15

Effect Main 2-factor 3-factor Total

df 6 7 2 15

3. Blocks of 16 units (a) Estimatable 2-factors: Same as in blocks of 8 units (b) Combine blocks 1–4 c 2000 by Chapman & Hall/CRC 

df 3 6 6 15

Plan 6A.6: 26 factorial in 32 units ( 1/2 replicate) Defining contrast: ABCDEF 1. Blocks of 4 units Estimatable 2-factors: All except AE, BF , and CD (confounded with blocks) Blocks

(1) (2) (1) ab abef ef acde acdf bcdf bcde AE, BF , CD,

(3) (4) (5) (6) ac bc ae af de df bf be abdf acef cd abcd bcef abde abcdef cdef ABD, ACF , ADF confounded Effect df Block 7 Main 6 2-factor 12 Higher order 6 Total 31

(7) ad ce abcf bdef

2. Blocks of 8 units Effect Block Main 2-factor Higher order Total

(a) Estimatable 2-factors: All except CD (b) Combine blocks 1 and 2; blocks 3 and 4; blocks 5 and 6; and blocks 7 and 8; CD, ABC, ABD confounded.

df 3 6 14 8 31

3. Blocks of 16 units (a) Estimatable 2-factors: All (b) Estimatable 3-factors: ABC = DEF is lost by confounding. The others are in alias pairs, e.g., ABD = CEF . (c) Combine blocks 1–4; and blocks 5–8. ABC confounded.

Effect Block Main 2-factor 3-factor Total

df 1 6 15 9 31

Effect Main 2-factor 3-factor Total

df 6 15 10 31

4. Blocks of 32 units (a) Estimatable 2-factors: All (b) Estimatable 3-factors: These are arranged in 10 alias pairs. (c) Combine blocks 1–8.

13.7.3

Plans of incomplete block designs

Plan 13.1: t = 7, k = 3, r = 3, b = 7, λ = 1 E = .78, Type II Block (1) (2) (3) (4)

I 7 1 2 3

c 2000 by Chapman & Hall/CRC 

Reps. II III 1 3 2 4 3 5 4 6

Block (5) (6) (7)

I 4 5 6

Reps. II III 5 7 6 1 7 2

(8) bd cf abce adef

Plan 13.2: t = 7, k = 4, r = 4, b = 7, λ = 2 E = .88, Type II Block (1) (2) (3) (4)

I 3 4 5 6

Reps. II III 5 6 6 7 7 1 1 2

IV 7 1 2 3

Block (5) (6) (7)

I 7 1 2

Reps. II III 2 3 3 4 4 5

IV 4 5 6

Plan 13.3: t = 11, k = 5, r = 5, b = 11, λ = 2 E = .88, Type I Block (1) (2) (3) (4) (5) (6)

I 1 7 9 11 10 8

II 2 1 8 9 11 7

Reps. III IV 3 4 6 10 1 6 7 1 5 8 2 3

V 5 3 2 4 1 11

Block (7) (8) (9) (10) (11)

I 2 6 3 5 4

II 6 3 4 10 5

Reps. III 4 11 10 9 8

IV 11 5 9 2 7

V 10 9 8 7 6

Plan 13.4: t = 11, k = 6, r = 6, b = 11, λ = 3 E = .92, Type I Block (1) (2) (3) (4) (5) (6)

I 6 5 4 3 2 1

II 7 8 5 10 3 6

Reps. III IV 8 9 4 11 7 3 2 6 9 7 10 4

V 10 2 11 5 4 9

VI 11 9 10 8 6 5

Block (7) (8) (9) (10) (11)

I 9 8 7 11 10

II 1 2 11 4 9

Reps. III IV 3 5 1 10 5 1 6 8 11 2

Plan 13.5: t = 13, k = 4, r = 4, b = 13, λ = 1 E = .81, Type I Block (1) (2) (3) (4) (5) (6) (7)

I 13 1 2 3 4 5 6

Reps. II III 1 3 2 4 3 5 4 6 5 7 6 8 7 9

c 2000 by Chapman & Hall/CRC 

IV 9 10 11 12 13 1 2

Block (8) (9) (10) (11) (12) (13)

I 7 8 9 10 11 12

Reps. II III 8 10 9 11 10 12 11 13 12 1 13 2

IV 3 4 5 6 7 8

V 8 7 6 1 3

VI 7 4 2 3 1

13.7.4

Main effect and interactions in factorial designs (T) A AB C AC BC ABC D AD BD ABD CD ACD BCD ABCD E AE BE ABE CE ACE BCE ABCE DE ADE BDE ABDE CDE ACDE BCDE ABCDE

Main effect and interactions in 22 , 23 , 24 , 25 , and 26 factorial designs

(1) a b ab c ac bc abc d ad bd abd cd acd bcd abcd e ae be abe ce ace bce abce de ade bde abde cde acde bcde abcde

+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +

−−+−++−−++−+−−+−++−+−−++−−+−++− +−−−−++−−++++−−−−++++−−++−−−−++ −+−−+−+−+−++−+−−+−++−+−+−+−−+−+ +++−−−−−−−−++++−−−−++++++++−−−− −−++−−+−++−−++−−++−−++−+−−++−−+ +−−++−−−−++−−++−−++−−++++−−++−− −+−+−+−−+−+−+−+−+−+−+−++−+−+−+− +++++++−−−−−−−−−−−−−−−−++++++++ −−+−++−+−−+−++−−++−+−−+−++−+−−+ +−−−−++++−−−−++−−++++−−−−++++−− −+−−+−++−+−−+−+−+−++−+−−+−++−+− +++−−−−++++−−−−−−−−++++−−−−++++ −−++−−++−−++−−+−++−−++−−++−−++− +−−++−−++−−++−−−−++−−++−−++−−++ −+−+−+−+−+−+−+−−+−+−+−+−+−+−+−+ +++++++++++++++−−−−−−−−−−−−−−−− −−+−++−−++−+−−++−−+−++−−++−+−−+ +−−−−++−−++++−−++−−−−++−−++++−− −+−−+−+−+−++−+−+−+−−+−+−+−++−+− +++−−−−−−−−++++++++−−−−−−−−++++ −−++−−+−++−−++−+−−++−−+−++−−++− +−−++−−−−++−−++++−−++−−−−++−−++ −+−+−+−−+−+−+−++−+−+−+−−+−+−+−+ +++++++−−−−−−−−++++++++−−−−−−−− −−+−++−+−−+−++−+−−+−++−+−−+−++− +−−−−++++−−−−++++−−−−++++−−−−++ −+−−+−++−+−−+−++−+−−+−++−+−−+−+ +++−−−−++++−−−−++++−−−−++++−−−− −−++−−++−−++−−++−−++−−++−−++−−+ +−−++−−++−−++−−++−−++−−++−−++−− −+−+−+−+−+−+−+−+−+−+−+−+−+−+−+− +++++++++++++++++++++++++++++++

c 2000 by Chapman & Hall/CRC 

F AF BF ABF CF ACF BCF ABCF DF ADF BDF ABDF CDF ACDF BCDF ABCDF EF AEF BEF ABEF CEF ACEF BCEF ABCEF DEF ADEF BDEF ABDEF CDEF ACDEF BCDEF ABCDEF

Main effect and interactions in 22 , 23 , 24 , 25 , and 26 factorial designs

(1) −++−+−−++−−+−++−+−−+−++−−++−+−−+ a −−++++−−++−−−−++++−−−−++−−++++−− b −+−++−+−+−+−−+−++−+−−+−+−+−++−+− ab −−−−++++++++−−−−++++−−−−−−−−++++ c −++−−++−+−−++−−++−−++−−+−++−−++− ac −−++−−++++−−++−−++−−++−−−−++−−++ bc −+−+−+−++−+−+−+−+−+−+−+−−+−+−+−+ abc − − − − − − − − + + + + + + + + + + + + + + + + − − − − − − − − d −++−+−−+−++−+−−++−−+−++−+−−+−++− ad −−++++−−−−++++−−++−−−−++++−−−−++ bd −+−++−+−−+−++−+−+−+−−+−++−+−−+−+ abd − − − − + + + + − − − − + + + + + + + + − − − − + + + + − − − − cd −++−−++−−++−−++−+−−++−−++−−++−−+ acd − − + + − − + + − − + + − − + + + + − − + + − − + + − − + + − − bcd − + − + − + − + − + − + − + − + + − + − + − + − + − + − + − + − abcd − − − − − − − − − − − − − − − − + + + + + + + + + + + + + + + + e −++−+−−++−−+−++−−++−+−−++−−+−++− ae −−++++−−++−−−−++−−++++−−++−−−−++ be −+−++−+−+−+−−+−+−+−++−+−+−+−−+−+ abe − − − − + + + + + + + + − − − − − − − − + + + + + + + + − − − − ce −++−−++−+−−++−−+−++−−++−+−−++−−+ ace − − + + − − + + + + − − + + − − − − + + − − + + + + − − + + − − bce − + − + − + − + + − + − + − + − − + − + − + − + + − + − + − + − abce − − − − − − − − + + + + + + + + − − − − − − − − + + + + + + + + de −++−+−−+−++−+−−+−++−+−−+−++−+−−+ ade − − + + + + − − − − + + + + − − − − + + + + − − − − + + + + − − bde − + − + + − + − − + − + + − + − − + − + + − + − − + − + + − + − abde − − − − + + + + − − − − + + + + − − − − + + + + − − − − + + + + cde − + + − − + + − − + + − − + + − − + + − − + + − − + + − − + + − acde − − + + − − + + − − + + − − + + − − + + − − + + − − + + − − + + bcde − + − + − + − + − + − + − + − + − + − + − + − + − + − + − + − + abcde − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − − −

c 2000 by Chapman & Hall/CRC 

(T) A B AB C AC BC ABC D AD BD ABD CD ACD BCD ABCD E AE BE ABE CE ACE BCE ABCE DE ADE BDE ABDE CDE ACDE BCDE ABCDE

Main effect and interactions in 22 , 23 , 24 , 25 , and 26 factorial designs

f af bf abf cf acf bcf abcf df adf bdf abdf cdf acdf bcdf abcdf ef aef bef abef cef acef bcef abcef def adef bdef abdef cdef acdef bcdef abcdef

+ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +

−−+−++−−++−+−−+−++−+−−++−−+−++− +−−−−++−−++++−−−−++++−−++−−−−++ −+−−+−+−+−++−+−−+−++−+−+−+−−+−+ +++−−−−−−−−++++−−−−++++++++−−−− −−++−−+−++−−++−−++−−++−+−−++−−+ +−−++−−−−++−−++−−++−−++++−−++−− −+−+−+−−+−+−+−+−+−+−+−++−+−+−+− +++++++−−−−−−−−−−−−−−−−++++++++ −−+−++−+−−+−++−−++−+−−+−++−+−−+ +−−−−++++−−−−++−−++++−−−−++++−− −+−−+−++−+−−+−+−+−++−+−−+−++−+− +++−−−−++++−−−−−−−−++++−−−−++++ −−++−−++−−++−−+−++−−++−−++−−++− +−−++−−++−−++−−−−++−−++−−++−−++ −+−+−+−+−+−+−+−−+−+−+−+−+−+−+−+ +++++++++++++++−−−−−−−−−−−−−−−− −−+−++−−++−+−−++−−+−++−−++−+−−+ +−−−−++−−++++−−++−−−−++−−++++−− −+−−+−+−+−++−+−+−+−−+−+−+−++−+− +++−−−−−−−−++++++++−−−−−−−−++++ −−++−−+−++−−++−+−−++−−+−++−−++− +−−++−−−−++−−++++−−++−−−−++−−++ −+−+−+−−+−+−+−++−+−+−+−−+−+−+−+ +++++++−−−−−−−−++++++++−−−−−−−− −−+−++−+−−+−++−+−−+−++−+−−+−++− +−−−−++++−−−−++++−−−−++++−−−−++ −+−−+−++−+−−+−++−+−−+−++−+−−+−+ +++−−−−++++−−−−++++−−−−++++−−−− −−++−−++−−++−−++−−++−−++−−++−−+ +−−++−−++−−++−−++−−++−−++−−++−− −+−+−+−+−+−+−+−+−+−+−+−+−+−+−+− +++++++++++++++++++++++++++++++

c 2000 by Chapman & Hall/CRC 

F AF BF ABF CF ACF BCF ABCF DF ADF BDF ABDF CDF ACDF BCDF ABCDF EF AEF BEF ABEF CEF ACEF BCEF ABCEF DEF ADEF BDEF ABDEF CDEF ACDEF BCDEF ABCDEF

Main effect and interactions in 22 , 23 , 24 , 25 , and 26 factorial designs

f +−−+−++−−++−+−−+−++−+−−++−−+−++− af ++−−−−++−−++++−−−−++++−−++−−−−++ bf +−+−−+−+−+−++−+−−+−++−+−+−+−−+−+ abf ++++−−−−−−−−++++−−−−++++++++−−−− cf +−−++−−+−++−−++−−++−−++−+−−++−−+ acf ++−−++−−−−++−−++−−++−−++++−−++−− bcf +−+−+−+−−+−+−+−+−+−+−+−++−+−+−+− abcf + + + + + + + + − − − − − − − − − − − − − − − − + + + + + + + + df +−−+−++−+−−+−++−−++−+−−+−++−+−−+ adf ++−−−−++++−−−−++−−++++−−−−++++−− bdf +−+−−+−++−+−−+−+−+−++−+−−+−++−+− abdf + + + + − − − − + + + + − − − − − − − − + + + + − − − − + + + + cdf +−−++−−++−−++−−+−++−−++−−++−−++− acdf + + − − + + − − + + − − + + − − − − + + − − + + − − + + − − + + bcdf + − + − + − + − + − + − + − + − − + − + − + − + − + − + − + − + abcdf + + + + + + + + + + + + + + + + − − − − − − − − − − − − − − − − ef +−−+−++−−++−+−−++−−+−++−−++−+−−+ aef ++−−−−++−−++++−−++−−−−++−−++++−− bef +−+−−+−+−+−++−+−+−+−−+−+−+−++−+− abef + + + + − − − − − − − − + + + + + + + + − − − − − − − − + + + + cef +−−++−−+−++−−++−+−−++−−+−++−−++− acef + + − − + + − − − − + + − − + + + + − − + + − − − − + + − − + + bcef + − + − + − + − − + − + − + − + + − + − + − + − − + − + − + − + abcef + + + + + + + + − − − − − − − − + + + + + + + + − − − − − − − − def +−−+−++−+−−+−++−+−−+−++−+−−+−++− adef + + − − − − + + + + − − − − + + + + − − − − + + + + − − − − + + bdef + − + − − + − + + − + − − + − + + − + − − + − + + − + − − + − + abdef + + + + − − − − + + + + − − − − + + + + − − − − + + + + − − − − cdef + − − + + − − + + − − + + − − + + − − + + − − + + − − + + − − + acdef + + − − + + − − + + − − + + − − + + − − + + − − + + − − + + − − bcdef + − + − + − + − + − + − + − + − + − + − + − + − + − + − + − + − abcdef + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + +

13.8

REFERENCES

1. W. G. Cochran and G. M. Cox, Experimental Designs, Second Edition, John Wiley & Sons, Inc, New York, 1957. 2. C. J. Colbourn and J. H. Dinitz, CRC Handbook of Combinatorial Designs, CRC Press, Boca Raton, FL, 1996, pages 578–581.

c 2000 by Chapman & Hall/CRC 

CHAPTER 14

Nonparametric Statistics Contents 14.1 14.2

Friedman test for randomized block design Kendall’s rank correlation coefficient 14.2.1 Kendall rank correlation coefficient table 14.3 Kolmogorov–Smirnoff tests 14.3.1 One-sample Kolmogorov–Smirnoff test 14.3.2 Two-sample Kolmogorov–Smirnoff test 14.3.3 Tables for Kolmogorov–Smirnoff tests 14.4 Kruskal–Wallis test 14.4.1 Tables for Kruskal–Wallis test 14.5 The runs test 14.5.1 Tables for the runs test 14.6 The sign test 14.6.1 Critical values for the sign test 14.7 Spearman’s rank correlation coefficient 14.7.1 Tables for Spearman’s rank correlation 14.8 Wilcoxon matched-pairs signed-ranks test 14.9 Wilcoxon rank–sum (Mann–Whitney) test 14.9.1 Tables for Wilcoxon U statistic 14.9.2 Critical values for Wilcoxon U statistic 14.10 Wilcoxon signed-rank test

Nonparametric, or distribution–free, statistical procedures generally assume very little about the underlying population(s). The test statistic used in each procedure is usually easy to compute and may involve qualitative measurements or measurements made on an ordinal scale. If both a parametric and nonparametric test are applicable, the nonparametric test is less efficient because it does not utilize all of the information in the sample. A larger sample size is required in order for the nonparametric test to have the same probability of a type II error.

c 2000 by Chapman & Hall/CRC 

14.1

FRIEDMAN TEST FOR RANDOMIZED BLOCK DESIGN

Assumptions: Let there be k independent random samples (treatments) from continuous distributions and b blocks. Hypothesis test: H0 : the k samples are from identical populations. Ha : at least two of the populations differ in location. Rank each observation from 1 (smallest) to k (largest) within each block. Equal observations are assigned the mean rank for their positions. Let Ri be the rank sum of the ith sample (treatment).

k  12 2 TS: Fr = R − 3b(k + 1) bk(k + 1) i=1 i RR: Fr ≥ χ2α,k−1 14.2

KENDALL’S RANK CORRELATION COEFFICIENT

Given  two sets containing ranked elements of the same size, consider each of the n2 = n(n−1) pairs of elements from within each set. Associate with each 2 pair (a) a score of +1 if the relative ranking of both samples is the same, or (b) a score of −1 if the relative  rankings are different. Kendall’s score, St , is defined as the total of these n2 individual scores. St will have a maximum value of n(n−1) if the two rankings are identical and a minimum value of 2 n(n−1) − 2 if the sets are ranked in exactly the opposite order. Kendall’s Tau is defined as I  n(n − 1)  τ = St (14.1) 2 and has the range −1 ≤ τ ≤ 1. The table on page 345 may be used to determine the exact probability associated with an occurrence (one-tailed) of a specific value of St . In this case the null hypothesis is the existence of an association between the two sets as extreme as an observed St . The tabled value is the probability that St is equaled or exceeded. Consider the sets of ranked elements: a = {4, 12, 6, 10} and b = {8, 7, 16, 2}. Kendall’s score is St = 0 for these sets since

Example 14.68 :

c 2000 by Chapman & Hall/CRC 

For For For For For For

the the the the the the

(1, 2) (1, 3) (1, 4) (2, 3) (2, 4) (3, 4)

term term term term term term

(with (with (with (with (with (with

a1 a1 a1 a2 a2 a3

< a2 < a3 < a4 > a3 > a4 < a4

and and and and and and

b1 b1 b1 b2 b2 b3

> b2 ) < b3 ) > b4 ) > b3 ) > b4 ) > b4 )

score score score score score score Total

is is is is is is is

−1 +1 −1 +1 +1 −1 0

Using the table on page 345 with n = 4 we find Prob [St ≥ 0] = .625. That is, an St value of 0 or larger would be expected 62.5% of the time.

14.2.1

Tables for Kendall rank correlation coefficient

The following table may be used to determine the exact probability associated with an occurrence (one-tailed) of a specific value of St . In this case the null hypothesis is the existence of an association between the two sets as extreme as an observed St . The tabled value is the probability that St is equaled or exceeded. Distribution of Kendall’s rank correlation coefficient in random rankings St 0 2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38

n=3 0.5000 0.1667

4 0.6250 0.3750 0.1667 0.0417

5 0.5917 0.4083 0.2417 0.1167 0.0417 0.0083

6 0.5000 0.3597 0.2347 0.1361 0.0681 0.0278 0.0083 0.0014

7 0.5000 0.3863 0.2810 0.1907 0.1194 0.0681 0.0345 0.0151 0.0054 0.0014 0.0002

8 0.5476 0.4524 0.3598 0.2742 0.1994 0.1375 0.0894 0.0543 0.0305 0.0156 0.0071 0.0028 0.0009 0.0002

9 0.5403 0.4597 0.3807 0.3061 0.2384 0.1792 0.1298 0.0901 0.0597 0.0376 0.0223 0.0124 0.0063 0.0029 0.0012 0.0004 0.0001

10 0.5000 0.4309 0.3637 0.3003 0.2422 0.1904 0.1456 0.1082 0.0779 0.0542 0.0363 0.0233 0.0143 0.0083 0.0046 0.0023 0.0011 0.0005 0.0002 0.0001

Note that each distribution is symmetric about St = 0: e.g., for n = 4, Prob [St = 2] = Prob [St = −2] = 0.375. Note also that St can only assume values with the same parity as n (for example, if n is even then Prob [St = odd] = 0); e.g., for n = 4, Prob [St = ±1] = Prob [St = ±3] = Prob [St = ±5] = 0.

c 2000 by Chapman & Hall/CRC 

14.3

KOLMOGOROV–SMIRNOFF TESTS

A one-sample Kolmogorov–Smirnoff test is used to compare an observed cumulative distribution function (computed from a sample) to a specific continuous distribution function. This is a special test of goodness of fit. A two-sample Kolmogorov–Smirnoff test is used to compare two observed cumulative distribution functions; the null hypothesis is that the two independent samples come from identical continuous distributions. 14.3.1

One-sample Kolmogorov–Smirnoff test

Suppose a sample of size n is drawn from a population with known cumulative distribution function F (x). The empirical distribution function, Fn (x), is defined by the sample and is a step function given by k when x(i) ≤ x < x(i+1) (14.2) n where k is the number of observations less than or equal to x and the {x(i) } are the order statistics. If the sample is drawn from the hypothesized distribution, then the empirical distribution function, Fn (x), should be close to F (x). Define the maximum absolute difference between the two distributions to be $ $ $ $ $ D = max $Fn (x) − F (x)$$ (14.3) Fn (x) =

For a two-tailed test the table on page 348 gives critical values for the sampling distribution of D under the null hypothesis. One should reject the hypothetical distribution F (x) if the value D exceeds the tabulated value. A corresponding one-tailed test is provided by the statistic   D+ = max Fn (x) − F (x)

(14.4)

Example 14.69 : The values {.5, .75, .9, .1} are observed from data that are presumed to be uniformly distributed on the interval (0, 1). Since the presumed distribution is uniform, we have F (x) = x. Figure 14.1 shows Fn (x) and F (x). To determine D, only the values of |Fn (x) − F (x)| for x at the endpoints (x = 0 and x = 1) and on each side of the sample values (since Fn (x) has discontinuities at the sample values) need to be considered. Constructing Table 14.1 results in D = .25. If α = .05 and n = 4 the table on page 348 yields a critical value of c = .624. Since D < c, the null hypothesis is not rejected.

14.3.2

Two-sample Kolmogorov–Smirnoff test

Suppose two independent samples of sizes n1 and n2 are drawn from a population with cumulative distribution function F (x). For each sample j an

c 2000 by Chapman & Hall/CRC 

x x=0 x = .1− x = .1+ x = .5− x = .5+ x = .75− x = .75+ x = .90− x = .90+ x=1

Fn (x) 0 0 .25 .25 .50 .50 .75 .75 .90 1

F (x) = x 0 .10 .10 .50 .50 .75 .75 .90 .90 1

|Fn (x) − F (x)| |0 − 0| = 0 |0 − .10| = .10 |.25 − .10| = .15 |.25 − .50| = .25 |.50 − .50| = 0 |.50 − .75| = .25 |.75 − .75| = 0 |.75 − .90| = .15 |.90 − .90| = 0 |1 − 1| = 0

Table 14.1: Table for Kolmogorov–Smirnoff computation. 1 0.8 0.6 0.4 0.2 0 0

0.2

0.4

0.6

0.8

1

Figure 14.1: Comparison of the sample distribution function (dotted curve) with the distribution function (solid line) for a uniform random variable on the interval (0, 1). empirical distribution function Fnj (x) is given by the step function, k j j ≤ x < xsample (14.5) when xsample (i) (i+1) n j where k is the number of observations less than or equal to x and the {xsample } (i) th are the order statistics for the j sample (for j = 1 or j = 2). If the two samples have been drawn from the same population, or from populations with the same distribution (the null hypothesis), then Fn1 (x) should be close to Fn2 (x). Define the maximum absolute difference between the two empirical distributions to be $ $ $ $ $ D = max $Fn1 (x) − Fn2 (x)$$ (14.6) Fnj (x) =

c 2000 by Chapman & Hall/CRC 

For a two-tailed test the table on page 350 gives critical values for the sampling distribution of D under the null hypothesis. The null hypothesis is rejected if the value of D exceeds the tabulated value. A corresponding one-tailed test is provided by the statistic   D+ = max Fn1 (x) − Fn2 (x) (14.7)

14.3.3

Tables for Kolmogorov–Smirnoff tests

14.3.3.1

Critical values, one-sample Kolmogorov–Smirnoff test

Critical values, one-sample Kolmogorov–Smirnov test One-sided test Two-sided test n=1 2 3 4 5

α = 0.10 α = 0.20 0.900 0.684 0.565 0.493 0.447

0.05 0.10 0.950 0.776 0.636 0.565 0.509

0.025 0.05 0.975 0.842 0.708 0.624 0.563

0.01 0.02 0.990 0.900 0.785 0.689 0.627

0.005 0.01 0.995 0.929 0.829 0.734 0.669

6 7 8 9 10

0.410 0.381 0.358 0.339 0.323

0.468 0.436 0.410 0.387 0.369

0.519 0.483 0.454 0.430 0.409

0.577 0.538 0.507 0.480 0.457

0.617 0.576 0.542 0.513 0.489

11 12 13 14 15 16 17 18 19 20

0.308 0.296 0.285 0.275 0.266 0.258 0.250 0.244 0.237 0.232

0.352 0.338 0.325 0.314 0.304 0.295 0.286 0.279 0.271 0.265

0.391 0.375 0.361 0.349 0.338 0.327 0.318 0.309 0.301 0.294

0.437 0.419 0.404 0.390 0.377 0.366 0.355 0.346 0.337 0.329

0.468 0.449 0.432 0.418 0.404 0.392 0.381 0.371 0.361 0.352

21 22 23 24 25

0.226 0.221 0.216 0.212 0.208

0.259 0.253 0.247 0.242 0.238

0.287 0.281 0.275 0.269 0.264

0.321 0.314 0.307 0.301 0.295

0.344 0.337 0.330 0.323 0.317

c 2000 by Chapman & Hall/CRC 

Critical values, one-sample Kolmogorov–Smirnov test One-sided test Two-sided test 26 27 28 29 30

α = 0.10 α = 0.20 0.204 0.200 0.197 0.193 0.190

0.05 0.10 0.233 0.229 0.225 0.221 0.218

0.025 0.05 0.259 0.254 0.250 0.246 0.242

0.01 0.02 0.290 0.284 0.279 0.275 0.270

0.005 0.01 0.311 0.305 0.300 0.295 0.290

0.187 0.184 0.182 0.179 0.177

0.214 0.211 0.208 0.205 0.202

0.238 0.234 0.231 0.227 0.224

0.266 0.262 0.258 0.254 0.251

0.285 0.281 0.277 0.273 0.269

0.174 0.172 0.170 0.168 0.165 1.07 √ n

0.199 0.196 0.194 0.191 0.189 1.22 √ n

0.221 0.218 0.215 0.213 0.210 1.36 √ n

0.247 0.244 0.241 0.238 0.235 1.52 √ n

0.265 0.262 0.258 0.255 0.252 1.63 √ n

31 32 33 34 35 36 37 38 39 40 Approximation for n > 40: 14.3.3.2

Critical values, two-sample Kolmogorov–Smirnoff test

Given the null hypothesis that the two distributions are the same (H0 : F1 (x) = F2 (x)), compute D = max |Fn1 (x) − Fn2 (x)|. (a) Reject H0 if D exceeds the value in the table on page 350. (b) Where ∗ appears in the table on page 350, do not reject H0 at the given significance level. (c) For large values of n1 and n2 , and various values of α, the approximate critical value of D is given in the table below. Level of significance Approximate critical value  2 α = 0.10 1.22 nn11+n n2  2 α = 0.05 1.36 nn11+n n2  2 α = 0.025 1.48 nn11+n n2  2 α = 0.01 1.63 nn11+n n2  2 α = 0.005 1.73 nn11+n n2  2 α = 0.001 1.95 nn11+n n2 c 2000 by Chapman & Hall/CRC 

The entries in the following table are expressed as rational numbers since all critical values of D are an integer divided by n1 n2 . For example, if n1 = 6 and n2 = 5, then  .108225 = Prob D  .047619 = Prob D  .025974 = Prob D  = Prob D  .004329 = Prob D

≥ ≥ ≥ ≥ ≥

       20 21 22 23 = Prob D ≥ = Prob D ≥ = Prob D ≥ 30 30 30 30  24 (least value of D for which α < 0.05) 30    25 26 (14.8) = Prob D ≥ 30 30      27 28 29 = Prob D ≥ = Prob D ≥ 30 30 30  30 (least value of D for which α < 0.01) 30

See P. J. Kim and R. I. Jennrich, Tables of the exact sampling distribution of the two-sample Kolmogorov–Smirnov criterion, Dmn , m ≤ n, pages 79– 170, in H. L. Harter and D. B. Brown (ed.), Selected Tables in Mathematical Statistics, Volume 1, American Mathematical Society, Providence, RI, 1973. Critical values for the Kolmogorov–Smirnov test of F1 (x) = F2 (x) (upper value for α ≤ .05, lower value for α ≤ .01) Sample Sample size n1 size n2 3 4 5 6 7 8 9 1 ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ ∗ 2 ∗ ∗ ∗ ∗ ∗ 16/16 18/18 ∗ ∗ ∗ ∗ ∗ ∗ ∗ 3 ∗ ∗ 15/15 18/18 21/21 21/24 24/27 ∗ ∗ ∗ ∗ ∗ 24/24 27/27 4 16/16 20/20 20/24 24/28 28/32 28/36 ∗ ∗ 24/24 28/28 32/32 32/36 5 ∗ 24/30 30/35 30/40 35/45 ∗ 30/30 35/35 35/40 40/45 6 30/36 30/42 34/48 39/54 36/36 36/42 40/48 45/54 7 42/49 40/56 42/63 42/49 48/56 49/63 8 48/64 46/72 56/64 55/72 9 54/81 63/81 10 11 12

c 2000 by Chapman & Hall/CRC 

10 ∗ ∗ 20/20 ∗ 27/30 30/30 30/40 36/40 40/50 45/50 40/60 48/60 46/70 53/70 48/80 60/80 53/90 70/90 70/100 80/100

11 ∗ ∗ 22/22 ∗ 30/33 33/33 33/44 40/44 39/55 45/55 43/66 54/66 48/77 59/77 53/88 64/88 59/99 70/99 60/110 77/110 77/121 88/121

12 ∗ ∗ 24/24 ∗ 30/36 36/35 36/48 44/48 43/60 50/60 48/72 60/72 53/84 60/84 60/96 68/96 63/108 75/108 66/120 80/120 72/132 86/132 96/144 84/144

14.4

KRUSKAL–WALLIS TEST

Assumptions: Suppose there are k > 2 independent random samples from continuous distributions, let ni (for i = 1, 2, . . . , k) be the number of observations in each sample, and let n = n1 + n2 + · · · + nk . Hypothesis test: H0 : the k samples are from identical populations. Ha : at least two of the populations differ. Rank all n observations from 1 (smallest) to n (largest). Equal observations are assigned the mean rank for their positions. Let Rij be the rank assigned to the j th observation in the ith sample, and let Ri be the total of the ranks in the ith sample. & % k  12 Ri2 − 3(n + 1) TS: H = n(n + 1) i=1 ni RR: H ≥ h where h is the critical value for the Kruskal–Wallis statistic (see table on page 352) such that Prob [H ≥ h] ≈ α. Note: (1) The Kruskal–Wallis procedure is equivalent to an analysis of variance of the ranks. Define the variance ratio as k 

VR =



Ri −R

2

ni k−1 i=1 ni R −R 2 k   ( ij i ) n−k i=1 j=1

(14.9)

where Ri = Ri /ni is the mean of the ranks assigned to the ith sample and R = (n+1)/2 is the overall mean. The Kruskal–Wallis test statistic, H, and VR are related by the equations VR =

H(n − k) , (k − 1)(n − 1 − H)

H=

(n − 1)(k − 1)VR . (n − k) + (k − 1)VR

(14.10)

(2) As n → ∞ and each ni /n → λi > 0, H has approximately a chi–square distribution with k − 1 degrees of freedom. Practically, if H0 is true, and either (a) k = 3, ni ≥ 6, i = 1, 2, 3 or ni ≥ 5, i = 1, 2, . . . , k then H has a chi–square distribution with k − 1 degrees of freedom. (3) The variance ratio, VR, has approximately an F distribution with k − 1 and n − k degrees of freedom. (b) k > 3,

c 2000 by Chapman & Hall/CRC 

Example 14.70 : Suppose that k = 3 treatments (A, B, and C) result in the following observations {1.2, 1.8, 1.7}, {0.9, 0.7}, and {1.0, 0.8}. (Therefore, n1 = 3, n2 = 2, n3 = 2, n = 7.) Ranking these values: Treatment Sample size, ni Ranks

Hence, H =

12 7(8)



182 3

Rank sums, Ri  2 2 + 42 + 62 − 3(8) =

A 3 5 7 6 18 33 7

B 2 3 1

C 2 4 2

4

6

≈ 4.714. From the table on page 352

with {ni } = {3, 2, 2}, we observe that Prob [H ≥ 4.714] = .0476. At the α = .05 level of significance, there is evidence to suggest at least two of the populations differ.

See R. L. Iman, D. Quade, and D. A. Alexander, Exact probability levels for the Kruskal–Wallis test, Selected Tables in Mathematical Statistics, Volume 3, American Mathematical Society, Providence, RI, 1975. 14.4.1

Tables for Kruskal–Wallis test

{ni } = {2, 1, 1} {ni } = {2, 2, 1} {ni } = {2, 2, 2} {ni } = {3, 2, 1} h P (H ≥ h) 2.700 0.5000

h P (H ≥ h) 3.600 0.2000

h P (H ≥ h) 4.571 0.0667 3.714 0.2000

h P (H ≥ h) 4.286 0.1000 3.857 0.1333

{ni } = {3, 2, 2} {ni } = {3, 3, 1} {ni } = {3, 3, 2} {ni } = {3, 3, 3} h P (H ≥ h) 5.357 0.0286 4.714 0.0476 4.500 0.0667 4.464 0.1048 3.929 0.1810 3.750 0.2190 3.607 0.2381

h P (H ≥ h) 5.143 0.0429 4.571 0.1000 4.000 0.1286 3.286 0.1571 3.143 0.2429 2.571 0.3286 2.286 0.4857

h P (H ≥ h) 6.250 0.0107 5.556 0.0250 5.361 0.0321 5.139 0.0607 5.000 0.0750 4.694 0.0929 4.556 0.1000

h P (H ≥ h) 7.200 0.0036 6.489 0.0107 5.956 0.0250 5.689 0.0286 5.600 0.0500 5.067 0.0857 4.622 0.1000

{ni } = {4, 2, 1} {ni } = {4, 2, 2} {ni } = {4, 3, 1} {ni } = {4, 3, 2} h P (H ≥ h) 4.821 0.0571 4.500 0.0762 4.018 0.1143 3.750 0.1333 3.696 0.1714 3.161 0.1905 2.893 0.2667 2.786 0.2857

h P (H ≥ h) 6.000 0.0143 5.500 0.0238 5.333 0.0333 5.125 0.0524 4.500 0.0905 4.458 0.1000 4.167 0.1048 4.125 0.1524

c 2000 by Chapman & Hall/CRC 

h P (H ≥ h) 5.833 0.0214 5.389 0.0357 5.208 0.0500 5.000 0.0571 4.764 0.0714 4.208 0.0786 4.097 0.0857 4.056 0.0929

h P (H ≥ h) 7.000 0.0048 6.444 0.0079 6.300 0.0111 6.111 0.0206 5.800 0.0302 5.500 0.0397 5.400 0.0508 4.444 0.1016

{ni } = {4, 3, 3} {ni } = {4, 4, 1} {ni } = {4, 4, 2} {ni } = {4, 4, 3} h P (H ≥ h) 8.018 0.0014 7.000 0.0062 6.745 0.0100 6.564 0.0171 6.018 0.0267 5.982 0.0343 5.727 0.0505 5.436 0.0619 5.064 0.0705 4.845 0.0810 4.700 0.1010

14.5

h P (H ≥ h) 6.667 0.0095 6.167 0.0222 6.000 0.0286 5.667 0.0349 5.100 0.0413 4.967 0.0476 4.867 0.0540 4.267 0.0698 4.167 0.0825 4.067 0.1016 3.900 0.1079

h P (H ≥ h) 7.855 0.0019 6.873 0.0108 6.545 0.0203 5.945 0.0279 5.645 0.0394 5.236 0.0521 4.991 0.0648 4.691 0.0800 4.555 0.0978 4.445 0.1029

h P (H ≥ h) 8.909 0.0005 7.144 0.0097 7.136 0.0107 6.659 0.0201 6.182 0.0296 6.167 0.0306 6.000 0.0400 5.576 0.0507 4.712 0.0902 4.477 0.1022

THE RUNS TEST

A run is a maximal subsequence of elements with a common property. Hypothesis test: H0 : the sequence is random. Ha : the sequence is not random. TS: V = the total number of runs RR: V ≥ v1 or V ≤ v2 where v1 and v2 are critical values for the runs test (see page 354) such that Prob [V ≥ v1 ] ≈ α/2 and Prob [V ≤ v2 ] ≈ α/2. The normal approximation: Let m be the number of elements with the property that occurs least and n be the number of elements with the other property. As m and n increase, V has approximately a normal distribution with µV =

2mn +1 m+n

and σV2 =

2mn(2mn − m − n) . (m + n)2 (m + n + 1)

(14.11)

The random variable V − µV σV has approximately a standard normal distribution. Z=

(14.12)

Example 14.71 : Suppose the following sequence of heads (H) and tails (T ) was obtained from flipping a coin: {H, H, T , T , H, T , H, T , T , T , T , H}. Is there any evidence to suggest the coin is biased? Solution: (S1) Place vertical bars at the end of each run. The data set may be written to easily count the number of runs. HH | T T | H | T | H | T T T T | H | c 2000 by Chapman & Hall/CRC 

(S2) Using this notation, there are 5 H’s, 7 T ’s, and 7 runs. (S3) The table on page 356 (using m = 5 and n = 7) indicates that 65% of the time one would expect there to be 7 runs or fewer. (S4) The table on page 356 (using m = 5 and n = 6) indicates that 42% of the time one would expect there be 6 runs or fewer. Alternatively, 58% (since 1 − 0.42 = 0.58) of the time there would be 7 runs or more. (S5) In neither case is there any evidence to suggest the coin is biased.

14.5.1

Tables for the runs test

Runs can be used to test data for randomness or to test the hypothesis that two samples come from the same distribution. A run is defined as a succession of identical elements which are followed and preceded by different elements or by no elements at all. Let m be the number of elements of one kind and n be the number of elements of the other kind. Let v equal the total number of runs among the n + m elements. The probability that exactly v runs occur is given by   n−1  m−1  2 (k−2)/2 (k−2)/2    n+m if k is even    n Prob [v runs ] =  (14.13)  n−1  m−1  +  n−1  m−1      (k−3)/2 (k−1)/2n+m(k−1)/2 (k−3)/2 if k is odd  n

The following tables give the sampling distribution for v for values of m and n less than or equal to 20. That is, the values listed in this table give the probability that v or fewer runs will occur. The table on page 364 gives percentage points of the distribution for larger sample sizes when m = n. The columns headed with 0.5%, 1%, 2.5%, 5% indicate the values of v such that v or fewer runs occur with probability less than that indicated; the columns headed with 97.5%, 99%, 99.5% indicate values of v for which the probability of v or more runs is less than 2.5%, 1%, 0.5%. For large values of m and n, particularly for m = n greater than 10, a normal approximation may be used, with the parameters given in equation (14.11).

c 2000 by Chapman & Hall/CRC 

Distribution of total number of runs v in samples of size (m, n) m, n 2, 2 2, 3 2, 4 2, 5 2, 6 2, 7 2, 8 2, 9 2, 10 2, 11 2, 12 2, 13 2, 14 2, 15 2, 16 2, 17 2, 18 2, 19 2, 20

v=2 0.3333 0.2000 0.1333 0.0952 0.0714 0.0556 0.0444 0.0364 0.0303 0.0256 0.0220 0.0190 0.0167 0.0147 0.0131 0.0117 0.0105 0.0095 0.0087

3 0.6667 0.5000 0.4000 0.3333 0.2857 0.2500 0.2222 0.2000 0.1818 0.1667 0.1538 0.1429 0.1333 0.1250 0.1176 0.1111 0.1053 0.1000 0.0952

4 1.0000 0.9000 0.8000 0.7143 0.6429 0.5833 0.5333 0.4909 0.4545 0.4231 0.3956 0.3714 0.3500 0.3309 0.3137 0.2982 0.2842 0.2714 0.2597

1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000

3, 3 3, 4 3, 5 3, 6 3, 7 3, 8 3, 9 3, 10 3, 11 3, 12 3, 13 3, 14 3, 15 3, 16 3, 17 3, 18 3, 19 3, 20

0.1000 0.0571 0.0357 0.0238 0.0167 0.0121 0.0091 0.0070 0.0055 0.0044 0.0036 0.0029 0.0025 0.0021 0.0018 0.0015 0.0013 0.0011

0.3000 0.2000 0.1429 0.1071 0.0833 0.0667 0.0545 0.0455 0.0385 0.0330 0.0286 0.0250 0.0221 0.0196 0.0175 0.0158 0.0143 0.0130

0.7000 0.5429 0.4286 0.3452 0.2833 0.2364 0.2000 0.1713 0.1484 0.1297 0.1143 0.1015 0.0907 0.0815 0.0737 0.0669 0.0610 0.0559

0.9000 0.8000 0.7143 0.6429 0.5833 0.5333 0.4909 0.4545 0.4231 0.3956 0.3714 0.3500 0.3309 0.3137 0.2982 0.2842 0.2714 0.2597

c 2000 by Chapman & Hall/CRC 

5

6

7

1.0000 0.9714 0.9286 0.8810 0.8333 0.7879 0.7455 0.7063 0.6703 0.6374 0.6071 0.5794 0.5539 0.5304 0.5088 0.4887 0.4701 0.4529

1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000

8

9

10

Distribution of total number of runs v in samples of size (m, n) m, n 4, 4 4, 5 4, 6 4, 7 4, 8 4, 9 4, 10 4, 11 4, 12 4, 13 4, 14 4, 15 4, 16 4, 17 4, 18 4, 19 4, 20

v=2 0.0286 0.0159 0.0095 0.0061 0.0040 0.0028 0.0020 0.0015 0.0011 0.0008 0.0007 0.0005 0.0004 0.0003 0.0003 0.0002 0.0002

3 0.1143 0.0714 0.0476 0.0333 0.0242 0.0182 0.0140 0.0110 0.0088 0.0071 0.0059 0.0049 0.0041 0.0035 0.0030 0.0026 0.0023

4 0.3714 0.2619 0.1905 0.1424 0.1091 0.0853 0.0679 0.0549 0.0451 0.0374 0.0314 0.0266 0.0227 0.0195 0.0170 0.0148 0.0130

5 0.6286 0.5000 0.4048 0.3333 0.2788 0.2364 0.2028 0.1758 0.1538 0.1357 0.1206 0.1078 0.0970 0.0877 0.0797 0.0727 0.0666

6 0.8857 0.7857 0.6905 0.6061 0.5333 0.4713 0.4186 0.3736 0.3352 0.3021 0.2735 0.2487 0.2270 0.2080 0.1913 0.1764 0.1632

7 0.9714 0.9286 0.8810 0.8333 0.7879 0.7455 0.7063 0.6703 0.6374 0.6071 0.5794 0.5539 0.5304 0.5088 0.4887 0.4701 0.4529

8 1.0000 0.9921 0.9762 0.9545 0.9293 0.9021 0.8741 0.8462 0.8187 0.7920 0.7663 0.7417 0.7183 0.6959 0.6746 0.6544 0.6352

1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000

5, 5 5, 6 5, 7 5, 8 5, 9 5, 10 5, 11 5, 12 5, 13 5, 14 5, 15 5, 16 5, 17 5, 18 5, 19 5, 20

0.0079 0.0043 0.0025 0.0016 0.0010 0.0007 0.0005 0.0003 0.0002 0.0002 0.0001 .04 983 .04 759 .04 594 .04 471 .04 376

0.0397 0.0238 0.0152 0.0101 0.0070 0.0050 0.0037 0.0027 0.0021 0.0016 0.0013 0.0010 0.0008 0.0007 0.0006 0.0005

0.1667 0.1104 0.0758 0.0536 0.0390 0.0290 0.0220 0.0170 0.0133 0.0106 0.0085 0.0069 0.0057 0.0047 0.0040 0.0033

0.3571 0.2619 0.1970 0.1515 0.1189 0.0949 0.0769 0.0632 0.0525 0.0441 0.0374 0.0320 0.0276 0.0239 0.0209 0.0184

0.6429 0.5216 0.4242 0.3473 0.2867 0.2388 0.2005 0.1698 0.1450 0.1246 0.1078 0.0939 0.0823 0.0724 0.0641 0.0570

0.8333 0.7381 0.6515 0.5758 0.5105 0.4545 0.4066 0.3654 0.3298 0.2990 0.2722 0.2487 0.2281 0.2098 0.1937 0.1793

0.9603 0.9113 0.8535 0.7933 0.7343 0.6783 0.6264 0.5787 0.5352 0.4958 0.4600 0.4276 0.3982 0.3715 0.3473 0.3252

0.9921 0.9762 0.9545 0.9293 0.9021 0.8741 0.8462 0.8187 0.7920 0.7663 0.7417 0.7183 0.6959 0.6746 0.6544 0.6352

1.0000 0.9978 0.9924 0.9837 0.9720 0.9580 0.9423 0.9253 0.9076 0.8893 0.8709 0.8524 0.8341 0.8161 0.7984 0.7811

6, 6 6, 7 6, 8 6, 9 6, 10 6, 11 6, 12 6, 13 6, 14 6, 15 6, 16 6, 17 6, 18 6, 19 6, 20

0.0022 0.0012 0.0007 0.0004 0.0002 0.0002 0.0001 .04 737 .04 516 .04 369 .04 268 .04 198 .04 149 .04 113 .05 869

0.0130 0.0076 0.0047 0.0030 0.0020 0.0014 0.0010 0.0007 0.0005 0.0004 0.0003 0.0002 0.0002 0.0001 0.0001

0.0671 0.0425 0.0280 0.0190 0.0132 0.0095 0.0069 0.0051 0.0039 0.0030 0.0023 0.0018 0.0014 0.0012 0.0009

0.1753 0.1212 0.0862 0.0629 0.0470 0.0357 0.0276 0.0217 0.0173 0.0139 0.0114 0.0093 0.0078 0.0065 0.0055

0.3918 0.2960 0.2261 0.1748 0.1369 0.1084 0.0869 0.0704 0.0575 0.0475 0.0395 0.0331 0.0280 0.0238 0.0203

0.6082 0.5000 0.4126 0.3427 0.2867 0.2418 0.2054 0.1758 0.1514 0.1313 0.1146 0.1005 0.0886 0.0785 0.0698

0.8247 0.7331 0.6457 0.5664 0.4965 0.4357 0.3832 0.3379 0.2990 0.2655 0.2365 0.2114 0.1896 0.1706 0.1540

0.9329 0.8788 0.8205 0.7622 0.7063 0.6538 0.6054 0.5609 0.5204 0.4835 0.4499 0.4195 0.3917 0.3665 0.3434

0.9870 0.9662 0.9371 0.9021 0.8636 0.8235 0.7831 0.7434 0.7049 0.6680 0.6329 0.5998 0.5685 0.5392 0.5118

c 2000 by Chapman & Hall/CRC 

9

10

Distribution of total number of runs v in samples of size (m, n) m, n v = 11 4, 4 4, 5 4, 6 4, 7 4, 8 4, 9 4, 10 4, 11 4, 12 4, 13 4, 14 4, 15 4, 16 4, 17 4, 18 4, 19 4, 20 5, 5 5, 6 5, 7 5, 8 5, 9 5, 10 5, 11 5, 12 5, 13 5, 14 5, 15 5, 16 5, 17 5, 18 5, 19 5, 20

1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000

6, 6 6, 7 6, 8 6, 9 6, 10 6, 11 6, 12 6, 13 6, 14 6, 15 6, 16 6, 17 6, 18 6, 19 6, 20

0.9978 0.9924 0.9837 0.9720 0.9580 0.9423 0.9253 0.9076 0.8893 0.8709 0.8524 0.8341 0.8161 0.7984 0.7811

12

13

1.0000 0.9994 0.9977 0.9944 0.9895 0.9830 0.9751 0.9659 0.9557 0.9447 0.9329 0.9207 0.9081 0.8952 0.8822

1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000

c 2000 by Chapman & Hall/CRC 

14 15 16 17 18 19 20 21

Distribution of total number of runs v in samples of size (m, n) m, n 7, 7 7, 8 7, 9 7, 10 7, 11 7, 12 7, 13 7, 14 7, 15 7, 16 7, 17 7, 18 7, 19 7, 20

v=2 0.0006 0.0003 0.0002 0.0001 .04 628 .04 397 .04 258 .04 172 .04 117 .05 816 .05 578 .05 416 .05 304 .05 225

3 0.0041 0.0023 0.0014 0.0009 0.0006 0.0004 0.0003 0.0002 0.0001 .04 938 .04 693 .04 520 .04 395 .04 304

4 0.0251 0.0154 0.0098 0.0064 0.0043 0.0030 0.0021 0.0015 0.0011 0.0008 0.0006 0.0005 0.0004 0.0003

5 0.0775 0.0513 0.0350 0.0245 0.0175 0.0128 0.0095 0.0072 0.0055 0.0043 0.0034 0.0027 0.0022 0.0018

6 0.2086 0.1492 0.1084 0.0800 0.0600 0.0456 0.0351 0.0273 0.0216 0.0172 0.0138 0.0112 0.0092 0.0075

7 0.3834 0.2960 0.2308 0.1818 0.1448 0.1165 0.0947 0.0777 0.0642 0.0536 0.0450 0.0381 0.0324 0.0278

8 0.6166 0.5136 0.4266 0.3546 0.2956 0.2475 0.2082 0.1760 0.1496 0.1278 0.1097 0.0947 0.0820 0.0714

9 0.7914 0.7040 0.6224 0.5490 0.4842 0.4276 0.3785 0.3359 0.2990 0.2670 0.2392 0.2149 0.1937 0.1751

10 0.9225 0.8671 0.8059 0.7433 0.6821 0.6241 0.5700 0.5204 0.4751 0.4340 0.3969 0.3634 0.3332 0.3060

8, 8 8, 9 8, 10 8, 11 8, 12 8, 13 8, 14 8, 15 8, 16 8, 17 8, 18 8, 19 8, 20

0.0002 .04 823 .04 457 .04 265 .04 159 .05 983 .05 625 .05 408 .05 272 .05 185 .05 128 .06 901 .06 643

0.0012 0.0007 0.0004 0.0003 0.0002 0.0001 .04 688 .04 469 .04 326 .04 231 .04 166 .04 122 .05 901

0.0089 0.0053 0.0033 0.0021 0.0014 0.0009 0.0006 0.0004 0.0003 0.0002 0.0002 0.0001 .04 946

0.0317 0.0203 0.0134 0.0090 0.0063 0.0044 0.0032 0.0023 0.0017 0.0013 0.0010 0.0008 0.0006

0.1002 0.0687 0.0479 0.0341 0.0246 0.0181 0.0134 0.0101 0.0077 0.0060 0.0047 0.0037 0.0029

0.2145 0.1573 0.1170 0.0882 0.0674 0.0521 0.0408 0.0322 0.0257 0.0207 0.0169 0.0138 0.0114

0.4048 0.3186 0.2514 0.1994 0.1591 0.1278 0.1034 0.0842 0.0690 0.0570 0.0473 0.0395 0.0332

0.5952 0.5000 0.4194 0.3522 0.2966 0.2508 0.2129 0.1816 0.1556 0.1340 0.1159 0.1006 0.0878

0.7855 0.7016 0.6209 0.5467 0.4800 0.4211 0.3695 0.3245 0.2856 0.2518 0.2225 0.1971 0.1751

9, 9 9, 10 9, 11 9, 12 9, 13 9, 14 9, 15 9, 16 9, 17 9, 18 9, 19 9, 20

.04 411 .04 217 .04 119 .05 680 .05 402 .05 245 .05 153 .06 979 .06 640 .06 427 .06 290 .06 200

0.0004 0.0002 0.0001 .04 714 .04 442 .04 281 .04 184 .04 122 .05 832 .05 576 .05 405 .05 290

0.0030 0.0018 0.0011 0.0007 0.0004 0.0003 0.0002 0.0001 .04 903 .04 638 .04 458 .04 333

0.0122 0.0076 0.0049 0.0032 0.0022 0.0015 0.0010 0.0007 0.0005 0.0004 0.0003 0.0002

0.0445 0.0294 0.0199 0.0137 0.0096 0.0068 0.0049 0.0036 0.0027 0.0020 0.0015 0.0012

0.1090 0.0767 0.0549 0.0399 0.0294 0.0220 0.0166 0.0127 0.0099 0.0077 0.0061 0.0048

0.2380 0.1786 0.1349 0.1028 0.0789 0.0612 0.0478 0.0377 0.0299 0.0240 0.0193 0.0157

0.3992 0.3186 0.2549 0.2049 0.1656 0.1347 0.1102 0.0907 0.0751 0.0626 0.0524 0.0441

0.6008 0.5095 0.4300 0.3621 0.3050 0.2572 0.2174 0.1842 0.1566 0.1336 0.1144 0.0983

10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10,

.04 108 .05 567 .05 309 .05 175 .05 102 .06 612 .06 377 .06 237 .06 152 .07 999 .07 666

0.0001 .04 595 .04 340 .04 201 .04 122 .05 765 .05 489 .05 320 .05 213 .05 145 .06 999

0.0010 0.0006 0.0003 0.0002 0.0001 .04 847 .04 557 .04 373 .04 255 .04 176 .04 124

0.0045 0.0027 0.0017 0.0011 0.0007 0.0005 0.0003 0.0002 0.0002 0.0001 .04 864

0.0185 0.0119 0.0078 0.0053 0.0036 0.0025 0.0018 0.0013 0.0009 0.0007 0.0005

0.0513 0.0349 0.0242 0.0170 0.0122 0.0088 0.0065 0.0048 0.0036 0.0028 0.0021

0.1276 0.0920 0.0670 0.0493 0.0367 0.0275 0.0209 0.0160 0.0124 0.0096 0.0076

0.2422 0.1849 0.1421 0.1099 0.0857 0.0673 0.0533 0.0425 0.0341 0.0276 0.0225

0.4141 0.3350 0.2707 0.2189 0.1775 0.1445 0.1180 0.0968 0.0798 0.0661 0.0550

10 11 12 13 14 15 16 17 18 19 20

c 2000 by Chapman & Hall/CRC 

Distribution of total number of runs v in samples of size (m, n) m, n 7, 7 7, 8 7, 9 7, 10 7, 11 7, 12 7, 13 7, 14 7, 15 7, 16 7, 17 7, 18 7, 19 7, 20

v = 11 0.9749 0.9487 0.9161 0.8794 0.8405 0.8009 0.7616 0.7233 0.6864 0.6512 0.6178 0.5862 0.5565 0.5286

12 0.9959 0.9879 0.9748 0.9571 0.9355 0.9109 0.8842 0.8561 0.8273 0.7982 0.7692 0.7407 0.7128 0.6857

13 0.9994 0.9977 0.9944 0.9895 0.9830 0.9751 0.9659 0.9557 0.9447 0.9329 0.9207 0.9081 0.8952 0.8822

14 1.0000 0.9998 0.9993 0.9981 0.9962 0.9935 0.9898 0.9852 0.9799 0.9738 0.9669 0.9595 0.9516 0.9433

1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000

8, 8 8, 9 8, 10 8, 11 8, 12 8, 13 8, 14 8, 15 8, 16 8, 17 8, 18 8, 19 8, 20

0.8998 0.8427 0.7822 0.7217 0.6634 0.6084 0.5573 0.5103 0.4674 0.4285 0.3931 0.3611 0.3322

0.9683 0.9394 0.9031 0.8618 0.8174 0.7718 0.7263 0.6818 0.6389 0.5981 0.5595 0.5232 0.4893

0.9911 0.9797 0.9636 0.9434 0.9201 0.8944 0.8672 0.8390 0.8104 0.7818 0.7536 0.7258 0.6988

0.9988 0.9958 0.9905 0.9823 0.9714 0.9580 0.9423 0.9248 0.9057 0.8855 0.8645 0.8429 0.8210

9, 9 9, 10 9, 11 9, 12 9, 13 9, 14 9, 15 9, 16 9, 17 9, 18 9, 19 9, 20

0.7620 0.6814 0.6050 0.5350 0.4721 0.4164 0.3674 0.3245 0.2871 0.2545 0.2261 0.2013

0.8910 0.8342 0.7731 0.7111 0.6505 0.5928 0.5389 0.4892 0.4437 0.4024 0.3650 0.3313

0.9555 0.9233 0.8851 0.8431 0.7991 0.7545 0.7104 0.6675 0.6264 0.5872 0.5503 0.5155

10, 10, 10, 10, 10, 10, 10, 10, 10, 10, 10,

0.5859 0.5000 0.4250 0.3607 0.3062 0.2602 0.2216 0.1893 0.1621 0.1392 0.1200

0.7578 0.6800 0.6050 0.5351 0.4715 0.4146 0.3641 0.3197 0.2809 0.2470 0.2175

0.8724 0.8151 0.7551 0.6950 0.6369 0.5818 0.5303 0.4828 0.4393 0.3997 0.3638

10 11 12 13 14 15 16 17 18 19 20

16

17

0.9998 0.9993 0.9981 0.9962 0.9935 0.9898 0.9852 0.9799 0.9738 0.9669 0.9595 0.9516 0.9433

1.0000 1.0000 0.9998 0.9994 0.9987 0.9976 0.9960 0.9939 0.9913 0.9881 0.9844 0.9803 0.9757

1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000

0.9878 0.9742 0.9551 0.9311 0.9031 0.8721 0.8390 0.8047 0.7699 0.7351 0.7008 0.6672

0.9970 0.9924 0.9851 0.9751 0.9625 0.9477 0.9309 0.9125 0.8929 0.8724 0.8513 0.8298

0.9996 0.9986 0.9965 0.9931 0.9880 0.9813 0.9729 0.9629 0.9515 0.9388 0.9250 0.9103

0.9487 0.9151 0.8751 0.8307 0.7839 0.7361 0.6886 0.6423 0.5978 0.5554 0.5155

0.9815 0.9651 0.9437 0.9180 0.8889 0.8574 0.8243 0.7904 0.7562 0.7223 0.6889

0.9955 0.9896 0.9804 0.9678 0.9519 0.9330 0.9115 0.8880 0.8629 0.8367 0.8097

c 2000 by Chapman & Hall/CRC 

15

18

19

1.0000 0.9998 0.9994 0.9987 0.9976 0.9960 0.9939 0.9913 0.9881 0.9844 0.9803 0.9757

1.0000 1.0000 0.9999 0.9998 0.9996 0.9991 0.9985 0.9976 0.9963 0.9948 0.9930 0.9908

1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000

0.9990 0.9973 0.9942 0.9896 0.9834 0.9755 0.9661 0.9551 0.9429 0.9296 0.9153

0.9999 0.9996 0.9988 0.9974 0.9952 0.9920 0.9879 0.9826 0.9763 0.9689 0.9606

1.0000 0.9999 0.9998 0.9996 0.9991 0.9985 0.9976 0.9963 0.9948 0.9930 0.9908

20

21

1.0000 1.0000 1.0000 0.9999 0.9999 0.9997 0.9994 0.9991 0.9985 0.9978 0.9969

1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000

Distribution of total number of runs v in samples of size (m, n) m, n 11, 11 11, 12 11, 13 11, 14 11, 15 11, 16 11, 17 11, 18 11, 19 11, 20

v=2 .05 284 .05 148 .06 801 .06 449 .06 259 .06 153 .07 931 .07 578 .07 366 .07 236

3 .04 312 .04 170 .05 961 .05 561 .05 337 .05 207 .05 130 .06 838 .06 549 .06 366

4 0.0003 0.0002 0.0001 .04 639 .04 396 .04 251 .04 162 .04 107 .05 714 .05 485

5 0.0016 0.0010 0.0006 0.0004 0.0002 0.0002 0.0001 .04 721 .04 500 .04 351

6 0.0073 0.0046 0.0030 0.0019 0.0013 0.0009 0.0006 0.0004 0.0003 0.0002

7 0.0226 0.0150 0.0101 0.0069 0.0048 0.0034 0.0025 0.0018 0.0013 0.0010

8 0.0635 0.0443 0.0313 0.0223 0.0161 0.0118 0.0087 0.0065 0.0049 0.0037

9 0.1349 0.0992 0.0736 0.0551 0.0416 0.0317 0.0244 0.0189 0.0148 0.0116

10 0.2599 0.2017 0.1569 0.1224 0.0960 0.0757 0.0600 0.0478 0.0383 0.0308

12, 12, 12, 12, 12, 12, 12, 12, 12,

12 13 14 15 16 17 18 19 20

.06 740 .06 385 .06 207 .06 115 .07 657 .07 385 .07 231 .07 142 .08 886

.05 888 .05 481 .05 269 .05 155 .06 920 .06 559 .06 347 .06 220 .06 142

.04 984 .04 556 .04 323 .04 193 .04 118 .05 734 .05 467 .05 303 .05 199

0.0005 0.0003 0.0002 0.0001 .04 769 .04 497 .04 328 .04 220 .04 150

0.0028 0.0017 0.0011 0.0007 0.0005 0.0003 0.0002 0.0001 .04 983

0.0095 0.0061 0.0040 0.0027 0.0018 0.0013 0.0009 0.0006 0.0005

0.0296 0.0201 0.0138 0.0096 0.0068 0.0048 0.0035 0.0025 0.0019

0.0699 0.0498 0.0358 0.0260 0.0191 0.0142 0.0106 0.0080 0.0061

0.1504 0.1126 0.0847 0.0640 0.0487 0.0373 0.0288 0.0223 0.0175

13, 13, 13, 13, 13, 13, 13, 13,

13 14 15 16 17 18 19 20

.06 192 .07 997 .07 534 .07 295 .07 167 .08 970 .08 576 .08 349

.05 250 .05 135 .06 748 .06 427 .06 251 .06 150 .07 921 .07 576

.04 302 .04 169 .05 972 .05 573 .05 346 .05 213 .05 134 .06 853

0.0002 0.0001 .04 636 .04 389 .04 243 .04 155 .04 100 .05 662

0.0010 0.0006 0.0004 0.0002 0.0002 0.0001 .04 682 .04 460

0.0038 0.0024 0.0016 0.0010 0.0007 0.0005 0.0003 0.0002

0.0131 0.0087 0.0058 0.0040 0.0027 0.0019 0.0014 0.0010

0.0341 0.0236 0.0165 0.0117 0.0084 0.0061 0.0045 0.0033

0.0812 0.0589 0.0430 0.0316 0.0234 0.0175 0.0132 0.0100

14, 14, 14, 14, 14, 14, 14,

14 15 16 17 18 19 20

.07 499 .07 258 .07 138 .08 754 .08 424 .08 244 .08 144

.06 698 .06 374 .06 206 .06 117 .07 679 .07 403 .07 244

.05 912 .05 507 .05 289 .05 169 .05 101 .06 612 .06 379

.04 597 .04 344 .04 203 .04 123 .05 757 .05 476 .05 304

0.0004 0.0002 0.0001 .04 829 .04 526 .04 339 .04 222

0.0015 0.0009 0.0006 0.0004 0.0002 0.0002 0.0001

0.0056 0.0036 0.0024 0.0016 0.0011 0.0007 0.0005

0.0157 0.0107 0.0073 0.0051 0.0035 0.0025 0.0018

0.0412 0.0291 0.0207 0.0149 0.0108 0.0079 0.0058

15, 15, 15, 15, 15, 15,

15 16 17 18 19 20

.07 129 .08 665 .08 354 .08 193 .08 108 .09 616

.06 193 .06 103 .07 566 .07 318 .07 183 .07 108

.05 272 .05 150 .06 848 .06 491 .06 290 .06 175

.04 191 .04 109 .05 639 .05 382 .05 233 .05 144

0.0001 .04 745 .04 450 .04 277 .04 173 .04 110

0.0006 0.0003 0.0002 0.0001 .04 873 .04 573

0.0023 0.0014 0.0009 0.0006 0.0004 0.0003

0.0070 0.0046 0.0031 0.0021 0.0014 0.0010

0.0199 0.0137 0.0095 0.0067 0.0047 0.0034

16, 16, 16, 16, 16,

16 17 18 19 20

.08 333 .08 171 .09 907 .09 493 .09 274

.07 532 .07 283 .07 154 .08 862 .08 493

.06 802 .06 440 .06 247 .06 142 .07 829

.05 604 .05 342 .05 198 .05 117 .06 707

.04 427 .04 250 .04 149 .05 909 .05 562

0.0002 0.0001 .04 754 .04 473 .04 302

0.0009 0.0006 0.0004 0.0002 0.0002

0.0030 0.0019 0.0013 0.0008 0.0006

0.0092 0.0062 0.0042 0.0029 0.0020

17, 17, 17, 17,

17 18 19 20

.09 857 .09 441 .09 233 .09 126

.07 146 .08 771 .08 419 .08 233

.06 234 .06 128 .07 712 .07 406

.05 188 .05 106 .06 607 .06 356

.04 142 .05 825 .05 488 .05 294

.04 718 .04 430 .04 262 .04 163

0.0003 0.0002 0.0001 .04 845

0.0012 0.0008 0.0005 0.0003

0.0041 0.0027 0.0018 0.0012

18, 18 .09 220 .08 397 .07 677 .06 577 .05 465 .04 250 0.0001 0.0005 0.0017 18, 19 .09 113 .08 209 .07 367 .06 322 .05 268 .04 148 .04 776 0.0003 0.0011 18, 20 .01 0596 .08 113 .07 204 .06 184 .05 157 .05 896 .04 482 0.0002 0.0007 19, 19 .01 0566 .08 108 .07 194 .06 175 .05 150 .05 856 .04 462 0.0002 0.0007 19, 20 .01 0290 .09 566 .07 105 .07 973 .06 857 .05 503 .04 280 0.0001 0.0005 20, 20 .01 0145 .09 290 .08 553 .07 527 .06 477 .05 288 .04 165 .04 710 0.0003 c 2000 by Chapman & Hall/CRC 

Distribution of total number of runs v in samples of size (m, n) m, n v = 11 11, 11 0.4100 11, 12 0.3350 11, 13 0.2735 11, 14 0.2235 11, 15 0.1831 11, 16 0.1504 11, 17 0.1240 11, 18 0.1027 11, 19 0.0853 11, 20 0.0712

12 0.5900 0.5072 0.4334 0.3690 0.3137 0.2665 0.2265 0.1928 0.1644 0.1404

13 0.7401 0.6650 0.5933 0.5267 0.4660 0.4116 0.3632 0.3205 0.2830 0.2500

14 0.8651 0.8086 0.7488 0.6883 0.6293 0.5728 0.5199 0.4708 0.4257 0.3846

15 0.9365 0.9008 0.8598 0.8154 0.7692 0.7225 0.6765 0.6317 0.5888 0.5480

16 0.9774 0.9594 0.9360 0.9078 0.8758 0.8410 0.8043 0.7666 0.7286 0.6908

17 0.9927 0.9850 0.9740 0.9598 0.9424 0.9224 0.9002 0.8763 0.8510 0.8247

18 0.9984 0.9960 0.9919 0.9857 0.9774 0.9669 0.9542 0.9395 0.9231 0.9051

19 0.9997 0.9990 0.9978 0.9958 0.9930 0.9891 0.9841 0.9781 0.9711 0.9631

20 1.0000 0.9999 0.9996 0.9991 0.9981 0.9967 0.9948 0.9922 0.9889 0.9849

21 1.0000 1.0000 0.9999 0.9999 0.9997 0.9994 0.9991 0.9985 0.9978 0.9969

12, 12, 12, 12, 12, 12, 12, 12, 12,

12 13 14 15 16 17 18 19 20

0.2632 0.2068 0.1628 0.1286 0.1020 0.0813 0.0651 0.0524 0.0424

0.4211 0.3475 0.2860 0.2351 0.1933 0.1591 0.1312 0.1085 0.0900

0.5789 0.5000 0.4296 0.3681 0.3149 0.2693 0.2304 0.1973 0.1693

0.7368 0.6642 0.5938 0.5277 0.4669 0.4118 0.3626 0.3189 0.2803

0.8496 0.7932 0.7345 0.6759 0.6189 0.5646 0.5137 0.4665 0.4231

0.9301 0.8937 0.8518 0.8062 0.7585 0.7101 0.6621 0.6153 0.5703

0.9704 0.9502 0.9251 0.8958 0.8632 0.8283 0.7919 0.7548 0.7176

0.9905 0.9816 0.9691 0.9528 0.9330 0.9101 0.8847 0.8572 0.8281

0.9972 0.9939 0.9886 0.9813 0.9718 0.9602 0.9465 0.9311 0.9140

0.9995 0.9985 0.9968 0.9940 0.9899 0.9844 0.9774 0.9690 0.9590

0.9999 0.9997 0.9992 0.9984 0.9971 0.9953 0.9929 0.9898 0.9860

13, 13, 13, 13, 13, 13, 13, 13,

13 14 15 16 17 18 19 20

0.1566 0.1189 0.0906 0.0695 0.0535 0.0415 0.0324 0.0254

0.2772 0.2205 0.1753 0.1396 0.1113 0.0890 0.0714 0.0575

0.4179 0.3475 0.2883 0.2389 0.1980 0.1643 0.1365 0.1138

0.5821 0.5056 0.4365 0.3751 0.3215 0.2752 0.2353 0.2012

0.7228 0.6525 0.5847 0.5212 0.4628 0.4098 0.3623 0.3200

0.8434 0.7880 0.7299 0.6714 0.6141 0.5592 0.5074 0.4592

0.9188 0.8811 0.8388 0.7934 0.7465 0.6992 0.6525 0.6072

0.9659 0.9446 0.9182 0.8873 0.8529 0.8159 0.7772 0.7377

0.9869 0.9764 0.9623 0.9446 0.9238 0.9001 0.8742 0.8465

0.9962 0.9921 0.9858 0.9771 0.9658 0.9520 0.9358 0.9174

0.9990 0.9976 0.9952 0.9917 0.9868 0.9805 0.9728 0.9635

14, 14, 14, 14, 14, 14, 14,

14 15 16 17 18 19 20

0.0871 0.0642 0.0476 0.0355 0.0266 0.0202 0.0153

0.1697 0.1306 0.1007 0.0779 0.0604 0.0471 0.0368

0.2798 0.2247 0.1804 0.1450 0.1167 0.0942 0.0763

0.4266 0.3576 0.2986 0.2487 0.2068 0.1720 0.1432

0.5734 0.5000 0.4336 0.3745 0.3227 0.2776 0.2387

0.7202 0.6519 0.5854 0.5226 0.4643 0.4110 0.3630

0.8303 0.7753 0.7183 0.6614 0.6058 0.5527 0.5027

0.9129 0.8749 0.8322 0.7863 0.7386 0.6903 0.6425

0.9588 0.9358 0.9081 0.8765 0.8418 0.8049 0.7667

0.9843 0.9727 0.9574 0.9382 0.9155 0.8898 0.8616

0.9944 0.9893 0.9820 0.9721 0.9598 0.9450 0.9281

15, 15, 15, 15, 15, 15,

15 16 17 18 19 20

0.0457 0.0328 0.0237 0.0173 0.0127 0.0094

0.0974 0.0728 0.0546 0.0412 0.0312 0.0237

0.1749 0.1362 0.1061 0.0830 0.0650 0.0512

0.2912 0.2362 0.1912 0.1546 0.1251 0.1014

0.4241 0.3576 0.3005 0.2519 0.2109 0.1766

0.5759 0.5046 0.4393 0.3806 0.3286 0.2831

0.7088 0.6424 0.5781 0.5174 0.4610 0.4095

0.8251 0.7710 0.7147 0.6581 0.6026 0.5493

0.9026 0.8638 0.8210 0.7754 0.7285 0.6813

0.9543 0.9305 0.9020 0.8693 0.8334 0.7952

0.9801 0.9672 0.9505 0.9303 0.9068 0.8806

16, 16, 16, 16, 16,

16 17 18 19 20

0.0228 0.0160 0.0113 0.0080 0.0058

0.0528 0.0385 0.0282 0.0207 0.0153

0.1028 0.0778 0.0591 0.0450 0.0345

0.1862 0.1465 0.1153 0.0908 0.0716

0.2933 0.2397 0.1956 0.1594 0.1300

0.4311 0.3659 0.3091 0.2603 0.2188

0.5689 0.5000 0.4369 0.3801 0.3297

0.7067 0.6420 0.5789 0.5188 0.4628

0.8138 0.7603 0.7051 0.6498 0.5959

0.8972 0.8584 0.8155 0.7697 0.7224

0.9472 0.9222 0.8928 0.8596 0.8237

17, 17, 17, 17,

17 18 19 20

0.0109 0.0075 0.0052 0.0036

0.0272 0.0194 0.0139 0.0100

0.0572 0.0422 0.0313 0.0233

0.1122 0.0859 0.0659 0.0506

0.1907 0.1514 0.1202 0.0955

0.3028 0.2495 0.2049 0.1680

0.4290 0.3659 0.3108 0.2631

0.5710 0.5038 0.4418 0.3854

0.6972 0.6341 0.5728 0.5146

0.8093 0.7567 0.7022 0.6474

0.8878 0.8486 0.8057 0.7604

18, 18 18, 19 18, 20

0.0050 0.0134 0.0303 0.0640 0.1171 0.2004 0.3046 0.4349 0.5651 0.6954 0.7996 0.0034 0.0094 0.0219 0.0479 0.0906 0.1606 0.2525 0.3729 0.5000 0.6338 0.7475 0.0023 0.0066 0.0159 0.0359 0.0701 0.1285 0.2088 0.3182 0.4398 0.5736 0.6940

19, 19 19, 20

0.0022 0.0064 0.0154 0.0349 0.0683 0.1256 0.2044 0.3127 0.4331 0.5669 0.6873 0.0015 0.0044 0.0109 0.0255 0.0516 0.0981 0.1650 0.2610 0.3729 0.5033 0.6271

20, 20

0.0009 0.0029 0.0075 0.0182 0.0380 0.0748 0.1301 0.2130 0.3143 0.4381 0.5619

c 2000 by Chapman & Hall/CRC 

Distribution of total number of runs v in samples of size (m, n) m, n v = 22 11, 11 11, 12 1.0000 11, 13 1.0000 11, 14 1.0000 11, 15 1.0000 11, 16 0.9999 11, 17 0.9998 11, 18 0.9996 11, 19 0.9994 11, 20 0.9991

23

24

25

1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000

12, 12, 12, 12, 12, 12, 12, 12, 12,

12 13 14 15 16 17 18 19 20

1.0000 1.0000 0.9999 0.9997 0.9993 0.9987 0.9978 0.9966 0.9950

13, 13, 13, 13, 13, 13, 13, 13,

13 14 15 16 17 18 19 20

14, 14, 14, 14, 14, 14, 14,

26

27

28

29

1.0000 1.0000 1.0000 0.9999 0.9998 0.9996 0.9994 0.9991

1.0000 1.0000 1.0000 1.0000 0.9999 0.9999 0.9998

1.0000 1.0000 1.0000 1.0000 1.0000

0.9998 0.9995 0.9988 0.9975 0.9957 0.9930 0.9894 0.9848

1.0000 0.9999 0.9997 0.9994 0.9989 0.9981 0.9969 0.9954

1.0000 1.0000 1.0000 0.9999 0.9997 0.9995 0.9991 0.9986

1.0000 1.0000 1.0000 1.0000 0.9999 0.9999 0.9998

1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000

14 15 16 17 18 19 20

0.9985 0.9967 0.9938 0.9894 0.9834 0.9756 0.9660

0.9996 0.9991 0.9981 0.9965 0.9941 0.9909 0.9867

0.9999 0.9998 0.9995 0.9990 0.9982 0.9970 0.9952

1.0000 1.0000 0.9999 0.9998 0.9996 0.9992 0.9987

1.0000 1.0000 1.0000 0.9999 0.9998 0.9997

1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 1.0000

15, 15, 15, 15, 15, 15,

15 16 17 18 19 20

0.9930 0.9872 0.9789 0.9678 0.9540 0.9375

0.9977 0.9954 0.9918 0.9866 0.9798 0.9712

0.9994 0.9987 0.9974 0.9953 0.9923 0.9881

0.9999 0.9997 0.9992 0.9985 0.9975 0.9959

1.0000 0.9999 0.9998 0.9996 0.9993 0.9987

1.0000 1.0000 1.0000 0.9999 0.9998 0.9997

1.0000 1.0000 1.0000 1.0000 1.0000 1.0000 0.9999 1.0000

16, 16, 16, 16, 16,

16 17 18 19 20

0.9772 0.9634 0.9457 0.9244 0.8996

0.9908 0.9840 0.9747 0.9626 0.9479

0.9970 0.9942 0.9900 0.9840 0.9761

0.9991 0.9981 0.9964 0.9938 0.9903

0.9998 0.9995 0.9989 0.9980 0.9965

1.0000 0.9999 0.9997 0.9994 0.9989

1.0000 1.0000 0.9999 0.9999 0.9997

1.0000 1.0000 1.0000 0.9999

17, 17, 17, 17,

17 18 19 20

0.9428 0.9172 0.8872 0.8534

0.9728 0.9578 0.9391 0.9168

0.9891 0.9816 0.9714 0.9584

0.9959 0.9925 0.9876 0.9808

0.9988 0.9975 0.9954 0.9924

0.9997 0.9992 0.9985 0.9972

0.9999 0.9998 0.9996 0.9992

1.0000 1.0000 0.9999 0.9998

18, 18 18, 19 18, 20

0.8829 0.9360 0.9697 0.9866 0.9950 0.9983 0.9995 0.9999 0.8438 0.9094 0.9540 0.9781 0.9911 0.9966 0.9990 0.9997 0.8010 0.8788 0.9345 0.9670 0.9856 0.9941 0.9980 0.9994

19, 19 19, 20

0.7956 0.8744 0.9317 0.9651 0.9846 0.9936 0.9978 0.9993 0.7444 0.8350 0.9048 0.9484 0.9756 0.9891 0.9959 0.9985

20, 20

0.6857 0.7870 0.8699 0.9252 0.9620 0.9818 0.9925 0.9971

c 2000 by Chapman & Hall/CRC 

Distribution of total number of runs v in samples of size (m, n) m, n v = 30 11, 11 11, 12 11, 13 11, 14 11, 15 11, 16 11, 17 11, 18 11, 19 11, 20

31

12, 12, 12, 12, 12, 12, 12, 12, 12,

12 13 14 15 16 17 18 19 20

13, 13, 13, 13, 13, 13, 13, 13,

13 14 15 16 17 18 19 20

14, 14, 14, 14, 14, 14, 14,

14 15 16 17 18 19 20

15, 15, 15, 15, 15, 15,

15 16 17 18 19 20

16, 16, 16, 16, 16,

16 17 18 19 20

1.0000 1.0000 1.0000 1.0000

17, 17, 17, 17,

17 18 19 20

1.0000 1.0000 1.0000 1.0000 0.9999 1.0000

32

33

34

18, 18 18, 19 18, 20

1.0000 1.0000 0.9999 1.0000 1.0000 0.9998 1.0000 1.0000

19, 19 19, 20

0.9998 1.0000 1.0000 0.9996 0.9999 1.0000 1.0000

20, 20

0.9991 0.9997 0.9999 1.0000 1.0000

c 2000 by Chapman & Hall/CRC 

35

36

37

The values listed in the previous tables indicate the probability that v or fewer runs will occur. For example, for two samples of size 4, the probability of three or fewer runs is 0.114. For sample size m = n, and m larger than 10, the following table can be used. The columns headed 0.5, 1, 2.5, and 5 give values of v such that v or fewer runs occur with probability less than the indicated percentage. For example, for m = n = 12, the probability of 8 or fewer runs is approximately 5%. The columns headed 95, 97.5, 99, and 99.5 give values of v for which the probability of v or more runs is less than 5, 2.5, 1, or 0.5 percent. Distribution of the total number of runs v in samples of size m = n. m=n 11 12 13 14 15 16 17 18 19 20 25 30 35 40 45 50 55 60 65 70 75 80 85 90 95 100

0.5 5 6 7 7 8 9 10 11 11 12 16 20 24 29 33 37 42 46 50 55 59 64 68 73 77 82

1.0 6 7 7 8 9 10 10 11 12 13 17 21 25 30 34 38 43 47 52 56 61 65 70 74 79 84

2.5 7 7 8 9 10 11 11 12 13 14 18 23 27 31 36 40 45 49 54 58 63 68 72 77 82 86

5.0 7 8 9 10 11 11 12 13 14 15 19 24 28 33 37 42 46 51 56 60 65 70 74 79 84 88

c 2000 by Chapman & Hall/CRC 

95.0 16 17 18 19 20 22 23 24 25 26 32 37 43 48 54 59 65 70 75 81 86 91 97 102 107 113

97.5 16 18 19 20 21 22 24 25 26 27 33 38 44 50 55 61 66 72 77 83 88 93 99 104 109 115

99.0 17 18 20 21 22 23 25 26 27 28 34 40 46 51 57 63 68 74 79 85 90 96 101 107 112 117

99.5 mean var (σ 2 ) s.d. (σ) 18 12 5.24 2.29 19 13 5.74 2.40 20 14 6.24 2.50 22 15 6.74 2.60 23 16 7.24 2.69 24 17 7.74 2.78 25 18 8.24 2.87 26 19 8.74 2.96 28 20 9.24 3.04 29 21 9.74 3.12 35 26 12.24 3.50 41 31 14.75 3.84 47 36 17.25 4.15 52 41 19.75 4.44 58 46 22.25 4.72 64 51 24.75 4.97 69 56 27.25 5.22 75 61 29.75 5.45 81 66 32.25 5.68 86 71 34.75 5.89 92 76 37.25 6.10 97 81 39.75 6.30 103 86 42.25 6.50 108 91 44.75 6.69 114 96 47.25 6.87 119 101 49.75 7.05

14.6

THE SIGN TEST

Assumptions: Let X1 , X2 , . . . , Xn be a random sample from a continuous distribution. Hypothesis test: H0 : µ ˜=µ ˜0 Ha : µ ˜>µ ˜0 ,

µ ˜<µ ˜0 ,

µ ˜ = µ ˜0

TS: Y = the number of Xi ’s greater than µ ˜0 . Under the null hypothesis, Y has a binomial distribution with parameters n and p = .5. RR: Y ≥ c1 ,

Y ≤ c2 ,

Y ≥ c or Y ≤ n − c

The critical values c1 , c2 , and c are obtained from the binomial distribution with parameters n and p = .5 to yield the desired significance level α. (See the table on page 366.) Sample values equal to µ ˜0 are excluded from the analysis and the sample size is reduced accordingly. The normal approximation: When n ≥ 10 and np ≥ 5 the binomial distribution can be approximated by a normal distribution with µY = np

and σY2 = np(1 − p)

(14.14)

The random variable Z=

Y − µY Y − np = σY np(1 − p)

(14.15)

has approximately a standard normal distribution when H0 is true and n ≥ 10 and np ≥ 5. 14.6.1

Table of critical values for the sign test

Let X1 , X2 , . . . , Xn be a random sample from a continuous distribution with hypothesized median µ ˜0 . The test statistic is Y , the number of Xi ’s greater than µ ˜0 . If the null hypothesis is true, the probability Xi is greater than the median is 1/2. The probability distribution for Y is given by the binomial probability function    n n 1 Prob [Y = y] = f (y) = . (14.16) y 2 The following table contains critical values k such that k    n  n 1 α Prob [Y ≤ k] = < . y 2 2 y=0

c 2000 by Chapman & Hall/CRC 

(14.17)

For a one-tailed test with significance level α, enter the table in the column headed by 2α. For larger values of n, approximate critical values may be found using equation (14.15). 1 2  k = np + np(1 − p)zα/2 (14.18) where zα is the critical value for the normal distribution. Critical values for the sign test n 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25

14.7

.01 0 0 0 0 0 0 0 0 0 0 0 1 1 1 2 2 2 3 3 3 4 4 4 5 5

probability .025 .05 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 1 1 1 1 1 1 2 2 2 2 2 2 3 3 3 3 4 3 4 4 4 4 5 4 5 5 5 5 6 6 6 6 7

.10 0 0 0 0 0 0 0 1 1 1 2 2 3 3 3 4 4 5 5 5 6 6 7 7 7

n 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50

.01 6 6 6 7 7 7 8 8 9 9 9 10 10 11 11 11 12 12 13 13 13 14 14 15 15

probability .025 .05 6 7 7 7 7 8 8 8 8 9 8 9 9 9 9 10 10 10 10 11 10 11 11 12 11 12 12 12 12 13 12 13 13 14 13 14 14 15 14 15 14 15 15 16 15 16 16 17 16 17

.10 8 8 9 9 10 10 10 11 11 12 12 13 13 13 14 14 15 15 16 16 16 17 17 18 18

SPEARMAN’S RANK CORRELATION COEFFICIENT

Suppose there are n pairs of observations from continuous distributions. Rank the observations in the two samples separately from smallest to largest. Equal observations are assigned the mean rank for their positions. Let ui be the rank of the ith observation in the first sample and let vi be the rank of the ith observation in the second sample. Spearman’s rank correlation coefficient, rS , is a measure of the correlation between ranks, calculated by using the ranks in place of the actual observations in the formula for the correlation coefficient r.

c 2000 by Chapman & Hall/CRC 



n 

rS = √

n 



n 



ui vi − ui vi n SSuv i=1 i=1 i=1 = F% G  n 2 & % n  n 2 & SSuu SSvv n G     2 2 H n n u − ui v − vi i=1

6 =1−

n 

i

i=1

i=1

i

i=1

d2i

i=1 n(n2 −

1)

where di = ui − vi .

(14.19)

The shortcut formula for rS that only uses the differences {di } is not exact when there are tied measurements. The approximation is good when the number of ties is small in comparison to n. Hypothesis test: H0 : ρS = 0 (no population correlation between ranks) Ha : ρS > 0,

ρS < 0,

ρS = 0

TS: rS RR: rS ≥ rS,α ,

rS ≤ −rS,α ,

|rS | ≥ rS,α/2

where rS,α is a critical value for Spearman’s rank correlation coefficient test (see page 367). The normal approximation: When H0 is true rS has approximately a normal distribution with 1 µrS = 0 and σr2S = . (14.20) n−1 The random variable √ rS − µrS rS − 0 Z= = √ (14.21) = rS n − 1 σrS 1/ n − 1 has approximately a standard normal distribution as n increases. 14.7.1

Tables for Spearman’s rank correlation coefficient

Spearman’s coefficient of rank correlation, rS , measures the correspondence between two rankings; see equation (14.19). The table below gives critical values for rS assuming the samples are independent; their derivation comes from the subsequent table.

c 2000 by Chapman & Hall/CRC 

Critical values of Spearman’s rank correlation coefficient n 4 5 6 7 8 9 10 11 12 13 14 15 20 25 30

α = 0.10 0.8000 0.7000 0.6000 0.5357 0.5000 0.4667 0.4424 0.4182 0.3986 0.3791 0.3626 0.3500 0.2977 0.2646 0.2400

α = 0.05 0.8000 0.8000 0.7714 0.6786 0.6190 0.5833 0.5515 0.5273 0.4965 0.4780 0.4593 0.4429 0.3789 0.3362 0.3059

α = 0.01 − 0.9000 0.8857 0.8571 0.8095 0.7667 0.7333 0.7000 0.6713 0.6429 0.6220 0.6000 0.5203 0.4654 0.4251

α = 0.001 − − − 0.9643 0.9286 0.9000 0.8667 0.8364 0.8182 0.7912 0.7670 0.7464 0.6586 0.5962 0.5479

 Let m represents the mean value Then  the following  2of the sum of squares.  tables give the probability that d ≥ S for S ≥ m, or that d2 ≤ S for  S ≤ m. The tables for n = 9 and n = 10 can be completed by symmetry. The values in the next table create the critical values in the last table. For example, taking n = 9 we note that (a) S = 26 (corresponding to a Spearman 26 rank correlation coefficient of 1 − 120 ≈ 0.7833) has a probability of 0.0086; and (b) S = 28 (corresponding a Spearman rank correlation coefficient of 28 1 − 120 ≈ 0.7667) has a probability of 0.0107. Hence, the critical value for n = 9 and α = 0.01, the least value of the coefficient whose probability is less that 0.01, is 0.7667.

c 2000 by Chapman & Hall/CRC 

Exact values for Spearman’s rank correlation coefficient 3

4

5

6

7

8

9

10

4

10

20

35

56

84

120

165

0.0083 0.0417 0.0667 0.1167 0.1750 0.2250 0.2583 0.3417 0.3917 0.4750 0.5250

0.0014 0.0083 0.0167 0.0292 0.0514 0.0681 0.0875 0.1208 0.1486 0.1778 0.2097

0.0002 0.0014 0.0034 0.0062 0.0119 0.0171 0.0240 0.0331 0.0440 0.0548 0.0694

0.0000 0.0002 0.0006 0.0011 0.0023 0.0036 0.0054 0.0077 0.0109 0.0140 0.0184

0.0000 0.0000 0.0001 0.0002 0.0004 0.0007 0.0010 0.0015 0.0023 0.0030 0.0041

0.0000 0.0000 0.0000 0.0000 0.0001 0.0001 0.0002 0.0003 0.0004 0.0006 0.0008

0.4750 0.3917 0.3417 0.2583 0.2250 0.1750 0.1167 0.0667 0.0417 0.0083

0.2486 0.2819 0.3292 0.3569 0.4014 0.4597 0.5000 0.5000 0.5000 0.4597

0.0833 0.1000 0.1179 0.1333 0.1512 0.1768 0.1978 0.2222 0.2488 0.2780

0.0229 0.0288 0.0347 0.0415 0.0481 0.0575 0.0661 0.0756 0.0855 0.0983

0.0054 0.0069 0.0086 0.0107 0.0127 0.0156 0.0184 0.0216 0.0252 0.0294

0.0011 0.0014 0.0019 0.0024 0.0029 0.0036 0.0044 0.0053 0.0063 0.0075

42 44 46 48 50 52 54 56 58 60

0.4014 0.3569 0.3292 0.2819 0.2486 0.2097 0.1778 0.1486 0.1208 0.0875

0.2974 0.3308 0.3565 0.3913 0.4198 0.4532 0.4817 0.5183 0.4817 0.4532

0.1081 0.1215 0.1337 0.1496 0.1634 0.1799 0.1947 0.2139 0.2309 0.2504

0.0333 0.0380 0.0429 0.0484 0.0540 0.0603 0.0664 0.0738 0.0809 0.0888

0.0087 0.0101 0.0117 0.0134 0.0153 0.0173 0.0195 0.0219 0.0245 0.0272

62 64 66 68 70 72 74 76 78 80

0.0681 0.0514 0.0292 0.0167 0.0083 0.0014

0.4198 0.3913 0.3565 0.3308 0.2974 0.2780 0.2488 0.2222 0.1978 0.1768

0.2682 0.2911 0.3095 0.3323 0.3517 0.3760 0.3965 0.4201 0.4410 0.4674

0.0969 0.1063 0.1149 0.1250 0.1348 0.1456 0.1563 0.1681 0.1793 0.1927

0.0302 0.0334 0.0367 0.0403 0.0441 0.0481 0.0524 0.0569 0.0616 0.0667

S 0 2 4 6 8 10 12 14 16 18 20

n=2  m=1

0.5000 0.1667 0.0417 0.5000 0.5000 0.1667 0.5000 0.2083 0.5000 0.3750 0.1667 0.4583 0.5417 0.4583 0.3750 0.2083 0.1667 0.0417

22 24 26 28 30 32 34 36 38 40

c 2000 by Chapman & Hall/CRC 

Exact values for Spearman’s rank correlation coefficient 3

4

5

6

7

8

9

10

4

10

20

35

56

84

120

165

80 82 84 86 88 90 92 94 96 98 100

0.1768 0.1512 0.1333 0.1179 0.1000 0.0833 0.0694 0.0548 0.0440 0.0331 0.0240

0.4674 0.4884 0.5116 0.4884 0.4674 0.4410 0.4201 0.3965 0.3760 0.3517 0.3323

0.1927 0.2050 0.2183 0.2315 0.2467 0.2603 0.2759 0.2905 0.3067 0.3218 0.3389

0.0667 0.0720 0.0774 0.0831 0.0893 0.0956 0.1022 0.1091 0.1163 0.1237 0.1316

102 104 106 108 110 112 114 116 118 120

0.0171 0.0119 0.0062 0.0034 0.0014 0.0002

0.3095 0.2911 0.2682 0.2504 0.2309 0.2139 0.1947 0.1799 0.1634 0.1496

0.3540 0.3718 0.3878 0.4050 0.4216 0.4400 0.4558 0.4742 0.4908 0.5092

0.1394 0.1478 0.1564 0.1652 0.1744 0.1839 0.1935 0.2035 0.2135 0.2241

122 124 126 128 130 132 134 136 138 140

0.1337 0.1215 0.1081 0.0983 0.0855 0.0756 0.0661 0.0575 0.0481 0.0415

0.4908 0.4742 0.4558 0.4400 0.4216 0.4050 0.3878 0.3718 0.3540 0.3389

0.2349 0.2459 0.2567 0.2683 0.2801 0.2918 0.3037 0.3161 0.3284 0.3410

142 144 146 148 150 152 154 156 158 160

0.0347 0.0288 0.0229 0.0184 0.0140 0.0109 0.0077 0.0054 0.0036 0.0023

0.3218 0.3067 0.2905 0.2759 0.2603 0.2467 0.2315 0.2183 0.2050 0.1927

0.3536 0.3665 0.3795 0.3925 0.4056 0.4191 0.4326 0.4458 0.4592 0.4730

S

n=2  m=1

c 2000 by Chapman & Hall/CRC 

14.8

WILCOXON MATCHED-PAIRS SIGNED-RANKS TEST

Assume we have a matched set of n observations {xi , yi }. Let di denote the differences di = xi − yi . Hypothesis test: H0 : there is no difference in the distribution of the xi ’s and the yi ’s Ha : there is a difference Rank all of the di ’s without regard to sign: the least value of |di | gets rank 1, the next largest value gets rank 2, etc. After determining the ranking, affix the signs of the differences to each rank. TS: T = the smaller sum of the like-signed ranks. RR: T ≥ c where c is found from the table on page 372. Example 14.72 : Suppose n = 10 values are as shown in the first two columns of the following table: xi 9 2 1 4 6 4 7 8 5 1

yi 8 2 3 2 3 0 4 5 4 0

di = xi − yi 1 0 −2 2 3 4 3 3 1 1

rank of |di | 2 – 4.5 4.5 7 9 7 7 2 2

signed rank of |di | 2 – −4.5 4.5 7 9 7 7 2 2  + R = 40.5  − R = −4.5

The subsequent columns show the differences, the ranks (note how ties are handled), and the signed ranks. The smaller of the two sums is T = 4.5. From the following table (with n = 10) we conclude that there is evidence of a difference in distributions at the .005 significance level.

See D. J. Sheskin, Handbook of Parametric and Nonparametric Statistical Procedures, CRC Press LLC, Boca Raton, FL, 1997, pages 291–301, 681.

c 2000 by Chapman & Hall/CRC 

Critical values for the Wilcoxon signed-ranks test and the matched-pairs signed-ranks test One sided Two sided n = 5 α = .05 α = .10 0 α = .025 α = .05 α = .01 α = .02 α = .005 α = .01

6 2 0

7 3 2 0

8 5 3 1 0

9 8 5 3 1

10 10 8 5 3

11 13 10 7 5

12 17 13 9 7

13 21 17 12 9

14 25 21 15 12

One sided Two sided n = 15 α = .05 α = .10 30 α = .025 α = .05 25 α = .01 α = .02 19 α = .005 α = .01 15

16 35 29 23 19

17 41 34 27 23

18 47 40 32 27

19 53 46 37 32

20 60 52 43 37

21 67 58 49 42

22 75 65 55 48

23 83 73 62 54

24 91 81 69 61

One sided Two sided n = 25 α = .05 α = .10 100 α = .025 α = .05 89 α = .01 α = .02 76 α = .005 α = .01 68

26 110 98 84 75

27 119 107 92 83

28 130 116 101 91

29 140 126 110 100

30 151 137 120 109

31 163 147 130 118

32 175 159 140 128

33 187 170 151 138

34 200 182 162 148

14.9

WILCOXON RANK–SUM (MANN–WHITNEY) TEST

Assumptions: Let X1 , X2 , . . . , Xm and Y1 , Y2 , . . . , Yn (with m ≤ n) be independent random samples from continuous distributions. Hypothesis test: H0 : µ ˜1 − µ ˜ 2 = ∆0 Ha : µ ˜1 − µ ˜ 2 > ∆0 ,

µ ˜1 − µ ˜ 2 < ∆0 ,

µ ˜1 − µ ˜2 = ∆0

Subtract ∆0 from each Xi . Combine the (Xi − ∆0 )’s and the Yj ’s into one sample and rank all of the observations. Equal differences are assigned the mean rank for their positions. m  TS: W = Ri i=1

where Ri is the rank of (Xi − ∆0 ) in the combined sample. RR: W ≥ c1 ,

W ≤ c2 ,

W ≥ c or W ≤ m(m + n + 1) − c

where c1 , c2 , and c are critical values for the Wilcoxon rank–sum statistic such that Prob [W ≥ c1 ] ≈ α, Prob [W ≤ c2 ] ≈ α, and Prob [W ≥ c] ≈ α/2. (In practice, we convert from W to U via equation (14.24) and look up U values.) The normal approximation: When both m and n are greater than 8, W has approximately a normal distribution with µW =

m(m + n + 1) 2

c 2000 by Chapman & Hall/CRC 

2 and σW =

mn(m + n + 1) . 12

(14.22)

The random variable W − µW σW has approximately a standard normal distribution. Z=

(14.23)

The Mann–Whitney U statistic: The rank–sum test may also be based on the test statistic m(m + 2n + 1) U= − W. (14.24) 2 When both m and n are greater than 8, U has approximately a normal distribution with mn mn(m + n + 1) 2 µU = and σU . (14.25) = 2 12 The random variable U − µU Z= (14.26) σU has approximately a standard normal distribution. Note that there are two tests commonly called the Mann–Whitney U test: one developed by Mann and Whitney and one developed by Wilcoxon. Although they employ different equations and different tables, the two versions yield comparable results. Example 14.73 : The Pennsylvania State Police theorize that cars travel faster during the evening rush hour versus the morning rush hour. Randomly selected cars were selected during each rush hour and there speeds were computed using radar. The data is given in the table below. Morning:

68 63

65 73

80 75

61 71

64

64

Evening:

70 75

70 74

71 81

72 72

72 74

71 71

Use the Mann–Whitney U test to determine if there is any evidence to suggest the median speeds are different. Use α = .05. Solution: (S1) Computations: m = 10, n = 12, W = 88, U = 87 (S2) Using the tables, the critical value for a two sided test with α = .05 is 29. (S3) The value of the test statistic is not in the rejection region. There is no evidence to suggest the median speeds are different.

14.9.1

Tables for Wilcoxon (Mann–Whitney) U statistic

Given two sample of sizes m and n (with m ≤ n) the Mann–Whitney U statistic (see equation (14.24)) is used to test the hypothesis that the two c 2000 by Chapman & Hall/CRC 

samples are from populations with the same median. Rank all of the observations in ascending order of magnitude. Let W be the sum of the ranks assigned to the sample of size m. Then U is defined as U=

m(m + 2n + 1) −W 2

(14.27)

The following tables present cumulative probability and are used to determine exact probabilities associated with this test statistic. If the null hypothesis is true, the body of the tables contains probabilities such that Prob [U ≤ u]. Only small values of u are shown in the tables since the probability distribution for U is symmetric. For example, for n = 3 and m = 2 the probability distribution of U values is: Prob [U = 0] = Prob [U = 1] = Prob [U = 5] = Prob [U = 6] = 0.1 Prob [U = 2] = Prob [U = 3] = Prob [U = 4] = 0.2 so that the distribution function is: Prob [U ≤ 0] = 0.1,

Prob [U ≤ 1] = 0.2,

Prob [U ≤ 2] = 0.4,

Prob [U ≤ 3] = 0.6, Prob [U ≤ 5] = 0.9,

Prob [U ≤ 4] = 0.8, Prob [U ≤ 6] = 1

Example 14.74 : Consider the two samples {13, 9} (m = 2) and {12, 16, 14} (n = 3). Arrange the combined samples in rank order and box the values from the first sample:

rank

9 1

12 2

13 3

14 4

16 5

(14.28)

Compute the U statistic: (a) W = 1 + 3 = 4 (b) U =

2(2 + 2 · 3 + 1 −4=5 2

Using the tables below (and the comment above): Prob [U ≤ 5] = .9. There is little evidence to suggest the medians are different.

u 0 1 2 3 4 5

n=3 m=1 2 0.250 0.100 0.500 0.200 0.750 0.400 0.600

c 2000 by Chapman & Hall/CRC 

3 0.050 0.100 0.200 0.350 0.500 0.650

u 0 1 2 3 4 5 6 7 8

m=1 0.200 0.400 0.600

n=4 2 0.067 0.133 0.267 0.400 0.600

3 0.029 0.057 0.114 0.200 0.314 0.429 0.571

4 0.014 0.029 0.057 0.100 0.171 0.243 0.343 0.443 0.557

u 0 1 2 3 4 5 6 7 8 9 10 11 12 13

u 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18

m=1 0.167 0.333 0.500 0.667

m=1 0.143 0.286 0.429 0.571

n=5 2 3 0.048 0.018 0.095 0.036 0.190 0.071 0.286 0.125 0.429 0.196 0.571 0.286 0.393 0.500 0.607

2 0.036 0.071 0.143 0.214 0.321 0.429 0.571

c 2000 by Chapman & Hall/CRC 

n=6 3 0.012 0.024 0.048 0.083 0.131 0.190 0.274 0.357 0.452 0.548

4 0.008 0.016 0.032 0.056 0.095 0.143 0.206 0.278 0.365 0.452 0.548

4 0.005 0.010 0.019 0.033 0.057 0.086 0.129 0.176 0.238 0.305 0.381 0.457 0.543

5 0.004 0.008 0.016 0.028 0.048 0.075 0.111 0.155 0.210 0.274 0.345 0.421 0.500 0.579

5 0.002 0.004 0.009 0.015 0.026 0.041 0.063 0.089 0.123 0.165 0.214 0.268 0.331 0.396 0.465 0.535

6 0.001 0.002 0.004 0.008 0.013 0.021 0.032 0.047 0.066 0.090 0.120 0.155 0.197 0.242 0.294 0.350 0.409 0.469 0.531

u 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25

m=1 0.125 0.250 0.375 0.500 0.625

2 0.028 0.056 0.111 0.167 0.250 0.333 0.444 0.556

c 2000 by Chapman & Hall/CRC 

n=7 3 4 0.008 0.003 0.017 0.006 0.033 0.012 0.058 0.021 0.092 0.036 0.133 0.055 0.192 0.082 0.258 0.115 0.333 0.158 0.417 0.206 0.500 0.264 0.583 0.324 0.394 0.464 0.536

5 0.001 0.003 0.005 0.009 0.015 0.024 0.037 0.053 0.074 0.101 0.134 0.172 0.216 0.265 0.319 0.378 0.438 0.500 0.562

6 0.001 0.001 0.002 0.004 0.007 0.011 0.017 0.026 0.037 0.051 0.069 0.090 0.117 0.147 0.183 0.223 0.267 0.314 0.365 0.418 0.473 0.527

7 0.000 0.001 0.001 0.002 0.003 0.006 0.009 0.013 0.019 0.027 0.036 0.049 0.064 0.082 0.104 0.130 0.159 0.191 0.228 0.267 0.310 0.355 0.402 0.451 0.500 0.549

u 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32

14.9.2

m=1 0.111 0.222 0.333 0.444 0.556

2 0.022 0.044 0.089 0.133 0.200 0.267 0.356 0.444 0.556

3 0.006 0.012 0.024 0.042 0.067 0.097 0.139 0.188 0.248 0.315 0.388 0.461 0.539

n=8 4 0.002 0.004 0.008 0.014 0.024 0.036 0.055 0.077 0.107 0.141 0.184 0.230 0.285 0.341 0.404 0.467 0.533

5 0.001 0.002 0.003 0.005 0.009 0.015 0.023 0.033 0.047 0.064 0.085 0.111 0.142 0.177 0.218 0.262 0.311 0.362 0.416 0.472 0.528

6 0.000 0.001 0.001 0.002 0.004 0.006 0.010 0.015 0.021 0.030 0.041 0.054 0.071 0.091 0.114 0.141 0.172 0.207 0.245 0.286 0.331 0.377 0.426 0.475 0.525

7 0.000 0.000 0.001 0.001 0.002 0.003 0.005 0.007 0.010 0.014 0.020 0.027 0.036 0.047 0.060 0.076 0.095 0.116 0.140 0.168 0.198 0.232 0.268 0.306 0.347 0.389 0.433 0.478 0.522

8 0.000 0.000 0.000 0.001 0.001 0.001 0.002 0.003 0.005 0.007 0.010 0.014 0.019 0.025 0.032 0.041 0.052 0.065 0.080 0.097 0.117 0.139 0.164 0.191 0.221 0.253 0.287 0.323 0.360 0.399 0.439 0.480 0.520

Critical values for Wilcoxon (Mann–Whitney) statistic

The following tables give critical values for U for significance levels of 0.00005, 0.0001, 0.005, 0.01, 0.025, 0.05, and 0.10 for a one-tailed test. For a two-tailed test, the significance levels are doubled. If an observed U is equal to or less than the tabular value, the null hypothesis may be rejected at the level of significance indicated at the head of the table.

c 2000 by Chapman & Hall/CRC 

Critical values of U in the Mann–Whitney test Critical values for the α = 0.10 level of significance m 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 8 19 20

n = 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

0 0 1 1 1 2 2 3 3 4 4 5 5 5 6 6 7 7

0 1 1 2 3 4 5 5 6 7 8 9 10 10 11 12 13 14 15

0 1 3 4 5 6 7 9 10 11 12 13 15 16 17 18 20 21 22

1 2 4 5 7 8 10 12 13 15 17 18 20 22 23 25 27 28 30

1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 34 36 38

1 4 6 8 11 13 16 18 21 23 26 28 31 33 36 38 41 43 46

2 5 7 10 13 16 19 22 24 27 30 33 36 39 42 45 48 51 54

2 5 9 12 15 18 22 25 28 31 35 38 41 45 48 52 55 58 62

3 6 10 13 17 21 24 28 32 36 39 43 47 51 54 58 62 66 70

3 7 11 15 19 23 27 31 36 40 44 48 52 57 61 65 69 73 78

4 8 12 17 21 26 30 35 39 44 49 53 58 63 67 72 77 81 86

4 9 13 18 23 28 33 38 43 48 53 58 63 68 74 79 84 89 94

5 10 15 20 25 31 36 41 47 52 58 63 69 74 80 85 91 97 102

5 10 16 22 27 33 39 45 51 57 63 68 74 80 86 92 98 104 110

5 11 17 23 29 36 42 48 54 61 67 74 80 86 93 99 106 112 119

6 12 18 25 31 38 45 52 58 65 72 79 85 92 99 106 113 120 127

6 13 20 27 34 41 48 55 62 69 77 84 91 98 106 113 120 128 135

7 14 21 28 36 43 51 58 66 73 81 89 97 104 112 120 128 135 143

7 15 22 30 38 46 54 62 70 78 86 94 102 110 119 127 135 143 151

Critical values of U in the Mann–Whitney test Critical values for the α = 0.05 level of significance m 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

n = 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

0 0 0 1 1 1 1 2 2 3 3 3 3 4 4 4

0 0 1 2 2 3 4 4 5 5 6 7 7 8 9 9 10 11

0 1 2 3 4 5 6 7 8 9 10 11 12 14 15 16 17 18

0 1 2 4 5 6 8 9 11 12 13 15 16 18 19 20 22 23 25

0 2 3 5 7 8 10 12 14 16 17 19 21 23 25 26 28 30 32

0 2 4 6 8 11 13 15 17 19 21 24 26 28 30 33 35 37 39

c 2000 by Chapman & Hall/CRC 

1 3 5 8 10 13 15 18 20 23 26 28 31 33 36 39 41 44 47

1 4 6 9 12 15 18 21 24 27 30 33 36 39 42 45 48 51 54

1 4 7 11 14 17 20 24 27 31 34 37 41 44 48 51 55 58 62

1 5 8 12 16 19 23 27 31 34 38 42 46 50 54 57 61 65 69

2 5 9 13 17 21 26 30 34 38 42 47 51 55 60 64 68 72 77

2 6 10 15 19 24 28 33 37 42 47 51 56 61 65 70 75 80 84

3 7 11 16 21 26 31 36 41 46 51 56 61 66 71 77 82 87 92

3 7 12 18 23 28 33 39 44 50 55 61 66 72 77 83 88 94 100

3 8 14 19 25 30 36 42 48 54 60 65 71 77 83 89 95 101 107

3 9 15 20 26 33 39 45 51 57 64 70 77 83 89 96 102 109 115

4 9 16 22 28 35 41 48 55 61 68 75 82 88 95 102 109 116 123

4 10 17 23 30 37 44 51 58 65 72 80 87 94 101 109 116 123 130

4 11 18 25 32 39 47 54 62 69 77 84 92 100 107 115 123 130 138

Critical values of U in the Mann–Whitney test Critical values for the α = 0.025 level of significance m 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

n = 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

0 0 0 0 1 1 1 1 1 2 2 2 2

0 1 1 2 2 3 3 4 4 5 5 6 6 7 7 8

0 1 2 3 4 4 5 6 7 8 9 10 11 11 12 13 14

0 1 2 3 5 6 7 8 9 11 12 13 14 15 17 18 19 20

1 2 3 5 6 8 10 11 13 14 16 17 19 21 22 24 25 27

1 3 5 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34

0 2 4 6 8 10 13 15 17 19 22 24 26 29 31 34 36 38 41

0 2 4 7 10 12 15 17 20 23 26 28 31 34 37 39 42 45 48

0 3 5 8 11 14 17 20 23 26 29 33 36 39 42 45 48 52 55

0 3 6 9 13 16 19 23 26 30 33 37 40 44 47 51 55 58 62

1 4 7 11 14 18 22 26 29 33 37 41 45 49 53 57 61 65 69

1 4 8 12 16 20 24 28 33 37 41 45 50 54 59 63 67 72 76

1 5 9 13 17 22 26 31 36 40 45 50 55 59 64 69 74 78 83

1 5 10 14 19 24 29 34 39 44 49 54 59 64 70 75 80 85 90

1 6 11 15 21 26 31 37 42 47 53 59 64 70 75 81 86 92 98

2 6 11 17 22 28 34 39 45 51 57 63 69 75 81 87 93 99 105

2 7 12 18 24 30 36 42 48 55 61 67 74 80 86 93 99 106 112

2 7 13 19 25 32 38 45 52 58 65 72 78 85 92 99 106 113 119

2 8 14 20 27 34 41 48 55 62 69 76 83 90 98 105 112 119 127

Critical values of U in the Mann–Whitney test Critical values for the α = 0.01 level of significance m 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

n = 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

0 0 0 0 0 0 1 1

0 0 1 1 1 2 2 2 3 3 4 4 4 5

0 1 1 2 3 3 4 5 5 6 7 7 8 9 9 10

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16

1 2 3 4 6 7 8 9 11 12 13 15 16 18 19 20 22

c 2000 by Chapman & Hall/CRC 

0 1 3 4 6 7 9 11 12 14 16 17 19 21 23 24 26 28

0 2 4 6 7 9 11 13 15 17 20 22 24 26 28 30 32 34

1 3 5 7 9 11 14 16 18 21 23 26 28 31 33 36 38 40

1 3 6 8 11 13 16 19 22 24 27 30 33 36 38 41 44 47

1 4 7 9 12 15 18 22 25 28 31 34 37 41 44 47 50 53

2 5 8 11 14 17 21 24 28 31 35 38 42 46 49 53 56 60

0 2 5 9 12 16 20 23 27 31 35 39 43 47 51 55 59 63 67

0 2 6 10 13 17 22 26 30 34 38 43 47 51 56 60 65 69 73

0 3 7 11 15 19 24 28 33 37 42 47 51 56 61 66 70 75 80

0 3 7 12 16 21 26 31 36 41 46 51 56 61 66 71 76 82 87

0 4 8 13 18 23 28 33 38 44 49 55 60 66 71 77 82 88 93

0 4 9 14 19 24 30 36 41 47 53 59 65 70 76 82 88 94 100

1 4 9 15 20 26 32 38 44 50 56 63 69 75 82 88 94 101 107

1 5 10 16 22 28 34 40 47 53 60 67 73 80 87 93 100 107 114

Critical values of U in the Mann–Whitney test Critical values for the α = 0.005 level of significance m 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

n = 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

0 0 0 1 1 1 2 2 2 2 03 03

0 0 1 1 2 2 3 3 4 5 5 6 6 7 8

0 1 1 2 3 4 5 6 7 7 8 9 10 11 12 13

0 1 2 3 4 5 6 7 9 10 11 12 13 15 16 17 18

0 1 3 4 6 7 9 10 12 13 15 16 18 19 21 22 24

1 2 4 6 7 9 11 13 15 17 18 20 22 24 26 28 30

0 1 3 5 7 9 11 13 16 18 20 22 24 27 29 31 33 36

0 2 4 6 9 11 13 16 18 21 24 26 29 31 34 37 39 42

0 2 5 7 10 13 16 18 21 24 27 30 33 36 39 42 45 48

1 3 6 9 12 15 18 21 24 27 31 34 37 41 44 47 51 54

1 3 7 10 13 17 20 24 27 31 34 38 42 45 49 53 57 60

1 4 7 11 15 18 22 26 30 34 38 42 46 50 54 58 63 67

2 5 8 12 16 20 24 29 33 37 42 46 51 55 60 64 69 73

2 5 9 13 18 22 27 31 36 41 45 50 55 60 65 70 74 79

2 6 10 15 19 24 29 34 39 44 49 54 60 65 70 75 81 86

2 6 11 16 21 26 31 37 42 47 53 58 64 70 75 81 87 92

0 3 7 12 17 22 28 33 39 45 51 57 63 69 74 81 87 93 99

0 3 8 13 18 24 30 36 42 48 54 60 67 73 79 86 92 99 105

Critical values of U in the Mann–Whitney test Critical values for the α = 0.001 level of significance m 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

n = 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

0 0 0 0

0 0 0 1 1 1 2 2 3 3 3

0 1 1 2 2 3 3 4 5 5 6 7 7

0 1 2 3 4 4 5 6 7 8 9 10 11 12

c 2000 by Chapman & Hall/CRC 

0 1 2 3 5 6 7 8 9 10 11 13 14 15 16

0 1 2 4 5 6 8 9 11 12 14 15 17 18 20 21

1 2 3 5 7 8 10 12 14 15 17 19 21 23 25 26

0 1 3 5 6 8 10 12 14 17 19 21 23 25 27 29 32

0 2 4 6 8 10 12 15 17 20 22 24 27 29 32 34 37

0 2 4 7 9 12 14 17 20 23 25 28 31 34 37 40 42

1 3 5 8 11 14 17 20 23 26 29 32 35 38 42 45 48

1 3 6 9 12 15 19 22 25 29 32 36 39 43 46 50 54

1 4 7 10 14 17 21 24 28 32 36 40 43 47 51 55 59

2 5 8 11 15 19 23 27 31 35 39 43 48 52 56 60 65

0 2 5 9 13 17 21 25 29 34 38 43 47 52 57 61 66 70

0 3 6 10 14 18 23 27 32 37 42 46 51 56 61 66 71 76

0 3 7 11 15 20 25 29 34 40 45 50 55 60 66 71 77 82

0 3 7 12 16 21 26 32 37 42 48 54 59 65 70 76 82 88

Critical values of U in the Mann–Whitney test Critical values for the α = 0.0005 level of significance m 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

14.10

n = 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20

0 0 0 1 1 1 2 2

0 0 1 1 2 2 3 3 4 4 5 5

0 1 2 2 3 4 5 5 6 7 8 8 9

0 1 2 3 4 5 6 7 8 9 10 11 13 14

0 1 2 4 5 6 7 9 10 11 13 14 15 17 18

0 1 2 4 5 7 8 10 11 13 15 16 18 20 21 23

0 2 3 5 7 8 10 12 14 16 18 20 22 24 26 28

1 2 4 6 8 10 12 15 17 19 21 24 26 28 31 33

1 3 5 7 10 12 15 17 20 22 25 27 30 33 35 38

0 2 4 6 9 11 14 17 20 23 25 28 31 34 37 40 43

0 2 5 7 10 13 16 19 22 25 29 32 35 39 42 45 49

0 3 5 8 11 15 18 21 25 28 32 36 39 43 46 50 54

1 3 6 9 13 16 20 24 27 31 35 39 43 47 51 55 59

1 4 7 10 14 18 22 26 30 34 39 43 47 51 56 60 65

1 4 8 11 15 20 24 28 33 37 42 46 51 56 61 65 70

2 5 8 13 17 21 26 31 35 40 45 50 55 60 65 70 76

2 5 9 14 18 23 28 33 38 43 49 54 59 65 70 76 81

WILCOXON SIGNED-RANK TEST

Assumptions: Let X1 , X2 , . . . , Xn be a random sample from a continuous symmetric distribution. Hypothesis test: H0 : µ ˜=µ ˜0 Ha : µ ˜>µ ˜0 ,

µ ˜<µ ˜0 ,

µ ˜ = µ ˜0

Rank the absolute differences |X1 − µ ˜0 |, |X2 − µ ˜0 |, . . . , |Xn − µ ˜0 |. Equal absolute differences are assigned the mean rank for their positions. TS: T+ = the sum of the ranks corresponding to the positive differences ˜0 ). (Xi − µ RR: T+ ≥ c1 ,

T+ ≤ c2 ,

T+ ≥ c or T+ ≤ n(n + 1) − c

where c1 , c2 , and c are critical values for the Wilcoxon signed-rank statistic (see table on page 372) such that Prob [T+ ≥ c1 ] ≈ α, Prob [T+ ≤ c2 ] ≈ α, and Prob [T+ ≥ c] ≈ α/2. Any observed difference (xi − µ ˜0 ) = 0 is excluded from the test and the sample size is reduced accordingly.

c 2000 by Chapman & Hall/CRC 

The normal approximation: When n ≥ 20, T+ has approximately a normal distribution with n(n + 1) n(n + 1)(2n + 1) µT+ = and σT2 + = . (14.29) 4 24 The random variable T+ − µT+ Z= (14.30) σT+ has approximately a standard normal distribution when H0 is true. See D. J. Sheskin, Handbook of Parametric and Nonparametric Statistical Procedures, CRC Press LLC, Boca Raton, FL, 1997, pages 83–94.

c 2000 by Chapman & Hall/CRC 

CHAPTER 15

Quality Control and Risk Analysis Contents 15.1

Quality assurance 15.1.1 Control charts 15.1.2 Abnormal distributions of points 15.2 Acceptance sampling 15.2.1 Sequential sampling 15.3 Reliability 15.3.1 Failure time distributions 15.4 Risk analysis and decision rules

15.1 15.1.1

QUALITY ASSURANCE Control charts Expression x chart R chart σ chart p chart c chart LCL UCL

Meaning control chart for means control chart for sample ranges control chart for standard deviations fraction defective chart number of defects chart lower control limit upper control limit

Suppose the population mean µ and the population standard deviation σ are unknown. For k samples of size n let xi and Ri be the sample mean and sample range for the ith sample, respectively. The average mean x and the average range R are defined by x=

k 1 xi k i=1

R=

k 1 Ri k i=1

(15.1)

Suppose the population proportion p or the number of defects is unknown. For k samples assume sample i has ni items, ci defects, and ei defective items. c 2000 by Chapman & Hall/CRC 

Note that a single defective item may have many defects. Define p and c by p=

e1 + e2 + · · · + ek n1 + n 2 + · · · + nk

Type of chart central line lower control limit (LCL) x chart (µ, σ known) µ µ − Aσ x chart (µ, σ unknown) x x − A2 R R chart (σ known) d2 σ D1 σ R chart (σ unknown) R D3 R p chart (based on past data)    p(1−p) p min 0, p − 3 n

c=

k 1 ci k i=1

(15.2)

upper control limit (UCL) µ + Aσ x + A2 R D2 σ D4 R  p + 3 p(1−p) n

number-of-defectives chart   np np − 3 np(1 − p) np + 3 np(1 − p) c chart √ √ c c−3 c c+3 c √ where n is the sample size, A = 3/ n, D1 = d2 − 3d3 , D2 = d2 + 3d3 , D3 = D1 /d2 , and D4 = D2 /d2 . Values of {A, A2 , d2 , d3 , D1 , D2 , D3 , D4 } are given in Table 15.11 . 1 This table reproduced, by permission, from ASTM Manual on Quality Control of Materials, American Society for Testing and Materials, Philadelphia, PA, 1951.

c 2000 by Chapman & Hall/CRC 

 c 2000 by Chapman & Hall/CRC

Chart for averages A A1 A2 2.121 3.760 1.880 1.732 2.394 1.023 1.500 1.880 0.729 1.342 1.596 0.577

Chart for standard deviations c2 B1 B2 B3 B4 0.5642 0 1.843 0 3.267 0.7236 0 1.858 0 2.568 0.7979 0 1.808 0 2.266 0.8407 0 1.756 0 2.089

d2 1.128 1.693 2.059 2.326

6 7 8 9 10

1.225 1.134 1.061 1.000 0.949

1.410 1.277 1.175 1.094 1.028

0.483 0.419 0.373 0.337 0.308

0.8686 0.8882 0.9027 0.9139 0.9227

0.026 0.105 0.167 0.219 0.262

1.711 1.672 1.638 1.609 1.584

0.030 0.118 0.185 0.239 0.284

1.970 1.882 1.815 1.761 1.716

2.534 2.704 2.847 2.970 3.078

0 0.205 0.387 0.546 0.687

5.078 5.203 5.307 5.394 5.469

0 0.076 0.136 0.184 0.223

2.004 1.924 1.864 1.816 1.777

11 12 13 14 15

0.905 0.866 0.832 0.802 0.775

0.973 0.925 0.884 0.848 0.816

0.285 0.266 0.249 0.235 0.223

0.9300 0.9359 0.9410 0.9453 0.9490

0.299 0.331 0.359 0.384 0.406

1.561 1.541 1.523 1.507 1.492

0.321 0.354 0.382 0.406 0.428

1.679 1.646 1.618 1.594 1.572

3.173 3.258 3.336 3.407 3.472

0.812 0.924 1.026 1.121 1.207

5.534 5.592 5.646 5.693 5.737

0.256 0.284 0.308 0.329 0.348

1.744 1.716 1.692 1.671 1.652

n 2 3 4 5

Chart for ranges D1 D2 D3 0 3.686 0 0 4.358 0 0 4.698 0 0 4.918 0

Table 15.1: Parameter values for control charts (reproduced by permission)

D4 3.267 2.575 2.282 2.115

15.1.2

Abnormal distributions of points in control charts

Abnormality Sequence

Description Seven or more consecutive points on one side of the center line. Denotes the average value has shifted. Fewer than seven consecutive points on one side of the center line, but most of the points are on that side.

Bias

• • • • Trend Approaching the limit

Periodicity

15.2

10 12 14 16

of 11 consecutive points or more of 14 consecutive points or more of 17 consecutive points or more of 20 consecutive points

Seven or more consecutive rising or falling points. Two out of three or three or more out of seven consecutive points are more than two thirds the distance between the center line and a control limit. The data points vary in a regular periodic pattern.

ACCEPTANCE SAMPLING

Expression AQL AOQ AOQL LTPD producer’s risk consumer’s risk

Meaning acceptable quality level average outgoing quality average outgoing quality limit (maximum value of AOQ for varying incoming quality) lot tolerance percent defective Type I error (percentage of “good” lots rejected) Type II error (percentage of “bad” lots accepted)

Military standard 105 D is a widely used sampling plan. There are three general levels of inspection corresponding to different consumer’s risks. (Inspection level II is usually chosen; level I uses smaller sample sizes and level II uses larger sample sizes.) There are also three types of inspections: normal, tightened, and reduced. Tables are available for single, double, and multiple sampling. To use MIL-STD-105 D for single sampling, determine the sample size code letter from Table 15.2. Using this sample size code letter find the sample size and the acceptance and rejection numbers from Table 15.3.

c 2000 by Chapman & Hall/CRC 

Lot or batch size 2 9 16 26 51 91 151 281 501 1,201 3,201 10,001 35,001 150,001 500,001

to to to to to to to to to to to to to to and

8 15 25 50 90 150 280 500 1,200 3,200 10,000 35,000 150,000 500,000 over

general I A A B C C D E F G H J K L M N

inspection levels II III A B B C C D D E E F F G G H H J J K K L L M M N N P P Q Q R

Table 15.2: Sample size code letters for MIL-STD-105 D Example 15.75 : Suppose that MIL-STD-105D is to be used with incoming lots of 1,000 items, inspection level II is to be used in conjunction with normal inspection, and an AQL of 2.5 percent is desired. How should the inspections be carried out? Solution: (S1) From table 15.2 the sample size code letter is J. (S2) From the table on page 388, for column J, the lot size should be 80. Using the row labeled 2.5 the acceptance number is 5 and the rejection number is 6. (S3) Thus, if a single sample of size 80 (selected randomly from each lot of 1,000 items) contains 5 or fewer defectives then the lot is to be accepted. If it contains 6 or more defectives, then the lot is to be rejected.

15.2.1

Sequential sampling

We need to inspect a lot. Suppose the two sequences of numbers {an } (for accept) and {rn } (for reject) are given. Elements from the lot are sequentially taken (n = 1, 2, . . . ); after each element is selected the total number of defective elements is determined. The lot is accepted after n elements if the number of defectives is less than or equal to an . The lot is rejected after n elements if the number of defectives is greater than or equal to rn . Sampling continues as long as the number of defectives in the sample of size n falls between an and rn .

c 2000 by Chapman & Hall/CRC 

AQL = Acceptable quality level (normal inspection). Ac|Re = Accept if Ac or fewer are found, reject if Re or more are found. ← = Use first sampling procedure to left. → = Use first sampling procedure to right. If sample size equals, or exceeds, lot or batch size, do 100 percent inspection.

Table 15.3: Master table for single sampling inspection (normal inspection) MIL-STD-105 D.

c 2000 by Chapman & Hall/CRC 

B 3 → → → → → → → → → → → → → 0|1 ← → 1|2 2|3 3|4 5|6 7|8 10|11 14|15 21|22 30|31 44|45

A 2 → → → → → → → → → → → → → → 0|1 → → 1|2 2|3 3|4 5|6 7|8 10|11 14|15 21|22 30|31

AQL 0.010 0.015 0.025 0.040 0.065 0.10 0.15 0.25 0.40 0.65 1.0 1.5 2.5 4.0 6.5 10 15 25 40 65 100 150 250 400 650 1000

→ → → → → → → → → → → → 0|1 ← → 1|2 2|3 3|4 5|6 7|8 10|11 14|15 21|22 30|31 44|45 ←

C 5 → → → → → → → → → → → 0|1 ← → 1|2 2|3 3|4 5|6 7|8 10|11 14|15 21|22 30|31 44|45 ← ←

D 8 → → → → → → → → → → 0|1 ← → 1|2 2|3 3|4 5|6 7|8 10|11 14|15 21|22 30|31 44|45 ← ← ←

→ → → → → → → → → 0|1 ← → 1|2 2|3 3|4 5|6 7|8 10|11 14|15 21|22 ← ← ← ← ← ←

F 20 → → → → → → → → 0|1 ← → 1|2 2|3 3|4 5|6 7|8 10|11 14|15 21|22 ← ← ← ← ← ← ←

G 32 → → → → → → → 0|1 ← → 1|2 2|3 3|4 5|6 7|8 10|11 14|15 21|22 ← ← ← ← ← ← ← ←

H 50 → → → → → → 0|1 ← → 1|2 2|3 3|4 5|6 7|8 10|11 14|15 21|22 ← ← ← ← ← ← ← ← ←

J 80 → → → → → 0|1 ← → 1|2 2|3 3|4 5|6 7|8 10|11 14|15 21|22 ← ← ← ← ← ← ← ← ← ←

K 125 → → → → 0|1 ← → 1|2 2|3 3|4 5|6 7|8 10|11 14|15 21|22 ← ← ← ← ← ← ← ← ← ← ←

L 200 → → → 0|1 ← → 1|2 2|3 3|4 5|6 7|8 10|11 14|15 21|22 ← ← ← ← ← ← ← ← ← ← ← ←

M 315

Sample size code letter and sample size E 13 → → 0|1 ← → 1|2 2|3 3|4 5|6 7|8 10|11 14|15 21|22 ← ← ← ← ← ← ← ← ← ← ← ← ←

N 500

→ 0|1 ← → 1|2 2|3 3|4 5|6 7|8 10|11 14|15 21|22 ← ← ← ← ← ← ← ← ← ← ← ← ← ←

0|1 ← → 1|2 2|3 3|4 5|6 7|8 10|11 14|15 21|22 ← ← ← ← ← ← ← ← ← ← ← ← ← ← ←

← ← 1|2 2|3 3|4 5|6 7|8 10|11 14|15 21|22 ← ← ← ← ← ← ← ← ← ← ← ← ← ← ← ←

P Q R 800 1250 2000

A sequential sampling plan with AQL= p0 , LTPD= p1 , producer’s risk α, and consumer’s risk β is given by      1−p0   log β  1−α + n log 1−p1       an =  1 log pp10 − log 1−p 1−p0     (15.3)  1−β 0 log α + n log 1−p 1−p1      rn =    p1 1−p1   log p − log 1−p 0

15.2.1.1

0

Sequential probability ratio tests

A sequential probability ratio test is a sequential sampling strategy in which the ratio of probabilities (based on the hypotheses) is used. Given two simple hypotheses (H0 and H1 ) and m observations, compute (a) P0m = Prob (observations | H0 ) (b) P1m = Prob (observations | H1 ) (c) vm = P1m /P0m and then follow the decision rule given by: 1−β (a) If vm ≥ then reject H0 . α β (b) If vm ≤ then reject H1 . 1−α β 1−β (c) If < vm < then make another observation. 1−α α Note that the number of samples taken is not fixed a priori, but is determined as sampling occurs. Example 15.76 : Lot acceptance Let θ denote the fraction of defective items. Two simple hypotheses are H0 : θ = θ0 = 0.05 and H1 : θ = θ1 = 0.15. Choose α = .05 and β = .10 (i.e., reject lot with θ = θ0 about 5% of the time, accept lot with θ = θ1 about 10% of the time). If, after m observations, there are d defective items, then  d  m−d m d θ1 1 − θ1 Pim = θi (1 − θi )m−d and vm = θ0 1 − θ0 d β or vm = 3d (0.895)m−d using the above numbers. The critical values are 1−α = 0.105 1−β and α = 18. The decision to perform another observation depends on whether or not

0.105 ≤ 3d (0.895)m−d ≤ 18 Taking logarithms, a (m − d, d) control chart can be drawn with the the following lines: d = 0.101(m − d) − 2.049 and d = 0.101(m − d) + 2.63. On the figure below, a sample path leading to rejection of H0 has been indicated: c 2000 by Chapman & Hall/CRC 

d

✻ reject H0

(defectives)

decision boundaries

✑ ✑

























✑✑ ✑





✑ ✑

sample



path

reject H1









m − d (non-defectives)

15.2.1.2

Two-sided mean test

Let X be normally distributed with unknown mean µ and known standard deviation σ. Consider the two simple hypotheses H0 : µ = µ0 and H1 : µ = µ1 . If Y is the sum of the first m observations of X, then a (Y, m) control chart is constructed with the two lines: µ0 + µ1 β σ2 Y = log m+ 2 µ1 − µ0 1−α (15.4) 1−β µ0 + µ1 σ2 log Y = m+ 2 µ1 − µ0 α 15.2.1.3

One-sided variance test

Consider testing the null hypothesis that σ = σ1 against the alternative that σ = σ2 > σ1 . If α is the assigned risk of rejecting the null hypothesis when σ equals σ1 and β is the assigned risk of accepting the null hypothesis when σ equal σ2 , then the sequential plan is as follows: 1. Define the quantities



1 g = 0.43429 − σ12 1−β a = log10 α 1−α b = log10 β log10 (σ22 /σ12 ) s= g

1 σ22



2a g 2b h1 = g

h2 =

(15.5)

2. Define the acceptance and rejection lines as (n)

= −h1 + s(n − 1)

(lower)

(n)

=

h2 + s(n − 1)

(upper)

Z1 Z2

The sequential test is carried out as follows: c 2000 by Chapman & Hall/CRC 

(15.6)

1. Let n stand for the number of sample items inspected 2. Let Z stand for the sum of squared deviations from the sample mean:  n 2 n   2 xi − xi n n  2 i=1 Z= (xi − x) = i=1 (15.7) n i=1 3. Test against limits (n)

(a) If Z < Z1 then accept the null hypothesis. (n) (b) If Z > Z2 then reject the null hypothesis. (c) If neither inequality is true, then take another sample. 15.3

RELIABILITY

1. The reliability of a product is the probability that the product will function within specified limits for at least a specified period of time. 2. A series system is one in which the entire system will fail if any of its components fail. 3. A parallel system is one in which the entire system will fail only if all of its components fail. 4. Let Ri denote the reliability of the ith component. 5. Let Rs denote the reliability of a series system. 6. Let Rp denote the reliability of a parallel system. The product law of reliabilities states Rs =

n )

Ri

(15.8)

i=1

The product law of unreliabilities states Rp = 1 −

n )

(1 − Ri )

(15.9)

i=1

15.3.1

Failure time distributions

1. Let the probability of a component failing between times t and t + ∆t be f (t)∆t. 2. The probability that a component will fail on the interval from 0 to t is  t F (t) = f (x) dx (15.10) 0

3. The reliability function is the probability that a component survives to time t R(t) = 1 − F (t)

c 2000 by Chapman & Hall/CRC 

(15.11)

4. The instantaneous failure rate, Z(t), is the average rate of failure in the interval from t to t + ∆t, given that the component survived to time t Z(t) =

f (t) f (t) = R(t) 1 − F (t)

(15.12)

Note the relationships: R(t) = e−

Jt 0

f (t) = Z(t)e−

Z(x) dx

Jt 0

Z(x) dx

(15.13)

β

If f (t) = αβtβ−1 eαt with α > 0 and β > 0, the probability distribution function for a Weibull random variable, then the failure rate is Z(t) = β αβtβ−1 and R(t) = e−αt . Note that failure rate decreases with time if β < 1 and increases with time if β > 1.

Example 15.77 :

15.3.1.1

Use of the exponential distribution

If the failure rate is a constant Z(t) = α (with α > 0) then f (t) = αe−αt (for t > 0) which is the probability density function for an exponential random variable. If a failed component is replaced with another having the same constant failure rate α, then the occurrence of failures is a Poisson process. The constant 1/α is called the mean time between failures (MTBF). The reliability function is R(t) = e−αt . If a series system has n components, each with constant failure rate {αi }, then

n  Rs (t) = exp − (15.14) αi i=1

The MTBF for the series system is µs µs =

1 µ1

+

1 µ2

1 + ··· +

1 µn

(15.15)

If a parallel system has n components, each with identical constant failure rate α, then the MTBF for the parallel system is µp   1 1 1 µp = 1 + + ··· + (15.16) α 2 n 15.4

RISK ANALYSIS AND DECISION RULES

Suppose knowledge of a specific state of a system is desired, and those states can be delineated as {θ1 , θ2 , . . . }. For example, in a weather application, the states might be θ1 for rain and θ2 for no rain. Decision rules are actions that may be taken based on the state of a system. For example, in deciding whether to go on a trip, there are the decision rules: stay home, go with an umbrella, and go without an umbrella.

c 2000 by Chapman & Hall/CRC 

Possible actions Stay home Go without an umbrella Go with an umbrella

a1 a2 a3

System state θ1 (rain) θ2 (no rain) 4 4 5 0 2 5

Table 15.4: An example loss function =(θ, a)

A loss function is an arbitrary function that depends on a specific state and a decision rule. For example, consider the loss function =(θ, a) given in Table 15.4. It is possible to determine the “best” decision, under different models, even without obtaining any data. • Minimax principle With this principle one should expect and prepare for the worst. That is, for each action it is possible to determine the minimum possible loss that may be incurred. This loss is assigned to each action; the action with the smallest (or minimum) maximum loss is the action chosen. For the data in Table 15.4 the maximum loss is 4 for action a1 and 5 for either of the actions a2 or a3 . Under a minimax principle, the chosen action would be a1 and the minimax loss would be 4. • Minimax principle for mixed actions It is possible to minimize the maximum loss when the action taken is a statistical distribution, p, of actions. Assume that action ai is taken with probability pi (with p1 +p2 +p3 = 1). Then the expected loss L(θi ) is given by L(θi ) = Ea [=(θi , a)] = p1 =(θi , a1 ) + p2 =(θi , a2 ) + p3 =(θi , a3 ). The data in Table 15.4 result in the following expected losses:         L(θ1 ) 4 5 2 + p2 + p3 (15.17) = p1 4 0 5 L(θ2 ) It can be shown that the minimax point of this mixed action case has to satisfy L(θ1 ) = L(θ2 ). Solving equation (15.17) with this constraint leads to 5p2 = 3p3 . Using this and p1 + p2 + p3 = 1 in equation (15.17) results in L(θ1 ) = L(θ2 ) = 4 − 7p3 /5. This indicates that p3 should be as large as possible. Hence, the maximum value is obtained by the mixed distribution p = ( 08 , 38 , 58 ). Hence, if action a2 is chosen 3/8’s of the time, and action a3 is chosen 5/8’s of the time, then the minimax loss is equal to L = 25/8. This is a smaller loss than using a pure strategy of only choosing a single action.

c 2000 by Chapman & Hall/CRC 

• Bayes actions If the probability distribution of the states {θ1 , θ2 , . . . } is given by the density function g(θi ), then the loss has  a known distribution with an expectation of B(a) = Ei [=(θi , a)] = i g(θi )=(θi , a). This quantity is known as the Bayes loss for action a. A Bayes action is an action that minimizes the Bayes loss. For example, assuming that the prior distribution is given by g(θ1 ) = 0.4 and g(θ2 ) = 0.6, then B(a1 ) = 4, B(a2 ) = 2, and B(a3 ) = 3.8. This leads to the choice of action a2 . A course of action can also be based on data about the states of interest. For example, a weather report Z will give data for the predictions of rain and no rain. Continuing the example, assume that the correctness of these predictions is given as follows: Predict rain Predict no rain

z1 z2

θ1 (rain) 0.8 0.2

θ2 (no rain) 0.1 0.9

That is, when it will rain, then the prediction is correct 80% of the time. A decision function is an assignment of data to actions. Since there are finitely many possible actions and finitely many possible values of Z, the number of decision functions is finite. In the example there are 32 = 9 possible decision functions, {d1 , d2 , . . . , d9 }; they are defined to be:

Predict z1 , take action Predict z2 , take action

d1 a1 a1

d2 a2 a2

Decision functions d3 d4 d5 d6 d7 a3 a1 a2 a1 a3 a3 a2 a1 a3 a1

d8 a2 a3

d9 a3 a2

The risk function R(θ, di ) is the expected value of the loss when a specific decision function is being used: R(θ, di ) = EZ [=(θ, di (Z))]. It is straightforward to compute the risk function for all values of {di } and {aj }. This results in the following values: Risk function evaluation Decision Function θ1 (rain) θ2 (no rain) d1 4 4 d2 5 0 d3 2 5 d4 4.2 0.4 d5 4.8 3.6 d6 3.6 4.9 d7 2.4 4.1 d8 4.4 4.5 d9 2.6 0.5 c 2000 by Chapman & Hall/CRC 

a1 a2 a3

θ1 (rain) 2 3 0

θ2 (no rain) 4 0 5

Table 15.5: The regret function r(θ, a) corresponding to the loss function in Table 15.4

This array can now be treated as though it gave the loss function in a no– data problem. The minimax principle for mixed action results in the “best” 7 solution being rule d3 for 17 ’s of the time and rule d9 for 10 17 ’s of the time. This 40 leads to a minimax loss of 17 . Before the data Z is received, the minimax loss 25 40 105 was 25 8 . Hence, the data Z is “worth” 8 − 17 = 136 in using the minimax approach. The regret function (also called the opportunity loss function) r(θ, a) is the loss, =(θ, a), minus the minimum loss for the given θ: r(θ, a) = =(θ, a) − minb =(θ, b). For each state, the least loss is determined if that state were known to be true. This is the contribution to loss that even a good decision cannot avoid. The quantity r(θ, a) represents the loss that could have been avoided had the state been known—hence the term regret. For the loss function example in Table 15.4, the minimum loss for θ = θ1 is 2, and the minimum loss for θ = θ2 is 0. Hence, the regret function is as given in Table 15.5. Most of the computations performed for a loss function could also be performed with the risk function. If the minimax principle is used to determine the “best” action, then, in this example, the “best” action is a2 .

c 2000 by Chapman & Hall/CRC 

CHAPTER 16

General Linear Models Contents 16.1 16.2

Notation The general linear model 16.2.1 The simple linear regression model 16.2.2 Multiple linear regression 16.2.3 One-way analysis of variance 16.2.4 Two-way analysis of variance 16.2.5 Analysis of covariance 16.3 Summary of rules for matrix operations 16.3.1 Linear combinations 16.3.2 Determinants 16.3.3 Inverse of a partitioned matrix 16.3.4 Eigenvalues 16.3.5 Differentiation involving vectors/matrices 16.3.6 Additional definitions and properties 16.4 Quadratic forms 16.4.1 Multivariate distributions 16.4.2 The principle of least squares 16.4.3 Minimum variance unbiased estimates 16.5 General linear hypothesis of full rank 16.5.1 Notation 16.5.2 Simple linear regression 16.5.3 Analysis of variance, one-way anova 16.5.4 Multiple linear regression 16.5.5 Randomized blocks (one observation per cell) 16.5.6 Quadratic form due to hypothesis 16.5.7 Sum of squares due to error 16.5.8 Summary 16.5.9 Computation procedure for hypothesis testing 16.5.10 Regression significance test 16.5.11 Alternate form of the distribution 16.6 General linear model of less than full rank 16.6.1 Estimable function and estimability 16.6.2 Linear hypothesis model of less than full rank 16.6.3 Constraints and conditions

c 2000 by Chapman & Hall/CRC 

16.1

NOTATION

In this chapter, matrices are denoted by bold–face capital letters; for example, if the matrix A has m rows and n columns, then A = Amn and   a11 a12 · · · a1n  a21 a22 · · · a2n    A= . .. . . . .  .. . ..  . am1 am2 · · · amn In general, column vectors will be denoted by lower–case bold–face letters. For example, xT = [x1 x2 · · · xn ],

βT = [β1 β2 · · · βk ].

(16.1)

If necessary, the number of rows in a column vector is indicated with a subscript, for example, xn has n rows. (1) Some special column vectors are T (vector of treatment totals), B (vector of block totals), 1 (vector of all ones), and 0 (vector of all zeros). (2) 1T A is a row vector whose entries are the column sums of A, and A1 denotes a column vector whose entries are the row sums of A. 1T A1 denotes the sum of all the elements in the matrix A. (3) AT denotes the transpose of A. (4) (A)ij = aij denotes the element in the ith row and the j th column of A. (5) The identity matrix is denoted by I. The order of the identity matrix may be indicated by a subscript, for example, In denotes an n × n identity matrix. (6) Dx denotes a diagonal matrix with entries x1 , x2 , . . . , xn (the subscript indicates the terms in the diagonal). (7) A tilde placed above a matrix indicates the matrix is triangular. The K is a lower triangular matrix and T K T is an upper triangular matrix T matrix:     t11 0 0 · · · 0 t11 t21 t31 · · · tn1  t21 t22 0 · · · 0   0 t22 t32 · · · tn2         T t t t · · · 0 K =  31 32 33 K = T T ,  0 0 t33 · · · tn3   .. .. .. . . ..   .. .. .. . . ..  . . . . . . . .  . .  tn1 tn2 tn3 · · · tnn 16.2 16.2.1

0

0

0

· · · tnn

THE GENERAL LINEAR MODEL The simple linear regression model

Let (x1 , y1 ), (x2 , y2 ), . . . , (xn , yn ) be n pairs of observations such that yi is an observed value of the random variable Yi . Assume there exist constants β0 c 2000 by Chapman & Hall/CRC 

and β1 such that Yi = β0 + β1 xi + 3i

(16.2)

where 31 , 32 , . . . , 3n are independent, normal random variables having mean 0 and variance σ 2 . Assumptions In terms of Yi ’s

In terms of 3i ’s

3i ’s are normally distributed

Yi ’s are normally distributed

E [3i ] = 0

E [Yi ] = β0 + β1 xi

Var [3i ] = σ

2

Var [Yi ] = σ 2

Cov [3i , 3j ] = 0, i = j

Cov [Yi , Yj ] = 0, i = j

Using equation (16.2): Y1 = β0 + β1 x1 + 31 Y2 = β0 + β1 x2 + 32 .. . Yn = β0 + β1 xn + 3n

(16.3)

This set of equations may be written in matrix form:       Y1 1 x1 31  Y2   1 x2     32         Y3    β0   +  33    =  1 x3   ..   ..  ..  ..  β1  .   .  .  .  Yn 1 xn 3n Y

=

X

β

+



The matrix X is the design matrix and may also be written as X = [1, x], where 1 is a column vector containing all 1’s and x is the column vector containing the xi ’s. The simple linear regression model, equation (16.2), is often written in the form Yi = µ + β1 (xi − x) + 3i

c 2000 by Chapman & Hall/CRC 

(16.4)

where µ = β0 + β1 x. This model may also be written in matrix form:       Y1 1 (x1 − x) 31  Y2   1 (x2 − x)     32         Y3      µ +  33    =  1 (x3 − x)   ..   ..  ..   β1 ..  .   .  .   . 1

Yn Y

=

(xn − x) X

3n β

+



where X = [1, (x − x1)]. 16.2.2

Multiple linear regression

Let there be n observations of the form (x1i , x2i , . . . , xki , yi ) such that yi is an observed value of the random variable Yi . Assume there exist constants β0 , β1 , . . . , βk such that Yi = β0 + β1 x1i + · · · + βk xki + 3i

(16.5)

where 31 , 32 , . . . , 3n are independent, normal random variables having mean 0 and variance σ 2 .

In terms of 3i ’s

Assumptions In terms of Yi ’s

3i ’s are normally distributed

Yi ’s are normally distributed

E [3i ] = 0

E [Yi ] = β0 + β1 x1i + · · · + βk xki

Var [3i ] = σ

2

Var [Yi ] = σ 2

Cov [3i , 3j ] = 0, i = j

Cov [Yi , Yj ] = 0, i = j

Using equation (16.5): Y1 = β0 + β1 x11 + β2 x21 + · · · + βk xk1 + 31 Y2 = β0 + β1 x12 + β2 x22 + · · · + βk xk2 + 32 .. . Yn = β0 + β1 x1n + β2 x2n + · · · + βk xkn + 3n

(16.6)

This set of equations may be written in matrix form:         β0 31 Y1 1 x11 x21 x31 · · · xk1  32   β1   Y2     1 x21 x22 x32 · · · xk2             ..  =  .. .. .. .. . . ..   β2  +  33   ..   .   . . ..  . .   . .  .   .  Yn 1 x1n x2n x3n · · · xkn βk 3n Y

=

c 2000 by Chapman & Hall/CRC 

X

β

+



where the design matrix X = [1, Xnk ] and Xnk is the matrix of observations on the independent variables. 16.2.3

One-way analysis of variance

Let there be k treatments (or populations), independent random samples of size ni (where i = 1, 2, . . . , k), from each population, and let N = n1 + n2 + · · · + nk . Let Yij be the j th random observation in the ith treatment group. Assume a fixed effects experiment model: Yij = µ + αi + 3ij

(16.7)

where µ is the grand mean, αi is the ith treatment effect, and 3ij is the random error term. The 3ij ’s are assumed to be independent, normally distributed, with mean 0 and variance σ 2 . Using equation (16.7): Y11 = µ + α1 Y12 = µ + α1 .. .. . . Y1n1 = µ + α1 Y21 = µ Y22 = µ .. .. . . Y2n2 = µ .. .. . . Yk1 = µ Yk2 = µ .. .. . . Yknk = µ

c 2000 by Chapman & Hall/CRC 

+ 311 + 312 .. . + 31n1 + α2 + α2 .. .

+ 321 + 322 .. .

+ α2 .. .

+ 32n2 .. . + αk + 3k1 + αk + 3k2 .. .. . . + αk + 3knk

This set of equations may be written in matrix form:     Y11 1 1 0 0 ··· 0 0 Y12  1 1 0 0 · · · 0 0      ..   .. .. .. .. . . .. ..   .  . . . . . . .     Y1n1  1 1 0 0 · · · 0 0         Y21  1 0 1 0 · · · 0 0      µ  Y22  1 0 1 0 · · · 0 0      α   ..   . . . . . . .   1 α2   .  =  .. .. .. .. . . .. ..    +       Y2n2  1 0 1 0 · · · 0 0  ..  .       .  . . . . . . .  ..   .. .. .. .. .. .. ..  αk         Yk1  1 0 0 0 · · · 0 1     Yk2  1 0 0 0 · · · 0 1      .  . . . . . . .  ..   .. .. .. .. . . .. ..  1 0 0 0 ··· 0 1 Yknk Y

=

X

The design matrix X may be written as  1n1 0n1  0n 2 1 n 2  XN k =  . ..  .. .

β

  311 312     ..   .    31n1      321    322     ..   .    32n2     .   ..      3k1    3k2     .   ..  3knk

+



X = [1N , XN k ] where  0n 1 · · · 0 n 1 0n 2 · · · 0 n 2   .. ..  . . 

(16.8)

0 n k 0n k 0n k · · · 1 n k  T and the parameter vector may be written as β = µ α . 16.2.4

Two-way analysis of variance

Let Yijk be the k th random observation for the ith level of factor A and the j th level of factor B. Assume there are nij observations for the ij factor combination: i = 1, 2, . . . , a, j = 1, 2, . . . , b, k = 1, 2, . . . , nij . For simplicity, consider a fixed effects experiment model: Yijk = µ + αi + βj + (αβ)ij + 3ijk

(16.9)

where µ is the grand mean, αi is the level i factor A effect, βj is the level j factor B effect, (αβ)ij is the level ij interaction effect, and 3ijk is the random error term. The 3ijk ’s are assumed to be independent, normally distributed, with mean 0 and variance σ 2 .

c 2000 by Chapman & Hall/CRC 

Suppose a = 2 and b = 3. The model may be written in matrix form:     3111 µ  ..   .   α1     311n11   α2            y11 1n11 1 0 1 0 0 1 0 0 0 0 0  3121   β1  y12    1n 1 0 0 1 0 0 1 0 0 0 0 β2  ..      12     .  y13  1n 1 0 0 0 1 0 0 1 0 0 0 β3     =  13    y21  1n 0 1 1 0 0 0 0 0 1 0 0 (αβ)11  + 312n12        21    .  y22  1n22 0 1 0 1 0 0 0 0 0 1 0 (αβ)12    ..      1n23 0 1 0 0 1 0 0 0 0 0 1  y23  (αβ)13   3231  (αβ)21       .  (αβ)22   ..  (αβ)23 323n23 Y 16.2.5

=

X

β

+



Analysis of covariance

The analysis of covariance procedure is a combination of analysis of variance and regression analysis. For example, consider a one-way classification with one independent variable. Let Yij be the j th random observation in the ith treatment group: i = 1, 2, . . . , a, j = 1, 2, . . . , ni . The model is Yij = µ + αi + βxij + 3ij

(16.10)

where µ is the grand mean, αi is the ith treatment effect, β is the regression coefficient of Y on X, and 3ij is the random error term. The 3ij ’s are assumed to be independent, normally distributed, with mean 0 and variance σ 2 . Suppose a = 3, using equation (16.10): Y11 Y12 .. .

= µ + = µ + .. .

α1 α1 .. .

+ βx11 + βx12 .. .

Y1n1 = µ + α1 Y21 .. .

+ 311 + 312 .. .

+ βx1n1 + 31n1

= µ .. .

+ α2 .. .

+ βx21 .. .

Y2n2 = µ

+ α2

+ βx2n2 + 32n2

Y31 .. .

= µ .. .

Y3n3 = µ

c 2000 by Chapman & Hall/CRC 

+ 321 .. .

+ α3 + βx31 + 331 .. .. .. . . . + α3 + βx3n3 + 33n3

If x1 , x2 , and x3 denote the observations on the independent variable in each treatment group, this set of equations may be written in matrix form:     Y11 311 Y12  312   .   .   .   .   .   .        Y1n1  31n1        µ      1n 1 1 0 0 x 1  Y21  321  α1   .  = 1n 0 1 0 x2  α2  +  .  2  .   .     .   .  1n3 0 0 1 x3 α3      Y2n2  32n2  β     Y  3   31   31   .   .   ..   ..  Y3n3 Y 16.3

33n3 =

X

β

+



SUMMARY OF RULES FOR MATRIX OPERATIONS

16.3.1

Linear combinations

Suppose X is a random vector: a vector whose elements are random variables. Let µ be the vector of means and let Σ be the variance–covariance matrix, denoted   E [X] = µ, Cov [X] = E XXT = Σ. (16.11) For any conforming matrix C, the linear combinations Y = CX have E [Y] = E [CX] = Cµ,

Cov [Y] = Cov [CX] = CΣCT .

The linear combinations Z = XT C have     E [Z] = E XT C = µT C, Cov [Z] = Cov XT C = CT ΣC. 16.3.2

(16.12)

(16.13)

Determinants and partitioning of determinants

The determinant of a square matrix X, denoted by |X| or det (X), is a scalar function of X defined as  det(X) = sgn(σ)x1,σ(1) x2,σ(2) · · · xn,σ(n) (16.14) σ

where the sum is taken over all permutations σ of {1, 2, . . . , n}. The signum function sgn(σ) is the number of successive transpositions required to change the permutation σ to the identity permutation. Note the properties of determinants: |A| |B| = |AB| and |A| = |AT |. Omitting the signum function in equation (16.14) yields the definition of the  permanent of X, given by per X = σ x1,σ(1) · · · xn,σ(n) . c 2000 by Chapman & Hall/CRC 

Suppose the matrix X can be partitioned, written as   X11 X12 X= . X21 X22 The determinant of X may be computed by $ $ $X11 X12 $ $ $ −1 −1 $ $ $ $ $X21 X22 $ = |X11 | X22 − X21 X11 X12 if X11 exists $ $ $ if X−1 exists. = |X22 | $X11 − X12 X−1 22 X21 22 16.3.3

(16.15)

(16.16)

Inverse of a partitioned matrix

Suppose the matrix X can be partitioned as in equation (16.15). The inverse of the matrix X may be written as % &−1 % & X11 X12 A B = where C D X21 X22

16.3.3.1 %

−1 A = [X11 − X12 X−1 22 X21 ]

−1 D = [X22 − X21 X−1 11 X12 ]

B = −X−1 11 X12 D

C = −X−1 22 X21 A

Symmetric case

X11 X12 XT 12 X22

%

&−1 =

A B

&

BT D

where

T −1 A = [X11 − X12 X−1 22 X12 ]

−1 −1 D = [X22 − XT 12 X11 X12 ]

−1 B = −AX12 X−1 22 = −X11 X12 D

16.3.4

Eigenvalues

If A is a k × k square matrix and I is the k × k identity matrix, then the scalers λ1 , λ2 , . . . , λk that satisfy the polynomial equation |A − λI| are the eigenvalues (or characteristic roots) of the matrix A. The equation |A − λI| is a function of λ and is the characteristic equation. Let ch(A) denote the characteristic roots of the matrix A and tr(A) denote the trace of A. (1) ch(AB) = ch(BA) except possibly for zero roots. (2) tr(AB) = tr(BA) n (3) If ch(A) = {λ}i=1 , then ch(A−1 ) = 1/λi and ch(I ± A) = 1 ± λi for i = 1, 2, . . . , n.

c 2000 by Chapman & Hall/CRC 

16.3.5 16.3.5.1

Differentiation involving vectors/matrices Definitions

(1) Let f be a real–valued function of x1 , x2 , . . . , xn . The symbol ∂f /∂x denotes a column vector whose ith element is ∂f /∂xi . (2) Let f be a real–valued function of x11 , x12 , . . . , x1n , x21 , . . . , x2n , . . . , xm1 , xm2 , . . . , xmn . The symbol ∂f /∂X denotes a matrix whose (i, j) entry is ∂f /∂xij . Note: If there are functional relationships between the elements of X (for example, in a symmetric matrix) these relationships are disregarded in the definition above. If xij = xji = yij (yij is the symbol for the distinct variable that occurs in two places in X) then ∂f /∂yij = (∂f /∂X)ij + (∂f /∂X)ij . (3) If y1 , y2 , . . . , yn are functions of x, then ∂y/∂x denotes the column vector whose ith entry is ∂yi /∂x. (4) If y11 , y12 , . . . , y1n , y21 , . . . , y2n , . . . , ym1 , . . . , ymn are functions of x, then ∂Y/∂x denotes the matrix whose (i, j) entry is ∂yij /∂x. (5) If each of the quantities y1 , y2 , . . . , yn is a function of the variables x1 , x2 , . . . , xm , then ∂yT /∂x denotes an m × n matrix whose (i, j) entry is ∂yj /∂xi . 16.3.5.2

Properties

Suppose a, b, e, x, y, and z are column vectors, and A, Q, X, and Y are matrices. ∂(xT x) = 2x (1) ∂x ∂(xT Qx) (2) = Qx + QT x ∂x (3) ∂(xT Qx)/∂x = 2Qx if Q is symmetric. (4) ∂(aT x)/∂x = a (5) ∂(aT Qx)/∂x = QT a (6) ∂ tr(AX)/∂X = AT (7) ∂ tr(XA)/∂X = AT (8) ∂ ln |X|/∂X = (XT )−1 if X is square and nonsingular. ∂y ∂zT ∂y (9) = · (Chain rule 1) ∂x ∂x ∂z ∂(xT A) (10) =A ∂x

c 2000 by Chapman & Hall/CRC 

(11) If e = B − AT x, then ∂(eT e) ∂eT ∂(eT e) = · (using property 9) ∂x ∂x ∂e = −2AT e (using properties 1 and 10) (12) If the scalar z is related to a scalar x through the variables yij , i = 1, 2, . . . , m; j = 1, 2, . . . , n, then     ∂z ∂Y ∂z ∂YT ∂z · = tr · = tr (16.17) ∂x ∂Y ∂x ∂YT ∂x This second chain rule is correct regardless of any functional relationships that may exist between the elements of Y. 16.3.6

Additional definitions and properties K K −1 is also lower triangular. (1) If T is lower triangular, then T (2) A matrix A is (a) positive definite if xT Ax > 0 for all x = 0.

(b) positive semi-definite if xT Ax ≥ 0 for all x. (3) For any matrix Q, the dimension of the row (column) space of Q is the row (column) rank of Q. (The row (column) rank of a matrix is also the number of linearly independent rows (columns) of that matrix.) The row rank and the column rank of any matrix Q are equal, and are the rank of the matrix Q. (4) If Q is a symmetric, positive-definite matrix, then there exists a unique K with positive diagonal entries such that Q = T KT K T . The real matrix T K matrix T may be obtained by using the forward Doolittle procedure: In each cycle, divide every element of the next–to–last row (the row which is immediately above the one beginning with 1) by the square–root of K T on the left and the leading (first) element. This technique produces T K −1 on the right–hand side. T (5) If Q is an n × n symmetric, positive semi-definite matrix of rank r, then K obtained via the forward Doolittle procedure will have the matrix T zeros to the right of the rth column. Q may be written as  " # K1 T K T TT Q= (16.18) T 1 2 T2 K 1 is triangular. This computational procedure is important where only T when determining characteristic roots. Suppose A and B are symmetric, and A is of rank r < n. To find the largest characteristic root of AB first obtain the representation  " # K1 T K T TT A= (16.19) T 1 2 T2 c 2000 by Chapman & Hall/CRC 

using the forward Doolittle procedure. The characteristic roots of AB may be found using " #  K  K T TT B T 1 ch(AB) = ch T (16.20) 1 2 T2 where the right–hand matrix is of small order and symmetric. 16.4

PRINCIPLE OF MINIMIZING QUADRATIC FORMS AND GAUSS MARKOV THEOREM

16.4.1

Multivariate distributions

Suppose X is a random variable with mean µ and variance σ 2 : E [X] = µ,

Var [X] = σ 2 .

(16.21)

(1) The standardized random variable Y = (X − µ)/σ has mean 0 and variance 1: E [Y ] = 0,

Var [Y ] = 1.

(16.22)

(2) If X is a normal random variable, then the random variable Z 2 = (X − µ)2 /σ 2 has a chi–square distribution with one degree of freedom. Suppose X is a random vector consisting of the p random variables {X1 , X2 , . . . , Xp }: that is XT = [X1 , X2 , . . . , Xp ]. Let µ be the vector of means and Σ be the variance–covariance matrix: E [X] = µ,

Cov [X] = Σ.

(16.23)

KA K T where A K is the lower triangular matrix obtained using the forLet Σ = A K −1 (X−µ) and note Σ−1 = (A K T )−1 A K −1 = ward Doolittle analysis. Let Y = A −1 T −1 K ) A K . (A K −1 E [X − µ] = 0. (1) E [Y] = A K −1 Cov [X − µ](A K −1 )T = A K −1 Cov [X − µ](A K T )−1 (2) Cov [Y] = A K −1 Cov [X](A K T )−1 = A KA K T (A K −1 A K T )−1 = I = A (3) The expression −1

L K −1 (X − µ) YT Y = (X − µ)T (bf A )T A = (X − µ)T Σ−1 (X − µ)

(16.24)

is the standard quadratic form. If X has a multivariate normal distribution, then YT Y has a chi–square distribution with p degrees of freedom.

c 2000 by Chapman & Hall/CRC 

16.4.2

The principle of least squares

Let Y be the random vector of responses, y be the vector of observed responses, β be the vector of regression coefficients, be the vector of random errors, and let X be the design matrix: 

         Y1 y1 β0 A1 1 x11 x21 · · · xk1  Y2   y2   β1   A2   1 x12 x22 · · · xk2            Y= .  y= .  β= .  = .  X= . . .. ..   ..   ..   ..   ..   .. .. . .  Yn yn βk An 1 x1n x2n · · · xkn

The model may now be written as Y = Xβ + where ∼ Nn (0, σ 2 In ) or equivalently Y ∼ Nn (Xβ, σ 2 In ). The sum of squared deviations about the true regression line is S(β) =

n 

[yi − (β0 + β1 x1i + · · · + βk xki )]2

i=1 T

(16.25)

= (y − β X )(y − βX) = e e T

T

T

where e is the vector of observed errors. To minimize equation (16.25): ∂(eT e) = −2XT (y − Xβ) ∂β

(16.26)

. T = (β.0 , β.1 , . . . , β.k ) that minimizes S(β) is the vector of least The vector β squares estimates. Setting equation (16.26) equal to zero:   . =0 XT y − Xβ (16.27) . = XT y. (XT X)β The normal equations are given by equation (16.27). If the matrix XT X is non–singular, then . = (XT X)−1 XT y. β 16.4.3

(16.28)

Minimum variance unbiased estimates

The minimum variance, unbiased, linear estimate of β is obtained by using a general form of the Gauss Markov theorem. Suppose Y = Xβ + ,

E [ ] = 0,

Cov [Y] = Cov [ ] = σ 2 V

(16.29)

where V is a square, symmetric, non–singular, n × n matrix with known entries. Therefore, the variance of Yi (for i = 1, 2, . . . , n) is known and Cov [Yi , Yj ], for all i = j, is known except for an arbitrary scalar multiple.

c 2000 by Chapman & Hall/CRC 

. where β . The best linear estimate of an arbitrary linear function cT β is cT β minimizes the quadratic form T V−1 . The standard quadratic form for is ( − E [ ])T (Cov [ ])−1 ( − E [ ]) = ( − 0)T (σ 2 V)−1 ( − 0) 1 = 2 T V−1 . σ

(16.30)

Minimizing the standard quadratic form is equivalent to minimizing 3T V−1 , as stated in the Gauss Markov Theorem. In this case the normal equations . = XT V−1 y. are XT V−1 Xβ 16.5

GENERAL LINEAR HYPOTHESIS OF FULL RANK

This section is concerned with the problem of testing hypotheses about certain parameters and the associated probability distributions. 16.5.1

Notation

A general null hypothesis is stated as Cβ = k where (1) C = Cnh m , (nh ≤ m), is the hypothesis matrix and is of rank nh . (2) β is an n×1 column vector of parameters as defined in the general linear model. (3) k is a vector of nh known elements, usually equal to 0. (4) nh is the number of degrees of freedom due to hypothesis; the number of rows in the hypothesis matrix C; the number of nonredundant statements in the null hypothesis. (5) ne is the number of degrees of freedom due to error and is equal to the number of observations minus the effective number of parameters. Note: A composite hypothesis should not contain: (1) contradictory statements like H0 : β1 = β2 and β1 = 2β2 simultaneously, (2) redundant statements like H0 : β1 = β2 16.5.2

and

3β1 = 3β2 .

Simple linear regression

Model : Let (x1 , y1 ), (x2 , y2 ), . . . , (xn , yn ) be n pairs of observations such that yi is an observed value of the random variable Yi . Assume there exist constants β0 and β1 such that Yi = β0 + β1 xi + 3i ,

parameter vector β = [β0 , β1 ]T

(16.31)

where 31 , 32 , . . . , 3n are independent, normal random variables having mean 0 and variance σ 2 . Examples:

c 2000 by Chapman & Hall/CRC 

(1) H0 : β0 = 0 Ha : β0 = 0 General linear hypothesis: nh = 1   β0 [1, 0] =0 β1 C

β

=0

(2) H0 : β1 = 0 Ha : β1 = 0 General linear hypothesis: nh = 1   β0 =0 [0, 1] β1 C

β

=0

(3) H0 : β0 = β1 = 0 simultaneously Ha : βi = 0 for some i General linear hypothesis: nh = 2      10 β0 0 = 01 β1 0 C

β

= 0

(4) H0 : β0 = β1 Ha : β0 = β1 General linear hypothesis: nh = 1   β0 [1, −1] =0 β1 C 16.5.3

β

=0

Analysis of variance, one-way anova

Model : Let there be k treatments, or populations, independent random samples of size ni , i = 1, 2, . . . , k, from each population, and let N = n1 + n2 + · · · + nk . Let Yij be the j th random observation in the ith treatment group. Assume a fixed effects experiment: Yij = µ + αi + 3ij ,

i = 1, 2, . . . , k,

j = 1, 2, . . . , ni

parameter vector β = [µ, α1 , α2 , . . . , αk ]T Examples: (1) H0 : α1 = α2 = · · · = αk Ha : αi = αj for some i = j General linear hypothesis: nh = k − 1

c 2000 by Chapman & Hall/CRC 

     µ 0 1 −1 0 0 ··· 0     α 1   0 1 0 −1 0 · · · 0  α2   0      1 0 0 −1 · · · 0  =  .. .. .. .. . . ..  α3   0  .  . . . . . . .   ..   ..  1 0 0 0 · · · −1 0 αk

 0 0  0   .. . 0

C(k−1)(k+1)

β

= 0k−1

(2) H0 : α1 = α2 = · · · = αk = 0 Ha : αi = 0 for at least one i General linear hypothesis: nh  0 1 0 0 0 ··· 0  0 0 1 0 0 ··· 0   0 0 0 1 0 ··· 0   0 0 0 0 1 ··· 0   .. .. .. .. .. . . .  . . . . . . .. 0

0

0

0

0

···

=k     0 µ  α1   0       α2   0       α3  =  0        ..   ..   .  . 1 0 αk

C(k)(k+1)

β

= 0k

(3) Suppose i = 1, 2, 3, 4. H0 : −α1 + 2α2 − α3 = 0 (Quadratic contrast of three effects) Ha : −α1 + 2α2 − α3 = 0 General linear hypothesis: nh = 1   µ α1     [0, −1, +2, −1, 0]  α2  = 0 α3  α4 C β =0 16.5.4

Multiple linear regression

Model : Let there be n observations of the form (x1i , x2i , . . . , xki , yi ) such that yi is an observed value of the random variable Yi . Assume there exist constants β0 , β1 , . . . , βk such that Yi = β0 + β1 x1i + · · · + βk xki + 3i parameter vector β = [β0 , β1 , β2 , . . . , βk ]T where 31 , 32 , . . . , 3n are independent, normal random variables having mean 0 and variance σ 2 .

c 2000 by Chapman & Hall/CRC 

Examples: (1) H0 : β1 = 0 Ha : β1 = 0 General linear hypothesis: nh = 1   β0 β1      [0, 1, 0, 0, · · · , 0] β2  = 0  ..  . βk C β =0 (2) H0 : β1 = β2 = β3 = · · · = βk = 0 Ha : βi = 0 for some i General linear hypothesis:  0 1 0 0 0 ···  0 0 1 0 0 ···   0 0 0 1 0 ···   0 0 0 0 1 ···   .. .. .. .. .. . .  . . . . . . 0

0

0

0

0

nh 0 0 0 0 .. .

···

C(k)(k+1)

1

=k 

   0 β0  β1   0       β2   0       β3  =  0        ..   ..  .  . βk

0

β

= 0k

(3) H0 : β1 = β2 = 0 Ha : βi = 0 for some i General linear hypothesis: nh = 2   β0    β1    0 1 0 0 0 · · · 0 β2  0  = 0 0 1 0 0 ··· 0  .  0  ..  βk C 16.5.5

β

= 0

Randomized blocks (one observation per cell)

Model : Let Yij be the random observation in the ith row and the j th column, i = 1, 2, 3 and j = 1, 2, 3, 4. Assume a fixed effects model: Yij = µ + αi + βj + 3ij parameter vector β = [µ, α1 , α2 , α3 , β1 , β2 , β3 , β4 ]T

c 2000 by Chapman & Hall/CRC 

Examples: (1) H0 : α1 = α2 = α3 Ha : αi = αj for some i = j General linear hypothesis: nh = 2



0 0

1 1

−1 0

0 −1

0 0

0 0

0 0



 µ  α1        α2     0   α3  = 0  0 0  β  1  β2     β3  β4

C

β

= 0

(2) H0 : −α1 + 2α2 − α3 = 0 (quadratic contrast) Ha : −α1 + 2α2 − α3 = 0 General linear hypothesis: nh = 1   µ  α1     α2     α3   [0, −1, 2, −1, 0, 0, 0, 0]   β1  = 0    β2     β3  β4 C 16.5.6

β

=0

Quadratic form due to hypothesis

For the general linear model Y = Xβ + ,

E [Y] = Xβ,

Var [Y] = σ 2 I

(16.32)

the normal equations are given by . = XT Y. (XT X)β

(16.33)

T

If the model is of full rank ((X X) has an inverse) then the estimate of β is . = (XT X)−1 XT Y β

c 2000 by Chapman & Hall/CRC 

(16.34)

. is given by and the variance–covariance matrix of β " # . = (XT X)−1 Var[XT Y](XT X)−1 Var β = (XT X)−1 XT Var [Y]X(XT X)−1 = σ 2 (XT X)−1 XT X(XT X)−1 = σ 2 (XT X)−1

(16.35)

. is an unbiased estimate of Cβ If the null hypothesis is H0 : Cβ = 0 then Cβ and " # . = Cβ = 0 E Cβ (16.36) " # . = C Var[β]C . T Var Cβ = σ 2 C(XT X)−1 CT , "

##−1 1 . Var Cβ = 2 [C(XT X)−1 CT ]−1 . σ Under the null hypothesis, the standard quadratic form is

(16.37)

"

(16.38)

1 .T T . (16.39) β C [C(XT X)−1 CT ]−1 Cβ. σ2 The expression in equation (16.39) is the sum of squares due to hypothesis (denoted SSH). If Y has a multivariate normal distribution, then SSH/σ 2 has a chi–square distribution with nh degrees of freedom. SSH =

16.5.7

Sum of squares due to error

. the error of For the general linear model Y = Xβ + let e = Y − Xβ, estimation. The sum of squares due to error is given by SSE =

n 

e2i = eT e

i=1

. T (Y − Xβ) . = (Y − Xβ) T

T

. XT Y + β . XT Xβ . = Y T Y − 2β . T XT Y + β . T XT X(XT X)−1 XT Y = Y T Y − 2β . T XT Y + β . T IXT Y = Y T Y − 2β . T XT Y. = YT Y − β

(16.40)

Thus, SSE is obtained by computing the sum of squares of all observations . and (YT Y) and subtracting the scalar product of the vector of estimates of β the vector on the right–hand side of the normal equations.

c 2000 by Chapman & Hall/CRC 

The sum of squares due to error may depend only on the model, and is determined once the model is stated. SSE is independent of any hypothesis which may be stated or tested. If Y has a multivariate normal distribution, then SSE/σ 2 has a chi–square distribution with ne degrees of freedom and is independent of any SSH. 16.5.8

Summary

For the general linear model Y = Xβ + ,

E [Y] = Xβ

(16.41)

suppose the model is of full rank (XT X is non–singular and thus has an inverse). If Var[Y] = σ 2 I

(homoscedasticity and independence)

(16.42)

then the normal equations are given by . T = XT Y (XT X)β

(16.43)

. = (XT X)−1 XT Y. β

(16.44)

and the estimate of β is

If the elements of Y are normally distributed, then the following hypothesis test may be conducted: H0 : Cβ = 0 Ha : Cβ = d ( = 0)

(16.45)

This hypothesis matrix has nh rows. If H0 is consistent and contains no redundancies then nh is the degrees of freedom due to hypothesis. 16.5.9

Computation procedure for hypothesis testing

A procedure for testing a hypothesis in a general linear model (equation (16.45)): (1) Obtain the sum of squares due to hypothesis:   . T CT C(XT X)−1 CT −1 Cβ. . SSH = β (16.46) (2) Obtain the sum of squares due to error: T

. XT Y. SSE = YT Y − β

(16.47)

(3) Let ne = n − np = (sample size) − (number of effective parameters in the model). (4) If the null hypothesis, H0 , is true, then SSH/nk SSE/ne c 2000 by Chapman & Hall/CRC 

(16.48)

has an F distribution with nk and ne degrees of freedom. 16.5.10

Regression significance test

For the general linear model Y = Xβ + ,

E [Y] = Xβ,

Var[Y] = σ 2 I

(16.49)

the normal equations, SSE, and an estimate of β are given by . = XT Y XT Xβ . T XT Y SSE = YT Y − β . = (XT X)−1 XT Y. β

(16.50)

Suppose the reduced model is given by Y = Xβr + ,

Cβr = 0.

(16.51)

The sum of squares due to hypothesis is given by SSH = SSE(R) − SSE. 16.5.11

(16.52)

Alternate form of the distribution

The quotient SSE (16.53) SSE + SSH has a beta distribution with parameters ne/2 and nh/2 . Using this distribution, the hypothesis tests are lower-tailed; reject H0 if the value of the test statistic B is smaller than the critical value. Thus the rejection region is given by n n  SSE e h B= ≤β , (usual notation) or, SSE(R) 2 2 ≤ β ∗ (nh , ne ) (beta percentage point), or n n  e h ≤I , (incomplete beta function). 2 2 B=

16.6

GENERAL LINEAR MODEL OF LESS THAN FULL RANK

A singular general linear model is not of full rank. For a general linear model Y = Xβ +

(16.54)

. = XT y XT Xβ

(16.55)

with normal equations

suppose the rank of the design matrix X is r (with r < m). Then the matrix (XT X) is singular and has no inverse; there are no unbiased estimates for each βi . However, there may exist unbiased estimates for certain functions of the βi ’s. c 2000 by Chapman & Hall/CRC 

16.6.1

Estimable function and estimability

Suppose wT β is a function of the βi ’s where wT is a given vector of weights. . such that An estimator for wT β is a linear function of the Y ’s, cT Y = wT β,  T  E c Y = wT β for all β (16.56)  T  and the variance is a minimum, i.e., Var c y = minimum. The unbiasedness constraints are   E cT Y = w T β cT E [Y] = wT β cT Xβ = wT β for all β cT X = w T cT X − wT = 0. (16.57)  T  Therefore, the variance, Var c Y = σ 2 cT c, must be a minimum subject to the constraints cT X = wT . The criterion function Φ is 1 Φ = cT c − (cT X − wT )λ and (16.58) 2 ∂Φ = c − Xλ. (16.59) ∂c In equation (16.59), set the derivative equal to zero to obtain ˆ. Xλ = c

(16.60)

ˆ XT Xλ = XT c

(16.61)

Premultiply by XT :

which is equal to w under the constraints. Therefore, XT Xλ = w.

(16.62)

Equations (16.60) and (16.62) are the conjugate normal equations. If X has rank r (with r < m) there will always be r columns that form a basis with the remaining m − r columns as an extension, linear combinations of the basis. For the general linear model Y = Xβ +

(16.63)

partition the elements of β and the columns of X such that T βT = [βT 1 , β2 ]

(16.64)

X = [X1 , X2 ]

(16.65)

with dimensions given by: β1 is r × 1, β2 is (m − r) × 1, X1 is r × r, and X2 is (m − r) × r, such that X1 is a basis for X. The columns of X2 must be c 2000 by Chapman & Hall/CRC 

linear combinations of those in X1 . Therefore, there exists a matrix Qr(m−r) such that X2 = X1 Q. If X1 and X2 are given, suppose X2 T X1 X2

= X1 Q and

= Q=

XT 1 X1 Q −1 T (XT X1 X2 . 1 X1 )

Often, Q may be found by inspection. Using equation (16.66) the matrix X may be written as       X = X 1 X 2 = X 1 X 1 Q = X 1 Ir Q and the conjugate normal equation (16.62) may be written as         XT XT XT w1 1 1 X1 1 X1 Q X1 X1 Q λ = λ= T T T T w2 Q QT XT X X Q X X Q 1 1 1 1 1

(16.66) (16.67)

(16.68)

(16.69)

where w1 is r × 1 and w2 is (m − r) × 1. Equating components:  T  X1 X1 XT 1 X1 Q λ = w1  T T  Q X1 X1 QT XT 1 X1 Q λ = w2 .

(16.71)

Premultiply equation (16.70) by QT to obtain  T T  T Q X1 X1 QT XT 1 X1 Q λ = Q w1 .

(16.72)

(16.70)

In order for the system to be consistent, the condition w2 = QT w1

(16.73)

must be true. Therefore, in the function wβ, the weight vector w must be of the form wT = [w1T , w2T ] where w2T = w1T Q.

(16.74)

If the vector of weights w is of this form then there is a linear unbiased (and mathematically consistent) estimate for the function wβ. Equation (16.74) is the condition of estimability of a linear A  function.  function wT β is estimable if wT can be written as wT = w1T , w2T where w2T is related to w2T in the same way as X2 is related to X1 . A parametric function is linearly estimable if there exists a linear combination of the observations whose expected value is equal to the function, i.e., if there exists an unbiased estimate. If the function wT β is estimable, equation (16.70) may be written as   (XT (16.75) 1 X1 ) I Q λ = w1 . The equation involving w2 may be disregarded since w2 is completely determined by the relation w2 = QT w1 . Therefore   −1 I Q λ = (XT w1 . (16.76) 1 X1 ) c 2000 by Chapman & Hall/CRC 

Rewriting the first conjugate normal equation (16.60) yields ˆ Xλ = c   ˆ X1 X2 λ = c   ˆ. X1 I Q λ = c

(16.77)

Using equation (16.76): −1 ˆ X1 (XT w1 = c 1 X1 )

and

(16.78)

T β = w (XT X )−1 XT y ˆy = w0 c 1 1 1 1

(16.79)

which is of the same form as in the non–singular case, except that X has been replaced by its basis X1 and in w only the first r elements are considered: w1 . The normal equations in the method of least squares, T

. = XT y (XT X)β

(16.80)

may be used formally in the reduced statement T .T (XT 1 X1 )β1 = X1 y.

16.6.2

(16.81)

General linear hypothesis model of less than full rank

For a general linear model of less than full rank: Y = Xβ +     β1 X X = + 1 2 β2     β1 = X1 X1 Q + β2

(16.82)

= X1 β1 + X1 Qβ2 + = X1 (β1 + Qβ2 ) + . Therefore, this general linear model may be written in the form Y = X1 β∗ + where β∗ = β1 + Qβ2 . 16.6.2.1

Sum of squares due to error

Since has not changed in this model, the normal equations are ∗

T . (XT 1 X1 )β = X1 y,

and the sum of squares due to error is ∗

. ]T XT y SSE = eT e = yT y − [β 1 −1 T = yT y − (XT X1 yXT 1 X1 ) 1 y.

The expression SSE/σ 2 has a chi–square distribution with n − r degrees of freedom where r is the rank of X. The effective number of parameters in the c 2000 by Chapman & Hall/CRC 

singular model is only r, while the remaining m−r parameters are determined in terms of the first r by the estimability condition. 16.6.2.2

Sum of squares due to hypothesis

Suppose the null hypothesis is given by H0 : Cβ = 0

where

  C = C1 C2

(16.83)

and C1 has dimension r × r and C2 has dimension r × (m − r). Equation (16.83) implies       β1 0 C1 C2 = . (16.84) 0 β2 The left–hand side of this equation must represent an estimable function, therefore C2 = C1 Q.

(16.85)

Equation (16.85) is the condition of testability: if  T c1  cT   2  C= .   ..  cT nh where (cT i β) is an estimable function (i = 1, 2, . . . , nh ), then the null hypothesis H0 : C1 β1 + C2 β2 = 0 may be written as H0 : C1 β1 + C1 Qβ2 = 0 ∗

H0 : C1 β = 0,

or simply ∗

where β = (β1 + Qβ2 ).

Therefore, a null hypothesis H0 : Cβ = 0 is testable if Cβ consists of nh estimable functions, i.e., if C2 = C1 Q, where C = [C1 , C2 ]. The sum of squares due to hypothesis is given by . ∗ )T CT [C1 (XT X1 )−1 CT ]−1 C1 β .∗ SSH = (β 1 1 1 ∗

. = (XT X1 )−1 XT y. The expression SSH/σ 2 has a chi–square diswhere β 1 1 tribution with nh degrees of freedom. If the null hypothesis is true, then SSH/nk SSE/ne has an F distribution with nk and ne degrees of freedom.

c 2000 by Chapman & Hall/CRC 

16.6.3

Constraints and conditions

If the general linear model is singular of rank r < m, then (m − r) constraints . ’s (the estimates) may be arbitrarily introduced, for example on the β i βˆr+1 = 0, . . . , βˆm = 0 m 

βˆi = 0,

i=1

m 

or

(16.86)

ni βˆi = 0.

(16.87)

i=1

This procedure reparameterizes the model. The constraining functions are fairly arbitrary, but they must not be estimable functions, otherwise the resulting model will still be singular. To apply the constraints in equation (16.86), delete the last (m − r) rows and columns of XT X and the last (m − r) elements of XT y. To apply the constraints in equation (16.87), add a constant to all elements of XT X. This has no effect on the value of estimable functions or test statistics. A different situation arises if conditions are placed on the parameters in a model, especially on interaction terms. In a two-factor, fixed effects experiment, the model is given by Yijk = µ + αi + βj + (αβ)ij + 3ijk

(16.88)

with assumptions on the interaction terms a 

(αβ)ij =

i=1

b 

(αβ)ij = 0.

(16.89)

j=1

The assumptions in equation (16.89) are often called natural constraints (even though the are neither natural nor constraints). These assumptions represent a set of conditions on the interactions, minimizing this effect (making SSH for interaction a minimum). Given these assumptions, the model is still singular, but can be made nonsingular by introducing the arbitrary constraints a 

α ˆ i = 0,

i=1

b 

βˆj = 0.

(16.90)

j=1

Using the different assumptions All αi ’s = 0,

All βj ’s = 0

would simplify the model to a one-way anova.

c 2000 by Chapman & Hall/CRC 

(16.91)

CHAPTER 17

Miscellaneous Topics Contents 17.1 17.2

Geometric probability Information and communication theory 17.2.1 Discrete entropy 17.2.2 Continuous entropy 17.2.3 Channel capacity 17.2.4 Shannon’s theorem 17.3 Kalman filtering 17.3.1 Extended Kalman filtering 17.4 Large deviations (theory of rare events) 17.4.1 Theory 17.4.2 Sample rate functions 17.4.3 Example: Insurance company 17.5 Markov chains 17.5.1 Transition function 17.5.2 Transition matrix 17.5.3 Recurrence 17.5.4 Stationary distributions 17.5.5 Random walks 17.5.6 Ehrenfest chain 17.6 Martingales 17.6.1 Examples of martingales 17.7 Measure theoretical probability 17.8 Monte Carlo integration techniques 17.8.1 Importance sampling 17.8.2 Hit-or-miss Monte Carlo method 17.9 Queuing theory 17.9.1 M/M/1 queue 17.9.2 M/M/1/K queue 17.9.3 M/M/2 queue 17.9.4 M/M/c queue 17.9.5 M/M/c/c queue c 2000 by Chapman & Hall/CRC 

17.9.6 M/M/c/K queue 17.9.7 M/M/∞ queue 17.9.8 M/Ek /1 queue 17.9.9 M/D/1 queue 17.10 Random matrix eigenvalues 17.10.1 Random matrix products 17.11 Random number generation 17.11.1 Pseudorandom number generation 17.11.2 Generating nonuniform random variables 17.11.3 References 17.12 Resampling methods 17.13 Self-similar processes 17.13.1 Definitions 17.13.2 Self-similar processes 17.14 Signal processing 17.14.1 Estimation 17.14.2 Matched filtering (Wiener filter) 17.14.3 Median filter 17.14.4 Mean filter 17.14.5 Spectral decompositions 17.15 Stochastic calculus 17.15.1 Brownian motion (Wiener processes) 17.15.2 Brownian motion expectations 17.15.3 Itˆ o lemma 17.15.4 Stochastic integration 17.15.5 Stochastic differential equations 17.15.6 Motion in a domain 17.15.7 Option Pricing 17.16 Classic and interesting problems 17.16.1 Approximating a distribution 17.16.2 Averages over vectors 17.16.3 Bertrand’s box “paradox” 17.16.4 Bertrand’s circle “paradox” 17.16.5 Bingo cards: nontransitive 17.16.6 Birthday problem 17.16.7 Buffon’s needle problem 17.16.8 Card problems 17.16.9 Coin problems 17.16.10 Coupon collectors problem 17.16.11 Dice problems 17.16.12 Ehrenfest urn model 17.16.13 Envelope problem “paradox” c 2000 by Chapman & Hall/CRC 

17.16.14 Gambler’s ruin problem 17.16.15 Gender distributions 17.16.16 Holtzmark distribution: stars in the galaxy 17.16.17 Large-scale testing 17.16.18 Leading digit distribution 17.16.19 Lotteries 17.16.20 Match box problem 17.16.21 Maximum entropy distributions 17.16.22 Monte Hall problem 17.16.23 Multi-armed bandit problem 17.16.24 Parking problem 17.16.25 Passage problems 17.16.26 Proofreading mistakes 17.16.27 Raisin cookie problem 17.16.28 Random sequences 17.16.29 Random walks 17.16.30 Relatively prime integers 17.16.31 Roots of a random polynomial 17.16.32 Roots of a random quadratic 17.16.33 Simpson paradox 17.16.34 Secretary call problem 17.16.35 Waiting for a bus 17.17 Electronic resources 17.17.1 Statlib 17.17.2 Uniform resource locators 17.17.3 Interactive demonstrations and tutorials 17.17.4 Textbooks, manuals, and journals 17.17.5 Free statistical software packages 17.17.6 Demonstration statistical software packages 17.18 Tables 17.18.1 Random deviates 17.18.2 Permutations 17.18.3 Combinations

17.1

GEOMETRIC PROBABILITY

1. Two points on a finite line: If A and B are uniformly chosen from the interval [0, 1), and X is the distance between A and B (that is, X = |A − B|) then the probability density of X if fX (x) = 2(1 − x) for 0 ≤ x ≤ 1.

c 2000 by Chapman & Hall/CRC 

2. Many points on a finite line: Suppose n − 1 values are randomly selected from a uniform distribution on the interval [0, 1). These n − 1 values determine n intervals. Pk (x) = Probability (exactly k intervals have length larger than x)  '   n n−1 = [1 − kx]n−1 − [1 − (k + 1)x]n−1 + k 1   , n−k · · · + (−1)s [1 − (k + s)x]n−1 s (17.1) 5 6 1 where s = − k . Using this, the probability that the largest interval x length exceeds x is (for 0 ≤ x ≤ 1):     n n 1 − P0 (x) = (1 − x)n−1 − (1 − 2x)n−1 + . . . (17.2) 1 2 3. Points in the plane: Suppose the number of points in any region of area A of the plane is a Poisson random variable with mean λA (i.e., λ is the density of the points). Given a fixed point P define R1 , R2 , . . . , to be the distance to the point nearest to P , second nearest to P , etc. The probability density function for Rs is (for 0 ≤ r ≤ ∞): fRs (r) =

2(λπ)s 2s−1 −λπr2 e r (s − 1)!

(17.3)

4. Points on a checkerboard: Consider the unit squares on a checkerboard and select one point uniformly in each square. The following results concern the distance between points, on average. (a) For adjacent squares (such as a black and white square) the mean distance between points is 1.088. (b) For diagonal squares (such as between two white squares) the mean between points is 1.473. 5. Points in three-dimensional space: Suppose the number of points in any volume V is a Poisson random variable with mean λV (i.e., λ is the density of the points). Given a fixed point P define R1 , R2 , . . . , to be the distance to the point nearest to P , second nearest to P , etc. The probability density function for Rs is (for 0 ≤ r ≤ ∞): s  3 43 λπ 3 4 fRs (r) = r3s−1 e− 3 λπr (17.4) (s − 1)! c 2000 by Chapman & Hall/CRC 

6. Points in a cube: Choose two points uniformly in a unit cube. The distance between these points has mean 0.66171 and standard deviation 0.06214. 7. Points in n-dimensional cubes: Let two points be selected randomly from a unit n-dimensional cube. The expected distance between the points, ∆(n), is ∆(1) = 1/3 ∆(2) ≈ 0.54141 . . . ∆(3) ≈ 0.66171 . . . ∆(4) ≈ 0.77766 . . .

∆(5) ≈ 0.87852 . . . ∆(6) ≈ 0.96895 . . . ∆(7) ≈ 1.05159 . . . ∆(8) ≈ 1.12817 . . .

8. Points on a circle: Select three points at random on a unit circle. These points determine a triangle with area A. The mean and variance of area are: 3 µA = ≈ 0.4775 2π  (17.5) 3 π2 − 6 2 σA = ≈ 0.1470 8π 2 9. Particle in a box: A particle is bouncing randomly in a square box with unit side. On average, how far does it travel between bounces? Suppose the particle is initially at some random position in the box and is traveling in a straight line in a random direction and rebounds normally at the edges. Let θ be the angle of the point’s initial vector. After traveling a distance r (where r  1; think of many adjacent boxes and the particle exits each box and enters the next box), the point has moved r cos θ horizontally and r sin θ vertically, and thus has struck r(sin θ + cos θ) + O(1) walls. Hence the average distance between walls is 1/(sin θ + cos θ). Averaging this over all angles results in √  √ 2 π/2 dθ 2 2 (17.6) = ln(1 + 2) ≈ 0.793515 π 0 sin θ + cos θ π See J. G. Berryman, Random close packing of hard spheres and disks, Physical Review A, 27, pages 1053–1061, 1983 and H. Solomon, Geometric Probability, SIAM, Philadelphia, PA, 1978. 17.2 17.2.1

INFORMATION AND COMMUNICATION THEORY Discrete entropy

Suppose X is a discrete random variable that assumes n distinct values. Let pX be the probability distribution for X and Prob [X = x] = px . The entropy c 2000 by Chapman & Hall/CRC 

Figure 17.1: Binary entropy function of the distribution is H(pX ) = −



px log2 px .

(17.7)

x

The units for entropy is bits. The maximum value of H(pX ) is log2 n and is obtained when X is a discrete uniform random variable that assumes n values. Entropy measures how much information is gained from observing the value of X. If X assumes only two values, pX = (p, 1 − p), and H(pX ) = H(p) = −p log2 p − (1 − p) log2 (1 − p)

(17.8)

with a maximum at p = 0.5. A plot of H(p) is in Figure 17.1. Given two discrete random variables X and Y , pX×Y is the joint distribution of X and Y . The mutual information of X and Y is defined by I(X, Y ) = H(pX ) + H(pY ) − H(pX×Y )

(17.9)

Note that I(X, Y ) ≥ 0 and that I(X, Y ) = 0 if and only if X and Y are independent. Mutual information gives the amount of information obtained about X after observing a value of Y (and vice versa). Example 17.78 : A coin weighing problem. There are 12 coins of which one is counterfeit, differing from the others by its weight. Using a balance but no weights, how many weighings are necessary to identify the counterfeit coin? Solution: (S1) Any of the 12 coins may turn out to be the counterfeit one, and it may be heavier or lighter than the genuine ones. Hence, there are 24 possible outcomes. For equal probabilities of these 24 outcomes, the entropy of the unknown result is then log2 24 ≈ 4.58. (S2) Each weighing process has three outcomes (equal weight, left side heavier, right side heavier). Using an assumption of equal probabilities gives an entropy of log2 3 ≈ 1.58 per weighing. (Note that other assumptions will produce a smaller entropy.) (S3) Therefore the minimal number of weighings cannot be less that 4.58/1.58 ≈ 2.90. Hence 3 weighings are needed. (In fact, 3 weighings are sufficient.)

c 2000 by Chapman & Hall/CRC 

17.2.2

Continuous entropy

For a d-dimensional continuous random variable X, the entropy is  h(X) = − p(x) log p(x) dx

(17.10)

Rd

Continuous entropy is not the the limiting case of the entropy of a discrete random variable. In fact, if X is the limit of the one-dimensional discrete random variable {Xn }, and the entropy of X is finite, then lim (H(Xn ) − n log 2) = h(X)

n→∞

(17.11)

If X and Y are continuous d-dimensional random variables with density functions p(x) and q(y), then the relative entropy is  p(x) H(X, Y) = dx (17.12) p(x) log q(x) d R A d-dimensional Gaussian random variable N (a, Γ) has the density function   1 1  exp − (x − a)T Γ−1 (x − a) g(x) = (17.13) 2 (2π)d/2 |Γ| where a is the vector of means and Γ is the positive definite covariance matrix. 1. If X = (X1 , X2 , . . . , Xd ) is a d-dimensional Gaussian random vector with distribution N (a, Γ) then   1 d h(X) = log (2πe) |Γ| (17.14) 2 2. If X and Y are d-dimensional Gaussian random vectors with distributions N (a, Γ) and N (b, ∆) then     1 |∆| H(X, Y) = log + tr Γ ∆−1 − Γ−1 2 |Γ|  (17.15) T −1 + (a − b) ∆ (a − b) 3. If X is a d-dimensional Gaussian random vector with distribution N (a, Γ), and if Y is a d-dimensional random vector with a continuous probability distribution having the same covariance matrix Γ, then h(X) ≤ h(Y) 17.2.3

(17.16)

Channel capacity

The transition probabilities are defined by  tx,y = Prob [X = x | Y = y]. The distribution pX determines pY by py = x tx,y px . The matrix T = (tx,y ) is c 2000 by Chapman & Hall/CRC 

the transition matrix. The matrix T defines a channel given by a transition diagram (input is X, output is Y ). For example (here X and Y only assume two values): tx1 ,y1 x0 t ✲ty ❍ ✯ 0 tx1 ,y0 ✟✟ ❍ ❍ ✟ ❍❍✟✟ ✟❍ ✟ t ❍ ✟ x0 ,y1❍❍ x1 t✟✟ ✲ty1 ❥ ❍ tx0 ,y0 The capacity of the channel is defined as C = max I(X, Y )

(17.17)

pX

A channel is symmetric if each row is a permutation of the first row and the transition matrix is a symmetric matrix. The capacity of a symmetric channel is C = log2 n − H(p), where p is the first row. The capacity of a symmetric channel is achieved with equally likely inputs. The channel below on the left is symmetric; both channels achieve capacity with equally likely inputs. Binary symmetric channel 1−p 0 t ✲t0 ✯ ✟ ❍❍ p ✟✟ ❍❍ ✟ ❍✟ ✟ ❍ ✟ ✟ p❍❍ ❍❍ 1 ✟ t ✟ ✲t1 ❥ 1−p C = 1 − H(p) 17.2.4

Binary erasure channel 1−p 0 t ✲t0  p   zt?  ✿ ✘✘✘ ✘ ✘ ✘✘p 1 t✘✘✘ ✲t1 1−p C =1−p

Shannon’s theorem

Let both X and Y be discrete random variables with values in an alphabet A. A code is a set of n-tuples (codewords) with entries from A that is in oneto-one correspondence with M messages. The rate R of the code is defined as n1 log2 M . Assume that the codeword is sent via a channel with transition matrix T by sending each vector element independently. Define e=

max

all codewords

Prob [codeword incorrectly decoded].

(17.18)

Then Shannon’s coding theorem states: (a) If R < C, then there is a sequence of codes with n → ∞ such that e → 0. (b) If R ≥ C, then e is always bounded away from 0.

c 2000 by Chapman & Hall/CRC 

17.3

KALMAN FILTERING

In the following model for k ≥ 0: xk+1 = Fk xk + Gk wk zk = HkT xk + vk

(17.19)

with the conditions: 1. The initial state x0 is a Gaussian random variable with mean x0 and covariance P0 , independent of {vk } and {wk }. 2. The {vk } and {wk } are independent, zero mean, Gaussian white processes with     E vk vlT = Rk δkl and E wk wlT = Qk δkl (17.20) an estimate of xk from observation of the zi ’s is desired. 1. Define Zk−1 to be the sequence of observed values {z0 , z1 , . . . , zk−1 } 2. Define the estimate of xk , conditioned on the z values (up to the (k−1)th .k/k−1 = E [xk | Zk−1 ]. Similarly, define the estimate of xk , value) to be x .k/k = E [xk | Zk ]. conditioned on the z values (up to the k th value) to be x 3. Define the error covariance matrix to be # "  T .k/k−1 xk − x .k/k−1 | Zk−1 . // Define Σk/k in Σk/k−1 = E xk − x a similar way. .k/k−1 for k = 0 (i.e., x .0/−1 ) to be x0 = E [x0 ], 4. By convention, define x i.e., the expected value of x0 given no measurements. Similarly, take Σ0/−1 to be P0 . The solution is given by (the intermediate matrix Kk is called the gain matrix ) .0/−1 = x0 x Σ0/−1 = P0 Ωk = HkT Σk/k−1 Hk + Rk Kk = Fk Σk/k−1 Hk Ω−1 k   .k+1/k = Fk − Kk HkT x .k/k−1 + Kz zk x   .k/k = x .k/k−1 .k/k−1 + Σk/k−1 Hk Ω−1 x zk − HkT x k

(17.21)

T Σk/k = Σk/k−1 − Σk/k−1 Hk Ω−1 k Hk Σk/k−1

Σk+1/k = Fk Σk/k FkT + Gk Qk GT k See B. D. O. Anderson and J. B. Moore, Optimal Filtering, Prentice–Hall, Inc., Englewood Cliffs, NJ, 1979.

c 2000 by Chapman & Hall/CRC 

17.3.1

Extended Kalman filtering

We have the following model for k ≥ 0: xk+1 = fk (xk ) + gk (xk )wk zk = hk (xk ) + vk

(17.22)

with the usual assumptions. We presume the nonlinear functions {fk , gk , hk } are sufficiently smooth and they can be expanded in Taylor series about the .k/k and x .k/k−1 as conditional means x .k/k ) + . . . fk (xk ) = fk (. xk/k ) + Fk (xk − x gk (xk ) = gk (. xk/k ) + · · · = Gk + . . . hk (xk ) = hk (. xk/k−1 ) +

HkT (xk

(17.23)

.k/k−1 ) + . . . −x

.k/k and x .k/k−1 Neglecting higher order terms and assuming knowledge of x enables us to approximate the original system as xk+1 = Fk xk + Gk wk + uk zk = HkT xk + vk + yk

(17.24)

where uk and yk are calculated from .k/k uk = fk (. xk/k ) − Fk x

and

.k/k−1 yk = hk (. xk/k−1 ) − HkT x

(17.25)

The Kalman filter for this approximate signal model is: .0/−1 = x0 x Σ0/−1 = P0 Ωk = HkT Σk/k−1 Hk + Rk Lk = Σk/k−1 Hk Ω−1 k   .k/k = x .k/k−1 + Lk zk − hk (. x xk/k−1 )

(17.26)

.k+1/k = fk (. x xk/k ) T Σk/k = Σk/k−1 − Σk/k−1 Hk Ω−1 k Hk Σk/k−1

Σk+1/k = Fk Σk/k FkT + Gk Qk GT k See B. D. O. Anderson and J. B. Moore, Optimal Filtering, Prentice–Hall, Inc., Englewood Cliffs, NJ, 1979. 17.4 17.4.1

LARGE DEVIATIONS (THEORY OF RARE EVENTS) Theory

1. Cram´ ers Theorem: Let {Xi } be a sequence of bounded, independent, identically distributed random variables with common mean m. Define

c 2000 by Chapman & Hall/CRC 

Mn to be the sample mean of the first n random variables: 1 (17.27) (X1 + X2 + · · · + Xn ) n The tails of the probability distribution for Mn decay exponentially, as n → ∞, at a rate given by the convex rate function I(x). Mn =

Prob [Mn > x] ∼ e−nI(x)

for x > m

−nI(x)

for x < m

Prob [Mn < x] ∼ e

(17.28)

2. Chernoff ’s Formula: The rate-function I(x) is related to the cumulant generating function λ(θ) (see page 38) via I(x) = max {xθ − λ(θ)} .

(17.29)

θ

3. Contraction Principle: If {Xn } satisfies a large deviation principle with rate function I and f is a continuous function, then {f (Xn )} satisfies a large deviation principle with rate function J, where J is given by J(y) = min [I(x) | f (x) = y] . 17.4.2

(17.30)

Sample rate functions

1. Let {Xi } be a sequence of Bernoulli random variables where p is the probability of obtaining a “1” and (1 − p)is the probability of obtaining a “0”. Then λ(θ) = ln p · eθ + (1 − p) · 1 and therefore I(x) = x ln

x 1−x + (1 − x) ln p 1−p

(The maximum value of I occurs when θ is θ =

ln x p

(17.31) −

ln(1−x) 1−p ).

2. If the random variables in the sequence {Xi } are all N(µ, σ 2 ) then  2 1 x−µ I(x) = . (17.32) 2 σ 17.4.3

Example: Insurance company

Suppose an insurance company collects daily premiums as a constant rate p, and has daily claims total Z ∼ N(µ, σ 2 ). The company would like to, naturally, avoid going bankrupt. The probability that the payments exceed T income after T days is the probability that k=1 Zk exceeds pT . For T large %

& T 1  Prob Zk > p ∼ e−T I(p) (17.33) T k=1

c 2000 by Chapman & Hall/CRC 

If an acceptable amount of risk is e−r , then e−T I(p) = e−r , or I(p) = r/T .  2 Using the rate function for a normal random variable, r = T2 p−µ , or σ 2r p=µ+σ . (17.34) T The safety loading is defined by -     p−µ 2r σ = (17.35) µ µ T 



17.5

 



safety loading

size of fluctuations







fixed by regulators

MARKOV CHAINS

A discrete parameter stochastic process is a collection of random variables {X(t), t = 0, 1, 2, . . . }. The values of X(t) are called the states of the process. The collection of states is called the state space. The values of t usually represent points in time. The number of states is either finite or countably infinite. A discrete parameter stochastic process is called a Markov chain if, for any set of n time points t1 < t2 < · · · < tn , the conditional distribution of X(tn ) given values for X(t1 ), X(t2 ), . . . , X(tn−1 ) depends only on X(tn−1 ). That is, Prob [X(tn ) ≤ xn | X(t1 ) = x1 , . . . , X(tn−1 ) = xn−1 ] = Prob [X(tn ) ≤ xn | X(tn−1 ) = xn−1 ]. (17.36) A Markov chain is said to be stationary if the value of the conditional probability P [X(tn+1 ) = xn+1 | X(tn ) = xn ] is independent of n. The following discussion will be restricted to stationary Markov chains. 17.5.1

Transition function

Let x and y be states and let {tn } be time points in T = {0, 1, 2, . . . }. The transition function, P (x, y), is defined by P (x, y) = Pn,n+1 (x, y) = Prob [X(tn+1 ) = y | X(tn ) = x]

(17.37)

P (x, y) is the probability that a Markov chain in state x at time n will be in state y at time  n + 1. Some properties of the transition function are: P (x, y) ≥ 0 and y P (x, y) = 1. The values of P (x, y) are commonly called the one-step transition probabilities.  The function π0 (x) = P (X(0) = x), with π0 (x) ≥ 0 and x π0 (x) = 1, is called the initial distribution of the Markov chain. It is the probability distribution when the chain is started. Thus, P [X(0) = x0 , X(1) = x1 , . . . , X(n) = xn ] = π0 (x0 )P0,1 (x0 , x1 )P1,2 (x1 , x2 ) · · · Pn−1,n (xn−1 , xn ). (17.38) c 2000 by Chapman & Hall/CRC 

17.5.2

Transition matrix

A convenient way to summarize the transition function of a Markov chain is by using the one-step transition matrix . It is defined as   P (0, 0) P (0, 1) . . . P (0, n) . . .  P (1, 0) P (1, 1) . . . P (1, n) . . .     .. .. .. .. . . . . . (17.39) P =   P (n, 0) P (n, 1) . . . P (n, n) . . .   .. .. .. . . . Define the n–step transition matrix by P (n) to be the matrix with entries P (n) (x, y) = Prob [X(tm+n ) = y | X(tm ) = x].

(17.40)

This can be written using the one-step transition matrix as P (n) = P n . Suppose the state space is finite. The one-step transition matrix is said to be regular if, for some positive power m, all of the elements of P m are strictly positive. Theorem 1 (Chapman–Kolmogorov equation) Let P (x, y) be the one-step transiton function of a Markov chain and define P 0 (x, y) = 1, if x = y, and 0, otherwise. Then, for any pair of nonnegative integers s and t such that s + t = n,  P n (x, y) = P s (x, z)P t (z, y). (17.41) z

17.5.3

Recurrence

Define the probability that a Markov chain starting in state x returns to state x for the first time after n steps by f n (x, x) = Prob [X(tn ) = x, X(tn−1 ) = x, . . . , X(t1 ) = x | X(t0 ) = x]. (17.42)  n It follows that P n (x, x) = k=0 f k (x, x)P n−k (x, x). A state x is said to be ∞ n recurrent if n=0 f (x, x) = 1. This means that a state x is recurrent if, after starting in x, the probability of returning to x after some finite length of time is one. A state which is not recurrent is said to be transient. Theorem 2 A state x of a Markov chain is recurrent if and only if ∞ n P (x, x) = ∞. n=1 Two states, x and y, are said to communicate if, for some n ≥ 0, P n (x, y) > 0. This theorem implies that if x is a recurrent state and x communicates with y, then y is also a recurrent state. A Markov chain is said to be irreducible if every state communicates with every other state and with itself.

c 2000 by Chapman & Hall/CRC 

Let x be a recurrent state and define Tx the (return time) to be the number of stages for a Markov chain to return to state x, having begun there. A recurrent state x is said to be null recurrent if E [Tx ] = ∞. A recurrent state that is not null recurrent is said to be positive recurrent. 17.5.4

Stationary distributions

Let {X(t), t = 0, 1, 2, . . . } be a Markov chain having a one-step  transition function of P (x, y). A function π(x) where each π(x) is nonnegative, x π(x)P (x, y) = π(y), and y π(y) = 1, is called a stationary distribution. If a Markov chain has a stationary distribution and limn→∞ P n (x, y) = π(y), then regardless of the initial distribution, π0 (x), the distribution of X(tn ) approaches π(x) as n becomes infinite. When this happens π(x) is often referred to as the steadystate distribution. The following categorizes those Markov chains that have stationary distributions. Theorem 3 Let XP denote the set of positive recurrent states of a Markov chain. 1. If XP is empty, the chain has no stationary distribution. 2. If XP is a nonempty irreducible set, the chain has a unique stationary distribution. 3. If XP is nonempty but not irreducible, the chain has an infinite number of distinct stationary distributions. The period of a state x is denoted by d(x) and is defined to be the greatest common divisor of all integers n ≥ 1 for which P n (x, x) > 0. If P n (x, x) = 0 for all n ≥ 1 then define d(x) = 0. If each state of a Markov chain has d(x) = 1 the chain is said to be aperiodic. If each state has period d > 1 the chain is said to be periodic with period d. The vast majority of Markov chains encountered in practice are aperiodic. An irreducible, positive recurrent, aperiodic Markov chain always possesses a steady-state distribution. An important special case occurs when the state space is finite. In this case, suppose that X = {1, 2, . . . , K}. Let π0 = {π0 (1), π0 (2), . . . , π0 (K)}. Theorem 4 Let P be a regular one-step transition matrix and π0 be an arbitrary vector of initial probabilities. Then limn→∞ π0 (x)P n = y, where K yP = y, and i=1 π0 (ti ) = 1. Example 17.79 : A Markov chainhaving three states, {0, 1, 2}, with a one-step transition matrix of  1/2 0 1/2 P =  1/4 3/4 0  is diagrammed as follows: 0 3/4 1/4

c 2000 by Chapman & Hall/CRC 

The one-step transition matrix gives a two–step transition matrix of   1/4 3/8 3/8 P (2) = P 2 =  5/16 9/16 1/8 3/ 16 9/16 1/8 The one-step transition matrix is regular. This Markov chain is irreducible, and all three states are recurrent. In addition, all three states are positive recurrent. Since all states have period 1, the chain is aperiodic. The steady-state distribution is π(0) = 3/11, π(1) = 6/11, and π(2) = 2/11.

17.5.5

Random walks

Let {η(t1 ), η(t2 ), . . . } be independent random variables having a common density f (x), and let t1 , t2 , . . . be integers. Let X(t0 ) be an integer–valued random variable that is independent of η(t1 ), η(t2 ), . . . , and X(tn ) = X0 + n η(t i ). The sequence {X(ti ), i = 0, 1, . . . } is called a random walk . An i=1 important special case is a simple random walk . It is defined by the following.   p if y = x − 1 P (x, y) = r if y = x , where p + q + r = 1, P (0, 0) = p + r.   q if y = x + 1 (17.43) Here, an object begins at a certain point in a lattice and at each step either stays at that point or moves to a neighboring lattice point. In the case of a one– or two–dimensional lattice it turns out that if a random walk begins at a lattice point, x, it will return to that point with probability 1. In the case of a three–dimensional lattice the probability that it will return to its starting point is approximately 0.3405. 17.5.6

Ehrenfest chain

A simple model of gas exchange between two isolated bodies is as follows. Suppose that there are two boxes, Box I and Box II, where Box I contains K molecules numbered 1, 2, . . . , K and Box II contains N − K molecules numbered K + 1, K + 2, . . . , N . A number is chosen at random from {1, 2, . . . , N }, c 2000 by Chapman & Hall/CRC 

and the molecule with that number is transferred from its box to the other one. Let X(tn ) be the number of molecules in Box I after n trials. Then the sequence {X(tn ), n = 0, 1, . . . } is a Markov chain with one–stage transition function of x  y = x − 1,  K x P (x, y) = 1 − (17.44) y = x + 1,  K   0 otherwise 17.6

MARTINGALES

A stochastic process {Zn | n ≥ 1} with E [|Zn |] < ∞ for all n is a (a) martingale process if E [Zn+1 | Z1 , Z2 , . . . , Zn ] = Zn (b) submartingale process if E [Zn+1 | Z1 , Z2 , . . . , Zn ] ≥ Zn (c) supermartingale process if E [Zn+1 | Z1 , Z2 , . . . , Zn ] ≤ Zn Azuma’s inequality: Let {Zn } be a martingale process with mean µ = E [Zn ]. Let Z0 = µ and suppose that −αi ≤ (Zi − Zi−1 ) ≤ βi for nonnegative constants {αi , βi } and i ≥ 1. Then, for any n ≥ 0 and a > 0: * n    2 (a) Prob [Xn − µ ≥ a] ≤ exp −2a2 (αi + βi ) i=1 * n    2 (b) Prob [Xn − µ ≤ −a] ≤ exp −2a2 (αi + βi ) i=1

17.6.1

Examples of martingales

(a) If {Xi } are independent, mean zero random variables, and Zn =

n 

Xi ,

i=1

then {Zn } is a martingale. (b) If {Xi } are independent random variables with E [Xi ] = 1, and Zn = n M Xi , then {Zn } is a martingale. i=1

(c) If {X, Yi } are arbitrary random variables with E [|X|] < ∞, and Zn = E [X | Y1 , Y2 , . . . , Yn ], then {Zn } is a martingale. 17.7

MEASURE THEORETICAL PROBABILITY

1. A σ-field of subsets of a set Ω is a collection F of subsets of Ω that contains φ (the empty set) as a member and is closed under complements and countable unions. If Ω is a topological space, the σ-field generated by the open subsets of Ω is called the Borel σ-field. 2. A probability measure P on a σ-field F of subsets of a set Ω is a function from F to the unit interval [0, 1] such that P (Ω) = 1 and P of a countable union of disjoint sets {Ai } equals the sum of P (Ai ). c 2000 by Chapman & Hall/CRC 

3. A probability space is a triple (Ω, F, P ), where Ω is a set, F is a σ-field of subsets of Ω, and P is a probability measure on F. 4. Given a probability space (Ω, F, P ) and a measurable space (Ψ, G), a random variable from (Ω, F, P ) to (Ψ, G) is a measurable function from (Ω, F, P ) to (Ψ, G). 5. A random variable X from (Ω, F, P ) to (Ψ, G) induces a probability measure on Ψ. The measure of a set A in G is simply P (X −1 (A)). This induced measure is called the distribution of X. 6. A real-valued function F defined on the set of real numbers R is called a distribution function for R if it is increasing and right-continuous and satisfies limx→−∞ F (x) = 0 and limx→∞ F (x) = 1. Let Q be the distribution of X where X is a real valued random variable. Then the function F : x → Q((−∞, x]) is a distribution function. We call F the distribution function of X. 17.8

MONTE CARLO INTEGRATION TECHNIQUES

Random numbers may be used to approximate the value of a definite integral. Let g be an integrable function and define the integral I by  I= g(x) dx, (17.45) B

where B is a bounded region that may be enclosed in a rectangular parallelepiped R with volume V (R). If 1B (x) represents the indicator function of B, ( 1 if x ∈ B 1B (x) = (17.46) 0 if x ∈ B then the integral I may be written as       1 I= g(x)1B (x) dx = g(x)1B (x)V (R) dx V (R) R R

(17.47)

Equation (17.47) may be interpreted as an expected value of the function h(X) = g(X)1B (X)V (R) where the random variable X is uniformly distributed on the parallelepiped R (i.e., it has density function 1/V (R)). The expected value of h(X) may be obtained by simulating random deviates from X, evaluating h at these points, and then computing the mean of the h values. If N trials are used, then the following estimate is obtained: N N 1  V (R)  I ≈ I. = h(xi ) = g(xi )1B (xi ) N i=1 N i=1

where each xi is uniformly distributed in R. c 2000 by Chapman & Hall/CRC 

(17.48)

See J. M. Hammersley and D. C. Handscomb, Monte Carlo Methods, John Wiley, 1965. 17.8.1

Importance sampling

Importance sampling is the term given to sampling from a non-uniform distribution so as to minimize the variance of the estimate for I in equation (17.45). Suppose a sample is selected from a distribution with density function f (x). The integral I may be written as      g(x) g(x) I= f (x) dx = Ef (17.49) f (x) f (x) B where Ef [·] denotes the expectation taken with respect to the density f (x). That is, I is the mean of g(x)/f (x) with respect to the distribution f (x). Associated with this mean is the variance: %'  2 ,2 &  2 g g(x) g (x) 2 2 σf = Ef − I = Ef = (17.50) −I dx − I 2 2 f (x) f B f (x) Approximations to I obtained by sampling from f (x) will have errors that scale with σf . A minimum variance estimator may be obtained by finding the density function f (x) such that σf2 is minimized. Using the calculus of variations the density function for the minimal estimator is fopt (x) = C|g(x)| = J

|g(x)| |g(x)| dx B

(17.51)

where the constant C is chosen so that fopt (x) is appropriately normalized. (Since fopt (x) is a density function, it must integrate to unity.) While finding fopt (x) is as difficult as determining the original integral I, equation (17.51) indicates that fopt (x) should have the same general behavior as |g(x)|. 17.8.2

Hit-or-miss Monte Carlo method

The hit-or-miss Monte Carlo method is very inefficient but is easy to understand. Suppose that 0 ≤ f (x) ≤ 1 when 0 ≤ x ≤ 1. Defining ( 0 if f (x) < y, g(x, y) = (17.52) 1 if f (x) > y, J1 J1J1 then I = 0 f (x) dx = 0 0 g(x, y) dy dx. This integral may be estimated by 1 n∗ I ≈ I. = g(ξ2i−1 , ξ2i ) = n i=1 n n

(17.53)

where the {ξi } are chosen independently and uniformly from the interval [0, 1]. The summation in equation (17.53) reduces to the number of points in the

c 2000 by Chapman & Hall/CRC 

unit square which are below the curve y = f (x) (this defines n∗ ) divided by the total number of sample points (i.e., n). 17.9

QUEUING THEORY

The following diagram and notation are used to define a queue.

A queue is represented as A/B/c/K/m/Z where (a) A and B represent the interarrival times and service times: D deterministic (constant) interarrival or service time distribution. Ek Erlang–k interarrival or service time distribution (a gamma distribution with α = (k − 1), β = 1/λk and density function f (t) = λk(λkt)k−1 e−λkt /(k − 1)! for t > 0. GI general independent interarrival time. G general service time distribution. Hk k–stage hyperexponential interarrival or service time distribution k  (density function is f (t) = αi µi e−µi t for t ≥ 0). i=1

M exponential interarrival or service time distribution. (b) c is the number of identical servers. (c) K is the system capacity. (d) m is the number in the source. (e) Z is the queue discipline: FCFS first come, first served (also known as FIFO). LIFO last in, first out. RSS service in random order. PRI priority service. If all variables are not present, the last three above have the default values: K = ∞, m = ∞, and Z is RSS. Note: The system includes both the queue and the service facility. The variables of interest are: c 2000 by Chapman & Hall/CRC 

(a) an : proportion of customers that find n customers already in the system when they arrive. (b) c: number of servers in the service facility. (c) dn : proportion of customers leaving behind n customers in the system. (d) K: maximum number of customers allowed in queueing system. (e) L: mean number of customers in the steady-state system, L = E [N ]. (f) Lq : mean number of customers in the steady-state queue, Lq = E [Nq ]. (g) λ: mean arrival rate of customers to the system (number per unit time), λ = 1/E [τ ]. (h) µ: mean service rate per server (number per unit time), µ = 1/E [s]. (i) N : random number of customers in system in steady state. (j) Na : random number of customers receiving service in steady state. (k) Nq : random number of customers in queue in steady state. (l) pn : proportion of time the system contains n customers. (m) πq (r): the queueing time that r percent of the customers do not exceed. (n) πw (r): the system time that r percent of the customers do not exceed. (o) q: random time a customer waits in the queue in order to begin service. (p) qn : probability that there are n customers in the system just before a customer enters. (q) ρ: server utilization, the probability that any particular server is busy. (r) s: random service time for one customer, E [s] = 1/µ. (s) τ : random interarrival time, E [τ ] = 1/λ. (t) u: traffic intensity (units are erlangs) u = λ/µ. (u) W : mean time of customers in the system in steady state, W = E [w]. (v) w: total waiting time in the system, including queue and service times, w = q + s. (w) Wq : mean time for customer in the queue in steady state, Wq = E [q]. Relationships between variables: (a) Little’s formula: L = λW and Lq = λWq . (b) For Poisson arrivals: pn = an . (c) If customers arrive one at a time, and are served one at a time: an = dn . (d) N = Nq + Ns

c 2000 by Chapman & Hall/CRC 

(e) W = Wq + Ws 17.9.1

M/M/1 queue

Assume λ < µ: (a) ρ = u/c = (λ/µ)/c (b) pn = (1 − ρ)ρn for n = 0, 1, 2, . . . (c) L = ρ/(1 − ρ) (d) Lq = ρ2 /(1 − ρ) (e) W = 1/µ(1 − ρ) (f) Wq = ρ/µ(1 − ρ) # "  100ρ (g) πq (r) = max 0, W log 100−r # "  100 (h) πw (r) = max 0, W log 100−r 17.9.2

M/M/1/K queue

Assume K ≥ 1 and N ≤ K: (a) ρ = (1 − pK )u ( (1−u)un if λ = µ 1−uK+1 (b) pn = 1/(K + 1) if λ = µ  K K+1 ]  u[1−(K+1)u +Ku K+1 (1−u)(1−u ) (c) L = K/2

and n = 0, 1, . . . , K and n = 0, 1, . . . , K if λ = µ if λ = µ

(d) Lq = L − (1 − p0 ) (e) λa = (1 − pK )λ is the actual arrival rate at which customers enter the system. (f) W = L/λa (g) Wq = Lq /λa (h) qn = pn /(1 − pK ) for n = 0, 1, . . . , K − 1 Note: pK is the probability that an arriving customer is lost since there is no room in the queue. 17.9.3

M/M/2 queue

(a) ρ = u/2 (b) p0 = (1 − ρ)/(1 + ρ) (c) pn = 2(1 − ρ)ρn /(1 + ρ) for n = 1, 2, 3, . . . , c (d) L = 2ρ/(1 − ρ2 ) (e) Lq = 2ρ3 /(1 − ρ2 ) (f) W = 1/µ(1 − ρ2 ) c 2000 by Chapman & Hall/CRC 

(g) Wq = ρ2 /µ(1 − ρ2 ) 17.9.4

M/M/c queue

Erlang’s C formula is the probability that all c servers are busy % &−1 c−1 c!(1 − ρ)  un C(c, u) = 1 + uc n! n=0

(17.54)

(a) ρ = u/c (b) u = λ/µ

% & c−1 n −1  u c!(1 − ρ) uc (c) p0 = C(c, u) = + uc c!(1 − ρ) n=0 n! ( un for n = 0, 1, . . . , c n! p0 (d) pn = n u for n ≥ c c!cn−c p0 (e) L = Lq + u uC(c,U ) c(1−ρ)

(f) Lq =

(g) W = Wq + 1/µ (h) Wq =

C(c,u) cµ(1−ρ)

? > C(c,U )] (i) πq (90) = max 0, ln[10 cµ(1−ρ) ? > C(c,U )] (j) πq (95) = max 0, ln[20 cµ(1−ρ)

17.9.5

M/M/c/c queue

Erlang’s B formula is the probability that all servers are busy % &−1 c c!  un B(c, u) = uc n=0 n! 

(a) pn =

n! un

c  n=0

un n!

(17.55)

−1

for n = 0, 1, . . . , c

(b) λa = λ(1−B(c, u)) is the average traffic rate experienced by the system. (c) ρ = λa /µc (d) L = u [1 − B(c, u)] (e) W = 1/µ 17.9.6

M/M/c/K queue &−1 % c K−c  un u c   u n (a) p0 = + n! c! n=1 c n=0

c 2000 by Chapman & Hall/CRC 

( un (b) pn =

n! c

u c!

p0  u n−c c

for n = 0, 1, . . . , c p0

(c) λa = λ(1 − pK ) (d) ρ = (1 − pK )u/c (e) L = Lq +

c−1 

for n = c + 1, . . . , K



npn + c 1 −

n=0 uc p0 u/c

 (f) Lq = 1− c!(1 − u/c)2 (g) W = L/λa

c−1 

pn n=0  u K−c+1 c

− (K − c + 1)

 u K−c c

u 1− c



(h) Wq = Lq /λa 17.9.7

M/M/∞ queue

(a) pn = e−n un /n! for n = 0, 1, 2, . . . (b) L = u (c) Lq = 0 17.9.8

M/Ek /1 queue     n   kj kj ρ kj n−j n−j−1 (a) pn = (1 − ρ) (−1) r+ for 1+ r n−j n−j−1 k j=0

n = 0, 1, . . . , where r = ρ/(k + ρ) (b) L = Lq + ρ (c) Lq = λWq (d) W = Wq + 1/µ   1+1/k ρ (e) Wq = µ(1−ρ) 2 17.9.9

M/D/1 queue

(a) p0 = (1 − ρ) (b) p1 = (1 − ρ) (eρ − 1)   n  (jρ)n−j−1 (jρ + n − j)ejρ (c) pn = (1 − ρ) (−1)n−j for n = 2, 3, . . . (n − j)! j=0 (d) L = λW = Lq + ρ (e) Lq = λWq =

ρ2 2(1−ρ)

1 µ ρ 2µ(1−ρ)

(f) W = Wq + (g) Wq =

c 2000 by Chapman & Hall/CRC 

17.10

RANDOM MATRIX EIGENVALUES

1. Let A be a n × n matrix whose entries are independent standard normal deviates. (a) The √ probability pn,k that A has k real eigenvalues has the form r + s 2, where r and s are rational. In particular, the probability that A has all real eigenvalues is pn,n = 2−n(n−1)/4 n

k

1

1

2

2

3 4

pn,k 1 2

1 √

1 2 √

1

1− 2 √ 1 4 2 √ 1 − 14 2

4

1 8

2

− 14 + 11 16 2 √ 9 11 8 − 16 2

0 3

0

(17.56)

1 2



≈ 0.707 ≈ 0.293 ≈ 0.354 ≈ 0.646 0.125 ≈ 0.722 ≈ 0.153

(b) The expected number of real eigenvalues of A is  n/2−1  √  (4k − 1)!    2 when n is even   (4k)!! k=0 En =   (4k − 3)!  √ (n−1)/2    when n is odd 1 + 2 (4k − 2)!! k=1    3 3 and En ∼ 2n π 1 − 8n − 128n2 + . . . as n → ∞.

(17.57)

(c) If λn denotes a real eigenvalue of A, then its marginal probability density fn (λ) is given by    1 1 Γ(n − 1, λ2 ) √ fn (λ) = En Γ(n − 1) 2π $ n−1 $ −λ2 /2   (17.58) $λ $e γ((n − 1)/2, λ2 /2) + Γ((n − 1)/2) Γ(n/2)2n/2 where γ(a, x) and Γ(a, x) are incomplete gamma functions (see page 519).

c 2000 by Chapman & Hall/CRC 

(d) If the elements of A have mean 0 and variance 1, and z is a scalar, then # "      2 E det A2 + z 2 I = E det (A + zI) = per J + z 2 I   = n! en z 2 (17.59)   1 1 2 = n! 1 F1 −1; n; − z I 2 2 where J is the matrix of all ones, “per” refers to the permanent n k of a matrix, en (x) = k=0 xk! is the truncated Taylor series for ex , and the hypergeometric function has a scalar multiple of the identity as its argument. 2. Let A be a random n × n matrix with randomly selected integer entries. Let P (p, n) be the probability that det(A) is not congruent to 0 modulo p. Then P (p, n) =

n ) 

1 − p−k



(17.60)

k=1

3. Let A be a random  n × n complex matrix uniformly distributed on the 2 sphere 1 = AF = i,j |Aij | . Then,     2  n−1 2 E det AH A | σmin =λ = λn−r (1 − nλ)n +r−1 r=0 (17.61) Γ(n2 )Γ(n + 1)Γ(n + 2) × Γ(r + 1)Γ(n − r)Γ(n2 + r − 1)Γ(n + 2 − r)     n −1 and if En = E det AH A then En = n2 +n−1 . The first few values of 1/En are {1, 10, 165, 3876, 118755, . . . }.

K (µ, σ 2 ) refers to the distribution 4. Define the following matrices where N X + iY where both X and Y are N (µ, σ 2 ): (a) Gaussian matrix: G(m, n), an m × n random matrix with iid elements which are N (0, 1). (b) Wishart matrix: W (m, n), symmetric random matrix AAT where A is G(m, n). (c) Gaussian orthogonal ensemble (GOE): an m × m random matrix (A + AT )/2 where A is G(m, m). K (d) Complex Gaussian matrix: G(m, n), an m × n random matrix with K (0, 1). iid elements which are N N (m, n), symmetric random matrix (e) Complex Wishart matrix: W H K AA where A is G(m, n).

c 2000 by Chapman & Hall/CRC 

W (m, n) :

)  2 ) πm 1 (n−m−1)/2 λi (λi − λj ) exp − λi Γm (m/2)Γm (n/2) 2 i
GOE : 2 N (m, n) : W

  1 2 ) (λi − λj ) exp − λi 2 i=1 Γ(i/2) i
1 Mn n/2

)  ) 2−mn π m(m−1) 1 λn−m λi (λi − λj )2 exp − i K K 2 Γm (n)Γm (m) i
GUE :   ) 2n(n−1)/2 Mn (λi − λj ) exp − λ2i Γ(i) i=1 i
π n/2

Table 17.1: Distribution functions for eigenvalues (f) Gaussian √ unitary ensemble (GUE): an m × m random matrix (A + H K A )/ 8 where A is G(m, m). Then (a) The joint densities of the eigenvalues λ1 ≥ · · · ≥ λm for the above random matrices are in Table 17.1 where the unlabeled sums and products run from i = 1 to m and the complex multivariate gamma function is defined by K m (a) = π m(m−1)/2 Γ

m )

Γ(a − i + 1)

(17.62)

i=1

For the Wishart cases, this is the joint distribution for the nonnegative eigenvalues. For the Gaussian cases, this is the joint distribution for the eigenvalues which may be anywhere on the real line. (b) The probability density function of the smallest eigenvalue of a matrix from W (m, m) is     m m−1 1 λ m+1 −1/2 −λm/2 fλmin (λ) = √ Γ e U λ ,− , 2 2 2 2 2π (17.63) Here, U (a, b, z) is the Tricomi function (see section 18.9.3).

c 2000 by Chapman & Hall/CRC 

(c) If κ is the condition number of a matrix from G(n, n), then the probability density function of κ/n converges to f (x) =

2x + 4 −2/x−2/x2 e x3

(17.64)

and E [log κ] = log n + c + o(1)

(17.65)

as n → ∞ where c ≈ 1.537. See A. Edelman, How many eigenvalues of a random matrix are real?, J. Amer. Math. Soc., 7, 1994, pages 247–267, and A. Edelman, Eigenvalues and condition numbers of random matrices, SIAM J. Matrix Anal. Appl., 9, 1988, pages 543–560. 17.10.1

Random matrix products

Let || · || be a non-negative real-valued function of matrices that satisfies ||AB|| ≤ ||A|| · ||B| when A and B are d × d matrices (i.e,, || · || could be a matrix norm). Let {Ai } be iid random d × d matrices in which all d2 elements of each Ai are iid normal random variables with mean 0 and variance 1. Then    1 d ||A1 A2 · · · At || lim log = 1+ψ (17.66) t→∞ t 2 2 where ψ is the digamma function (see page 518). See J. E. Cohen and C. M. Newman, The stability of large matrices and their products, Ann. Probab., 1984, 12, pages 283–310. 17.10.1.1

Vibonacci numbers

The Fibonacci numbers are defined by fn = fn−1 + fn−2 and f1 = f2 = 1. This formula can represented as the matrix product:      fn−1 0 1 fn−2 = (17.67) 1 1 fn−1 fn As n → ∞, the ratio

fn+1 fn

As n → ∞, the ratio

vn+1 vn

∼ φ ≈ 1.61803 . . . (the golden mean). The vibonacci numbers are defined by vn = vn−1 ± vn−1 where the ± sign is chosen randomly (50% of the time it is − and 50% of the time it is +). This formula can represented as the matrix product: & % & %  0 1 vn−2   50% of the time      1 1 vn−1 vn−1 = % (17.68) &% & vn   0 1 v  n−2   50% of the time  1 −1 vn−1 ∼ 1.1319882 . . . (Viswanath’s constant).

c 2000 by Chapman & Hall/CRC 

See B. Hayes, The vibonacci numbers, American Scientist, 87, July–August 1999, pages 296–301. 17.11

RANDOM NUMBER GENERATION

17.11.1

Methods of pseudorandom number generation

Depending on the application, either integers in some range or floating point numbers in [0, 1) are the desired output from a pseudorandom number generator (PRNG). Since most PRNGs use integer recursions, a conversion into integers in a desired range or into a floating point number in [0, 1) is required. If xn is an integer produced by some PRNG in the range 0 ≤ xn ≤ M − 1, then an in the range 0 ≤ xn ≤ N − 1, with N ≤ M , is given by P O integer yn = NMxn . If N # M , then yn = xn (mod N ) may be used. Alternately, if a floating point value in [0, 1) is desired, let yn = xn /M . 17.11.1.1

Linear congruential generators

Perhaps the oldest generator still in use is the linear congruential generator (LCG). The underlying integer recursion for LCGs is: xn = axn−1 + b

(mod M ).

(17.69)

Equation (17.69) defines a periodic sequence of integers modulo M starting with x0 , the initial seed. The constants of the recursion are referred to as the modulus, M , multiplier, a, and additive constant, b. If M = 2m , a very efficient implementation is possible. Alternately, there are theoretical reasons why choosing M prime is optimal. Hence, the only moduli that are used in practical implementations are M = 2m or the prime M = 2p − 1 (i.e., M is a Mersenne prime). With a Mersenne prime, modular multiplication can be implemented at about twice the computational cost of multiplication modulo 2p . Equation (17.69) yields a sequence {xn } whose period, denoted Per(xn ), depends on M , a, and b. The values of the maximal period for the three most common cases used and the conditions required to obtain them are: a primitive root of M 3 or 5 (mod 8) 1 (mod 4)

b anything 0 1 (mod 2)

M prime 2m 2m

Per(xn ) M −1 2m−2 2m

A major shortcoming of LCGs modulo a power-of-two compared with prime modulus LCGs derives from the following theorem for LCGs: Theorem 5 Define the following LCG sequence: xn = axn−1 + b (mod M1 ). If M2 divides M1 and if yn = xn (mod M2 ), then yn satisfies yn = ayn−1 + b (mod M2 ).

c 2000 by Chapman & Hall/CRC 

Theorem 5 implies that the k least-significant bits of any power-of-two modulus LCG with Per(xn ) = 2m = M has Per(yn ) = 2k , 0 < k ≤ m. Since a long period is crucial in PRNGs, when these types of LCGs are employed in a manner that makes use of only a few least-significant bits, their quality may be compromised. When M is prime, no such problem arises. Since LCGs are in such common usage, the table below contains a list of parameter values mentioned in the literature. The Park–Miller LCG is widely considered to be a minimally acceptable PRNG.

17.11.1.2

a

b

M

Source

75 131 16333 3432 171

0 0 25887 6789 0

231 − 1 235 215 9973 30269

Park–Miller Neave Oakenfull Oakenfull Wichman–Hill

Shift register generators

Another popular method of generating pseudorandom numbers is to use binary shift register sequences to produce pseudorandom bits. A binary shift register sequence (SRS) is defined by a binary recursion of the type: xn = xn−j1 ⊕ xn−j2 ⊕ · · · ⊕ xn−jk ,

j1 < j2 < · · · < jk = =.

(17.70)

where ⊕ is the exclusive “or” operation. Note that x ⊕ y ≡ x + y (mod 2). Thus the new bit, xn , is produced by adding k previously computed bits together modulo 2. The implementation of this recurrence requires keeping the last = bits from the sequence in a shift register, hence the name. The longest possible period is equal to the number of nonzero =-dimensional binary vectors, namely, 2A − 1. A sufficient condition for achieving Per(xn ) = 2A − 1 is that the characteristic polynomial corresponding to equation (17.70) be primitive modulo 2. Since primitive trinomials of nearly all degrees of interest have been found, SRSs are usually implemented using two-term recursions of the form: xn = xn−k ⊕ xn−A ,

0 < k < =.

(17.71)

In these two-term recursions, k is the lag and = is the register length. Proper choice of the pair (=, k) leads to SRSs with Per(xn ) = 2A − 1. Here is a list with suitable (=, k) pairs: (5,2) (31,3) 17.11.1.3

Primitive trinomial exponents (7,1) (7,3) (17,3) (17,5) (31,6) (31,7) (31,13 (127,1)

(17,6) (521,32)

Lagged-Fibonacci generators

Another way of producing pseudorandom numbers is to use the lagged-Fibonacci methods. The term “lagged-Fibonacci” refers to two term recurrences of the c 2000 by Chapman & Hall/CRC 

form: xn = xn−k & xn−A ,

0 < k < =,

(17.72)

where & refers to three common methods of combination: (1) addition modulo 2m , (2) multiplication modulo 2m , or (3) bit-wise exclusive ‘OR’ing of mlong bit vectors. Combination method (3) can be thought of as a special implementation of a two-term SRS. Using combination method (1) leads to additive lagged-Fibonacci sequences (ALFSs). If xn satisfies: xn = xn−k + xn−A

(mod 2m ),

0 < k < =,

(17.73)

. then the maximal period is Per(xn ) = (2 − 1)2 ALFS are especially suitable for producing floating point deviates using the real valued recursion yn = yn−k + yn−A (mod 1). This circumvents the need to convert from integers to floating point values and allows floating point hardware to be used. One caution with ALFS is that theorem 5 holds, and so the low-order bits have periods that are shorter than the maximal period. However, this is not nearly the problem as in the LCG case. With ALFSs the j least-significant bits will have period (2A − 1)2j−1 , so if = is large there really is no problem. Note that one can use the table of primitive trinomial exponents to find (=, k) pairs that give maximal period ALF sequences. A

17.11.2

m−1

Generating nonuniform random variables

In order to select random observations from an arbitrary distribution, suppose X J x has probability density function f (x) and distribution function F (x) = f (u) du. In the following we write “Y is U [0, 1)” to mean Y is uniformly −∞ distributed on the interval [0, 1). Two general techniques for converting uniform random variables into those from other distributions are: 1. The inverse transform method. If Y is U [0, 1), then the random variable X = F −1 (Y ) has probability density function f (x). 2. The acceptance–rejection method. Suppose the density can be written as f (x) = Ch(x)g(x) where h(x) is the density of a computableJ random variable, g(x) satisfies the inequality ∞ 0 < g(x) ≤ 1, and C −1 = −∞ h(u)g(u) du is a normalization constant. If X is U [0, 1), Y has density h(x), and if X < g(Y ), then X has density f (x). Therefore, generate {X, Y } pairs, reject both if X ≥ g(Y ) and return X if X < g(Y ). Examples of the inverse transform method: c 2000 by Chapman & Hall/CRC 

1. (Exponential distribution) The exponential distribution with rate λ has f (x) = λe−λx (for x ≥ 0) and F (x) = 1 − e−λx . Thus u = F (x) can be solved to give x = F −1 (u) = −λ−1 ln(1 − u). If U is U [0, 1) then so is 1 − U . Hence X = −λ−1 ln U is exponentially distributed with rate λ. 2. (Normal distribution) Let Zi be normally distributed with f (z) = √12π e−z  The polar transformation produces random variables R = Z12 + Z22 (exponentially distributed with λ = 2) and Θ√= arctan(Z2 /Z1 ) (uniformly√distributed on [0, 2π)). Inverting: Z1 = −2 ln X1 cos 2πX2 and Z2 = −2 ln X1 sin 2πX2 are normally distributed when X1 and X2 are U [0, 1). (This is the Box–Muller technique.) Examples of the rejection method: 1. (Exponential distribution with λ = 1) (a) Generate random numbers {Ui }N i=1 uniform on [0, 1]), stopping at N = min{n | U1 ≥ U2 ≥ Un−1 < Un }. (b) If N is even accept that run, and go to step (c). If N is odd reject the run, and return to step (a). (c) Set X equal to the number of failed runs plus the first random number in the successful run. 2. (Normal distribution) (a) Select two random variables (V1 , V2 ) from U [0, 1). Form R = V12 + V22 . (b) If R > 1 then reject the - (V1 , V2 ) pair and select another pair. ln R (c) If R < 1 then X = V1 −2 has a N (0, 1) distribution. R 3. (Normal distribution) (a) Select two exponentially distributed random variables with rate 1: (V1 , V2 ). (b) If V2 ≥ (V1 − 1)2 /2, then reject the (V1 , V2 ) pair and select another pair. (c) Otherwise, V1 has a a N (0, 1) distribution. 4. (Cauchy distribution) To generate values of X from f (x) = −∞ < x < ∞:

1 π(1+x2 )

on

(a) Generate random numbers U1 , U2 (uniform on [0, 1)) and set Y1 = U1 − 12 , Y2 = U2 − 12 . (b) If Y12 +Y22 ≤ 14 then return X = Y1 /Y2 . Otherwise return to step 1. To generate values of X from a Cauchy distribution with parameters β β and θ; f (x) = for −∞ < x < ∞; construct X as above π [β 2 + (x − θ)2 ] and then use βX + θ. c 2000 by Chapman & Hall/CRC 

2

/2

.

17.11.2.1

Discrete random variables

In general, the density function of a discrete random variable can be represented as a vector p = (p0 , p1 , . . . , pn−1 , pn ) by defining the probabilities Prob [x = j] = pj (for j = 0, . . . , n). The distribution function can be defined j by the vector c = (c0 , c1 , . . . , cn−1 , 1) where cj = i=0 pi . Given this representation of F (x) we can apply the inverse transform by computing X to be U [0, 1), and then finding the index j so that cj ≤ X < cj+1 . In this case event j will have occurred. For example: 1. (Binomial distribution) The binomial distribution with  n trials, with each trial having probability of success p, has pj = nj pj (1 − p)n−j for j = 0, . . . , n. (a) As an example, consider the result of flipping a fair coin. In 2 flips, the probability of obtaining (0, 1, 2) heads are p = ( 14 , 12 , 14 ). Hence c = ( 14 , 34 , 1). If x (chosen from U [0, 1)) turns out to be say, 0.4, then “1 head” is returned (since 14 ≤ 0.4 < 12 ). (b) Note that when n is large it is costly to compute the density and distribution vectors. When n is large and relatively few binomially distributed pseudorandom numbers are desired, an alternative is to use the normal approximation to the binomial. n (c) Alternately, one can form the sum i=1 'Ui + p(, where each Ui is U [0, 1). 2. (Geometric distribution) To simulate a value 7 8 from Prob [X = i] = p(1 − log U i−1 p) for i ≥ 1; use X = 1 + . log(1 − p) 3. (Poisson distribution) The Poisson distribution with mean λ has pj = λj eλ /j! for j ≥ 0. The Poisson distribution counts the number of events in a unit time interval if the times are exponentially distributed with rate λ. Thus if the times {Ti } are exponentially distributed with rate J λ, then J will be Poisson distributed with mean λ when i=0 Ti ≤ J+1 1 ≤ i=0 Ti . Since Ti = −λ−1 ln Ui , where Ui is U [0, 1), the previous MJ MJ+1 equation may be written as i=0 Ui ≥ e−λ ≥ i=0 Ui . Hence, we can random variables by iteratively computing PJ = MJ compute Poisson −λ . The first such J that makes this inequality i=0 Ui until PJ < e true will have a Poisson distribution. Random variables can be simulated using the following table (each U and Ui are uniform on the interval [0, 1)):

c 2000 by Chapman & Hall/CRC 

Using uniform random deviates to create random deviates from different distributions Distribution Binomial Cauchy

Density pj =

  n j p (1 − p)n−j j

f (x) =

π(x2

σ + σ2 )

Exponential

f (x) = λe−λx

Pareto

f (x) =

Rayleigh 17.11.2.2

aba xa+1

x −x2 /2σ2 e σ

Formula for deviate n 

'Ui + p(

i=1

σ tan(πU ) −λ−1 ln U b U 1/a √ σ − ln U

Testing pseudorandom numbers

The prudent way to check a complicated computation that makes use of pseudorandom numbers is to run it several times with different types of pseudorandom number generators and see if the results appear consistent across the generators. The fact that this is not always possible or practical has led researchers to develop statistical tests of randomness that should be passed by general purpose pseudorandom number generators. Some common tests are the spectral test, the equidistribution test, the serial test, the runs test, the coupon collector test (section 17.16.10), and the birthday spacing test (section 17.16.6). 17.11.3

References

1. L. Devroye, Non-Uniform Random Variate Generation, Springer–Verlag, New York, 1986. 2. S. K. Park and K. W. Miller, Random number generators: good ones are hard to find, Communications of the ACM, October 1988, Volume 31, Number 10, pages 1192–1201. 17.12

RESAMPLING METHODS

Assume the set S = {x1 , x2 , . . . , xn } contains n sample values of the random deviate X. Resampling techniques use the values in S repeatedly to obtain an estimate of a statistic, and also the variance of that estimate. For example, when computing the sample mean of the values in S, the value of the sample standard deviation indicates the possible range of values for the true mean. However, when computing the sample median, there is no natural way (other than resampling) to find the variance of the estimate. c 2000 by Chapman & Hall/CRC 

Suppose an estimate of the statistic θ(X) is desired. Let θ. be a procedure (estimator) used for estimating θ. 1. Bootstrap To apply the bootstrap technique, m random sets of the same size are drawn from the set S with replacement, and θ. is calculated for each sample. The bootstrapped estimate for θ(X) is the mean of the values of θ. for each random set. Consider a sample of four data points: {1, 3, 5, 9}. The estimated median from this sample is 4.

Example 17.80 :

To estimate the variability of the estimate of the median, repeatedly sample with replacement from the four data points. For example: (a) {5, 9, 9, 9}, median is 9

(f) {3, 9, 9, 5}, median is 7

(b) {1, 5, 1, 3}, median is 2

(g) {5, 3, 9, 1}, median is 4

(c) {9, 3, 3, 9}, median is 6

(h) {3, 3, 9, 3}, median is 3

(d) {3, 1, 5, 1}, median is 2

(i) {9, 5, 3, 3}, median is 4

(e) {5, 1, 9, 5}, median is 5

(j) {3, 9, 9, 1}, median is 6

These 10 estimates of the median have a mean of 4.8 and a standard deviation 2.3. This indicates the approximate variability of the estimate of the median.

2. Jackknife To apply the jackknife technique, θ. is computed on the set S and also on the n − 1 sets obtained from S by sequentially deleting each element. Hence, θ. is computed for 1 sample of size n and n samples of size n − 1. The jackknife estimate of θ(X) is the mean of the n values of θ. that have been obtained. See B. Efron, ”Bootstrap methods, another look at the jackknife,” Annals of Statistics, 7, 1979, pages 1–26 and B. Efron, The Jackknife, the Bootstrap and Other Resampling Plans, Society for Industrial and Applied Mathematics, Philadelphia, 1982. 17.13 17.13.1

SELF-SIMILAR PROCESSES Definitions

(a) A function L(x) is slowly varying if, for all x > x0 , lim

t→∞

L(tx) = 1. L(t)

(17.74)

Slowly varying functions include L(x) = c + o(1) for x > 0, L(x) = log x for x > 1, and L(x) = 1/ log x for x > 1. c 2000 by Chapman & Hall/CRC 

(b) A random variable X has a heavy tailed distribution if Prob [X > x] = x−α L(x) for α > 0 and x > x0 where L(x) is a slowly varying function. 17.13.2

Self-similar processes

A process {Xt }t=0,1,2,... is asymptotically self-similar if the autocorrelation function, r(k), has the form r(k) ∼ k −(2−2H) L(k)

as k → ∞

(17.75)

where L(x) is a slowly varying function and the Hurst parameter H satisfies 1/2 < H < 1. The process is exactly self-similar if  1 r(k) = (k + 1)2H − 2k 2H + (k − 1)2H . (17.76) 2 Note: White noise has r(k) = 0, which corresponds to H = 1/2. (m)

For any process {Xt }t=0,1,2,... , the aggregated version {Xt }t=0,1,2,... is constructed by partitioning {Xt } into non–overlapping blocks of m sequential (m) elements and constructing a single element Xt from the mean: (m)

Xt

=

tm  1 Xi m i=tm−m+1

(17.77)

(m)

Note: {Xt } represents viewing {Xt } on a time scale that is a factor of m coarser. (m) For a typical process, as m increases the autocorrelation of {Xt } decreases (m) until, in the limit, the elements of {Xt } are uncorrelated. For a self-similar (m) process, the processes {Xt } and {Xt } have the same autocorrelation function. 17.14 17.14.1

SIGNAL PROCESSING Estimation

Let {et } be a white noise process (so that E [et ] = µ, Var [et ] = σ 2 , and Cov [et , es ] = 0 for s = t). Suppose that ∞{Xt } is a time series. A nonanticipating linear model presumes that u=0 hu Xt−u = et , where {hu } the ∞ are constants. This can be written H(z)Xt = et where H(z) = u=0 hu z u and z n Xt = Xt−n . Alternately, Xt = H −1 (z)et . In practice, several types of models are used: 1. AR(k): autoregressive model of order k. This assumes that H(z) = 1 + a1 z + · · · + ak zk and so Xt + a1 Xt−1 + · · · + ak Xt−k = et 2. MA(l): moving average of order l. This assumes that c 2000 by Chapman & Hall/CRC 

(17.78)

H −1 (z) = 1 + b1 z + · · · + bk zk and so Xt = et + b1 et−1 + · · · + bl et−l

(17.79)

3. ARMA(k, l): mixed autoregressive/moving average of order (k, l). This 1+b1 z+···+bk zk assumes that H −1 (z) = 1+a and so 1 z+···+ak zk Xt + a1 Xt−1 + · · · + ak Xt−k = et + b1 et−1 + · · · + bl et−l 17.14.2

(17.80)

Matched filtering (Wiener filter)

Let S(t) represent a signal to be recovered, let N (t) represent noise, and let Y (t) =  S(t) + N (t) represent the observable. A prediction of the signal is ∞

Sp (t) = K(z)Y (t − z) dz, where K(z) is a filter. The mean square error is 0   E (S(t) − Sp (t))2 ; this is minimized by the optimal (Wiener) filter Kopt (z). When X and Y are stationary, define their autocorrelation functions to be RXX (t − s) = E [X(t)X(s)] and RY Y (t − s) = E [Y (t)Y (s)]. If F represents the Fourier transform, then the optimal filter is given by F [Kopt (t)] =

1 F [RXX (t)] 2π F [RY Y (t)]

(17.81)

For example, if X and N are uncorrelated, then F [Kopt (t)] =

1 F [RXX (t)] 2π F [RXX (t)] + F [RN N (t)]

In the case of no noise: F [Kopt (t)] = 17.14.3

1 2π ,

(17.82)

Kopt (t) = δ(t), and Sp (t) = Y (t).

Median filter

A median filter replaces a value in a data set with the median of the entries surrounding that value. In the one-dimensional case it consists of sliding a window of an odd number of elements along the signal, replacing the center sample by the median of the samples in the window. The median is a stronger “central indicator” than the average. The median is hardly affected by one or two discrepant values among the data values in the region. Consequently, median filtering is very effective at removing various kinds of noise. In two-dimensional data sets (such as images) median filtering is a nonlinear signal enhancement technique for the smoothing of signals, the suppression of impulse noise, and preserving of edges. 17.14.4

Mean filter

A mean filter or averaging filter replaces the values in a data set with the average of the entries surrounding that value. Thought of as a convolution filter, it is represented by a kernel which represents the shape and size of the neighborhood to be sampled when calculating the mean. In two-dimensional data sets (such as images) a 3 × 3 square kernel is often used: c 2000 by Chapman & Hall/CRC 

1/9

1/9

1/9

1/9

1/9

1/9

1/9

1/9

1/9

A problem with the mean filter is that it blurs edges and other sharp details. An alternative is to use a median filter. 17.14.5

Spectral decomposition of stationary random functions

Any stationary function X(t) can be written as  ∞ X(t) − µX = eiωt dΦ(ω) −∞





If the correlation function satisfies the equation −∞

increments dΦ(ω) satisfy

(17.83)

|KX (t)| dt < ∞ then the

E [dΦ(ω)] = 0 ∗

E [dΦ (ω1 ) dΦ(ω2 )] = SX (ω)δ(ω1 − ω2 ) dω1 dω2

(17.84)

where ∗ denotes the complex conjugate. Here, SX (ω) is the spectral density of X(t) and δ(x) denotes the δ-function. The correlation function and the spectral density are related by mutually inverse Fourier transforms:  ∞  ∞ 1 iωt KX (t) = e SX (ω) dω, and SX (ω) = e−iωt KX (t) dt 2π −∞ −∞ (17.85) Note that: (a) If X(t) is a real function, then SX (ω) = SX (−ω). (b) The spectral density of X  (t) is related to SX (ω) by: SX  (ω) = ω 2 SX (ω) 17.15 17.15.1

(17.86)

STOCHASTIC CALCULUS Brownian motion (Wiener processes)

Brownian motion W (t) is a Gaussian random process that has  a mean givenby its starting point, E [W (t)] = W0 = W (t0 ), a variance of E (W (t) − W0 )2 = t − t0 , and a covariance of E [W (t)W (s)] = min(t, s). The sample paths of W (t) are continuous but not differentiable. Brownian motion is also called a Wiener process. Formally: Let (Ω, B, Pr) be a Lebesgue probability space, and let (R, F, m) represent the real numbers with Lebesgue measure m. Then a Brownian motion is a function X(t, ω) : R+ × Ω → R satisfying three conditions: c 2000 by Chapman & Hall/CRC 

1. For any 0 < s < t, (X(t, ω) − X(s, ω)) has a Gaussian distribution with mean zero and variance m([s, t)); 2. If t0 < t1 < · · · < tk , then (X(tj , ω) − X(tj−1 , ω))j=1,2,...,k is an independent system; 3. X(0, ω) = 0 for all ω ∈ Ω. 17.15.2

Brownian motion expectations

Define the following types of Brownian motions: 1. Brownian motion Ws (µ) 2. Brownian motion with drift Ws = µs + Ws 3. Reflecting Brownian motion |Ws | For each, let the Brownian motion start at the location x. Define the following types of stopping times at which a process X will stop: 1. An exponential stopping time is the time τ given by Prob [τ > t] = e−λt for λ > 0. (Here τ is independent of the process X). 2. A first hitting time is the first time that a process reaches some value; for example, Hz = min{s | Ws = z} is the first time when a Brownian motion reaches the value z. 3. A first exit time is the first time that a process leaves a region; for example, Ha,b = min{s | Ws ∈ (a, b)} is the first time that a Brownian motion exits the interval (a, b). We have the following expectations involving Brownian motion: 1. Brownian motion Ws : (a) Unconstrained     2 i. E eiβWt = exp iβx − β2 t -      γ2t t ii. E exp −γ sup Ws = exp −γx + erfc γ 2 2 0
(c) First hitting time: √   i. E e−αHz = e−|x−z| 2α %    J∞ = e−γx 1 − γ(x − z) ii. E exp −γ sup Ws 0
for z ≤ x   iii. E exp γ

inf

% = eγx 1 − γ(z − x)

Ws



iv. E exp −γ

H Jz

J∞ z−x

x≤z %

e

x−z

 0
&

  √2γ βx  I e   0  √|β| ,  2γ βz   I0 |β| e

& e2βWs ds

0

=

−γv

v

& −γv

e

v

dv

(z − x)β ≥ 0

√   2γ βx  K0 |β| e     √2γ βz  , (z − x)β ≤ 0 K0

|β|

e

(d) First exit time  √    cosh (b−2x+a) α/2  √  i. E e−αHa,b = cosh (b−a) α/2 # " βWHa,b b−x iβa iβb = b−a e + x−a ii. E e b−a e (µ)

2. Brownian motion with drift Ws

= µs + Ws

(a) Unconstrained " #  (µ) i. E eiβWt = exp iβ(x + µt) −

β2 t 2



(b) Exponential stopping times " # (µ) 2λ i. E eiβWτ = eiβx 2λ − 2iβµ + β 2     2λ + µ2 − µ −γx (µ)  = ii. E exp −γ sup Ws e 0
(c) First hitting time (Hz = min{s | Ws = z})      i. E e−αHz = exp µ(z − x) − |z − x| 2α + µ2 (µ)

(d) First exit time (Ha,b = min{s | Ws ∈ (a, b)})   µ(a−x) sinh((b−x)η)+eµ(b−x) sinh((x−a)η) i. E e−αHz = e sinh((b−a)η)  2 where η = 2α + µ   (µ) sinh ((b − x)|µ|) µ(a−x)+iβa iW ii. E e Ha,b = e sinh ((b − a)|µ|) sinh ((x − a)|µ|) µ(b−x)+iβb + e sinh ((b − a)|µ|) 3. Reflecting Brownian motion |Ws | (a) Exponential stopping times √ " #  √ 2λ √ −β|Wτ | −βx − 2λx i. E e 2λe − βe = for 0 ≤ x 2λ − β 2 c 2000 by Chapman & Hall/CRC 

dv

for

See A. N. Borodin and P. Salminen, Handbook of Brownian Motion — Facts and Formulae, Birkh¨ auser Verlag, Boston, 1996. 17.15.3

Itˆ o lemma

Consider the process f (W ) = {F (Wt ) | t ≥ 0} where f is a given smooth function and W is a Brownian motion. When f has two derivatives, the Itˆ o formula is given by:  t  1 t  f (Wt ) − f (W0 ) = f  (Ws ) dWs + f (Ws ) ds (17.87) 2 0 0 17.15.4

Stochastic integration

If W (t) is a Wiener process and G(t, W (t)) is an arbitrary function, then Jt the stochastic integral I = t0 G(s, W (s)) dW (s) is defined as a limiting sum. Divide the interval [t0 , t] into n sub-intervals: t0 ≤ t1 ≤ · · · ≤ tn−1 ≤ tn = t, and choose points {τi } that lie in each sub-interval: ti−1 ≤ τi ≤ ti . The stochastic integral I is defined as the limit of partial sums, I = limn→∞ Sn , n with Sn = i=1 G(τi , W (τi ))[W (ti ) − W (ti−1 )]. Consider, for example, the special case of G(t) = W (t). Then the expectation of Sn is % n &  E [Sn ] = E W (τi )[W (ti ) − W (ti−1 )] i=1

= =

n  i=1 n 

[min(τi , ti ) − min(τi , ti−1 )]

(17.88)

(τi − ti−1 ).

i=1

n If τi = αti + (1 − α)ti−1 (where 0 < α < 1), then E [Sn ] = i=1 (ti − ti−1 )α = (t − t0 )α. Hence, the value of Sn depends on α. For consistency, some specific choice must be made for the points {τi }. J 1. For the Itˆ o stochastic integral (indicated by I ), we choose τi = ti−1 (i.e., α = 0 in the above). That is:  t G(s, W (s)) dW (s) I t0 ( n  (17.89)  = ms-lim G(ti−1 , W (ti−1 ))[W (ti ) − W (ti−1 )] , n→∞

i=1

where “ms-lim” refers to the mean square limit.

c 2000 by Chapman & Hall/CRC 

J 2. For the Stratonovich stochastic integral (indicated by S ), we choose τi = (ti + ti−1 )/2 (i.e., α = 1/2 in the above). That is:  t G(W (s), x) dW (s) (17.90) S t0 ( n      ti + ti−1 = ms-lim [W (ti ) − W (ti−1 )] . G ti−1 , w n→∞ 2 i=1 Example 17.81 :

Jt

W (s) dW (s). The Itˆ o and Stratonovich t0 evaluations are   Jt 1. I t0 W (s) dW (s) = W 2 (t) − W 2 (t0 ) − (t − t0 ) /2 .   Jt 2. S t0 W (s) dW (s) = W 2 (t) − W 2 (t0 ) /2.

17.15.5

Consider the integral

Stochastic differential equations

If a differential equation contains random terms, then the solution can only be described statistically. A linear ordinary differential equation for, say, x(t) with linearly appearing white Gaussian noise terms can be converted to a parabolic partial differential equation whose solution is the probability density of x(t). This equation is called a Fokker–Planck equation or a forward Kolmogorov equation. Consider the linear differential system for the m component vector X(t) d X(t) = b(t, X) + σ(t, X) N(t), dt X(t0 ) = y,

(17.91)

where σ(t, X) is a real m × n matrix and N(t) is a vector of n independent white noise terms. That is, E [Ni (t)] = 0, E [Ni (t)Nj (t + τ )] = δij δ(τ ),

(17.92)

where δij is the Kronecker delta, and δ(τ ) is the delta function. The Fokker– Planck equation corresponding to equation (17.91) is m m  ∂P ∂ ∂2 1  (bi P ) + (aij P ) , =− ∂t ∂xi 2 i,j=1 ∂xi ∂xj i=1

(17.93)

where P = P (t, x) is a probability density and the matrix A = (aij ) is defined by A(t, x) = σ(t, x)σ T (t, x). Any statistical information about X(t) that could be ascertained from equation (17.91) may be derived from P (t, x). In one dimension, the stochastic differential equation dX = f (X) + g(X)N (t), dt c 2000 by Chapman & Hall/CRC 

(17.94)

with X(0) = z, corresponds to the Fokker–Planck equation ∂P ∂ 1 ∂2 2 (g (x)P ), = − (f (x)P ) + ∂t ∂x 2 ∂x2 for P (t, x) with P (0, x) = δ(x − z).

(17.95)

Example 17.82 : The Langevin equation X  + βX  = N (t),

(17.96)



with the initial conditions X(0) = 0 and X (0) = u0 can be written as the vector system        d X U 0 0 N1 (t) = + , −βU 0 1 N2 (t) dt U (17.97)     X 0 = . U t=0 u0 The Fokker–Planck equation for P (t, x, u), the joint probability density of X and U at time t, is ∂P ∂ ∂ 1 ∂2P , = − (uP ) + (βuP ) + ∂t ∂x ∂u 2 ∂u2 P (0, x, u) = δ(x)δ(u − u0 ).

(17.98)

This equation has the solution    T 1 x − µx x − µx D , exp − P (t, x, u) = u − µu u − µu det D  where D =

 σxx σxu , and the parameters {µx , µu , σxx , σxu , σuu } are given by σxu σuu  u0  µx = 1 − e−βt , β µu = u0 e−βt ,   t 2  1  2 σxx = 2 − 3 1 − e−βt + 1 − e−2βt , 3 β β 2β   1  1  2 −βt −2βt σxu = 2 1 − e − 1 − e , β 2β 2   1 2 σuu = 1 − e−2βt . 2β

17.15.6

(17.99)

(17.100)

Motion in a domain

Consider a particle starting at y and randomly moving in a domain Ω. If the probability density of the location evolves according to m m  ∂P ∂P 1  ∂2P = L[P ] = − bi (y) + aij (y) , ∂t ∂yi 2 i,j=1 ∂yi ∂yj i=1

then c 2000 by Chapman & Hall/CRC 

(17.101)

(a) The expectation of the exit time w(y) is the solution of L[w] = −1 in Ω, with w = 0 on ∂Ω. (b) The probability u(y) that the exit occurs on the boundary segment Γ is the solution ( of L[u] = 0 in Ω with 1 for y ∈ Γ u(y) = . 0 for y ∈ Ω/Γ 17.15.7

Option Pricing

Let S represent the price of a share of stock, and presume that S follows a geometric Brownian motion dS = µS dt + σS dω, where t is time, µ is a constant, and σ is a constant called the volatility. Let V (S, t) be the value of a derivative security whose payoff is solely a function of S and t. Construct a portfolio consisting of V and ∆ shares of stock. The value P of this portfolio is P = V + S∆. The random component of the portfolio increment (dP ) can be removed by choosing ∆ = −∂V /∂S. The concept of arbitrage says that dP = rP dt, where r is the (constant) risk-free bank interest rate. Together this results in the Black–Scholes equation for option pricing (note that no transaction costs are included): ∂V ∂2V ∂V 1 (17.102) + rS + σ 2 S 2 2 − rV = 0 ∂t ∂S 2 ∂S If the asset pays a continuous dividend of DS dt (i.e., this is proportional to the asset value S during the time period dt), then the equation is modified to become ∂V ∂V 1 ∂2V + (r − D)S + σ 2 S 2 2 − rV = 0 (17.103) ∂t ∂S 2 ∂S If E is the exercise price of the option, and T is the expiry date (the only date on which the option can be exercised), then the solution of the modified Black–Scholes equation is V (S, t) = e−D(T −t) SN (d1 ) − Ee−r(T −t) Φ(d2 )

(17.104)

where d1 =

log(S/E) + (r − D + σ 2 /2)(T − t) √ , σ T −t

√ d2 = d1 − σ T − t, (17.105)

and Φ is the cumulative probability distribution for the normal distribution.

c 2000 by Chapman & Hall/CRC 

17.16 17.16.1

CLASSIC AND INTERESTING PROBLEMS Approximating a distribution

Given the moments of a distribution (µ, σ, and {γi }), the asymptotic probability density function is given by   γ1 γ2 (4) γ2 f (t) = φ(x) − φ(3) (x) + φ (x) + 1 φ(6) (x) 6 24 72 (17.106)   γ3 (5) γ1 γ2 (7) γ3 − φ (x) + φ (x) + 1 φ(9) (x) + . . . 120 144 1296 where φ(x) = 17.16.2

2 2 1 √ e−(x−µ) /2σ is the normal density function. σ 2π

Averages over vectors

Let f (n) denote the average of the function f as the unit vector n varies uniformly in all directions in three dimensions. If a, b, c, and d are constant three-dimensional vectors, then 2

2

|a · n| = |a| /3 (a · n)(b · n) = (a · b)/3 (a · n)n = a/3 2

(17.107)

2

|a × n| = 2 |a| /3 (a × n) · (b × n) = 2a · b/3 (a · n)(b · n)(c · n)(d · n) = [(a · b)(c · d) + (a · c)(b · d) + (a · d)(b · c)]/15

Now let f (n) denote the average of the function f as the unit vector n varies uniformly in all directions in two dimensions. If a and b are constant twodimensional vectors, then 2

2

|a · n| = |a| /2 (a · n)(b · n) = (a · b)/2

(17.108)

(a · n)n = a/2 17.16.3

Bertrand’s box “paradox”

Suppose there are three small boxes, each with two drawers, each containing a coin. One box has two gold coins, another has two silver coins, and the third has one gold and one silver coin. Suppose a box and a drawer are selected at random, and a gold coin is discovered. What is the probability that the other drawer also contains a gold coin?

c 2000 by Chapman & Hall/CRC 

If a gold coin is found, even though the box selected is not the one with two silver coins, it is not equally likely to be one of the remaining two boxes. Given a gold coin has been found, there are two ways this could happen if selecting a drawer from the G/G box, and only one way if selecting from the the G/S box. Therefore, the probability the box is G/G is 2/3 and the probability the other coin in the selected box is gold is 2/3. 17.16.4

Bertrand’s circle “paradox”

Consider the following problem: Given a circle of unit radius, find the probability that a randomly chosen chord will be longer than the side of an inscribed equilateral triangle. There are at least three solutions to this problem: (a) Randomly choose two points on a circle and measure the distance between them. The result depends only on the position of the second point relative to the first one (i.e., the first point can be fixed). Consider chords that emanate from a fixed first point; 1/3 of the resulting chords will be longer that the side of an equilateral triangle. Therefore, the probability is 1/3. (b) A chord is completely specified by its midpoint. If the length of a chord exceeds the side of an equilateral triangle, then the midpoint must be inside a smaller circle of radius 1/2. The area of the smaller circle is 1/4 of the area of the original (unit) circle. Therefore, the probability is 1/4. (c) A chord is completely specified by its midpoint. If the length of a chord exceeds the side of an equilateral triangle, then the midpoint is within 1/2 of the center of the circle. If the midpoints are distributed uniformly over the radius (instead of over the area, as assumed in ((b)), the probability is 1/2. It all depends on how randomly is defined. See M. Kac and S. M. Ulam, Mathematics and Logic, Dover Publications, NY, 1968. 17.16.5

Bingo cards: nontransitive

Consider the 4 bingo cards shown below (labeled A, B, C, and D), which were created by D. E. Knuth. Two players each select a bingo card. Numbers from 1 to 6 are randomly drawn without replacement, as they are in standard bingo. If a selected number is on a card, it is marked with a bean. The first player to complete a horizontal row wins. It can be shown that, statistically, card A beats card B, card B beats card C, card C beats card D, and card D beats card A. A 1 3

B 2 4

c 2000 by Chapman & Hall/CRC 

2 5

C 4 6

1 4

D 3 5

1 2

5 6

17.16.6

Birthday problem

The probability that n people have different birthdays is       364 363 366 − n qn = · ··· 365 365 365

(17.109)

Let pn = 1 − qn . For 23 people the probability of at least two people having the same birthday is more than half (p23 = 1 − q23 > 1/2). n pn

10 0.117

20 0.411

23 0.507

30 0.706

40 0.891

50 0.970

The number of people needed to have a 50% chance of two people having the same birthday is 23. The number of people needed to have a 50% chance of three people having the same birthday is 88. For four, five, and six people having the same birthday the number of people necessary is 187, 313, and 460. The number of people needed so that there is a 50% chance that two people have a birthday within one day of each other is 14. In general, in a n-day year the probability that p people all have birthdays at least k days apart (so k = 1 is the original birthday problem) is   n − p(k − 1) − 1 (p − 1)! probability = (17.110) p−1 np−1 The non-equiprobable case was solved in M. Klamkin and D. Newman, Extensions of the birthday surprise, J. Comb. Theory, 3 (1967), pages 279–282. 17.16.7

Buffon’s needle problem

A needle of length L is placed at random on a plane on which are drawn parallel lines a distance D apart. Assume that L < D so that only one intersection is possible. The probability P that the needle intersects a line is 2L P = (17.111) πD If a convex object of perimeter s, with a maximum diameter less than D, is randomly placed on the ruled plane, then the probability P that the object intersects a line is P = s/πD. If a needle of length L is dropped on a rectilinear ruled grid, with line separations of a and b (with L < a and L < b), then the probability that the needle 2L(a + b) − L2 will land on a line is . abπ 17.16.8 Card problems 17.16.8.1

Shuffling cards

A riffle shuffle takes a deck of n cards, splits it into two stacks, and then recombines them. We assume that the probability that the split occurs after c 2000 by Chapman & Hall/CRC 

the k th card has a binomial distribution. The recombination occurs by taking cards, one by one, off of the bottom of the two stacks, and placing them onto one stack; if there are k1 and k2 cards in the two stacks, then the probability 1 of taking the next card from the first (resp. second) stack is k1k+k (resp. 2 k2 k1 +k2 ). Let Ωn = {π1 , π2 , . . . , πn! } denote the possible permutations of n cards. Let fX (πi ) be the probability that operator X (a riffle shuffle) produces the ordering πi . The variation distance between the process X and a uniform distribution on Ωn is $ $ 1  $$ 1 $$ ||fX − u|| = (π) − f (17.112) $ X 2 n! $ π∈Ωn

where the 12 is used so that value falls in the interval [0, 1]. For the riffle shuffle equation (17.112) becomes $ k $  n $ 2 +n−r 1 1 1 $$ ||fX − u|| = (17.113) An,r $$ − n 2 r=1 2nk n! $ where the Eulerian numbers {An,r } are defined by An,1 = 1 and a−1  n + a − r  n An,a = a − An,r n r=1

(17.114)

Table 17.2 contains numerical values for a n = 52 card deck. Until 5 shuffles have occurred, the output of X is very far from random. After 5 shuffles, the distance from the random process is essentially halved each time a shuffle occurs. See D. Bayer and P. Dianconis, Trailing the dovetail shuffle to its lair, Annals of Applied Probability, 2(2), 1992, pages 294–313. 17.16.8.2

Card games

(a) Poker hands   The number of distinct 13-card poker hands is 52 5 = 2,598,960. Hand royal flush straight flush four of a kind full house flush straight three of a kind two pair one pair

c 2000 by Chapman & Hall/CRC 

Probability

Odds

−6

649,739:1 72,192:1 4,164:1 693:1 508:1 254:1 46:1 20:1 1.37:1

1.54 × 10 1.39 × 10−5 2.40 × 10−4 1.44 × 10−3 1.97 × 10−3 3.92 × 10−3 0.0211 0.0475 0.423

Number of riffle shuffles 1 2 3 4 5 6 7 8 9 10 11 12 13 14

Variation distance 1 1 1 0.99999953 0.924 0.614 0.334 0.167 0.0854 0.0429 0.0215 0.0108 0.00538 0.00269

Table 17.2: Number of riffle shuffles versus variation distance

(b) Bridge hands   The number of distinct 13-card bridge hands is 52 13 = 635,013,559,600. In bridge, the honors are the ten, jack, queen, king, or ace. Obtaining the three top cards (ace, king, and queen) of three suits and the ace, king, queen, and jack of the remaining suit is called 13 top honors. Obtaining all cards of the same suit is called a 13-card suit. Obtaining 12 cards of the same suit with ace high and the 13th card not an ace is called a 12-card suit, ace high. Obtaining no honors is called a Yarborough. Hand Probability Odds 13 top honors 13-card suit 12-card suit, ace high Yarborough four aces nine honors 17.16.9 17.16.9.1

6.30 × 10−12 6.30 × 10−12 2.72 × 10−9 5.47 × 10−4 2.64 × 10−3 9.51 × 10−3

158,753,389,899:1 158,753,389,899:1 367,484,698:1 1,827:1 378:1 104:1

Coin problems Even odds from a biased coin

If you have a coin biased toward heads, it is possible to get the equivalent of a fair coin with several tosses of the unfair coin. Toss the biased coin twice. If both tosses give the same result, repeat this process (throw out the two tosses and start again). Otherwise, take the first of the two results as the toss of an unbiased coin.

c 2000 by Chapman & Hall/CRC 

17.16.9.2

Two heads in a row

What is the probability in n flips of a fair coin that there will be two heads in a row? Consider strings of n H’s and T ’s. We count the number of strings that do not contain HH and subtract it from the total number of such strings (which is 2n ). There must be no more than n/2 H’s; otherwise two heads would be adjacent. If a string contains i H’s, with no two of them in a row, then these H’s must be placed between (or around) the (n − i) T ’s present:  there are n−i+1 ways to do this. Hence, the total number of strings that do i n/2   n − i + 1 not contain HH is = Fn+2 (where Fm is the mth Fibonacci i i=0 number). Hence, the probability that n coin tosses will contain a HH is: 2n − Fn+2 . 2n 17.16.10

Coupon collectors problem

There are n coupons that can be collected. At each time a random coupon is selected, with replacement. How long must one wait until they have a specified collection of coupons? Define Wn,j to be the number of time steps required until j different coupons are seen; then E [Wn,j ] = n

j  i=1

Var [Wn,j ] = n

j  i=1

1 n−i+1

(17.115)

i−1 (n − i + 1)2

When j = n, then all coupons are being collected and E [Wn,n ] = nHn with Hn = 1 + 1/2 + 1/3 + · · · + 1/n. As n → ∞, E [Wn,n ] ∼ n log n. Typical values are shown below.  n E [Wn,n ] Var [Wn,n ] 2 5 10 50 100 200

c 2000 by Chapman & Hall/CRC 

3 11.4 29.3 225 519 1,176

1.41 5.02 11.2 62 126 254

In the unequal probabilities case, with pi being the probability of obtaining coupon i, &  ∞% n )   −pi t E [Wn,n ] = 1−e dt (17.116) 1− 0

i=1

17.16.11

Dice problems

17.16.11.1

Dice: nontransitive (Efron)

(a) Let A have the {2, 3, 3, 9, 10, 11} (b) Let B have the {0, 1, 7, 8, 8, 8} (c) Let C have the {5, 5, 6, 6, 6, 6} (d) Let D have the {4, 4, 4, 4, 12, 12}

two six-sided dice with sides of {0, 0, 4, 4, 4, 4} and two six-sided dice with sides of {3, 3, 3, 3, 3, 3} and two six-sided dice with sides of {2, 2, 2, 2, 6, 6} and two six-sided dice with sides of {1, 1, 1, 5, 5, 5} and

Then the odds of (a) A winning against B (b) B winning against C

(c) C winning against D (d) D winning against A

are all 2:1. 17.16.11.2

Dice: distribution of sums

When rolling two dice, the probability distribution of the sum is 6 − |s − 7| for 2 ≤ s ≤ 12 36 When rolling three dice, the probability distribution of the sum is   1 (s − 1)(s − 2) for 3 ≤ s ≤ 8 1 2 2 Prob [sum of s] = −s + 21s − 83 for 9 ≤ s ≤ 14 216   [1 2 (19 − s)(20 − s) for 15 ≤ s ≤ 18 Prob [sum of s] =

(17.117)

(17.118)

For 2 dice, the most common roll is a 7 (probability 1/6). For 3 dice, the most common rolls are 10 and 11 (probability 1/8 each). For 4 dice, the most common roll is a 14 (probability 73/648). 17.16.11.3

Dice: same distribution

Ordinary six-sided dice have the numbers {1, 2, 3, 4, 5, 6} on the sides. Rolling two dice and summing the numbers face up creates a distribution of values from 2 to 12 (see equation (17.117)). That same distribution of sums can be obtained by rolling two cubes, with the following numbers on their sides: c 2000 by Chapman & Hall/CRC 

• {0, 1, 1, 2, 2, 3} and {2, 4, 5, 6, 7, 9} • {0, 2, 3, 4, 5, 7} and {2, 3, 3, 4, 4, 5} • {1, 2, 2, 3, 3, 4} and {1, 3, 4, 5, 6, 8} (Sicherman dice) 17.16.12

Ehrenfest urn model

Suppose there are two urns and 2R numbered balls, and suppose that there are R + a balls in urn I and R − a balls in urn II. Each second a ball is chosen at random and moved from its urn into the other urn. The procedure is repeated. Consider the case of R = 10, 000. If a = 0, then the expected time to return to the initial configuration is 175 seconds. If a = R, then the expected time to return to the initial configuration is approximately 106000 years. 17.16.13

Envelope problem “paradox”

The envelope exchange problem is “Someone has prepared two envelopes containing money. One contains twice as much money as the other. You have decided to pick one envelope, but then the following argument occurs to you: ‘Suppose my chosen envelope contains X, then the other envelope either contains X/2 or 2X. Both cases are equally likely, expectation if I take the other en so my 1 velope is 12 X (2X) = 1.25X, which is higher than + 2 2 my current X. Hence, I should change my mind and take the other envelope.’ But then I can apply the argument all over again. There must be sometime wrong.” There is, in fact, no contradiction. Switching the envelope or not switching the envelope results in the same expected return. See R. Christensen and J. Utts, Bayesian resolution of the ‘Exchange Paradox’, The American Statistician, 46 (4), 1992, pages 274–276. 17.16.14

Gambler’s ruin problem

A gambler starts with z dollars. For each gamble: with probability p she wins one dollar, with probability q she loses one dollar. Gambling stops when she has either zero dollars, or a dollars.

c 2000 by Chapman & Hall/CRC 

If qz denotes the probability of eventually stopping at z = a (“gambler’s success”) then  (q/p)a − (q/p)z   if p = q   (q/p)a − 1 qz = (17.119)   z  1 − if p = q = 12 a For example: fair game

biased game 17.16.15

p 0.5 0.5 0.5 0.5 0.4 0.4

q 0.5 0.5 0.5 0.5 0.6 0.6

z 9 90 900 9000 90 90

a 10 100 1000 10000 100 99

qz .900 .900 .900 .900 .017 .667

Gender distributions

For these problems, there is a 50/50 chance of male or female on each birth. 1. Hospital deliveries: Every day a large hospital delivers 1000 babies and a small hospital delivers 100 babies. Which hospital has a better chance of having the same number of boys as girls? The small If 2n babies are born, then the probability of an even 2none.  n . This is a decreasing function of n. As n goes to infinity 22n the probability of an even split approaches zero (although it is still the most likely event).

split is

2. Family planning Suppose that in large society of people, every family continues to have children until they have a girl, then they stop having children. After many generations of families, what is the ratio of males to females? The ratio will be 50-50; half of all conceptions will be male, half female. 3. If a person has two children and (a) The older one is a girl, then the probability that both children are girls is 1/2. (b) At least one is a girl, then the probability that both children are girls is 1/3.

c 2000 by Chapman & Hall/CRC 

17.16.16

Holtzmark distribution: stars in the galaxy

The force f on a unit mass due to a star of mass m at a location r is f = Gm r. |r|3 The probability density for the magnitude of the force, f = |f |, due to a uniform distribution of stars in the galaxy, with random masses, with λ stars per unit volume, is given by 1 H(β) 4πa2 β 2 # " 4λ 3/2 a= (2πG) E m3/2 15 β = f /a2/3  ∞ 3/2 2 H(β) = x e−(x/β) sin x dx πβ 0

(17.120)

 4β 2 as β → 0  3π + O(β 4 )  H(β) ∼  15 2 β −5/2 + O(β −4 ) as β → ∞ 8 π

(17.121)

p(f ) =

Note that

See S. Chandrasekhar, Stochastic problems in physics and astronomy, Reviews of Modern Physics, 15, number 1, January 1943, pages 1–89. 17.16.17

Large-scale testing

17.16.17.1

Infrequent success

Suppose that a disease occurs in one person out of every 1000. Suppose that a test for this disease has a type I and a type II error of .01 (that is, α = β = 0.01). Imagine that 100,000 people are tested. Of the 100 people who have the disease, 99 will be diagnosed as having it. Of the 99,900 people who do not have the disease, 999 will be diagnosed as having it. Hence, only 99 1098 ≈ 9% of the people who test positive for the disease actually have it. 17.16.17.2

Pooling of blood samples

A large population of persons is to be screened for the presence of some condition using a blood test which registers positive if a sample contains blood from a person having the condition and registers negative otherwise. To implement the screening, the blood specimens of groups of k persons are pooled, and the pooled blood samples are tested. If a pooled sample registers negative, the group is cleared; otherwise each person in the group is tested. If the probability of a random person having a positive test is p, what group size k ∗ minimizes the expected number of tests per person? If f (k) is the

c 2000 by Chapman & Hall/CRC 

expected number of tests per person, then ( 1 f (k) = 1 − (1 − p)k + 1/k

if k = 1 if k ≥ 2

(17.122)

If {{x}} = x − 'x( denotes that fractional part of x, then  1/3 (a) If p > 1 − 13 ≈ 0.3066 then k ∗ = 1.  1/3 then either (b) If p < 1 − 13 O −1/2 P ∗ • k =1+ p ; or O −1/2 P ∗ • k =2+ p . O −1/2 P >> ?? O P p −1/2 (c) If p < O −1/2 P  −1/2  , then k ∗ = 1 + p−1/2 . 2 p + p (d) It is never the case that k ∗ = 2. Some values of k ∗ : p k∗

.0001 101

.0005 45

.001 32

.005 15

.01 11

.05 5

.1 4

.2 3

.3 1

.4 1

.5 1

See S. M. Samuels, The exact solution to the two-stage group-testing problem, Technometrics, 1978, 20, pages 497–500. 17.16.18

Leading digit distribution

17.16.18.1

Ratio of uniform numbers

If X and Y are chosen uniformly from the interval (0, 1], what is the probability that the ratio Y /X starts with a 1? What is the probability that the ratio Y /X starts with a 9? (a) If Y /X has a leading digit of 1, then Y /X must lie in one of the intervals {. . . , [0.1, 0.2), [1, 2), [10, 20), . . . }. This corresponds to Y /X lying in one of several triangles with height 1 and bases on either the right or top edges of the square. The triangles are defined by the lines (Y = 0.1X, Y = 0.2X, and X = 1), (Y = X, Y = X, and Y = 1), (Y = 10X, Y = 20Xl, and Y = 1), . . . . The bases along the right edge have lengths 0.1 + 0.01 + · · · = 1/9. The bases along the top edge have lengths 0.5 + 0.05 + · · · = 5/9. For a total base length of 6/9 and a height of 1, the total area is 1/3. The total area of the square is 1, hence the probability that Y /X starts with a 1 is 1/3. (b) In this case, similar triangles can be drawn; on the right edge the total length is 1/9. On the top edge the length is 1/90 + 1/900 + · · · = 1/81. Total base length is 10/81 and the total area (and hence probability of a leading 9) is 5/81 ≈ 0.061728.

c 2000 by Chapman & Hall/CRC 

17.16.18.2

Benford’s law

Let Pd be the probability that the leading digit of a set of numbers is d. Assuming that the probability distribution of leading digits is scale-invariant (common, for example, for data that are dimensional), then Pd is approxi. Hence, mately given by Pd ≈ log10 d+1 d P1 0.301 17.16.19

P2 0.176

P3 0.125

P4 0.097

P5 0.079

P6 0.067

P7 0.058

P8 0.051

P9 0.048

Lotteries

Consider a collection of n numbers. A player chooses a subset T , called a ticket, of these numbers of size k. The house chooses a subset S of these numbers of size p (these are the winning numbers). The payoff depends on the size of the intersection between S and T . If the intersection is i elements, then the payoff is wi . If n, p, and k are fixed, then   1. There are np possible tickets. 2. The number of ways that i of the player’s k choices are among the   selected p values is ki n−k p−i . 3. The probability that i of the player’s k choices are among the selected (k)(n−k) p values is i np−i . (p) kn−k p  i p−i 4. The expected return is wi n . i=0

p

5. Each i-subset has the same probability

(pi) of appearing in the p-subset (ni)

chosen. It is possible to obtain a large number of tickets and obtain few matches. If n = 49 and p = 6 then there are 13,983,816 possible tickets.    1. There are 60 43 = 6, 096, 454 tickets that have no matches with the 6 winning numbers.    643 2. There are 60 43 = 11, 872, 042 tickets that have, at most, 6 + 1 5 only a 1-match with the winning numbers.    643 643  3. There are 60 43 + + 6 1 5 2 4 = 13, 723, 192 tickets that have, at most, only a 2-match with the winning numbers. A lottery wheel is a system of buying multiple tickets to guarantee a minimum match. For example, if n = 49 and p = 6 then it is possible to obtain 174 tickets and guarantee at least a 3-match. See C. J. Colbourn and J. H. Dinitz, CRC Handbook of Combinatorial Designs, CRC Press, Boca Raton, FL, 1996, pages 578–581. c 2000 by Chapman & Hall/CRC 

17.16.20

Match box problem

A certain mathematician carries one match box in his right pocket and one in his left pocket. When he wants a match, he selects a pocket at random. Each box initially contains N matches. The probability ur that a selected box is empty on the rth trial is   N − r −2N +r ur = 2 (17.123) N 17.16.21

Maximum entropy distributions

Define Jthe entropy of a probability density function to be J = − f (x) log[f (x)] dx. (a) Among all probability density functions restricted to the interval [−1, 1], the one of minimum entropy ( has the form f (x) = C; that is, it is the 1 for −1 ≤ x ≤ 1 uniform distribution: f (x) = 2 0 otherwise (b) Among all probability density functions restricted to the interval [−1, 1] and having a given mean of µ, the( one of minimum entropy has the λ eλx for −1 ≤ x ≤ 1 form f (x) = Ceλx ; that it f (x) = 2 sinh λ where 0 otherwise λ µ = λ−tanh λ tanh λ . (c) Among all probability density functions restricted to the interval [−1, 1] that have a mean of zero and a given variance, the one of minimum 2 entropy has the form f (x) = Ceλ1 x+λ2 x See F. E. Udwadia, Some results on maximum entropy distributions for parameters known to lie in certain intervals, SIAM Review, 31, Number 1, March 1989, SIAM, Philadelphia, pages 103–109. 17.16.22

Monte Hall problem

Consider a game in which there are three doors: one door has a prize, two doors have no prize. A player selects one of the three doors. Suppose someone opens one of the two unselected doors that contains no prize. The player is then asked if he would like to exchange the originally selected door for the remaining unopened door. To increase the chance of winning, the player should switch doors: without switching the probability of winning is 1/3; by switching the probability of winning is 2/3. 17.16.23

Multi-armed bandit problem

In the multi-armed bandit problem, a gambler must decide which arm of K non-identical slot machines to play in a sequence of trials so as to maximize his reward. This problem has received much attention because of the simple model it provides of the trade-off between exploration (trying out each arm c 2000 by Chapman & Hall/CRC 

to find the best one) and exploitation (playing the arm believed to give the best payoff). 17.16.24

Parking problem

Let E(x) denote the expected number of cars of length 1 which can be parked on a block of length x if cars park randomly (and with a uniform distribution in the available space). From the integral equation  x−1 2 E(x) = 1 + E(t) dt (17.124) x−1 0 consider the Laplace transform    t  ∞  1 − e−u e−s ∞ du dt. e−sx E(x) dx = exp −2 s s u 0 s

(17.125)

This implies E(x) ∼ cx as x → ∞ (where c ≈ 0.7476). 17.16.25

Passage problems

Given a random function X(t), a passage (time) at a given level a is when a graph of X(t) crosses the horizontal line X = A from below. Define the derivative of X to be V (t) = dX(t) dt , and let f (x, v | t) be the probability density of x and v at time t (define the probability density for X to be f (x)). The probability that a passage (time) lies in the time interval dt about the time t is p(a | t) dt and  ∞ p(a | t) = f (a, v | t) v dv (17.126) 0

(a) For normal functions

$  + 1 1 (a − x)2 ∂ 2 Kx (t1 , t2 ) $$ p(a | t) = exp − 2π 2σx2 Kx (t, t) ∂t1 ∂t2 $t1 =t2 =t (17.127)

where Kx is the correlation function. (b) For normal stationary functions p(a | t) = p(a) =

  1 σv (a − x)2 exp − 2π σx 2σx2

(17.128)

(c) The number of passages of a stationary function during a time interval T is N a = T p(a). (d) The average duration τa of a passage of a stationary function is J∞ f (x) dx τa = a (17.129) p(a)

c 2000 by Chapman & Hall/CRC 

For normal stationary functions:     a−x σx (a − x)2 1 − Φ τa = π exp − σv 2σx2 σx

(17.130)

(e) Finding the average number of minima of a random differentiable function can be reduced to passage problems since a maxima is achieved when the first derivative passes through zero from above. For a small average number of passages during a time interval T , the approximate probability Q for non-occurrence of any run during this interval is Q = e−N a (i.e., the number of passages in the given interval can be approximated as a Poisson distribution). 17.16.26

Proofreading mistakes

Suppose that proofreader A finds a mistakes, and proofreader B finds b mistakes. Assume that there are c overlaps, mistakes that both A and B found. ab and the approximate number The approximate number of total mistakes is c (a − c)(b − c) of mistakes missed by both A and B is . c 17.16.27

Raisin cookie problem

A baker creates enough cookie dough for 1000 raisin cookies. The number of raisins to be added to the dough, R, is to be determined. 1. In order to be 99% certain that the first cookie will have at least one  999 R raisin, then 1 − 1000 ≥ 0.99, or R ≥ 4603. 2. In order to be 99% certain that every cookie will have at least one raisin, then P (C, R) ≥ 0.99, where C is the number of cookies and C   P (C, R) = C −R i=0 Ci (−1)i (C − i)R . Or, R ≥ 11508. 17.16.28

Random sequences

17.16.28.1

Long runs

For a biased coin with a probability of heads p, let Rn denote that length of the longest run of heads in the first n tosses of the coin. It has been shown: % & Rn Prob lim =1 =1 (17.131) n→∞ log1/p (n) See R. Arratia and M. S. Waterman, An Erd¨ os–R´enyi law with shifts, Advances in Mathematics, 55, 1985, pages 13–23.

c 2000 by Chapman & Hall/CRC 

17.16.28.2

Waiting times: two types of characters

In a Bernoulli process with a probability of success p and a probability of failure q = 1 − p, the probability that a run of α consecutive successes occurs before a run of β consecutive failures is   pα−1 1 − q β probability = α−1 (17.132) p − q β−1 The expected waiting time until either run occurs is   (1 − pα ) 1 − q β expected waiting time = α p q + pq β − pα q β

(17.133)

The waiting times for strings of two characters and for strings of many types of characters can be very different. 17.16.28.3

Waiting times: many types of characters

Assume there are N different characters, each equally likely to occur. Let S = (a1 , a2 , · · · , an ) be a sequence of characters, some of which may be equal. If Sm = (a1 , a2 , . . . , am ), for 1 ≤ m ≤ n, appears both at the beginning and at the end of S, then Sm is an overlapping subsequence of S. The mean waiting time until S appears depends on the lengths of the overlapping subsequences. If the lengths of the overlapping subsequences are {k1 , k2 , . . . , kr } then, expected waiting time =

r 

N kr

(17.134)

i=0

Example 17.83 :

Choosing letters randomly from the alphabet, it will take on the

average, (a) 263 letters to get APE (b) 263 + 26 letters to get DAD (c) 2611 + 264 + 26 letters to get ABRACADABRA.

Example 17.84 : Flipping a fair coin, it will take on the average, (a) 24 flips to get HHTT (b) 24 + 21 flips to get HTHH (c) 24 + 22 flips to get TH TH (d) 24 + 23 + 22 + 21 flips to get HHHH

See G. Blom, On the mean number of random digits until a given sequence occurs, J. Appl. Prob., 19, 1982, pages 136–143 and G. Blom and D. Thorurn, How many random digits are required until given sequences are obtained?, J. Appl. Prob., 19, 1982, pages 518–531.

c 2000 by Chapman & Hall/CRC 

[AA] = 1 0 1 0 = 10 A =THTH A =THTH [BB] = 1 0 0 1 = 9 B =HTHH B =HTHH

[AB] = 0 1 0 1 = 5 A=THTH B =HTHH [BA] = 0 0 0 0 = 0 B =HTHH A=THTH

Figure 17.2: Computation of leading numbers

17.16.28.4

First random sequence

Let a fair coin be flipped until a given sequence appears. Let two competitors, P and Q, choose sequences of heads and tails; the winner is the one whose sequence appears first. Conway’s leading number algorithm indicates who is likely to win: Given two sequences A = (a1 , . . . , am ) and B = (b1 , . . . , bm ), define the leading number of A over B as a binary integer 3n 3n−1 · · · 31 via  1 3i =

if 1 ≤ i ≤ min(m, n) and the two sequence (am−i+1 , . . . , am ) and (b1 , . . . , bi ) are identical  0 otherwise (17.135)

Then, let [AB] denote the leading number of A over B. The odds for B to precede A in a symmetric Bernoulli process are ([AA] − [AB]) : ([BB] − [BA])

(17.136)

Example 17.85 : Consider the two different 4-tuples A =THTH and B =HTHH. Figure 17.2 contains the needed leading number computations so that ([AA] − [AB]) : ([BB] − [BA]) (in equation 17.136) becomes (10 − 5) : (9 − 0) or 5 : 9. Hence the odds that A occurs before B is 9/14 (see section 3.3.5). Note that sequence A = THTH has a waiting time (see section 17.16.28.2) of 20 while the sequence B = HTHH has a waiting time of 18. Hence, in this case, an event that is less frequent in the long run is likely to happen before a more frequent event.

Table 17.3 has the probability of Q winning in a double game. Table 17.4 has the probability of Q winning in a triplet game. When playing a triplet game, note that whatever triplet the first player chooses, the second player can choose a better one.

c 2000 by Chapman & Hall/CRC 

P \Q HH HT TH TT

HH — 1/2 3/4 1/2

HT 1/2 — 1/2 1/4

TH 1/4 1/2 — 1/2

TT 1/2 3/4 1/2 —

Table 17.3: Probability of P winning in a double game P \ Q HHH HHT HTH HTT THH THT TTH TTT 2/5 2/5 1/8 5/12 3/10 1/2 1/2 HHH — 2/3 1/4 5/8 1/2 7/10 1/2 2/3 HHT — 1/3 1/2 1/2 3/8 7/12 3/5 1/2 HTH — 1/3 1/2 1/2 3/4 7/8 3/5 1/2 HTT — 3/4 1/2 1/2 1/3 3/5 7/8 1/2 THH — 3/8 1/2 1/2 1/2 3/5 7/12 1/3 THT — 1/2 5/8 1/4 2/3 2/3 7/10 1/2 TTH — 3/10 5/12 1/8 2/5 2/5 1/2 1/2 TTT — Table 17.4: Probability of P winning in a triplet game

17.16.29

Random walks

17.16.29.1

Random walk on a grid

Consider a random walk of uniform step lengths on a one-, two-, or threedimensional grid. Each step goes in any of the 2, 4, or 8 directions randomly. In one and two dimensions, the random walk will, with probability 1, return to the starting location. In three dimensions the probability of a return to the origin is approximately 0.34053. The probability of return to the origin in d dimensions (for d = 4, 5, . . . ) is J ∞   d −t 1 p(d) = 1 − u(d) where u(d) = 0 I0 dt e dt. The approximate probabilities of returning to the origin are as follows: dimension probability 17.16.29.2

4 0.20

5 0.136

6 0.105

7 0.0858

8 0.0729

Random walk in two dimensions (Rayleigh problem)

Consider a two-dimensional random walk in which each step, xi , is of length a is taken in a random direction. The position after  n steps is (X1 , X2 ) = and n X . The radial distance after n steps is R = X12 + X22 . If p(x) is the i n i=1 probability density for a single step and pn (x) is the probability density after

c 2000 by Chapman & Hall/CRC 

n steps, then p(x) =

1 δ(|x| − a) 2πa  ∞



−∞

−∞ ∞

φ(ξ) = 1 2π  < r] = r



(17.137) n

ρ J0 (ρ |x|) [J0 (ρa)] dρ

pn (x) = Prob [RN

p(x) eiξ·x dx = J0 (a |ξ|)

0 ∞

n

J1 (ru) [J0 (au)] du

0

Note that E [X1 ] = E [X2 ] = 0 E [X1 X2 ] = 0   na2   E X12 = E X22 = 2   E Rn2 = na2

(17.138)

If, instead, the ith step has length ai , then    ∞ Prob [RN < r] = r J1 (ru) J0 (a1 u) · J0 (a2 u) · · · J0 (an u) du (17.139) 0

17.16.29.3

Random walk in three dimensions

Consider a three-dimensional random walk in which each step, Xi , is of length a and is taken in a random direction. If p(x) is the probability density for a single step and pn (x) is the probability density after n steps, then p(x) =

1 δ(|x| − a) 2 4πa   ∞





1 px (x)eiξ·x dx = sin(aξ) a |ξ| −∞ −∞ −∞ n   ∞ sin(|x| ρ) 1 2 sin(aρ) pn (x) = dρ ρ 2π 2 0 aρ |x| ρ φ(ξ) =

Note that

  p2 (x) =

c 2000 by Chapman & Hall/CRC 

1 8πa2 |x|  0

for |x| ≤ 2a for |x| > 2a

(17.140)

(17.141)

 1    3   8πa 3a − |x| p3 (x) =    16πa3 |x|   0 17.16.29.4

for 0 ≤ |x| ≤ a for a < |x| ≤ 3a

(17.142)

for 3a < |x|

Self-avoiding walks

Consider a random walk of n unit length steps on a d-dimensional grid. A walk in which no grid point is visited twice is called a self-avoiding walk ; the number of such walks is denoted c(n). Known values include: c(0) = 1, c(1) = 2d, and c(2) = 2d(d − 1). For each d there is a non-zero constant 1/n µd (the connective constant) such that c(n) ∼ (µd ) as n → ∞. Current bounds include: 2.62 < µ2 < 2.70 4.57 < µ3 < 4.75 6.74 < µ4 < 6.82 8.82 < µ5 < 8.87 10.87 < µ6 < 10.89

(17.143)

See J. Noonan, New upper bounds for the connective constants of self-avoiding walks, J. Statist. Physics, 91, 1998, pages 871–888. 17.16.30

Relatively prime integers

Given two integers chosen at random, the probability that they are relatively −1 prime is [ζ(2)] = 6/π 2 ≈ 0.608. Given three integers chosen at random, the −1 probability they have no common factor other than 1 is [ζ(3)] ≈ 1.202−1 ≈ 0.832. 17.16.31

Roots of a random polynomial

Consider the polynomial f (x; a) = a0 +a1 x+· · ·+an−1 xn−1 where the {ai } are chosen randomly. To determine the expected number of real roots, suppose: (a) The {ai } are chosen uniformly distributed on the interval (−1, 1); or (b) The {ai } are chosen equal to +1 and −1 with equal probability; or (c) The {ai } are chosen to be independent standard normal coefficients then the expected number of real roots is  + 4 1 1 (n + 1)2 t2n E [number of real roots] = − dt 2 2 π 0 (1 − t ) (1 − t2n+2 )2   2 2 1 as n → ∞ ∼ log n + C + +O π nπ n2 (17.144) c 2000 by Chapman & Hall/CRC 

where C = 0.6257358072 . . . . Consider a random polynomial of degree n with coefficients that are independent and identically distributed normal random variables. Define m = 0 to be the mean divided by the standard deviation (m = µ/σ). Then the expected number of real roots as n → ∞ is   1 C +1 2+γ 2 1 log n + − − log |m| + O (17.145) π 2 π π n where γ is Euler’s constant. See M. Kac, On the average number of real roots of a random algebraic equation, Bulletin of the AMS, 49, pages 314–320, 1943, and A. T. BharuchaReid and M. Sambandham, Random Polynomials, Academic Press, New York, Chapter 4, pages 49–102, 1986. 17.16.32

Roots of a random quadratic

If A, B, and C are independent random variables uniformly distributed on (0, 1) then the probability that Ax2 + Bx + C = 0 has real roots is (5 + 3 ln 4)/36. 17.16.33

Simpson paradox

Consider the comparison of a new treatment with an old treatment, and suppose that the following results are obtained: All patients New treatment Old treatment

improved 20 24

not improved 20 16

percent improved 50% 60%

On the basis of these data, one would be inclined to say that the new treatment is not better than the old treatment. Now consider a disaggregation of these data into the following two sub-groups: Young patients New treatment Old treatment

improved 12 3

not improved 18 7

percent improved 40% 30%

Old patients New treatment Old treatment

improved 8 21

not improved 2 9

percent improved 80% 70%

From these data, one would be inclined to say that the new treatment is better than the old treatment for both young and old patients. This is an example of Simpson’s paradox, and it is not a small sample effect. Simpson’s paradox reflects the fact that it is possible for all three of the following inequalities to hold simultaneously:

c 2000 by Chapman & Hall/CRC 

Prob [I | A, B] > Prob [I | A, B  ] Prob [I | A , B] > Prob [I | A , B  ] Prob [I | B] < Prob [I | B  ]

(17.146)

However, if Prob [A | B] = Prob [A | B  ] then it is not possible for the 3 equations in (17.146) to hold. Practically, this means that the relative proportions (new treatment versus old treatment) should be the same for both new and old patients. 17.16.34

Secretary call problem

Suppose a secretary must be hired and n applicants are in line. Each applicant will be interviewed one at a time. Following each interview an immediate decision is made to hire or reject the candidate. Once a candidate has left, he cannot be called back. The following strategy may be used to find the best secretary. For sufficiently large n, the interviewer should consider n/e candidates, and then select the next candidate that is better than any seen previously. Using this strategy, the probability of selecting the best candidate is approximately 1/e. See J. S. Rose, The secretary problem with a call option, Operations Research Letters, Vol. 3, No. 5, December 1984. 17.16.35

Waiting for a bus

Suppose buses pass a certain corner with an average time between them of 20 minutes. What is the average time that one would expect to wait for a bus? If the buses are exactly 20 minutes apart, then the average waiting time is 10 minutes. If the buses’ arrival pattern is a Poisson distribution, then the average waiting time is 20 minutes. If the buses arrival pattern is a hyperexponential distribution, then the average waiting time may exceed 20 minutes. 17.17

ELECTRONIC RESOURCES

Thanks are extended to John C. Pezzullo for supplying the source for much of this section. 17.17.1

Statlib

Statlib, located at http://lib.stat.cmu.edu is a system for distributing statistical software, datasets, and information by electronic mail, FTP and WWW. The main web page includes pointers to: (a) The apstat collection; a nearly complete set of algorithms published in Applied Statistics. c 2000 by Chapman & Hall/CRC 

(b) The CMLIB collection of non-proprietary Fortran subprograms solving a variety of mathematical and statistical problems (originally produced by NIST). (c) The DASL library of datafiles and stories that illustrate the use of basic statistics methods. (d) A collection of interesting datasets, from classics like the Stanford heart transplant data, to the complete data from several textbooks. For example, over 6Mb of data and descriptions are available from Case Studies in Biometry, by N. Lange, L. Ryan, L. Billard, and D. Brillinger, John Wiley & Sons, New York, 1994. (e) A collection of designs, programs, and algorithms for creating designs for statistical experiments. (f) Algorithms from Applied Statistics Algorithms by P. Griffiths and I. D. Hill, Ellis Horwood, Chichester, 1985. (g) A collection of software related to articles published in the Journal of the American Statistical Association. (h) The jcgs archive of contributed datasets and software and abstracts from the Journal of Computational Graphics and Statistics. (i) The jqt collection of algorithms from articles published in the Journal of Quality Technology. (j) P-Stat functions and related software. (k) Software and macros for the Genstat language. (l) Macros, software, and algorithms for the GLIM statistical package. (m) Software and extensions for the S (Splus) language. Over 130 separate packages including many novel statistical ideas. (n) The Sapaclisp collection of Lisp functions that can be used for computations described in Spectral Analysis for Physical Applications: Multitaper and Conventional Univariate Techniques, by D. B. Percival and A. T. Walden, Cambridge University Press, Cambridge, England, 1993. (o) The MacAnova statistical system (for Mac, Windows, and Unix). (p) Macros for Minitab. (q) Materials related to the Stata statistical package. (r) Lisp-Stat, which is an extensible environment for statistical computing and dynamic graphics based on Lisp. 17.17.2

Uniform resource locators

The following is a list of interesting Uniform Resource Locators (URL): (a) http://www.stat.berkeley.edu/~stark/SticiGui/Text/ gloss.htm A very useful glossary.

c 2000 by Chapman & Hall/CRC 

(b) http://www.yahoo.com/Science/Mathematics/ A very large list of useful sites relating to mathematics. It is perhaps the best place to start researching an arbitrary mathematical question not covered elsewhere in this list. (c) http://daisy.uwaterloo.ca/~alopez-o/math-faq/ math-faq.html The FAQ (frequently asked questions) listing from the newsgroup sci.math. (d) http://e-math.ams.org/ The American Mathematical Society home page, with information about AMS-TEX, the Combined Membership List of the AMS, Math Reviews subject classifications, preprints, etc. (e) http://www.siam.org/ The Society for Industrial and Applied Mathematics. (f) http://gams.cam.nist.gov/ The Guide to Available Mathematical Software (GAMS). (g) http://www.york.ac.uk/depts/maths/histstat/welcome.htm The early history of statistics: a collection of seminal papers by Bayes, Pascal, Laplace, Legendre, and others. (h) http://www-sci.lib.uci.edu/HSG/RefCalculators.html#STAT Martindale’s reference desk—calculators on–line—statistics: the largest compendia of calculating web pages. (i) http://www.fedstats.gov FedStats: master page maintained by the Federal Interagency Council on Statistical Policy to provide easy access to the full range of statistics and information produced by more than 70 agencies in the United States Government. (j) http://members.aol.com/johnp71/javastat.html A collection of web pages with links to many statistical sites on the web, maintained by John C. Pezzullo. 17.17.3

Interactive demonstrations and tutorials

(a) Statistics tutorials http://204.215.60.174/tutindex.html These briefly explain the use and interpretation of standard statistical analysis techniques, using the WINKS program from TexaSoft. Tutorials include: • • • •

descriptive statistics program Friedman’s test independent group t-test Kruskal–Wallis test

c 2000 by Chapman & Hall/CRC 

• • • •

Mann–Whitney test one-way ANOVA paired t-test Pearson’s correlation coefficient

• simple linear regression

• single sample t-test

(b) Simulation and demonstrations http://www.ruf.rice.edu/~lane/stat_sim/index.html From the creator of the HyperStat Online statistics book. More than a dozen applets; including: • 2 × 2 contingency tables • histograms • mean and median • a small effect size can make a • normal approximation to large difference • bin widths binomial distribution • chi–square test of deviations • regression by eye from expected frequencies • reliability and regression • components of r analysis • confidence interval for a • repeated measures proportion • restriction of range • sampling distribution • confidence intervals simulation • cross validation (c) Virtual fly lab http://vflylab.calstatela.edu/edesktop/VirtApps/VflyLab/ IntroVflyLab.html Simulate classic drosophila genetics experiments: Specify characteristics of a hypothetical male and female fly; simulate the offspring of a mating; select a pair of offspring to breed; simulate the offspring of their mating; and compare the observed frequencies with those predicted from Mendelian genetics. http://www.users.on.net/zhcchz/java/quincunx/quincunx.1.html Illustration of the central limit theorem in Java. (d) http://www.math.csusb.edu/faculty/stanton/m262/ probstat.html Miscellaneous Java demos. (e) http://www.ms.uky.edu/~mai/java/AppletIndex.html Five demonstrations of how samples of increasing size approach a theoretical distribution: • Empirical • Kaplan–Meier • Nelson–Aalen

• interval censored • doubly-censored data

(f) http://www.ms.uky.edu/~mai/java/AppletIndex.html Three demonstrations of famous random processes: • 1-dimensional and 2-dimensional Brownian motion • Buffon needle experiment • Galton’s ball–drop quincunx (g) http://www.ms.uky.edu/~lancastr/java/javapage.html Three bootstrap demonstrations all using the Exp(1) distribution: c 2000 by Chapman & Hall/CRC 

• central limit theorem • parametric bootstrap of sample mean • nonparametric bootstrap of sample mean (h) http://www-stat.stanford.edu/~naras/jsm/NormalDensity/ NormalDensity.html Experiments with the normal distribution. (i) http://www.stat.uiuc.edu/~stat100/java/guess/ PPApplet.html Linear regression and correlation demo in Java— click a bunch of points onto the screen; as each is entered the computer immediately computes and displays an adjusted regression line (with equation and correlation coefficient). (j) http://stat-www.berkeley.edu/users/stark/Java/ Correlation.html Similar to above, but lets you simulate points from distributions with known correlation coefficient. (k) http://www.ruf.rice.edu/~lane/stat_sim/sampling_dist/ sampling_demo.html Simulation of various sampling distributions in Java. (l) http://www.stat.uiuc.edu/~stat100/java/box/BoxApplet.html Simulation in Java of drawing objects (with replacement) from a “box” whose contents you can define. (m) http://www.stat.sc.edu/~west/javahtml/epid.html Simulate the course of an epidemic in Java. (n) http://huizen.dds.nl/~berrie/ A collection of QuickTime movies illustrating various statistical concepts. (o) http://members.aol.com/Trane64/java/CardPowersim.html Monte–Carlo p versus sample size simulation for survey questionnaire results, with graphical output (in Java). 17.17.4

Online textbooks, reference manuals, and journals

(a) Glossary of statistical terms used in evidence-based medicine http://www.musc.edu/muscid/glossary.html Can be printed out into a handy 12-page mini-handbook of statistical concepts and terminology. (b) HyperStat statistics textbook http://ruf.rice.edu/~lane/hyperstat/contents.html A well-designed and well-constructed “hyper-text-book.” (c) Statistics at Square One http://www.bmj.com/collections/statsbk/index.shtml An excellent online textbook. c 2000 by Chapman & Hall/CRC 

(d) Electronic Statistics Textbook http://www.statsoft.com/textbook/stathome.html (by StatSoft) Very extensive and well-organized (can also be downloaded for quicker access from your hard drive). (e) Instat Guide http://www.graphpad.com/instatman Instat Guide to Choosing and Interpreting Statistical Tests, from Graphpad. (f) Glossary of over 30 statistical terms http://www.animatedsoftware.com/statglos/statglos.htm Explanations vary from short definitions to extensive explanations, with graphics and hyperlinks. (g) Introduction to Probability http://www.geom.umn.edu/docs/snell/chance/teaching_aids/ probability_book/pdf.html A complete book by Grinstead and Snell in Adobe Acrobat and Postscript format. It can be downloaded all at once, or chapter by chapter. (h) Graphpad web site of statistical resources http://www.graphpad.com/www/ Short articles, book chapters, bibliographies, and (commercial) software. Well-written, down-to-earth, and helpful. (i) InterStat (Statistics on the internet) http://interstat.stat.vt.edu/intersta.htm/ An online journal where one can publish or read about any aspect of statistical research or innovative method. Abstracts of articles can be viewed; from there you can read or download the article (or any comments written about it) in pdf and ps formats. (j) Journal of Statistics Education http://www.stat.ncsu.edu/info/jse/ A refereed electronic journal on postsecondary teaching of statistics. (k) http://www.psychstat.smsu.edu/sbk00.htm Introductory Statistics: Concepts, Models, and Applications by David W. Stockburger. In-depth coverage, with extensive use of web technology (animated graphics, interactive calculating pages). (l) http://www.psychstat.smsu.edu/multibook/mlt00.htm Multivariate Statistics: Concepts, Models, and Applications, by David W. Stockburger In-depth coverage, with extensive use of web technology (animated graphics, interactive calculating pages). (m) http://trochim.human.cornell.edu/kb/kbhome.htm The Knowledge Base— An Online Research Methods Textbook, by William M. K. Trochim. An online textbook for an introductory course in research methods. c 2000 by Chapman & Hall/CRC 

(n) http://www.cern.ch/Physics/DataAnalysis/BriefBook/ Data Analysis BriefBook : A condensed handbook, or an extended glossary, written in encyclopedic format, covering subjects in statistics, computing, analysis, and related fields; meant to be both introduction and reference for data analysts, scientists and engineers (from CERN). (o) http://www.cne.gmu.edu/modules/dau/stat/ Statistics refresher tutorial. (p) http://www.wrightslaw.com/advoc/articles/ tests_measurements.html Statistics of performance measurement, designed for parents of children with special educational needs, but also worthwhile for a good introductory presentation of basic statistical concepts. (q) http://nimitz.mcs.kent.edu/~blewis/stat/scon.html Guide to Basic Laboratory Statistics (in Java). (r) http://duke.usask.ca/~rbaker/stats.html Basic Principles of Statistical Analysis. 17.17.5

Free statistical software packages

A selection of free software packages. (a) ViSta http://forrest.psych.unc.edu/research/ViSta.html A VIsual STAtistics program for Windows, Mac, and Unix. Features a structured desktop. (b) WebStat http://www.stat.sc.edu/~west/webstat/ Java-based statistical computing environment for the web. Requires a browser, but can be run offline. (c) MacANOVA http://www.stat.umn.edu/~gary/macanova/macanova.home.html Available for Windows and Mac. This is a complete programming language. (d) Scilab http://www-rocq.inria.fr/scilab/ Available for Windows, Mac, and Unix. This is a programming language with MatLab-like syntax, hundreds of built-in functions and libraries, 3D graphics, and symbolic capabilities through a Maple interface. (e) SISA http://home.clara.net/sisa/pasprog.htm Simple Interactive Statistical Analysis for DOS from Daan Uitenbroek. Collection of modules for many types of hypothesis testing, sample size calculations, and survey design. Includes many analyses not usually found elsewhere. c 2000 by Chapman & Hall/CRC 

(f) Anderson statistical archives http://odin.mdacc.tmc.edu/anonftp/page_2.html A large collection of statistical programs for DOS and Mac with Fortran and C source, from the Biomathematics Department of the M. D. Anderson Cancer Center. (g) STPLAN http://odin.mdacc.tmc.edu/anonftp/page_2.html Performs power, sample size, and related calculations needed to plan studies. Covers a wide variety of situations, including studies whose outcomes involve the binomial, poisson, normal, and log-normal distributions, or are survival times or correlation coefficients. For DOS and Mac with Fortran and C source. (h) EpiInfo http://www.cdc.gov/epo/epi/epiinfo.htm A set of programs for word processing, data management, and epidemiologic analysis, designed for public health professionals. Consists of Epi Info (forms design, data entry, data management), Epi Map (generated geographical, map-based output), SSS1 (Box–Jenkins time series analysis, MMWR graphs, trend analysis, and 2-source comparisons). (i) Rweb http://www.math.montana.edu/Rweb/ A Web based interface to the ”R” statistical programming language (similar to S or S-plus). (j) ARIMA ftp://ftp.census.gov/pub/ts/x12a/ A seasonal adjustment program for PC and Unix, developed by the Census Bureau. (k) G*Power http://www.psychologie.uni-trier.de:8000/projects/gpower.html A general Power Analysis program for DOS and Macintosh. Performs high-precision analysis for t-tests, F-tests, chi–square tests. Computes power, sample sizes, alpha, beta, and alpha/beta ratios. Contains webbased tutorial and reference manual. (l) JDB http://www.isi.edu/~johnh/SOFTWARE/JDB/index.html Relational database and elementary statistics for unix. Useful for manipulating experimental data (joining files, cleaning data, reformatting for input into other programs). Computes basic statistics (mean, std. dev., confidence intervals, quartiles, n-tiles, percentiles, histograms, correlations, z-scores, t-scores). (m) EasyMA http://www.spc.univ-lyon1.fr/~mcu/easyma/

c 2000 by Chapman & Hall/CRC 

DOS program for the meta-analysis of clinical trials results. Developed to help physicians and medical researchers to synthesize evidence in clinical or therapeutic research. (n) Meta-analysis 5.3 http://www.yorku.ca/faculty/academic/schwarze/meta_e.htm DOS software for meta-analysis, perhaps the most frequently used metaanalysis software in the world. Can select the analysis of exact p values or effect sizes (d or r, with a cluster size option). Plots a stem-and-leaf display of correlation coefficients. A utility menu allows various transformations and preliminary computations that are typically required before meta-analysis can be performed. (o) First Bayes http://www.maths.nott.ac.uk/personal/aoh/1b.html A windows application for elementary Bayesian Statistics. Performs most standard, elementary Bayesian analyses, including: plotting and summarizing distributions, defining and examining arbitrary mixtures of distributions, analysis of two kinds of linear model (one or more normal samples with common but unknown variance, and simple linear regression), examination of marginal distributions for arbitrary linear combinations of the location parameters, and the generation of predictive distributions. (p) MANET http://www1.math.uni-augsburg.de/Manet/ Missings Are Now Equally Treated: Mac software for interactive graphics tools for data sets with missing values. Generates missing values chart, histograms and barcharts, boxplots and dotplots, scatterplots, mosaic plots, polygon plots, highlighted boxplots, interactive trellis displays, traces, context-sensitive interrogation, cues, redframing, selection sequences. (q) WinSAAM http://www.winsaam.com/#WinSAAM Windows version of SAAM (System Analysis and Modeling Software). Lets you create mathematical models, design and simulate experiments, and analyze data. Models can contain differential equations. Graphic and tabular output is provided. (r) Boomer http://www.cpb.uokhsc.edu/common/anonymous/MFB/toc.html Non-linear regression program for analysis of pharmacokinetic and pharmacodynamic data. Includes normal fitting, Bayesian estimation, or simulation-only, with integrated or differential equation models. Allows selection of weighting schemes and methods for numerical integration. Available for Macintosh and PC. Online manual, tutorial, and sample data sets. c 2000 by Chapman & Hall/CRC 

(s) Serpik Graph http://www.chat.ru/~mserpik/index.htm A very sophisticated function graphing program. 17.17.6

Demonstration statistical software packages

The following are a selection of “demonstration versions” or “student versions” of commercial packages. They can be freely downloaded, but are usually restricted or limited in some way. (a) AssiStat http://users.aol.com/micrometr/assistat.htm A Windows package, designed to complement a typical statistical package. It covers topics not normally in a primary analysis package such as correction of correlations for restriction in range or less-than-perfect reliability. (b) CART http://www.salford-systems.com/ Salford Systems flagship decision-tree software, combines an easy-to-use GUI with advanced features for data mining, data pre-processing, and predictive modeling. (c) InStat http://www.graphpad.com/instat3/instatdemo.htm INstant STATistics: A package from GraphPad Software for Windows and Mac. Demo version disables printing, saving, and exporting capabilities. (d) MiniTab http://www.minitab.com/products/minitab/rel12/index.htm aA program for Windows. Good coverage of industrial quality control analyses. (e) Prism http://www.graphpad.com/prism/pdemo.htm A package from from GraphPad Software for Windows and Mac. Performs basic biostatistics, fits curves and creates publication quality scientific graphs in one complete package. (f) QuickStat http://www.camcode.com/arcus.html A program for Windows. Does not assume statistical knowledge. Gives advice on experimental design, analysis, and interpretation from the integrated statistical help system. Results presented in plain language. For the more experienced user, Arcus is an extremely useful statistical toolbox which covers biomedical statistical methods. (g) Rasch http://mesa.spc.uchicago.edu/ Measurement software deals with the various nuances of constructing optimal rating scales from a number of (usually) dichotomous measurec 2000 by Chapman & Hall/CRC 

ments, such as responses to questions in a survey or test. (h) STATISTICA demo http://www.statsoft.com/download.html#demo Has many features of the “Basic Statistics” module of STATISTICA. (i) Statlets http://www.statlets.com/ A Java statistics program. (j) SPSS demos http://www.spss.com/cool/index.htm Numerous demo packages from products acquired by SPSS (all for the PC except one): • • • • • • • • • • 17.18 17.18.1

allCLEAR 3.5 and 4.5 GOLDMineR DeltaGraph (Mac) LogXact 2.1 PeakFit 4.06 QI Analyst 3.5 Remark Office OMR 3.0 SamplePower 1.2 SigmaGel 1.0 SPSS Diamond

• • • • • • • • •

SigmaPlot 4.0 SigmaScan Pro 4.0 SigmaStat 2.0 SmartViewer StatXact 3.1 SYSTAT 7.0 TableCurve 2D 4.07 TableCurve 3D 3.01 WesVar Complex Samples

TABLES Random deviates

Consider the two data sequences Scontinuous unif and Sstd normal defined by: (a) Continuous uniform random numbers on the interval (0, 1) are generated by a linear congruential generator with parameters: x0 = 1, a = 99, b = 0, and M = 220 ; see section 17.11.1.1. The sequence begins {1, 99, 9801, 970299, 639185, 364755, . . . }. Dividing this sequence by M yields numbers on the interval (0, 1). Hence, Scontinuous unif = {.00000095, .000094, .0093, .9253, .6095, .3478, . . . } (b) Sstd normal is then obtained by using the Box–Muller method (see section 17.11.2). Specifically, if the values in Scontinuous uniform are written as {u1 , v1 , u2 , v2 , . . . } then standard normal variables are given by √ −2 ln ui cos 2πvi . Hence, Sstd normal = {5.266, 2.727, −.574, −.763, −.140, . . . }. The following random units are obtained from the third digit of the first 2,500 values in Scontinuous unif . The standard normal values are the first 500 values in Sstd normal .

c 2000 by Chapman & Hall/CRC 

2,500 Random units Line Col = (1) 1 00959 2 08149 3 49324 4 03207 5 43855

(2) 77100 50395 30579 57615 79617

(3) 36484 26971 12667 90513 46906

(4) 06101 29957 28910 02808 90009

(5) 34734 25235 42885 68168 48557

(6) 18924 85468 78563 12164 21458

(7) 28001 85629 10340 81531 31805

(8) 87674 57477 77508 90806 99078

(9) 16684 68311 96233 01306 15841

(10) 07027 33626 84084 94806 12303

6 7 8 9 10

56825 91822 96171 81089 00214

52535 62169 27730 47387 98330

70234 10152 38041 46900 43526

98035 29526 12364 96147 10401

96147 48393 85443 68169 39687

93443 07763 55515 37767 16238

08513 12402 26148 82207 43098

81271 43920 85117 74778 63975

04047 06443 49046 21357 59317

29078 44055 28379 54056 90397

11 12 13 14 15

77471 69857 91127 17793 44797

67351 73393 97088 53476 23004

83417 24342 07052 95434 91138

65376 43527 71213 42814 43018

79343 49655 52345 82423 14720

15268 46460 20899 29613 10925

37451 12109 50674 77770 53861

36632 36747 29653 99328 22495

52965 80829 04919 90929 14266

29032 80614 99244 23462 06873

16 17 18 19 20

80021 23856 56168 72348 17468

02262 96770 45893 46166 91307

28002 76740 93645 92883 08718