Introduction to Statistical Thought

Michael Lavine

November 26, 2005

ii

c 2005 by Michael Lavine

Copyright
 Contents

List of Figures vii

List of Tables xi

Preface xiii

1 Probability 1
1.1 Basic Probability . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Probability Densities . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3 Parametric Families of Distributions . . . . . . . . . . . . . . . . 12
1.3.1 The Binomial Distribution . . . . . . . . . . . . . . . . . . 13
1.3.2 The Poisson Distribution . . . . . . . . . . . . . . . . . . 16
1.3.3 The Exponential Distribution . . . . . . . . . . . . . . . . 18
1.3.4 The Normal Distribution . . . . . . . . . . . . . . . . . . 20
1.4 Centers, Spreads, Means, and Moments . . . . . . . . . . . . . . 26
1.5 Joint, Marginal and Conditional Probability . . . . . . . . . . . . 34
1.6 Association, Dependence, Independence . . . . . . . . . . . . . . 44
1.7 Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
1.7.1 Calculating Probabilities . . . . . . . . . . . . . . . . . . . 49
1.7.2 Evaluating Statistical Procedures . . . . . . . . . . . . . . 53
1.8 R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
1.9 Some Results for Large Samples . . . . . . . . . . . . . . . . . . . 66
1.10 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70

2 Modes of Inference 79
2.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
2.2 Data Description . . . . . . . . . . . . . . . . . . . . . . . . . . . 80
2.2.1 Summary Statistics . . . . . . . . . . . . . . . . . . . . . . 80
2.2.2 Displaying Distributions . . . . . . . . . . . . . . . . . . . 84
2.2.3 Exploring Relationships . . . . . . . . . . . . . . . . . . . 98
2.3 Likelihood . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
2.3.1 The Likelihood Function . . . . . . . . . . . . . . . . . . . 109
2.3.2 Likelihoods from the Central Limit Theorem . . . . . . . 118

iii
iv CONTENTS

2.3.3 Likelihoods for several parameters . . . . . . . . . . . . . 122

2.4 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 130
2.4.1 The Maximum Likelihood Estimate . . . . . . . . . . . . 131
2.4.2 Accuracy of Estimation . . . . . . . . . . . . . . . . . . . 132
2.4.3 The sampling distribution of an estimator . . . . . . . . . 135
2.5 Bayesian Inference . . . . . . . . . . . . . . . . . . . . . . . . . . 140
2.6 Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 148
2.7 Hypothesis Testing . . . . . . . . . . . . . . . . . . . . . . . . . . 154
2.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 163

3 Regression 173
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 173
3.2 Normal Linear Models . . . . . . . . . . . . . . . . . . . . . . . . 180
3.2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . 180
3.2.2 Inference for Linear Models . . . . . . . . . . . . . . . . . 191
3.3 Generalized Linear Models . . . . . . . . . . . . . . . . . . . . . . 203
3.3.1 Logistic Regression . . . . . . . . . . . . . . . . . . . . . . 203
3.3.2 Poisson Regression . . . . . . . . . . . . . . . . . . . . . . 211
3.4 Predictions from Regression . . . . . . . . . . . . . . . . . . . . . 213
3.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 218

4 More Probability 227

4.1 More Probability Density . . . . . . . . . . . . . . . . . . . . . . 227
4.2 Random Vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . 228
4.2.1 Densities of Random Vectors . . . . . . . . . . . . . . . . 228
4.2.2 Moments of Random Vectors . . . . . . . . . . . . . . . . 229
4.2.3 Functions of Random Vectors . . . . . . . . . . . . . . . . 230
4.3 Representing Distributions . . . . . . . . . . . . . . . . . . . . . . 233
4.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 238

5 Special Distributions 241

5.1 Binomial and Negative Binomial . . . . . . . . . . . . . . . . . . 241
5.2 Multinomial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 250
5.3 Poisson . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 251
5.4 Uniform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 256
5.5 Gamma, Exponential, Chi Square . . . . . . . . . . . . . . . . . . 259
5.6 Beta . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 265
5.7 Normal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 266
5.7.1 The Univariate Normal Distribution . . . . . . . . . . . . 266
5.7.2 The Multivariate Normal Distribution . . . . . . . . . . . 271
5.8 t and F . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 279
5.8.1 The t distribution . . . . . . . . . . . . . . . . . . . . . . 279
5.8.2 The F distribution . . . . . . . . . . . . . . . . . . . . . . 283
5.9 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 283
CONTENTS v

6 More Models 291

6.1 Hierarchical Models . . . . . . . . . . . . . . . . . . . . . . . . . 291
6.2 Markov Chains . . . . . . . . . . . . . . . . . . . . . . . . . . . . 292
6.3 Contingency Tables . . . . . . . . . . . . . . . . . . . . . . . . . . 292
6.4 Survival analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . 292
6.5 The Poisson process . . . . . . . . . . . . . . . . . . . . . . . . . 292
6.6 Change point models . . . . . . . . . . . . . . . . . . . . . . . . . 292
6.7 Evaluating and enhancing models . . . . . . . . . . . . . . . . . . 293

7 Mathematical Statistics 295

7.1 Properties of Statistics . . . . . . . . . . . . . . . . . . . . . . . . 295
7.1.1 Sufficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . 295
7.1.2 Consistency, Bias, and Mean-squared Error . . . . . . . . 298
7.1.3 Efficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . 299
7.1.4 Asymptotic Normality . . . . . . . . . . . . . . . . . . . . 299
7.1.5 Robustness . . . . . . . . . . . . . . . . . . . . . . . . . . 299
7.2 Transformations of Parameters . . . . . . . . . . . . . . . . . . . 299
7.3 Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 300
7.4 More Hypothesis Testing . . . . . . . . . . . . . . . . . . . . . . . 300
7.4.1 p values . . . . . . . . . . . . . . . . . . . . . . . . . . . . 300
7.4.2 The Likelihood Ratio Test . . . . . . . . . . . . . . . . . . 300
7.4.3 The Chi Square Test . . . . . . . . . . . . . . . . . . . . . 300
7.4.4 Power . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 300
7.5 Exponential families . . . . . . . . . . . . . . . . . . . . . . . . . 300
7.6 Location and Scale Families . . . . . . . . . . . . . . . . . . . . . 300
7.7 Functionals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 300
7.8 Invariance . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 300
7.9 Asymptotics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 300
7.10 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 306

Bibliography 309

Index 311

Index of Examples 314

vi CONTENTS
 List of Figures

1.1 pdf for time on hold at Help Line . . . . . . . . . . . . . . . . . . 6

1.2 pY for the outcome of a spinner . . . . . . . . . . . . . . . . . . . 8
1.3 (a): Ocean temperatures; (b): Important discoveries . . . . . . . 10
1.4 Change of variables . . . . . . . . . . . . . . . . . . . . . . . . . . 12
1.5 Binomial probabilities . . . . . . . . . . . . . . . . . . . . . . . . 15
1.6 P[X = 3 | λ] as a function of λ . . . . . . . . . . . . . . . . . . . 18
1.7 Exponential densities . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.8 Normal densities . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.9 Ocean temperatures at 45◦ N, 30◦ W, 1000m depth . . . . . . . . . 23
1.10 Normal samples and Normal densities . . . . . . . . . . . . . . . 25
1.11 hydrographic stations off the coast of Europe and Africa . . . . . 28
1.12 Water temperatures . . . . . . . . . . . . . . . . . . . . . . . . . 29
1.13 Water temperatures with standard deviations . . . . . . . . . . . 35
1.14 Permissible values of N and X . . . . . . . . . . . . . . . . . . . 38
1.15 Features of the joint distribution of (X, Y ) . . . . . . . . . . . . . 42
1.16 Lengths and widths of sepals and petals of 150 iris plants . . . . 45
1.17 correlations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
1.18 1000 simulations of θ̂ for n.sim = 50, 200, 1000 . . . . . . . . . . 52
1.19 1000 simulations of θ̂ under three procedures . . . . . . . . . . . 56
1.20 Monthly concentrations of CO2 at Mauna Loa . . . . . . . . . . . 56
1.21 1000 simulations of a FACE experiment . . . . . . . . . . . . . . 60
1.22 Histograms of craps simulations . . . . . . . . . . . . . . . . . . . 71

2.1 Quantiles. The black circles are the .05, .5, and .9 quantiles.
Panels a and b are for a sample; panels c and d are for a distribution. 82
2.2 Histograms of tooth growth . . . . . . . . . . . . . . . . . . . . . 86
2.3 Histograms of tooth growth . . . . . . . . . . . . . . . . . . . . . 87
2.4 Histograms of tooth growth . . . . . . . . . . . . . . . . . . . . . 88
2.5 calorie contents of beef hot dogs . . . . . . . . . . . . . . . . . . 91
2.6 Strip chart of tooth growth . . . . . . . . . . . . . . . . . . . . . 93
2.7 Quiz scores from Statistics 103 . . . . . . . . . . . . . . . . . . . 96
2.8 QQ plots of water temperatures (◦ C) at 1000m depth . . . . . . 97
2.9 Mosaic plot of UCBAdmissions . . . . . . . . . . . . . . . . . . . 101

vii
viii LIST OF FIGURES

2.10 Mosaic plot of UCBAdmissions . . . . . . . . . . . . . . . . . . . 102

2.11 Durations of and waiting time for eruptions of Old Faithful . . . 105
2.12 Waiting time versus duration in the Old Faithful dataset . . . . . 106
2.13 Time series of duration and waiting time at Old Faithful . . . . . 107
2.14 Time series of duration and waiting time at Old Faithful . . . . . 108
2.15 Temperature versus latitude for different values of longitude . . . 110
2.16 Temperature versus longitude for different values of latitude . . . 111
2.17 LikelihoodP function for the proportion of red cars . . . . . . . . . 113
2.18 `(θ) after yi = 40 in 60 quadrats. . . . . . . . . . . . . . . . . 115
2.19 Likelihood for Slater School . . . . . . . . . . . . . . . . . . . . . 117
2.20 Marginal and exact likelihoods for Slater School . . . . . . . . . . 119
2.21 Marginal likelihood for mean CEO salary . . . . . . . . . . . . . 121
2.22 FACE Experiment: data and likelihood . . . . . . . . . . . . . . 124
2.23 Likelihood function for Quiz Scores . . . . . . . . . . . . . . . . . 126
2.24 Log of the likelihood function for (λ, θf ) in Example 2.12 . . . . 129
2.25 Likelihood function for the probability of winning craps . . . . . 134
2.26 Sampling distribution of the sample mean and median . . . . . . 136
2.27 Histograms of the sample mean for samples from Bin(n, .1) . . . 138
2.28 Prior, likelihood and posterior in the seedlings example . . . . . 144
2.29 Prior, likelihood and posterior densities for λ with n = 1, 4, 16 . . 146
2.30 Prior, likelihood and posterior densities for λ with n = 60 . . . . 147
2.31 Prior, likelihood and posterior density for Slater School . . . . . 149
2.32 Plug-in predictive distribution for seedlings . . . . . . . . . . . . 151
2.33 Predictive distributions for seedlings after n = 0, 1, 60 . . . . . . 153
2.34 pdf of the Bin(100, .5) distribution . . . . . . . . . . . . . . . . . 158
2.35 pdfs of the Bin(100, .5) (dots) and N(50, 5) (line) distributions . . 158
2.36 Approximate density of summary statistic t . . . . . . . . . . . . 160
2.37 Number of times baboon father helps own child . . . . . . . . . . 164
2.38 Histogram of simulated values of w.tot . . . . . . . . . . . . . . . 165

3.1 Four regression examples . . . . . . . . . . . . . . . . . . . . . . . 174

3.2 1970 draft lottery. Draft number vs. day of year . . . . . . . . . 177
3.3 Draft number vs. day of year with smoothers . . . . . . . . . . . 178
3.4 Total number of New seedlings 1993 – 1997, by quadrat. . . . . . 179
3.5 Calorie content of hot dogs . . . . . . . . . . . . . . . . . . . . . 181
3.6 Density estimates of calorie contents of hot dogs . . . . . . . . . 183
3.7 The PlantGrowth data . . . . . . . . . . . . . . . . . . . . . . . . 185
3.8 Ice cream consumption versus mean temperature . . . . . . . . . 190
3.9 Likelihood functions for (µ, δM , δP ) in the Hot Dog example. . . 196
3.10 pairs plot of the mtcars data . . . . . . . . . . . . . . . . . . . . 198
3.11 mtcars — various plots . . . . . . . . . . . . . . . . . . . . . . . 200
3.12 likelihood functions for β1 , γ1 , δ1 and δ2 in the mtcars example. 202
3.13 Pine cones and O-rings . . . . . . . . . . . . . . . . . . . . . . . . 205
3.14 Pine cones and O-rings with regression curves . . . . . . . . . . . 206
3.15 Likelihood function for the pine cone data . . . . . . . . . . . . . 209
3.16 Actual vs. fitted and residuals vs. fitted for the seedling data . . 214
LIST OF FIGURES ix

3.17 Diagnostic plots for the seedling data . . . . . . . . . . . . . . . . 215

3.18 Actual mpg and fitted values from three models . . . . . . . . . . 217
3.19 Happiness Quotient of bankers and poets . . . . . . . . . . . . . 221

4.1 The (X1 , X2 ) plane and the (Y1 , Y2 ) plane . . . . . . . . . . . . . 233

4.2 pmf’s, pdf’s, and cdf’s . . . . . . . . . . . . . . . . . . . . . . . . 235

5.1 The Binomial pmf . . . . . . . . . . . . . . . . . . . . . . . . . . 246

5.2 The Negative Binomial pmf . . . . . . . . . . . . . . . . . . . . . 249
5.3 Poisson pmf for λ = 1, 4, 16, 64 . . . . . . . . . . . . . . . . . . . 254
5.4 Rutherford and Geiger’s Figure 1 . . . . . . . . . . . . . . . . . . 258
5.5 Gamma densities . . . . . . . . . . . . . . . . . . . . . . . . . . . 260
5.6 Exponential densities . . . . . . . . . . . . . . . . . . . . . . . . . 263
5.7 Beta densities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 267
5.8 Water temperatures (◦ C) at 1000m depth . . . . . . . . . . . . . 268
5.9 Bivariate Normal density . . . . . . . . . . . . . . . . . . . . . . 273
5.10 Bivariate Normal density . . . . . . . . . . . . . . . . . . . . . . 276
5.11 t densities for four degrees of freedom and the N(0, 1) density . . 282

6.1 Graphical representation of hierarchical model for fMRI . . . . . 292

7.1 The Be(.39, .01) density . . . . . . . . . . . . . . . . . . . . . . . 304

7.2 Densities of Ȳin . . . . . . . . . . . . . . . . . . . . . . . . . . . . 305
7.3 Densities of Zin . . . . . . . . . . . . . . . . . . . . . . . . . . . . 307
x LIST OF FIGURES
 List of Tables

1.1 Party Affiliation and Referendum Support . . . . . . . . . . . . 36

1.2 Steroid Use and Test Results . . . . . . . . . . . . . . . . . . . . 37

2.1 New and Old seedlings in quadrat 6 in 1992 and 1993 . . . . . . 127

3.1 Correspondence between Models 3.3 and 3.4 . . . . . . . . . . . 184

3.2 β’s for Figure 3.14 . . . . . . . . . . . . . . . . . . . . . . . . . . 207

5.1 Rutherford and Geiger’s data . . . . . . . . . . . . . . . . . . . . 257

xi
xii LIST OF TABLES
 Preface

This book is intended as an upper level undergraduate or introductory graduate

textbook in statistical thinking with a likelihood emphasis for students with a
good knowledge of calculus and the ability to think abstractly. By “statistical
thinking” is meant a focus on ideas that statisticians care about as opposed to
technical details of how to put those ideas into practice. By “likelihood empha-
sis” is meant that the likelihood function and likelihood principle are unifying
ideas throughout the text. Another unusual aspect is the use of statistical soft-
ware as a pedagogical tool. That is, instead of viewing the computer merely
as a convenient and accurate calculating device, we use computer calculation
and simulation as another way of explaining and helping readers understand the
underlying concepts.
Our software of choice is R. R and accompanying manuals are available for
free download from http://www.r-project.org. You may wish to download
An Introduction to R to keep as a reference. It is highly recommended that
you try all the examples in R. They will help you understand concepts, give
you a little programming experience, and give you facility with a very flexible
statistical software package. And don’t just try the examples as written. Vary
them a little; play around with them; experiment. You won’t hurt anything and
you’ll learn a lot.

xiii
xiv LIST OF TABLES
 Chapter 1

Probability

1.1 Basic Probability

Let X be a set and F a collection of subsets of X . A probability measure, or just
a probability, on (X , F) is a function µ : F → [0, 1]. In other words, to every set
in F, µ assigns a probability between 0 and 1. We call µ a set function because
its domain is a collection of sets. But not just any set function will do. To be
a probability µ must satisfy

1. µ(∅) = 0 (∅ is the empty set.),

2. µ(X ) = 1, and

3. if A1 and A2 are disjoint then µ(A1 ∪ A2 ) = µ(A1 ) + µ(A2 ).

One can show that property 3 holds for any finite collection of disjoint sets, not
just two; see Exercise 1. It is common practice, which we adopt in this text,
to assume more — that property 3 also holds for any countable collection of
disjoint sets.
When X is a finite or countably infinite set (usually integers) then µ is said
to be a discrete probability. When X is an interval, either finite or infinite, then
µ is said to be a continuous probability. In the discrete case, F usually contains
all possible subsets of X . But in the continuous case, technical complications
prohibit F from containing all possible subsets of X . See Casella and Berger
[2002] or Schervish [1995] for details. In this text we deemphasize the role of F
and speak of probability measures on X without mentioning F.
In practical examples X is the set of outcomes of an “experiment” and µ is
determined by experience, logic or judgement. For example, consider rolling
a six-sided die. The set of outcomes is {1, 2, 3, 4, 5, 6} so we would assign
X ≡ {1, 2, 3, 4, 5, 6}. If we believe the die to be fair then we would also as-
sign µ({1}) = µ({2}) = · · · = µ({6}) = 1/6. The laws of probability then imply

1
2 CHAPTER 1. PROBABILITY

various other values such as

µ({1, 2}) = 1/3

µ({2, 4, 6}) = 1/2
etc.

Often we omit the braces and write µ(2), µ(5), etc. Setting µ(i) = 1/6 is not
automatic simply because a die has six faces. We set µ(i) = 1/6 because we
believe the die to be fair.
We usually use the word “probability” or the symbol P in place of µ. For
example, we would use the following phrases interchangeably:

• The probability that the die lands 1

• P(1)

• P[the die lands 1]

• µ({1})

We also use the word distribution in place of probability measure.

The next example illustrates how probabilities of complicated events can be
calculated from probabilities of simple events.

Example 1.1 (The Game of Craps)

Craps is a gambling game played with two dice. Here are the rules, as explained on
the website www.online-craps-gambling.com/craps-rules.html.

For the dice thrower (shooter) the object of the game is to throw a
7 or an 11 on the first roll (a win) and avoid throwing a 2, 3 or 12 (a
loss). If none of these numbers (2, 3, 7, 11 or 12) is thrown on the first
throw (the Come-out roll) then a Point is established (the point is the
number rolled) against which the shooter plays. The shooter continues
to throw until one of two numbers is thrown, the Point number or a
Seven. If the shooter rolls the Point before rolling a Seven he/she wins,
however if the shooter throws a Seven before rolling the Point he/she
loses.

Ultimately we would like to calculate P(shooter wins). But for now, let’s just
calculate

P(shooter wins on Come-out roll) = P(7 or 11) = P(7) + P(11).

Using the language of page 1, what is X in this case? Let d 1 denote the number
showing on the first die and d2 denote the number showing on the second die. d1
1.1. BASIC PROBABILITY 3

and d2 are integers from 1 to 6. So X is the set of ordered pairs (d1 , d2 ) or

(6, 6) (6, 5) (6, 4) (6, 3) (6, 2) (6, 1)

(5, 6) (5, 5) (5, 4) (5, 3) (5, 2) (5, 1)
(4, 6) (4, 5) (4, 4) (4, 3) (4, 2) (4, 1)
(3, 6) (3, 5) (3, 4) (3, 3) (3, 2) (3, 1)
(2, 6) (2, 5) (2, 4) (2, 3) (2, 2) (2, 1)
(1, 6) (1, 5) (1, 4) (1, 3) (1, 2) (1, 1)

If the dice are fair, then the pairs are all equally likely. Since there are 36 of them,
we assign P(d1 , d2 ) = 1/36 for any combination (d1 , d2 ). Finally, we can calculate

P(7 or 11) = P(6, 5) + P(5, 6) + P(6, 1) + P(5, 2)

+ P(4, 3) + P(3, 4) + P(2, 5) + P(1, 6) = 8/36 = 2/9.

The previous calculation uses desideratum 3 for probability measures. The different
pairs (6, 5), (5, 6), . . . , (1, 6) are disjoint, so the probability of their union is the
sum of their probabilities.

Example 1.1 illustrates a common situation. We know the probabilities

of some simple events like the rolls of individual dice, and want to calculate
the probabilities of more complicated events like the success of a Come-out
roll. Sometimes those probabilities can be calculated mathematically as in the
example. Other times it is more convenient to calculate them by computer
simulation. We frequently use R to calculate probabilities. To illustrate, Exam-
ple 1.2 uses R to calculate by simulation the same probability we found directly
in Example 1.1.

Example 1.2 (Craps, continued)

To simulate the game of craps, we will have to simulate rolling dice. That’s like
randomly sampling an integer from 1 to 6. The sample() command in R can do
that. For example, the following snippet of code generates one roll from a fair,
six-sided die and shows R’s response:

> sample(1:6,1)
[1] 1
>

When you start R on your computer, you see >, R’s prompt. Then you can type a
command such as sample(1:6,1) which means “take a sample of size 1 from the
numbers 1 through 6”. (It could have been abbreviated sample(6,1).) R responds
with [1] 1. The [1] says how many calculations R has done; you can ignore it.
The 1 is R’s answer to the sample command; it selected the number “1”. Then
it gave another >, showing that it’s ready for another command. Try this several
times; you shouldn’t get “1” every time.
Here’s a longer snippet that does something more useful.
4 CHAPTER 1. PROBABILITY

> x <- sample ( 6, 10, replace=T ) # take a sample of

# size 10 and call it x
> x # print the ten values
[1] 6 4 2 3 4 4 3 6 6 2

> sum ( x == 3 ) # how many are equal to 3?

[1] 2
>

Note

• # is the comment character. On each line, R ignores all text after #.

• We have to tell R to take its sample with replacement. Otherwise, when R

selects “6” the first time, “6” is no longer available to be sampled a second
time. In replace=T, the T stands for True.

• <- does assignment. I.e., the result of sample ( 6, 10, replace=T ) is

assigned to a variable called x. The assignment symbol is two characters: <
followed by -.

• A variable such as x can hold many values simultaneously. When it does, it’s
called a vector. You can refer to individual elements of a vector. For example,
x[1] is the first element of x. x[1] turned out to be 6; x[2] turned out to
be 4; and so on.

• == does comparison. In the snippet above, (x==3) checks, for each element
of x, whether that element is equal to 3. If you just type x == 3 you will see
a string of T’s and F’s (True and False), one for each element of x. Try it.

• The sum command treats T as 1 and F as 0.

• R is almost always tolerant of spaces. You can often leave them out or add
extras where you like.

On average, we expect 1/6 of the draws to equal 1, another 1/6 to equal 2, and
so on. The following snippet is a quick demonstration. We simulate 6000 rolls of a
die and expect about 1000 1’s, 1000 2’s, etc. We count how many we actually get.
This snippet also introduces the for loop, which you should try to understand now
because it will be extremely useful in the future.

> x <- sample(6,6000,replace=T)

> for ( i in 1:6 ) print ( sum ( x==i ))

[1] 995
[1] 1047
1.1. BASIC PROBABILITY 5

[1] 986
[1] 1033
[1] 975
[1] 964
>

Each number from 1 through 6 was chosen about 1000 times, plus or minus a little
bit due to chance variation.
Now let’s get back to craps. We want to simulate a large number of games, say
1000. For each game, we record either 1 or 0, according to whether the shooter
wins on the Come-out roll, or not. We should print out the number of wins at the
end. So we start with a code snippet like this:

# make a vector of length 1000, filled with 0’s

wins <- rep ( 0, 1000 )
for ( i in 1:1000 ) {
simulate a Come-out roll
if shooter wins on Come-out, wins[i] <- 1
}

sum ( wins ) # print the number of wins

Now we have to figure out how to simulate the Come-out roll and decide whether
the shooter wins. Clearly, we begin by simulating the roll of two dice. So our snippet
expands to

# make a vector of length 1000, filled with 0’s

wins <- rep ( 0, 1000 )
for ( i in 1:1000 ) {
d <- sample ( 1:6, 2, replace=T )
if ( sum(d) == 7 || sum(d) == 11 ) wins[i] <- 1
}
sum ( wins ) # print the number of wins

The “||” stands for “or”. So that line of code sets wins[i] <- 1 if the sum of
the rolls is either 7 or 11. When I ran this simulation R printed out 219. The
calculation in Example 1.1 says we should expect around (2/9) × 1000 ≈ 222 wins.
Our calculation and simulation agree about as well as can be expected from a
simulation. Try it yourself a few times. You shouldn’t always get 219. But you
should get around 222 plus or minus a little bit due to the randomness of the
simulation.
Try out these R commands in the version of R installed on your computer. Make
sure you understand them. If you don’t, print out the results. Try variations. Try
any tricks you can think of to help you learn R.
6 CHAPTER 1. PROBABILITY

1.2 Probability Densities

So far we have dealt with discrete probabilities, or the probabilities of at most a
countably infinite number of outcomes. For discrete probabilities, X is usually
a set of integers, either finite or infinite. Section 1.2 deals with the case where
X is an interval, either of finite or infinite length. Some examples are
Medical trials the time until a patient experiences a relapse
Sports the length of a javelin throw
Ecology the lifetime of a tree
Manufacturing the diameter of a ball bearing
Computing the amount of time a Help Line customer spends on hold
Physics the time until a uranium atom decays
Oceanography the temperature of ocean water at a specified latitude, longi-
tude and depth
Probabilities for such outcomes are called continuous. For example, let Y
be the time a Help Line caller spends on hold. The random variable Y is often
modelled with a density similar to that in Figure 1.1.
p(y)

Figure 1.1: pdf for time on hold at Help Line

The curve in the figure is a probability density function or pdf. The pdf is
large near y = 0 and monotonically decreasing, expressing the idea that smaller
1.2. PROBABILITY DENSITIES 7

values of y are more likely than larger values. (Reasonable people may disagree
about whether this pdf accurately represents callers’ experience.) We typically
use the symbols p, π or f for pdf’s. We would write p(50), π(50) or f (50) to
denote the height of the curve at y = 50. For a pdf, probability is the same as
area under the curve. For example, the probability that a caller waits less than
60 minutes is Z 60
P[Y < 60] = p(t) dt.
0
Every pdf must satisfy two properties.
1. p(y) ≥ 0 for all y.
R∞
2. −∞ p(y) dy = 1.
The first property holds because, if p(y) < 0 on the interval (a, b) then P[Y ∈
Rb
(a, b)] = a p(y) dy < 0; and we can’t have probabilities less than 0. The second
R∞
property holds because P[Y ∈ (−∞, ∞)] = −∞ p(y) dy = 1.
One peculiar fact about any continuous random variable Y is that P[Y =
a] = 0, for every a ∈ R. That’s because
Z a+
P[Y = a] = lim P[Y ∈ [a, a + ]] = lim pY (y) dy = 0.
→0 →0 a

Consequently, for any numbers a < b,

P[Y ∈ (a, b)] = P[Y ∈ [a, b)] = P[Y ∈ (a, b]] = P[Y ∈ [a, b]].

The use of “density” in statistics is entirely analagous to its use in physics.

In both fields
mass
density = (1.1)
volume
In statistics, we interpret density as probability density, mass as probability mass
and volume as length of interval. In both fields, if the density varies from place
to place (In physics it would vary within an object; in statistics it would vary
along the real line.) then the density at a particular location is the limit of
Equation 1.1 as volume → 0.
Probability density functions are derivatives of probabilities. For any fixed
number a Z b
d d
P[X ∈ (a, b]] = fX (x) dx = fX (b). (1.2)
db db a
Similarly, d/da P[X ∈ (a, b]] = −fX (a).
Sometimes we can specify pdf’s for continuous random variables based on
the logic of the situation, just as we could specify discrete probabilities based
on the logic of dice rolls. For example, let Y be the outcome of a spinner that
is marked from 0 to 1. Then Y will be somewhere in the unit interval, and all
parts of the interval are equally likely. So the pdf pY must look like Figure 1.2.
8 CHAPTER 1. PROBABILITY

0.0 0.4 0.8

p(y)

0.0 0.2 0.4 0.6 0.8 1.0

Figure 1.2: pY for the outcome of a spinner

Figure 1.2 was produced by the following snippet.

plot ( c(0,1), c(1,1), xlab="y", ylab="p(y)",
ylim=c(0,1.1), type="l" )

Note:
• c(0,1) collects 0 and 1 and puts them into the vector (0,1). Likewise,
c(1,1) creates the vector (1,1).
• plot(x,y,...) produces a plot. The plot(c(0,1), c(1,1), ...)
command above plots the points (x[1],y[1]) = (0,1) and (x[2],y[2]) =
(1,1).
• type="l" says to plot a line instead of individual points.
• xlab and ylab say how the axes are labelled.
• ylim=c(0,1.1) sets the limits of the y-axis on the plot. If ylim is not
specified then R sets the limits automatically. Limits on the x-axis can be
specified with xlim.

At other times we use probability densities and distributions as models for

data, and estimate the densities and distributions directly from the data. Fig-
1.2. PROBABILITY DENSITIES 9

ure 1.3 shows how that works. The upper panel of the figure is a histogram
of 112 measurements of ocean temperature at a depth of 1000 meters in the
North Atlantic near 45◦ North latitude and 20◦ degrees West longitude. Ex-
ample 1.5 will say more about the data. Superimposed on the histogram is a
pdf f . We think of f as underlying the data. The idea is that measuring a
temperature at that location is like randomly drawing a value from f . The 112
measurements, which are spread out over about a century of time, are like 112
independent draws from f . Having the 112 measurements allows us to make a
good estimate of f . If oceanographers return to that location to make additional
measurements, it would be like making additional draws from f . Because we
can estimate f reasonably well, we can predict with some degree of assurance
what the future draws will be like.
The bottom panel of Figure 1.3 is a histogram of the discoveries data
set that comes with R and which is, as R explains, “The numbers of ‘great’
inventions and scientific discoveries in each year from 1860 to 1959.” It is
overlaid with a line showing the Poi(3.1) distribution. (Named distributions will
be introduced in Section 1.3.) It seems that the number of great discoveries each
year follows the Poi(3.1) distribution, at least approximately. If we think the
future will be like the past then we should expect future years to follow a similar
pattern. Again, we think of a distribution underlying the data. The number of
discoveries in a single year is like a draw from the underlying distribution. The
figure shows 100 years, which allow us to estimate the underlying distribution
reasonably well.

Figure 1.3 was produced by the following snippet.

par ( mfrow=c(2,1) )

good <- abs ( med.1000$lon + 20 ) < 1 &

abs ( med.1000$lat - 45 ) < 1
hist ( med.1000$temp[good], xlab="temperature", ylab="",
main="", prob=T, xlim=c(5,11) )
m <- mean ( med.1000$temp[good] )
s <- sqrt ( var ( med.1000$temp[good] ) )
x <- seq ( 5, 11, length=40 )
lines ( density(med.1000$temp[good]) )

hist ( discoveries, xlab="discoveries", ylab="", main="",

prob=T, breaks=seq(-.5,12.5,by=1) )
lines ( 0:12, dpois(0:12, 3.1), type="b" )

Note:
• par sets R’s graphical parameters. mfrow=c(2,1) tells R to make an array
of multiple f igures in a 2 by 1 layout.
10 CHAPTER 1. PROBABILITY

0.4
0.2
0.0

5 6 7 8 9 10 11

temperature
0.20
0.10
0.00

0 2 4 6 8 10 12

discoveries

Figure 1.3: (a): Ocean temperatures at 1000m depth near 45◦ N latitude, -20◦
longitude; (b) Numbers of important discoveries each year 1860–1959
1.2. PROBABILITY DENSITIES 11

• med.1000 is a data set of North Atlantic ocean temperatures at a depth of

1000 meters. med.1000$lon and med.1000$lat are the longitude and lat-
itude of the measurements. med.1000$temp are the actual temperatures.
• abs stands for absolute value.
• good <- ... calls those points good whose longitude is between -19 and
-21 and whose latitude is between 44 and 46.
• hist() makes a histogram. prob=T turns the y-axis into a probability
scale (area under the histogram is 1) instead of counts.
• mean() calculates the mean. var() calculates the variance. Section 1.4
defines the mean and variance of distributions. Section 2.2.1 defines the
mean and variance of data sets.
• lines() adds lines to an existing plot.
• density() estimates a density from a data set.

It is often necessary to transform one variable into another as, for example,
Z = g(X) for some specified function g. We might know pX (The subscript
indicates which random variable we’re talking about.) and want to calculate pZ .
Here we consider only monotonic functions g, so there is an inverse X = h(Z).

Theorem 1.1. Let X be a random variable with pdf pX . Let g be a differen-

tiable, monotonic, invertible function and define Z = g(X). Then the pdf of Z
is  −1 
 d g (t) 
pZ (t) = pX (g −1 (t))  
dt 

Proof. If g is an increasing function then

d d
pZ (b) = P[Z ∈ (a, b]] = P[X ∈ (h(a), h(b)]]
db db
Z h(b)
d
= pX (x) dx = h0 (b)pX (h(b))
db h(a)

The proof when g is decreasing is left as an exercise.

To illustrate, suppose that X is a random variable with pdf pX (x) = 2x on

the unit interval. Let Z = 1/X. What is pZ (z)? The inverse transformation is
X = 1/Z. Its derivative is dx/dz = −z −2 . Therefore,
 −1   
 d g (z)  2  1 
pZ (z) = pX (g −1 (z))   = −  = 2
d z  z  z2  z3
12 CHAPTER 1. PROBABILITY

And the possible values of Z are from 1 to ∞. So pZ (z) = 2/z 3 on the interval
(1, ∞). As a partial check, we can verify that the integral is 1.
Z ∞ ∞
2 1 
dz = − 2  = 1.
1 z3 z 1
Theorem 1.1 can be explained by Figure 1.4. The figure shows an x, a z,
and the function z = g(x). A little interval is shown around x; call it Ix . It gets
mapped by g into a little interval around z; call it Iz . The density is
P[Z ∈ Iz ] P[X ∈ Ix ] length(Ix )
pZ (z) ≈ = ≈ pX (x) |h0 (z)| (1.3)
length(Iz ) length(Ix ) length(Iz )
The approximations in Equation 1.3 are exact as the lengths of Ix and Iz de-
crease to 0.
If g is not one-to-one, then it is often possible to find subsets of R on which
g is one-to-one, and work separately on each subset.

Z=g(X)
z

Figure 1.4: Change of variables

1.3 Parametric Families of Distributions

Probabilities often depend on one or more unknown numerical constants. Sup-
pose, for example, that we have a biased coin. Let θ be the chance that it
lands H. Then P(H) = θ; but we might not know θ; it is an unknown nu-
merical constant. In this case we have a family of probability measures, one
1.3. PARAMETRIC FAMILIES OF DISTRIBUTIONS 13

for each value of θ, and we don’t know which one is right. When we need to
be explicit that probabilities depend on θ, we use the notation, for example,
P(H | θ) or P(H | θ = 1/3). The vertical bar is read “given” or “given that”. So
P(H | θ = 1/3) is read “the probability of Heads given that θ equals 1/3” and
P(H | θ) is read “the probability of Heads given θ.” This notation means

P(H | θ = 1/3) = 1/3,

P(T | θ = 1/3) = 2/3,
P(T | θ = 1/5) = 4/5

and so on. Instead of “given” we also use the word “conditional”. So we would
say “the probability of Heads conditional on θ”, etc.
The unknown constant θ is called a parameter. The set of possible values
for θ is denoted Θ (upper case θ). For each θ there is a probability measure µθ .
The set of all possible probability measures (for the problem at hand),

{µθ : θ ∈ Θ},

is called a parametric family of probability measures. The rest of this chapter

introduces four of the most useful parametric families of probability measures.

1.3.1 The Binomial Distribution

Statisticians often have to consider observations of the following type.
• A repeatable event results in either a success or a failure.
• Many repetitions are observed.
• Successes and failures are counted.
• The number of successes helps us learn about the probability of success.
Such observations are called binomial. Some examples are
Medical trials A new treatment is given to many patients. Each is either
cured or not.
Toxicity tests Many laboratory animals are exposed to a potential carcinogen.
Each either develops cancer or not.
Ecology Many seeds are planted. Each either germinates or not.
Quality control Many supposedly identical items are subjected to a test.
Each either passes or not.
Because binomial experiments are so prevalent there is specialized language to
describe them. Each repetition is called a trial ; the number of trials is usually
denoted N ; the unknown probability of success is usually denoted either p or θ;
the number of successes is usually denoted X. We write X ∼ Bin(N, p). The
14 CHAPTER 1. PROBABILITY

symbol “∼” is read is distributed as; we would say “X is distributed as Binomial

N p” or “X has the Binomial N, p distribution”. Some important assumptions
about binomial experiments are that N is fixed in advance, θ is the same for
every trial, and the outcome of any trial does not influence the outcome of any
other trial. When N = 1 we say X has a Bernoulli(θ) distribution and write
X ∼ Bern(θ); the individual trials in a binomial experiment are called Bernoulli
trials.
When a binomial experiment is performed, X will turn out to be one of
the integers from 0 to N . We need to know the associated probabilities; i.e.
P[X = k | θ] for each value of k from 0 to N . These probabilities are given by
Equation 1.4 whose derivation is given in Section 5.1.
 
N k
P[X = k | θ] = θ (1 − θ)N −k (1.4)
k



The term N k is called a binomial coefficient and is read “N choose k”.
N
k =
N!
k!(N −k)! and is equal to the number of subsets of size k that can be formed from
a group of N distinct items. In case k = 0 or k = N , 0! is defined to be 1.
Figure 1.5 shows binomial probabilities for N ∈ {3, 30, 300} and p ∈ {.1, .5, .9}.

Example 1.3 (Craps, continued)

This example continues the game of craps. See Examples 1.1 and 1.2.
What is the probability that at least one of the next four players wins on his
Come-out roll?
This is a Binomial experiment because
1. We are looking at repeated trials. Each Come-out roll is a trial. It results in
either success, or not.
2. The outcome of one trial does not affect the other trials.
3. We are counting the number of successes.
Let X be the number of success. There are four trials, so N = 4. We calculated
the probability of success in Example 1.1; it’s p = 2/9. So X ∼ Bin(4, 2/9). The
probability of success in at least one Come-out roll is
P[success in at least one Come-out roll] = P[X ≥ 1]
4 4  
X X 4
= P[X = i] = (2/9)i (7/9)4−i ≈ 0.634 (1.5)
i=1 i=1
i
A convenient way to re-express Equation 1.5 is
P[X ≥ 1] = 1 − P[X = 0],
which can be quickly calculated in R. The dbinom() command computes Bino-
mial probabilities. To compute Equation 1.5 we would write
1 - dbinom(0,4,2/9)
The 0 says what value of X we want. The 4 and the 2/9 are the number of
trials and the probability of success. Try it. Learn it.
1.3. PARAMETRIC FAMILIES OF DISTRIBUTIONS 15

N=3 N=3 N=3

p = 0.1 p = 0.5 p = 0.9

0.35
probability

probability

probability
0.6

0.6
0.15
0.0

0.0
0 1 2 3 0 1 2 3 0 1 2 3

x x x

N = 30 N = 30 N = 30
p = 0.1 p = 0.5 p = 0.9
0.15
0.20

0.20
probability

probability

probability
0.00

0.00

0.00

0 10 30 0 10 30 0 10 30

x x x

N = 300 N = 300 N = 300

p = 0.1 p = 0.5 p = 0.9
0.04
probability

probability

probability
0.06

0.06
0.00

0.00

0.00

0 200 0 200 0 200

x x x

Figure 1.5: Binomial probabilities

16 CHAPTER 1. PROBABILITY

1.3.2 The Poisson Distribution

Another common type of observation occurs in the following situation.

• There is a domain of study, usually a block of space or time.

• Events arise seemingly at random in the domain.

• There is an underlying rate at which events arise.

Such observations are called Poisson after the 19th century French mathemati-
cian Siméon-Denis Poisson. The number of events in the domain of study helps
us learn about the rate. Some examples are

Ecology Tree seedlings emerge from the forest floor.

Computer programming Bugs occur in computer code.

Quality control Defects occur along a strand of yarn.

Genetics Mutations occur in a population.

Traffic flow Cars arrive at an intersection.

Customer service Customers arrive at a service counter.

The rate at which events occur is often called λ; the number of events that occur
in the domain of study is often called X; we write X ∼ Poi(λ). Important
assumptions about Poisson observations are that two events cannot occur at
exactly the same location in space or time, that the occurence of an event at
location `1 does not influence whether an event occurs at any other location `2 ,
and the rate at which events arise does not vary over the domain of study.
When a Poisson experiment is observed, X will turn out to be a nonnegative
integer. The associated probabilities are given by Equation 1.6.

λk e−λ
P[X = k | λ] = . (1.6)
k!
One of the main themes of statistics is the quantitative way in which data
help us learn about the phenomenon we are studying. Example 1.4 shows how
this works when we want to learn about the rate λ of a Poisson distribution.

Example 1.4 (Seedlings in a Forest)

Tree populations move by dispersing their seeds. Seeds become seedlings, seedlings
become saplings, and saplings become adults which eventually produce more seeds.
Over time, whole populations may migrate in response to climate change. One
instance occurred at the end of the Ice Age when species that had been sequestered
in the south were free to move north. Another instance may be occurring today
in response to global warming. One critical feature of the migration is its speed.
Some of the factors determining the speed are the typical distances of long range
1.3. PARAMETRIC FAMILIES OF DISTRIBUTIONS 17

seed dispersal, the proportion of seeds that germinate and emerge from the forest
floor to become seedlings, and the proportion of seedlings that survive each year.
To learn about emergence and survival, ecologists return annually to forest
quadrats (square meter sites) to count seedlings that have emerged since the pre-
vious year. One such study was reported in Lavine et al. [2002]. A fundamental
quantity of interest is the rate λ at which seedlings emerge. Suppose that, in one
quadrat, three new seedlings are observed. What does that say about λ?
Different values of λ yield different values of P[X = 3 | λ]. To compare different
values of λ we see how well each one explains the data X = 3; i.e., we compare
P[X = 3 | λ] for different values of λ. For example,

13 e−1
P[X = 3 | λ = 1] = ≈ 0.06
3!
23 e−2
P[X = 3 | λ = 2] = ≈ 0.18
3!
33 e−3
P[X = 3 | λ = 3] = ≈ 0.22
3!
43 e−4
P[X = 3 | λ = 4] = ≈ 0.14
3!
In other words, the value λ = 3 explains the data almost four times as well as
the value λ = 1 and just a little bit better than the values λ = 2 and λ = 4.
Figure 1.6 shows P[X = 3 | λ] plotted as a function of λ. The figure suggests
that P[X = 3 | λ] is maximized by λ = 3. The suggestion can be verified by
differentiating Equation 1.6 with respect to lambda, equating to 0, and solving.
The figure also shows that any value of λ from about 0.5 to about 9 explains the
data not too much worse than λ = 3.

Figure 1.6 was produced by the following snippet.

lam <- seq ( 0, 10, length=50 )
y <- dpois ( 3, lam )
plot ( lam, y, xlab="lambda", ylab="P[x=3]", type="l" )

Note:
• seq stands for “sequence”. seq(0,10,length=50) produces a sequence of
50 numbers evenly spaced from 0 to 10.
• dpois calculates probabilities for Poisson distributions the way dbinom
does for Binomial distributions.
• plot produces a plot. In the plot(...) command above, lam goes on the
x-axis, y goes on the y-axis, xlab and ylab say how the axes are labelled,
and type="l" says to plot a line instead of indvidual points.
18 CHAPTER 1. PROBABILITY

0.00 0.10 0.20

P[x=3]

0 2 4 6 8 10

lambda

Figure 1.6: P[X = 3 | λ] as a function of λ

Making and interpreting plots is a big part of statistics. Figure 1.6 is a good
example. Just by looking at the figure we were able to tell which values of λ are
plausible and which are not. Most of the figures in this book were produced in
R.

1.3.3 The Exponential Distribution

It is often necessary to model a continuous random variable X whose density
decreases away from 0. Some examples are

Customer service time on hold at a help line

Neurobiology time until the next neuron fires

Seismology time until the next earthquake

Medicine remaining years of life for a cancer patient

Ecology dispersal distance of a seed

In these examples it is expected that most calls, times or distances will be

short and a few will be long. So the density should be large near x = 0 and
decreasing as x increases. A useful pdf for such situations is the Exponential
density
1 x
p(x) = e− λ for x > 0. (1.7)
λ
1.3. PARAMETRIC FAMILIES OF DISTRIBUTIONS 19

We say X has an exponential distribution with parameter λ and write X ∼

Exp(λ). Figure 1.7 shows exponential densities for several different values of λ.
10

lambda = 2
lambda = 1
lambda = 0.2
8

lambda = 0.1
6
p(x)

4
2
0

0.0 0.5 1.0 1.5 2.0

Figure 1.7: Exponential densities

Figure 1.7 was produced by the following snippet.

x <- seq ( 0, 2, length=40 ) # 40 values from 0 to 2
lam <- c ( 2, 1, .2, .1 ) # 4 different values of lambda
y <- matrix ( NA, 40, 4 ) # y values for plotting
for ( i in 1:4 )
y[,i] <- dexp ( x, 1/lam[i] ) # exponential pdf
matplot ( x, y, type="l", xlab="x", ylab="p(x)", col=1 )
legend ( 1.2, 10, paste ( "lambda =", lam ),
20 CHAPTER 1. PROBABILITY

lty=1:4, cex=.75 )

• We want to plot the exponential density for several different values of λ

so we choose 40 values of x between 0 and 2 at which to do the plotting.
• Next we choose 4 values of λ.
• Then we need to calculate and save p(x) for each combination of x and
λ. We’ll save them in a matrix called y. matrix(NA,40,4) creates the
matrix. The size of the matrix is 40 by 4. It is filled with NA, or Not
Available.
• dexp calculates the exponential pdf. The argument x tells R the x values
at which to calculate the pdf. x can be a vector. The argument 1/lam[i]
tells R the value of the parameter. R uses a different notation than this
book. Where this book says Exp(2), R says Exp(.5). That’s the reason for
the 1/lam[i].

• matplot plots one matrix versus another. The first matrix is x and the
second is y. matplot plots each column of y against each column of x. In
our case x is vector, so matplot plots each column of y, in turn, against
x. type="l" says to plot lines instead of points. col=1 says to use the
first color in R’s library of colors.
• legend (...) puts a legend on the plot. The 1.2 and 10 are the x and
y coordinates of the upper left corner of the legend box. lty=1:4 says to
use line types 1 through 4. cex=.75 sets the character expansion factor
to .75. In other words, it sets the font size.
• paste(..) creates the words that go into the legend. It pastes together
”lambda =” with the four values of lam.

1.3.4 The Normal Distribution

It is often necessary to model a continuous random variable Y whose density is
mound-shaped. Some examples are

Biological Anthropology Heights of people

Oceanography Ocean temperatures at a particular location

Quality Control Diameters of ball bearings

Education SAT scores

In each case the random variable is expected to have a central value around
which most of the observations cluster. Fewer and fewer observations are farther
1.3. PARAMETRIC FAMILIES OF DISTRIBUTIONS 21

and farther away from the center. So the pdf should be unimodal — large in
the center and decreasing in both directions away from the center. A useful pdf
for such situations is the Normal density
1 1 y−µ 2
p(y) = √ e− 2 ( σ ) . (1.8)
2πσ
We say Y has a Normal distribution with mean µ and standard deviation σ
and write Y ∼ N(µ, σ). Figure 1.8 shows Normal densities for several different
values of (µ, σ). As illustrated by the figure, µ controls the center of the density;
each pdf is centered over its own value of µ. On the other hand, σ controls the
spread. pdf’s with larger values of σ are more spread out; pdf’s with smaller σ
are tighter.

Figure 1.8 was produced by the following snippet.

x <- seq ( -6, 6, len=100 )
y <- cbind ( dnorm ( x, -2, 1 ),
dnorm ( x, 0, 2 ),
dnorm ( x, 0, .5 ),
dnorm ( x, 2, .3 ),
dnorm ( x, -.5, 3 )
)
matplot ( x, y, type="l", col=1 )
legend ( -6, 1.3, paste( "mu =", c(-2,0,0,2,-.5),
"; sigma =", c(1,2,.5,.3,3) ),
lty=1:5, col=1, cex=.75 )

• dnorm(...) computes the Normal pdf. The first argument is the set of
x values; the second argument is the mean; the third argument is the
standard deviation.

As a further illustration, Figure 1.9 shows a histogram of 105 ocean temper-

atures (◦ C) recorded in the Atlantic Ocean from about 1938 to 1997 at a depth
of 1000 meters, near 45 degrees North latitude and 30 degrees West longitude.
The N(5.87, .72) density is superimposed on the histogram. The Normal density
represents the data moderately well. We will study ocean temperatures in much
more detail in a series of examples beginning with Example 1.5.

Figure 1.9 was produced by

hist ( y, prob=T, xlab="temperature", ylab="density",
ylim=c(0,.6), main="" )
t <- seq ( 4, 7.5, length=40 )
lines ( t, dnorm ( t, mean(y), sd(y) ) )
22 CHAPTER 1. PROBABILITY

mu = −2 ; sigma = 1
mu = 0 ; sigma = 2
1.2

mu = 0 ; sigma = 0.5
mu = 2 ; sigma = 0.3
mu = −0.5 ; sigma = 3
1.0
0.8
y

0.6
0.4
0.2
0.0

−6 −4 −2 0 2 4 6

Figure 1.8: Normal densities

1.3. PARAMETRIC FAMILIES OF DISTRIBUTIONS 23

0.6
0.4
density

0.2
0.0

4.0 4.5 5.0 5.5 6.0 6.5 7.0 7.5

temperature

Figure 1.9: Ocean temperatures at 45◦ N, 30◦ W, 1000m depth. The N(5.87, .72)
density.
24 CHAPTER 1. PROBABILITY

• The 105 temperatures are in a vector y.

• hist produces a histogram. The argument prob=T causes the vertical
scale to be probability density instead of counts.
• The line t <- ... sets 40 values of t in the interval [4, 7.5] at which to
evaluate the Normal density for plotting purposes.
• lines displays the Normal density.

As usual, you should try to understand the R commands.

The function rnorm(n,mu,sig) generates a random sample from a Normal

distribution. n is the sample size; mu is the mean; and sig is the standard devi-
ation. To demonstrate, we’ll generate a sample of size 100 from the N(5.87, .72)
density, the density in Figure 1.9, and compare the sample histogram to the
theoretical density. Figure 1.10(a) shows the comparison. It shows about how
good a fit can be expected between a histogram and the Normal density, for a
sample of size around 100 in the most ideal case when the sample was actually
generated from the Normal distribution. It is interesting to consider whether
the fit in Figure 1.9 is much worse.

Figure 1.10(a) was produced by

samp <- rnorm ( 100, 5.87, .72 )
y.vals <- seq ( 4, 7.5, length=40 )
hist ( samp, prob=T, main="(a)",
xlim=c(4,7.5), xlab="degrees C",
ylim=c(0,.6), ylab="density" )
lines ( y.vals, dnorm(y.vals,5.87,.72) )

When working with Normal distributions it is extremely useful to think in

terms of units of standard deviation, or simply standard units. One standard
unit equals one standard deviation. In Figure 1.10(a) the number 6.6 is about
1 standard unit above the mean, while the number 4.5 is about 2 standard units
below the mean. To see why that’s a useful way to think, Figure 1.10(b) takes
the sample from Figure 1.10(a), multiplies by 9/5 and adds 32, to simulate tem-
peratures measured in ◦ F instead of ◦ C. The histograms in panels (a) and (b)
are slightly different because R has chosen the bin boundaries differently; but
the two Normal curves have identical shapes. Now consider some temperatures,
say 6.5◦ C = 43.7◦ F and 4.5◦ C = 40.1◦ F. Corresponding temperatures occupy
corresponding points on the plots. A vertical line at 6.5 in panel (a) divides
the density into two sections exactly congruent to the two sections created by
a vertical line at 43.7 in panel (b). A similar statement holds for 4.5 and 40.1.
The point is that the two density curves have exactly the same shape. They are
identical except for the scale on the horizontal axis, and that scale is determined
1.3. PARAMETRIC FAMILIES OF DISTRIBUTIONS 25

0.6
0.4
(a)
density

0.2
0.0

4.0 4.5 5.0 5.5 6.0 6.5 7.0 7.5

degrees C

(b)
0.30
density

0.15
0.00

39 40 41 42 43 44 45

degrees F

(c)
0.8
density

0.4
0.0

−2 −1 0 1 2

standard units

Figure 1.10: (a): A sample of size 100 from N(5.87, .72) and the N(5.87, .72) den-
sity. (b): A sample of size 100 from N(42.566, 1.296) and the N(42.566, 1.296)
density. (c): A sample of size 100 from N(0, 1) and the N(0, 1) density.
26 CHAPTER 1. PROBABILITY

by the standard deviation. Standard units are a scale-free way of thinking about
the picture.
To continue, we converted the temperatures in panels (a) and (b) to stan-
dard units, and plotted them in panel (c). Once again, R made a slightly
different choice for the bin boundaries, but the Normal curves all have the same
shape.

Panels (b) and (c) of Figure 1.10 were produced by

y2samp <- samp * 9/5 + 32
y2.vals <- y.vals * 9/5 + 32
hist ( y2samp, prob=T, main="(b)",
xlim=c(39.2,45.5), xlab="degrees F",
ylim=c(0,1/3), ylab="density" )
lines ( y2.vals, dnorm(y2.vals,42.566,1.296) )

zsamp <- (samp-5.87) / .72

z.vals <- (y.vals-5.87) / .72
hist ( zsamp, prob=T, main="(c)",
xlim=c(-2.6,2.26), xlab="standard units",
ylim=c(0,.833), ylab="density" )
lines ( z.vals, dnorm(z.vals,0,1) )

Let Y ∼ N(µ, σ) and define a new random variable Z = (Y − µ)/σ. Z is in

standard units. It tells how many standard units Y is above or below its mean
µ. What is the distribution of Z? The easiest way to find out is to calculate
pZ , the density of Z, and see whether we recognize it. From Theorem 1.1,
1 1 2
pZ (z) = pY (σz + µ)σ = √ e− 2 z
2π
which we recognize as the N(0, 1) density. I.e., Z ∼ N(0, 1). The N(0, 1) distri-
bution is called the standard Normal distribution.

1.4 Centers, Spreads, Means, and Moments

Recall Figure 1.3 (pg. 10). In each panel there is a histogram of a data set along
with an estimate of the underlying pdf or pmf p. In each case we have found
a distribution that matches the data reasonably well, but the distributions we
have drawn are not the only ones that match well. We could make modest
changes to either distribution and still have a reasonably good match. But
whatever pdf we propose for the top panel should be roughly mound shaped
with a center around 8◦ and a spread that ranges from about 6◦ to about 10◦ .
And in the bottom panel we would want a distribution with a peak around 2
or 3 and a longish right hand tail.
1.4. CENTERS, SPREADS, MEANS, AND MOMENTS 27

In either case, the details of the distribution matter less than these central
features. So statisticians often need to refer to the center, or location, of a
sample or a distribution and also to its spread. Section 1.4 gives some of the
theoretical underpinnings for talking about centers and spreads of distributions.

Example 1.5
Physical oceanographers study physical properties such as temperature, salinity,
pressure, oxygen concentration, and potential vorticity of the world’s oceans. Data
about the oceans’ surface can be collected by satellites’ bouncing signals off the sur-
face. But satellites cannot collect data about deep ocean water. Until as recently as
the 1970s, the main source of data about deep water came from ships that lower in-
struments to various depths to record properties of ocean water such as temperature,
pressure, salinity, etc. (Since about the 1970s oceanographers have begun to employ
neutrally buoyant floats. A brief description and history of the floats can be found
on the web at www.soc.soton.ac.uk/JRD/HYDRO/shb/float.history.html.)
Figure 1.11 shows locations, called hydrographic stations, off the coast of Europe
and Africa where ship-based measurements were taken between about 1910 and
1990. The outline of the continents is apparent on the right-hand side of the figure
due to the lack of measurements over land.
Deep ocean currents cannot be seen but can be inferred from physical properties.
Figure 1.12 shows temperatures recorded over time at a depth of 1000 meters at
nine different locations. The upper right panel in Figure 1.12 is the same as the top
panel of Figure 1.3. Each histogram in Figure 1.12 has a black circle indicating the
“center” or “location” of the points that make up the histogram. These centers are
good estimates of the centers of the underlying pdf’s. The centers range from a low
of about 5◦ at latitude 45 and longitude -40 to a high of about 9 ◦ at latitude 35
and longitude -20. (By convention, longitudes to the west of Greenwich, England
are negative; longitudes to the east of Greenwich are positive.) It’s apparent from
the centers that for each latitude, temperatures tend to get colder as we move from
east to west. For each longitude, temperatures are warmest at the middle latitude
and colder to the north and south. Data like these allow oceanographers to deduce
the presence of a large outpouring of relatively warm water called the Mediterranean
tongue from the Mediterranean Sea into the Atlantic ocean. The Mediterranean
tongue is centered at about 1000 meters depth and 35 ◦ N latitude, flows from east
to west, and is warmer than the surrounding Atlantic waters into which it flows.

There are many ways of describing the center of a data sample. But by far
the most common is the mean. The mean of a sample, or of any list of numbers,
is just the average.
Definition 1.1 (Mean of a sample). The mean of a sample, or any list of
numbers, x1 , . . . , xn is
1X
mean of x1 , . . . , xn = xi . (1.9)
n
The black circles in Figure 1.12 are means. The mean of x1 , . . . , xn is often
28 CHAPTER 1. PROBABILITY

60
50
latitude

40
30
20

−40 −30 −20 −10 0

longitude

Figure 1.11: hydrographic stations off the coast of Europe and Africa
1.4. CENTERS, SPREADS, MEANS, AND MOMENTS 29

latitude = 45 latitude = 45 latitude = 45

longitude = −40 longitude = −30 longitude = −20

20
60

20
40

10
10
20

5
5
0

0
4 6 8 4 6 8 4 6 8

temperature temperature temperature

n = 213 n = 105 n = 112

latitude = 35 latitude = 35 latitude = 35

longitude = −40 longitude = −30 longitude = −20
12
20

15
15

0 2 4 6 8

10
10

5
5
0

4 6 8 4 6 8 4 6 8

temperature temperature temperature

n = 37 n = 24 n = 44

latitude = 25 latitude = 25 latitude = 25

longitude = −40 longitude = −30 longitude = −20
15

15
20

10

10
5 10

5
0

4 6 8 4 6 8 4 6 8

temperature temperature temperature

n = 47 n = 35 n = 27

Figure 1.12: Water temperatures (◦ C) at 1000m depth, latitude 25, 35, 45

degrees North longitude 20, 30, 40 degrees West
30 CHAPTER 1. PROBABILITY

denoted x̄. Means are often a good first step in describing data that are unimodal
and roughly symmetric.
Similarly, means are often useful in describing distributions. For example,
the mean of the pdf in the upper panel of Figure 1.3 is about 8.1, the same as
the mean of the data in same panel. Similarly, in the bottom panel, the mean of
the Poi(3.1) distribution is 3.1, the same as the mean of the discoveries data.
Of course we chose the distributions to have means that matched the means of
the data.
For some other examples, consider the Bin(n, p) distributions shown in Fig-
ure 1.5. The center of the Bin(30, .5) distribution appears to be around 15, the
center of the Bin(300, .9) distribution appears to be around 270, and so on. The
mean of a distribution, or of a random variable, is also called the expected value
or expectation and is written E(X).
Definition 1.2 (Mean of a random variable). Let X be a random variable
with cdf FX and pdf pX . Then the mean of X (equivalently, the mean of FX )
is (P
i P[X = i] if X is discrete
E(X) = R i (1.10)
x pX (x) dx if X is continuous

The logic of the definition is that E(X) is a weighted average of the possible
values of X. Each value is weighted by its importance, or probability. In
addition to E(X), another common notation for the mean of a random variable
X is µX .
Let’s look at some of the families of probability distributions that we have
already studied and calculate their expectations.
Binomial If X ∼ Bin(n, p) then
n
X
E(X) = i P[x = i]
i=0
n  
X n i
= i p (1 − p)n−i
i=0
i
n  
X n i
= i p (1 − p)n−i
i=1
i (1.11)
n
X (n − 1)!
= np pi−1 (1 − p)n−i
i=1
(i − 1)!(n − i)!
n−1
X (n − 1)!
= np pj (1 − p)n−1−j
j=0
j!(n − 1 − j)!
= np

The first five

Pn−1equalities are just algebra. The sixth is worth remembering.
The sum j=0 · · · is the sum of the probabilities of the Bin(n − 1, p)
1.4. CENTERS, SPREADS, MEANS, AND MOMENTS 31

distribution. Therefore the sum is equal to 1. You may wish to compare

E(X) to Figure 1.5.

Poisson If X ∼ Poi(λ) then

E(X) = λ.
The derivation is left as Exercise 18.

Exponential If X ∼ Exp(λ) then

Z ∞ Z ∞
−1
E(X) = x p(x) dx = λ xe−x/λ dx = 1/λ.
0 0

Use integration by parts.

Normal If X ∼ N(µ, σ) then E(X) = µ. The derivation is left as Exercise 18.

Statisticians also need to measure and describe the spread of distributions,

random variables and samples. In Figure 1.12, the spread would measure how
much variation there is in ocean temperatures at a single location, which in
turn would tell us something about how heat moves from place to place in
the ocean. Spread could also describe the variation in the annual numbers of
“great” discoveries, the range of typical outcomes for a gambler playing a game
repeatedly at a casino or an investor in the stock market, or the uncertain effect
of a change in the Federal Reserve Bank’s monetary policy, or even why different
patches of the same forest have different plants on them.
By far the most common measures of spread are the variance and its square
root, the standard deviation.

Definition 1.3 (Variance). The variance of a sample y1 , . . . , yn is

X
Var(y1 , . . . , yn ) = n−1 (yi − ȳ)2 .

The variance of a random variable Y is

Var(Y ) = E((Y − µY )2 )

Definition 1.4 (Standard deviation). The standard deviation of a sample

y1 , . . . , yn is q X
SD(y1 , . . . , yn ) = n−1 (yi − ȳ))2 .
The standard deviation of a random variable Y is
p
SD(Y ) = E((Y − µY )2 ).

The variance (standard deviation) of Y is often denoted σY2 (σY ). The

variances of common distributions will be derived later in the book.
Caution: for reasons which we don’t go into here,Pmany books define the
variance of a sample as Var(y1 , . . . , yn ) = (n − 1)−1 (yi − ȳ)2 . For large n
32 CHAPTER 1. PROBABILITY

there is no practical difference between the two definitions. And the definition
of variance of a random variable remains unchanged.
While the definition of the variance of a random variable highlights its in-
terpretation as deviations away from the mean, there is an equivalent formula
that is sometimes easier to compute.
Theorem 1.2. If Y is a random variable, then Var(Y ) = E(Y 2 ) − (EY )2 .
Proof.

Var(Y ) = E((Y − EY )2 )
= E(Y 2 − 2Y EY + (EY )2 )
= E(Y 2 ) − 2(EY )2 + (EY )2
= E(Y 2 ) − (EY )2

To develop a feel for what the standard deviation measures, Figure 1.13
repeats Figure 1.12 and adds arrows showing ± 1 standard deviation away from
the mean. Standard deviations have the same units as the original random
variable; variances have squared units. E.g., if Y is measured in degrees, then
SD(Y ) is in degrees but Var(Y ) is in degrees2 . Because of this, SD is easier to
interpret graphically. That’s why we were able to depict SD’s in Figure 1.13.
Most “mound-shaped” samples, that is, samples that are unimodal and
roughly symmetric, follow this rule of thumb:
• about 2/3 of the sample falls within about 1 standard deviation of the
mean;
• about 95% of the sample falls within about 2 standard deviations of the
mean.
The rule of thumb has implications for predictive accuracy. If x1 , . . . , xn are a
sample from a mound-shaped distribution, then one would predict that future
observations will be around x̄ with, again, about 2/3 of them within about one
SD and about 95% of them within about two SD’s.
Definition 1.5 (Moment). The r’th moment of a sample y1 , . . . , yn or random
variable Y is defined as
X
n−1 (yi − ȳ)r (for samples)
E((Y − µY )r ) (for random variables)

Variances are second moments. Moments above the second have little appli-
cability.
R has built-in functions to compute means and variances and can compute
other moments easily. Note that R uses the divisor n − 1 in its definition of
variance.
1.4. CENTERS, SPREADS, MEANS, AND MOMENTS 33

# Use R to calculate moments of the Bin(100,.5) distribution

x <- rbinom ( 5000, 100, .5 )
m <- mean ( x ) # the mean
v <- var ( x ) # the variance
s <- sqrt ( v ) # the SD
mean ( (x-m)^3 ) # the third moment
our.v <- mean ( (x-m)^2 ) # our variance
our.s <- sqrt ( our.v ) # our standard deviation
print ( c ( v, our.v ) ) # not quite equal
print ( c ( s, our.s ) ) # not quite equal

• rbinom(...) generates random draws from the binomial distribution.

The 5000 says how many draws to generate. The 100 and .5 say that the
draws are to be from the Bin(100, .5) distribution.
R P
Let h be a function. Then E[h(Y )] = h(y)p(y) dy ( h(y)p(y) in the
discrete case) is the expected value of h(Y ) and is called a generalized moment.
There are sometimes two ways to evaluate E[h(Y )]. One is to evaluate the
integral.
R The other is to let X = h(Y ), find pX , and then evaluate E[X] =
xpX (x) dx. For example, let Y have pdf fY (y) = 1 for y ∈ (0, 1), and let
X = h(Y ) = exp(Y ).
Method 1 Z 1
1
E[h(Y )] = exp(y) dy = exp(y)|0 = e − 1.
0

Method 2

pX (x) = pY (log(x)) dy/dx = 1/x

Z e Z e
E[X] = xpx (x) dx = 1 dx = e − 1
1 1

If h is a linear function then E[h(Y )] has a particularly appealing form.

Theorem 1.3. If X = a + bY then E[X] = a + bE[Y ].

Proof. We prove the continuous case; the discrete case is left as an exercise.
Z
EX = (a + by)fY (y) dy
Z Z
= a fY (y) dy + b yfy (y) dy

= a + bEY

There is a corresponding theorem for variance.

Theorem 1.4. If X = a + bY then Var(X) = b2 Var(Y ).

34 CHAPTER 1. PROBABILITY

Proof. We prove the continuous case; the discrete case is left as an exercise. Let
µ = E[Y ].
Var(X) = E[(a + bY − (a + bµ))2 ]
= E[b(Y − µ)2 ]
= b2 Var(Y )

1.5 Joint, Marginal and Conditional Probability

Statisticians often have to deal simultaneously with the probabilities of several
events, quantities, or random variables. For example, we may classify voters
in a city according to political party affiliation and support for a school bond
referendum. Let A and S be a voter’s affiliation and support, respectively.
( (
D if Democrat Y if in favor
A= and S=
R if Republican. N if opposed

Suppose a polling organization finds that 80% of Democrats and 35% of Re-
publicans favor the bond referendum. The 80% and 35% are called conditional
probabilities because they are conditional on party affiliation. The notation
for conditional probabilities is pS | A . As usual, the subscript indicates which
random variables we’re talking about. Specifically,
pS | A (Y | D) = 0.80; pS | A (N | D) = 0.20;
pS | A (Y | R) = 0.35; pS | A (N | R) = 0.65.
We say “the conditional probability that S = N given A = D is 0.20”, etc.
Suppose further that 60% of voters in the city are Democrats. Then 80% of
60% = 48% of the voters are Democrats who favor the referendum. The 48% is
called a joint probability because it is the probability of (A = D, S = Y ) jointly.
The notation is pA,S (D, Y ) = .48. Likewise, pA,S (D, N ) = .12; pA,S (R, Y ) =
.14; and pA,S (R, N ) = 0.26. Table 1.1 summarizes the calculations. The quan-
tities .60, .40, .62, and .38 are called marginal probabilities. The name derives
from historical reasons, because they were written in the margins of the ta-
ble. Marginal probabilities are probabilities for one variable alone, the ordinary
probabilities that we’ve been talking about all along.
The event A = D can be partitioned into the two smaller events (A = D, S =
Y ) and (A = D, S = N ). So
pA (D) = .60 = .48 + .12 = pA,S (D, Y ) + pA,S (D, N ).
The event A = R can be partitioned similarly. Too, the event S = Y can be
partitioned into (A = D, S = Y ) and (A = R, S = Y ). So
pS (Y ) = .62 = .48 + .14 = pA,S (D, Y ) + pA,S (R, Y ).
1.5. JOINT, MARGINAL AND CONDITIONAL PROBABILITY 35

latitude = 45 latitude = 45 latitude = 45

longitude = −40 longitude = −30 longitude = −20

20
60

20
40

10
10
20

5
5
0

0
4 6 8 4 6 8 4 6 8

temperature temperature temperature

n = 213 n = 105 n = 112

latitude = 35 latitude = 35 latitude = 35

longitude = −40 longitude = −30 longitude = −20
12
20

15
15

0 2 4 6 8

10
10

5
5
0

4 6 8 4 6 8 4 6 8

temperature temperature temperature

n = 37 n = 24 n = 44

latitude = 25 latitude = 25 latitude = 25

longitude = −40 longitude = −30 longitude = −20
15

15
20

10

10
5 10

5
0

4 6 8 4 6 8 4 6 8

temperature temperature temperature

n = 47 n = 35 n = 27

Figure 1.13: Water temperatures (◦ C) at 1000m depth, latitude 25, 35, 45

degrees North, longitude 20, 30, 40 degrees West, with standard deviations
36 CHAPTER 1. PROBABILITY

For Against
Democrat 48% 12% 60%
Republican 14% 26% 40%
62% 38%

Table 1.1: Party Affiliation and Referendum Support

These calculations illustrate a general principle: To get a marginal prob-

ability for one variable, add the joint probabilities for all values of the other
variable. The general formulae for working simultaneously with two discrete
random variables X and Y are
fX,Y (x, y) = fX (x) · fY | X (y | x) = fY (y) · fX | Y (x | y) (1.12)
X X
fX (x) = fX,Y (x, y) fY (y) = fX,Y (x, y)
y x

Sometimes we know joint probabilities and need to find marginals and condi-
tionals; sometimes it’s the other way around. And sometimes we know fX and
fY | X and need to find fY or fX | Y . The following story is an example of the
latter. It is a common problem in drug testing, disease screening, polygraph
testing, and many other fields.
The participants in an athletic competition are to be randomly tested for
steroid use. The test is 90% accurate in the following sense: for athletes who use
steroids, the test has a 90% chance of returning a positive result; for non-users,
the test has a 10% chance of returning a positive result. Suppose that only 30%
of athletes use steroids. An athlete is randomly selected. Her test returns a
positive result. What is the probability that she is a steroid user?
This is a problem of two random variables, U , the steroid use of the athlete
and T , the test result of the athlete. Let U = 1 if the athlete uses steroids;
U = 0 if not. Let T = 1 if the test result is positive; T = 0 if not. We want
fU | T (1 | 1). We can calculate fU | T if we know fU,T ; and we can calculate fU,T
because we know fU and fT | U . Pictorially,
fU , fT | U −→ fU,T −→ fU | T
The calculations are
fU,T (0, 0) = (.7)(.9) = .63 fU,T (0, 1) = (.7)(.1) = .07
fU,T (1, 0) = (.3)(.1) = .03 fU,T (1, 1) = (.3)(.9) = .27

so

fT (0) = .63 + .03 = .66 fT (1) = .07 + .27 = .34

and finally

fU | T (1 | 1) = fU,T (1, 1)/fT (1) = .27/.34 ≈ .80.

1.5. JOINT, MARGINAL AND CONDITIONAL PROBABILITY 37

T =0 T =1
U =0 .63 .07 .70
U =1 .03 .27 .30
.66 .34

Table 1.2: Steroid Use and Test Results

In other words, even though the test is 90% accurate, the athlete has only an
80% chance of using steroids. If that doesn’t seem intuitively reasonable, think
of a large number of athletes, say 100. About 30 will be steroid users of whom
about 27 will test positive. About 70 will be non-users of whom about 7 will
test positive. So there will be about 34 athletes who test positive, of whom
about 27, or 80% will be users.
Table 1.2 is another representation of the same problem. It is important to
become familiar with the concepts and notation in terms of marginal, condi-
tional and joint distributions, and not to rely too heavily on the tabular repre-
sentation because in more complicated problems there is no convenient tabular
representation.
Example 1.6 is a further illustration of joint, conditional, and marginal dis-
tributions.

Example 1.6 (Seedlings)

Example 1.4 introduced an observational experiment to learn about the rate of
seedling production and survival at the Coweeta Long Term Ecological Research
station in western North Carolina. For a particular quadrat in a particular year, let
N be the number of new seedlings that emerge. Suppose that N ∼ Poi(λ) for
some λ > 0. Each seedling either dies over the winter or survives to become an old
seedling the next year. Let θ be the probability of survival and X be the number of
seedlings that survive. Suppose that the survival of any one seedling is not affected
by the survival of any other seedling. Then X ∼ Bin(N, θ). Figure 1.14 shows
the possible values of the pair (N, X). The probabilities associated with each of
the points in Figure 1.14 are denoted fN,X where, as usual, the subscript indicates
which variables we’re talking about. For example, fN,X (3, 2) is the probability that
N = 3 and X = 2.
The next step is to figure out what the joint probabilities are. Consider, for
example, the event N = 3. That event can be partitioned into the four smaller
events (N = 3, X = 0), (N = 3, X = 1), (N = 3, X = 2), and (N = 3, X = 3).
So
fN (3) = fN,X (3, 0) + fN,X (3, 1) + fN,X (3, 2) + fN,X (3, 3)
The Poisson model for N says fN (3) = P[N = 3] = e−λ λ3 /6. But how is the
total e−λ λ3 /6 divided into the four parts? That’s where the Binomial model for X
comes in. The division is made according to the Binomial probabilities
       
3 3 3 2 3 3
(1 − θ)3 θ(1 − θ)2 θ (1 − θ) θ
0 1 2 3
38 CHAPTER 1. PROBABILITY

...
6
4
X

...
2
0

0 1 2 3 4 5 6 7

Figure 1.14: Permissible values of N and X, the number of new seedlings and
the number that survive.

The e−λ λ3 /6 is a marginal probability like the 60% in the affiliation/support prob-
lem. The binomial probabilities above are conditional probabilities like the 80% and
20%; they are conditional on N = 3. The notation is fX|N (2 | 3) or P[X = 2 | N =
3]. The joint probabilities are

e−λ λ3 e−λ λ3
fN,X (3, 0) = (1 − θ)3 fN,X (3, 1) = 3θ(1 − θ)2
6 6
e−λ λ3 2 e−λ λ3 3
fN,X (3, 2) = 3θ (1 − θ) fN,X (3, 3) = θ
6 6

In general,

 
e−λ λn n x
fN,X (n, x) = fN (n)fX | N (x | n) = θ (1 − θ)n−x
n! x

An ecologist might be interested in fX , the pdf for the number of seedlings that
will be recruited into the population in a particular year. For a particular number
x, fX (x) is like looking in Figure 1.14 along the horizontal line corresponding to
X = x. To get fX (x) ≡ P[X = x], we must add up all the probabilities on that
1.5. JOINT, MARGINAL AND CONDITIONAL PROBABILITY 39

line.
∞  
X X e−λ λn n x
fX (x) = fN,X (n, x) = θ (1 − θ)n−x
n n=x
n! x
∞
X e−λ(1−θ) (λ(1 − θ))n−x e−λθ (λθ)x
=
n=x
(n − x)! x!
∞
e−λθ (λθ)x X e−λ(1−θ) (λ(1 − θ))z
=
x! z=0
z!
e−λθ (λθ)x
=
x!
P
The last equality follows since z · · · = 1 because it is the sum of probabilities
from the Poi(λ(1 − θ)) distribution. The final result is recognized as a probability
from the Poi(λ∗ ) distribution where λ∗ = λθ. So X ∼ Poi(λ∗ ).
In the derivation we used the substitution z = n − x. The trick is worth remem-
bering. If N ∼ Poi(λ), and if the individual events that N counts are randomly
divided into two types X and Z (e.g., survivors and non-survivors) according to
a binomial distributuion, then (1) X ∼ Poi(λθ) and Z ∼ Poi(λ(1 − θ)) and (2)
X ⊥ Z. Another way to put it: if (1) X ∼ Poi(λX ) and Z ∼ Poi(λZ ) and (2)
X ⊥ Z then N ≡ X + Z ∼ Poi(λX + λZ ).

For continuous random variables, conditional and joint densities are written
pX | Y (x | y) and pX,Y (x, y) respectively and, analgously to Equation 1.12 we
have

pX,Y (x, y) = pX (x)pY | X (y | x) = pY (y)pX | Y (x | y) (1.13)

Z ∞ Z ∞
pX (x) = pX,Y (x, y) dy pY (y) = pX,Y (x, y) dx
−∞ −∞

The logic is the same as for discrete random variables. In order for (X = x, Y =
y) to occur we need either of the following.

1. First X = x occurs then Y = y occurs. The “probability” of that hap-

pening is just pX (x) × pY | X (y | x), the “probability” that X = x occurs
times the “probability” that Y = y occurs under the condition that X = x
has already occured. “Probability” is in quotes because, for continuous
random variables, the probability is 0. But “probability” is a useful way
to think intuitively.

2. First Y = y occurs then X = x occurs. The reasoning is similar to that

in item 1.

Just as for single random variables, probabilities are integrals

R of the density.
If
R A is a region in the (x, y) plane, P[(X, Y ) ∈ A] = A p(x, y) dx dy, where
A
. . . indicates a double integral over the region A.
40 CHAPTER 1. PROBABILITY

Just as for discrete random variables, the unconditional density of a random

variable is called its marginal density; pX and pY are marginal densities. Let
B ⊂ R be a set. Since a density is “the function thatR must be integrated to
calculate a probability”, on one hand, P[X ∈ B] = B pX (x) dx. On the other
hand,
Z Z 
P[X ∈ B] = P[(X, Y ) ∈ B × R] = pX,Y (x, y) dy dx
B R
R
which implies pX (x) = R pX,Y (x, y) dy.
An example will help illustrate. A customer calls the computer Help Line.
Let X be the amount of time he spends on hold and Y be the total duration
of the call. The amount of time a consultant spends with him after his call is
answered is W = Y − X. Suppose the joint density is pX,Y (x, y) = e−y in the
region 0 < x < y < ∞.
1. What is the marginal density of X?
Z Z ∞ ∞

p(x) = p(x, y) dy = e−y dy = −e−y  = e−x
x x

2. What is the marginal density of Y ?

Z Z y
p(y) = p(x, y) dx = e−y dx = ye−y
0

3. What is the conditional density of X given Y ?

p(x, y)
p(x | y) = = y −1
p(y)

4. What is the conditional density of Y given X?

p(x, y)
p(y | x) = = ex−y
p(x)

5. What is the marginal density of W ?

Z ∞ Z x+w
d d
p(w) = P[W ≤ w] = e−y dydx
dw dw 0 x
Z ∞ x+w 
d −y 
= −e  dx
dw 0 x
Z ∞
d
= e−x (1 − e−w ) dx = e−w
dw 0

Figure 1.15 illustrates the Help Line calculations. For questions 1 and 2, the
answer comes from using Equations 1.13. The only part deserving comment is
1.5. JOINT, MARGINAL AND CONDITIONAL PROBABILITY 41

the limits of integration. In question 1, for example, for any particular value
X = x, Y ranges from x to ∞, as can be seen from panel (a) of the figure. That’s
where the limits of integration come from. In question 2, for any particular y,
X ∈ (0, y), which are the limits of integration. Panel (d) shows the conditional
density of X given Y for three different values of Y . We see that the density of
X is uniform on the interval (0, y). See Section 5.4 for discussion of this density.
Panel (d) shows the conditional density of Y given X for three different values
of X. It shows, first, that Y > X and second, that the density of Y decays
exponentially. See Section 1.3.3 for discussion of this density. Panel (f) shows
the region of integration for question 5. Take the time to understand the method
being used to answer question 5.
When dealing with a random variable X, sometimes its pdf is given to us
and we can calculate its expectation:
Z
E(X) = xp(x) dx.

(The integral is replaced by a sum if X is discrete.) Other times X arises more

naturally as part of a pair (X, Y ) and its expectation is
Z
E(X) = xp(x, y) dxdy.

The two formulae are, of course, equivalent. But when X does arise as part of
a pair, there is still another way to view p(x) and E(X):
Z



pX (x) = pX | Y (x | y) pY (y) dy = E pX | Y (x | y) (1.14)
Z Z 
E(X) = xpX | Y (x | y) pY (y) dxdy = E (E(X | Y )) . (1.15)

The notation deserves some explanation. For any numberRx, pX | Y (x | y) is a

function of y, say g(y). The middle term in Equation 1.14 is g(y)p(y)dy, which
equals E(g(y)), which is the right hand term. Similarly,
R E(X | Y ) is a function
of Y , say h(Y ). The middle term in Equation 1.15 is h(y)p(y)dy, which equals
E(h(y)), which is the right hand term.

Example 1.7 (Seedlings, continued)

Examples 1.4 and 1.6 discussed (N, X), the number of new seedlings in a for-
est quadrat and the number of those that survived over the winter. Example 1.6
suggested the statistical model

N ∼ Poi(λ) and X | N ∼ Bin(N, θ).

Equation 1.15 shows that E(X) can be computed as

E(X) = E (E(X | N )) = E(N θ) = θE(N ) = θλ.

42 CHAPTER 1. PROBABILITY

(a) (b)

0.8
3

p(x)
2
y

0.4
1

0.0
0

0 1 2 3 4 0 2 4 6 8 10

x x

(c) (d)

2.0
y=1
y=2
p(x|y)
0.2

y=3
p(y)

1.0
0.0

0.0

0 2 4 6 8 10 0.0 1.0 2.0 3.0

y x

(e) (f)

x=1
6
0.8

x=2
p(y|x)

y−x=w
4

x=3
y
0.4

y−x=0
0.0

0 2 4 6 8 10 0 1 2 3 4

y x

Figure 1.15: (a): the region of R2 where (X, Y ) live; (b): the marginal density
of X; (c): the marginal density of Y ; (d): the conditional density of X given Y
for three values of Y ; (e): the conditional density of Y given X for three values
of X; (f ): the region W ≤ w
1.5. JOINT, MARGINAL AND CONDITIONAL PROBABILITY 43

Example 1.8 (Craps, continued)

Examples 1.1, 1.2 and 1.3 introduced the game of craps. Example 1.8 calculates
the chance of winning.
(
0 if shooter wins
Let X =
1 if shooter loses

X has a Bernoulli distribution. We are trying to find

P[shooter wins] = pX (1) = E(X).

(Make sure you see why pX (1) = E(X).) Let Y be the outcome of the Come-out
roll. Equation 1.15 says

E(X) = E (E(X | Y ))
= E(X | Y = 2) P[Y = 2] + E(X | Y = 3) P[Y = 3]
+ E(X | Y = 4) P[Y = 4] + E(X | Y = 5) P[Y = 5]
+ E(X | Y = 6) P[Y = 6] + E(X | Y = 7) P[Y = 7]
+ E(X | Y = 8) P[Y = 8] + E(X | Y = 9) P[Y = 9]
+ E(X | Y = 10) P[Y = 10] + E(X | Y = 11) P[Y = 11]
+ E(X | Y = 12) P[Y = 12]
1 2 3
= 0× +0× + E(X | Y = 4)
36 36 36
4 5
+ E(X | Y = 5) + E(X | Y = 6)
36 36
6 5 4
+1× + E(X | Y = 8) + E(X | Y = 9)
36 36 36
3 2 1
+ E(X | Y = 10) + 1 × +0× .
36 36 36
So it only remains to find E(X | Y = y) for y = 4, 5, 6, 8, 9, 10. The calculations
are all similar. We will do one of them to illustrate. Let w = E(X | Y = 5) and let
z denote the next roll of the dice. Once 5 has been established as the point, then a
roll of the dice has three possible outcomes: win (if z = 5), lose (if z = 7), or roll
again (if z is anything else). Therefore

w = 1 × 4/36 + 0 × 6/36 + w × 26/36

(10/36)w = 4/36
w = 4/10.

After similar calculations for the other possible points we find

3 3 4 4 5 5 6 5 5 4 4 3 3 2
E(X) = + + + + + + + ≈ .493.
9 36 10 36 11 36 36 11 36 10 36 9 36 36
Craps is a very fair game; the house has only a slight edge.
44 CHAPTER 1. PROBABILITY

1.6 Association, Dependence, Independence

It is often useful to describe or measure the degree of association between two
random variables X and Y . The R dataset iris provides a good example. It
contains the lengths and widths of setals and petals of 150 iris plants. The first
several lines of iris are
Sepal.Length Sepal.Width Petal.Length Petal.Width Species
1 5.1 3.5 1.4 0.2 setosa
2 4.9 3.0 1.4 0.2 setosa
3 4.7 3.2 1.3 0.2 setosa
Figure 1.16 shows each variable plotted against every other variable. It is evident
from the plot that petal length and petal width are very closely associated with
each other, while the relationship between sepal length and sepal width is much
weaker. Statisticians need a way to quantify the strength of such relationships.

Figure 1.16 was produced by the following line of R code.

pairs ( iris[,1:4] )

• pairs() produces a pairs plot, a matrix of scatterplots of each pair of

variables. The names of the variables are shown along the main diagonal
of the matrix. The (i, j)th plot in the matrix is a plot of variable i versus
variable j. For example, the upper right plot has sepal length on the
vertical axis and petal width on the horizontal axis.

By far the most common measures of association are covariance and corre-
lation.
Definition 1.6. The covariance of X and Y is

Cov(X, Y ) ≡ E((X − µX )(Y − µY ))

In R, cov measures the covariance in a sample. Thus, cov(iris[,1:4])

produces the matrix
Sepal.Length Sepal.Width Petal.Length Petal.Width
Sepal.Length 0.68569351 -0.04243400 1.2743154 0.5162707
Sepal.Width -0.04243400 0.18997942 -0.3296564 -0.1216394
Petal.Length 1.27431544 -0.32965638 3.1162779 1.2956094
Petal.Width 0.51627069 -0.12163937 1.2956094 0.5810063
in which the diagonal entries are variances and the off-diagonal entries are co-
variances.
The measurements in iris are in centimeters. To change to millimeters we
would multiply each measurement by 10. Here’s how that affects the covariances.
1.6. ASSOCIATION, DEPENDENCE, INDEPENDENCE 45

2.0 3.0 4.0 0.5 1.5 2.5

7.5
6.5
Sepal.Length

5.5
4.5
2.0 2.5 3.0 3.5 4.0

Sepal.Width

7
6
5

Petal.Length
4
3
2
1
0.5 1.0 1.5 2.0 2.5

Petal.Width

4.5 6.0 7.5 1 3 5 7

Figure 1.16: Lengths and widths of sepals and petals of 150 iris plants
46 CHAPTER 1. PROBABILITY

> cov ( 10*iris[,1:4] )

Sepal.Length Sepal.Width Petal.Length Petal.Width
Sepal.Length 68.569351 -4.243400 127.43154 51.62707
Sepal.Width -4.243400 18.997942 -32.96564 -12.16394
Petal.Length 127.431544 -32.965638 311.62779 129.56094
Petal.Width 51.627069 -12.163937 129.56094 58.10063

Each covariance has been multiplied by 100 because each variable has been
multiplied by 10. In fact, this rescaling is a special case of the following theorem.

Theorem 1.5. Let X and Y be random variables. Then Cov(aX + b, cY + d) =

ac Cov(X, Y ).

Proof.

Cov(aX + b, cY + d) = E ((aX + b − (aµX + b))(cY + d − (cµY + d)))

= E(ac(X − µX )(Y − µY )) = ac Cov(X, Y )

Theorem 1.5 shows that Cov(X, Y ) depends on the scales in which X and
Y are measured. A scale-free measure of association would also be useful. Cor-
relation is the most common such measure.

Definition 1.7. The correlation between X and Y is

Cov(X, Y )
Cor(X, Y ) ≡
SD(X) SD(Y )

cor measures correlation. The correlations in iris are

> cor(iris[,1:4])
Sepal.Length Sepal.Width Petal.Length Petal.Width
Sepal.Length 1.0000000 -0.1175698 0.8717538 0.8179411
Sepal.Width -0.1175698 1.0000000 -0.4284401 -0.3661259
Petal.Length 0.8717538 -0.4284401 1.0000000 0.9628654
Petal.Width 0.8179411 -0.3661259 0.9628654 1.0000000

which confirms the visually impression that sepal length, petal length, and petal
width are highly associated with each other, but are only loosely associated with
sepal width.
Theorem 1.6 tells us that correlation is unaffected by linear changes in mea-
surement scale.

Theorem 1.6. Let X and Y be random variables. Then Cor(aX + b, cY + d) =

Cor(X, Y ).

Proof. See Exercise 38.

1.6. ASSOCIATION, DEPENDENCE, INDEPENDENCE 47

Correlation doesn’t measure all types of association; it only measures clus-

tering around a straight line. The first two columns of Figure 1.17 show data
sets that cluster around a line, but with some scatter above and below the line.
These data sets are all well described by their correlations, which measure the
extent of the clustering; the higher the correlation, the tighter the points clus-
ter around the line and the less they scatter. Negative values of the correlation
correspond to lines with negative slopes. The last column of the figure shows
some other situations. The first panel of the last column is best described as
having two isolated clusters of points. Despite the correlation of .96, the panel
does not look at all like the last panel of the second column. The second and
third panels of the last column show data sets that follow some nonlinear pat-
tern of association. Again, their correlations are misleading. Finally, the last
panel of the last column shows a data set in which most of the points are tightly
clustered around a line but in which there are two outliers. The last column
demonstrates that correlations are not good descriptors of nonlinear data sets,
or data sets with outliers.
Correlation measures linear association between random variables. But
sometimes we want to say whether two random variables have any association
at all, not just linear.
Definition 1.8. Two two random variables, X and Y , are said to be indepen-
dent if p(X | Y ) = p(X), for all values of Y . If X and Y are not independent
then they are said to be dependent.
If X and Y are independent then it is also true that p(Y | X) = p(Y ). The
interpretation is that knowing one of the random variables does not change the
probability distribution of the other. If X and Y are independent (dependent)
we write X ⊥ Y (X 6⊥ Y ). If X and Y are independent then Cov(X, Y ) =
Cor(X, Y ) = 0. The converse is not true. Also, if X ⊥ Y then p(x, y) =
p(x)p(y). This last equality is usually taken to be the definition of independence.
Do not confuse independent with mutually exclusive. Let X denote the
outcome of a die roll and let A = 1 if X ∈ {1, 2, 3} and A = 0 if X ∈ {4, 5, 6}. A
is called an indicator variable because it indicates the occurence of a particular
event. There is a special notation for indicator variables:
A = 1{1,2,3} (X).
1{1,2,3} is an indicator function. 1{1,2,3} (X) is either 1 or 0 according to whether
X is in the subscript. Let B = 1{4,5,6} (X), C = 1{1,3,5} (X), D = 1{2,4,6} (X)
and E = 1{1,2,3,4} (X). A and B are dependent because P[A] = .5 but P[A | B] =
0. D and E are independent because P[D] = P[D | E] = .5. You can also check
that P[E] = P[E | D] = 2/3. Do not confuse dependence with causality. A and
B are dependent, but neither causes the other.
For an example, recall the Help Line story on page 40. X and Y were
the amount of time on hold and the total length of the call, respectively. The
difference was W = Y − X. We found p(x | y) = y −1 . Because p(x | y) depends
on y, X 6⊥ Y . Similarly, p(y | x) depends on x. Does that make sense? Would
knowing something about X tell us anything about Y ?
48 CHAPTER 1. PROBABILITY

cor = −0.9 cor = 0.3 cor = 0.96

cor = −0.5 cor = 0.5 cor = 0.67

cor = −0.2 cor = 0.8 cor = −0.37

cor = 0 cor = 0.95 cor = 0.62

Figure 1.17: correlations

1.7. SIMULATION 49

What about X and W ? Are they independent?

d d
p(w | x) = P[W ≤ w | X = x] = P[Y ≤ w + x | X = x] = e−w
dw dw
which does not depend on x. Therefore X ⊥ W . Does that make sense? Would
knowing something about X tell us anything about W ?

Example 1.9 (Seedlings, continued)

Examples 1.4, 1.6, and 1.7 were about new seedlings in forest quadrats. Suppose
that ecologists observe the number of new seedlings in a quadrat for k successive
years; call the observations N1 , . . . , Nk . If the seedling arrival rate is the same
every year, then we could adopt the model Ni ∼ Poi(λ). I.e., λ is the same for
every year. If λ is known, or if we condition on λ, then the number of new seedlings
in one year tells us nothing about the number of new seedlings in another year, we
would model the Ni ’s as being independent, and their joint density, conditional on
λ, would be
k k P
Y Y e−λ λni e−kλ λ ni
p(n1 , . . . , nk | λ) = p(ni | λ) = = Q
i=1 i=1
n i ! ni

But if λ is unknown then we might treat it like a random variable. It would be a

random variable if, for instance, different locations in the forest had different λ’s
we chose a location at random or if we thought of Nature as randomly choosing
λ for our particular location. In that case the data from early years, N 1 , . . . , Nm ,
say, yield information about λ and therefore about likely values of N m+1 , . . . , Nk ,
so the Ni ’s are dependent. In fact,
Z Z −kλ P ni
e λ
p(n1 , . . . , nk ) = p(n1 , . . . , nk | λ)p(λ) dλ = Q p(λ) dλ
ni
So, whether the Ni ’s are independent is not a question with a single right answer.
Instead, it depends on our perspective. But in either case, we would say the N i ’s
are conditionally independent given λ.

1.7 Simulation
We have already seen, in Example 1.2, an example of computer simulation
to estimate a probability. More broadly, simulation can be helpful in several
types of problems: calculating probabilities, assessing statistical procedures, and
evaluating integrals. These are explained and exemplified in the next several
subsections.

1.7.1 Calculating Probabilities

Probabilities can often be estimated by computer simulation. Simulations are
especially useful for events so complicated that their probabilities cannot be eas-
ily calculated by hand, but composed of smaller events that are easily mimicked
50 CHAPTER 1. PROBABILITY

on computer. For instance, in Example 1.2 we wanted to know the probability

that the shooter in a craps game rolls either 7 or 11 on the Come Out roll.
Although it’s easy enough to calculate this probability exactly, we did it by
simulation in the Example.
Expected values can also be estimated by simulations. Let Y be a random
variable and suppose we want to estimate the expected value of some function
µg ≡ E(g(Y )). We can write a computer program to simulate Y many times.
To keep track of the simulations we use the notation y (j) for the j’th simulated
value of Y . Let n be the number of simulations. Then
X
µ̂g = n−1 g(y (i) )

is a sensible estimate of µg . In fact, the Law of Large Numbers tells us that

X
lim n−1 g(y (i) ) = µg .
n→∞

So as we do a larger and larger simulation we get a more and more accurate

estimate of µg .
Probabilities can be computed as special cases of expectations. Suppose
we want to calculate P[Y ∈ S] for some set S. Define X ≡ 1S (Y ). Then
P[Y ∈ S] = E(X) and is sensibly estimated by

number of occurences of E X
= n−1 x(i) .
number of trials
Example 1.10 illustrates with the game of Craps.

Example 1.10 (Craps, continued)

Example 1.8 calculated the chance of winning the game of Craps. Here is the R
code to calculate the same probability by simulation.

makepoint <- function ( point ) {

determined <- F
while ( !determined ) { # roll until outcome is determined
roll <- sum ( sample ( 6, 2, replace=T ) )
if ( roll == point ) {
made <- T
determined <- T
} else if ( roll == 7 ) {
made <- F
determined <- T
}
} # end while
return ( made )
} # end makepoint

sim.craps <- function () {

1.7. SIMULATION 51

roll <- sum ( sample ( 6, 2, replace=T ) )

if ( roll==7 | roll==11 )
win <- T
else if ( roll==2 | roll==3 | roll==12 )
win <- F
else
win <- makepoint ( roll )

return ( win )
}

n.sim <- 1000

wins <- 0
for ( i in 1:n.sim )
wins <- wins + sim.craps()
print ( wins/n.sim )

• “!” is R’s symbol for not. If determined is T then !determined is F.

• while(!determined) begins a loop. The loop will repeat as many times

as necessary until !determined is F.

Try the example code a few times. See whether you get about 49% as Exam-
ple 1.8 suggests.
Along with the estimate itself, it is useful to estimate the accuracy of µ̂g as an
estimate of µg . If the simulations are independent and there are many P of them
then Var(µ̂g ) = n−1 Var(g(Y )), Var(g(Y )) can be p estimated by n−1 ((g(y) −
P
µ̂g )2 ) and SD(g(Y )) can be estimated by n−1/2 ((g(y) − µ̂g )2 ). Because
−1/2
SD’s decrease in proportion to n , it takes a 100 fold increase in n to get,
for example, a 10 fold increase in accuracy.
Similar reasoning applies to probabilities, but when we are simulating the
occurence or nonoccurence of an event, then the simulations are Bernoulli trials,
so we have a more explicit formula for the variance and SD.

Example 1.11 (Craps, continued)

How accurate is the simulation in Example 1.10?
The simulation keeps track of X, the number of successes in n.sim trials. Let
θ be the true probability of success. (Example 1.8 found θ ≈ .49, but in most
practical applications we won’t know θ.)

X ∼ Bin(n.sim, θ)
Var(X) ≈ n.sim(θ)(1 − θ)
p
SD(X/n.sim) ≈ (θ)(1 − θ)/n.sim
52 CHAPTER 1. PROBABILITY

and, by the Central Limit Theorem if n.sim is large,

θ̂ = X/n.sim ∼ N(θ, (θ(1 − θ)/n.sim)1/2 )

What does this mean in practical terms? How accurate is the simulation when
n.sim = 50, or 200, or 1000, say? To illustrate we did 1000 simulations with
n.sim = 50, then another 1000 with n.sim = 200, and then another 1000 with
n.sim = 1000.
The results are shown as a boxplot in Figure 1.18. In Figure 1.18 there are three
boxes, each with whiskers extending vertically. The box for n.sim = 50 shows that
the median of the 1000 θ̂’s was just about .50 (the horizontal line through the box),
that 50% of the θ̂’s fell between about .45 and .55 (the upper and lower ends of the
box), and that almost all of the θ̂’s fell between about .30 and .68 (the extent of
the whiskers). In comparison, the 1000 θ̂’s for n.sim = 200 are spread out about
half as much, and the 1000 θ̂’s for n.sim = 1000 are spread out about half as
much again. The factor of about a half comes from the n.sim.5 in the formula for
SD(θ̂). When n.sim increases by a factor of about 4, the SD decreases by a factor
of about 2. See the notes for Figure 1.18 for a further description of boxplots.
0.7
0.5
0.3

50 200 1000

n.sim

Figure 1.18: 1000 simulations of θ̂ for n.sim = 50, 200, 1000

Here is the R code for the simulations and Figure 1.18.

N <- 1000
1.7. SIMULATION 53

n.sim <- c(50, 200, 1000)

theta.hat <- matrix ( NA, N, length(n.sim) )

for ( i in seq(along=n.sim) ) {
for ( j in 1:N ) {
wins <- 0
for ( k in 1:n.sim[i] )
wins <- wins + sim.craps()
theta.hat[j,i] <- wins / n.sim[i]
}
}

boxplot ( theta.hat ~ col(theta.hat), names=n.sim,

xlab="n.sim" )

• matrix forms a matrix. The form is matrix(x,nrows,ncols) where x are

the entries in the matrix and nrows and ncols are the numbers of rows
and columns.
• seq(along=n.sim) is the same as 1:length(n.sim) except that it be-
haves more sensibly in case length(n.sim) is 0.
• A boxplot is one way to display a data set. It produces a rectangle, or
box, with a line through the middle. The rectangle contains the central
50% of the data. The line indicates the median of the data. Extending
vertically above and below the box are dashed lines called whiskers. The
whiskers contain most of the outer 50% of the data. A few extreme data
points are plotted singly. See Example 2.3 for another use of boxplots and
a fuller explanation.

1.7.2 Evaluating Statistical Procedures

Simulation can sometimes be useful in deciding whether a particular experiment
is worthwhile or in choosing among several possible experiments or statistical
procedures. For a fictitious example, consider ABC College, where, of the 10,000
students, 30% of the students are members of a sorority or fraternity (greeks)
and 70% are independents. There is an upcoming election for Head of the
student governance organization. Two candidates are running, D and E. Let
θG = proportion of greeks supporting D
θI = proportion of independents supporting D
θ = proportion of all students supporting D
A poll is commisioned to estimate θ and it is agreed that the pollster will sample
100 students. Three different procedures are proposed.
54 CHAPTER 1. PROBABILITY

1. Randomly sample 100 students. Estimate

θ̂1 = proportion of polled students who favor D

2. Randomly sample 100 students. Estimate

θ̂G = proportion of polled greeks supporting D

θ̂I = proportion of polled independents supporting D
θ̂2 = .3θ̂G + .7θ̂I

3. Randomly sample 30 greeks and 70 independents. Estimate

θ̂G = proportion of polled greeks supporting D

θ̂I = proportion of polled independents supporting D
θ̂3 = .3θ̂G + .7θ̂I

Which procedure is best? One way to answer the question is by exact calcula-
tion, but another way is by simulation. In the simulation we try each procedure
many times to see how accurate it is, on average. We must choose some “true”
values of θG , θI and θ under which to do the simulation. Here is some R code
for the simulation.
# choose "true" theta.g and theta.i
theta.g <- .8
theta.i <- .4
prop.g <- .3
prop.i <- 1 - prop.g
theta <- prop.g * theta.g + prop.i * theta.i

sampsize <- 100

n.times <- 1000 # should be enough

theta.hat <- matrix ( NA, n.times, 3 )
for ( i in 1:n.times ) {
theta.hat[i,1] <- sim.1()
theta.hat[i,2] <- sim.2()
theta.hat[i,3] <- sim.3()
}

print ( apply(theta.hat,2,mean) )
boxplot ( theta.hat ~ col(theta.hat) )

sim.1 <- function() {

x <- rbinom(1,sampsize,theta)
return ( x / sampsize )
1.7. SIMULATION 55

sim.2 <- function() {

n.g <- rbinom ( 1, sampsize, prop.g )
n.i <- sampsize - n.g
x.g <- rbinom ( 1, n.g, theta.g )
x.i <- rbinom ( 1, n.i, theta.i )
t.hat.g <- x.g / n.g
t.hat.i <- x.i / n.i
return ( prop.g * t.hat.g + (1-prop.g) * t.hat.i )
}

sim.3 <- function() {

n.g <- sampsize * prop.g
n.i <- sampsize * prop.i
x.g <- rbinom ( 1, n.g, theta.g )
x.i <- rbinom ( 1, n.i, theta.i )
t.hat.g <- x.g / n.g
t.hat.i <- x.i / n.i
return ( prop.g * t.hat.g + (1-prop.g) * t.hat.i )
}

• apply applies a function to a matrix. In the code above, apply(...)

applies the function mean to dimension 2 of the matrix theta.hat. That
is, it returns the mean of of each column of theta.hat

The boxplot, shown in Figure 1.19 shows little practical difference between
the three procedures.
The next example shows how simulation was used to evaluate whether an
experiment was worth carrying out.

Example 1.12 (FACE)

The amount of carbon dioxide, or CO2 , in the Earth’s atmosphere has been steadily
increasing over the last century or so. You can see the increase yourself in the
co2 data set that comes with R. Typing ts.plot(co2) makes a time series plot,
reproduced here as Figure 1.20. Typing help(co2) gives a brief explanation. The
data are the concentrations of CO 2 in the atmosphere measured at Mauna Loa each
month from 1959 to 1997. The plot shows a steadily increasing trend imposed on a
regular annual cycle. The primary reason for the increase is burning of fossil fuels.
CO2 is a greenhouse gas that traps heat in the atmosphere instead of letting it
radiate out, so an increase in atmospheric CO2 will eventually result in an increase
in the Earth’s temperature. But what is harder to predict is the effect on the Earth’s
plants. Carbon is a nutrient needed by plants. It is possible that an increase in CO 2
will cause an increase in plant growth which in turn will partly absorb some of the
extra carbon.
56 CHAPTER 1. PROBABILITY

0.35 0.50 0.65

1 2 3

Figure 1.19: 1000 simulations of θ̂ under three possible procedures for conduct-
ing a poll
350
co2

320

1960 1970 1980 1990

Time

Figure 1.20: Monthly concentrations of CO2 at Mauna Loa

1.7. SIMULATION 57

To learn about plant growth under elevated CO 2 , ecologists began by conduct-

ing experiments in greenhouses. In a greenhouse, two sets of plants could be grown
under conditions that are identical except for the amount of CO 2 in the atmosphere.
But the controlled environment of a greenhouse is quite unlike the uncontrolled nat-
ural environment, so, to gain verisimilitude, experiments soon moved to open-top
chambers. An open-top chamber is a space, typically a few meters in diameter,
enclosed by a solid, usually plastic, wall and open at the top. CO 2 can be added to
the air inside the chamber. Because the chamber is mostly enclosed, not much CO 2
will escape, and more can be added as needed. Some plants can be grown in cham-
bers with excess CO2 others in chambers without and their growth compared. But
as with greenhouses, open-top chambers are not completely satisfactory. Ecologists
wanted to conduct experiments under even more natural conditions.
To that end, in the late 1980’s the Office of Biological and Environmental Re-
search in the U.S. Department of Energy (DOE) began supporting research using a
technology called FACE, or Free Air CO2 Enrichment developed at the Brookhaven
National Laboratory. As the lab’s webpage www.face.bnl.gov/face1.htm ex-
plains
“FACE provides a technology by which the microclimate around
growing plants may be modified to simulate climate change conditions.
Typically CO2-enriched air is released from a circle of vertical pipes into
plots up to 30m in diameter, and as tall as 20 m.
“Fast feedback control and pre-dilution of CO2 provide stable, ele-
vated [CO2] simulating climate change conditions.
“No containment is required with FACE equipment and there is no
significant change in natural air-flow. Large FACE plots reduce effects
of plot edge and capture fully-functioning, integrated ecosystem-scale
processes. FACE Field data represent plant and ecosystems responses
to concentrations of atmospheric CO2 expected in the mid-twenty-first
century.”
See the website for pictures and more information. In a FACE experiment, CO 2 is
released into some treatment plots. The level of CO2 inside the plot is continually
monitored. More CO2 is released as needed to keep the amount of CO2 in the
atmosphere at some prespecified level, typically the level that is expected in the
mid-21st century. Other plots are reserved as control plots. Plant growth in the
treatment plots is compared to that in the control plots.
Because a FACE site is not enclosed, CO2 continually drifts out of the site and
needs to be replenished. Keeping enough CO2 in the air is very costly and is, in
fact, the major expense in conducting FACE experiments.
The first several FACE sites were in Arizona (sorghum, wheat, cotton), Switzer-
land (rye grass, clover) and California (native chaparral). All of these contained
low-growing plants. By the early 1990’s, ecologists wanted to conduct a FACE ex-
periment in a forest, and such an experiment was proposed by investigators at Duke
University, to take place in Duke Forest. But before the experiment could be funded
the investigators had to convince the Department of Energy (DOE) that it would be
worthwhile. In particular, they wanted to demonstrate that the experiment would
58 CHAPTER 1. PROBABILITY

have a good chance of uncovering whatever growth differences would exist between
treatment and control. The demonstration was carried out by computer simulation.
The code for that demonstration, slightly edited for clarity, is given at the end of
this Example and explained below.

1. The experiment would consist of 6 sites, divided into 3 pairs. One site in
each pair would receive the CO2 treatment; the other would be a control.
The experiment was planned to run for 10 years. Investigators had identified
16 potential sites in Duke Forest. The above ground biomass of those sites,
measured before the experiment began, is given in the line b.mass <- c (
... ).
2. The code simulates 1000 repetitions of the experiment. That’s the meaning
of nreps <- 1000.
3. The above ground biomass of each site is stored in M.actual.control and
M.actual.treatment. There must be room to store the biomass of each site
for every combination of (pair,year,repetition). The array(...) command
creates a multidimensional matrix, or array, filled with NA’s. The dimensions
are given by c(npairs,nyears+1,nreps).
4. A site’s actual biomass is not known exactly but is measured with error. The
simulated measurements are stored in M.observed.control and
M.observed.treatment.
5. Each repetition begins by choosing 6 sites from among the 16 available. Their
observed biomass goes into temp. The first three values are assigned to
M.observed.control and the last three to M.observed.treatment. All this
happens in a loop for(i in 1:nreps).
6. Investigators expected that control plots would grow at an average rate of 2%
per year and treatment plots at an average of something else. Those values
are called betaC and betaT. The simulation was run with betaT = 1.04,
1.06, 1.08 (shown below) and 1.10. Each site would have its own growth
rate which would be slightly different from betaC or betaT. For control sites,
those rates are drawn from the N(betaC, 0.1 ∗ (betaC − 1)) distribution and
stored in beta.control, and similarly for the treatment sites.
7. Measurement errors of biomass were expected to have an SD around 5%.
That’s sigmaE. But at each site in each year the measurement error would be
slightly different. The measurement errors are drawn from the N(1, sigmaE)
distribution and stored in errors.control and errors.treatment.
8. Next we simulate the actual biomass of the sites. For the first year where we
already have measurements that’s

M.actual.control [ , 1, ] <- ...

M.actual.treatment [ , 1, ] <- ...
1.7. SIMULATION 59

For subsequent years the biomass in year i is the biomass in year i-1 mul-
tiplied by the growth factor beta.control or beta.treatment. Biomass is
simulated in the loop for(i in 2:(nyears+1)).
9. Measured biomass is the actual biomass multiplied by measurement error. It
is simulated by

M.observed.control <- ...

M.observed.treatment <- ...

10. The simulations for each year were analyzed each year by a two-sample t-test
which looks at the ratio
biomass in year i
biomass in year 1
to see whether it is significantly larger for treatment sites than for control
sites. See Section xyz for details about t-tests. For our purposes here, we have
replaced the t-test with a plot, Figure 1.21, which shows a clear separation
between treatment and control sites after about 5 years.

The DOE did decide to fund the proposal for a FACE experiment in Duke Forest,
at least partly because of the demonstration that such an experiment would have a
reasonably large chance of success.

########################################################
# A power analysis of the FACE experiment
#

# Initial measured biomass of potential FACE sites in g/m2:

b.mass <- c ( 17299.1, 17793.1, 23211.7, 23351.8, 24278,
25335.9, 27001.5, 27113.6, 30184.3, 30625.5,
33496.2, 33733.76, 35974.3, 38490.8, 40319.6,
44903 )

npairs <- 3
nyears <- 10
nreps <- 1000

M.observed.control <- array ( NA, c(npairs,nyears+1,nreps) )

M.actual.control <- array ( NA, c(npairs,nyears+1,nreps) )
M.observed.treatment <- array ( NA, c(npairs,nyears+1,nreps) )
M.actual.treatment <- array ( NA, c(npairs,nyears+1,nreps) )

# Specify the initial levels of biomass

for ( i in 1:nreps ) {
temp <- sample ( b.mass, size=2*npairs )
M.observed.control [,1,i] <- temp [1:npairs]
60 CHAPTER 1. PROBABILITY

2.0
growth rate

1.5
1.0

2 4 6 8 10

year

Figure 1.21: 1000 simulations of a FACE experiment. The x-axis is years.

The y-axis shows the mean growth rate (biomass in year i / biomass in year
1) of control plants (lower solid line) and treatment plants (upper solid line).
Standard deviations are shown as dashed lines.
1.7. SIMULATION 61

M.observed.treatment [,1,i] <- temp [(npairs+1):(2*npairs)]

}

# Specify the betas

betaC <- 1.02

betaT <- 1.08

beta.control <- matrix ( rnorm ( npairs*nreps, betaC,

0.1*(betaC-1) ),
npairs, nreps )
beta.treatment <- matrix ( rnorm ( npairs*nreps, betaT,
0.1*(betaT-1) ),
npairs, nreps )

#############################################################
# measurement errors in biomass
sigmaE <- 0.05

errors.control <- array ( rnorm ( npairs*(nyears+1)*nreps,

1, sigmaE ),
c(npairs,nyears+1,nreps) )
errors.treatment <- array ( rnorm ( npairs*(nyears+1)*nreps,
1, sigmaE ),
c(npairs,nyears+1,nreps) )
##############################################################
##############################################################
# Generate 10 years of data. The model for generation is:
# M.actual[i,j,] : above ground biomass in ring i, year j
# M.actual[i,j+1,] = beta[i] * M.actual[i,j,]

# We actually observe M.observed[i,j] = M.actual[i,j] * error

# Start with M.observed[i,1] and generate M.actual[i,1]

M.actual.control [ , 1, ] <-
M.observed.control [ , 1, ] / errors.control[ , 1, ]

M.actual.treatment [ , 1, ] <-
M.observed.treatment [ , 1, ] / errors.treatment[ , 1, ]

# Now, for the rest of the years, generate M.actual

for ( i in 2:(nyears+1) ) {
M.actual.control [ , i, ] <-
M.actual.control [ , i-1, ] * beta.control

M.actual.treatment [ , i, ] <-
62 CHAPTER 1. PROBABILITY

M.actual.treatment [ , i-1, ] * beta.treatment

}

# The initial observed data, corresponding to the [,1,],

# doesn’t need to be recomputed. But the following two
# statements are cleaner without subscripts.
M.observed.control <- M.actual.control * errors.control
M.observed.treatment <- M.actual.treatment * errors.treatment
##############################################################

##############################################################
# two-sample t-test on (M.observed[j]/M.observed[1]) removed
# plot added

ratio.control <- matrix ( NA, nyears, npairs*nreps )

ratio.treatment <- matrix ( NA, nyears, npairs*nreps )

for ( i in 2:(nyears+1) ) {
ratio.control [ i-1, ] <-
as.vector ( M.observed.control[,i,]
/ M.observed.control[,1,] )
ratio.treatment [ i-1, ] <-
as.vector ( M.observed.treatment[,i,]
/ M.observed.treatment[,1,] )
}

mean.control <- apply ( ratio.control, 1, mean )

mean.treatment <- apply ( ratio.treatment, 1, mean )
sd.control <- sqrt ( apply ( ratio.control, 1, var ) )
sd.treatment <- sqrt ( apply ( ratio.treatment, 1, var ) )

plot ( 2:11, mean.control, type="l", ylim=c(.9,2.4),

xlab="year", ylab="growth rate" )
lines ( 2:11, mean.treatment )
lines ( 2:11, mean.control - sd.control, lty=2 )
lines ( 2:11, mean.control + sd.control, lty=2 )
lines ( 2:11, mean.treatment - sd.treatment, lty=2 )
lines ( 2:11, mean.treatment + sd.treatment, lty=2 )

1.8 R
This section introduces a few more of the R commands we will need to work
fluently with the software. They are introduced in the context of studying a
dataset on the percent bodyfat of 252 men. You should download the data onto
your own computer and try out the analysis in R to develop your familiarity with
1.8. R 63

what will prove to be a very useful tool. The data can be found at StatLib,
an on-line repository of statistical data and software. The data were originally
contributed by Roger Johnson of the Department of Mathematics and Computer
Science at the South Dakota School of Mines and Technology. The StatLib
website is lib.stat.cmu.edu. If you go to StatLib and follow the links to
datasets and then bodyfat you will find a file containing both the data and
an explanation. Copy just the data to a text file named bodyfat.dat on your
own computer. The file should contain just the data; the first few lines should
look like this:
1.0708 12.3 23 ...
1.0853 6.1 22 ...
1.0414 25.3 22 ...
The following snippet shows how to read the data into R and save it into
bodyfat.
bodyfat <- read.table ( "bodyfat.dat",
col.names = c ( "density", "percent.fat", "age", "weight",
"height", "neck.circum", "chest.circum", "abdomen.circum",
"hip.circum", "thigh.circum", "knee.circum", "ankle.circum",
"bicep.circum", "forearm.circum", "wrist.circum" ) )
dim ( bodyfat ) # how many rows and columns in the dataset?
names ( bodyfat ) # names of the columns

• read.table(...) reads data from a text file into a dataframe. A

dataframe is a flexible way to represent data because it can be treated
as either a matrix or a list. Type help(read.table) to learn more. The
first argument, "bodyfat.dat", tells R what file to read. The second ar-
gument, col.names = c ( "density", ... ), tells R the names of the
columns.
• dim gives the dimension of a matrix, a dataframe, or anything else that
has a dimension. For a matrix or dataframe, dim tells how many rows and
columns.
• names gives the names of things. names(bodyfat) should tell us the names
density, percent.fat, . . . . It’s used here to check that the data were
read the way we intended.

Individual elements of matrices can be accessed by two-dimensional sub-

scripts such as bodyfat[1,1] or bodyfat[3,7] in which the subscripts refer
to the row and column of the matrix. (Try this out to make sure you know how
two dimensional subscripts work.) If the columns of the matrix have names,
then the second subscript can be a name, as in bodyfat[1,"density"] or
bodyfat[3,"chest.circum"]. Often we need to refer to an entire column
at once, which can be done by omitting the first subscript. For example,
bodyfat[,2] refers to the entire set of 252 measurements of percent body fat.
64 CHAPTER 1. PROBABILITY

A dataframe is a list of columns. Because bodyfat has 15 columns its length,

length(bodyfat), is 15. Members of a list can be accessed by subscripts with
double brackets, as in bodyfat[[1]]. Each member of bodyfat is a vector of
length 252. Individual measurements can be accessed as in bodyfat[[1]][1] or
bodyfat[[3]][7]. If the list members have names, then they can be accessed
as in bodyfat$percent.fat. Note the quotation marks used when treating
bodyfat as a matrix and the lack of quotation marks when treating bodyfat as a
list. The name following the ”$” can be abbreviated, as long as the abbreviation
is unambiguous. Thus bodyfat$ab works, but bodyfat$a fails to distinguish
between age and abdomen.circum.
Begin by displaying the data.

par ( mfrow=c(5,3) ) # establish a 5 by 3 array of plots

for ( i in 1:15 ) {
hist ( bodyfat[[i]], xlab="", main=names(bodyfat)[i] )
}

Although it’s not our immediate purpose, it’s interesting to see what the rela-
tionships are among the variables. Try pairs(bodyfat).
To illustrate some of R’s capabilities and to explore the concepts of marginal,
joint and conditional densities, we’ll look more closely at percent fat and its
relation to abdomen circumference. Begin with a histogram of percent fat.

fat <- bodyfat$per # give these two variables short names

abd <- bodyfat$abd # so we can refer to them easily
par ( mfrow=c(1,1) ) # just need one plot now, not 15
hist ( fat )

We’d like to rescale the vertical axis to make the area under the histogram equal
to 1, as for a density. R will do that by drawing the histogram on a “density”
scale instead of a “frequency” scale. While we’re at it, we’ll also make the labels
prettier. We also want to draw a Normal curve approximation to the histogram,
so we’ll need the mean and standard deviation.

hist ( fat, xlab="", main="percent fat", freq=F )

mu <- mean ( fat )
sigma <- sqrt ( var ( fat )) # standard deviation
lo <- mu - 3*sigma
hi <- mu + 3*sigma
x <- seq ( lo, hi, length=50 )
lines ( x, dnorm ( x, mu, sigma ) )

That looks better, but we can do better still by slightly enlarging the axes.
Redraw the picture, but use

hist ( fat, xlab="", main="percent fat", freq=F,

xlim=c(-10, 60), ylim=c(0,.06) )
1.8. R 65

The Normal curve fits the data reasonably well. A good summary of the data
is that it is distributed approximately N(19.15, 8.37).
Now examine the relationship between abdomen circumference and percent
body fat. Try the following command.

plot ( abd, fat, xlab="abdomen circumference",

ylab="percent body fat" )

The scatter diagram shows a clear relationship between abdomen circumference

and body fat in this group of men. One man doesn’t fit the general pattern; he
has a circumference around 148 but a body fat only around 35%, relatively low
for such a large circumference. To quantify the relationship between the vari-
ables, let’s divide the men into groups according to circumference and estimate
the conditional distribution of fat given circumference. If we divide the men
into twelfths we’ll have 21 men per group.

cut.pts <- quantile ( abd, (0:12)/12 )

groups <- cut ( abd, cut.pts, labels=1:12 )
boxplot ( fat ~ groups,
xlab="quantiles of abdomen circumference",
ylab="percent body fat" )

Note:

• A quantile is a generalization of median. For example, the 1/12-th quantile

of abd is the number q such that 1/12 of all the abd measurements are
less than q and 11/12 are greater than q. (A more careful definition would
say what to do in case of ties.) The median is the .5 quantile. We have
cut our data according to the 1/12, 2/12, . . . , 12/12 quantiles of abd.

• If you don’t see what the cut(abd,...) command does, print out cut.pts
and groups, then look at them until you figure it out.

• Boxplots are a convenient way to compare different groups of data. In this

case there are 12 groups. Each group is represented on the plot by a box
with whiskers. The box spans the first and third quartiles (.25 and .75
quantiles) of fat for that group. The line through the middle of the box
is the median fat for that group. The whiskers extend to cover most of
the rest of the data. A few outlying fat values fall outside the whiskers;
they are indicated as individual points.

• ”fat ~ groups” is R’s notation for a formula. It means to treat fat

as a function of groups. Formulas are extremely useful and will arise
repeatedly.

The medians increase in not quite a regular pattern. The irregularities are
probably due to the vagaries of sampling. We can find the mean, median and
variance of fat for each group with
66 CHAPTER 1. PROBABILITY

mu.fat <- tapply ( fat, groups, mean )

me.fat <- tapply ( fat, groups, median )
sd.fat <- sqrt ( tapply ( fat, groups, var ) )
cbind ( mu.fat, me.fat, sd.fat )
Note: cbind means ”bind together in columns”. There is an analgous command
rbind.
Finally, let’s make a figure similar to Figure 1.12.
x <- seq ( 0, 50, by=1 )
par ( mfrow=c(4,3) )
for ( i in 1:12 )
good <- groups == i
hist ( fat[good], xlim=c(0,50), ylim=c(0,.1),
breaks=seq(0,50,by=5), freq=F,
xlab="percent fat", main="" )
y <- dnorm ( x, mu.fat[i], sd.fat[i] )
lines ( x, y )

The Normal curves seem to fit well. We saw earlier that the marginal (Marginal
means unconditional.) distribution of percent body fat is well approximated by
N(19.15, 8.37). Here we see that the conditional distribution of percent body
fat, given that abdomen circumference is in between the (i − 1)/12 and i/12
quantiles is N(mu.fat[i], sd.fat[i]). If we know a man’s abdomen circum-
ference even approximately then (1) we can estimate his percent body fat more
accurately and (2) the typical estimation error is smaller. [add something
about estimation error in the sd section]

1.9 Some Results for Large Samples

It is intuitively obvious that large samples are better than small, that more data
is better than less, and, less obviously, that with enough data one is eventually
led to the right answer. These intuitive observations have precise mathematical
statements in the form of Theorems 1.12, 1.13 and 1.14. We state those theorems
here so we can use them throughout the rest of the book. They are examined
in more detail in Section 7.9.
Definition 1.9 (Random Sample). A collection y1 , . . . , yn of random vari-
ables is said to be a random sample of size n from pdf or pmf f if and only
if
1. yi ∼ f for each i = 1, 2, . . . , n and
Qn
2. the yi ’s are mutually independent, i.e. f (y1 , . . . , yn ) = i=1 f (yi ).
The collection (y1 , . . . , yn ) is called a data set. We write
y1 , . . . , yn ∼ i.i.d. f
1.9. SOME RESULTS FOR LARGE SAMPLES 67

where i.i.d. stands for independent and identically distributed.

Many introductory statistics texts describe how to collect random samples,
many pitfalls that await, and many pratfalls taken in the attempt. We omit
that discussion here and refer the interested reader to our favorite introductory
text on the subject, Freedman et al. [1998] which has an excellent description
of random sampling in general as well as detailed discussion of the US census
and the Current Population Survey. R R
Suppose y1 , y2 , . . . , ∼ i.i.d. f . Let µ = yf (y) dy and σ 2 = (y − µ)2 f (y) dy
be the mean and variance of f . Typically µ and σ are unknown and we take the
sample in order to learn about them. We will often use ȳn ≡ (y1 + · · · + yn )/n,
the mean of the first n observations, to estimate µ. Some questions to consider
are
• For a sample of size n, how accurate is ȳn as an estimate of µ?
• Does ȳn get closer to µ as n increases?
• How large must n be in order to achieve a desired level of accuracy?
Theorems 1.12 and 1.14 provide answers to these questions. Before stating them
we need some preliminary results about the mean and variance of ȳn .
Theorem 1.7. Let x1 , . . . , xn be random variables with means µ1 , . . . , µn . Then
E[x1 + · · · + xn ] = µ1 + · · · + µn .
Proof. It suffices to prove the case n = 2.
Z Z
E[x1 + x2 ] = (x1 + x2 )f (x1 , x2 ) dx1 dx2
ZZ ZZ
= x1 f (x1 , x2 ) dx1 dx2 + x2 f (x1 , x2 ) dx1 dx2

= µ1 + µ2

Corollary 1.8. Let y1 , . . . , yn be a random sample from f with mean µ. Then

E[ȳn ] = µ.
Proof. The corollary follows from Theorems 1.3 and 1.7.
Theorem 1.9. Let x1 , . . . , xn be independent random variables with means
µ1 , . . . , µn and SDs σ1 , . . . , σn . Then Var[x1 + · · · + xn ] = σ12 + · · · + σn2 .
Proof. It suffices to prove the case n = 2. Using Theorem 1.2,

Var(X1 + X2 ) = E((X1 + X2 )2 ) − (µ1 + µ2 )2

= E(X12 ) + 2E(X1 X2 ) + E(X22 ) − µ21 − 2µ1 µ2 − µ22



= E(X12 ) − µ21 + E(X22 ) − µ22 + (E(X1 X2 ) − µ1 µ2 )
= σ12 + σ22 + (E(X1 X2 ) − µ1 µ2 ) .
68 CHAPTER 1. PROBABILITY

But if X1 ⊥ X2 then
ZZ
E(X1 X2 ) = x1 x2 f (x1 , x2 ) dx1 dx2
Z Z
= x1 x2 f (x2 ) dx2 f (x1 ) dx1
Z
= µ2 x1 f (x1 ) dx1 = µ1 µ2 .

So Var(X1 + X2 ) = σ12 + σ22 .

Note that Theorem 1.9 requires independence while Theorem 1.7 does not.
Corollary 1.10. Let y1 , . . . , yn be a random sample from f with variance σ 2 .
Then Var(ȳn ) = σ 2 /n.
Proof. The corollary follows from Theorems 1.4 and 1.9.
Theorem 1.11 (Chebychev’s Inequality). Let X be a random variable with
mean µ and SD σ. Then for any  > 0,
P[|X − µ| ≥ ] ≤ σ 2 /2 .
Proof.
Z
2
σ = (x − µ)2 f (x) dx
Z µ− Z µ+ Z ∞
2 2
= (x − µ) f (x) dx + (x − µ) f (x) dx + (x − µ)2 f (x) dx
−∞ µ− µ+
Z µ− Z ∞
≥ (x − µ)2 f (x) dx + (x − µ)2 f (x) dx
−∞ µ+
Z µ− Z ∞
2 2
≥ f (x) dx +  f (x) dx
−∞ µ+
= 2 P[|X − µ| ≥ ].

Theorems 1.12 and 1.14 are the two main limit theorems of statistics. They
provide answers, at least probabilistically, to the questions on page 67.
Theorem 1.12 (Weak Law of Large Numbers). Let y1 , . . . , yn be a random
sample from a distribution with mean µ and variance σ 2 . Then for any  > 0,
lim P[|ȳn − µ| < ] = 1. (1.16)
n→∞

Proof. Apply Chebychev’s Inequality (Theorem 1.11) to ȳn .

lim P[|ȳn − µ| < ] = lim 1 − P[|ȳn − µ| ≥ ] ≥ lim 1 − σ 2 /n2 = 1.
n→∞ n→∞ n→∞
1.9. SOME RESULTS FOR LARGE SAMPLES 69

Another version of Theorem 1.12 is called the Strong Law of Large Numbers.

Theorem 1.13 (Strong Law of Large Numbers). Let y1 , . . . , yn be a ran-

dom sample from a distribution with mean µ and variance σ 2 . Then for any
 > 0,

P[ lim |Ȳn − µ| < ] = 1;

n→∞

i.e.,

P[ lim Ȳn = µ] = 1.
n→∞

It is beyond the scope of this section to explain the difference between the
WLLN and the SLLN. See Section 7.9.

Theorem 1.14 (Central Limit Theorem). Let y1 , . . . , yn be √ a random sam-

ple from f with mean µ and variance σ 2 . Let zn = (ȳn − µ)/(σ/ n). Then, for
any numbers a < b,
Z b
1 2
lim P[zn ∈ [a, b]] = √ e−w /2 dw.
n→∞ a 2π

I.e. The limiting distribution of zn is N(0, 1).

The Law of Large Numbers is what makes simulations work and why large
samples are better than small. It says that as the number of simulation grows or
as the sample size grows, (n → ∞), the average of the simulations or the average
of the sample gets closer and closer to the true value (X̄n → µ). For instance, in
Example 1.11, where we used simulation to estimate P[Shooter wins] in Craps,
the estimate became more and more accurate as the number of simulations
increased from 50, to 200, and then to 1000. The Central Limit Theorem helps
us look at those simulations more closely.
Colloquially, the Central Limit Theorem says that

lim pzn = N(0, 1).

n→∞

Its practical application is that, for large n,

pzn ≈ N(0, 1),

which in turn means

√
pȳn ≈ N(µ, σ/ n).
In Example 1.11 we simulated the game of Craps for n.sim = 50, 200, and
1000. Those simulations are depicted in Figure 1.22. The upper panel shows
a histogram of 1000 simulations, all using n.sim = 50. For a single simulation
with n.sim = 50 let X1 , . . . , X50 be the outcomes of those simulations. Each
70 CHAPTER 1. PROBABILITY
√
Xi ∼ Bern(.493), so µ = .493 and σ = .493 ∗ .507 ≈ .5. Therefore, according
to the Central Limit Theorem, when n.sim = 50
√
X̄n ∼ N(µ, σ/ n) ≈ N(.493, .071)

This is the Normal density plotted in the upper panel of Figure 1.22. We see
that the N(.493, .071) is a good approximation to the histogram. And that’s
because θ̂ = X̄50 ∼ N(.493, .071), approximately. The Central Limit Theorem
says that the approximation will be good for “large” n. In this case n = 50
is large enough. (Section 7.1.4 will discuss the question of when n is “large
enough”.)
Similarly,

X̄n ∼ N(.493, .035) when n.sim = 200

X̄n ∼ N(.493, .016) when n.sim = 1000.

These densities are plotted in the middle and lower panels of Figure 1.22.
The Central Limit Theorem makes three statements about the distribution
of ȳn (zn ) in large samples:

1. E[ȳn ] = µ (E[zn ] = µ),

√
2. SD(ȳn ) = σ/ n (SD(zn ) = 1), and

3. ȳn (zn ) has, approximately, a Normal distribution.

The first two of these are already known from Theorems 1.7 and 1.9. It’s
the third point that is key to the Central Limit Theorem. Another surprising
implication from the Central Limit Theorem is that the distributions of ȳn and
zn in large samples are determined solely by µ and σ; no other features of f
matter.

1.10 Exercises
1. Show: if µ is a probability measure then for any integer n ≥ 2, and disjoint
sets A1 , . . . , An
[n Xn
µ( Ai ) = µ(Ai ).
i=1 i=1

2. Simulating Dice Rolls

(a) simulate 6000 dice rolls. Count the number of 1’s, 2’s, . . . , 6’s.
(b) You expect about 1000 of each number. How close was your result
to what you expected?
(c) About how often would you expect to get more that 1030 1’s? Run
an R simulation to estimate the answer.
1.10. EXERCISES 71

0 1 2 3 4 5 6
Density

0.3 0.4 0.5 0.6 0.7

theta hat
n.sim = 50
8 10
Density

6
4
2
0

0.3 0.4 0.5 0.6 0.7

theta hat
n.sim = 200
20
Density

10
5
0

0.3 0.4 0.5 0.6 0.7

theta hat
n.sim = 1000

Figure 1.22: Histograms of craps simulations. Solid curves are Normal approx-
imations according to the Central Limit Theorem.
72 CHAPTER 1. PROBABILITY

3. The Game of Risk In the board game Risk players place their armies
in different countries and try eventually to control the whole world by
capturing countries one at a time from other players. To capture a country,
a player must attack it from an adjacent country. If player A has A ≥ 2
armies in country A, she may attack adjacent country D. Attacks are
made with from 1 to 3 armies. Since at least 1 army must be left behind
in the attacking country, A may choose to attack with a minimum of 1
and a maximum of min(3, A − 1) armies. If player D has D ≥ 1 armies
in country D, he may defend himself against attack using a minimum of
1 and a maximum of min(2, D) armies. It is almost always best to attack
and defend with the maximum permissible number of armies.
When player A attacks with a armies she rolls a dice. When player D
defends with d armies he rolls d dice. A’s highest die is compared to D’s
highest. If both players use at least two dice, then A’s second highest is
also compared to D’s second highest. For each comparison, if A’s die is
higher than D’s then A wins and D removes one army from the board;
otherwise D wins and A removes one army from the board. When there
are two comparisons, a total of two armies are removed from the board.

• If A attacks with one army (she has two armies in country A, so may
only attack with one) and D defends with one army (he has only one
army in country D) what is the probability that A will win?
• Suppose that Player 1 has two armies each in countries C1 , C2 , C3
and C4 , that Player 2 has one army each in countries B1 , B2 , B3
and B4 , and that country Ci attacks country Bi . What is the chance
that Player 1 will be successful in at least one of the four attacks?

4. (a) Justify the last step of Equation 1.2.

(b) Justify the last step of the proof of Theorem 1.1.
(c) Prove Theorem 1.1 when g is a decreasing function.

5. Y is a random variable. Y ∈ (−1, 1). The pdf is p(y) = ky 2 for some

constant, k.

(a) Find k.
(b) Use R to plot the pdf.
(c) Let Z = −Y . Find the pdf of Z. Plot it.

6. U is a random variable on the interval [0, 1]; p(u) = 1.

(a) V = U 2 . On what interval does V live? Plot V as a function of U .

Find the pdf of V . Plot pV (v) as a function of v.
(b) W = 2U . On what interval does W live? Plot W as a function of U .
Find the pdf of W . Plot pW (w) as a function of w.
1.10. EXERCISES 73

(c) X = − log(U ). On what interval does X live? Plot X as a function

of U . Find the pdf of X. Plot pX (x) as a function of x.
7. Let X ∼ Exp(λ) and let Y = cX for some constant c.
(a) Write down the density of X.
(b) Find the density of Y .
(c) Name the distribution of Y .
8. A teacher randomly selects a student from a Sta 103 class. Let X be the
number of math courses the student has completed. Let Y = 1 if the stu-
dent is female and Y = 0 if the student is male. Fifty percent of the class
is female. Among the women, thirty percent have completed one math
class, forty percent have completed two math classes and thirty percent
have completed three. Among the men, thirty percent have completed
one math class, fifty percent have completed two math classes and twenty
percent have completed three.
(a) True or False: X and Y are independent.
(b) Find E[X|Y = 1].
9. Sue is studying the Bin(25,.4) distribution. In R she types

y <- rbinom(50, 25, .4)

m1 <- mean(y)
m2 <- sum(y) / 25
m3 <- sum ( (y-m1)^2 ) / 50

(a) Is y a number, a vector or a matrix?

(b) What is the approximate value of m1?
(c) What is the approximate value of m2?
(d) What was Sue trying to estimate with m3?

10. The random variables X and Y have joint pdf fX,Y (x, y) = 1 in the
triangle of the XY -plane determined by the points (-1,0), (1,0), and (0,1).
Hint: Draw a picture.
(a) Find fX (.5).
(b) Find fY (y).
(c) Find fY | X (y | X = .5).
(d) Find E[Y | X = .5].
(e) Find fY (.5).
(f) Find fX (x).
(g) Find fX | Y (x | Y = .5).
74 CHAPTER 1. PROBABILITY

(h) Find E[X | Y = .5].

11. Verify the claim in Example 1.4 that argmaxλ P[x = 3 | λ] = 3. Hint:
differentiate Equation 1.6.
R
12. (a) p is the pdf of a continuous random variable w. Find R p(s) ds.
R
(b) Find R p(s) ds for the pdf in Equation 1.7.
13. Page 7 says “Every pdf must satisfy two properties . . . ” and that one of
them is “p(y) ≥ 0 for all y.” Explain why that’s not quite right.
14.
1 1 2
p(y) = √ e− 2 y
2π
R0
is the pdf of a continuous random variable y. Find −∞ p(s) ds.
15. When spun, an unbiased spinner points to some number y ∈ (0, 1]. What
is p(y)?
16. Some exercises on the densities of tranformed variables. One of them
should illustrate the need for the absolute value of the Jacobian.
17. Prove Theorem 1.1 when g is decreasing.
18. (a) Prove: if X ∼ Poi(λ) then E(X) = λ. Hint: use the same trick we
used to derive the mean of the Binomial distribution.
(b) Prove: if X ∼ N(µ, σ) then E(X) = µ. Hint: change variables in the
integral.
19. (a) Prove: if X ∼ Bin(n, p) then Var(X) = np(1 − p). Hint: use Theo-
rem 1.9.
(b) Prove: if X ∼ Poi(λ) then Var(X) = λ. Hint: use the same trick
we used to derive the mean of the Binomial distribution and Theo-
rem 1.2.
(c) If X ∼ Exp(λ), find Var(X). Hint: use Theorem 1.2.
(d) If X ∼ N(µ, σ), find Var(X). Hint: use Theorem 1.2.
20. Let X1 ∼ Bin(50, .1), X2 ∼ Bin(50, .9) and X1 ⊥ X2 . Define Y = X1 +X2 .
Does Y have the Bin(100, .5) distribution? Why or why not?
21. Let X1 ∼ Bin(50, .5), X2 ∼ Bin(50, .5) and X1 ⊥ X2 . Define Y1 = X1 +X2
and Y2 = 2X1 . Which has the bigger mean: Y1 or Y2 ? Which has the
bigger variance: Y1 or Y2 ? Justify your answer.
22. Consider customers arriving at a service counter. Interarrival times often
have a distribution that is approximately exponential with a parameter λ
that depends on conditions specific to the particular counter. I.e., p(t) =
λe−λt . Assume that successive interarrival times are independent of each
other. Let T1 be the arrival time of the next customer and T2 be the
additional time until the arrival of the second customer.
1.10. EXERCISES 75

(a) What is the joint density of (T1 , T2 )?

(b) Let S = T1 + T2 , the time until the next two customers arrive. What
is P[S ≤ 5]; i.e. the probability that at least 2 customers arrive within
the next 5 minutes?
(c) What is E(S)?
23. A gambler plays at a roulette table for two hours betting on Red at each
spin of the wheel. There are 60 spins during the two hour period. What
is the distribution of
(a) z, the number of times the gambler wins,
(b) y, the number of times the gambler loses,
(c) x, the number of times the gambler is jostled by the person standing
behind,
(d) w, the gambler’s net gain?
24. If human DNA contains xxx bases, and if each base mutates with probabil-
ity p over the course of a lifetime, what is the average number of mutations
per person? What is the variance of the number of mutations per person?
25. Isaac is in 5th grade. Each sentence he writes for homework has a 90%
chance of being grammatically correct. The correctness of one sentence
does not affect the correctness of any other sentence. He recently wrote a
10 sentence paragraph for a writing assignment. Write a formula for the
chance that no more than two sentences are grammatically incorrect.
26. Teams A and B play each other in the World Series of baseball. Team A
has a 60% chance of winning each game. What is the chance that B wins
the series? (The winner of the series is the first team to win 4 games.)
27. A basketball player shoots ten free throws in a game. She has a 70%
chance of making each shot. If she misses the shot, her team has a 30%
chance of getting the rebound.
(a) Let m be the number of shots she makes. What is the distribution of
m? What are its expected value and variance? What is the chance
that she makes somewhere between 5 and 9 shots, inclusive?
(b) Let r be the number of rebounds her team gets from her free throws.
What is the distribution of r? What are its expected value and
variance? What is the chance that r ≥ 1?
28. Let (x, y) have joint density
RR function fx,y . There are two ways to find Ey.
One way is to evaluate yfx,y (x, y) dxdy. The other is to start
R with the
joint density fx,y , find the marginal density fy , then evaluate yfy (y) dy.
Show that these two methods give the same answer.
29. Prove Theorem 1.3 (pg. 33) in the discrete case.
76 CHAPTER 1. PROBABILITY

30. Prove Theorem 1.7 (pg. 67) in the continuous case.

31. A researcher randomly selects mother-daughter pairs. Let xi and yi be

the heights of the i’th mother and daughter, respectively. True or False:

(a) xi and xj are independent

(b) xi and yj are independent
(c) yi and yj are independent
(d) xi and yi are independent

32. As part of his math homework Isaac had to roll two dice and record the
results. Let X1 be the result of the first die and X2 be the result of the
second. What is the probability that X1=1 given that X1 + X2 = 5?

33. A doctor suspects a patient has the rare medical condition DS, or dis-
staticularia, the inability to learn statistics. DS occurs in .01% of the
population, or one person in 10,000. The doctor orders a diagnostic test.
The test is quite accurate. Among people who have DS the test yields a
positive result 99% of the time. Among people who do not have DS the
test yields a positive result only 5% of the time.
For the patient in question, the test result is positive. Calculate the prob-
ability that the patient has DS.

34. For various reasons, researchers often want to know the number of peo-
ple who have participated in “embarassing” activities such as illegal drug
use, cheating on tests, robbing banks, etc. An opinion poll which asks
these questions directly is likely to elicit many untruthful answers. To get
around the problem, researchers have devised the method of randomized
response. The following scenario illustrates the method.
A pollster identifies a respondent and gives the following instructions.
“Toss a coin, but don’t show it to me. If it lands Heads, answer question
(a). If it lands tails, answer question (b). Just answer ’yes’ or ’no’. Do
not tell me which question you are answering.
Question (a): Does your telephone number end in an even digit?
Question (b): Have you ever used cocaine?”
Because the respondent can answer truthfully without revealing his or her
cocaine use, the incentive to lie is removed. Researchers hope respondents
will tell the truth.
You may assume that respondents are truthful and that telephone numbers
are equally likely to be odd or even. Let p be the probability that a
randomly selected person has used cocaine.

(a) What is the probability that a randomly selected person answers

”yes”?
1.10. EXERCISES 77

(b) Suppose we survey 100 people. Let X be the number who answer
”yes”. What is the distribution of X?

35. This exercise is based on a computer lab that another professor uses to
teach the Central Limit Theorem. It was originally written in MATLAB but
here it’s translated into R.
Enter the following R commands:

u <- matrix ( runif(250000), 1000, 250 )

y <- apply ( u, 2, mean )

These create a 1000x250 (a thousand rows and two hundred fifty columns)
matrix of random draws, called u and a 250-dimensional vector y which
contains the means of each column of U.
Now enter the command hist(u[,1]). This command takes the first
column of u (a column vector with 1000 entries) and makes a histogram.
Print out this histogram and describe what it looks like. What distribution
is the runif command drawing from?
Now enter the command hist(y). This command makes a histogram
from the vector y. Print out this histogram. Describe what it looks like
and how it differs from the one above. Based on the histogram, what
distribution do you think y follows?
You generated y and u with the same random draws, so how can they
have different distributions? What’s going on here?

36. Suppose that extensive testing has revealed that people in Group A have
IQ’s that are well described by a N(100, 10) distribution while the IQ’s of
people in Group B have a N(105, 10) distribution. What is the probability
that a randomly chosen individual from Group A has a higher IQ than a
randomly chosen individual from Group B?

(a) Write a formula to answer the question. You don’t need to evaluate
the formula.
(b) Write some R code to answer the question.

37. The so-called Monte Hall or Let’s Make a Deal problem has caused much
consternation over the years. It is named for an old television program.
A contestant is presented with three doors. Behind one door is a fabu-
lous prize; behind the other two doors are virtually worthless prizes. The
contestant chooses a door. The host of the show, Monte Hall, then opens
one of the remaining two doors, revealing one of the worthless prizes. Be-
cause Monte is the host, he knows which doors conceal the worthless prizes
and always chooses one of them to reveal, but never the door chosen by
the contestant. Then the contestant is offered the choice of keeping what
78 CHAPTER 1. PROBABILITY

is behind her original door or trading for what is behind the remaining
unopened door. What should she do?
There are two popular answers.
• There are two unopened doors, they are equally likely to conceal the
fabulous prize, so it doesn’t matter which one she chooses.
• She had a 1/3 probability of choosing the right door initially, a 2/3
chance of getting the prize if she trades, so she should trade.

(a) Create a simulation in R to discover which answer is correct.

(b) Show using formal arguments of conditional probability which answer
is correct.

Make sure your answers to (a) and (b) agree!

38. Prove Theorem 1.6 (pg. 46).
 Chapter 2

Modes of Inference

2.1 Data
This chapter takes up the heart of statistics: making inferences, quantitatively,
from data. The data, y1 , . . . , yn are assumed to be a random sample from a
population.
In Chapter 1 we reasoned from f to Y . That is, we made statements like “If
the experiment is like . . . , then f will be . . . , and (y1 , . . . , yn ) will look like . . . ”
or “E(Y ) must be . . . ”, etc. In Chapter 2 we reason from Y to f . That is, we
make statements such as R “Since (y1 , . . . , yn ) turned out to be . . . it seems that f
is likely to be . . . ”, or “ yf (y) dy is likely to be around . . . ”, etc. This is a basis
for knowledge: learning about the world by observing it. Its importance cannot
be overstated. The field of statistics illuminates the type of thinking that allows
us to learn from data and contains the tools for learning quantitatively.
Reasoning from Y to f works because samples are usually like the popula-
tions from which they come. For example, if f has a mean around 6 then most
reasonably large samples from f also have a mean around 6, and if our sample
has a mean around 6 then we infer that f likely has a mean around 6. If our
sample has an SD around 10 then we infer that f likely has an SD around 10,
and so on. So much is obvious. But can we be more precise? If our sample
has a mean around 6, then can we infer that f likely has a mean somewhere
between, say, 5.5 and 6.5, or can we only infer that f likely has a mean between
4 and 8, or even worse, between about -100 and 100? When can we say any-
thing quantitative at all about the mean of f ? The answer is not obvious, and
that’s where statistics comes in. Statistics provides the quantitative tools for
answering such questions.
This chapter presents several generic modes of statistical analysis.
Data Description Data description can be visual, through graphs, charts,
etc., or numerical, through calculating sample means, SD’s, etc. Display-
ing a few simple features of the data y1 , . . . , yn can allow us to visualize
those same features of f . Data description requires few a priori assump-
tions about f .

79
80 CHAPTER 2. MODES OF INFERENCE

Likelihood In likelihood inference we assume that f is a member of a para-

metric family of distributions {fθ : θ ∈ Θ}. Then inference about f is
the same as inference about the parameter θ, and different values of θ are
compared according to how well fθ explains the data.

Estimation The goal of estimation is to estimate various aspects of f , such as

its mean, median, SD, etc. Along with the estimate, statisticians try to
give quantitative measures of how accurate the estimates are.

Bayesian Inference Bayesian inference is a way to account not just for the
data y1 , . . . , yn , but also for other information we may have about f .

Prediction Sometimes the goal of statistical analysis is not to learn about f

per se, but to make predictions about y’s that we will see in the future.
In addition to the usual problem of not knowing f , we have the additional
problem that even if we knew f , we still wouldn’t be able to predict future
y’s exactly.

Decision Making Often, decisions have to be made on the basis of what we

have learned about f . In addition, making good decisions requires ac-
counting for the potential gains and losses of each decision.

2.2 Data Description

There are many ways, both graphical and numerical, to describe data sets.
Sometimes we’re interested in means, sometimes variations, sometimes trends
through time, and there are good ways to describe and display all these aspects
and many more. Simple data description is often enough to shed light on an
underlying scientific problem. The subsections of Section 2.2 show some basic
ways to describe various types of data.

2.2.1 Summary Statistics

One of the simplest ways to describe a data set is by a low dimensional summary.
For instance, in Example 1.5 on ocean temperatures there were multiple mea-
surements of temperatures from each of 9 locations. The measurements P from
each location were summarized by the sample mean ȳ = n−1 yi ; compar-
isons of the 9 sample means helped oceanographers deduce the presence of the
Mediterranean tongue. Similarly, the essential features of many data sets can be
captured in a one-dimensional or low-dimensional summary. Such a summary
is called a statistic. The examples below refer to a data set y1 , . . . , yn of size n.

Definition 2.1 (Statistic). A statistic is any function, possibly vector valued,

of the data.

The most important statistics are measures of location and dispersion. Im-
portant examples of location statistics include
2.2. DATA DESCRIPTION 81

P
mean The mean of the data is ȳ ≡ n−1 yi . R can compute means:

y <- 1:10
mean(y)

median A median of the data is any number m such that at least half of the
yi ’s are less than or equal to m and at least half of the yi ’s are greater
than or equal to m. We say “a” median instead of “the” median because a
data set with an even number of observations has an interval of medians.
For example, if y <- 1:10, then every m ∈ [5, 6] is a median. When R
computes a median it computes a single number by taking the midpoint
of the interval of medians. So median(y) yields 5.5.

quantiles For any p ∈ [0, 1], the p-th quantile of the data should be, roughly
speaking, the number q such that pn of the data points are less than q
and (1 − p)n of the data points are greater than q.
Figure 2.1 illustrates the idea. Panel a shows a sample of 100 points
plotted as a stripchart (page 92). The black circles on the abcissa are the
.05, .5, and .9 quantiles; so 5 points (open circles) are to the left of the
first vertical line, 50 points are on either side of the middle vertical line,
and 10 points are to the right of the third vertical line. Panel b shows
the empirical cdf of the sample. The values .05, .5, and .9 are shown
as squares on the vertical axis; the quantiles are found by following the
horizontal lines from the vertical axis to the cdf, then the vertical lines
from the cdf to the horizontal axis. Panels c and d are similar, but show
the distribution from which the sample was drawn instead of showing the
sample itself. In panel c, 5% of the mass is to the left of the first black
circle; 50% is on either side of the middle black circle; and 10% is to the
right of the third black dot. In panel d, the open squares are at .05, .5,
and .9 on the vertical axis; the quantiles are the circles on the horizontal
axis.
Denote the p-th quantile as qp (y1 , . . . , yn ), or simply as qp if the data set
is clear from the context. With only a finite sized sample qp (y1 , . . . , yn )
cannot be found exactly. So the algorithm for finding quantiles works as
follows.

1. Sort the yi ’s in ascending order. Label them y(1) , . . . , y(n) so that

y(1) ≤ · · · ≤ y(n) .

2. Set q0 ≡ y(1) and q1 ≡ y(n) .

3. y(2) through y(n−1) determine n − 1 subintervals in [y(1) , y(n) ]. So,
for i = 1, . . . , n − 2, set q n−1
i ≡ y(i+1) .
i i+1
4. For p ∈ ( n−1 , n−1 ) let qp be any number in the interval (q i , q i+1 ).
n−1 n−1
82 CHAPTER 2. MODES OF INFERENCE

If p is a “nice” number then qp is often given a special name. For example,

q.5 is the median; (q.25 , q.5 , q.75 ), the first, second and third quartiles, is
a vector-valued statistic of dimension 3; q.1 , q.2 , . . . are the deciles; q.78 is
the 78’th percentile.
i i+1
R can compute quantiles. When faced with p ∈ ( n−1 , n−1 ) R does linear
interpolation. E.g. quantile(y,c(.25,.75)) yields (3.25, 7.75).

The vector (y(1) , . . . , y(n) ) defined in step 1 of the algorithm for quantiles is
an n-dimensional statistic called the order statistic. y(i) by itself is called
the i’th order statistic.

a b

0.8
F(y)

0.4
0.0

0 2 4 6 8 0 2 4 6 8

y y

c d
0.8
0.15

F(y)
p(y)

0.4
0.00

0.0

0 2 4 6 8 0 2 4 6 8

y y

Figure 2.1: Quantiles. The black circles are the .05, .5, and .9 quantiles. Panels a
and b are for a sample; panels c and d are for a distribution.
2.2. DATA DESCRIPTION 83

Figure 2.1 was created with the following R code.

plot.ecdf(y, xlab="y", ylab="F(y)", xlim=c(0,10),
main="b" )
for ( q in quant )
segments ( c(0,quantile(y,q)), c(q,0),
rep(quantile(y,q),2), rep(q,2) )
points ( x=quantile(y,quant), y=rep(0,nquant), pch=19 )
points ( x=rep(0,nquant), y=quant, pch=22 )

y <- seq(0,10,length=100)
plot ( y, dgamma(y,3), type="l", xlim=c(0,10), ylab="p(y)",
main="c" )
points ( x=qgamma(quant,3), y=rep(0,nquant), pch=19 )

plot ( y, pgamma(y,3), type="l", ylab="F(y)", main="d" )

for ( q in quant )
segments ( c(0,qgamma(q,3)), c(q,0), rep(qgamma(q,3),2),
rep(q,2) )
points ( x=qgamma(quant,3), y=rep(0,nquant), pch=19 )
points ( x=rep(0,nquant), y=quant, pch=22 )

Dispersion statistics measure how spread out the data are. Since there are many
ways to measure dispersion there are many dispersion statistics. Important
dispersion statistics include
standard deviation The sample standard deviation or SD of a data set is
rP
(yi − ȳ)2
s≡
n
Note: some statisticians prefer
rP
(yi − ȳ)2
s≡
n−1
for reasons which do not concern us here. If n is large there is little
difference between the two versions of s.
variance The sample variance is
P
2 (yi − ȳ)2
s ≡
n
Note: some statisticians prefer
P
2 (yi − ȳ)2
s ≡
n−1
84 CHAPTER 2. MODES OF INFERENCE

for reasons which do not concern us here. If n is large there is little

difference between the two versions of s2 .
interquartile range The interquartile range is q.75 − q.25
Presenting a low dimensional statistic is useful if we believe that the statis-
tic is representative of the whole population. For instance, in Example 1.5,
oceanographers believe the data they have collected is representative of the
long term state of the ocean. Therefore the sample means at the nine locations
in Figure 1.12 are representative of the long term state of the ocean at those
locations. More formally, for each location we can imagine a population of tem-
peratures, one temperature for each moment in time. That population has an
unknown pdf f . Even though are data are not really a random sample from f
(The sampling times were not chosen randomly, among other problems.) we can
think of them that way without making too serious an error. The histograms in
Figure 1.12 are estimates of the f ’s for the nine locations. The mean of each f
is what oceanographers call a climatological mean, or an average which, because
it is taken over a long period of time, represents the climate. The nine sam-
ple means are estimates of the nine climatological mean temperatures at those
nine locations. Simply presenting the sample means reveals some interesting
structure in the data, and hence an interesting facet of physical oceanography.
Often, more than a simple data description or display is necessary; the statis-
tician has to do a bit of exploring the data set. This activity is called exploratory
data analysis or simply eda. It is hard to give general rules for eda, although
displaying the data in many different ways is often a good idea. The statistician
must decide what displays and eda are appropriate for each data set and each
question that might be answered by the data set. That is one thing that makes
statistics interesting. It cannot be reduced to a set of rules and procedures.
A good statistician must be attuned to the potentially unique aspects of each
analysis. We now present several examples to show just a few of the possible
ways to explore data sets by displaing them graphically. The examples reveal
some of the power of graphical display in illuminating data and teasing out what
it has to say.

2.2.2 Displaying Distributions

Instead of reducing a data set to just a few summary statistics, it is often helpful
to display the full data set. But reading a long list of numbers is usually not
helpful; humans are not good at assimilating data in that form. We can learn a
lot more from a graphical representation of the data.

Histograms The next examples uses histograms to display the full distribu-
tion of some data sets. Visual comparison of the histograms reveals structure
in the data.

Example 2.1 (Tooth Growth)

The R statistical language comes with many data sets. Type data() to see what
2.2. DATA DESCRIPTION 85

they are. This example works with the data set ToothGrowth on the effect of
vitamin C on tooth growth in guinea pigs. You can get a description by typing
help(ToothGrowth). You can load the data set into your R session by typing
data(ToothGrowth). ToothGrowth is a dataframe of three columns. The first
few rows look like this:
len supp dose
1 4.2 VC 0.5
2 11.5 VC 0.5
3 7.3 VC 0.5

Column 1, or len, records the amount of tooth growth. Column 2, supp, records
whether the guinea pig was given vitamin C in ascorbic acid or orange juice. Col-
umn 3, dose, records the dose, either 0.5, 1.0 or 2.0 mg. Thus there are six groups
of guinea pigs in a two by three layout. Each group has ten guinea pigs, for a total
of sixty observations. Figure 2.2 shows histograms of growth for each of the six
groups. From Figure 2.2 it is clear that dose affects tooth growth.

Figure 2.2 was produced by the following R code.

supp <- unique ( ToothGrowth$supp )
dose <- unique ( ToothGrowth$dose )
par ( mfcol=c(3,2) )
for ( i in 1:2 )
for ( j in 1:3 ) {
good <- ( ToothGrowth$supp == supp[i]
& ToothGrowth$dose == dose[j] )
hist ( ToothGrowth$len[good], breaks=seq(0,34,by=2),
xlab="", ylab="",
main=paste(supp[i], ", ", dose[j], sep="") )
}

• unique(x) returns the unique values in x. For example, if x <- c(1,1,2)

then unique(x) would be 1 2.

Figure 2.3 is similar to Figure 2.2 but laid out in the other direction. (Notice
that it’s easier to compare histograms when they are arranged vertically rather than
horizontally.) The figures suggest that delivery method does have an effect, but not
as strong as the dose effect. Notice also that Figure 2.3 is more difficult to read
than Figure 2.2 because the histograms are too tall and narrow. Figure reffig:tooth3
repeats Figure 2.3 but using less vertical distance; it is therefore easier to read. Part
of good statistical practice is displaying figures in a way that makes them easiest to
read and interpret.
The figures alone have suggested that dose is the most important effect, and
delivery method less so. A further analysis could try to be more quantitative: what
86 CHAPTER 2. MODES OF INFERENCE

VC, 0.5 OJ, 0.5

3.0

5
4
2.0

3
2
1.0

1
0.0

0
0 5 10 20 30 0 5 10 20 30

VC, 1 OJ, 1

0.0 0.5 1.0 1.5 2.0

4
3
2
1
0

0 5 10 20 30 0 5 10 20 30

VC, 2 OJ, 2
0.0 0.5 1.0 1.5 2.0

3.0
2.0
1.0
0.0

0 5 10 20 30 0 5 10 20 30

Figure 2.2: Histograms of tooth growth by delivery method (VC or OJ) and
dose (0.5, 1.0 or 2.0).
2.2. DATA DESCRIPTION 87

3.0 VC, 0.5 VC, 1 VC, 2

2.0
4
2.5

1.5
3
2.0
1.5

1.0
2
1.0

0.5
1
0.5
0.0

0.0
0

0 10 25 0 10 25 0 10 25

OJ, 0.5 OJ, 1 OJ, 2

2.0

3.0
5

2.5
4

1.5

2.0
3

1.0

1.5
2

1.0
0.5
1

0.5
0.0

0.0
0

0 10 25 0 10 25 0 10 25

Figure 2.3: Histograms of tooth growth by delivery method (VC or OJ) and
dose (0.5, 1.0 or 2.0).
88 CHAPTER 2. MODES OF INFERENCE

VC, 0.5 VC, 1 VC, 2

3.0

2.0
4
3
1.5

1.0
2
1
0.0

0.0
0

0 10 25 0 10 25 0 10 25

OJ, 0.5 OJ, 1 OJ, 2

2.0

3.0
4

1.0

1.5
2

0.0

0.0
0

0 10 25 0 10 25 0 10 25

Figure 2.4: Histograms of tooth growth by delivery method (VC or OJ) and
dose (0.5, 1.0 or 2.0).
2.2. DATA DESCRIPTION 89

is the typical size of each effect, how sure can we be of the typical size, and how
much does the effect vary from animal to animal. The figures already suggest
answers, but a more formal analysis is deferred to Section 2.7.

Figures 1.12, 2.2, and 2.3 are histograms. The abscissa has the same scale as
the data. The data are divided into bins. The ordinate shows the number of data
points in each bin. (hist(...,prob=T) plots the ordinate as probability rather
than counts.) Histograms are a powerful way to display data because they give
a strong visual impression of the main features of a data set. However, details
of the histogram can depend on both the number of bins and on the cut points
between bins. For that reason it is sometimes better to use a display that does
not depend on those features, or at least not so strongly. Example 2.2 illustrates.

Density Estimation

Example 2.2 (Hot Dogs)

In June of 1986, Consumer Reports published a study of hot dogs. The data are
available at DASL, the Data and Story Library, a collection of data sets for free use
by statistics students. DASL says the data are
“Results of a laboratory analysis of calories and sodium content of major
hot dog brands. Researchers for Consumer Reports analyzed three types
of hot dog: beef, poultry, and meat (mostly pork and beef, but up to
15% poultry meat).”
You can download the data from http://lib.stat.cmu.edu/DASL/Datafiles/
Hotdogs.html. The first few lines look like this:
Type Calories Sodium
Beef 186 495
Beef 181 477

This example looks at the calorie content of beef hot dogs. (Later examples will
compare the calorie contents of different types of hot dogs.)
Figure 2.5(a) is a histogram of the calorie contents of beef hot dogs in the
study. From the histogram one might form the impression that there are two major
varieties of beef hot dogs, one with about 130–160 calories or so, another with about
180 calories or so, and a rare outlier with fewer calories. Figure 2.5(b) is another
histogram of the same data but with a different bin width. It gives a different
impression, that calorie content is evenly distributed, approximately, from about
130 to about 190 with a small number of lower calorie hot dogs. Figure 2.5(c)
gives much the same impression as 2.5(b). It was made with the same bin width
as 2.5(a), but with cut points starting at 105 instead of 110. These histograms
illustrate that one’s impression can be influenced by both bin width and cut points.
Density estimation is a method of reducing dependence on cut points. Let
x1 , . . . , x20 be the calorie contents of beef hot dogs in the study. We think of
x1 , . . . , x20 as a random sample from a density f representing the population of all
90 CHAPTER 2. MODES OF INFERENCE

beef hot dogs. Our goal is to estimate f . For any fixed number x, how shall we
estimate f (x)? The idea is to use information local to x to estimate f (x). We first
describe a basic version, then add two refinements to get kernel density estimation
and the density() function in R.
Let n be the sample size (20 for the hot dog data). Begin by choosing a number
h > 0. For any number x the estimate fˆbasic (x) is defined to be
n
1 X fraction of sample points within h of x
fˆbasic (x) ≡ 1(x−h,x+h)(xi ) =
2nh i=1 2h

fˆbasic has at least two apparently undesirable features.

1. fˆbasic (x) gives equal weight to all data points in the interval (x − h, x + h)
and has abrupt cutoffs at the ends of the interval. It would be better to give
the most weight to data points closest to x and have the weights decrease
gradually for points increasingly further away from x.
2. fˆbasic (x) depends critically on the choice of h.
We deal with these problems by introducing a weight function that depends on
distance from x. Let g0 be a probability density function. Usually g0 is chosen to
be symmetric and unimodal, centered at 0. Define
1X
fˆ(x) ≡ g0 (x − xi )
n
Choosing g0 to be a probability density ensures that fˆ is also a probability density
because Z ∞ XZ ∞
ˆ dx = 1
f(x) g0 (x − xi ) dx = 1 (2.1)
−∞ n i −∞
When g0 is chosen to be a continuous function it deals nicely with problem 1 above.
In fact, fˆbasic comes from taking g0 to be the uniform density on (−h, h).
To deal with problem 2 we rescale g0 . Choose a number h > 0 and define a
new density g(x) = h−1 g0 (x/h). A little thought shows that g differsR from g0 by a
rescaling of the horizontal axis; the factor h−1 compensates to make g = 1. Now
define the density estimate to be
1X 1 X
fˆh (x) ≡ g(x − xi ) = g0 ((x − xi )/h)
n nh
h is called the bandwidth. Of course fˆh still depends on h. It turns out that
dependence on bandwidth is not really a problem. It is useful to view density
estimates for several different bandwidths. Each reveals features of f at different
scales. Figures 2.5(d), (e), and (f) are examples. Panel (d) was produced by the
default bandwidth; panels (e) and (f) were produced with 1/4 and 1/2 the default
bandwidth. Larger bandwidth makes a smoother estimate of f ; smaller bandwidth
makes it rougher. None is exactly right. It is useful to look at several.
2.2. DATA DESCRIPTION 91

(a) (b)
5

8
4

6
3

4
2

2
1
0

0
120 140 160 180 120 140 160 180

calories calories

(c) (d)
5
4

0.010
density
3
2
1

0.000
0

120 140 160 180 100 150 200

calories calories

(e) (f)
0.030

0.020
density

density
0.015

0.010
0.000

0.000

100 140 180 100 140 180

calories calories

Figure 2.5: (a), (b), (c): histograms of calorie contents of beef hot dogs; (d),
(e), (f ): density estimates of calorie contents of beef hot dogs.
92 CHAPTER 2. MODES OF INFERENCE

Figure 2.5 was produced with

hotdogs <- read.table ( "data/hotdogs/data", header=T )
cal.beef <- hotdogs$Calories [ hotdogs$Type == "Beef" ]
par ( mfrow=c(3,2) )
hist ( cal.beef, main="(a)", xlab="calories", ylab="" )
hist ( cal.beef, breaks=seq(110,190,by=20), main="(b)",
xlab="calories", ylab="" )
hist ( cal.beef, breaks=seq(105,195,by=10), main="(c)",
xlab="calories", ylab="" )

plot ( density ( cal.beef ), main="(d)", xlab="calories",

ylab="density" )
plot ( density ( cal.beef, adjust=1/4 ), main="(e)",
xlab="calories", ylab="density" )
plot ( density ( cal.beef, adjust=1/2 ), main="(f)",
xlab="calories", ylab="density" )

• In panel (a) R used its default method for choosing histogram bins.
• In panels (b) and (c) the histogram bins were set by
hist ( ..., breaks=seq(...) ).
• density() produces a kernel density estimate.
• R uses a Gaussian kernel by default which means that g 0 above is the N(0, 1)
density.
• In panel (d) R used its default method for choosing bandwidth.
• In panels (e) and (f) the bandwidth was set to 1/4 and 1/2 the default by
density(..., adjust=...).

Stripcharts and Dotplots Figure 2.6 uses the ToothGrowth data to illus-
trate stripcharts, also called dotplots, an alternative to histograms. In the top
panel there are three rows of points corresponding to the three doses of ascorbic
acid. Each point is for one animal. The abscissa shows the amount of tooth
growth; the ordinate shows the dose. The panel is slightly misleading because
points with identical coordinates are plotted directly on top of each other. In
such situations statisticians often add a small amount of jitter to the data, to
avoid overplotting. The middle panel is a repeat of the top, but with jitter
added. The bottom panel shows tooth growth by delivery method. Compare
Figure 2.6 to Figures 2.2 and 2.3. Which is a better display for this particular
data set?
2.2. DATA DESCRIPTION 93

2 (a)
dose

1
0.5

5 10 15 20 25 30 35

growth

(b)
2
dose

1
0.5

5 10 15 20 25 30 35

growth

(c)
VC
method

OJ

5 10 15 20 25 30 35

growth

Figure 2.6: (a) Tooth growth by dose, no jittering; (b) Tooth growth by dose
with jittering; (c) Tooth growth by delivery method with jittering
94 CHAPTER 2. MODES OF INFERENCE

Figure 2.6 was produced with the following R code.

par ( mfrow=c(3,1) )
stripchart ( ToothGrowth$len ~ ToothGrowth$dose, pch=1,
main="(a)", xlab="growth", ylab="dose" )
stripchart ( ToothGrowth$len ~ ToothGrowth$dose,
method="jitter", main="(b)", xlab="growth",
ylab="dose", pch=1 )
stripchart ( ToothGrowth$len ~ ToothGrowth$supp,
method="jitter", main="(c)", xlab="growth",
ylab="method", pch=1 )

Boxplots An alternative, useful for comparing many distributions simultane-

ously, is the boxplot. Example 2.3 uses boxplots to compare scores on 24 quizzes
in a statistics course.

Example 2.3 (Quiz Scores)

In the spring semester of 2003, 58 students completed Statistics 103 at Duke Uni-
versity. Figure 2.7 displays their grades.
There were 24 quizzes during the semester. Each was worth 10 points. The
upper panel of the figure shows the distribution of scores on each quiz. The abscissa
is labelled 1 through 24, indicating the quiz number. For each quiz, the figure shows
a boxplot. For each quiz there is a box. The horizontal line through the center of
the box is the median grade for that quiz. We can see that the median score on
Quiz 2 is around 7, while the median score on Quiz 3 is around 4. The upper end
of the box is the 75th percentile (3rd quartile) of scores; the lower end of the box is
the 25th percentile (1st quartile). We can see that about half the students scored
between about 5 and 8 on Quiz 2, while about half the students scored between
about 2 and 6 on Quiz 3. Quiz 3 was tough.
Each box may have whiskers, or dashed lines which extend above and below the
box. The exact definition of the whiskers is not important, but they are meant to
include most of the data points that don’t fall inside the box. (In R, by default, the
whiskers extend to the most extreme data point which is no more that 1.5 times the
interquartile range away from the median.) Finally, there may be some individual
points plotted above or below each boxplot. These indicate outliers, or scores that
are extremely high or low relative to other scores on that quiz. Many quizzes had
low outliers; only Quiz 5 had a high outlier.
Box plots are extremely useful for comparing many sets of data. We can easily
see, for example, that Quiz 5 was the most difficult (75% of the class scored 3 or
less.) while Quiz 1 was the easiest (over 75% of the class scored 10.)
There were no exams or graded homeworks. Students’ grades were determined
by their best 20 quizzes. To compute grades, each student’s scores were sorted, the
first 4 were dropped, then the others were averaged. Those averages are displayed
in a stripchart in the bottom panel of the figure. It’s easy to see that most of the
2.2. DATA DESCRIPTION 95

class was scored between about 5 and 9 and that 4 scores were much lower than
the rest of the class.

Figure 2.7 was produced by the following R code.

... # read in the data
colnames(scores) <- paste("Q",1:24,sep="")
# define column names
boxplot ( data.frame(scores), main="Individual quizzes" )

scores[is.na(scores)] <- 0 # replace missing scores

# with 0’s
temp <- apply ( scores, 1, sort ) # sort
temp <- temp[5:24,] # drop the 4 lowest
scores.ave <- apply ( temp, 2, mean ) # find the average

stripchart ( scores.ave, "jitter", pch=1, xlab="score",

xlim=c(0,10), main="Student averages" )

QQ plots Sometimes we want to assess whether a data set is well modelled

by a Normal distribution and, if not, how it differs from Normal. One obvious
way to assess Normality is by looking at histograms or density estimates. In
fact, the answer is often not obvious from the figure. A better way to assess
Normality is with QQ plots. Figure 2.8 illustrates for the nine histograms of
ocean temperatures in Figure 1.12.
Each panel in Figure 2.8 was created with the ocean temperatures near a
particular (latitude, longitude) combination. Consider, for example, the upper
right panel which was constructed from the n = 213 points x1 , . . . , x213 taken
near (45, -40). Those points are sorted, from smallest to largest, to create the or-
der statistic (x(1) , . . . , x(213) ). Then they are plotted against E[(z(1) , . . . , z(213) )],
the expected order statistic from a Normal distribution. If the xi s are approx-
imately Normal then the QQ plot will look approximately linear. The slope of
the line indicates the standard deviation.
In Figure 2.8 most of the panels do look approximately linear, indicating
that a Normal model is reasonable. But some of the panels show departures
from Normality. In the upper left and lower left panels, for example, the plots
looks roughly linear except for the upper right corners which show some data
points much warmer than expected if they followed a Normal distribution. In
contrast, the coolest temperatures in the lower middle panel are not quite as
cool as expected from a Normal distribution.

Figure 2.8 was produced with

lats <- c ( 45, 35, 25 )
96 CHAPTER 2. MODES OF INFERENCE

Individual quizzes

10
8
6
4
2
0

Q1 Q4 Q7 Q11 Q15 Q19 Q23

Student averages

0 2 4 6 8 10

score

Figure 2.7: Quiz scores from Statistics 103

2.2. DATA DESCRIPTION 97

latitude = 45 latitude = 45 latitude = 45

longitude = −40 longitude = −30 longitude = −20

10
9

6.5
8

9
7

8
5.5
6

7
5

6
4.5
4

−3 −1 1 3 −2 0 2 −2 0 2

n = 213 n = 105 n = 112

latitude = 35 latitude = 35 latitude = 35

longitude = −40 longitude = −30 longitude = −20

10.5
9.5
7.5

9.0

9.5
8.5
6.5

8.0

8.5
5.5

−2 0 2 −2 0 1 2 −2 0 2

n = 37 n = 24 n = 44

latitude = 25 latitude = 25 latitude = 25

longitude = −40 longitude = −30 longitude = −20
7.6
8.0
8.5

7.5

7.2
7.5

7.0

6.8
6.5

6.5

−2 0 2 −2 0 2 −2 0 2

n = 47 n = 35 n = 27

Figure 2.8: QQ plots of water temperatures (◦ C) at 1000m depth

98 CHAPTER 2. MODES OF INFERENCE

lons <- c ( -40, -30, -20 )

par ( mfrow=c(3,3) )
for ( i in 1:3 )
for ( j in 1:3 ) {
good <- abs ( med.1000$lon - lons[j] ) < 1 &
abs ( med.1000$lat - lats[i] ) < 1
qqnorm ( med.1000$temp[good], xlab="", ylab="",
sub=paste("n = ", sum(good), sep=""),
main = paste ( "latitude =", lats[i], "\n longitude =",
lons[j] ))
}

2.2.3 Exploring Relationships

Sometimes it is the relationships among several random variables that is of
interest. For example, in discrimination cases the focus is on the relationship
between race or gender on one hand and employment or salary on the other
hand. Subsection 2.2.3 shows several graphical ways to display relationships.
We begin with Example 2.4, an analysis of potential discrimination in ad-
mission to UC Berkely graduate school.

Example 2.4
In 1973 UC Berkeley investigated its graduate admissions rates for potential sex
bias. Apparently women were more likely to be rejected than men. The data set
UCBAdmissions gives the acceptance and rejection data from the six largest grad-
uate departments on which the study was based. Typing help(UCBAdmissions)
tells more about the data. It tells us, among other things:

...
Format:

A 3-dimensional array resulting from cross-tabulating 4526

observations on 3 variables. The variables and their levels
are as follows:

No Name Levels
1 Admit Admitted, Rejected
2 Gender Male, Female
3 Dept A, B, C, D, E, F
...

The major question at issue is whether there is sex bias in admissions. To investigate
we ask whether men and women are admitted at roughly equal rates.
Typing UCBAdmissions gives the following numerical summary of the data.
2.2. DATA DESCRIPTION 99

, , Dept = A

Gender
Admit Male Female
Admitted 512 89
Rejected 313 19

, , Dept = B

Gender
Admit Male Female
Admitted 353 17
Rejected 207 8

, , Dept = C

Gender
Admit Male Female
Admitted 120 202
Rejected 205 391

, , Dept = D

Gender
Admit Male Female
Admitted 138 131
Rejected 279 244

, , Dept = E

Gender
Admit Male Female
Admitted 53 94
Rejected 138 299

, , Dept = F

Gender
Admit Male Female
Admitted 22 24
Rejected 351 317

For each department, the twoway table of admission status versus sex is dis-
played. Such a display, called a crosstabulation, simply tabulates the number of
entries in each cell of a multiway table. It’s hard to tell from the crosstabulation
whether there is a sex bias and, if so, whether it is systemic or confined to just a few
100 CHAPTER 2. MODES OF INFERENCE

departments. Let’s continue by finding the marginal (aggregated by department as

opposed to conditional given deparment) admissions rates for men and women.
> apply(UCBAdmissions, c(1, 2), sum)
Gender
Admit Male Female
Admitted 1198 557
Rejected 1493 1278
The admission rate for men is 1198/(1198 + 1493) = 44.5% while the admission
rate for women is 557/(557 + 1493) = 30.4%, much lower. A mosaic plot, created
with
mosaicplot(apply(UCBAdmissions, c(1, 2), sum),
main = "Student admissions at UC Berkeley")
is a graphical way to display the discrepancy. The left column is for admitted
students; the heights of the rectangles show how many admitted students were male
and how many were female. The right column is for rejected students; the heights
of the rectangles show how many were male and female. If sex and admission status
were independent, i.e., if there were no sex bias, then the proportion of men among
admitted students would equal the proportion of men among rejected students and
the heights of the left rectangles would equal the heights of the right rectangles.
The apparent difference in heights is a visual representation of the discrepancy in
sex ratios among admitted and rejected students. The same data can be viewed as
discrepant admission rates for men and women by transposing the matrix:
mosaicplot(t(apply(UCBAdmissions, c(1, 2), sum)),
main = "Student admissions at UC Berkeley")
The existence of discrepant sex ratios for admitted and rejected students is
equivalent to the existence of discrepant admission rates for males and females and
to dependence of sex and admission rates. The lack of discrepant ratios is equivalent
to independence of sex and admission rates.
Evidently UC Berkeley admitted men and women at different rates. But graduate
admission decisions are not made by a central admissions office; they are made by
the individual departments to which students apply. So our next step is to look at
admission rates for each department separately. We can look at the crosstabulation
on page 98 or make mosaic plots for each department separately (not shown here)
with
## Mosaic plots for individual departments
for(i in 1:6)
mosaicplot(UCBAdmissions[,,i],
xlab = "Admit", ylab = "Sex",
main = paste("Department", LETTERS[i]))
The plots show that in each department men and women are admitted at roughly
equal rates. The following snippet calculates and prints the rates. It confirms the
2.2. DATA DESCRIPTION 101

Student admissions at UC Berkeley

Admitted Rejected
Male
Gender
Female

Admit

Figure 2.9: Mosaic plot of UCBAdmissions

102 CHAPTER 2. MODES OF INFERENCE

Student admissions at UC Berkeley

Male Female

Admitted
Admit
Rejected

Gender

Figure 2.10: Mosaic plot of UCBAdmissions

2.2. DATA DESCRIPTION 103

rough equality except for department A which admitted women at a higher rate
than men.

for ( i in 1:6 ) { # for each department

temp <- UCBAdmissions[,,i] # that department’s data
m <- temp[1,1] / (temp[1,1]+temp[2,1]) # Men’s admission rate
w <- temp[1,2] / (temp[1,2]+temp[2,2]) # Women’s admission rate
print ( c ( m, w ) ) # print them
}

Note that departments A and B which had high admission rates also had large num-
bers of male applicants while departments C, D, E and F which had low admission
rates had large numbers of female applicants. The generally accepted explanation
for the discrepant marginal admission rates is that men tended to apply to depart-
ments that were easy to get into while women tended to apply to departments that
were harder to get into. A more sinister explanation is that the university gave
more resources to departments with many male applicants, allowing them to admit
a greater proportion of their applicants. The data we’ve analyzed are consistent
with both explanations; the choice between them must be made on other grounds.
One lesson here for statisticians is the power of simple data displays and sum-
maries. Another is the need to consider the unique aspects of each data set. The
explanation of different admissions rates for men and women could only be discov-
ered by someone familiar with how universities and graduate schools work, not by
following some general rules about how to do statistical analyses.

The next example is about the duration of eruptions and interval to the next
eruption of the Old Faithful geyser. It explores two kinds of relationships —
the relationship between duration and eruption and also the relationship of each
variable with time.

Example 2.5 (Old Faithful)

Old Faithful is a geyser in Yellowstone National Park and a great tourist attraction.
As Denby and Pregibon [1987] explain, “From August 1 to August 8, 1978, rangers
and naturalists at Yellowstone National Park recorded the duration of eruption and
interval to the next eruption (both in minutes) for eruptions of Old Faithful between
6 a.m. and midnight. The intent of the study was to predict the time of the next
eruption, to be posted at the Visitor’s Center so that visitors to Yellowstone can
usefully budget their time.” The R dataset faithful contains the data. In addition
to the references listed there, the data and analyses can also be found in Weisberg
[1985] and Denby and Pregibon [1987]. The latter analysis emphasizes graphics,
and we shall follow some of their suggestions here.
We begin exploring the data with histograms of durations and intervals. These
are shown in Figure 2.11. The histograms suggest bimodal distributions. For dura-
tion there seems to be one bunch of data around two minutes and another around
four or five minutes. For interval, the modes are around 50 minutes and 80 min-
utes. A plot of interval versus duration, Figure 2.12, suggests that the bimodality
is present in the joint distribution of the two variables. Because the data were
104 CHAPTER 2. MODES OF INFERENCE

collected over time, it might be useful to plot the data in the order of collection.
That’s Figure 2.13. The horizontal scale in Figure 2.13 is so compressed that it’s
hard to see what’s going on. Figure 2.14 repeats Figure 2.13 but divides the time
interval into two subintervals to make the plots easier to read. The subintervals
overlap slightly. The persistent up-and-down character of Figure 2.14 shows that,
for the most part, long and short durations are interwoven, as are long and short
intervals. (Figure 2.14 is potentially misleading. The data were collected over an
eight day period. There are eight separate sequences of eruptions with gaps in
between. The faithful data set does not tell us where the gaps are. Denby and
Pregibon [1987] tell us where the gaps are and use the eight separate days to find
errors in data transcription.) Just this simple analysis, a collection of four figures,
has given us insight into the data that will be very useful in predicting the time of
the next eruption.
Figures 2.11, 2.12, 2.13, and 2.14 were produced with the following R code.
data(faithful)
attach(faithful)

par ( mfrow=c(2,1) )
hist ( eruptions, prob=T, main="a" )
hist ( waiting, prob=T, main="b" )

par ( mfrow=c(1,1) )
plot ( eruptions, waiting, xlab="duration of eruption",
ylab="time to next eruption" )

par ( mfrow=c(2,1) )
plot.ts ( eruptions, xlab="data number", ylab="duration",
main="a" )
plot.ts ( waiting, xlab="data number", ylab="waiting time",
main="b" )

par ( mfrow=c(4,1) )
plot.ts ( eruptions[1:150], xlab="data number",
ylab="duration", main="a1" )
plot.ts ( eruptions[130:272], xlab="data number",
ylab="duration", main="a2" )
plot.ts ( waiting[1:150], xlab="data number",
ylab="waiting time", main="b1")
plot.ts ( waiting[130:272], xlab="data number",
ylab="waiting time", main="b2")

Figures 2.15 and 2.16 introduce coplots, a tool for visualizing the relation-
ship among three variables. They represent the ocean temperature data from
2.2. DATA DESCRIPTION 105

0.4 a
Density

0.2
0.0

2 3 4 5

eruptions

b
0.04
Density

0.02
0.00

40 50 60 70 80 90 100

waiting

Figure 2.11: Histograms of (a): durations of eruptions and (b): waiting time
until the next eruption in the Old Faithful dataset
106 CHAPTER 2. MODES OF INFERENCE

90
80
time to next eruption

70
60
50

1.5 2.0 2.5 3.0 3.5 4.0 4.5 5.0

duration of eruption

Figure 2.12: Waiting time versus duration in the Old Faithful dataset
2.2. DATA DESCRIPTION 107

4.5 a
duration

3.5
2.5
1.5

0 50 100 150 200 250

data number

b
90
waiting time

70
50

0 50 100 150 200 250

data number

Figure 2.13: (a): duration and (b): waiting time plotted against data number
in the Old Faithful dataset
108 CHAPTER 2. MODES OF INFERENCE

a1

duration

3.5
1.5

0 50 100 150

data number

a2
duration

4.0
2.0

0 20 40 60 80 100 120 140

data number

b1
waiting time

80
50

0 50 100 150

data number

b2
waiting time

80
50

0 20 40 60 80 100 120 140

data number

Figure 2.14: (a1), (a2): duration and (b1), (b2): waiting time plotted against
data number in the Old Faithful dataset
2.3. LIKELIHOOD 109

Example 1.5. In Figure 2.15 there are six panels in which temperature is plotted
against latitude. Each panel is made from the points in a restriced range of lon-
gitude. The upper panel, the one spanning the top of the Figure, shows the six
different ranges of longitude. For example, the first longitude range runs from
about -10 to about -17. Points whose longitude is in the interval (−17, −10) go
into the upper right panel of scatterplots. These are the points very close to
the mouth of the Mediterranean Sea. Looking at that panel we see that tem-
perature increases very steeply from South to North, until about 35◦ , at which
point they start to decrease steeply as we go further North. That’s because
we’re crossing the Mediterranean tongue at a point very close to its source.
The other longitude ranges are about (−20, −13), (−25, −16), (−30, −20),
(−34, −25) and (−40, −28). They are used to create the scatterplot panels in the
upper center, upper left, lower right, lower center, and lower right, respectively.
The general impression is

• temperatures decrease slightly as we move East to West,

• the angle in the scatterplot becomes slightly shallower as we move East

to West, and

• there are some points that don’t fit the general pattern.

Notice that the longitude ranges are overlapping and not of equal width. The
ranges are chosen by R to have a little bit of overlap and to put roughly equal
numbers of points into each range.
Figure 2.16 reverses the roles of latitude and longitude. The impression is
that temperature increases gradually from West to East. These two figures give
a fairly clear picture of the Mediterranean tongue.

Figures 2.15 and 2.16 were produced by

coplot ( temp ~ lat | lon )
coplot ( temp ~ lon | lat )

2.3 Likelihood
2.3.1 The Likelihood Function
It often happens that we observe data from a distribution that is not known
precisely but whose general form is known. For example, we may know that
the data come from a Poisson distribution, X ∼ Poi(λ), but we don’t know the
value of λ. We may know that X ∼ Bin(n, θ) but not know θ. Or we may know
that the values of X are densely clustered around some central value and sparser
on both sides, so we decide to model X ∼ N(µ, σ), but we don’t know the values
of µ and σ. In these cases there is a whole family of probability distributions
110 CHAPTER 2. MODES OF INFERENCE

Given : lon
−40 −35 −30 −25 −20 −15 −10

20 30 40 50 20 30 40 50

15
10
5
temp
15
10
5

20 30 40 50

lat

Figure 2.15: Temperature versus latitude for different values of longitude

2.3. LIKELIHOOD 111

Given : lat
20 25 30 35 40 45 50

−40 −30 −20 −10 −40 −30 −20 −10

15
10
5
temp
15
10
5

−40 −30 −20 −10

lon

Figure 2.16: Temperature versus longitude for different values of latitude

112 CHAPTER 2. MODES OF INFERENCE

indexed by either λ, θ, or (µ, σ). We call λ, θ, or (µ, σ) the unknown parameter ;

the family of distributions is called a parametric family. Often, the goal of the
statistical analysis is to learn about the value of the unknown parameter. Of
course, learning which value of the parameter is the true one, or which values
of the parameter are plausible in light of the data, is the same as learning
which member of the family is the true one, or which members of the family are
plausible in light of the data.
The different values of the parameter, or the different members of the fam-
ily, represent different theories or hypotheses about nature. A sensible way to
discriminate among the theories is according to how well they explain the data.
Recall the Seedlings data (Examples 1.4, 1.6, 1.7 and 1.9) in which X was the
number of new seedlings in a forest quadrat, X ∼ Poi(λ), and different values
of λ represent different theories or hypotheses about the arrival rate of new
seedlings. When X turned out to be 3, how well a value of λ explains the data
is measured by Pr[X = 3 | λ]. This probability, as a function of λ, is called the
likelihood function and denoted `(λ). It says how well each value of λ explains
the datum X = 3. Figure 1.6 (pg. 18) is a plot of the likelihood function.
In a typical problem we know the data come from a parametric family in-
dexed by a parameter θ, i.e. X1 , . . . , Xn ∼ i.i.d.f (x | θ), but we don’t know θ.
The joint density of all the data is
Y
f (X1 , . . . , Xn | θ) = f (Xi | θ). (2.2)

Equation 2.2, as a function of θ, is the likelihood function. We sometimes write

f (Data | θ) instead of indicating each individual datum. To emphasize that we
are thinking of a function of θ we may also write the likelihood function as `(θ)
or `(θ | Data).
The interpretation of the likelihood function is always in terms of ratios.
If, for example, `(θ1 )/`(θ2 ) > 1, then θ1 explains the data better than θ2 . If
`(θ1 )/`(θ2 ) = k, then θ1 explains the data k times better than θ2 . To illustrate,
suppose students in a statistics class conduct a study to estimate the fraction
of cars on Campus Drive that are red. Student A decides to observe the first 10
cars and record X, the number that are red. Student A observes

N R, R, N R, N R, N R, R, N R, N R, N R, R

and records X = 3. She did a Binomial experiment;
3her statistical model is

X ∼ Bin(10, θ); her likelihood function is `A (θ) = 10 7
3 θ (1 − θ) . It is plotted in
Figure 2.17. Because only ratios matter, the likelihood function can be rescaled
by any arbitrary positive constant. In Figure 2.17 it has been rescaled so that
the maximum is 1. The interpretation of Figure 2.17 is that values of θ around
θ ≈ 0.3 explain the data best, but that any value of θ in the interval from
about 0.1 to about 0.6 explains the data not too much worse than the best. I.e.,
θ ≈ 0.3 explains the data only about 10 times better than θ ≈ 0.1 or θ ≈ 0.6,
and a factor of 10 is not really very much. On the other hand, values of θ less
than about 0.05 or greater than about 0.7 explain the data much worse than
θ ≈ 0.3.
2.3. LIKELIHOOD 113

1.0
0.8
likelihood function

0.6
0.4
0.2
0.0

0.0 0.2 0.4 0.6 0.8 1.0

Figure 2.17: Likelihood function `(θ) for the proportion θ of red cars on Campus
Drive
114 CHAPTER 2. MODES OF INFERENCE

Figure 2.17 was produced by the following snippet.

theta <- seq ( 0, 1, by=.01 ) # some values of theta
y <- dbinom ( 3, 10, theta ) # calculate l(theta)
y <- y / max(y) # rescale
plot ( theta, y, type="l", xlab=expression(theta),
ylab="likelihood function" )

• expression is R’s way of getting mathematical symbols and formulae into

plot labels. For more information, type help(plotmath).

To continue the example, Student B decides to observe cars until the third
red one drives by and record Y , the total number of cars that drive by until the
third red one. Students A and B went to Campus Drive at the same time and
observed the same cars. B records Y = 10. For B the likelihood function is
`B (θ) = P[Y = 10 | θ]
= P[2 reds among first 9 cars] × P[10’th car is red]
 
9 2
= θ (1 − θ)7 × θ
2
 
9 3
= θ (1 − θ)7
2



`B differs from `A by the multiplicative constant 92 / 10 3 . But since multiplica-
tive constants don’t matter, A and B really have the same likelihood function
and hence exactly the same information about θ. Student B would also use
Figure 2.17 as the plot of her likelihood function.
Student C decides to observe every car for a period of 10 minutes and record
Z1 , . . . , Zk where k is the number of cars that drive by in 10 minutes and
each Zi is either 1 or 0 according to whether the i’th car is red. When C went
to Campus Drive with A and B, only 10 cars drove by in the first 10 minutes.
Therefore C recorded exactly the same data as A and B. Her likelihood function
is

`C (θ) = (1 − θ)θ(1 − θ)(1 − θ)(1 − θ)θ(1 − θ)(1 − θ)(1 − θ)θ = θ 3 (1 − θ)7

`C is proportional to `A and `B and hence contains exactly the same information

and looks exactly like Figure 2.17. So even though the students planned different
experiments they ended up with the same data, and hence the same information
about θ.
The next example follows the Seedling story and shows what happens to the
likelihood function as data accumulates.

Example 2.6 (Seedlings, cont.)

Examples 1.4, 1.6, 1.7, and 1.9 reported data from a single quadrat on the number
2.3. LIKELIHOOD 115

of new seedlings to emerge in a given year. In fact, ecologists collected data from
multiple quadrats over multiple years. In the first year there were 60 quadrats and
a total of 40 seedlings so the likelihood function was
`(λ) ≡ p(Data | λ)
= p(y1 , . . . , y60 | λ)
60
Y
= p(yi | λ)
1
60
Y e−λ λyi
=
1
yi !
−60λ 40
∝e λ
Q
Note that yi ! is a multiplicative factor that does not
P depend on λ and so is
irrelevant to `(λ). Note also thatP `(λ) depends only on yi , not on the individual
yi ’s. I.e., we only need to know yi = 40; we don’t need to know the individual
yi ’s. `(λ) is plotted in Figure 2.18. Compare to Figure 1.6 (pg. 18). Figure 2.18 is
much more peaked. That’s because it reflects much more information, 60 quadrats
instead of 1. The extra information pins down the value of λ much more accurately.
0.8
likelihood

0.4
0.0

0.0 0.5 1.0 1.5 2.0

P
Figure 2.18: `(θ) after yi = 40 in 60 quadrats.

Figure 2.18 was created with

lam <- seq ( 0, 2, length=50 )
116 CHAPTER 2. MODES OF INFERENCE

lik <- dpois ( 40, 60*lam )

lik <- lik / max(lik)
plot ( lam, lik, xlab=expression(lambda),
ylab="likelihood", type="l" )

The next example is about a possible cancer cluster in California.

Example 2.7 (Slater School)

This example was reported in Brodeur [1992]. See Lavine [1999] for further analysis.
The Slater school is an elementary school in Fresno, California where teachers
and staff were “concerned about the presence of two high-voltage transmission
lines that ran past the school . . . .” Their concern centered on the “high incidence
of cancer at Slater. . . .” To address their concern, Dr. Raymond Neutra of the
California Department of Health Services’ Special Epidemiological Studies Program
conducted a statistical analysis on the

“eight cases of invasive cancer, . . . , the total years of employment

of the hundred and forty-five teachers, teachers’ aides, and staff mem-
bers, . . . , [and] the number of person-years in terms of National Cancer
Institute statistics showing the annual rate of invasive cancer in Amer-
ican women between the ages of forty and forty-four — the age group
encompassing the average age of the teachers and staff at Slater —
[which] enabled him to calculate that 4.2 cases of cancer could have
been expected to occur among the Slater teachers and staff members
. . . .”

For our purposes we can assume that X, the number of invasive cancer cases
at the Slater School has the Binomial distribution X ∼ Bin(145, θ). We observe
x = 8. The likelihood function

`(θ) ∝ θ8 (1 − θ)137 (2.3)

is pictured in Figure 2.19. From the Figure it appears that values of θ around .05
or .06, explain the data better than values less than .05 or greater than .06, but
that values of θ anywhere from about .02 or .025 up to about .11 explain the data
reasonably well.

Figure 2.19 was produced by the following R code.

theta <- seq ( 0, .2, length=100 )
lik <- dbinom ( 8, 145, theta )
lik <- lik / max(lik)
plot ( theta, lik, xlab=expression(theta),
ylab="likelihood", type="l", yaxt="n" )
2.3. LIKELIHOOD 117

likelihood

0.00 0.05 0.10 0.15 0.20

Figure 2.19: Likelihood for Slater School

The first line of code creates a sequence of 100 values of θ at which to compute
`(θ), the second line does the computation, the third line rescales so the maximum
likelihood is 1, and the fourth line makes the plot.

Examples 2.6 and 2.7 show how likelihood functions are used. They reveal
which values of a parameter the data support (equivalently, which values of a
parameter explain the data well) and values they don’t support (which values
explain the data poorly). There is no hard line between support and non-
support. Rather, the plot of the likelihood functions shows the smoothly varying
levels of support for different values of the parameter.
Because likelihood ratios measure the strength of evidence for or against one
hypothesis as opposed to another, it is important to ask how large a likelihood
ratio needs to be before it can be considered strong evidence. Or, to put it
another way, how strong is the evidence in a likelihood ratio of 10, or 100, or
1000, or more? One way to answer the question is to construct a reference
experiment, one in which we have an intuitive understanding of the strength of
evidence and can calculate the likelihood; then we can compare the calculated
likelihood to the known strength of evidence.
For our reference experiment imagine we have two coins. One is a fair coin,
the other is two-headed. We randomly choose a coin. Then we begin conduct a
sequence of coin tosses to learn which coin was selected. Suppose the tosses yield
n consecutive Heads. P[n Heads | fair] = 2−n ; P[n Heads | two-headed] = 1. So
the likelihood ratio is 2n . That’s our reference experiment. A likelihood ratio
118 CHAPTER 2. MODES OF INFERENCE

around 8 is like tossing three consecutive Heads; a likelihood ratio around 1000
is like tossing ten consecutive Heads.
In Example 2.7 argmax `(θ) ≈ .055 and `(.025)/`(.055) ≈ .13 ≈ 1/8, so the
evidence against θ = .025 as opposed to θ = .055 is about as strong as the
evidence against the fair coin when three consecutive Heads are tossed. The
same can be said for the evidence against θ = .1. Similarly, `(.011)/`(.055) ≈
`(.15)/`(.055) ≈ .001, so the evidence against θ = .011 or θ = .15 is about as
strong as 10 consecutive Heads. A fair statement of the evidence is that θ’s in
the interval from about θ = .025 to about θ = .1 explain the data not much
worse than the maximum of θ ≈ .055. But θ’s below about .01 or larger than
about .15 explain the data not nearly as well as θ’s around .055.

2.3.2 Likelihoods from the Central Limit Theorem

Sometimes it is not possible to compute the likelihood function exactly, either
because it is too difficult or because we don’t know what it is. But we can often
compute an approximate likelihood function using the Central Limit Theorem.
The following example is the simplest case, but typifies the more exotic cases
we will see later on.
Suppose we sample X1 , X2 , . . . , Xn from a probability density f . We don’t
know what f is; we don’t even know what parametric family it belongs to. As-
sume that f has a mean µ and an SD σ (I.e, assume that the mean and variance
are finite.) and that we would like to learn about µ. If (µ, σ) are the only
unknown parameters then the likelihood function is `(µ, σ) = f (Data | µ, σ) =
Q
f (Xi | µ, σ). But we don’t know f and can’t calculate `(µ, σ).
However, we can reason as follows.

1. Most of the information in the data for learning about µ is containined in

X̄. That is, X̄ tells us a lot about µ and the deviations δi ≡ Xi − X̄, i =
1, . . . , n tell us very little.

2. If n is large then the Central Limit Theorem tells us

√
X̄ ∼ N(µ, σ/ n), approximately

3. We can estimate σ 2 from the data by

X
σ̂ 2 = s2 = δi2 /n

4. And therefore the function

 2 !
1 µ − X̄
`M (µ) ∝ exp − √ (2.4)
2 σ̂/ n

is a good approximation to the likelihood function.

2.3. LIKELIHOOD 119

Function 2.4 is called a marginal likelihood function. Tsou and Royall [1995]
show that marginal likelihoods are good approximations to true likelihoods and
can be used to make accurate inferences, at least in cases where the Central Limit
Theorem applies. We shall use marginal likelihoods throughout this book.

Example 2.8 (Slater School, continued)

We redo the Slater School example (Example 2.7) to illustrate the marginal likeli-
hood and see how it compares to the exact likelihood. In that example the X i ’s were
1’s and 0’s indicating which teachers got cancer. There were 8 1’s out of 145 teach-
ers, so X̄ = 8/145 ≈ .055. Also, σ̂ 2 = (8(137/145)2 + 137(8/145)2))/145 ≈ .052,
so σ̂ ≈ .23. We get
 2 !
1 µ − .055
`M (µ) ∝ exp − √ (2.5)
2 .23/ 145

Figure 2.20 shows the marginal and exact likelihood functions. The marginal likeli-
hood is a reasonably good approximation to the exact likelihood.
0.8

marginal
likelihood

exact
0.4
0.0

0.00 0.05 0.10 0.15 0.20

Figure 2.20: Marginal and exact likelihoods for Slater School

Figure 2.20 was produced by the following snippet.

theta <- seq ( 0, .2, length=100 )
lik <- dbeta ( theta, 9, 138 )
lik.mar <- dnorm ( theta, 8/145,
sqrt((8*(137/145)^2 + 137*(8/145)^2)/145)/sqrt(145) )
120 CHAPTER 2. MODES OF INFERENCE

lik <- lik/max(lik)

lik.mar <- lik.mar/max(lik.mar)
matplot ( theta, cbind(lik,lik.mar), xlab=expression(mu),
ylab="likelihood", type="l", lty=c(2,1), col=1 )
legend ( .1, 1, c("marginal", "exact"), lty=c(1,2) )

Example 2.9 (CEO salary)

How much are corporate CEO’s paid? Forbes magazine collected data in 1993 that
can begin to answer this question. The data are available on-line at DASL, the Data
and Story Library, a collection of data sets for free use by statistics students. DASL
says

“Forbes magazine published data on the best small firms in 1993. These
were firms with annual sales of more than five and less than $350 million.
Firms were ranked by five-year average return on investment. The data
extracted are the age and annual salary of the chief executive officer for
the first 60 ranked firms. In question are the distribution patterns for
the ages and the salaries.”

You can download the data from http://lib.stat.cmu.edu/DASL/Datafiles/

ceodat.html. The first few lines look like this:

AGE SAL

53 145
43 621
33 262

In this example we treat the Forbes data as a random sample of size n = 60 of CEO
salaries for small firms. We’re interested in the average salary µ. Our approach is
to calculate the marginal likelihood function ` M (µ).
Figure 2.21(a) shows a stripchart of the data. Evidently, most salaries are
in the range of $200 to $400 thousand dollars, but with a long right-hand tail.
Because the right-hand tail is so much larger than the left, the data are not even
approximately Normally distributed. But the Central Limit Theorem tells us that X̄
is approximately Normally distributed, so the method of marginal likelihood applies.
Figure 2.21(b) displays the marginal likelihood function ` M (µ).

Figure 2.21 was produced by the following snippet.

ceo <- read.table ("data/ceo_salaries/data",header=T)

par ( mfrow=c(2,1) )
stripchart ( ceo$SAL, "jitter", pch=1, main="(a)",
2.3. LIKELIHOOD 121

(a)

0 200 400 600 800 1000

Salary (thousands of dollars)

(b)
1.0
likelihood

0.6
0.2

340 360 380 400 420 440 460

mean salary

Figure 2.21: Marginal likelihood for mean CEO salary

122 CHAPTER 2. MODES OF INFERENCE

xlab="Salary (thousands of dollars)" )

m <- mean ( ceo$SAL, na.rm=T )

s <- sqrt ( var(ceo$SAL,na.rm=T) / (length(ceo$SAL)-1) )
x <- seq ( 340, 470, length=40 )
y <- dnorm ( x, m, s )
y <- y / max(y)
plot ( x, y, type="l", xlab="mean salary",
ylab="likelihood", main="(b)" )

• In s <- sqrt ... the (length(ceo$SAL)-1) is there to account for one

missing data point.
• y <- y / max(y) doesn’t accomplish much and could be omitted.
• The data strongly support the conclusion that the mean salary is between
about $350 and $450 thousand dollars. That’s much smaller than the range
of salaries on display in Figure 2.21(a). Why?
• Is inference about the mean salary useful in this data set? If not, what would
be better?

2.3.3 Likelihoods for several parameters

What if there are two unknown parameters? Then the likelihood is a function
of two variables. For example, if the Xi ’s are a sample from N(µ, σ) then the
likelihood is a function of (µ, σ). The next example illustrates the point.

Example 2.10 (FACE, continued)

This example continues Example 1.12 about a FACE experiment in Duke Forest.
There were six rings; three were treated with excess CO 2 . The dominant canopy
tree in the FACE experiment is pinus taeda, or loblolly pine. Figure 2.22a is a
histogram of the final basal area of each loblolly pine in 1998 divided by its initial
basal area in 1996. It shows that the trees in Ring 1 grew an average of about 30%
but with variability that ranged from close to 0% on the low end to around 50% or
60% on the high end. Because the data are clustered around a central value and
fall off roughly equally on both sides they can be well approximated by a Normal
distribution. But with what mean and SD? What values of (µ, σ) might reasonably
produce the histogram in Figure 2.22a?
2.3. LIKELIHOOD 123

The likelihood function is

n
Y
`(µ, σ) = f (Xi | µ, σ)
1
n
Y 1 1 2
= √ e− 2σ2 (Xi −µ)
1
2πσ
1
Pn 2
∝ σ −n e− 2σ2 1 (Xi −µ)

Figure 2.22b is a contour plot of the likelihood function. The dot in the center,
where (µ, σ) ≈ (1.27, .098), is where the likelihood function is highest. That is the
value of (µ, σ) that best explains the data. The next contour line is drawn where
the likelihood is about 1/4 of its maximum; then the next is at 1/16 the maximum,
the next at 1/64, and the last at 1/256 of the maximum. They show values of
(µ, σ) that explain the data less and less well.
Ecologists are primarily interested in µ because they want to compare the µ’s
from different rings to see whether the excess CO2 has affected the average growth
rate. (They’re also interested in the σ’s, but that’s a secondary concern.) But ` is
a function of both µ and σ, so it’s not immediately obvious that the data tell us
anything about µ by itself. To investigate further, Figure 2.22c shows slices through
the likelihood function at σ = .09, .10, and.11, the locations of the dashed lines in
Figure 2.22b. The three curves are almost identical. Therefore, the relative support
for different values of µ does not depend very much on the value of σ, and therefore
we are justified in interpreting any of the curves in Figure 2.22c as a “likelihood
function” for µ alone, showing how well different values of µ explain the data. In
this case, it looks as though values of µ in the interval (1.25, 1.28) explain the data
much better than values outside that interval.

Figure 2.22 was produced with

par ( mfrow=c(2,2) ) # a 2 by 2 array of plots
x <- ba98$BA.final / ba96$BA.init
x <- x[!is.na(x)]
hist ( x, prob=T, xlab="basal area ratio",
ylab="", main="(a)" )
mu <- seq ( 1.2, 1.35, length=50 )
sd <- seq ( .08, .12, length=50 )
lik <- matrix ( NA, 50, 50 )
for ( i in 1:50 )
for ( j in 1:50 )
lik[i,j] = prod ( dnorm ( x, mu[i], sd[j] ) )
lik <- lik / max(lik)
contour ( mu, sd, lik, levels=4^(-4:0), drawlabels=F,
xlab=expression(mu), ylab=expression(sigma),
main="(b)" )
abline ( h = c (.09, .1, .11 ), lty=2 )
124 CHAPTER 2. MODES OF INFERENCE

(a) (b)

0.12
5
4

0.11
3

0.10
σ
2

0.09
1

0.08
0

1.0 1.2 1.4 1.6 1.20 1.30

basal area ratio µ

(c)
1.0
0.8
likelihood

0.6
0.4
0.2
0.0

1.20 1.30

Figure 2.22: FACE Experiment, Ring 1. (a): (1998 final basal area) ÷
(1996 initial basal area); (b): contours of the likelihood function. (c): slices
of the likelihood function.
2.3. LIKELIHOOD 125

lik.09 <- lik[,13] / max(lik[,13])

lik.10 <- lik[,26] / max(lik[,26])
lik.11 <- lik[,38] / max(lik[,38])
matplot ( mu, cbind ( lik.09, lik.10, lik.11 ), type="l",
col=1, main="(c)",
xlab=expression(mu), ylab="likelihood" )

• The line x <- x[!is.na(x)] is there because some data is missing. This
line selects only those data that are not missing and keeps them in x. When
x is a vector, is.na(x) is another vector, the same length as x, with TRUE
or FALSE, indicating where x is missing. The ! is “not”, or negation, so
x[!is.na(x)] selects only those values that are not missing.
• The lines mu <- ... and sd <- ... create a grid of µ and σ values at
which to evaluate the likelihood.
• The line lik <- matrix ( NA, 50, 50 ) creates a matrix for storing the
values of `(µ, σ) on the grid. The next three lines are a loop to calculate the
values and put them in the matrix.
• The line lik <- lik / max(lik) rescales all the values in the matrix so the
maximum value is 1. Rescaling makes it easier to set the levels in the next
line.
• contour produces a contour plot. contour(mu,sd,lik,...) specifies the
values on the x-axis, the values on the y-axis, and a matrix of values on the
grid. The levels argument says at what levels to draw the contour lines,
while drawlabels=F says not to print numbers on those lines. (Make your
own contour plot without using drawlabels to see what happens.)
• abline is used for adding lines to plots. You can say either abline(h=...)
or abline(v=...) to get horizontal and vertical lines, or
abline(intercept,slope) to get arbitrary lines.
• lik.09, lik.10, and lik.11 pick out three columns from the lik matrix.
They are the three columns for the values of σ closest to σ = .09, .10, .11.
Each column is rescaled so its maximum is 1.

Example 2.11 (Quiz Scores, continued)

This example continues Example 2.3 about scores in Statistics 103. Figure 2.7 shows
that most students scored between about 5 and 10, while 4 students were well below
the rest of the class. In fact, those students did not show up for every quiz so their
averages were quite low. But the remaining students’ scores were clustered together
in a way that can be adequately described by a Normal distribution. What do the
data say about (µ, σ)?
Figure 2.23 shows the likelihood function. The data support values of µ from
126 CHAPTER 2. MODES OF INFERENCE

about 7.0 to about 7.6 and values of σ from about 0.8 to about 1.2. A good
description of the data is that most of it follows a Normal distribution with (µ, σ)
in the indicated intervals, except for 4 students who had low scores not fitting the
general pattern. Do you think the instructor should use this analysis to assign letter
grades and, if so, how?

1.3
1.2
1.1
1.0
σ

0.9
0.8
0.7

6.8 7.0 7.2 7.4 7.6

Figure 2.23: Likelihood function for Quiz Scores

Figure 2.23 was produced by

x <- sort(scores.ave)[5:58]
mu <- seq ( 6.8, 7.7, length=60 )
sig <- seq ( .7, 1.3, length=60 )
lik <- matrix ( NA, 60, 60 )
2.3. LIKELIHOOD 127

for ( i in 1:60 )
for ( j in 1:60 )
lik[i,j] <- prod ( dnorm ( x, mu[i], sig[j] ) )
lik <- lik/max(lik)
contour ( mu, sig, lik, xlab=expression(mu),
ylab=expression(sigma) )

Examples 2.10 and 2.11 have likelihood contours that are roughly circular,
indicating that the likelihood function for one parameter does not depend very
strongly on the value of the other parameter, and so we can get a fairly clear
picture of what the data say about one parameter in isolation. But in other data
sets two parameters may be inextricably entwined. Example 2.12 illustrates the
problem.

Example 2.12 (Seedlings, continued)

Examples 1.4, 1.6, and 1.7 introduced an observational study by ecologists to learn
about tree seedling emergence and survival. Some species, Red Maple or acer
rubrum for example, get a mark called a bud scale scar when they lose their leaves
over winter. By looking for bud scale scars ecologists can usually tell whether an
acer rubrum seedling is New (in its first summer), or Old (already survived through
at least one winter). When they make their annual observations they record the
numbers of New and Old acer rubrum seedlings in each quadrat. Every Old seedling
in year t must have been either a New or an Old seedling in year t − 1.
Table 2.1 shows the 1992–1993 data for quadrat 6. Clearly the data are incon-
sistent; where did the Old seedling come from in 1993? When confronted with
this paradox the ecologists explained that some New seedlings emerge from the
ground after the date of the Fall census but before the winter. Thus they are not
counted in the census their first year, but develop a bud scale scar and are counted
as Old seedlings in their second year. One such seedling must have emerged in
1992, accounting for the Old seedling in 1993.

Year No. of New seedlings No. of Old seedlings

1992 0 0
1993 0 1

Table 2.1: Numbers of New and Old seedlings in quadrat 6 in 1992 and 1993.

How shall we model the data? Let NiT be the true number of New seedlings in
year i, i.e., including those that emerge after the census; and let N iO be the observed
number of seedlings in year i, i.e., those that are counted in the census. As in
Example 1.4 we model NiT ∼ Poi(λ). Furthermore, each seedling has some chance
θf of being found in the census. (Nominally θf is the proportion of seedlings that
emerge before the census, but in fact it may also include a component accounting
128 CHAPTER 2. MODES OF INFERENCE

for the failure of ecologists to find seedlings that have already emerged.) Treating
the seedlings as independent and all having the same θ f leads to the model NiO ∼
Bin(NiT , θf ). The data are the NiO ’s; the NiT ’s are not observed. What do the
data tell us about the two parameters (λ, θf )?
O
Ignore the Old seedlings for now and just look at 1992 data N 1992 = 0. Dropping
the subscript 1992, the likelihood function is

`(λ, θf ) = P[N O = 0 | λ, θf ]
X∞
= P[N O = 0, N T = n | λ, θf ]
n=0
X∞
= P[N T = n | λ] P[N O = 0 | N T , θf ]
n=0
∞ (2.6)
X e−λ λn
= (1 − θf )n
n=0
n!
∞ n
X e−λ(1−θf ) (λ(1 − θf ))
=
n=0
eλθf n!
= e−λθf

Figure 2.24a plots log10 `(λ, θf ). (We plotted log10 ` instead of ` for variety.)
The contour lines are not circular. To see what that means, focus on the curve
log10 `(λ, θf ) = −1 which runs from about (λ, θf ) = (2.5, 1) to about (λ, θf ) =
(6, .4). Points (λ, θf ) along that curve explain the datum N O = 0 about 1/10 as
well as the m.l.e.(The m.l.e. is any pair where either λ = 0 or θ f = 0.) Points below
and to the left of that curve explain the datum better than 1/10 of the maximum.
The main parameter of ecological interest is λ, the rate at which New seedlings
tend to arrive. The figure shows that values of λ as large as 6 can have reasonably
large likelihoods and hence explain the data reasonably well, at least if we believe
that θf might be as small as .4. To investigate further, Figure 2.24b is similar
to 2.24a but includes values of λ as large as 1000. It shows that even values of
λ as large as 1000 can have reasonably large likelihoods if they’re accompanied
by sufficiently small values of θf . In fact, arbitrarily large values of λ coupled with
sufficiently small values of θf can have arbitrarily large likelihoods. So from the data
alone, there is no way to rule out extremely large values of λ. Of course extremely
large values of λ don’t make ecological sense, both in their own right and because
extremely small values of θf are also not sensible. Scientific background information
of this type is incorporated into statistical analysis often through Bayesian inference
(Section 2.5). But the point here is that λ and θf are linked, and the data alone
does not tell us much about either parameter individually.

Figure 2.24(a) was produced with the following snippet.

lam <- seq ( 0, 6, by=.1 )

th <- seq ( 0, 1, by=.02 )
2.3. LIKELIHOOD 129

0.8
(a)
θf

0.4
0.0

0 1 2 3 4 5 6

(b)
0.8
θf

0.4
0.0

0 200 400 600 800 1000

Figure 2.24: Log of the likelihood function for (λ, θf ) in Example 2.12
130 CHAPTER 2. MODES OF INFERENCE

lik <- matrix ( NA, length(lam), length(th) )

for ( i in seq(along=lam) )
for ( j in seq(along=th) )
lik[i,j] <- exp ( -lam[i]*th[j] )
contour ( lam, th, log10(lik),
levels=c(0,-.2,-.6,-1,-1.5,-2),
xlab=expression(lambda),
ylab=expression(theta[f]), main="(a)" )

• log10 computes the base 10 logarithm.

Figure 2.24(b) was produced with the following snippet.

lam2 <- seq ( 0, 1000, by=1 )
lik2 <- matrix ( NA, length(lam2), length(th) )
for ( i in seq(along=lam2) )
for ( j in seq(along=th) )
lik2[i,j] <- exp ( -lam2[i]*th[j] )
contour ( lam2, th, log10(lik2), levels=c(0,-1,-2,-3),
xlab=expression(lambda),
ylab=expression(theta[f]), main="(b)" )

We have now seen two examples (2.10 and 2.11) in which likelihood contours
are roughly circular and one (2.12) in which they’re not. By far the most com-
mon and important case is similar to Example 2.10 because it applies when the
Central Limit Theorem applies. That is, there are many instances in which we
are trying to make an inference about a parameter θ and can invoke the Central
Limit Theorem saying that for some statistic t, t ∼ N(θ, σt ) approximately and
where we can estimate σt . In these cases we can, if necessary, ignore any other
parameters in the problem and make an inference about θ based on `M (θ).

2.4 Estimation
Sometimes the purpose of a statistical analysis is to judge the amount of support
given by y to various values of θ. Once y has been observed, then p(y | θ) is a
function, called the likelihood function and denoted `(θ), only of θ. For each
value of θ, `(θ) says how well that particular θ explains the data. A fundamental
principle of estimation is that if `(θ1 ) > `(θ2 ) then θ1 explains the data better
than θ2 and therefore the data give more support to θ1 than θ2 . In fact, the
amount of support is directly proportional to `(θ). So if `(θ1 ) = 2`(θ2 ) then the
data support θ1 twice as much as θ2 .
An informed guess at the value of θ is called an estimate and denoted θ̂. If
we had to estimate θ based solely on y, we would choose the value of θ for which
2.4. ESTIMATION 131

`(θ) was largest. That’s the subject of Section 2.4.1.

2.4.1 The Maximum Likelihood Estimate

In many statistics problems there is a unique value of θ that maximizes `(θ).
This value is called the maximum likelihood estimate, or m.l.e. of θ and denoted
θ̂.
θ̂ ≡ argmaxθ p(y | θ) = argmaxθ `(θ).
For instance, in Example 2.7 and Figure 2.19 θ was the rate of cancer oc-
curence and we calculated `(θ) based on y = 8 cancers in 145 people. Figure 2.19
suggests that the m.l.e. is about θ̂ ≈ .05.
When `(θ) is differentiable, the m.l.e. can be found by differentiating and
equating to zero. In Example 2.7 the likelihood was `(θ) ∝ θ 8 (1 − θ)137 . The
derivative is
d`(θ)
∝ 8θ7 (1 − θ)137 − 137θ8 (1 − θ)136
dθ (2.7)
= θ7 (1 − θ)136 [8(1 − θ) − 137θ]

Equating to 0 yields

0 = 8(1 − θ) − 137θ
145θ = 8
θ = 8/145 ≈ .055

So θ̂ ≈ .055 is the m.l.e. Of course if the mode is flat, there are multiple modes,
the maximum occurs at an endpoint, or ` is not differentiable, then more care
is needed.
Equation 2.7 shows more generally the m.l.e. for Binomial data. Simply
replace 137 with n and 8 with y to get θ̂ = y/n. In the Exercises you will be
asked to find the m.l.e. for data from other types of distributions.
There is a trick that is often useful for finding m.l.e.’s. Because log is a
monotone function, argmax `(θ) = argmax log(`(θ)),
Q so the m.l.e.Pcan be found
by maximizing log `. For i.i.d. data, `(θ) = p(yi | θ), log `(θ) = log p(xi | θ),
and it is often easier to differentiate the sum than the product. For the Slater
example the math would look like this:

log `(θ) = 8 log θ + 137 log(1 − θ)

d log `(θ) 8 137
= −
dθ θ 1−θ
137 8
=
1−θ θ
137θ = 8 − 8θ
8
θ= .
145
132 CHAPTER 2. MODES OF INFERENCE

Equation 2.7 shows that if y1 , . . . , yn ∼ Bern(θ) then the m.l.e. of θ is

X
θ̂ = n−1 yi = sample mean

The Exercises ask you to show the following.

1. If y1 , . . . , yn ∼ N(µ, σ) then the m.l.e. of µ is
X
µ̂ = n−1 yi = sample mean

2. If y1 , . . . , yn ∼ Poi(λ) then the m.l.e. of λ is

X
λ̂ = n−1 yi = sample mean

3. If y1 , . . . , yn ∼ Exp(λ) then the m.l.e. of λ is

X
λ̂ = n−1 yi = sample mean

2.4.2 Accuracy of Estimation

Finding the m.l.e. is not enough. Statisticians want to quantify the accuracy of
θ̂ as an estimate of θ. In other words, we want to know what other values of
θ, in addition to θ̂, have reasonably high likelihood (provide a reasonably good
explanation of the data). And what does “reasonable” mean? Section 2.4.2
addresses this question.
As we saw from the reference experiment in section 2.3, the evidence is not
very strong against any value of θ such that `(θ) > `(θ̂)/10. So when considering
estimation accuracy it is useful to think about sets such as
( )
`(θ)
LS.1 ≡ θ : ≥ .1
`(θ̂)

LS stands for likelihood set. More generally, for any α ∈ (0, 1) we define the
likelihood set of level α to be
( )
`(θ)
LSα ≡ θ : ≥α
`(θ̂)

LSα is the set of θ’s that explain the data reasonbly well, and therefore the
set of θ’s best supported by the data, where the quantification of “reasonable”
and “best” are determined by α. The notion is only approximate and meant
as a heuristic reference; in reality there is no strict cutoff between reasonable
and unreasonable values of θ. Also, there is no uniquely best value of α. We
frequently use α ≈ .1 for convenience and custom.
In many problems the likelihood function `(θ) is continuous and unimodal,
i.e. strictly decreasing away from θ̂, and goes to 0 as θ → ±∞, as in Figures 2.18
2.4. ESTIMATION 133

and 2.19. In these cases, θ ≈ θ̂ ⇒ `(θ) ≈ `(θ̂). So values of θ close to θ̂ explain

the data almost as well as and are about as plausible as θ̂ and LSα is an interval

LSα = [θl , θu ]

where θl and θu are the lower and upper endpoints, respectively, of the interval.
In Example 2.8 (Slater School) θ̂ = 8/145, so we can find `(θ̂) on a calculator,
or by using R’s built-in function

dbinom ( 8, 145, 8/145 )

which yields about .144. Then θl and θu can be found by trial and error. Since
dbinom(8,145,.023) ≈ .013 and dbinom(8,145,.105) ≈ .015, we conclude
that LS.1 ≈ [.023, .105] is a rough likelihood interval for θ. Review Figure 2.19
to see whether this interval makes sense.
The data in Example 2.8 could pin down θ to an interval of width about .08.
In general, an experiment will pin down θ to an extent determined by the amount
of information in the data. As data accumulates so does information and the
ability to determine θ. Typically the likelihood function becomes increasingly
more peaked as n → ∞, leading to increasingly accurate inference for θ. We
saw that in Figures 1.6 and 2.18. Example 2.13 illustrates the point further.

Example 2.13 (Craps, continued)

Example 1.10 introduced a computer simulation to learn the probability θ of winning
the game of craps. In this example we use that simulation to illustrate the effect
of gathering ever increasing amounts of data. We’ll start by running the simulation
just a few times, and examining the likelihood function `(θ). Then we’ll add more
and more simulations and see what happens to `(θ).
The result is in Figure 2.25. The flattest curve is for 3 simulations, and the curves
become increasingly peaked for 9, 27, and 81 simulations. After only 3 simulations
LS.1 ≈ [.15, .95] is quite wide, reflecting the small amount of information. But after
9 simulations `(θ) has sharpened so that LS.1 ≈ [.05, .55] is much smaller. After
27 simulations LS.1 has shrunk further to about [.25, .7], and after 81 it has shrunk
even further to about [.38, .61].

Figure 2.25 was produced with the following snippet.

n.sim <- c ( 3, 9, 27, 81 )
th <- seq ( 0, 1, length=200 )
lik <- matrix ( NA, 200, length(n.sim) )

for ( i in seq(along=n.sim) ) {
wins <- 0
for ( j in 1:n.sim[i] )
wins <- wins + sim.craps()
lik[,i] <- dbinom ( wins, n.sim[i], th )
lik[,i] <- lik[,i] / max(lik[,i])
134 CHAPTER 2. MODES OF INFERENCE

1.0
0.8
likelihood

0.6
0.4
0.2
0.0

0.0 0.2 0.4 0.6 0.8 1.0

Figure 2.25: Likelihood function for the probability θ of winning a game of

craps. The four curves are for 3, 9, 27, and 81 simulations.
2.4. ESTIMATION 135

matplot ( th, lik, type="l", col=1, lty=1:4,

xlab=expression(theta), ylab="likelihood" )

In Figure 2.25 the likelihood function looks increasingly like a Normal density
as the number of simulations increases. That is no accident; it is the typical
behavior in many statistics problems. Section 2.4.3 explains the reason.

2.4.3 The sampling distribution of an estimator

The estimator θ̂ is a function of the data y1 , . . . , yn . If we repeat the experiment
and get new data we also get a new θ̂. So θ̂ is a random variable and has a
distribution called the sampling distribution of θ̂ and denoted Fθ̂ . We studied
Fθ̂ in Example 1.11 where we used simulation to estimate the probability θ of
winning a game of craps. For each sample size of n = 50, 200, 1000 we did 1000
simulations. Each simulation yielded a different θ̂. Those 1000 θ̂’s are a random
sample of size 1000 from Fθ̂ . Figure 1.18 showed boxplots of the simulations.
Now we examine the sampling distribution of θ̂ in more detail. There are
at least two reasons for doing so. First, Fθ̂ is another way, in addition to
likelihood sets, of assessing the accuracy of θ̂ as an estimator of θ. If Fθ̂ is
tightly concentrated around θ then θ̂ is highly accurate. Conversely, if Fθ̂ is
highly dispersed, or not centered around θ, then θ̂ is an inaccurate estimator.
Second, we may want to compare two possible estimators. I.e., if there are
two potential estimators θ̂1 and θ̂2 , we can compare Fθ̂1 and Fθ̂2 and use the
estimator whose sampling distribution is most tightly concentrated around θ.
To illustrate, let’s suppose we sample y1 , . . . , yn from distribution FY , and
want to estimate θ ≡ E[Y ]. We consider two potential estimators, the sample
P
mean θ̂1 = (1/n) yi and the sample median θ̂2 . To see which estimator is
better we do a simulation, as shown in the following snippet. The simulation
is done at four different sample sizes, n = 4, 16, 64, 256, to see whether sample
size matters. Here we’ll let FY be N(0, 1). But the choice between θ̂1 and θ̂2
might depend on what FY is, so a more thorough investigation would consider
other choices of FY .
We do 1000 simulations at each sample size. Figure 2.26 shows the result.
The figure suggests that the sampling distributions of both θ̂1 and θ̂2 are cen-
tered at the true value of θ. The distribution of θ̂1 is slightly less variable than
that of θ̂2 , but not enough to make much practical difference.

Figure 2.26 was produced by the following snippet.

sampsize <- c ( 4, 16, 64, 256 )
n.sim <- 1000
136 CHAPTER 2. MODES OF INFERENCE

(a) (b)

0.5
1

0.0
0
−1

−1.0
−2

mean median mean median

(c) (d)
0.2
0.4

0.0
0.0

−0.2
−0.4

mean median mean median

Figure 2.26: Sampling distribution of θ̂1 , the sample mean and θ̂2 , the sample
median. Four different sample sizes. (a): n=4; (b): n=16; (c): n=64; (d):
n=256
2.4. ESTIMATION 137

par ( mfrow=c(2,2) )
for ( i in seq(along=sampsize) ) {
y <- matrix ( rnorm ( n.sim*sampsize[i], 0, 1 ),
nrow=sampsize[i], ncol=n.sim )
that.1 <- apply ( y, 2, mean )
that.2 <- apply ( y, 2, median )
boxplot ( that.1, that.2, names=c("mean","median"),
main=paste("(",letters[i],")",sep="") )
abline ( h=0, lty=2 )
}

For us, comparing θ̂1 to θ̂2 is only a secondary point of the simulation. The
main point is three-fold.
1. An estimator is a random variable and has a distribution.
2. Statisticians study conditions under which one estimator is better than
another.
3. Simulation is useful.
When the m.l.e. is the sample mean, as it is when FY is a Bernoulli, Normal,
Poisson or Exponential distribution, the Central Limit Theorem tells us that
in large samples, θ̂ is approximately Normally distributed. Therefore, in these
cases, its distribution can be well described by its mean and SD. Approximately,

θ̂ ∼ N(µθ̂ , σθ̂ ).

where
µθ̂ = µY
σY (2.8)
σθ̂ = √
n
both of which can be easily estimated from the sample. So we can use the
sample to compute a good approximation to the sampling distribution of the
m.l.e.
To see that more clearly, let’s make 1000 simulations of the m.l.e. in n =
5, 10, 25, 100 Bernoulli trials with p = .1. We’ll make histograms of those sim-
ulations and overlay them with kernel density estimates and Normal densities.
The parameters of the Normal densities will be estimated from the simulations.
Results are shown in Figure 2.27.

Figure 2.27 was produced by the following snippet.

sampsize <- c ( 5, 10, 25, 100 )
n.sim <- 1000
p.true <- .1
138 CHAPTER 2. MODES OF INFERENCE

(a) (b)
12

12
density

density
0 2 4 6 8

0 2 4 6 8
0.0 0.2 0.4 0.6 0.0 0.2 0.4 0.6

θ^ θ^

(c) (d)
12

12
density

density
0 2 4 6 8

0 2 4 6 8

0.0 0.2 0.4 0.6 0.0 0.2 0.4 0.6

θ^ θ^

Figure 2.27: Histograms of θ̂, the sample mean, for samples from Bin(n, .1).
Dashed line: kernel density estimate. Dotted line: Normal approximation. (a):
n=4; (b): n=16; (c): n=64; (d): n=256
2.4. ESTIMATION 139

par(mfrow=c(2,2))
for ( i in seq(along=sampsize) ) {
# n.sim Bernoulli samples of sampsize[i]
y <- matrix ( rbinom ( n.sim*sampsize[i], 1, p.true ),
nrow=n.sim, ncol=sampsize[i] )

# for each sample, compute the mean

t.hat <- apply ( y, 1, mean )

# histogram of theta hat

hist ( t.hat, prob=T,
xlim=c(0,.6), xlab=expression(hat(theta)),
ylim=c(0,14), ylab="density",
main=paste ( "(", letters[i], ")", sep="" )
)

# kernel density estimate of theta hat

lines ( density ( t.hat ), lty=2 )

# Normal approximation to density of theta hat,

# calculated from the first sample
m <- mean(y[1,])
sd <- sd(y[1,])/sqrt(sampsize[i])
t <- seq ( min(t.hat), max(t.hat), length=40 )
lines ( t, dnorm ( t, m, sd ), lty=3 )
}

Notice that the Normal approximation is not very good for small n. That’s
because the underlying distribution FY is highly skewed, nothing at all like a
Normal distribution. In fact, R was unable to compute the Normal approxima-
tion for n = 5. But for large n, the Normal approximation is quite good. That’s
the Central Limit Theorem kicking in. For any n, we can use the sample to
estimate the parameters in Equation 2.8. For small n, those parameters don’t
help us much. But for n = 256, they tell us a lot about the accuracy of θ̂, and
the Normal approximation computed from the first sample is a good match to
the sampling distribution of θ̂.
The SD of an estimator is given a special name. It’s called the standard
error or SE of the estimator because it measures the typical size of estimation
errors |θ̂ − θ|. When θ̂ ∼ N(µθ̂ , σθ̂ ), approximately, then σθ̂ is the SE. For any
Normal distribution, about 95% of the mass is within ±2 standard deviations
of the mean. Therefore,
Pr[|θ̂ − θ| ≤ 2σθ̂ ] ≈ .95
In other words, estimates are accurate to within about two standard errors
140 CHAPTER 2. MODES OF INFERENCE

about 95% of the time, at least when Normal theory applies.

We have now seen two ways of assessing estimation accuracy — through
`(θ) and through Fθ̂ . Often these two apparently different approaches almost
coincide. That happens under the following conditions.
√
1. When θ̂ ∼ N(θ, σθ̂ ), and σθ̂ ≈ σ/ n, an approximation often justified by
the Central Limit Theorem, then we can estimate θ to within about ±2σθ ,
around 95% of the time. So the interval (θ̂ − 2σθ̂ , θ̂ + 2σθ̂ ) is a reasonable
estimation interval.
2. When most of the information in the data come from the sample mean,
and in other
 cases whena marginal likelihood argument applies, then
 2
1 θ−√Ȳ
`(θ) ≈ exp − 2 σ̂/ n (Equation 2.4) and LS.1 ≈ (θ̂ − 2σθ̂ , θ̂ + 2σθ̂ ).
So the two intervals are about the same.

2.5 Bayesian Inference

The essence of Bayesian inference is using probability distributions to describe
our state of knowledge of some parameter of interest, θ. We construct p(θ),
either a pmf or pdf, to reflect our knowledge by making p(θ) large for those
values of θ that seem most likely, and p(θ) small for those values of θ that seem
least likely, according to our state of knowledge. Although p(θ) is a probability
distribution, it doesn’t necessarily mean that θ is a random variable. Rather,
p(θ) encodes our state of knowledge. And different people can have different
states of knowledge, hence different probability distributions. For example,
suppose you toss a fair coin, look at it, but don’t show it to me. The outcome is
not random; it has already occured and you know what it is. But for me, each
outcome is equally likely. I would encode my state of knowledge by assigning
P(H) = P(T) = 1/2. You would encode your state of knowledge by assigning
either P(H) = 1 or P(T) = 1 according to whether the coin was Heads or Tails.
After I see the coin I would update my probabilities to be the same as yours.
For another common example, consider horse racing. When a bettor places
a bet at 10 to 1, she is paying $1 for a ticket that will pay $10 if the horse wins.
Her expected payoff for that bet is −$1 + P[horse wins] × $10. For that to be a
good deal she must think that P[horse wins] ≥ .1. Of course other bettors may
disagree.
Here are some other examples in which probability distributions must be
assessed.
• In deciding whether to fund Head Start, legislators must assess whether
the program is likely to be beneficial and, if so, the degree of benefit.
• When investing in the stock market, investors must assess the future prob-
ability distributions of stocks they may buy.
• When making business decisions, firms must assess the future probability
distributions of outcomes.
2.5. BAYESIAN INFERENCE 141

• Weather forecasters assess the probability of rain.

• Public policy makers must assess whether the observed increase in average
global temperature is anthropogenic and, if so, to what extent.

• Doctors and patients must assess and compare the distribution of out-
comes under several alternative treatments.

• At the Slater School, Example 2.7, teachers and administrators must assess
their probability distribution for θ, the chance that a randomly selected
teacher develops invasive cancer.

Information of many types goes into assessing probability distributions. But

it is often useful to divide the information into two types: general background
knowledge and information specific to the situation at hand. How do those two
types of information combine to form an overall distribution for θ? Often we
begin by summarizing just the background information as p(θ), the marginal
distribution of θ. The specific information at hand is data which we can model
as p(y1 , . . . , yn | θ), the conditional distribution of y1 , . . . , yn given θ. Next, the
marginal and conditional densities are combined to give the joint distribution
p(y1 , . . . , yn , θ). Finally, the joint distribution yields p(θ | y1 , . . . , yn ) the con-
ditional distribution of θ given y1 , . . . , yn . And p(θ | y1 , . . . , yn ) represents our
state of knowledge accounting for both the background information and the
data specific to the problem at hand. p(θ) is called the prior distribution and
p(θ | y1 , . . . , yn ) is the posterior distribution.
A common application is in medical screening exams. Consider a patient
being screened for a rare disease, one that affects 1 in 1000 people, say. The
disease rate in the population is background information; the patient’s response
on the screening exam is data specific to this particular patient. Define an
indicator variable D by D = 1 if the patient has the disease and D = 0 if not.
Define a second random variable T by T = 1 if the test result is positive and
T = 0 if the test result is negative. And suppose the test that is 95% accurate
in the sense that P[T = 1 | D = 1] = P[T = 0 | D = 0] = .95. Finally, what is the
chance that the patient has the disease given that the test is positve? In other
words, what is P[D = 1 | T = 1]?
We have the marginal distribution of D and the conditional distribution of
T given D. The procedure is to find the joint distribution of (D, T), then the
142 CHAPTER 2. MODES OF INFERENCE

conditional distribution of D given T. The math is

P[D = 1 and T = 1]
P[D = 1 | T = 1] =
P[T = 1]
P[D = 1 and T = 1]
=
P[T = 1 and D = 1] + P[T = 1 and D = 0]
P[D = 1] P[T = 1 | D = 1]
=
P[D = 1] P[T = 1 | D = 1] + P[D = 0] P[T = 1 | D = 0]
(.001)(.95)
=
(.001)(.95) + (.999)(.05)
.00095
= ≈ .019.
.00095 + .04995
(2.9)

That is, a patient who tests positive has only about a 2% chance of having the
disease, even though the test is 95% accurate.
Many people find this a surprising result and suspect a mathematical trick.
But a quick heuristic check says that out of 1000 people we expect 1 to have the
disease, and that person to test positive; we expect 999 people not to have the
disease and 5% of those, or about 50, to test positive; so among the 51 people
who test postive, only 1, or a little less than 2%, has the disease. The math is
correct. This is an example where most people’s intuition is at fault and careful
attention to mathematics is required in order not to be led astray.
What is the likelihood function in this example? There are two possible
values of the parameter, hence only two points in the domain of the likelihood
function, D = 0 and D = 1. So the likelihood function is

`(0) = .05; `(1) = .95

Here’s another way to look at the medical screening problem, one that highlights
the multiplicative nature of likelihood.

P[D = 1 | T = 1] P[D = 1 and T = 1]

=
P[D = 0 | T = 1] P[D = 0 and T = 1]
P[D = 1] P[T = 1 | D = 1]
=
P[D = 0] P[T = 1 | D = 0]
  
P[D = 1] P[T = 1 | D = 1]
=
P[D = 0] P[T = 1 | D = 0]
  
1 .95
=
999 .05
≈ .019

The LHS of this equation is the posterior odds of having the disease. The
penultimate line shows that the posterior odds is the product of the prior odds
and the likelihood ratio. Specifically, to calculate the posterior, we need only the
2.5. BAYESIAN INFERENCE 143

likelihood ratio, not the absolute value of the likelihood function. And likelihood
ratios are the means by which prior odds get transformed into posterior odds.
Let’s look more carefully at the mathematics in the case where the distribu-
tions have densities. Let y denote the data, even though in practice it might be
y1 , . . . , yn .
p(θ, y)
p(θ | y) =
p(y)
p(θ, y)
=R (2.10)
p(θ, y) dθ
p(θ)p(y | θ)
=R
p(θ)p(y | θ) dθ
Equation 2.10 is the same as Equation 2.9, only in more general terms. Since
we are treating the data as given and p(θ | y) as a function of θ, we are justified
in writing
p(θ)`(θ)
p(θ | y) = R
p(θ)`(θ) dθ
or
p(θ)`(θ)
p(θ | y) =
c
R
where c = p(θ)`(θ) dθ is a constant that does not depend on θ. (An integral
with respect to θ does not depend on θ; after integration it does not contain
θ.) The effect of the constantR c is to rescale the function in the numerator so
that it integrates to 1. I.e., p(θ | y) dθ = 1. And since c plays this role, the
likelihood function can absorb an arbitrary constant which will ultimately be
compensated for by c. One often sees the expression
p(θ | y) ∝ p(θ)`(θ) (2.11)
where the unmentioned constant of proportionality is c.
We can find
R c either through Equation 2.10 or by using Equation 2.11, then
setting c = [ p(θ)`(θ) dθ]−1 . Example 2.14 illustrates the second approach.

Example 2.14 (Seedlings, continued)

Recall the Seedlings examples (1.4, 1.6, 1.7, 1.9, 2.6, and 2.12) which modelled
the number of New seedling arrivals as Poi(λ). Prior to the experiment ecologists
knew quite a bit about regeneration rates of acer rubrum in the vicinity of the
experimental quadrats. They estimated that New seedlings would arise at a rate
most likely around .5 to 2 seedlings per quadrat per year and less likely either more
or less than that. Their knowledge could be encoded in the prior density displayed
in Figure 2.28 which is p(λ) = 2λ2 e−2λ . (This is the Gam(3, 1/2) density; see
Section 5.5.) Figure 2.28 also displays the likelihood function p(y | λ) ∝ λ 3 e−λ
found in Example 1.4 and Figure 1.6. Therefore, according to Equation 2.11, the
posterior density is p(λ | y) ∝ λ5 e−3λ . In Section 5.5 we will see that this is the
Gam(6, 1/3) density, up to a constant of proportionality. Therefore c in this example
must be the constant that appears in the Gamma density: c = 1/[5! × (1/3) 6 ].
144 CHAPTER 2. MODES OF INFERENCE

0.5

prior
likelihood
posterior
0.4
0.3
0.2
0.1
0.0

0 1 2 3 4 5

Figure 2.28: Prior, likelihood and posterior densities for λ in the seedlings
example after the single observation y = 3
2.5. BAYESIAN INFERENCE 145

In Figure 2.28 the posterior density is more similar to the prior density than
to the likelihood function. But the analysis deals with only a single data point.
Let’s see what happens as data accumulates. If we have observations y1 , . . . , yn ,
the likelihood function becomes
Y Y e−λ λyi P
`(λ) = p(yi | λ) = ∝ e−nλ λ yi
yi !
To see what this means in practical terms, Figure 2.29 shows (a): the same
prior we used in Example 2.14, (b): `(λ) for n = 1, 4, 16, and (c): the posterior
for n = 1, 4, 16, always with ȳ = 3.
1. As n increases the likelihood function becomes increasingly peaked. That’s
because as n increases, the amount of information about λ increases, and
we know λ with increasing accurracy. The likelihood function becomes
increasingly peaked around the true value of λ and interval estimates
become increasingly narrow.
2. As n increases the posterior density becomes increasingly peaked and be-
comes increasingly like `(λ). That’s because as n increases, the amount of
information in the data increases and the likelihood function becomes in-
creasingly peaked. Meanwhile, the prior density remains as it was. Even-
tually the data contains much more information than the prior, so the
likelihood function becomes much more peaked than the prior and the
likelihood dominates. So the posterior, the product of prior and likeli-
hood, looks increasingly like the likelihood.
Another way
Pn to look at it is through the loglikelihood log `(λ) = c +
log p(λ) + 1 p(yi | λ). As n → ∞ there is an increasing number of terms
in the sum, so the sum eventually becomes much larger and much more
important than log p(λ).
In practice, of course, ȳ usually doesn’t remain constant as n increases. We
saw in Example 1.6 that there were 40 new seedlings in 60 quadrats. With this
data the posterior density is

p(θ | y1 , . . . , y60 ) ∝ λ42 e−62λ (2.12)

which is the Gam(43, 1/62) density. It is pictured in Figure 2.30. Compare to

Figure 2.28.
Example 2.15 shows Bayesian statistics at work for the Slater School. See
Lavine [1999] for further analysis.

Example 2.15 (Slater School, cont.)

At the time of the analysis reported in Brodeur [1992] there were two other lines
of evidence regarding the effect of power lines on cancer. First, there were some
epidemiological studies showing that people who live near power lines or who work
as power line repairmen develop cancer at higher rates than the population at large,
though only slightly higher. And second, chemists and physicists who calculate the
146 CHAPTER 2. MODES OF INFERENCE

a b

0.5

0.8
0.4

0.6
likelihood
0.3
prior

0.4
0.2

0.2
0.1
0.0

0.0
0 1 2 3 4 5 0 1 2 3 4 5

λ λ

c
1.0
0.8
posterior

0.6

n=1
n=4
n = 16
0.4
0.2
0.0

0 1 2 3 4 5

Figure 2.29: a:Prior, b:likelihood and c:posterior densities for λ with n = 1, 4, 16

2.5. BAYESIAN INFERENCE 147

prior
likelihood
posterior
3
density

2
1
0

0.0 0.5 1.0 1.5 2.0

P
Figure 2.30: Prior, likelihood and posterior densities for λ with n = 60, yi =
40.
148 CHAPTER 2. MODES OF INFERENCE

size of magnetic fields induced by power lines (the supposed mechanism for inducing
cancer) said that the small amount of energy in the magnetic fields is insufficient
to have any appreciable affect on the large biological molecules that are involved in
cancer genesis. These two lines of evidence are contradictory. How shall we assess
a distribution for θ, the probability that a teacher hired at Slater School develops
cancer?
Recall from page 116 that Neutra, the state epidemiologist, calculated “4.2
cases of cancer could have been expected to occur” if the cancer rate at Slater
were equal to the national average. Therefore, the national average cancer rate for
women of the age typical of Slater teachers is 4.2/145 ≈ .03. Considering the view
of the physicists, our prior distribution should have a fair bit of mass on values of
θ ≈ .03. And considering the epidemiological studies and the likelihood that effects
would have been detected before 1992 if they were strong, our prior distribution
should put most of its mass below θ ≈ .06. For the sake of argument let’s adopt
the prior depicted in Figure 2.31. It’s formula is
Γ(20)Γ(400) 19
p(θ) = θ (1 − θ)399 (2.13)
Γ(420)
which we will see in Section 5.6 is the Be(20, 400) density. The likelihood function
is `(θ) ∝ θ8 (1 − θ)137 (Equation 2.3, Figure 2.19). Therefore the posterior density
p(θ | y) ∝ θ27 (1 − θ)536 which we will see in Section 5.6 is the Be(28, 537) density.
Therefore we can easily write down the constant and get the posterior density
Γ(28)Γ(537) 27
p(θ | y) = θ (1 − θ)536
Γ(565)
which is also pictured in Figure 2.31.

Examples 2.14 and 2.15 have the convenient feature that the prior density
had the same form — λa e−bλ in one case and θ a (1 − θ)b in the other — as
the likelihood function, which made the posterior density and the constant c
particularly easy to calculate. This was not a coincidence. The investigators
knew the form of the likelihood function and looked for a convenient prior of the
same form that approximately represented their prior beliefs. This convenience,
and whether choosing a prior density for this property is legitimate, are topics
which deserve serious thought but which we shall not take up at this point.

2.6 Prediction
Sometimes the goal of statistical analysis is to make predictions for future ob-
servations. Let y1 , . . . , yn , yf be a sample from p(· | θ). We observe y1 , . . . , yn
but not yf , and want a prediction for yf . There are three common forms that
predictions take.
point predictions A point prediction is a single guess for yf . It might be a
predictive mean, predictive median, predictive mode, or any other type of
point prediction that seems sensible.
2.6. PREDICTION 149

40

prior
likelihood
30

posterior
20
10
0

0.00 0.05 0.10 0.15 0.20

Figure 2.31: Prior, likelihood and posterior density for Slater School
150 CHAPTER 2. MODES OF INFERENCE

interval predictions An interval prediction or predictive interval, is an inter-

val of plausible values for yf . A predictive interval is accompanied by
a probability. For example, we might say that “The interval (0, 5) is a
90% predictive interval for yf ” which would mean Pr[yf ∈ (0, 5)] = .90.
In a given problem there are, for two reasons, many predictive intervals.
First, there are 90% intervals, 95% intervals, 50% intervals, and so on.
And second, there are many predictive intervals with the same probabil-
ity. For instance, if (0, 5) is a 90% predictive interval, then it’s possible
that (−1, 4.5) is also a 90% predictive interval.

predictive distributions A predictive distribution is a probability distribu-

tion for yf . From a predictive distribution, different people could compute
point predictions or interval predictions, each according to their needs.

In the real world, we don’t know θ. After all, that’s why we collected data
y1 , . . . , yn . But for now, to clarify the types of predictions listed above, let’s pre-
tend that we do know θ. Specifically, let’s pretend that we know y1 , . . . , yn , yf ∼
i.i.d. N(−2, 1).
The main thing to note, since we know θ (in this case, the mean and SD of
the Normal distribution), is that y1 , . . . , yn don’t help us at all. That is, they
contain no information about yf that is not already contained in the knowledge
of θ. In other words, y1 , . . . , yn and yf are conditionally independent given θ.
In symbols:
p(yf | θ, y1 , . . . , yn ) = p(yf | θ).
Therefore, our prediction should be based on the knowledge of θ alone, not on
any aspect of y1 , . . . , yn .
A sensible point prediction for yf is ŷf = −2, because -2 is the mean, median,
and mode of the N(−2, 1) distribution. Some sensible 90% prediction intervals
are (−∞, −0.72), (−3.65, −0.36) and (−3.28, ∞). We would choose one or the
other depending on whether we wanted to describe the lowest values that yf
might take, a middle set of values, or the highest values. And, of course, the
predictive distribution of yf is N(−2, 1). It completely describes the extent of
our knowledge and ability to predict yf .
In real problems, though, we don’t know θ. The simplest way to make a
prediction consists of two steps. First use y1 , . . . , yn to estimate θ, then make
predictions based on p(yf | θ̂). Predictions made by this method are called plug-
in predictions. In the example of the previous paragraph, if y1 , . . . , yn yielded
µ̂ = −2 and σ̂ = 1, then predictions would be exactly as described above.
For an example with discrete data, refer to Examples 1.4 and 1.6 in which
λ is the arrival rate of new seedlings. We found λ̂ = 2/3. The entire plug-in
predictive distribution is displayed in Figure 2.32. ŷf = 0 is a sensible point
prediction. The set {0, 1, 2} is a 97% plug-in prediction interval or prediction
set (because ppois(2,2/3) ≈ .97); the set {0, 1, 2, 3} is a 99.5% interval.
There are two sources of uncertainty in making predictions. First, because
yf is random, we couldn’t predict it perfectly even if we knew θ. And second,
we don’t know θ. In any given problem, either one of the two might be the
2.6. PREDICTION 151

0.4
probability

0.2
0.0

0 2 4 6 8 10

Figure 2.32: Plug-in predictive distribution yf ∼ Poi(λ = 2/3) for the seedlings
example

more important source of uncertainty. The first type of uncertainty can’t be

eliminated. But in theory, the second type can be reduced by collecting an
increasingly large sample y1 , . . . , yn so that we know θ with ever more accuracy.
Eventually, when we know θ accurately enough, the second type of uncertainty
becomes negligible compared to the first. In that situation, plug-in predictions
do capture almost the full extent of predictive uncertainty.
But in many practical problems the second type of uncertainty is too large
to be ignored. Plug-in predictive intervals and predictive distributions are too
optimistic because they don’t account for the uncertainty involved in estimating
θ. A Bayesian approach to prediction can account for this uncertainty. The
prior distribution of θ and the conditional distribution of y1 , . . . , yn , yf given θ
provide the full joint distribution of y1 , . . . , yn , yf , θ, which in turn provides the
conditional distribution of yf given y1 , . . . , yn . Specifically,
Z
p(yf | y1 , . . . , yn ) = p(yf , θ | y1 , . . . , yn ) dθ
Z
= p(θ | y1 , . . . , yn )p(yf | θ, y1 , . . . , yn ) dθ (2.14)
Z
= p(θ | y1 , . . . , yn )p(yf | θ) dθ

Equation 2.14 is just the yf marginal density derived from the joint density of
(θ, yf ), all densities being conditional on the data observed so far. To say it
152 CHAPTER 2. MODES OF INFERENCE
R R
another way, the predictive density p(yf ) is p(θ, yf ) dθ = p(θ)p(yf | θ) dθ,
but where p(θ) is really the posterior p(θ | y1 , . . . , yn ). The role of y1 , . . . , yn is
to give us the posterior density of θ instead of the prior.
The predictive distribution in Equation 2.14 will be somewhat more dis-
persed than the plug-in predictive distribution. If we don’t know much about θ
then the posterior will be widely dispersed and Equation 2.14 will be much more
dispersed than the plug-in predictive distribution. On the other hand, if we know
a lot about θ then the posterior distribution will be tight and Equation 2.14 will
be only slightly more dispersed than the plug-in predictive distribution.

Example 2.16 (Seedlings, cont.)

Refer to Examples 1.4 and 2.14 about Y , the number of new seedlings emerging
each year in a forest quadrat. Our model is Y ∼ Poi(λ). The prior (page 143) was
λ ∼ Γ(3, 2). Before collecting any data our predictive distribution would be based
on that prior. For any number y we could calculate
Z
pYf (y) ≡ P[Yf = y] = pYf | Λ (y | λ)pΛ (λ) dλ
Z y −λ 3
λ e 2
= λ2 e−2λ dλ
y! Γ(3)
Z
23
= λy+2 e−3λ dλ
y!Γ(3)
Z
23 Γ(x + 3) 3y+3
= λy+2 e−3λ dλ
y!Γ(3)3y+3 Γ(y + 3)
   3  y
y+2 2 1
= ,
y 3 3
which is the NegBin(3, 2/3) distribution. So Yf ∼ NegBin(3, 2/3).
In the first quadrat we found y1 = 3 and the posterior distribution Λ | y1 ∼
Γ(6, 3). So after the first observation the predictive distribution is
Z
pYf | Y1 (y | 3) = pYf | Λ (y | λ)pΛ | Y1 (λ | 3) dλ

which is, by calculations similar to those above, the NegBin(6, 1/2) distribution. So
Yf | y1 ∼ NegBin(6, 1/2).
Finally, when we collected data from 60 quadrats, we found Λ | y 1 , . . . , y60 ∼
Gam(43, 62). Therefore the predictive distribution is
Yf | y1 , . . . , y60 ∼ NegBin(43, 43/62).
A priori, and after only n = 1 observation, λ is not know very precisely; both
types of uncertainty are important; and the Bayesian predictive distribution is notica-
bly different from the plug-in predictive distribution. But after n = 60 observations
λ is known fairly well; the second type of uncertainty is negligible; and the Bayesian
predictive distribution is very similar to the plug-in predictive distribution.
2.6. PREDICTION 153

0.5

n=0
n=1
n=60
plug−in
0.4
0.3
0.2
0.1
0.0

0 2 4 6 8 10

Figure 2.33: Predictive distributions of yf in the seedlings example after samples

of size n = 0, 1, 60, and the plug-in predictive
154 CHAPTER 2. MODES OF INFERENCE

2.7 Hypothesis Testing

Scientific inquiry often takes the form of hypothesis testing. In each instance
there are two hypotheses — the null hypothesis H0 and the alternative hypoth-
esis Ha .

medicine

• H0 : the new drug and the old drug are equally effective.
• Ha : the new drug is better than the old.

public health

• H0 exposure to high voltage electric lines is benign.

• Ha exposure to high voltage electric lines promotes cancer.

public policy

• H0 : Head Start has no effect.

• Ha : Head Start is beneficial.

astronomy

• H0 : The sun revolves around the Earth.

• Ha : The Earth revolves around the sun.

physics

• H0 : Newtonian mechanics holds.

• Ha : Relativity holds.

public trust

• H0 : Winning lottery numbers are random.

• Ha : Winning lottery numbers have patterns.

ESP

• H0 : There is no ESP.
• Ha : There is ESP.

ecology

• H0 : Forest fires are irrelevant to forest diversity.

• Ha : Forest fires enhance forest diversity.
2.7. HYPOTHESIS TESTING 155

By tradition H0 is the hypothesis that says nothing interesting is going on

or the current theory is correct, while Ha says that something unexpected is
happening or our current theories need updating. Often the investigator is
hoping to disprove the null hypothesis and to suggest the alternative hypothesis
in its place.
It is worth noting that while the two hypotheses are logically exclusive,
they are not logically exhaustive. For instance, it’s logically possible that forest
fires decrease diversity even though that possibility is not included in either
hypothesis. So one could write Ha : Forest fires decrease forest diversity, or even
Ha : Forest fires change forest diversity. Which alternative hypothesis is chosen
makes little difference for the theory of hypothesis testing, though it might make
a large difference to ecologists.
Statisticians have developed several methods called hypothesis tests. We
focus on just one for the moment, useful when H0 is specific. The fundamental
idea is to see whether the data are “compatible” with the specific H 0 . If so, then
there is no reason to doubt H0 ; if not, then there is reason to doubt H0 and
possibly to consider Ha in its stead. The meaning of “compatible” can change
from problem to problem but typically there is a three step process.

1. Formulate a scientific null hypothesis and translate it into statistical terms.

2. Choose a low dimensional statistic, say w = w(y1 , . . . , yn ) such that the

distribution of w is specified under H0 and likely to be different under Ha .

3. Calculate, or at least approximate, the distribution of w under H0 .

4. Check whether the observed value of w, calculated from y1 , . . . , yn , is

compatible with its distribution under H0 .

How would this work in the examples listed at the beginning of the chapter?
What follows is a very brief description of how hypothesis tests might be carried
out in some of those examples. To focus on the key elements of hypothesis
testing, the descriptions have been kept overly simplistic. In practice, we would
have to worry about confounding factors, the difficulties of random sampling,
and many other issues.

public health Sample a large number of people with high exposure to power
lines. For each person, record Xi , a Bernoulli random variable indicating
whether that person has cancer. Model X1 , . . . , Xn ∼ i.i.d. Bern(θ1 ).
Repeat for a sample of people with low exposure; getting Y1 , . . . , Yn ∼
i.i.d. Bern(θ2 ). Estimate θ1 and θ2 . Let w = θ̂1 − θ̂2 . H0 says E[w] = 0.
Either the Binomial distribution or the Central Limit Theorem tells us
the SD’s of θ̂1 and θ̂2 , and hence the SD of w. Ask How many SD’s is
w away from its expected value of 0. If it’s off by many SD’s, more than
about 2 or 3, that’s evidence against H0 .

public policy Test a sample children who have been through Head Start.
Model their test scores as X1 , . . . , Xn ∼ i.i.d. N(µ1 , σ1 ). Do the same
156 CHAPTER 2. MODES OF INFERENCE

for children who have not been through Head Start, getting Y1 , . . . , Yn ∼
i.i.d. N(µ2 , σ2 ). H0 says µ1 = µ2 . Let w = µ̂1 − µ̂2 . The parameters
µ1 , µ2 , σ1 , σ2 can all be estimated from the data; therefore w can be cal-
culated and its SD estimated. Ask How many SD’s is w away from its
expected value of 0. If it’s off by many SD’s, more than about 2 or 3, that’s
evidence against H0 .
ecology We could either do an observational study, beginning with one sample
of plots that had had frequent forest fires in the past and another sample
that had had few fires. Or we could do an experimental study, beginning
with a large collection of plots and subjecting half to a regime of regular
burning and the other half to a regime of no burning. In either case we
would measure and compare species diversity in both sets of plots. If
diversity is similar in both groups, there is no reason to doubt H0 . But if
diversity is sufficiently different (Sufficient means large compared to what
is expected by chance under H0 .) that would be evidence against H0 .
To illustrate in more detail, let’s consider testing a new blood pressure medi-
cation. The scientific null hypothesis is that the new medication is not any more
effective than the old. We’ll consider two ways a study might be conducted and
see how to test the hypothesis both ways.
Method 1 A large number of patients are enrolled in a study and their blood
pressure is measured. Half are randomly chosen to receive the new medication
(treatment); half receive the old (control). After a prespecified amount of time,
their blood pressure is remeasured. Let YC,i be the change in blood pressure
from the beginning to the end of the experiment for the i’th control patient
and YT,i be the change in blood pressure from the beginning to the end of the
experiment for the i’th treatment patient. The model is
2
YC,1 , . . . YC,n ∼ i.i.d. fC ; E[YC,i ] = µC ; Var(YC,i ) = σC
YT,1 , . . . YT,n ∼ i.i.d. fT ; E[YT,i ] = µT ; Var(YT,i ) = σT2

for some unknown means µC and µT and variances σC and σT . The translation
of the hypotheses into statistical terms is

H 0 : µT = µ C
Ha : µT 6= µC

Because we’re testing a difference in means, let w = ȲT − ȲC . If the sample
size n is reasonably large, then the Central Limit Theorem says approximately
2 2
w ∼ N(0, σw ) under H0 with σw 2
= (σT2 + σC )/n. The mean of 0 comes from H0 .
2
The variance σw comes from adding variances of independent random variables.
σT2 and σC2
and therefore σw2
can be estimated from the data. So we can calculate
w from the data and see whether it is within about 2 or 3 SD’s of where H0
says it should be. If it isn’t, that’s evidence against H 0 .
Method 2 A large number of patients are enrolled in a study and their
blood pressure is measured. They are matched together in pairs according to
2.7. HYPOTHESIS TESTING 157

relevant medical characteristics. The two patients in a pair are chosen to be as

similar to each other as possible. In each pair, one patient is randomly chosen
to receive the new medication (treatment); the other receives the old (control).
After a prespecified amount of time their blood pressure is measured again. Let
YT,i and YC,i be the change in blood pressure for the i’th treatment and i’th
control patients. The researcher records
(
1 if YT,i > YC,i
Xi =
0 otherwise

The model is
X1 , . . . , Xn ∼ i.i.d. Bern(p)
for some unknown probability p. The translation of the hypotheses into statis-
tical terms is

H0 : p = .5
Ha : p 6= .5
P
Let w = Xi . Under H0 , w ∼ Bin(n, .5). To test H0 we plot the Bin(n, .5)
distribution and see where w falls on the plot . Figure 2.34 shows the plot for
n = 100. If w turned out to be between about 40 and 60, then there would be
little reason to doubt H0 . But on the other hand, if w turned out to be less
than 40 or greater than 60, then we would begin to doubt. The larger |w − 50|,
the greater the cause for doubt.
This blood pressure example exhibits a feature common to many hypothesis
tests. First, we’re testing a difference in means. I.e., H0 and Ha disagree about
a mean, in this case the mean change in blood pressure from the beginning to
the end of the experiment. So we take w to be the difference in sample means.
Second, since the experiment is run on a large number of people, the Central
Limit Theorem says that w will be approximately Normally distributed. Third,
we can calculate or estimate the mean µ0 and SD σ0 under H0 . So fourth,
we can compare the value of w from the data to what H0 says its distribution
should be.
In Method 1 above, that’s just what we did. In Method 2 above, we didn’t
use the Normal approximation; we used the Binomial distribution. But we could
have used the approximation.√ From facts about the Binomial distribution we
know µ0 = n/2 and σ0 = n/2 under H0 . For n = 100, Figure 2.35 compares
the exact Binomial distribution to the Normal approximation.
In general, when the Normal approximation is valid, we compare w to the
N(µ0 , σ0 ) density, where µ0 is calculated according to H0 and σ0 is either calcu-
lated according to H0 or estimated from the data. If t ≡ |w − µ0 |/σ0 is bigger
than about 2 or 3, that’s evidence against H0 .
The following example shows hypothesis testing at work.

Example 2.17 (Tooth Growth, continued)

This continues Example 2.1. Let’s concentrate on a particular dosage, say dose =
158 CHAPTER 2. MODES OF INFERENCE

0.08
0.04
0.00

30 40 50 60 70

Figure 2.34: pdf of the Bin(100, .5) distribution

0.08
0.04
0.00

30 40 50 60 70

Figure 2.35: pdfs of the Bin(100, .5) (dots) and N(50, 5) (line) distributions
2.7. HYPOTHESIS TESTING 159

0.5, and test the null hypothesis that, on average, the delivery method (supp) makes
no difference to tooth growth, as opposed to the alternative that it does make a
difference. Those are the scientific hypotheses. The data for testing the hypothesis
are x1 , . . . , x10 , the 10 recordings of growth when supp = VC and y1 , . . . , y10 , the
10 recordings of growth when supp = OJ. The xi ’s are 10 independent draws from
one distribution; the yi ’s are 10 independent draws from another:

x1 , . . . , x10 ∼ i.i.d. fVC

y1 , . . . , y10 ∼ i.i.d. fOJ

Define the two means to be µVC ≡ E[xi ] and µOJ ≡ E[yi ]. The scientific hypothesis
and its alternative, translated into statistical terms become

H0 : µVC = µOJ
Ha : µVC 6= µOJ

Those are the hypotheses in statistical terms.

Because we’re testing a difference in means, we choose our one dimensional
summary statistic to be t = |x̄ − ȳ|. Small values of t support H 0 ; large values
support Ha . But how small is small; how large is large? The Central Limit Theorem
says
 
σVC
x̄ ∼ N µVC , √
n
 
σOJ
ȳ ∼ N µOJ , √
n

approximately, so that under H0 ,

r !
2 + σ2
σVC OJ
t ∼ N 0, ,
n

approximately. The statistic t can be calculated, its SD estimated, and its approx-
imate density plotted as in Figure 2.36. We can see from the Figure, or from the
fact that t/σt ≈ 3.2 that the observed value of t is moderately far from its expected
value under H0 . The data provide moderately strong, evidence against H 0 .

Figure 2.36 was produced with the following R code.

x <- ToothGrowth$len[ ToothGrowth$supp=="VC"
& ToothGrowth$dose==0.5 ]
y <- ToothGrowth$len[ ToothGrowth$supp=="OJ"
& ToothGrowth$dose==0.5 ]
160 CHAPTER 2. MODES OF INFERENCE

0.15
0.00

−6 −4 −2 0 2 4 6

Figure 2.36: Approximate density of summary statistic t. The black dot is the
value of t observed in the data.

t <- abs ( mean(x) - mean(y) )

sd <- sqrt ( ( var(x) + var(y) )/length(x) )
tvals <- seq ( -4*sd, 4*sd, len=80 )
plot ( tvals, dnorm(tvals,0,sd), type="l",
xlab="", ylab="", main="" )
points ( t, 0, pch=16, cex=1.5 )

The points(...) adds the observed value of t to the plot.

In the next example it is difficult to estimate the distribution of w under H0 ;

so we use simulation to work it out.

Example 2.18 (Baboons)

Because baboons are promiscuous, when a baby is born it is not obvious, at least
to humans, who the father is. But do the baboons themselves know who the father
is? Buchan et al. [2003] report a study of baboon behavior that attempts to answer
that question. For more information see http://www.princeton.edu/~baboon.
Baboons live in social groups comprised of several adult males, several adult females,
and juveniles. Researchers followed several groups of baboons periodically over a
period of several years to learn about baboon behavior. The particular aspect of
behavior that concerns us here is that adult males sometimes come to the aid of
juveniles. If adult males know which juveniles are their own children, then it’s at
2.7. HYPOTHESIS TESTING 161

least possible that they tend to aid their own children more than other juveniles.
The data set baboons (available on the web site)1 contains data on all the recorded
instances of adult males helping juveniles. The first four lines of the file look like
this.
Recip Father Maleally Dadpresent Group
ABB EDW EDW Y OMO
ABB EDW EDW Y OMO
ABB EDW EDW Y OMO
ABB EDW POW Y OMO

1. Recip identifies the juvenile who received help. In the four lines shown here,
it is always ABB.
2. Father identifes the father of the juvenile. Researchers know the father
through DNA testing of fecal samples. In the four lines shown here, it is
always EDW.
3. Maleally identifies the adult male who helped the juvenile. In the fourth line
we see that POW aided ABB who is not his own child.
4. Dadpresent tells whether the father was present in the group when the ju-
venile was aided. In this data set it is always Y.
5. Group identifies the social group in which the incident occured. In the four
lines shown here, it is always OMO.

Let w be the number of cases in which a father helps his own child. The snippet
dim ( baboons )
sum ( baboons$Father == baboons$Maleally )

reveals that there are n = 147 cases in the data set, and that w = 87 are cases in
which a father helps his own child. The next step is to work out the distribution of
w under H0 : adult male baboons do not know which juveniles are their children.
Let’s examine one group more closely, say the OMO group. Typing
baboons[baboons$Group == "OMO",]

displays the relevant records. There are 13 of them. EDW was the father in 9, POW
was the father in 4. EDW provided the help in 9, POW in 4. The father was the
ally in 9 cases; in 4 he was not. H0 implies that EDW and POW would distribute
their help randomly among the 13 cases. If H0 is true, i.e., if EDW distributes his 9
helps and POW distributes his 4 helps randomly among the 13 cases, what would
be the distribution of W , the number of times a father helps his own child? We
1 We have slightly modified the data to avoid some irrelevant complications.
162 CHAPTER 2. MODES OF INFERENCE

can answer that question by a simulation in R. (We could also answer it by doing
some math or by knowing the hypergeometric distribution, but that’s not covered
in this text.)
dads <- baboons$Father [ baboons$Group == "OMO" ]
ally <- baboons$Maleally [ baboons$Group == "OMO" ]
N.sim <- 1000
w <- rep ( NA, N.sim )
for ( i in 1:N.sim ) {
perm <- sample ( dads )
w[i] <- sum ( perm == ally )
}
hist(w)
table(w)

Try out the simulation for yourself. It shows that the observed number in the data,
w = 9, is not so unusual under H0 .
What about the other social groups? If we find out how many there are, we can
do a similar simulation for each. Let’s write an R function to help.

g.sim <- function (group, N.sim) {

dads <- baboons$Father [ baboons$Group == group ]
ally <- baboons$Maleally [ baboons$Group == group ]
w <- rep ( NA, N.sim )
for ( i in 1:N.sim ) {
perm <- sample ( dads )
w[i] <- sum ( perm == ally )
}
return(w)
}

Figure 2.37 shows histograms of g.sim for each group, along with a dot showing the
observed value of w in the data set. For some of the groups the observed value of w,
though a bit on the high side, might be considered consistent with H 0 . For others,
the observed value of w falls outside the range of what might be reasonably expected
by chance. In a case like this, where some of the evidence is strongly against H 0
and some is only weakly against H0 , an inexperienced statistician might believe the
overall case against H0 is not very strong. But that’s not true. In fact, every one of
the groups contributes a little evidence against H 0 , and the total evidence against
H0 is very strong. To see this, we can combine the separate simulations into one.
The following snippet of code does this. Each male’s help is randomly reassigned
to a juvenile within his group. The number of times when a father helps his own
child is summed over the different groups. Simulated numbers are shown in the
histogram in Figure 2.38. The dot in the figure is at 84, the actual number of
instances in the full data set. Figure 2.38 suggests that it is almost impossible that
2.8. EXERCISES 163

the 84 instances arose by chance, as H0 would suggest. We should reject H0 and

reach the conclusion that (a) adult male baboons do know who their own children
are, and (b) they give help preferentially to their own children.

Figure 2.37 was produced with the following snippet.

groups <- unique ( baboons$Group )
n.groups <- length(groups)
par ( mfrow=c(3,2) )
for ( i in 1:n.groups ) {
good <- baboons$Group == groups[i]
w.obs <- sum ( baboons$Father[good]
== baboons$Maleally[good] )
w.sim <- g.sim ( groups[i], N.sim )
hist ( w.sim, xlab="w", ylab="", main=groups[i],
xlim=range(c(w.obs,w.sim)) )
points ( w.obs, 0, pch=16, cex=1.5 )
print ( w.obs )
}

Figure 2.38 was produced with the following snippet.

w.obs <- rep ( NA, n.groups )
w.sim <- matrix ( NA, n.groups, N.sim )
for ( i in 1:n.groups ) {
good <- baboons$Group == groups[i]
w.obs[i] <- sum ( baboons$Father[good]
== baboons$Maleally[good] )
w.sim[i,] <- g.sim ( groups[i], N.sim )
}
w.obs.tot <- sum ( w.obs )
w.sim.tot <- apply ( w.sim, 2, sum )
hist ( w.sim.tot, xlab="w.tot", ylab="",
xlim=range(c(w.obs.tot,w.sim.tot)) )
points ( w.obs.tot, 0, pch=16, cex=1.5 )
print ( w.obs.tot )

2.8 Exercises
1. (a) Justify Equation 2.1.
(b) Show that the function g(x) defined just below Equation 2.1 is a
probability density. I.e., show that it integrates to 1.
164 CHAPTER 2. MODES OF INFERENCE

OMO VIV

400
400

200
200
0

0
5 6 7 8 9 10 11 8 10 12 14 16

w w

NYA WEA

300
300

200
200

100
100
0

0 1 2 3 4 5 15 20 25 30

w w

LIN
100 150
50
0

8 10 14 18

Figure 2.37: Number of times baboon father helps own child in Example 2.18.
Histograms are simulated according to H0 . Dots are observed data.
2.8. EXERCISES 165

200
100
0

50 60 70 80

w.tot

Figure 2.38: Histogram of simulated values of w.tot. The dot is the value
observed in the baboon data set.

2. This exercise uses the ToothGrowth data from Examples 2.1 and 2.17.

(a) Estimate the effect of delivery mode for doses 1.0 and 2.0. Does it
seem that delivery mode has a different effect at different doses?
(b) Does it seem as though delivery mode changes the effect of dose?
(c) For each delivery mode, make a set of three boxplots to compare the
three doses.

3. This exercise uses data from 272 eruptions of the Old Faithful geyser in
Yellowstone National Park. The data are in the R dataset faithful.
One column contains the duration of each eruption; the other contains
the waiting time to the next eruption.

(a) Plot eruption versus waiting. Is there a pattern? What is going

on?
(b) Try ts.plot(faithful$eruptions[1:50]). Try other sets of erup-
tions, say ts.plot(faithful$eruptions[51:100]). (There is noth-
ing magic about 50, but if you plot all 272 eruptions then the pattern
might be harder to see. Choose any convenient number that lets you
see what’s going on.) What is going on?

4. This Exercise relies on Example 2.7 about the Slater school. There were
8 cancers among 145 teachers. Figure 2.19 shows the likelihood function.
166 CHAPTER 2. MODES OF INFERENCE

Suppose the same incidence rate had been found among more teachers.
How would that affect `(θ)? Make a plot similar to Figure 2.19, but
pretending that there had been 80 cancers among 1450 teachers. Compare
to Figure 2.19. What is the result? Does it make sense? Try other numbers
if it helps you see what is going on.

5. This exercise continues Exercise 34 in Chapter 1. Let p be the fraction of

the population that uses illegal drugs.

(a) Suppose researchers know that p ≈ .1. Jane and John are given the
randomized response question. Jane answers “yes”; John answers
“no”. Find the posterior probability that Jane uses cocaine; find the
posterior probability that John uses cocaine.
(b) Now suppose that p is not known and the researchers give the ran-
domized response question to 100 people. Let X be the number who
answer “yes”. What is the likelihood function?
(c) What is the mle of p if X=50, if X=60, if X=70, if X=80, if X=90?

6. This exercise deals with the likelihood function for Poisson distributions.

(a) Let x1 , . . . , xn ∼ i.i.d. Poi(λ). Find `(λ) in terms of x1 , . . . , xn .

P
(b) Show that `(λ) depends only on xi and not on the specific values
of the individual xi ’s.
(c) Let y1 , . . . , yn be a sample from Poi(λ). Show that λ̂ = ȳ is the m.l.e.
(d) Find the m.l.e. in Example 1.4.

7. The book Data Andrews and Herzberg [1985] contains lots of data sets
that have been used for various purposes in statistics. One famous data set
records the annual number of deaths by horsekicks in the Prussian Army
from 1875-1894 for each of 14 corps. Download the data from statlib at
http://lib.stat.cmu.edu/datasets/Andrews/T04.1. (It is Table 4.1
in the book.) Let Yij be the number of deaths in year i, corps j, for
i = 1875, . . . , 1894 and j = 1, . . . , 14. The Yij s are in columns 5–18 of the
table.

(a) What are the first four columns of the table?

(b) What is the last column of the table?
(c) What is a good model for the data?
(d) Suppose you model the data as i.i.d. Poi(λ). (Yes, that’s a good
answer to the previous question.)
i. Plot the likelihood function for λ.
ii. Find λ̂.
iii. What can you say about the rate of death by horsekick in the
Prussian calvary at the end of the 19th century?
2.8. EXERCISES 167

(e) Is there any evidence that different corps had different death rates?
How would you investigate that possibility?

8. Use the data from Example 2.7. Find the m.l.e. for θ.

9. X1 , . . . , Xn ∼ Normal(µ, 1). Multiple choice: The m.l.e. µ̂ is found from

the equation
d d
(a) dµ dx f (x1 , . . . , xn |µ) =0
d
(b) dµ f (x1 , . . . , xn |µ) =0
d
(c) dx f (x1 , . . . , xn |µ) =0

10. This exercise deals with the likelihood function for Normal distributions.

(a) Let y1 , . . . , yn ∼ i.i.d. N(µ, 1). Find `(µ) in terms of y1 , . . . , yn .

P
(b) Show that `(µ) depends only on yi and not on the specific values
of the individual yi ’s.
(c) Let n = 10 and choose a value for µ. Use R to generate a sample of
size 10 from N(µ, 1). Plot the likelihood function. How accurately
can you estimate µ from a sample of size 10?
(d) Let y1 , . . . , y10 ∼ i.i.d. N(µ, σ) where σ is known but not necessarily
equal to 1. Find `(µ) in terms of y1 , . . . , y10 and σ.
(e) Let y1 , . . . , y10 ∼ i.i.d. N(µ, σ) where µ is known but σ is unknown.
Find `(σ) in terms of y1 , . . . , y10 and µ.

11. Let y1 , . . . , yn be a sample from N(µ, 1). Show that µ̂ = ȳ is the m.l.e.

12. Let y1 , . . .P
, yn be a sample from N(µ, σ) where µ is known. Show that
σ̂ 2 = n−1 (yi − µ)2 is the m.l.e.

13. Recall the discoveries data from page 9 on the number of great dis-
coveries each year. Let Yi be the number of great discoveries in year i
and suppose Yi ∼ Poi(λ). Plot the likelihood function `(λ). Figure 1.3
suggested that λ ≈ 3.1 explained the data reasonably well. How sure can
we be about the 3.1?

14. Justify each step of Equation 2.6.

15. Page 135 discusses a simulation experiment comparing the sample mean
and sample median as estimators of a population mean. Figure 2.26 shows
the results of the simulation experiment. Notice that the vertical scale
decreases from panel (a) to (b), to (c), to (d). Why? Give a precise
mathematical formula for the amount by which the vertical scale should
decrease. Does the actual decrease agree with your formula?

16. In the medical screening example on page 141, find the probability that
the patient has the disease given that the test is negative.
168 CHAPTER 2. MODES OF INFERENCE

17. A drug testing example

18. Country A suspects country B of having hidden chemical weapons. Based
on secret information from their intelligence agency they calculate
P[B has weapons] = .8. But then country B agrees to inspections, so A
sends inspectors. If there are no weapons then of course the inspectors
won’t find any. But if there are weapons then they will be well hidden,
with only a 20% chance of being found. I.e.,

P[finding weapons|weapons exist] = .2. (2.15)

No weapons are found. Find the probability that B has weapons. I.e.,
find
Pr[B has weapons|no weapons are found].

19. Let T be the amount of time a customer spends on Hold when calling the
computer help line. Assume that T ∼ exp(λ) where λ is unknown. A
sample of n calls is randomly selected. Let t1 , . . . , tn be the times spent
on Hold.
(a) Choose a value of λ for doing simulations.
(b) Use R to simulate a sample of size n = 10.
(c) Plot `(λ) and find λ̂.
(d) About how accurately can you determine λ?
P
(e) Show that `(λ) depends only on ti and not on the values of the
individual ti ’s.
20. The Great Randi is a professed psychic and claims to know the outcome
of coin flips. This problem concerns a sequence of 20 coin flips that Randi
will try to guess (or not guess, if his claim is correct).
(a) Take the prior P[Randi is psychic] = .01.
i. Before any guesses have been observed, find
P[first guess is correct] and P[first guess is incorrect].
ii. After observing 10 consecutive correct guesses, find the updated
P[Randi is psychic].
iii. After observing 10 consecutive correct guesses, find
P[next guess is correct] and P[next guess is incorrect].
iv. After observing 20 consecutive correct guesses, find
P[next guess is correct] and P[next guess is incorrect].
(b) Two statistics students, a skeptic and a believer discuss Randi after
class.
Believer: I believe her, I think she’s psychic.
Skeptic: I doubt it. I think she’s a hoax.
Believer: How could you be convinced? What if Randi guessed 10 in
2.8. EXERCISES 169

a row? What would you say then?

Skeptic: I would put that down to luck. But if she guessed 20 in a
row then I would say P[Randi can guess coin flips] ≈ .5.
Find the skeptic’s prior probability that Randi can guess coin flips.
(c) Suppose that Randi doesn’t claim to guess coin tosses perfectly, only
that she can guess them at better than 50%. 100 trials are conducted.
Randi gets 60 correct. Write down H0 and Ha appropriate for testing
Randi’s claim. Do the data support the claim? What if 70 were
correct? Would that support the claim?
(d) The Great Sandi, a statistician, writes the following R code to calcu-
late a probability for Randi.
y <- rbinom ( 500, 100, .5)
sum ( y == 60 ) / 500

What is Sandi trying to calculate? Write a formula (Don’t evaluate

it.) for the quantity Sandi is trying to calculate.
21. Let w be the fraction of free throws that Shaquille O’Neal (or any other
player of your choosing) makes during the next NBA season. Find a
density that approximately represents your prior opinion for w.
22. Let t be the amount of time between the moment when the sun first
touches the horizon in the afternoon and the moment when it sinks com-
pletely below the horizon. Without making any observations, assess your
distribution for t.
23. Assess your prior distribution for b, the proportion of M&M’s that are
brown. Buy as many M&M’s as you like and count the number of browns.
Calculate your posterior distribution.
24. (a) Let y ∼ N(θ, 1) and let the prior distribution for θ be θ ∼ N(0, 1).
When y has been observed, what is the posterior distribution for θ?
(b) Let y ∼ N(θ, σy ) and let the prior distribution for θ be θ ∼ N(m, σ).
When y has been observed, what is the posterior distribution for θ?
(c) Let y1 , . . . , yn be a sample of size n from N(θ, σy ) and let the prior
distribution for θ be θ ∼ N(m, σ). When y1 , . . . , yn have been ob-
served, what is the posterior distribution for θ?
(d) An example with data.
25. Refer to the discussion of predictive intervals on page 150. Justify the
claim that (−∞, −.72), (−3.65, −0.36), and (−3.28, ∞) are 90% prediction
intervals. Find the corresponding 80% prediction intervals.
26. (a) Following Example 2.16, find Pr[yf = k | y1 , . . . yn ] for k = 1, 2, 3, 4.
(b) Using the results from part (a), make a plot analagous to Figure 2.32.
170 CHAPTER 2. MODES OF INFERENCE

27. Suppose you want to test whether the random number generator in R
generates each of the digits 0, 1, . . . , 9 with probability 0.1. How could
you do it? You may consider first testing whether R generates 0 with the
right frequency, then repeating the analysis for each digit.
28. (a) Repeat the analysis of Example 2.17, but for dose = 1 and dose = 2.
(b) Test the hypothesis that increasing the dose from 1 to 2 makes no
difference in tooth growth.
(c) Test the hypothesis that the effect of increasing the dose from 1 to 2
is the same for supp = VC as it is for supp = OJ.
(d) Do the answers to parts (a), (b) and (c) agree with your subjective
assessment of Figures 2.2, 2.3, and 2.6?
29. Continue Exercise 7 from Chapter 5. The autoganzfeld trials resulted in
X = 122.
(a) What is the parameter in this problem?
(b) Plot the likelihood function.
(c) Test the “no ESP, no cheating” hypothesis.
(d) Adopt and plot a reasonable and mathematically tractable prior dis-
tribution for the parameter. Compute and plot the posterior distri-
bution.
(e) What do you conclude?
30. Three biologists named Asiago, Brie, and Cheshire are studying a muta-
tion in morning glories, a species of flowering plant. The mutation causes
the flowers to be white rather than colored. But it is not known whether
the mutation has any effect on the plants’ fitness. To study the question,
each biologist takes a random sample of morning glories having the muta-
tion, counts the seeds that each plant produces, and calculates a likelihood
set for the average number of seeds produced by mutated morning glories.
Asiago takes a sample of size nA = 100 and calculates a LS.1 set. Brie
takes a sample of size nB = 400 and calculates a LS.1 set. Cheshire takes
a sample of size nC = 100 and calculates a LS.2 set.
(a) Who will get the longer interval, Asiago or Brie? About how much
longer will it be? Explain.
(b) Who will get the longer interval, Asiago or Cheshire? About how
much longer will it be? Explain.
31. In the 1990’s, a committee at MIT wrote A Study on the Status of
Women Faculty in Science at MIT. In 1994 there were 15 women
among the 209 tenured women the in six departments of the School of
Science. They found, among other things, that the amount of resources
(money, lab space, etc.) given to women was, on average, less than the
2.8. EXERCISES 171

amount given to men. The report goes on to pose the question: Given
the tiny number of women faculty in any department one might ask if it is
possible to obtain significant data to support a claim of gender differences
....
What does statistics say about it? Focus on a single resource, say labo-
ratory space. The distribution of lab space is likely to be skewed. I.e.,
there will be a few people with lots more space than most others. So
let’s model the distribution of lab space with an Exponential distribu-
tion. Let x1 , . . . , x15 be the amounts of space given to tenured women, so
xi ∼ Exp(λw ) for some unknown parameter λw . Let M be the average
lab space given to tenured men. Assume that M is known to be 100,
from the large number of tenured men. If there is no discrimination, then
λw = 1/100 = .01. (λw is 1/E(xi ).)
Chris Stats writes the following R code.

y <- rexp(15,.01)
m <- mean(y)
s <- sqrt( var(y ))
lo <- m - 2*s
hi <- m + 2*s

What is y supposed to represent? What is (lo,hi) supposed to represent?

Now Chris puts the code in a loop.

n <- 0
for ( i in 1:1000 ) {
y <- rexp(15,.01)
m <- mean(y)
s <- sqrt ( var(y ))
lo <- m - 2*s
hi <- m + 2*s
if ( lo < 100 & hi > 100 ) n <- n+1
}
print (n/1000)

What is n/1000 supposed to represent? If a sample size of 15 is sufficiently

large for the Central Limit Theorem to apply, then what, approximately,
is the value of n/1000?
32. Refer to the R code in Example 2.1. Why was it necessary to have a brace
(“{”) after the line
for ( j in 1:3 )
but not after the line
for ( i in 1:2 )?
172 CHAPTER 2. MODES OF INFERENCE
 Chapter 3

Regression

3.1 Introduction
Regression is the study of how the distribution of one variable, Y , changes
according to the value of another variable, X. R comes with many data sets
that offer regression examples. Four are shown in Figure 3.1.

1. The data set attenu contains data on several variables from 182 earth-
quakes, including hypocenter-to-station distance and peak acceleration.
Figure 3.1 (a) shows acceleration plotted against distance. There is a
clear relationship between X = distance and the distribution of Y =
acceleration. When X is small, the distribution of Y has a long right-
hand tail. But when X is large, Y is always small.

2. The data set airquality contains data about air quality in New York
City. Ozone levels Y are plotted against temperature X in Figure 3.1 (b).
When X is small then the distribution of Y is concentrated on values
below about 50 or so. But when X is large, Y can range up to about 150
or so.

3. Figure 3.1 (c) shows data from mtcars. Weight is on the abcissa and the
type of transmission (manual=1, automatic=0) is on the ordinate. The
distribution of weight is clearly different for cars with automatic transmis-
sions than for cars with manual transmissions.

4. The data set faithful contains data about eruptions of the Old Faith-
ful geyser in Yellowstone National Park. Figure 3.1 (d) shows Y =
time to next eruption plotted against X = duration of current eruption.
Small values of X tend to indicate small values of Y .

Figure 3.1 was produced by the following R snippet.

par ( mfrow=c(2,2) )

173
174 CHAPTER 3. REGRESSION

(a) (b)

0.8

150
0.6
Acceleration

100
ozone
0.4

50
0.2
0.0

0
0 100 300 60 70 80 90

Distance temperature

(c) (d)
90
1
Manual Transmission

80
waiting

70
60
50
0

2 3 4 5 1.5 2.5 3.5 4.5

Weight eruptions

Figure 3.1: Four regression examples

3.1. INTRODUCTION 175

data ( attenu )
plot ( attenu$dist, attenu$accel, xlab="Distance",
ylab="Acceleration", main="(a)", pch="." )

data ( airquality )
plot ( airquality$Temp, airquality$Ozone, xlab="temperature",
ylab="ozone", main="(b)", pch="." )

data ( mtcars )
stripchart ( mtcars$wt ~ mtcars$am, pch=1, xlab="Weight",
method="jitter", ylab="Manual Transmission",
main="(c)" )

data ( faithful )
plot ( faithful, pch=".", main="(d)" )

Both continuous and discrete variables can turn up in regression problems.

In the attenu, airquality and faithful datasets, both X and Y are continu-
ous. In mtcars, it seems natural to think of how the distribution of Y = weight
varies with X = transmission, in which case X is discrete and Y is continuous.
But we could also consider how the fraction of cars Y with automatic transmis-
sions varies as a function of X = weight, in which case Y is discrete and X is
continuous.
In many regression problems we just want to display the relationship be-
tween X and Y . Often a scatterplot or stripchart will suffice, as in Figure 3.1.
Other times, we will use a statistical model to describe the relationship. The
statistical model may have unknown parameters which we may wish to estimate
or otherwise make inference for. Examples of parametric models will come later.
Our study of regression begins with data display.
In many instances a simple plot is enough to show the relationship between
X and Y . But sometimes the relationship is obscured by the scatter of points.
Then it helps to draw a smooth curve through the data. Examples 3.1 and 3.2
illustrate.

Example 3.1 (1970 Draft Lottery)

The result of the 1970 draft lottery is available at DASL . The website explains:

“In 1970, Congress instituted a random selection process for the

military draft. All 366 possible birth dates were placed in plastic capsules
in a rotating drum and were selected one by one. The first date drawn
from the drum received draft number one and eligible men born on that
date were drafted first. In a truly random lottery there should be no
relationship between the date and the draft number.”
176 CHAPTER 3. REGRESSION

Figure 3.2 shows the data, with X = day of year and Y = draft number. There
is no apparent relationship between X and Y .

Figure 3.2 was produced with the following snippet.

plot ( draft$Day.of.year, draft$Draft.No,
xlab="Day of year", ylab="Draft number" )

More formally, a relationship between X and Y usually means that the expected
value of Y is different for different values of X. (We don’t consider changes in SD or
other aspects of the distribution here.) Typically, when X is a continuous variable,
changes in Y are smooth, so we would adopt the model

E[Y | X] = g(X) (3.1)

for some unknown smooth function g.

R has a variety of built-in functions to estimate g. These functions are called
scatterplot smoothers, for obvious reasons. Figure 3.3 shows the draft lottery data
with two scatterplot smoother estimates of g. Both estimates show a clear trend,
that birthdays later in the year were more likely to have low draft numbers. check
this: Following discovery of this trend, the procedure for drawing draft numbers
was changed in subsequent years.

Figure 3.3 was produced with the following snippet.

x <- draft$Day.of.year
y <- draft$Draft.No
plot ( x, y, xlab="Day of year", ylab="Draft number" )
lines ( lowess ( x, y ) )
lines ( supsmu ( x, y ), lty=2 )

• lowess (locally weighted scatterplot smoother) and supsmu (super smoother)

are two of R’s scatterplot smoothers. In the figure, the lowess curve is less
wiggly than the supsmu curve. Each smoother has a tuning parameter that
can make the curve more or less wiggly. Figure 3.3 was made with the default
values for both smoothers.

Example 3.2 (Seedlings, continued)

As mentioned in Example 1.6, the seedlings study was carried out at the Coweeta
Long Term Ecological Research station in western North Carolina. There were five
plots at different elevations on a hillside. Within each plot there was a 60m×1m
strip running along the hillside divided into 60 1m×1m quadrats. It is possible
that the arrival rate of New seedlings and the survival rates of both Old and New
seedlings are different in different plots and different quadrats. Figure 3.4 shows the
3.1. INTRODUCTION 177

300
Draft number

200
100
0

0 100 200 300

Day of year

Figure 3.2: 1970 draft lottery. Draft number vs. day of year
178 CHAPTER 3. REGRESSION

300
Draft number

200
100
0

0 100 200 300

Day of year

Figure 3.3: 1970 draft lottery. Draft number vs. day of year. Solid curve fit by
lowess; dashed curve fit by supsmu.
3.1. INTRODUCTION 179

total number of New seedlings in each of the quadrats in one of the five plots. The
lowess curve brings out the spatial trend: low numbers to the left, a peak around
quadrat 40, and a slight falling off by quadrat 60.
25
total new seedlings

20
15
10
5
0

0 10 20 30 40 50 60

quadrat index

Figure 3.4: Total number of New seedlings 1993 – 1997, by quadrat.

Figure 3.4 was produced by

plot ( total.new, xlab="quadrat index",
ylab="total new seedlings" )
lines ( lowess ( total.new ) )

In a regression problem the data are pairs (xi , yi ) for i = 1, . . . , n. For each
i, yi is a random variable whose distribution depends on xi . We write

yi = g(xi ) + i . (3.2)

Equation 3.2 expresses yi as a systematic or explainable part g(xi ) and an

unexplained part i . g is called the regression function. Often the statistician’s
180 CHAPTER 3. REGRESSION

goal is to estimate g. As usual, the most important tool is a simple plot, similar
to those in Figures 3.1 through 3.4.
Once we have an estimate, ĝ, for the regression function g (either by a
scatterplot smoother or by some other technique) we can calculate ri ≡ yi −
ĝ(xi ). The ri ’s are estimates of the i ’s and are called residuals. The i ’s
themselves are called errors. Because the ri ’s are estimates they are sometimes
written with the “hat” notation:

ˆi = ri = estimate of i

Residuals are used to evaluate and assess the fit of models for g, a topic which
is beyond the scope of this book.
In regression we use one variable to explain or predict the other. It is cus-
tomary in statistics to plot the predictor variable on the x-axis and the predicted
variable on the y-axis. The predictor is also called the independent variable, the
explanatory variable, the covariate, or simply x. The predicted variable is called
the dependent variable, or simply y. (In Economics x and y are sometimes called
the exogenous and endogenous variables, respectively.) Predicting or explaining
y from x is not perfect; knowing x does not tell us y exactly. But knowing x
does tell us something about y and allows us to make more accurate predictions
than if we didn’t know x.
Regression models are agnostic about causality. In fact, instead of using x to
predict y, we could use y to predict x. So for each pair of variables there are two
possible regressions: using x to predict y and using y to predict x. Sometimes
neither variable causes the other. For example, consider a sample of cities and
let x be the number of churches and y be the number of bars. A scatterplot of
x and y will show a strong relationship between them. But the relationship is
caused by the population of the cities. Large cities have large numbers of bars
and churches and appear near the upper right of the scatterplot. Small cities
have small numbers of bars and churches and appear near the lower left.
Scatterplot smoothers are a relatively unstructured way to estimate g. Their
output follows the data points more or less closely as the tuning parameter allows
ĝ to be more or less wiggly. Sometimes an unstructured approach is appropriate,
but not always. The rest of Chapter 3 presents more structured ways to estimate
g.

3.2 Normal Linear Models

3.2.1 Introduction
In Section 1.3.4 we studied the Normal distribution, useful for continuous popu-
lations having a central tendency with roughly equally sized tails. In Section 3.2
we generalize to the case where there are many Normal distributions with dif-
ferent means which depend in a systematic way on another variable. We begin
our study with an example in which there are three distinct distributions.
3.2. NORMAL LINEAR MODELS 181

Example 3.3 (Hot Dogs, continued)

Figure 3.5 displays calorie data for three types of hot dogs. It appears that poultry
hot dogs have, on average, slightly fewer calories than beef or meat hot dogs. How
should we model these data?
Poultry
Meat
Beef

100 120 140 160 180

calories

Figure 3.5: Calorie content of hot dogs

Figure 3.5 was produced with

stripchart ( hotdogs$Calories ~ hotdogs$Type, pch=1,
xlab="calories" )

There are 20 Beef, 17 Meat and 17 Poultry hot dogs in the sample. We think of
them as samples from much larger populations. Figure 3.6 shows density estimates
of calorie content for the three types. For each type of hot dog, the calorie contents
cluster around a central value and fall off to either side without a particularly long
left or right tail. So it is reasonable, at least as a first attempt, to model the three
distributions as Normal. Since the three distributions have about the same amount
182 CHAPTER 3. REGRESSION

of spread we model them as all having the same SD. We adopt the model
B1 , . . . , B20 ∼ i.i.d. N(µB , σ)
M1 , . . . , M17 ∼ i.i.d. N(µM , σ) (3.3)
P1 , . . . , P17 ∼ i.i.d. N(µP , σ),
where the Bi ’s, Mi ’s and Pi ’s are the calorie contents of the Beef, Meat and Poultry
hot dogs respectively. Figure 3.6 suggests
µB ≈ 150; µM ≈ 160; µP ≈ 120; σ ≈ 30.
An equivalent formulation is
B1 , . . . , B20 ∼ i.i.d. N(µ, σ)
M1 , . . . , M17 ∼ i.i.d. N(µ + δM , σ) (3.4)
P1 , . . . , P17 ∼ i.i.d. N(µ + δP , σ)
Models 3.3 and 3.4 are mathematically equivalent. Each has three parameters
for the population means and one for the SD. They describe exactly the same set
of distributions and the parameters of either model can be written in terms of the
other. The equivalence is shown in Table 3.3. For the purpose of further exposition
we adopt Model 3.4.
We will see later how to carry out inferences regarding the parameters. For now
we stop with the model.

Figure 3.6 was produced with the following snippet.

par ( mfrow=c(3,1))
plot ( density ( hotdogs$C[hotdogs$T=="Beef"], bw=20 ),
xlim=c(50,250), yaxt="n", ylab="", xlab="calories",
main="Beef" )
plot ( density ( hotdogs$C[hotdogs$T=="Meat"], bw=20 ),
xlim=c(50,250), yaxt="n", ylab="", xlab="calories",
main="Meat" )
plot ( density ( hotdogs$C[hotdogs$T=="Poultry"], bw=20 ),
xlim=c(50,250), yaxt="n", ylab="", xlab="calories",
main="Poultry" )

• hotdogs$C and hotdogs$T illustrate a convenient feature of R, that compo-

nents of a structure can be abbreviated. Instead of typing hotdogs$Calories
and hotdogs$Type we can use the abbreviations. The same thing applies to
arguments of functions.
• density ( ..., bw=20 ) specifies the bandwidth of the density estimate.
Larger bandwidth gives a smoother estimate; smaller bandwidth gives a more
wiggly estimate. Try different bandwidths to see what they do.
3.2. NORMAL LINEAR MODELS 183

Beef

50 100 150 200 250

calories

Meat

50 100 150 200 250

calories

Poultry

50 100 150 200 250

calories

Figure 3.6: Density estimates of calorie contents of hot dogs

184 CHAPTER 3. REGRESSION

Model 3.3 Model 3.4 Interpretation Approximate value

mean calorie content of
µB µ 150
Beef hot dogs
mean calorie content of
µM µ + δM 160
Meat hot dogs
mean calorie content of
µP µ + δP 120
Poultry hot dogs
mean calorie difference be-
µM − µ B δM tween Beef and Meat hot- 10
dogs
mean calorie difference be-
µP − µ B δP tween Beef and Poultry -30
hotdogs
SD of calorie content
σ σ within a single type of hot 30
dog

Table 3.1: Correspondence between Models 3.3 and 3.4

The PlantGrowth data set in R provides another example. As R explains,

the data previously appeared in Dobson [1983] and are
“Results from an experiment to compare yields (as measured by
dried weight of plants) obtained under a control and two different
treatment conditions.”
The first several lines are
weight group
1 4.17 ctrl
2 5.58 ctrl
3 5.18 ctrl
Figure 3.7 shows the whole data set.
It appears that plants grown under different treatments tend to have different
weights. In particular, plants grown under Treatment 1 appear to be smaller
on average than plants grown under either the Control or Treatment 2. What
statistical model should we adopt?

Figure 3.7 was produced with the following snippet.

stripchart ( PlantGrowth$weight ~ PlantGrowth$group, pch=1,
xlab="weight" )

First, we think of the 10 plants grown under each condition as a sample

from a much larger population of plants that could have been grown. Second,
3.2. NORMAL LINEAR MODELS 185

trt2
trt1
ctrl

3.5 4.0 4.5 5.0 5.5 6.0

weight

Figure 3.7: The PlantGrowth data

a look at the data suggests that the weights in each group are clustered around
a central value, approximately symmetrically without an especially long tail in
either direction. So we model the weights as having Normal distributions.
But we should allow for the possibility that the three populations have dif-
ferent means. (We do not address the possibility of different SD’s here.) Let µ
be the population mean of plants grown under the Control condition, δ1 and δ2
be the extra weight due to Treatment 1 and Treatment 2 respectively, and σ be
the SD. We adopt the model

WC,1 , . . . , WC,10 ∼ i.i.d. N(µ, σ)

WT1 ,1 , . . . , WT1 ,10 ∼ i.i.d. N(µ + δ1 , σ) (3.5)
WT2 ,1 , . . . , WT2 ,10 ∼ i.i.d. N(µ + δ2 , σ).

There is a mathematical structure shared by 3.4, 3.5 and many other sta-
tistical models, and some common statistical notation to describe it. We’ll use
the hot dog data to illustrate.

Example 3.4 (Hot Dogs, continued)

Example 3.4 continues Example 2.2. First, there is the main variable of interest,
often called the response variable and denoted Y . For the hot dog data Y is calorie
content. (Another analysis could be made in which Y is sodium content.)
The distribution of Y is different under different circumstances. In this example,
Y has a Normal distribution whose mean depends on the type of hot dog. In general,
the distribution of Y will depend on some quantity of interest, called a covariate,
186 CHAPTER 3. REGRESSION

regressor, or explanatory variable. Covariates are often called X.

The data consists of multiple data points, or cases. We write Y i and Xi for the
i’th case. It is usual to represent the data as a matrix with one row for each case.
One column is for Y ; the other columns are for explanatory variables. For the hot
dog data the matrix is
Type Calories Sodium
Beef 186 495
Beef 181 477
...
Meat 140 428
Meat 138 339
Poultry 129 430
Poultry 132 375
...

(For analysis of calories, the third column is irrelevant.)

Rewriting the data matrix in a slightly different form reveals some mathematical
structure common to many models. There are 54 cases in the hotdog study. Let
(Y1 , . . . , Y54 ) be their calorie contents. For each i from 1 to 54, define two new
variables X1,i and X2,i by
(
1 if the i’th hot dog is Meat,
X1,i =
0 otherwise

and
(
1 if the i’th hot dog is Poultry,
X2,i =
0 otherwise.

X1,i and X2,i are indicator variables. Two indicator variables suffice because, for
the i’th hot dog, if we know X1,i and X2,i , then we know what type it is. (More
generally, if there are k populations, then k − 1 indicator variables suffice.) With
these new variables, Model 3.4 can be rewritten as
Yi = µ + δM X1,i + δP X2,i + i (3.6)
for i = 1, . . . , 54, where
1 , . . . , 54 ∼ i.i.d. N(0, σ).
Equation 3.6 is actually 54 separate equations, one for each case. We can write
them succinctly using vector and matrix notation. Let
Y = (Y1 , . . . , Y54 )t ,
B = (µ, δM , δP )t ,
E = (1 , . . . , 54 )t ,
3.2. NORMAL LINEAR MODELS 187

(The transpose is there because, by convention, vectors are column vectors.) and
 
1 0 0
1 0 0
 
 .. .. .. 
. . .
 
1 0 0
 
1 1 0
 
X = . . .
 .. .. .. 
 
1 1 0
 
1 0 1
 
. . .
 .. .. .. 
1 0 1
X is a 54 × 3 matrix. The first 20 lines are for the Beef hot dogs; the next 17 are
for the Meat hot dogs; and the final 17 are for the Poultry hot dogs. Equation 3.6
can be written
Y = XB + E (3.7)

Equations similar to 3.6 and 3.7 are common to many statistical models.
For the PlantGrowth data (page 184) let
Yi = weight of i’th plant,
(
1 if i’th plant received treatment 1
X1,i =
0 otherwise
(
1 if i’th plant received treatment 2
X2,i =
0 otherwise
Y = (Y1 , . . . , Y30 )t
B = (µ, δ1 , δ2 )t
E = (1 , . . . , 30 )t

and
 
1 0 0
1 0 0
 
 .. .. .. 
. . .
 
1 0 0
 
1 1 0
 
X = . .. .. 
 .. . .
 
1 1 0
 
1 0 1
 
. .. .. 
 .. . .
1 0 1
188 CHAPTER 3. REGRESSION

Then analogously to 3.6 and 3.7 we can write

Yi = µ + δ1 X1,i + δ2 X2,i + i (3.8)

and

Y = XB + E. (3.9)

Notice that Equation 3.6 is nearly identical to Equation 3.8 and Equation 3.7
is identical to Equation 3.9. Their structure is common to many statistical
models. Each Yi is written as the sum of two parts. The first part, XB,
(µ + δM X1,i + δP X2,i for the hot dogs; µ + δ1 X1,i + δ2 X2,i for PlantGrowth) is
called systematic, deterministic, or signal and represents the explainable differ-
ences between populations. The second part, E, or i , is random, or noise, and
represents the differences between hot dogs or plants within a single population.
The i ’s are called errors. In statistics, the word “error” does not indicate a
mistake; it simply means the noise part of a model, or the part left unexplained
by covariates. Modelling a response variable as

response = signal + noise

is a useful way to think and will recur throughout this book.

In 3.6 the signal µ+δM X1,i +δP X2,i is a linear function of (µ, δM , δP ). In 3.8
the signal µ + δ1 X1,i + δ2 X2,i is a linear function of (µ, δ1 , δ2 ). Models in which
the signal is a linear function of the parameters are called linear models.
In our examples so far, X has been an indicator. For each of a finite number
of X’s there has been a corresponding population of Y ’s. As the next example
illustrates, linear models can also arise when X is a continuous variable.

Example 3.5 (Ice Cream Consumption)

This example comes from DASL, which says
“Ice cream consumption was measured over 30 four-week periods
from March 18, 1951 to July 11, 1953. The purpose of the study was
to determine if ice cream consumption depends on the variables price,
income, or temperature. The variables Lag-temp and Year have been
added to the original data.”
You can download the data from
http://lib.stat.cmu.edu/DASL/Datafiles/IceCream.html.
The first few lines look like this:
date IC price income temp Lag-temp Year
1 .386 .270 78 41 56 0
2 .374 .282 79 56 63 0
3 .393 .277 81 63 68 0
3.2. NORMAL LINEAR MODELS 189

The variables are

date Time period (1-30) of the study (from 3/18/51 to 7/11/53)

IC Ice cream consumption in pints per capita

Price Price of ice cream per pint in dollars

Income Weekly family income in dollars

Temp Mean temperature in degrees F

Lag-temp Temp variable lagged by one time period

Year Year within the study (0 = 1951, 1 = 1952, 2 = 1953)

Figure 3.8 is a plot of consumption versus temperature. It looks as though an

equation of the form

consumption = β0 + β1 temperature + error (3.10)

would describe the data reasonably well. This is a linear model, not because con-
sumption is a linear function of temperature, but because it is a linear function of
(β0 , β1 ). To write it in matrix form, let

Y = (IC1 , . . . , IC30 )t
B = (β0 , β1 )t
E = (1 , . . . , 30 )t

and
 
1 temp1
1 temp2 
 
X = . .. 
 .. . 
1 temp30

The model is
Y = XB + E. (3.11)

Equation 3.11 is a linear model, identical to Equations 3.7 and 3.9.

Equation 3.7 (equivalently, 3.9 or 3.11) is the basic form of all linear models.
Linear models are extremely useful because they can be applied to so many
kinds of data sets. Section 3.2.2 investigates some of their theoretical properties
and R’s functions for fitting them to data.
190 CHAPTER 3. REGRESSION

0.55
0.50
0.45
consumption

0.40
0.35
0.30
0.25

30 40 50 60 70

temperature

Figure 3.8: Ice cream consumption (pints per capita) versus mean temperature
(◦ F)
3.2. NORMAL LINEAR MODELS 191

3.2.2 Inference for Linear Models

Section 3.2.1 showed some graphical displays of data that were eventually de-
scribed by linear models. Section 3.2.2 treats more formal inference for linear
models. We begin by deriving the likelihood function.
Linear models are described by Equation 3.7 (equivalently, 3.9 or 3.11) which
we repeat here for convenience:

Y = XB + E. (3.12)

In general there is an arbitrary number of cases, say n, and an arbitrary number

of covariates, say p. Equation 3.12 is shorthand for the collection of univariate
equations

Yi = β0 + β1 X1,i + · · · + βp Xp,i + i (3.13)

or equivalently,

Yi ∼ N(µi , σ)

P
for i = 1, . . . , n where µi = β0 + j βj Xj,i and the i ’s are i.i.d. N(0, σ). There
are p + 2 parameters: (β0 , . . . , βp , σ). The likelihood function is

n
Y
`(β0 , . . . , βp , σ) = p(yi | β0 , . . . , βp , σ)
i=1
n
Y 1 1 yi −µi 2
= √ e− 2 ( σ )
i=1
2πσ
n yi −(β0 + βj Xj,i ) 2
P
1
“ ”
Y −1
= √ e 2 σ

i=1
2πσ

− n 1
P P 2
= 2πσ 2 2 e− 2σ2 i (yi −(β0 + βj Xj,i )) (3.14)

Likelihood 3.14 is a function of the p + 2 parameters. To find the m.l.e.’s we

could differentiate 3.14 with respect to each parameter in turn, set the deriva-
tives equal to 0, and solve. But it is easier to take the log of 3.14 first, then
differentiate and solve.
 2
1 X X
log `(β0 , . . . , βp , σ) = C − n log σ − 2 yi − (β0 + βj Xi,j )
2σ i j
192 CHAPTER 3. REGRESSION

for some irrelevant constant C, so we get the system of equations

1 X X

2
yi − (β0 + βj Xi,j ) = 0
σ i j
1 X X

2
yi − (β0 + βj Xi,j ) Xi,1 = 0
σ i j
.. (3.15)
.
1 X X

y i − (β 0 + βj Xi,j ) Xi,p = 0
σ2 i j
n 1 X X
2
− + 3 yi − (β0 + βj Xi,j ) = 0
σ σ i j

The first p + 1 of these can be multiplied by σ 2 , yielding p + 1 linear equations

in the p + 1 unknown β’s. Because they’re linear, they can be solved by linear
algebra. The solution is
B̂ = (Xt X)−1 Xt Y,
using the notation of Equation 3.12.
For each i ∈ {1, . . . , n}, let

ŷi = β̂0 + x1i β̂1 + · · · + xpi β̂p .

The ŷi ’s are called fitted values. The residuals are

ri = yi − ŷi
= yi − β̂0 + x1i β̂1 + · · · + xpi β̂p

and are estimates of the errors i . Finally, referring to the last line of Equa-
tion 3.15, the m.l.e. σ̂ is found from
n 1 X X
2
0=− + 3 yi − (β0 + βj Xi,j )
σ σ i j
n 1 X 2
=− + 3 r
σ σ i i

so
1X 2
σ̂ 2 = ri
n

and
P  12
ri2
σ̂ = (3.16)
n
3.2. NORMAL LINEAR MODELS 193

In addition to the m.l.e.’s we often want to look at the likelihood function

to judge, for example, how accurately each β can be estimated. The likelihood
function for a single βi comes from the Central Limit Theorem. We will not
work out the math here but, fortunately, R will do all the calculations for us.
We illustrate with the hot dog data.

Example 3.6 (Hot Dogs, continued)

Estimating the parameters of a model is called fitting a model to data. R has built-in
commands for fitting models. The following snippet fits Model 3.7 to the hot dog
data. The syntax is similar for many model fitting commands in R, so it is worth
spending some time to understand it.
hotdogs.fit <- lm ( hotdogs$Calories ~ hotdogs$Type )

• lm stands for linear model.

• ~ stands for “is a function of”. It is used in many of R’s modelling commands.
y ~ x is called a formula and means that y is modelled as a function of x.
In the case at hand, Calories is modelled as a function of Type.
• lm specifies the type of model.
• R automatically creates the X matrix in Equation 3.12 and estimates the
parameters.
• The result of fitting the model is stored in a new object called hotdogs.fit.
Of course we could have called it anything we like.
• lm can have an argument data, which specifies a dataframe. So instead of

hotdogs.fit <- lm ( hotdogs$Calories ~ hotdogs$Type )

we could have written

hotdogs.fit <- lm ( Calories ~ Type, data=hotdogs )

You may want to try this to see how it works.

To see hotdogs.fit, use R’s summary function. It’s use and the resulting output
are shown in the following snippet.
> summary(hotdogs.fit)

Call:
lm(formula = hotdogs$Calories ~ hotdogs$Type)
194 CHAPTER 3. REGRESSION

Residuals:
Min 1Q Median 3Q Max
-51.706 -18.492 -5.278 22.500 36.294

Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 156.850 5.246 29.901 < 2e-16 ***
hotdogs$TypeMeat 1.856 7.739 0.240 0.811
hotdogs$TypePoultry -38.085 7.739 -4.921 9.4e-06 ***
---
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1

Residual standard error: 23.46 on 51 degrees of freedom

Multiple R-Squared: 0.3866, Adjusted R-squared: 0.3626
F-statistic: 16.07 on 2 and 51 DF, p-value: 3.862e-06

The most important part of the output is the table labelled Coefficients:. There
is one row of the table for each coefficient. Their names are on the left. In this table
the names are Intercept, hotdogs$TypeMeat, and hotdogs$TypePoultry. The
first column is labelled Estimate. Those are the m.l.e.’s. R has fit the model

Yi = β0 + β1 X1,i + β2 X2,i + i

where X1 and X2 are indicator variables for the type of hot dog. The model implies

Yi = β 0 +  i for beef hotdogs

Yi = β 0 + β 1 +  i for meat hotdogs
Yi = β 0 + β 2 +  i for poultry hotdogs

Therefore the names mean

β0 = Intercept = mean calorie content of beef hot dogs

β1 = hotdogs$TypeMeat = mean difference between beef and meat hot dogs
β2 = hotdogs$TypePoultry = mean difference between
beef and poultry hot dogs

From the Coefficients table the estimates are

β̂0 = 156.850
β̂1 = 1.856
β̂2 = −38.085

The next column of the table is labelled Std. Error. It contains the SD’s of
the estimates. In this case, β̂0 has an SD of about 5.2; β̂1 has an SD of about
3.2. NORMAL LINEAR MODELS 195

7.7, and β̂2 also has an SD of about 7.7. The Central Limit Theorem says that
approximately, in large samples

β̂0 ∼ N(β0 , σβ0 )

β̂1 ∼ N(β1 , σβ1 )
β̂2 ∼ N(β2 , σβ2 )

The SD’s in the table are estimates of the SD’s in the Central Limit Theorem.
Figure 3.9 plots the likelihood functions. The interpretation is that β 0 is likely
somewhere around 157, plus or minus about 10 or so; β 1 is somewhere around 2,
plus or minus about 15 or so; and β2 is somewhere around -38, plus or minus about
15 or so. (Compare to Table 3.3.) In particular, there is no strong evidence that
Meat hot dogs have, on average, more or fewer calories than Beef hot dogs; but
there is quite strong evidence that Poultry hot dogs have considerably fewer.

Figure 3.9 was produced with the following snippet.

m <- c ( 156.85, 1.856, -38.085 )
s <- c ( 5.246, 7.739, 7.739 )

par ( mfrow=c(2,2) )

x <- seq ( m[1]-3*s[1], m[1]+3*s[1], length=40 )

plot ( x, dnorm(x,m[1],s[1]), type="l",
xlab=expression(mu), ylab="likelihood", yaxt="n" )

x <- seq ( m[2]-3*s[2], m[2]+3*s[2], length=40 )

plot ( x, dnorm(x,m[2],s[2]), type="l",
xlab=expression(delta[M]),
ylab="likelihood", yaxt="n" )

x <- seq ( m[3]-3*s[3], m[3]+3*s[3], length=40 )

plot ( x, dnorm(x,m[3],s[3]), type="l",
xlab=expression(delta[P]),
ylab="likelihood", yaxt="n" )

The summary also gives an estimate of σ. The estimate is labelled Residual

standard error. In this case, σ̂ ≈ 23.46.1 So our model says that for each type
of hot dog, the calorie contents have approximately a Normal distribution with SD
about 23 or so. Compare to Figure 3.5 to see whether the 23.46 makes sense.
1
P R, like most statistical software, does not report the m.l.e. but reports instead σ̂ ≡
( ri2 /(n − p − 1))1/2 . Compare to Equation 3.16 for the m.l.e. in which the denomina-
tor is n. The situation is similar to the sample SD on page 83. When n  p there is little
difference between the two estimates.
196 CHAPTER 3. REGRESSION

likelihood

140 150 160 170 likelihood −20 0 10

µ δM
likelihood

−60 −40 −20

δP

Figure 3.9: Likelihood functions for (µ, δM , δP ) in the Hot Dog example.
3.2. NORMAL LINEAR MODELS 197

Regression is sometimes used in an exploratory setting, when scientists want

to find out which variables are related to which other variables. Often there is
a response variable Y , (imagine, for example, performance in school) and they
want to know which other variables affect Y (imagine, for example, poverty,
amount of television watching, computer in the home, parental involvment,
etc.) Example 3.7 illustrates the process.

Example 3.7 (mtcars)

This example uses linear regression to explore the R data set mtcars (See Figure 3.1,
panel (c)) more thoroughly with the goal of modelling mpg (miles per gallon) as a
function of the other variables. As usual, type data(mtcars) to load the data into
R and help(mtcars) for an explanation. As R explains:

“The data was extracted from the 1974 Motor Trend US magazine,
and comprises fuel consumption and 10 aspects of automobile design
and performance for 32 automobiles (1973-74 models).”

In an exploratory exercise such as this, it often helps to begin by looking at the

data. Accordingly, Figure 3.10 is a pairs plot of the data, using just the continuous
variables.

Figure 3.10 was produced by

pairs ( mtcars[,c(1,3:7)] )

Clearly, mpg is related to several of the other variables. Weight is an obvious

and intuitive example. The figure suggests that the linear model

mpg = β0 + β1 wt +  (3.17)

is a good start to modelling the data. Figure 3.11(a) is a plot of mpg vs. weight plus
the fitted line. The estimated coefficients turn out to be β̂0 ≈ 37.3 and β̂1 ≈ −5.34.
The interpretation is that mpg decreases by about 5.34 for every 1000 pounds of
weight. Note: this does not mean that if you put a 1000 pound weight in your car
your mileage mpg will decrease by 5.34. It means that if car A weighs about 1000
pounds less than car B, then we expect car A to get an extra 5.34 miles per gallon.
But there are likely many differences between A and B besides weight. The 5.34
accounts for all of those differences, on average.
We could just as easily have begun by fitting mpg as a function of horsepower
with the model
mpg = γ0 + γ1 hp +  (3.18)
We use γ’s to distinguish the coefficients in Equation 3.18 from those in Equa-
tion 3.17. The m.l.e.’s turn out to be γ̂0 ≈ 30.1 and γ̂1 ≈ −0.069. Figure 3.11(b)
shows the corresponding scatterplot and fitted line. Which model do we prefer?
Choosing among different possible models is a major area of statistical practice
198 CHAPTER 3. REGRESSION

100 400 3.0 4.5 16 22

25
mpg

10
400

disp
100

250
hp

50
4.5

drat
3.0

4
wt
2
22

qsec
16

10 25 50 250 2 4

Figure 3.10: pairs plot of the mtcars data. Type help(mtcars) in R for an
explanation.
3.2. NORMAL LINEAR MODELS 199

with a large literature that can be highly technical. In this book we show just a few
considerations.
One way to judge models is through residual plots, which are plots of residuals
versus either X variables or fitted values. If models are adequate, then residual
plots should show no obvious patterns. Patterns in residual plots are clues to model
inadequacy and how to improve models. Figure 3.11(c) and (d) are residuals plots
for mpg.fit1 (mpg vs. wt) and mpg.fit2 (mpg vs. hp). There are no obvious
patterns in panel (c). In panel (d) there is a suggestion of curvature. For fitted
values between about 15 and 23, residuals tend to be low but for fitted values less
than about 15 or greater than about 23, residuals tend to be high (The same pattern
might have been noted in panel (b).) suggesting that mpg might be better fit as a
nonlinear function of hp. We do not pursue that suggestion further at the moment,
merely noting that there may be a minor flaw in mpg.fit2 and we therefore slightly
prefer mpg.fit1.
Another thing to note from panels (c) and (d) is the overall size of the residuals.
In (c), they run from about -6 to about +4, while in (d) they run from about
-6 to about +6. That is, the residuals from mpg.fit2 tend to be slightly larger
in absolute value than the residuals from mpg.fit1, suggesting that wt predicts
mpg slightly better than does hp. That impression can be confirmed by getting
the summary of both fits and checking σ̂. From mpg.fit1 σ̂ ≈ 3.046 while from
mpg.fit2 σ̂ ≈ 3.863. I.e., from wt we can predict mpg to within about 6 or so
(two SD’s) while from hp we can predict mpg only to within about 7.7 or so. For
this reason too, we slightly prefer mpg.fit1 to mpg.fit2.
What about the possibility of using both weight and horsepower to predict mpg?
Consider

mpg.fit3 <- lm ( mpg ~ wt + hp, data=mtcars )

• The formula y ~ x1 + x2 means fit y as a function of both x1 and x2. In

our example that means

mpg = δ0 + δ1 wt1 + δ2 hp2 +  (3.19)

A residual plot from model 3.19 is shown in Figure 3.11 (e). The m.l.e.’s are
δ̂0 ≈ 37.2, δ̂1 ≈ −3.88, δ̂2 ≈ −0.03, and σ̂ ≈ 2.6. Since the residual plot looks
curved, Model 3.17 has residuals about as small as Model 3.19, and Model 3.17 is
more parsimonious than Model 3.19 we slightly prefer Model 3.17.

Figure 3.11 (a) was produced with

plot ( mtcars$wt, mtcars$mpg, xlab="weight", ylab="mpg" )
mpg.fit1 <- lm ( mpg ~ wt, data=mtcars )
abline ( coef(mpg.fit1) )
200 CHAPTER 3. REGRESSION

(a) (b)

30

30
mpg

mpg
20

20
10

10
2 3 4 5 50 150 250

weight horsepower

(c) (d)
0 2 4 6

6
resid

resid

2
−2
−4

−6

10 15 20 25 30 10 15 20 25

fitted values from fit1 fitted values from fit2

(e)
6
4
2
resid

0
−4

10 15 20 25 30

fitted values from fit3

Figure 3.11: mtcars — (a): mpg vs. wt; (b): mpg vs. hp; (c): residual plot
from mpg~ wt; (d): residual plot from mpg~ hp; (e): residual plot from mpg~
wt+hp
3.2. NORMAL LINEAR MODELS 201

Figure 3.11, panels (c) and (d) were produced with

# panel c
plot ( fitted(mpg.fit1), resid(mpg.fit1), main="(c)",
xlab="fitted values from fit1", ylab="resid" )
# panel d
plot ( fitted(mpg.fit2), resid(mpg.fit2),
xlab="fitted values from fit2", ylab="resid",
main="(d)" )

In Example 3.7 we fit three models for mpg, repeated here with their original
equation numbers.

mpg = β0 + β1 wt +  (3.17)
mpg = γ0 + γ1 hp +  (3.18)
mpg = δ0 + δ1 wt1 + δ2 hp +  (3.19)

What is the connection between, say, β1 and δ1 , or between γ1 and δ2 ? β1 is the

average mpg difference between two cars whose weights differ by 1000 pounds.
Since heavier cars tend to be different than lighter cars in many ways, not just in
weight, β1 captures the net effect on mpg of all those differences. On the other
hand, δ1 is the average mpg difference between two cars of identical horsepower
but whose weights differ by 1000 pounds. Figure 3.12 shows the likelihood
functions of these four parameters. The evidence suggests that β1 is probably
in the range of about -7 to about -4, while δ1 is in the range of about -6 to -2.
It’s possible that β1 ≈ δ1 . On the other hand, γ1 is probably in the interval
(−.1, −.04) while δ2 is probably in the interval (−.05, 0). It’s quite likely that
γ1 6≈ δ2 . Scientists sometimes ask the question “What is the effect of variable
X on variable Y ?” That question does not have an unambiguous answer; the
answer depends on which other variables are accounted for and which are not.

Figure 3.12 was produced with

par ( mfrow=c(2,2) )
x <- seq ( -8, -1.5, len=60 )
plot ( x, dnorm(x,-5.3445,.5591), type="l",
xlab=expression(beta[1]), ylab="", yaxt="n" )
x <- seq ( -.1, 0, len=60 )
plot ( x, dnorm(x,-.06823,.01012), type="l",
xlab=expression(gamma[1]), ylab="", yaxt="n" )
x <- seq ( -8, -1.5, len=60 )
plot ( x, dnorm(x,-3.87783,.63273), type="l",
xlab=expression(delta[1]), ylab="", yaxt="n" )
202 CHAPTER 3. REGRESSION

−8 −6 −4 −2 −0.10 −0.04 0.00

β1 γ1

−8 −6 −4 −2 −0.10 −0.04 0.00

δ1 δ2

Figure 3.12: likelihood functions for β1 , γ1 , δ1 and δ2 in the mtcars example.

3.3. GENERALIZED LINEAR MODELS 203

x <- seq ( -.1, 0, len=60 )

plot ( x, dnorm(x,-.03177,.00903), type="l",
xlab=expression(delta[2]), ylab="", yaxt="n" )

3.3 Generalized Linear Models

3.3.1 Logistic Regression
Look again at panel (c) in Figure 3.1 on page 174. The dependent variable
is binary, as opposed to the continuous dependent variables in panels (a), (b)
and (d). In (a), (b) and (d) we modelled Y | X as having a Normal distribution;
regression was a model for E[Y | X], the mean of that Normal distribution as a
function of X. In (c) Y | X has a Binomial distribution. We still use the term
“regression” for a model of E[Y | X]. When Y is binary, regression is a model
for the probability of success θ as a function of X.
Figure 3.13 shows two more scatterplots where Y is a binary variable. The
data are described in the next two examples.

Example 3.8 (FACE, continued)

Refer to Examples 1.12 and 2.10 about the FACE experiment to assess the effects
of excess CO2 on the growth of the forest. To describe the size of trees, ecologists
sometimes use Diameter at Breast Height, or DBH. DBH was recorded every year
for each loblolly pine tree in the FACE experiment. One potential effect of elevated
CO2 is for the trees to reach sexual maturity and hence be able to reproduce earlier
than otherwise. If they do mature earlier, ecologists would like to know whether
that’s due only to their increased size, or whether trees will reach maturity not just
at younger ages, but also at smaller sizes. Sexually mature trees can produce pine
cones but immature trees cannot. So to investigate sexual maturity, a graduate
student counted the number of pine cones on each tree. For each tree let X be its
DBH and Y be either 1 or 0 according to whether the tree has pine cones.
Figure 3.13(a) is a plot of Y versus X for all the trees in Ring 1. It does appear
that larger trees are more likely to have pine cones.

Example 3.9 (O-rings)

“On January 28, 1986 America was shocked by the destruction of the space shuttle
Challenger, and the death of its seven crew members.” So begins the website
http://www.fas.org/spp/51L.html of the Federation of American Scientist’s
Space Policy Project.
Up until 1986 the space shuttle orbiter was lifted into space by a pair of booster
rockets, one on each side of the shuttle, that were comprised of four sections
stacked vertically on top of each other. The joints between the sections were sealed
by O-rings. On January 28, 1986 the temperature at launch time was so cold that
the O-rings became brittle and failed to seal the joints, allowing hot exhaust gas
204 CHAPTER 3. REGRESSION

to come into contact with unburned fuel. The result was the Challenger disaster.
An investigation ensued. The website http://science.ksc.nasa.gov/shuttle/
missions/51-l/docs contains links to
1. a description of the event,
2. a report (Kerwin) on the initial attempt to determine the cause,
3. a report (rogers-commission) of the presidential investigative commission that
finally did determine the cause, and
4. a transcript of the operational recorder voice tape.
One of the issues was whether NASA could or should have forseen that cold weather
might diminish performance of the O-rings.
After launch the booster rockets detach from the orbiter and fall into the ocean
where they are recovered by NASA, taken apart and analyzed. As part of the
analysis NASA records whether any of the O-rings were damaged by contact with
hot exhaust gas. If the probability of damage is greater in cold weather then, in
principle, NASA might have forseen the possibility of the accident which occurred
during a launch much colder than any previous launch.
Figure 3.13(b) plots Y = presence of damage against X = temperature for
the launches prior to the Challenger accident. The figure does suggest that colder
launches are more likely to have damaged O-rings. What is wanted is a model for
probability of damage as a function of temperature, and a prediction for probability
of damage at 37◦ F, the temperature of the Challenger launch.

Fitting straight lines to Figure 3.13 doesn’t make sense. In panel (a) what
we need is a curve such that
1. E[Y | X] = P[Y = 1 | X] is close to 0 when X is smaller than about 10 or
12 cm., and
2. E[Y | X] = P[Y = 1 | X] is close to 1 when X is larger than about 25 or
30 cm.
In panel (b) we need a curve that goes in the opposite direction.
The most commonly adopted model in such situations is

eβ0 +β1 x
E[Y | X] = P[Y = 1 | X] = (3.20)
1 + eβ0 +β1 x
Figure 3.14 shows the same data as Figure 3.13 with some curves added accord-
ing to Equation 3.20. The values of β0 and β1 are in Table 3.3.1.

Figure 3.14 was produced by the following snippet.

par ( mfrow=c(2,1) )
plot ( cones$dbh[ring1], mature[ring1], xlab="DBH",
ylab="pine cones present", main="(a)" )
3.3. GENERALIZED LINEAR MODELS 205

(a)
pine cones present

0.8
0.4
0.0

5 10 15 20 25

DBH

(b)
damage present

0.8
0.4
0.0

55 60 65 70 75 80

temperature

Figure 3.13: (a): pine cone presence/absence vs. dbh. (b): O-ring damage vs.
launch temperature
206 CHAPTER 3. REGRESSION

(a)
pine cones present

0.8
0.4
0.0

5 10 15 20 25

DBH

(b)
damage present

0.8
0.4
0.0

55 60 65 70 75 80

temperature

Figure 3.14: (a): pine cone presence/absence vs. dbh. (b): O-ring damage vs.
launch temperature, with some logistic regression curves
3.3. GENERALIZED LINEAR MODELS 207

β0 β1
solid -8 .45
(a) dashed -7.5 .36
dotted -5 .45
solid 20 -.3
(b) dashed 15 -.23
dotted 18 -.3

Table 3.2: β’s for Figure 3.14

x <- seq ( 4, 25, length=40 )

b0 <- c(-8, -7.5, -5)
b1 <- c ( .45, .36, .45 )
for ( i in 1:3 )
lines ( x, exp(b0[i] + b1[i]*x)/(1 + exp(b0[i] + b1[i]*x) ),
lty=i )

plot ( orings$temp, orings$damage>0, xlab="temperature",

ylab="damage present", main="(b)" )
x <- seq ( 50, 82, length=40 )
b0 <- c(20, 15, 18)
b1 <- c ( -.3, -.23, -.3 )
for ( i in 1:3 )
lines ( x, exp(b0[i] + b1[i]*x)/(1 + exp(b0[i] + b1[i]*x) ),
lty=i )

Model 3.20 is known as logistic regression. Let the i’th observation have
covariate xi and probability of success θi = E[Yi | xi ]. Define
 
θi
φi ≡ log .
1 − θi

φi is called the logit of θi . The inverse transformation is

e φi
θi = .
1 + e φi
The logistic regression model is

φ i = β 0 + β 1 xi .

This is called a generalized linear model or glm because it is a linear model for
φ, a transformation of E(Y | x) rather than for E(Y | x) directly. The quantity
β0 + β1 x is called the linear predictor. If β1 > 0, then as x → +∞, θ → 1 and
208 CHAPTER 3. REGRESSION

as x → −∞, θ → 0. If β1 < 0 the situation is reversed. β0 is like an intercept;

it controls how far to the left or right the curve is. β1 is like a slope; it controls
how quickly the curve moves between its two asymptotes.
Logistic regression and all generalized linear models differ from linear re-
gression in two ways: the regression function is nonlinear and the distribution
of Y | x is not Normal. These differences imply that the methods we used to
analyze linear models are not correct for generalized linear models. We need to
derive the likelihood function and find new calculational algorithms.
The likelihood function is derived from first principles.
Y
p(Y1 , . . . , Yn | x1 , . . . , xn , β0 , β1 ) = p(Yi | xi , β0 , β1 )
i
Y
= θixi (1 − θi )1−xi
i
Y Y
= θi (1 − θi )
i:xi =1 i:xi =0
Y eβ0 +β1 xi Y 1
= β +β x β0 +β1 xi
i:xi =1
1 + e 0 1 i
i:x =0
1 + e
i

This is a rather complicated function of the two variables (β0 , β1 ). However, a

Central Limit Theorem applies to give a likelihood function for β0 and β1 that
is accurate when n is reasonable large. The theory is beyond the scope of this
book, but R will do the calculations for us. We illustrate with the pine cone
data from Example 3.8. Figure 3.15 shows the likelihood function.

Figure 3.15 was produced by the following snippet.

mature <- cones$X2000[ring1] > 0
b0 <- seq ( -11, -4, length=60 )
b1 <- seq ( .15, .5, length=60 )
lik <- matrix ( NA, 60, 60 )
for ( i in 1:60 )
for ( j in 1:60 ) {
linpred <- b0[i] + b1[j]*cones$dbh[ring1]
theta <- exp(linpred) / (1+exp(linpred))
lik[i,j] <- prod ( theta^mature * (1-theta)^(1-mature) )
}
lik <- lik/max(lik)
contour ( b0, b1, lik, xlab=expression(beta[0]),
ylab=expression(beta[1]) )

• mature is an indicator variable for whether a tree has at least one pine
cone.
3.3. GENERALIZED LINEAR MODELS 209

0.50
0.45
0.40
0.35
β1

0.30
0.25
0.20
0.15

−11 −10 −9 −8 −7 −6 −5 −4

β0

Figure 3.15: Likelihood function for the pine cone data

210 CHAPTER 3. REGRESSION

• The lines b0 <- ... and b1 <- ... set some values of (β0 , β1 ) at which
to evaluate the likelihood. They were chosen after looking at the output
from fitting the logistic regression model.
• lik <- ... creates a matrix to hold values of the likelihood function.
• linpred is the linear predictor. Because cones$dbh[ring1] is a vec-
tor, linpred is also a vector. Therefore theta is also a vector, as is
theta^mature * (1-theta)^(1-mature). It will help your understand-
ing of R to understand what these vectors are.

One notable feature of Figure 3.15 is the diagonal slope of the contour el-
lipses. The meaning is that we do not have independent information about β0
and β1 . For example if we thought, for some reason, that β0 ≈ −9, then we
could be fairly confident that β1 is in the neighborhood of about .4 to about
.45. But if we thought β0 ≈ −6, then we would believe that β1 is in the neigh-
borhood of about .25 to about .3. More generally, if we knew β0 , then we could
estimate β1 to within a range of about .05. But since we don’t know β0 , we
can only say that β1 is likely to be somewhere between about .2 and .6. The
dependent information for (β0 , β1 ) means that our marginal information for β1
is much less precise than our conditional information for β1 given β0 . That
imprecise marginal information is reflected in the output from R, shown in the
following snippet which fits the model and summarizes the result.

cones <- read.table ( "data/pinecones.dat", header=T )

ring1 <- cones$ring == 1
mature <- cones$X2000[ring1] > 0

fit <- glm ( mature ~ cones$dbh[ring1], family=binomial )

summary ( fit )
...
Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -7.46684 1.76004 -4.242 2.21e-05 ***
cones$dbh[ring1] 0.36151 0.09331 3.874 0.000107 ***
...

• cones ... reads in the data. There is one line for each tree. The first
few lines look like this.

ring ID xcoor ycoor spec dbh X1998 X1999 X2000

1 1 11003 0.71 0.53 pita 19.4 0 0 0
2 1 11004 1.26 2.36 pita 14.1 0 0 4
3 1 11011 1.44 6.16 pita 19.4 0 6 0
3.3. GENERALIZED LINEAR MODELS 211

ID is a unique identifying number for each tree; xcoor and ycoor are
coordinates in the plane; spec is the species; pita stands for pinus taeda
or loblolly pine, X1998, X1999 and X2000 are the numbers of pine cones
each year.

• ring1 ... is an indicator variable for trees in Ring 1.

• mature ... indicates whether the tree had any cones at all in 2000. It is
not a precise indicator of maturity.

• fit ... fits the logistic regression. glm fits a generalized linear model.
The argument family=binomial tells R what kind of data we have. In
this case it’s binomial because y is either a success or failure.

• summary(fit) shows that (β̂0 , β̂1 ) ≈ (−7.5, 0.36). The SD’s are about 1.8
and .1. These values guided the choice of b0 and b1 in creating Figure 3.15.
It’s the SD of about .1 that says we can estimate β1 to within an interval
of about .4, or about ±2SD’s.

3.3.2 Poisson Regression

Section 3.3.1 dealt with the case where the response variable Y was Bernoulli.
Another common situation is where the response Y is a count. In that case it
is natural to adopt, at least provisionally, a model in which Y has a Poisson
distribution: Y ∼ Poi(λ). When there are covariates X, then λ may depend on
X. It is common to adopt the regression

log λ = β0 + β1 x (3.21)

Model 3.21 is another example of a generalized linear model. Example 3.10

illustrates its use.

Example 3.10 (Seedlings, continued)

Several earlier examples have discussed data from the Coweeta LTER on the emer-
gence and survival of red maple (acer rubrum) seedlings. Example 3.2 showed
that the arrival rate of seedlings seemed to vary by quadrat. Refer especially to
Figure 3.4. Example 3.10 follows up that observation more quantitatively.
Roughly speaking, New seedlings arise in a two-step process. First, a seed falls
out of the sky, then it germinates and emerges from the ground. We may reasonably
assume that the emergence of one seedling does not affect the emergence of another
(They’re too small to interfere with each other.) and hence that the number of New
seedlings has a Poi(λ) distribution. Let Yij be the number of New seedlings observed
in quadrat i and year j. Here are two fits in R, one in which λ varies by quadrat
and one in which it doesn’t.

new <- data.frame ( count=count,

quadrat=as.factor(quadrat),
year=as.factor(year)
212 CHAPTER 3. REGRESSION

)
fit0 <- glm ( count ~ 1, family=poisson, data=new )
fit1 <- glm ( count ~ quadrat, family=poisson, data=new )

• The command data.frame creates a dataframe. R describes dataframes

as
“tightly coupled collections of variables which share many of
the properties of matrices and of lists, used as the fundamental
data structure by most of R’s modeling software.”
We created a dataframe called new, having three columns called count,
quadrat, and year. Each row of new contains a count (of New seedlings), a
quadrat number and a year. There are as many rows as there are observations.
• The command as.factor turns its argument into a factor. That is, instead
of treating quadrat and year as numerical variables, we treat them as indica-
tor variables. That’s because we don’t want a quadrat variable running from
1 to 60 implying that the 60th quadrat has 60 times as much of something
as the 1st quadrat. We want the quadrat numbers to act as labels, not as
numbers.
• glm stands for generalized linear model. The family=poisson argument says
what kind of data we’re modelling. data=new says the data are to be found
in a dataframe called new.
• The formula count ~ 1 says to fit a model with only an intercept, no covari-
ates.
• The formula count ~ quadrat says to fit a model in which quadrat is a
covariate. Of course that’s really 59 new covariates, indicator variables for 59
of the 60 quadrats.

To examine the two fits and see which we prefer, we plotted actual versus fitted
values and residuals versus fitted values in Figure 3.16. Panels (a) and (b) are from
fit0. Because there may be overplotting, we jittered the points and replotted them
in panels (c) and (d). Panels (e) and (f) are jittered values from fit1. Comparison
of panels (c) to (e) and (d) to (f) shows that fit1 predicts more accurately and
has smaller residuals than fit0. That’s consistent with our reading of Figure 3.4.
So we prefer fit1.
Figure 3.17 continues the story. Panel (a) shows residuals from fit1 plotted
against year. There is a clear difference between years. Years 1, 3, and 5 are high
while years 2 and 4 are low. So perhaps we should use year as a predictor. That’s
done by
fit2 <- glm ( count ~ quadrat+year, family=poisson,
data=new )
3.4. PREDICTIONS FROM REGRESSION 213

Panels (b) and (c) show diagnostic plots for fit2. Compare to similar panels in
Figure 3.16 to see whether using year makes an appreciable difference to the fit.

Figure 3.16 was created with the following snippet.

par ( mfrow=c(3,2) )
plot ( fitted(fit0), new$count, xlab="fitted values",
ylab="actual values", main="(a)" )
abline ( 0, 1 )
plot ( fitted(fit0), residuals(fit0), xlab="fitted values",
ylab="residuals", main="(b)" )
plot ( jitter(fitted(fit0)), jitter(new$count),
xlab="fitted values", ylab="actual values",
main="(c)" )
abline ( 0, 1 )
plot ( jitter(fitted(fit0)), jitter(residuals(fit0)),
xlab="fitted values", ylab="residuals", main="(d)" )
plot ( jitter(fitted(fit1)), jitter(new$count),
xlab="fitted values", ylab="actual values",
main="(e)" )
abline ( 0, 1 )
plot ( jitter(fitted(fit1)), jitter(residuals(fit1)),
xlab="fitted values", ylab="residuals", main="(f)" )

The following snippet shows how Figure 3.17 was made in R.

par ( mfrow=c(2,2) )
plot ( new$year, residuals(fit1),
xlab="year", ylab="residuals", main="(a)" )
plot ( jitter(fitted(fit2)), jitter(new$count),
xlab="fitted values", ylab="actual values",
main="(b)" )
abline ( 0, 1 )
plot ( jitter(fitted(fit2)), jitter(residuals(fit2)),
xlab="fitted values", ylab="residuals", main="(c)" )

3.4 Predictions from Regression

From a regression equation, if we have estimates of the β’s we can
1. plug in the values of the x’s we have to get fitted values, and
2. plug in the values of x for new or future cases to get predicted values.
214 CHAPTER 3. REGRESSION

(a) (b)

8
15
actual values

6
residuals
10

4
2
5

0
0

0.6 0.8 1.0 1.2 0.6 0.8 1.0 1.2

fitted values fitted values

(c) (d)

8
15
actual values

6
residuals
10

4
2
5

0
0

0.865 0.875 0.885 0.895 0.865 0.875 0.885 0.895

fitted values fitted values

(e) (f)
15

3
actual values

residuals
10

1
−1
5

−3
0

0 1 2 3 4 5 0 1 2 3 4 5

fitted values fitted values

Figure 3.16: Actual vs. fitted and residuals vs. fitted for the New seedling data.
(a) and (b): fit0. (c) and (d): jittered values from fit0. (e) and (f ): jittered
values from fit1.
3.4. PREDICTIONS FROM REGRESSION 215

4 (a) (b)

15
3
2

actual values

10
residuals

1
0
−1

5
−3

1 2 3 4 5 0 2 4 6 8

year fitted values

(c)
3
2
1
residuals

0
−1
−2
−3

0 2 4 6 8

fitted values

Figure 3.17: New seedling data. (a): residuals from fit1 vs. year. (b): actual
vs. fitted from fit2. (c): residuals vs. fitted from fit2.
216 CHAPTER 3. REGRESSION

We illustrate with the mtcars data.

Example 3.11 (mtcars, continued)

Example 3.7 concluded with a comparison of three models for mpg. Here we con-
tinue that comparison by seeing whether the models make substantially different
predictions for any of the cars in the data set. For each car we know its weight and
horsepower and we have estimates of all the parameters in Equations 3.17, 3.18,
and 3.19, so we can compute its fitted values from all three models. In symbols,
ŷi = β̂0 + β̂1 wti (from 3.17)
ŷi = γ̂0 + γ̂1 hpi (from 3.18)
ŷi = δ̂0 + δ̂1 wti + δ̂2 hpi (from 3.19)
We plot the fitted values against each other to see whether there are any noticable
differences. Figure 3.18 displays the result. Figure 3.18 shows that the mpg.fit1
and mpg.fit3 produce fitted values substantially similar to each other and agreeing
fairly well with actual values, while mpg.fit2 produces fitted values that differ
somewhat from the others and from the actual values, at least for a few cars. This
is another reason to prefer mpg.fit1 and mpg.fit3 to mpg.fit2. In Example 3.7
this lack of fit showed up as a higher σ̂ for mpg.fit2 than for mpg.fit1.

Figure 3.18 was made with the following snippet.

fitted.mpg <- cbind ( fitted(mpg.fit1), fitted(mpg.fit2),
fitted(mpg.fit3), mtcars$mpg )
pairs ( fitted.mpg, labels = c ( "fitted from wt",
"fitted from hp", "fitted from both", "actual mpg" ) )

• fitted(xyz) extracts fitted values. xyz can be any model previously fitted
by lm, glm, or other R functions to fit models.

In Example 3.5 we posited model 3.10:

y = β 0 + β1 x +  (3.22)
where x was mean temperature during the week and y was ice cream consump-
tion during the week. Now we want to fit the model to the data and use the fit
to predict consumption. In addition, we want to say how accurate the predic-
tions are. Let xf be the predicted mean temperature for some future week and
yf be consumption. xf is known; yf is not. Our model says
yf ∼ N(µf , σ)

where

µ f = β 0 + β 1 xf
3.4. PREDICTIONS FROM REGRESSION 217

10 20 10 20 30

30
25
20
fitted from wt

15
10
25
20

fitted from hp
15
10

30
25
20

fitted from both

15
10
30
25

actual mpg
20
15
10

10 20 30 10 20 30

Figure 3.18: Actual mpg and fitted values from three models
218 CHAPTER 3. REGRESSION

µf is unknown because β0 and β1 are unknown. But (β0 , β1 ) can be estimated

from the data, so we can form an estimate

µ̂f = β̂0 + β̂1 xf

How accurate is µ̂f as an estimate of µf ? The answer depends on how accurate

(β̂0 , β̂1 ) are as estimates of (β0 , β1 ). Advanced theory about Normal distribu-
tions, beyond the scope of this book, tells us

µ̂f ∼ N(µf , σfit )

for some σfit . σfit may depend on xf ; we have omitted the dependency from the
notation.
µf is the average ice cream consumption in all weeks whose mean temper-
ature is xf . So µ̂f is also an estimator of yf . But in any particular week the
actual consumption won’t exactly equal µf . Our model says

yf = µ f + 

where  ∼ N(0, σ). So in any given week yf will differ from µf by an amount
up to about ±2σ or so. The uncertainty in estimating yf has two components:
the uncertainty of µf and the variability due to .

3.5 Exercises
1. (a) Use the attenu , airquality and faithful datasets to reproduce
Figures 3.1 (a), (b) and (d).
(b) Add lowess and supsmu fits.
(c) Figure out how to use the tuning parameters and try out several
different values. (Use the help or help.start functions.)

2. With the mtcars dataset, use a scatterplot smoother to plot the relation-
ship between weight and displacement. Does it matter which we think of
as X and which as Y ? Is one way more natural than the other?

3. Download the 1970 draft data from DASL and reproduce Figure 3.3. Use
the tuning parameters (f for lowess; span for supsmu) to draw smoother
and wigglier scatterplot smoothers.

4. How could you test whether the draft numbers in Example 3.1 were gener-
ated uniformly? What would H0 be? What would be a good test statistic
w? How would estimate the distribution of w under H0 ?

5. Using the information in Example 3.6 estimate the mean calorie content
of meat and poultry hot dogs.

6. Refer to Examples 2.2, 3.4, and 3.6.

3.5. EXERCISES 219

(a) Formulate statistical hypotheses for testing whether the mean calorie
content of Poultry hot dogs is equal to the mean calorie content of
Beef hot dogs.
(b) What statistic will you use?
(c) What should that statistic be if H0 is true?
(d) How many SD’s is it off?
(e) What do you conclude?
(f) What about Meat hot dogs?

7. Refer to Examples 2.2, 3.4, and 3.6. Figure 3.5 shows plenty of overlap
in the calorie contents of Beef and Poultry hot dogs. I.e., there are many
Poultry hot dogs with more calories than many Beef hot dogs. But Fig-
ure 3.9 shows very little support for values of δP near 0. Can that be
right? Explain?

8. Examples 2.2, 3.4, and 3.6 analyze the calorie content of Beef, Meat, and
Poultry hot dogs. Create a similar analysis, but for sodium content. Your
analysis should cover at least the following steps.

(a) A stripchart similar to Figure 3.5 and density estimates similar to

Figure 3.6.
(b) A model similar to Model 3.4, including definitions of the parameters.
(c) Indicator variables analogous to those in Equation 3.6.
(d) A model similar to Model 3.7, including definitions of all the terms.
(e) A fit in R, similar to that in Example 3.6.
(f) Parameter estimates and SD’s.
(g) Plots of likelihood functions, analagous to Figure 3.9.
(h) Interpretation.

9. Analyze the PlantGrowth data from page 184. State your conclusion
about whether the treatments are effective. Support you conclusion with
analysis.

10. Analyze the Ice Cream data from Example 3.5. Write a model similar to
Model 3.7, including definitions of all the terms. Use R to fit the model.
Estimate the coefficients and say how accurate your estimates are. If
temperature increases by about 5 ◦ F, about how much would you expect
ice cream consumption to increase? Make a plot similar to Figure 3.8, but
add on the line implied by Equation 3.10 and your estimates of β0 and β1 .

11. Verify the claim that for Equation 3.18 γ̂0 ≈ 30, γ̂1 ≈ −.07 and σ̂ ≈ 3.9.
220 CHAPTER 3. REGRESSION

12. Does a football filled with helium travel further than one filled with air?
DASL has a data set that attempts to answer the question. Go to DASL,
http://lib.stat.cmu.edu/DASL, download the data set Helium foot-
ball and read the story. Use what you know about linear models to analyze
the data and reach a conclusion. You must decide whether to include data
from the first several kicks and from kicks that appear to be flubbed. Does
your decision affect your conclusion?
13. Use the PlantGrowth data from R. Refer to page 184 and Equation 3.5.
(a) Estimate µC , µT 1 , µT 2 and σ.
(b) Test the hypothesis µT 1 = µC .
(c) Test the hypothesis µT 1 = µT 2 .
14. Jack and Jill, two Duke sophomores, have to choose their majors. They
both love poetry so they might choose to be English majors. Then their
futures would be full of black clothes, black coffee, low paying jobs, and
occasional volumes of poetry published by independent, non-commercial
presses. On the other hand, they both see the value of money, so they
could choose to be Economics majors. Then their futures would be full
of power suits, double cappucinos, investment banking and, at least for
Jack, membership in the Augusta National golf club. But which would
make them more happy?
To investigate, they conduct a survey. Not wanting to embarass their
friends and themselves, Jack and Jill go up Chapel Hill to interview poets
and investment bankers. In all of Chapel Hill there are 90 poets but only
10 investment bankers. J&J interview them all. From the interviews J&J
compute the Happiness Quotient or HQ of each subject. The HQ’s are
in Figure 3.19. J&J also record two indicator variables for each person:
Pi = 1 or 0 (for poets and bankers); Bi = 1 or 0 (for bankers and poets).
Jill and Jack each write a statistical model:

Jill: HQi = α0 + α1 Pi + i Jack: HQi = β1 Pi + β2 Bi + i

(a) Say in words what are α0 , α1 , β1 and β2 .

(b) In their data set J&J find HQ = 43 among poets, HQ = 44 among
bankers and σ̂ 2 = 1. (Subjects report disappointment with their
favorite basketball team as the primary reason for low HQ.) Find
sensible numerical estimates of α0 , α1 , β1 and β2 .

15. Is poverty related to academic performance in school? The file

schools_poverty

at this text’s website contains relevant data from the Durham, NC school
system in 2001. The first few lines are
3.5. EXERCISES 221

p
b

39 40 41 42 43 44 45

HQ

Figure 3.19: Happiness Quotient of bankers and poets

pfl eog type

1 66 65 e
2 32 73 m
3 65 65 e

Each school in the Durham public school system is represented by one

line in the file. The variable pfl stands for percent free lunch. It records
the percentage of the school’s student population that qualifies for a free
lunch program. It is an indicator of poverty. The variable eog stands for
end of grade. It is the school’s average score on end of grade tests and
is an indicator of academic success. Finally, type indicates the type of
school — e, m, or h for elementary, middle or high school, respectively.
You are to investigate whether pfl is predictive of eog.

(a) Read the data into R and plot it in a sensible way. Use different plot
symbols for the three types of schools.
(b) Does there appear to be a relationship between pfl and eog? Is the
relationship the same for the three types of schools? Decide whether
the rest of your analysis should include all types of schools, or only
one or two.
(c) Using the types of schools you think best, remake the plot and add
a regression line. Say in words what the regression line means.
222 CHAPTER 3. REGRESSION

(d) During the 2000-2001 school year Duke University, in Durham, NC,
sponsored a tutoring program in one of the elementary schools. Many
Duke students served as tutors. From looking at the plot, and as-
suming the program was successful, can you figure out which school
it was?
16. Load mtcars into an R session. Use R to find the m.l.e.’s (β̂0 , β̂1 ). Confirm
that they agree with the line drawn in Figure 3.11(a). Starting from
Equation 3.17, derive the m.l.e.’s for β0 and β1 .
17. Get more current data similar to mtcars. Carry out a regression analysis
similar to Example 3.7. Have relationships among the variables changed
over time? What are now the most important predictors of mpg?
18. Repeat the logistic regression of am on wt, but use hp instead of wt.
19. A researcher randomly selects cities in the US. For each city she records
the number of bars yi and the number of churches zi . In the regression
equation zi = β0 + β1 yi do you expect β1 to be positive, negative, or
around 0?
20. Jevons’ coins?
21. (a) Jane writes the following R code:
x <- runif ( 60, -1, 1 )

Describe x. Is it a number, a vector, or a matrix? What is in it?

(b) Now she writes
y <- x + rnorm ( 60 )
myfit <- lm ( y ~ x )

Make an intelligent guess of what she found for β̂0 and β̂1 .
(c) Using advanced statistical theory she calculates

SD(β̂0 ) = .13
SD(β̂1 ) = .22

Finally she writes

in0 <- 0
in1 <- 0
for ( i in 1:100 ) {
x <- runif ( 60, -1, 1 )
y <- x + rnorm ( 60 )
fit <- lm ( y ~ x )
if ( abs(fit$coef[1]) > .26 ) in0 <- in0 + 1
3.5. EXERCISES 223

if ( abs(fit$coef[2]-1) > .44 ) in1 <- in1 + 1

}

Make an intelligent guess of in0 and in1 after Jane ran this code.

22. The Army is testing a new mortar. They fire a shell up at an angle
of 60◦ and track its progress with a laser. Let t1 , t2 , . . . , t100 be equally
spaced times from t1 = (time of firing) to t100 = (time when it lands).
Let y1 , . . . , y100 be the shell’s heights and z1 , . . . , z100 be the shell’s dis-
tance from the howitzer (measured horizontally along the ground) at times
t1 , t2 , . . . , t100 . The yi ’s and zi ’s are measured by the laser. The measure-
ments are not perfect; there is some measurement error. In answering
the following questions you may assume that the shell’s horizontal speed
remains constant until it falls to ground.

(a) True or False: The equation

y i = β 0 + β 1 ti +  i

should fit the data well.

(b) True or False: The equation

yi = β0 + β1 ti + β2 t2i + i (3.23)

should fit the data well.

(c) True or False: The equation

z i = β 0 + β 1 ti +  i (3.24)

should fit the data well.

(d) True or False: The equation

zi = β0 + β1 ti + β2 t2i + i (3.25)

should fit the data well.

(e) True or False: The equation

yi = β 0 + β 1 zi +  i (3.26)

should fit the data well.

(f) True or False: The equation

yi = β0 + β1 zi + β2 zi2 + i (3.27)

should fit the data well.

(g) Approximately what value did the Army find for β̂0 in Part (b)?
224 CHAPTER 3. REGRESSION

(h) Approximately what value did the Army find for β̂2 in Part (d)?
23. Some nonstatisticians (not readers of this book, we hope) do statistical
analyses based almost solely on numerical calculations and don’t use plots.
R comes with the data set anscombe which demonstrates the value of plots.
Type data(anscombe) to load the data into your R session. It is an 11 by
8 dataframe. The variable names are x1, x2, x3, x4, y1, y2, y3, and y4.
(a) Start with x1 and y1. Use lm to model y1 as a function of x1. Print
a summary of the regression so you can see β̂0 , β̂1 , and σ̂.
(b) Do the same for the other pairs: x2 and y2, x3 and y3, x4 and y4.
(c) What do you conclude so far?
(d) Plot y1 versus x1. Repeat for each pair. You may want to put all
four plots on the same page. (It’s not necessary, but you should know
how to draw the regression line on each plot. Do you?)
(e) What do you conclude?
(f) Are any of these pairs well described by linear regression? How would
you describe the others? If the others were not artificially constructed
data, but were real, how would you analyze them?
24. Here’s some R code:

x <- rnorm ( 1, 2, 3 )
y <- -2*x + 1 + rnorm ( 1, 0, 1 )

(a) What is the marginal distribution of x?

(b) Write down the marginal density of x.
(c) What is the conditional distribution of y given x?
(d) Write down the conditional density of y given x.
(e) Write down the joint density of (x,y).

Here’s more R code:

N.sim <- 1000

w <- rep ( NA, N.sim )
for ( i in 1:N.sim ) {
x <- rnorm ( 50, 2, 3 )
y <- -2*x + 1 + rnorm ( 50, 0, 1 )
fit <- lm ( y ~ x )
w[i] <- fit$coef[2]
}
z1 <- mean ( w )
z2 <- sqrt ( var ( w + 2 ) )
3.5. EXERCISES 225

What does z1 estimate? What does z2 estimate?

25. A statistician thinks the regression equation yi = β0 + β1 xi + i fits her

data well. She would like to learn β1 . She is able to measure the yi ’s
accurately but can measure the xi ’s only approximately. In fact, she can
measure wi = xi + δi where δi ∼ N(0, .1). So she can fit the regression
equation yi = β0∗ + β1∗ wi + ∗i . Note that (β0∗ , β1∗ ) might be different than
(β0 , β1 ) because they’re for the wi ’s, not the xi ’s. So the statistician writes
the following R code.

N.sim <- 1000

b.0 <- -10:10
b.1 <- -10:10
n <- 50
for ( i in 1:21 )
for ( j in 1:21 ) {
val <- rep ( NA, N.sim )
for ( k in 1:N.sim ) {
x <- rnorm ( n )
w <- x + rnorm ( n, 0, sqrt(.1) )
y <- b.0[i] + x * b.1[j] + rnorm ( n, 3 )
fit <- lm ( y ~ w )
val[k] <- fit$coef[2]
}
m <- mean ( val )
sd <- sqrt ( var ( val ) )
print ( c ( m, sd, (m-b.1[j])/sd ) )
}

What is she trying to do? The last time through the loop, the print
statement yields [1] 9.086723 0.434638 -2.101237. What does this
show?

26. The purpose of this exercise is to familiarize yourself with plotting logistic
regression curves and getting a feel for the meaning of β0 and β1 .

(a) Choose some values of x. You will want between about 20 and 100
evenly spaced values. These will become the abscissa of your plot.
(b) Choose some values of β0 and β1 . You are trying to see how different
values of the β’s affect the curve. So you might begin with a single
value of β1 and several values of β0 , or vice versa.
(c) For each choice of (β0 , β1 ) calculate the set of θi = eβ0 +β1 xi /(1 +
eβ0 +β1 xi ) and plot θi versus xi . You should get sigmoidal shaped
curves. These are logistic regression curves.
226 CHAPTER 3. REGRESSION

(d) You may find that the particular x’s and β’s you chose do not yield
a visually pleasing result. Perhaps all your θ’s are too close to 0 or
too close to 1. In that case, go back and choose different values. You
will have to play around until you find x’s and β’s compatible with
each other.
27. Carry out a logistic regression analysis of the O-ring data. What does your
analysis say about the probability of O-ring damage at 36◦ F, the temper-
ature of the Challenger launch. How relevant should such an analysis have
been to the decision of whether to postpone the launch?
28. This exercise refers to Example 3.10.
(a) Why are the points lined up vertically in Figure 3.16, panels (a)
and (b)?
(b) Why do panels (c) and (d) appear to have more points than pan-
els (a) and (b)?
(c) If there were no jittering, how many distinct values would there be
on the abscissa of panels (c) and (d)?
(d) Download the seedling data. Fit a model in which year is a predic-
tor but quadrat is not. Compare to fit1. Which do you prefer?
Which variable is more important: quadrat or year? Or are they
both important?
 Chapter 4

More Probability

4.1 More Probability Density

Section 1.2 on page 6 introduced probability densities. Section 4.1 discusses
them further and gives a formal definition.
Let X be a continuous random  variable with cdf FX . Theorem 1.2 on page 7
d
implies that fX (x) = db FX (b)b=x and therefore that we can define the pdf by
0 d

fX (x) ≡ FX (x) = db FX (b)b=x . In fact, this definition is a little too restrictive.
The key property of pdf’s is that the probability of a set A is given by the
integral of the pdf. I.e.,
Z
P[X ∈ A] = fX (x) dx
A
∗
But if f is a function
R that
R differs from fX at only countably many points then,
for any set A, A f ∗ = A fX , so we could just as well have defined
Z
P[X ∈ A] = f ∗ (x) dx
A

There are infinitely many functions having the same integrals as fX and f ∗ .
These functions differ from each other on “sets of measure zero”, terminology
beyond our scope but defined in books on measure theory. For our purposes
we can think of sets of measure zero as sets containing at most countably many
points. In effect, the pdf of X can be arbitrarily changed on sets of measure
zero. It does not matter which of the many equivalent functions we use as the
probability density of X. Thus, we define
Definition 4.1. Any function f such that, for all sets A,
Z
P[X ∈ A] = f (x) dx
A

is called a probability density function, or pdf, for the random variable X. Any
such function may be denoted fX .

227
228 CHAPTER 4. MORE PROBABILITY

Definition 4.1 can be used in an alternate proof of Theorem 1.1 on page 11.
The central step in the proof is just a change-of-variable in an integral, showing
that Theorem 1.1 is, in essence, just a change of variables. For convenience we
restate the theorem before reproving it.
Theorem 1.1 Let X be a random variable with pdf pX . Let g be a differ-
entiable, monotonic, invertible function and define Z = g(X). Then the pdf of
Z is  −1 
−1
 d g (t) 
pZ (t) = pX (g (t))  
dt 

R
Proof. For any set A, P[Z ∈ g(A)] = P[X ∈ A] = A pX (x) dx. Let z = g(x)
and change variables in the integral to get
Z  
−1
 dx 
P[Z ∈ g(A)] = pX (g (z))   dz
g(A) dz
R
I.e., P[Z ∈ g(A)] = g(A) something dz. Therefore something must be pZ (z).
Hence, pZ (z) = pX (g −1 (z))|dx/dz|.

4.2 Random Vectors

It is often useful, even essential, to talk about several random variables simul-
taneously. We have seen many examples throughout the text beginning with
Section 1.5 on joint, marginal, and conditional probabilities. Section 4.2 reviews
the basics and sets out new probability theory for multiple random variables.
Let X1 , . . . , Xn be a set of n random variables. The n-dimensional vector
X~ = (X1 , . . . , Xn ) is called a multivariate random variable or random vector. As
explained below, X ~ has a pdf or pmf, a cdf, an expected value, and a covariance
matrix, all analogous to univariate random variables.

4.2.1 Densities of Random Vectors

~ has a pdf, written
When X1 , . . . , Xn are continuous then X

pX~ (x1 , . . . , xn ).

As in the univariate case, the pdf is any function whose integral yields proba-
bilities. That is, if A is a region in Rn then
Z Z
P[X ~ ∈ A] = · · · p ~ (x1 , · · · , xn ) dx1 . . . dxn
X
A

~ = (X1 , X2 )
For example, let X1 ∼ Exp(1); X2 ∼ Exp(2); X1 ⊥ X2 ; and X
and suppose we want to find P[|X1 − X2 | ≤ 1]. Our plan for solving this
4.2. RANDOM VECTORS 229

problem is to find the joint density pX~ , then integrate pX~ over the region A
where |X1 − X2 | ≤ 1. Because X1 ⊥ X2 , the joint density is
1
pX~ (x1 , x2 ) = pX1 (x1 )pX2 (x2 ) = e−x1 × e−x2 /2
2
To find the region A over which to integrate, it helps to plot the X1 -X2 plane.
Making the plot is left as an exercise.
ZZ
P[|X1 − X2 | ≤ 1] = pX~ (x1 , x2 ) dx1 dx2
A
Z 1 Z x1 +1 Z ∞ Z x1 +1
1 1
= e−x1 e−x2 /2 dx2 dx1 + e−x1 e−x2 /2 dx2 dx1
2 0 0 2 1 x1 −1
≈ 0.47 (4.1)
The random variables (X1 , . . . , Xn ) are said to be mutually independent or
jointly independent if
pX~ (x1 , . . . , xn ) = pX1 (x1 ) × · · · × pXn (xn )
for all vectors (x1 , . . . , xn ).
Mutual independence implies pairwise independence. I.e., if (X1 , . . . , Xn )
are mutually independent, then any pair (Xi , Xj ) are also independent. The
proof is left as an exercise. It is curious but true that pairwise independence
does not imply joint independence. For an example, consider the discrete three-
dimensional distribution on X ~ = (X1 , X2 , X3 ) with

P[(X1 , X2 , X3 )] = (0, 0, 0)
= P[(X1 , X2 , X3 )] = (1, 0, 1)
(4.2)
= P[(X1 , X2 , X3 )] = (0, 1, 1)
= P[(X1 , X2 , X3 )] = (1, 1, 0) = 1/4
It is easily verified that X1 ⊥ X2 , X1 ⊥ X3 , and X2 ⊥ X3 but that X1 , X2 ,
and X3 are not mutually independent. See Exercise 6.

4.2.2 Moments of Random Vectors

~ is a random vector, its expected value is also a vector.
When X
~ ≡ (E[X1 ], . . . , E[Xn ])
E[X]
When X ~ ≡ (X1 , . . . , Xn ) is a random vector, instead of a variance it has a
covariance matrix. The ij’th entry of the covariance matrix is Cov(Xi , Xj ).
The notation is
 2 
σ1 σ12 · · · σ1n
 σ12 σ22 · · · σ2n 
~
Cov(X) ≡ ΣX~ =  .

.. ..

.. 
 .. . . . 
σ1n σ2n ··· σn2
230 CHAPTER 4. MORE PROBABILITY

where σij = Cov(Xi , Xj ) and σi2 = Var(Xi ). Sometimes σi2 is also denoted σii .

4.2.3 Functions of Random Vectors

Section 4.2.3 considers functions of random vectors. If g is an arbitrary function
that maps X ~ to R then
Z Z
~
E[g(X)] = · · · g(X) ~ dx1 · · · dxn

~ When g is a
but it’s hard to say much in general about the variance of g(X).
linear function we can go farther, but first we need a lemma.
Lemma 4.1. Let X1 and X2 be random variables and Y = X1 + X2 . Then
1. E[Y ] = E[X1 ] + E[X2 ]
2. Var(Y ) = Var(X1 ) + Var(X2 ) + 2 Cov(X1 , X2 )
Proof. Left as exercise.
Now we can deal with linear combinations of random vectors.
Theorem 4.2. Let ~a = (a1 , . . . , an ) be an n-dimensional vector and define
~ = P ai Xi . Then,
Y = ~a0 X
P P
1. E[Y ] = E[ ai Xi ] = ai E[Xi ]
P Pn−1 Pn
2. Var(Y ) = a2i Var(Xi ) + 2 i=1 a0 ΣX~ ~a
j=i+1 ai aj Cov(Xi , Xj ) = ~

Proof. Use Lemma 4.1 and Theorems 1.3 (pg. 33) and 1.4 (pg. 33). See Exer-
cise 8.
The next step is to consider several linear combinations simultaneously. For
some k ≤ n, and for each i = 1, . . . , k, let
X
Yi = ai1 X1 + · · · ain Xn = ~
aij Xj = ~a0i X
j

where the aij ’s are arbitrary constants and ~ai = (ai1 , . . . , ain ). ~ =
Let Y
(Y1 , . . . , Yk ). In matrix notation,
~ = AX
Y ~
where A is the k × n matrix of elements aij . Covariances of the Yi ’s are given
by
~ ~a0 X)
Cov(Yi , Yj ) = Cov(~a0i X, ~
j
n
XX n
= Cov(aik Xk , aj` Xj )
k=1 `=1
Xn n−1
X n
X
= aik ajk σk2 + (aik aj` + ajk ai` )σk`
k=1 k=1 `=k+1

= ~a0i ΣX~ ~aj

4.2. RANDOM VECTORS 231

Combining the previous result with Theorem 4.2 yields Theorem 4.3.
Theorem 4.3. Let X ~ be a random vector of dimension n with mean E[X]~ =µ
~ = Σ; let A be a k × n matrix of rank k; and let
and covariance matrix Cov(X)
~ ~
Y = AX. Then
~ ] = Aµ, and
1. E[Y
~ ) = AΣA0
2. Cov(Y
Finally, we take up the question of multivariate transformations, extending
the univariate version, Theorem 1.1 (pg. 11). Let X ~ = (X1 , . . . , Xn ) be an n-
dimensional continuous random vector with pdf fX~ . Define a new n-dimensional
random vector Y~ = (Y1 , . . . , Yn ) = (g1 (X),
~ . . . , gn (X))
~ where the gi ’s are differ-
entiable functions and where the the transformation g : X ~ 7→ Y ~ is invertible.
~?
What is fY~ , the pdf of Y
Let J be the so-called Jacobian matrix of partial derivatives.
 ∂Y1 ∂Y1 
∂X1 ··· ∂Xn
 ∂Y2 ··· ∂Y2 
 ∂X1 ∂Xn 
J = . .. .. 
 .. . . 
∂Yn ∂Yn
∂X1 ··· ∂Xn

and |J| be the absolute value of the determinant of J.

Theorem 4.4.
fY~ (~y) = fX~ (g −1 (y))|J|−1
Proof. The proof follows the alternate proof
R R of Theorem 1.1 on page 228. For
~ ∈ g(A)] = P[X
any set A, P[Y ~ ∈ A] = ··· p ~ (~x) dx1 · · · dxn . Let ~y = g(~x)
A X
and change variables in the integral to get
Z Z
~ −1
P[Y ∈ g(A)] = · · · pX~ (g −1 (~y )) |J| dy1 · · · dyn
g(A)

R R
~ ∈ g(A)] = ···
I.e., P[Y g(A) something dy1 · · · dyn . Therefore something must
be pY~ (~y ). Hence, pY~ (~y ) = pX~ (g −1 (~y ))|J|−1 .
To illustrate the use of Theorem 4.4 we solve again an example previously
given on page 228, which we restate here. Let X1 ∼ Exp(1); X2 ∼ Exp(2);
X1 ⊥ X2 ; and X ~ = (X1 , X2 ) and suppose we want to find P[|X1 − X2 | ≤ 1].
We solved this problem previously by finding the joint density of X ~ = (X1 , X2 ),
then integrating over the region where |X1 − X2 | ≤ 1. Our strategy this time is
to define new variables Y1 = X1 − X2 and Y2 , which is essentially arbitrary, find
~ = (Y1 , Y2 ), then integrate over the region where |Y1 | ≤ 1.
the joint density of Y
We define Y1 = X1 − X2 because that’s the variable we’re interested in. We
need a Y2 because Theorem 4.4 is for full rank transformations from Rn to Rn .
232 CHAPTER 4. MORE PROBABILITY

The precise definition of Y2 is unimportant, as long as the transformation from

~ to Y
X ~ is differentiable and invertible. For convenience, we define Y2 = X2 .
With these definitions,
!  
∂Y1 ∂Y1
∂X ∂X 1 −1
J = ∂Y21 ∂Y22 =
∂X1 ∂X2
0 1
|J| = 1
X1 = Y 1 + Y 2

and

X2 = Y 2

From the solution on page 228 we know pX~ (x1 , x2 ) = e−x1 × 21 e−x2 /2 , so
pY~ (y1 , y2 ) = e−(y1 +y2 ) × 12 e−y2 /2 = 21 e−y1 e−3y2 /2 . Figure 4.1 shows the region
over which to integrate.
ZZ
P[|X1 − X2 | ≤ 1] = P[|Y1 | ≤ 1] = pY~ (y1 , y2 ) dy1 dy2
A
Z 0 Z ∞ Z 1 Z ∞
1 1
= e−y1 e−3y2 /2 dy2 dy1 + e−y1 e−3y2 /2 dy2 dy1
2 −1 −y1 2 0 0
Z 0 h i∞ Z 1 h i∞
1 1
= e−y1 −e−3y2 /2 dy1 + e−y1 −e−3y2 /2 dy1
3 −1 −y1 3 0 0
Z 0 Z 1
1 1
= ey1 /2 dy1 + e−y1 dy1
3 −1 3 0
2 y1 /2 0 1 −y1 1
= e  − e 0
3 −1 3
2 h i 1  
= 1 − e−1/2 + 1 − e−1
3 3
≈ 0.47 (4.3)

Figure 4.1 was produced by the following snippet.

par ( mar=c(0,0,0,0) )
plot ( c(0,6), c(0,2), type="n", xlab="", ylab="", xaxt="n",
yaxt="n", bty="n" )
polygon ( c(1,1.9,1.9,1), c(1,1,1.9,1.9), col=gray(.8),
border=NA )
polygon ( c(1,1.2,1.9,1.9,1.7,1), c(1,1,1.7,1.9,1.9,1.2),
col=gray(.5), border=NA )
segments ( 0, 1, 1.9, 1, lwd=3 ) # x1 axis
segments ( 1, 0, 1, 1.9, lwd=3 ) # x2 axis
4.3. REPRESENTING DISTRIBUTIONS 233

X2 Y2

X1 Y1

Figure 4.1: The (X1 , X2 ) plane and the (Y1 , Y2 ) plane. The light gray regions
are where X~ and Y
~ live. The dark gray regions are where |X1 − X2 | ≤ 1.

text ( c(2,1), c(1,2), c(expression(bold(X[1])),

expression(bold(X[2]))) )
polygon ( c(5,5.9,5.9,4), c(1,1,1.9,1.9), col=gray(.8),
border=NA )
polygon ( c(5,5.2,5.2,4.8,4.8), c(1,1,1.9,1.9,1.2),
col=gray(.5), border=NA )
segments ( 4, 1, 5.9, 1, lwd=3 ) # y1 axis
segments ( 5, 0, 5, 1.9, lwd=3 ) # y2 axis
text ( c(6,5), c(1,2), c(expression(bold(Y[1])),
expression(bold(Y[2]))) )

arrows ( 2.5, 1, 3.5, 1, length=.2, lwd=2 )

The point of the example, of course, is not that we care about the answer, but
that we do care about the method. Functions of random variables and random
vectors are common in statistics and probability. There are many methods to
deal with them. The method of transforming the pdf is one that is often useful.

4.3 Representing Distributions

We usually describe a random variable Y through pY — its pmf if Y is discrete
or its pdf if Y is continuous. But there are at least two alternatives. First, any
random variable Y can be described by its cumulative distribution function, or
234 CHAPTER 4. MORE PROBABILITY

cdf, FY which is defined by

(P
c
P[Y = y] if Y is discrete
FY (c) ≡ P[Y ≤ c] = R cy=−∞ (4.4)
−∞
p(y) dy if Y is continuous.

Equation 4.4 defines the cdf in terms of the pmf or pdf. It is also possible
to go the other way. If Y is continuous, then for any number b ∈ R
Z b
P(Y ≤ b) = F (b) = p(y) dy
−∞

which shows by the Fundamental Theorem of Calculus that p(y) = F 0 (y). On

the other hand, if Y is discrete, Then P[Y = y] = P[Y ≤ y] − P[Y < y] =
FY (y) − FY (y − ). (We use the notation FY (y − ) to mean the limit of FY (z) as z
approaches y from below. It is also written lim↑0 FY (y − )). Thus the reverse
of Equation 4.4 is
(
FY (y) − FY (y − ) if Y is discrete
pY (y) = (4.5)
FY0 (y) if Y is continuous

Equation 4.5 is correct except in one case which seldom arises in practice. It
is possible that FY (y) is a continuous but nondifferentiable function, in which
case Y is a continuous random variable, but Y does not have a density. In this
case there is a cdf FY without a corresponding pmf or pdf.
Figure 4.2 shows the pmf and cdf of the Bin(10, .7) distribution and the pdf
and cdf of the Exp(1) distribution.

Figure 4.2 was produced by the following snippet.

par ( mfrow=c(2,2) )
y <- seq ( -1, 11, by=1 )
plot ( y, dbinom ( y, 10, .7 ), type="p", ylab="pmf",
main="Bin (10, .7)" )
plot ( y, pbinom ( y, 10, .7 ), type="p", pch=16,
ylab="cdf", main="Bin (10, .7)" )
segments ( -1:10, pbinom ( -1:10, 10, .7 ),
0:11, pbinom ( -1:10, 10, .7 ) )
y <- seq ( 0, 5, len=50 )
plot ( y, dexp ( y, 1 ), type="l", ylab="pdf", main="Exp(1)" )
plot ( y, pexp ( y, 1 ), type="l", ylab="cdf", main="Exp(1)" )

• segments ( x0, y0, x1, y1) draws line segments. The line segments
run from (x0,y0) to (x1,y1). The arguments may be vectors.
4.3. REPRESENTING DISTRIBUTIONS 235

Bin (10, .7) Bin (10, .7)

0.8
0.20
pmf

cdf
0.10

0.4
0.00

0.0

0 4 8 0 4 8

y y

Exp(1) Exp(1)
0.8
0.8
pdf

cdf

0.4
0.4

0.0
0.0

0 1 2 3 4 5 0 1 2 3 4 5

y y

Figure 4.2: pmf’s, pdf’s, and cdf’s

236 CHAPTER 4. MORE PROBABILITY

The other alternative representation for Y is its moment generating function

or mgf MY . The moment generating function is defined as
(P
tY ety pY (y) if Y is discrete
MY (t) = E[e ] = R yty (4.6)
e pY (y) if Y is continuous
MY is also known as the Laplace transform of pY .
Because we define the mgf as a sum or integral there is the question of
whether the sum or integral is finite and hence whether the mgf is well defined.
In Equation 4.6, the mgf is always defined at t = 0. (See Exercise 10.) But even
if MY (t) is not well defined (the integral or sum is not absolutely convergent) for
large t, what matters for statistical practice is whether MY (t) is well defined in
a neighborhood of 0, i.e. whether there exists a δ > 0 such that MY (t) exists for
t ∈ (−δ, δ). The moment generating function gets its name from the following
theorem.
Theorem 4.5. If Y has mgf MY defined in a neighborhood of 0, then
(n) dn 
E[Y n ] = MY (0) ≡ n MY (t)

dt 0

Proof. We provide the proof for the case n = 1. The proof for larger values of
n is similar.
 Z 
d d
ety pY (y) dy 
 
MY (t) =
dt 0 dt 0
Z
d ty 
= e  pY (y) dy
dt 0
Z 
= yety  pY (y) dy

0
Z
= ypY (y) dy

= E[Y ]

The second line of the proof has the form

Z Z
d d
f (t, y) dy = f (t, y) dy,
dt dt
an equality which is not necessarily true. It is true for “nice” functions f ; but
establishing exactly what “nice” means requires measure theory and is beyond
the scope of this book. We will continue to use the equality without thorough
justification.
One could, if one wished, calculate and plot MY (t), though there is usually
little point in doing so. The main purpose of moment generating functions is in
proving theorems and not, as their name might suggest, in deriving moments.
And mgf’s are useful in proving theorems mostly because of the following two
results.
4.3. REPRESENTING DISTRIBUTIONS 237

Theorem 4.6. Let X and Y be two random variables with moment generating
functions (assumed to exist) MX and MY . If MX (t) = MY (t) for all t in some
neighborhood of 0, then FX = FY ; i.e., X and Y have the same distribution.
Theorem 4.7. Let Y1 , . . . be a sequence of random variables with moment gen-
erating functions (assumed to exist) MY1 , . . . . Define M (t) = limn→∞ MYn (t).
If the limit exists for all t in a neighborhood of 0, and if M (t) is a moment
generating function, then there is a unique cdf FY such that
FY (y) = lim FYn (y)
n→∞

for all y where FY is continuous.

Theorems 4.6 and 4.7 both assume that the necessary mgf’s exist. It is
inconvenient that not all distributions have mgf’s. One can avoid the problem
by using characteristic functions (also known as Fourier transforms) instead of
moment generating functions. The characteristic function is defined as
CY (t) = E[eitY ]
√
where i = −1. All distributions have characteristic functions, and the char-
acteristic function completely characterizes the distribution, so characteristic
functions are ideal for our purpose. However, dealing with complex numbers
presents its own inconveniences. We shall not pursue this topic further. Proofs
of Theorems 4.6 and 4.7 and similar results for characteristic functions are omit-
ted but may be found in more advanced books.
Two more useful results are theorems 4.8 and 4.9.
Theorem 4.8. Let X be a random variable, a, b be constants, and define Y =
aX + b. Then MY (t) = ebt MX (at).
Proof.
h i
MY (t) = E e(aX+b)t
 
= ebt E eatX
= ebt MX (at)

Theorem 4.9. Let X and Y be independent random variables. Define Z =

X + Y . Then
MZ (t) = MX (t)MY (t)
Proof.
h i
MZ (t) = E e(X+Y )t = E[eXt eY t ] = E[eXt ]E[eY t ] = MX (t)MY (t)

Corollary 4.10. Let Y1 , . . . , Yn be a collection of i.i.d. random variables each

with mgf MY . Define X = Y1 + · · · + Yn . Then
n
MX (t) = [MY (t)]
238 CHAPTER 4. MORE PROBABILITY

4.4 Exercises
1. Refer to Equation 4.1 on page 229.

(a) To help visualize the joint density pX~ , make a contour plot. You
will have to choose some values of x1 , some values of x2 , and then
evaluate pX~ (x1 , x2 ) on all pairs (x1 , x2 ) and save the values in a
matrix. Finally, pass the values to the contour function. Choose
values of x1 and x2 that help you visualize pX~ . You may have to
choose values by trial and error.
(b) Draw a diagram that illustrates how to find the region A and the
limits of integration in Equation 4.1.
(c) Supply the missing steps in Equation 4.1. Make sure you understand
them. Verify the answer.
(d) Use R to verify the answer to Equation 4.1 by simulation.

2. Refer to Example 1.6 on page 37 on tree seedlings where N is the number

of New seedlings that emerge in a given year and X is the number that
survive to the next year. Find P[X ≥ 1].

3. (X1 , X2 ) have a joint distribution that is uniform on the unit circle. Find
p(X1 ,X2 ) .

4. The random vector (X, Y ) has pdf p(X,Y ) (x, y) ∝ ky for some k > 0 and
(x, y) in the triangular region bounded by the points (0, 0), (−1, 1), and
(1, 1).

(a) Find k.
(b) Find P[Y ≤ 1/2].
(c) Find P[X ≤ 0].
(d) Find P[|X − Y | ≤ 1/2].

5. Prove the assertion on page 229 that mutual independence implies pairwise
independence.

(a) Begin with the case of three random variables X ~ = (X1 , X2 , X3 ).

Prove that if X1 , X2 , X3 are mutually independent, then any two of
them are independent.
(b) Generalize to the case X ~ = (X1 , . . . , Xn ).

6. Refer to Equation 4.2 on page 229. Verify that X1 ⊥ X2 , X1 ⊥ X3 , and

X2 ⊥ X3 but that X1 , X2 , and X3 are not mutually independent.

7. Prove Lemma 4.1

8. Fill in the proof of Theorem 4.2 on page 230.

4.4. EXERCISES 239

9. Let the random vector (U, V ) be distributed uniformly on the unit square.
Let X = U V and Y = U/V .
(a) Draw the region of the X–Y plane where the random vector (X, Y )
lives.
(b) Find the joint density of (X, Y ).
(c) Find the marginal density of X.
(d) Find the marginal density of Y .
(e) Find P[Y > 1].
(f) Find P[X > 1].
(g) Find P[Y > 1/2].
(h) Find P[X > 1/2].
(i) Find P[XY > 1].
(j) Find P[XY > 1/2].
10. Just below Equation 4.6 is the statement “the mgf is always defined at
t = 0.” For any random variable Y , find MY (0).
11. Provide the proof of Theorem 4.5 for the case n = 2.
12. Refer to Theorem 4.9. Where in the proof is the assumption X ⊥ Y used?
240 CHAPTER 4. MORE PROBABILITY
 Chapter 5

Special Distributions

Statisticians often make use of standard parametric families of probability dis-

tributions. A parametric family is a collection of probability distributions distin-
guished by, or indexed by, a parameter. An example is the Binomial distribution
introduced in Section 1.3.1. There were N trials. Each had a probability θ of
success. Usually θ is unknown and could be any number in (0, 1). There is one
Bin(N, θ) distribution for each value of θ; θ is a parameter; the set of probability
distributions
{Bin(N, θ) : θ ∈ (0, 1)}
is a parametric family of distributions.
We have already seen four parametric families — the
Binomial (Section 1.3.1), Poisson (Section 1.3.2), Exponential (Section 1.3.3),
and Normal (Section 1.3.4) distributions. Chapter 5 examines these in more
detail and introduces several others.

5.1 The Binomial and Negative Binomial Dis-

tributions
The Binomial Distribution Statisticians often deal with situations in which
there is a collection of trials performed under identical circumstances; each
trial results in either success or failure. Typical examples are coin flips (Heads
or Tails), medical trials (cure or not), voter polls (Democrat or Republican),
basketball free throws (make or miss). Conditions for the Binomial Distribution
are
1. the number of trials n is fixed in advance,
2. the probability of success θ is the same for each trial, and
3. trials are conditionally independent of each other, given θ.
Let the random variable X be the number of successes in such a collection of
trials. Then X is said to have the Binomial distribution with parameters (n, θ),

241
242 CHAPTER 5. SPECIAL DISTRIBUTIONS

written X ∼ Bin(n, θ). The possible values of X are the integers 0, 1, . . . , n.

Figure 1.5 shows examples of Binomial pmf’s for several combinations of n and
θ. Usually θ is unknown and the trials are performed in order to learn about θ.
Obviously, large values of X are evidence that θ is large and small values
of X are evidence that θ is small. But to evaluate the evidence quantitatively
we must be able to say more. In particular, once a particular value X = x has
been observed we want to quantify how well it is explained by different possible
values of θ. That is, we want to know p(x | θ).
Theorem 5.1. If X ∼ Bin(n, θ) then
 
n x
pX (x) = θ (1 − θ)n−x
x
for x = 0, 1, . . . , n.
Proof. When the n trials of a Binomial experiment are carried out there will
be a sequence of successes (1’s) and failures (0’s) such as 1000110 · · · 100. Let
S = {0, 1}n be the set of such sequences and, for each x ∈ {0, 1, . . . , n}, let Sx
be the subset of S consisting of sequences with x 1’s and n − x 0’s. If s ∈ Sx
then Pr(s) = θx (1−θ)n−x . In particular, all s’s in Sx have the same probability.
Therefore,

pX (x) = P(X = x) = P(Sx )

= (size of Sx ) · θx (1 − θ)n−x
 
n x
= θ (1 − θ)n−x
x

The special case n = 1 is important enough to have its own name. When
n = 1 then X is said to have a Bernoulli distribution with parameter θ. We
write X ∼ Bern(θ). If X ∼ Bern(θ) then pX (x) = θx (1 − θ)1−x for x ∈ {0, 1}.
Experiments that have two possible outcomes are called Bernoulli trials.
Suppose X1 ∼ Bin(n1 , θ), X2 ∼ Bin(n2 , θ) and X1 ⊥ X2 . Let X3 = X1 +X2 .
What is the distribution of X3 ? Logic suggests the answer is X3 ∼ Bin(n1 +
n2 , θ) because (1) there are n1 + n2 trials, (2) the trials all have the same
probability of success θ, (3) the trials are independent of each other (the reason
for the X1 ⊥ X2 assumption) and (4) X3 is the total number of successes.
Theorem 5.3 shows a formal proof of this proposition. But first we need to
know the moment generating function.
Theorem 5.2. Let X ∼ Bin(n, θ). Then
 n
MX (t) = θet + (1 − θ)

Proof. Let Y ∼ Bern(θ). Then

MY (t) = E[etY ] = θet + (1 − θ).

5.1. BINOMIAL AND NEGATIVE BINOMIAL 243

Pn
Now let X = i=1 Yi where the Yi ’s are i.i.d. Bern(θ) and apply Corollary 4.10.

Theorem 5.3. Suppose X1 ∼ Bin(n1 , θ); X2 ∼ Bin(n1 , θ); and X1 ⊥ X2 . Let

X3 = X1 + X2 . Then X3 ∼ Bin(n1 + n2 , θ).

Proof.

MX3 (t) = MX1 (t)MX2 (t)

  n1  t  n2
= θet + (1 − θ) θe + (1 − θ)
 n1 +n2
= θet + (1 − θ)

The first equality is by Theorem 4.9; the second is by Theorem 5.2. We recognize
the last expression as the mgf of the Bin(n1 + n2 , θ) distribution. So the result
follows by Theorem 4.6.

The mean of the Binomial distribution was calculated in Equation 1.11.

Theorem 5.4 restates that result and gives the variance and standard deviation.

Theorem 5.4. Let X ∼ Bin(n, θ). Then

1. E[X] = nθ.

2. Var(X) = nθ(1 − θ).

p
3. SD(X) = nθ(1 − θ).

Proof. The proof for E[X] was given earlier. If X ∼ Bin(n, θ), then X =
P n
i=1 Xi where Xi ∼ Bern(θ) and the Xi ’s are mutually independent. Therefore,
by Theorem 1.9, Var(X) = n Var(Xi ). But

Var(Xi ) = E(Xi2 ) − E(Xi )2 = θ − θ2 = θ(1 − θ).

So Var(X) = nθ(1 − θ). The result for SD(X) follows immediately.

R comes with built-in functions for working with Binomial distributions.

You can get the following information by typing help(dbinom), help(pbinom),
help(qbinom), or help(rbinom). There are similar functions for working with
other distributions, but we won’t repeat their help pages here.

Usage:

dbinom(x, size, prob, log = FALSE)

pbinom(q, size, prob, lower.tail = TRUE, log.p = FALSE)
qbinom(p, size, prob, lower.tail = TRUE, log.p = FALSE)
rbinom(n, size, prob)

Arguments:
244 CHAPTER 5. SPECIAL DISTRIBUTIONS

x, q: vector of quantiles.

p: vector of probabilities.

n: number of observations. If ‘length(n) > 1’, the length is

taken to be the number required.

size: number of trials.

prob: probability of success on each trial.

log, log.p: logical; if TRUE, probabilities p are given as log(p).

lower.tail: logical; if TRUE (default), probabilities are P[X <= x],

otherwise, P[X > x].

Details:

The binomial distribution with ‘size’ = n and ‘prob’ = p has

density

p(x) = choose(n,x) p^x (1-p)^(n-x)

for x = 0, ..., n.

If an element of ‘x’ is not integer, the result of ‘dbinom’ is

zero, with a warning. p(x) is computed using Loader’s algorithm,
see the reference below.

The quantile is defined as the smallest value x such that

F(x) >= p, where F is the distribution function.

Value:

‘dbinom’ gives the density, ‘pbinom’ gives the distribution

function, ‘qbinom’ gives the quantile function and ‘rbinom’
generates random deviates.

If ‘size’ is not an integer, ‘NaN’ is returned.

References:

Catherine Loader (2000). Fast and Accurate Computation of

Binomial Probabilities; manuscript available from <URL:
http://cm.bell-labs.com/cm/ms/departments/sia/catherine/dbinom>
5.1. BINOMIAL AND NEGATIVE BINOMIAL 245

See Also:

‘dnbinom’ for the negative binomial, and ‘dpois’ for the Poisson
distribution.

Examples:

# Compute P(45 < X < 55) for X Binomial(100,0.5)

sum(dbinom(46:54, 100, 0.5))

## Using "log = TRUE" for an extended range :

n <- 2000
k <- seq(0, n, by = 20)
plot (k, dbinom(k, n, pi/10, log=TRUE), type=’l’, ylab="log density",
main = "dbinom(*, log=TRUE) is better than log(dbinom(*))")
lines(k, log(dbinom(k, n, pi/10)), col=’red’, lwd=2)
## extreme points are omitted since dbinom gives 0.
mtext("dbinom(k, log=TRUE)", adj=0)
mtext("extended range", adj=0, line = -1, font=4)
mtext("log(dbinom(k))", col="red", adj=1)
Figure 5.1 shows the Binomial pmf for several values of x, n, and p. Note
that for a fixed p, as n gets larger the pmf looks increasingly like a Normal pdf.
That’s the Central Limit Theorem. Let Y1 , . . . , Yn P∼ i.i.d. Bern(p). Then the
distribution of X is thePsame as the distribution of Yi and the Central Limit
Theorem tells us that Yi looks increasingly Normal as n → ∞.
Also, for a fixed n, the pmf looks more Normal when p = .5 than when
p = .05. And that’s because convergence under the Central Limit Theorem is
faster when the distribution of each Yi is more symmetric.

Figure 5.1 was produced by

par ( mfrow=c(3,2) )
n <- 5
p <- .05
x <- 0:5
plot ( x, dbinom(x,n,p), ylab="p(x)", main="n=5, p=.05" )
...

The Negative Binomial Distribution Rather than fix in advance the num-
ber of trials, experimenters will sometimes continue the sequence of trials until
a prespecified number of successes r has been achieved. In this case the total
number of failures N is the random variable and is said to have the Negative
Binomial distribution with parameters (r, θ), written N ∼ NegBin(r, θ). (Warn-
ing: some authors say that the total number of trials, N + r, has the Negative
246 CHAPTER 5. SPECIAL DISTRIBUTIONS

n=5, p=.05 n=5, p=.5

0.8

0.25
0.6
p(x)

p(x)
0.4

0.15
0.2

0.05
0.0

0 1 2 3 4 5 0 1 2 3 4 5

x x

n=20, p=.05 n=20, p=.5

0.05 0.10 0.15

0.3
0.2
p(x)

p(x)
0.1
0.0

0 1 2 3 4 5 6 8 10 12 14

x x

n=80, p=.05 n=80, p=.5

0.15

0.06
p(x)

p(x)
0.05

0.02

0 2 4 6 8 30 35 40 45 50

x x

Figure 5.1: The Binomial pmf

5.1. BINOMIAL AND NEGATIVE BINOMIAL 247

Binomial distribution.) One example is a gambler who decides to play the daily
lottery until she wins. The specified number of successes is r = 1. The number
of losses N until she wins is random. When r = 1, N is said to have a Geomet-
ric distribution with parameter θ; we write N ∼ Geo(θ). Often, θ is unknown.
Large values of N are evidence that θ is small; small values of N are evidence
that θ is large. The probability function is

pN (k) = P(N = k)
= P(r − 1 successes in the first k + r − 1 trials
and k + r’th trial is a success)
 
k+r−1 r
= θ (1 − θ)k
r−1
for k = 0, 1, . . . .
Let N1 ∼ NegBin(r1 , θ), . . . , Nt ∼ NegBin(rt , θ), and N1 , . . . , Nt be in-
dependent
P of P each other. Then one can imagine a sequence of trials of length
Ni having ri successes. N1 is the trial on which the r1 ’th success occurs;
N1 + · · · + Nk is P the trial on which the r1 + · · · + rP
k ’th success occurs. It is
evident that N ≡ Ni is the trial on which the r ≡ ri ’th success occurs and
therefore that N ∼ NegBin(r, θ).
Theorem 5.5. If Y ∼ NegBin(r, θ) then E[Y ] = r(1 − θ)/θ and Var(Y ) =
r(1 − θ)/θ2 .
Proof. It suffices to prove the result for r = 1. Then the result for r > 1 will
follow by Theorems 1.7 and 1.9. For r = 1,
∞
X
E[N ] = n P[N = n]
n=0
X∞
= n(1 − θ)n θ
n=1
∞
X
= θ(1 − θ) n(1 − θ)n−1
n=1
∞
X d
= −θ(1 − θ) (1 − θ)n
n=1
dθ
∞
d X
= −θ(1 − θ) (1 − θ)n
dθ n=1
d 1−θ
= −θ(1 − θ)
dθ θ
−1
= −θ(1 − θ) 2
θ
1−θ
=
θ
248 CHAPTER 5. SPECIAL DISTRIBUTIONS

The trick of writing each term as a derivative, then switching the order of
summation and derivative is occasionally useful. Here it is again.
∞
X
E(N 2 ) = n2 P[N = n]
n=0
∞
X
= θ(1 − θ) (n(n − 1) + n) (1 − θ)n−1
n=1
X∞ ∞
X
= θ(1 − θ) n(1 − θ)n−1 + θ(1 − θ)2 n(n − 1)(1 − θ)n−2
n=1 n=1
∞
1−θ X d2
= + θ(1 − θ)2 2
(1 − θ)n
θ n=1
θ
∞
1−θ d2 X
= + θ(1 − θ)2 2 (1 − θ)n
θ θ n=1
1−θ d2 1 − θ
= + θ(1 − θ)2 2
θ θ θ
1−θ θ(1 − θ)2
= +2
θ θ3
2 − 3θ − 2θ 2
=
θ2
Therefore,
1−θ
Var(N ) = E[N 2 ] − (E[N ])2 = .
θ2

The R functions for working with the negative Binomial distribution are
dnbinom, pnbinom, qnbinom, and rnbinom. Figure 5.2 displays the Negative
Binomial pdf and illustrates the use of qnbinom.

Figure 5.2 was produced with the following snippet.

r <- c ( 1, 5, 30 )
p <- c ( .1, .5, .8 )
par ( mfrow=c(3,3) )
for ( i in seq(along=r) )
for ( j in seq(along=p) ) {
lo <- qnbinom(.01,r[i],p[j])
hi <- qnbinom(.99,r[i],p[j])
x <- lo:hi
plot ( x, dnbinom(x,r[i],p[j]), ylab="probability", xlab="N",
main = paste ( "r=", r[i], ", ",
"theta =", p[j], sep="" ) )
}
5.1. BINOMIAL AND NEGATIVE BINOMIAL 249

r=1, theta =0.1 r=1, theta =0.5 r=1, theta =0.8

0.8
0.08

0.4

0.6
probability

probability

probability

0.4
0.04

0.2

0.2
0.00

0.0
0 20 40 0 2 4 6 0.0 1.0 2.0

N N N

r=5, theta =0.1 r=5, theta =0.5 r=5, theta =0.8

0.02 0.06 0.10 0.14
0.020

0.25
probability

probability

probability
0.010

0.15
0.05
0.000

20 60 100 0 4 8 14 0 2 4

N N N

r=30, theta =0.1 r=30, theta =0.5 r=30, theta =0.8

0.05
0.006

0.02 0.06 0.10

probability

probability

probability
0.03
0.002

0.01

200 350 15 30 45 2 6 12

N N N

Figure 5.2: The Negative Binomial pmf

250 CHAPTER 5. SPECIAL DISTRIBUTIONS

• lo and hi are the limits on the x-axis of each plot. The use of qbinom
ensures that each plot shows at least 98% of its distribution.

5.2 The Multinomial Distribution

The multinomial distribution generalizes the binomial distribution in the fol-
lowing way. The binomial distribution applies when the outcome of a trial has
two possible values; the multinomial distribution applies when the outcome of
a trial has more than two possible outcomes. Some examples are
Clinical Trials In clinical trials, each patient is administered a treatment, usu-
ally an experimental treatment or a standard, control treatment. Later,
each patient may be scored as either success, failure, or censored. Censor-
ing occurs because patients don’t show up for their appointments, move
away, or can’t be found for some other reason.
Craps After the come-out roll, each successive roll is either a win, loss, or
neither.
Genetics Each gene comes in several variants. Every person has two copies of
the gene, one maternal and one paternal. So the person’s status can be
described by a pair like {a, c} meaning that she has one copy of type a
and one copy of type c. The pair is called the person’s genotype. Each
person in a sample can be considered a trial. Geneticists may count how
many people have each genotype.
Political Science In an election, each person prefers either the Republican
candidate, the Democrat, the Green, or is undecided.
In this case we count the number of outcomes of each type. If there are k
possible outcomes then the result is a vector y1 , . . . , yk where yi is the number
of times that outcome i occured and y1 + · · · + yk = n is the number of trials.
Let p ≡ (p1 , . . . , pk ) be the probabilities of the k categories and n be the
number of trials. We write Y ∼ Mult(n, p). In particular, Y ≡ (Y1 , . . . , Yk ) is a
vector of length k. Because Y is a vector, so is its expectation

E[Y ] = µ = (µ1 , . . . , µk ) ≡ (E[Y1 ], . . . , E[Yk ]) = (np1 , . . . , npk ).

The i’th coordinate, Yi , is a random variable in its own right. Because Yi

counts the number of times outcome i occurred in n trials, its distribution is

Yi ∼ Bin(n, pi ). (See Exercise ??.) (5.1)

Although the Yi ’s are all Binomial, they are not independent. After all, if
Y1 = n, then Y2 = · · · = Yk = 0, so the Yi ’s must be dependent. What is their
joint pmf? What is the conditional distribution of, say, Y2 , . . . , Yk given Y1 ?
The next two theorems provide the answers.
5.3. POISSON 251

Theorem 5.6. If Y ∼ Mult(n, p) then

 
n
fY (y1 , . . . , yk ) = py1 · · · pykk
y1 · · · yk 1
n


where y1 ···yk is the multinomial coefficient
 
n n!
=Q
y1 · · · yk yk !

Proof. When the n trials of a multinomial experiment are carried out, there will
be a sequence of outcomes such as abkdbg · · · f , where the letters indicate the
outcomes of individual trials. One such sequence is

· · a} |b ·{z
|a ·{z · · }b · · · k
| ·{z
· · k}
y1 times y2 times yk times
Q
The probability of this particular sequence is pyi i . Every sequence with y1
a’s, . . . , yk k’s has the same probability. So
 Y
Y n
fY (y1 , . . . , yk ) = (number of such sequences) × pyi i = pyi i .
y1 · · · yk

Theorem 5.7. If Y ∼ Mult(n, p) then

(Y2 , . . . , Yk | Y1 = y1 ) ∼ Mult(n − y1 , (p∗2 , . . . , p∗k ))

where p∗i = pi /(1 − p1 ) for i = 2, . . . , k.

Proof. See Exercise 19.

R’s functions for the multinomial distribution are rmultinom and dmultinom.
rmultinom(m,n,p) draws a sample of size m. p is a vector of probabilities. The
result is a k × m matrix. Each column is one draw, so each column sums to n.
The user does not specify k; it is determined by k = length(p).

5.3 The Poisson Distribution

The Poisson distribution is used to model counts in the following situation.

• There is a domain of study, usually a block of space or time.

• Events arise at seemingly random locations in the domain.

• There is an underlying rate at which events arise.

• The rate does not vary over the domain.
252 CHAPTER 5. SPECIAL DISTRIBUTIONS

• The occurence of an event at any location `1 is independent of the oc-

curence of an event at any other location `2 .
Let y be the total number of events that arise in the domain. Y has a Poisson
distribution with rate parameter λ, written Y ∼ Poi(λ). The pmf is
e−λ λy
pY (y) = for y = 0, 1, . . .
y!
The mean was derived in Chapter 1, Exercise 18a. It is
E[Y ] = λ.
Theorem 5.8. Let Y ∼ Poi(λ). Then
t
MY (t) = eλ(e −1)

Proof.
∞ ∞
X X e−λ λy
MY (t) = E[etY ] = ety pY (y) = ety
y=0 y=0
y!
∞ ∞ t
X e−λ (λet )y e−λ X e−λe (λet )y
= =
y=0
y! e−λet y=0
y!
t
= eλ(e −1)

Theorem 5.9. Let Y ∼ Poi(λ). Then

Var(Y ) = λ
Proof. Just for fun (!) we will prove the theorem two ways — first directly and
then with moment generating functions.

Proof 1.
∞
2
X e−λ λy
E[Y ] = y2
y=0
y!
∞ ∞
X e−λ λy X e−λ λy
= y(y − 1) + y
y=0
y! y=0
y!
∞
X e−λ λy
= y(y − 1) +λ
y=2
y!
∞
X e−λ λz+2
= +λ
z=0
z!
2
=λ +λ
So Var(Y ) = E[Y 2 ] − (E[Y ])2 = λ.
5.3. POISSON 253

Proof 2.
d2 
E[Y 2 ] =

M Y (t) 
dt2 t=0
d2 λ(et −1) 
= 2e 
dt t=0
d t λ(et −1) 
= λe e 
hdt t
t=0
t
i
= λet eλ(e −1)
+ λ2 e2t eλ(e −1)
t=0
= λ + λ2

So Var(Y ) = E[Y 2 ] − (E[Y ])2 = λ.

Theorem 5.10. Let YPi ∼ Poi(λi ) forP
i = 1, . . . , n and let the Yi s be mutually
n n
independent. Let Y = 1 Yi and λ = 1 λi . Then Y ∼ Poi(λ).
Proof. Using Theorems 4.9 and 5.8 we have
Y Y t t
MY (t) = MYi (t) = eλi (e −1) = eλ(e −1)

which is the mgf of the Poi(λ) distribution.

Suppose, for i = 1, . . . , n Yi is the number of events occurring on a domain
Di ; Y i ∼ P
Poi(λi ). Suppose the Di ’s are disjoint and the Yi ’s are independent.
Let Y = Yi be the number of events arising P on D = ∪Di . The logic of the
situation suggests that Y ∼ Poi(λ) where λ = λi . Theorem 5.10 assures us
that everything works correctly; that Y does indeed have the Poi(λ) distribution.
Figure 5.3 shows the Poisson pmf for λ = 1, 4, 16, 64. As λ increases the
pmf looks increasingly Normal. That’s a consequence of Theorem 5.10 and the
Central Limit Theorem. When Y ∼ Poi(λ) Theorem 5.10 tells us we can think
Pλ
of Y as Y = i=1 Yi where each Yi ∼ Poi(1). (λ must be an integer for this to be
precise.) Then the Central Limit Theorem tells us that Y will be approximately
Normal when λ is large.
One of the early uses of the Poisson distribution was in The probability
variations in the distribution of α particles by Rutherford and Geiger 1910.
(An α particle is a Helium nucleus, or two protons and two neutrons.)

Example 5.1 (Rutherford and Geiger)

The phenomenon of radioactivity was beginning to be understood in the early 20 th
century. In their 1910 article, Rutherford and Geiger write
“In counting the α particles emitted from radioactive substances . . . [it]
is of importance to settle whether . . . variations in distribution are in
agreement with the laws of probability, i.e. whether the distribution of
α particles on an average is that to be anticipated if the α particles
are expelled at random both in regard to space and time. It might
be conceived, for example, that the emission of an α particle might
254 CHAPTER 5. SPECIAL DISTRIBUTIONS

λ=1 λ=4

0.20
0.3

0.15
0.2
pY(y)

pY(y)

0.10
0.1

0.05
0.0

0.00
0 2 4 6 0 2 4 6 8

y y

λ = 16 λ = 64
0.08

0.04
pY(y)

pY(y)
0.04

0.02
0.00

0.00

10 15 20 25 50 60 70 80

y y

Figure 5.3: Poisson pmf for λ = 1, 4, 16, 64

5.3. POISSON 255

precipitate the disintegration of neighbouring atoms, and so lead to a

distribution of α particles at variance with the simple probability law.”

So Rutherford and Geiger are going to do three things in their article. They’re
going to count α particle emissions from some radioactive substance; they’re going
to derive the distribution of α particle emissions according to theory; and they’re
going to compare the actual and theoretical distributions.
Here they describe their experimental setup.

“The source of radiation was a small disk coated with polonium, which
was placed inside an exhausted tube, closed at one end by a zinc sulphide
screen. The scintilations were counted in the usual way . . . the number
of scintillations . . . corresponding to 1/8 minute intervals were counted
....
“The following example is an illustration of the result obtained. The
numbers, given in the horizontal lines, correspond to the number of
scintillations for successive intervals of 7.5 seconds.

Total per minute.

1st minute : 3 7 4 4 2 3 2 0. . . 25
2nd ” 5 2 5 4 3 5 4 2 .. 30
3rd ” 5 4 1 3 3 1 5 2 .. 24
4th ” 8 2 2 2 3 4 2 6 .. 31
5th ” 7 4 2 6 4 5 10 4 .. 42

Average for 5 minutes . . . 30.4

True average . . . . . . . . . . . 31.0

And here they describe their theoretical result.

“The distribution of α particles according to the law of probability was

kindly worked out for us by Mr. Bateman. The mathematical theory
is appended as a note to this paper. Mr. Bateman has shown that if
x be the true average number of particles for any given interval falling
on the screen from a constant source, the probability
n
that n α particles
are observed in the same interval is given by xn! e−x . n is here a whole
number, which may have all positive values from 0 to ∞. The value of
x is determined by counting a large number of scintillations and dividing
by the number of intervals involved. The probability for n α particles
in the given interval can then at once be calculated from the theory.”

Refer to Bateman [1910] for his derivation. Table 5.1 shows their data. As Ruther-
ford and Geiger explain:

“For convenience the tape was measured up in four parts, the results
of which are given separately in horizontal columns I. to IV.
256 CHAPTER 5. SPECIAL DISTRIBUTIONS

“For example (see column I.), out of 792 intervals of 1/8 minute, in
which 3179 α particles were counted, the number of intervals 3 α par-
ticles was 152. Combining the four columns, it is seen that out of 2608
intervals containing 10,097 particles, the number of times that 3 α par-
ticles were observed was 525. The number calculated from the equation
was the same, viz. 525.”
Finally, how did Rutherford and Geiger compare their actual and theoretical dis-
tributions? They did it with a plot, which we reproduce as Figure 5.4. Their
conclusion:
“It will be seen that, on the whole, theory and experiment are in excel-
lent accord. . . . We may consequently conclude that the distribution of
α particles in time is in agreement with the laws of probability and that
the α particles are emitted at random. . . . Apart from their bearing on
radioactive problems, these results are of interest as an example of a
method of testing the laws of probability by observing the variations in
quantities involved in a spontaneous material process.”

5.4 The Uniform Distribution

The Discrete Uniform Distribution The discrete uniform distribution is
the distribution that gives equal weight to each integer 1, . . . , n. We write
Y ∼ U(1, n). The pmf is
p(y) = 1/n (5.2)
for y = 1, . . . , n. The discrete uniform distribution is used to model, for example,
dice rolls, or any other experiment in which the outcomes are deemed equally
likely. The only parameter is n. It is not an especially useful distribution in
practical work but can be used to illustrate concepts in a simple setting. For an
applied example see Exercise 22.

The Continuous Uniform Distribution The continuous uniform distri-

bution is the distribution whose pdf is flat over the interval [a, b]. We write
Y ∼ U(a, b). Although the notation might be confused with the discrete uni-
form, the context will indicate which is meant. The pdf is

p(y) = 1/(b − a)

for y ∈ [a, b]. The mean, variance, and moment generating function are left as
Exercise 23.
Suppose we observe a random sample y1 , . . . , yn from U(a, b). What is the
m.l.e.(â, b̂)? The joint density is
( n
1
b−a if a ≤ y(1) and b ≥ y(n)
p(y1 , . . . , yn ) =
0 otherwise
 5.4. UNIFORM
Table 5.1: Rutherford and Geiger’s data

Number of 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 Number of Number of Average

α particles α particles intervals number
I .......... 15 56 106 152 170 122 88 50 17 12 3 0 0 1 0 3179 792 4.01
II . . . . . . . . . 17 39 88 116 120 98 63 37 4 9 4 1 0 0 0 2334 596 3.92
III . . . . . . . . 15 56 97 139 118 96 60 26 18 3 3 1 0 0 0 2373 632 3.75
IV . . . . . . . . 10 52 92 118 124 92 62 26 6 3 0 2 0 0 1 2211 588 3.76
Sum . . . . . . 57 203 383 525 532 408 273 139 45 27 10 4 0 1 1 10097 2608 3.87
Theoretical 54 210 407 525 508 394 254 140 68 29 11 4 1 4 1
values

257
258 CHAPTER 5. SPECIAL DISTRIBUTIONS

500
400
Number of Groups

300
200
100
0

0 2 4 6 8 10 12

Number of Particles in Interval

Figure 5.4: Rutherford and Geiger’s Figure 1 comparing theoretical (solid line)
to actual (open circles) distribution of α particle counts.
5.5. GAMMA, EXPONENTIAL, CHI SQUARE 259

which is maximized, as a function of (a, b), if b−a is as small as possible without

making the joint density 0. Thus, â = y(1) and b̂ = y(n) .

5.5 The Gamma, Exponential, and Chi Square

Distributions
“Γ” is the upper case Greek letter Gamma. The gamma function is a special
mathematical function defined on R+ as
Z ∞
Γ(α) = tα−1 e−t dt
0

Information about the gamma function can be found in mathematics texts and
reference books. For our purposes, the key facts are:

Γ(α + 1) = αΓ(α) for α > 1

Γ(n) = (n − 1)! for positive integers n
√
Γ(1/2) = π

For any positive numbers α and β, the Gamma(α, β) distribution has pdf
1
p(y) = y α−1 e−y/β for y ≥ 0 (5.3)
Γ(α)β α

We write Y ∼ Gam(α, β).

Figure 5.5 shows Gamma densities for four values of α and four values of β.
• In each panel of Figure 5.5 the curves for different α’s have different shapes.
Sometimes α is called the shape parameter of the Gamma distribution.
• The four panels look identical except for the axes. I.e., the four curves
with α = .5, one from each panel, have the same shape but different scales.
The different scales correspond to different values of β. For this reason
β is called a scale parameter. One can see directly from Equation 5.3
that β is a scale parameter because p(y) depends on y only through the
ratio y/β. The idea of scale parameter is embodied in Theorem 5.11. See
Section 7.6 for more on scale parameters.

Figure 5.5 was produced by the following snippet.

par ( mfrow=c(2,2) )

shape <- c ( .5, 1, 2, 4 )

scale <- c ( .5, 1, 2, 4 )

for ( i in seq(along=scale) ) {
260 CHAPTER 5. SPECIAL DISTRIBUTIONS

a b

3.0

1.5
alpha = 0.5 alpha = 0.5
alpha = 1 alpha = 1
alpha = 2 alpha = 2
2.0

1.0
alpha = 4 alpha = 4
p(y)

p(y)
1.0

0.5
0.0

0.0
0 1 2 3 4 5 0 2 4 6 8

y y

c d
0.8

0.4

alpha = 0.5 alpha = 0.5

alpha = 1 alpha = 1
0.6

0.3

alpha = 2 alpha = 2
alpha = 4 alpha = 4
p(y)

p(y)
0.4

0.2
0.2

0.1
0.0

0.0

0 5 10 15 20 0 10 20 30 40

y y

Figure 5.5: Gamma densities — a: β = .5; b: β = 1; c: β = 2; d: β = 4;

5.5. GAMMA, EXPONENTIAL, CHI SQUARE 261

ymax <- scale[i]*max(shape) + 3*sqrt(max(shape))*scale[i]

y <- seq ( 0, ymax, length=100 )
den <- NULL
for ( sh in shape )
den <- cbind ( den, dgamma(y,shape=sh,scale=scale[i]) )
matplot ( y, den, type="l", main=letters[i], ylab="p(y)" )
legend ( ymax*.1, max(den[den!=Inf]),
paste ( "alpha = ", shape, sep="" ) )
}

Theorem 5.11. Let X ∼ Gam(α, β) and let Y = cX. Then Y ∼ Gam(α, cβ).

Proof. Use Theorem 1.1.

1
pX (x) = xα−1 e−x/β ,
Γ(α)β α

Since Y = cX, x = y/c and dx/dy = 1/c, so

1
pY (y) = (y/c)α−1 e−y/cβ
cΓ(α)β α
1
= (y)α−1 e−y/cβ
Γ(α)(cβ)α

which is the Gam(α, cβ) density. Also see Exercise 10.

The mean, mgf, and variance are recorded in the next several theorems.

Theorem 5.12. Let Y ∼ Gam(α, β) Then E[Y ] = αβ.

Proof.
Z ∞
1
E[Y ] = yα
y α−1 e−y/β dy
0 Γ(α)β
Z
Γ(α + 1)β ∞ 1
= y α e−y/β dy
Γ(α) 0 Γ(α + 1)β α+1
= αβ.

The last equality follows because (1) Γ(α + 1) = αΓ(α), and (2) the integrand
is a Gamma density so the integral is 1.

The last trick in the proof — recognizing an integrand as a density and

concluding that the integral is 1 — is very useful. Here it is again.

Theorem 5.13. Let Y ∼ Gam(α, β). Then the moment generating function is
MY (t) = (1 − tβ)−α for t < 1/β.
262 CHAPTER 5. SPECIAL DISTRIBUTIONS

Proof.
Z ∞
1
MY (t) = ety α
y α−1 e−y/β dy
0 Γ(α)β
β
( 1−tβ )α Z ∞ 1 1−tβ
= β
y α−1 e−y β dy
βα 0 Γ(α)( 1−tβ ) α

= (1 − tβ)−α

Theorem 5.14. Let Y ∼ Gam(α, β). Then

Var(Y ) = αβ 2

and √
SD(Y ) = αβ.
Proof. See Exercise 11.

The Exponential Distribution We often have to deal with situations such

as
• the lifetime of an item
• the time until a specified event happens
The most fundamental probability distribution for such situations is the
exponential distribution. Let Y be the time until the item dies or the event
occurs. If Y has an exponential distribution then for some λ > 0 the pdf of Y
is
pY (y) = λ−1 e−y/λ for y ≥ 0.
and we write Y ∼ Exp(λ). This density is pictured in Figure 5.6 (a repeat
of Figure 1.7) for four values of λ. The exponential distribution is the special
case of the Gamma distribution when α = 1. The mean and SD are given by
Theorems 5.12 and 5.14.
Each exponential density has its maximum at y = 0 and decreases monoton-
ically. The value of λ determines the value pY (0 | λ) and the rate of decrease.
Usually λ is unknown. Small values of y are evidence for large values of λ; large
values of y are evidence for small values of λ.

Example 5.2 (Radioactive Decay)

It is well known that some chemical elements are radioactive. Every atom of a
radioactive element will eventually decay into smaller components. E.g., uranium-
238 (by far the most abundant uranium isotope, 238 U) decays into thorium-234 and
an α particle while plutonium-239 (the isotope used in nuclear weapons, 239 Pu)
decays into uranium-235 (235 U) and an α particle. (See http://www.epa.gov/
radiation/radionuclides for more information.)
5.5. GAMMA, EXPONENTIAL, CHI SQUARE 263

10

lambda = 2
lambda = 1
lambda = 0.2
8

lambda = 0.1
6
p(x)

4
2
0

0.0 0.5 1.0 1.5 2.0

Figure 5.6: Exponential densities

264 CHAPTER 5. SPECIAL DISTRIBUTIONS

The time Y at which a particular atom decays is a random variable that has an
exponential distribution. Each radioactive isotope has its own distinctive value of
λ. A radioactive isotope is usually characterized by its median lifetime, or half-life,
instead of λ. The half-life is the value m which satisfies P[Y ≤ m] = P[Y ≥ m] =
0.5. The half-life m can be found by solving
Z m
λ−1 e−y/λ dy = 0.5.
0

The answer is m = λ log 2.

Uranium-238 has a half-life of 4.47 billion years. Thus its λ is about 6.45 billion.
Plutonium-239 has a half-life of 24,100 years. Thus its λ is about 35,000.

Exponential distributions have an interesting and unique memoryless prop-

erty. To demonstrate, we examine the Exp(λ) distribution as a model for T ,
the amount of time a computer Help line caller spends on hold. Suppose the
caller has already spent t minutes on hold; i.e., T ≥ t. Let S be the remaing
time on hold; i.e., S = T − t. What is the distribution of S given T > t? For
any number r > 0,

P[T ≥ t + r, T ≥ t]
P[S > r | T ≥ t] = P[T ≥ t + r | T ≥ t] =
P[T ≥ t]
P[T ≥ t + r] λ−1 e−(t+r)/λ
= = = e−r/λ .
P[T ≥ t] λ−1 e−t/λ

In other words, S has an Exp(λ) distribution (Why?) that does not depend on
the currently elapsed time t (Why?). This is a unique property of the Expo-
nential distribution; no other continuous distribution has it. Whether it makes
sense for the amount of time on hold is a question that could be verified by
looking at data. If it’s not sensible, then Exp(λ) is not an accurate model for
T.

Example 5.3
Some data here that don’t look exponential

The Poisson process There is a close relationship between Exponential,

Gamma, and Poisson distributions. For illustration consider a company’s cus-
tomer call center. Suppose that calls arrive according to a rate λ such that

1. in a time interval of length T , the number of calls is a random variable

with distribution Poi(λT ) and

2. if I1 and I2 are disjoint time intervals then the number of calls in I1 is

independent of the number of calls in I2 .

When calls arrive in this way we say the calls follow a Poisson process.
5.6. BETA 265

Suppose we start monitoring calls at time t0 . Let t1 be the time of the first
call after t0 and Y = t1 − t0 , the time until the first call. Y is a random variable.
What is its distribution? For any positive number y,
Pr[Y > y] = Pr[no calls in [t0 , t0 + y]] = e−λy
where the second equality follows by the Poisson assumption. But
Pr[Y > y] = e−λy ⇒ Pr[Y ≤ y] = 1 − e−λy ⇒ pY (y) = λe−λy ⇒ Y ∼ Exp(1/λ)
What about the time to the second call? Let t2 be the time of the second call
after t0 and Y2 = t2 − t0 . What is the distribution of Y2 ? For any y > 0,
Pr[Y2 > y] = Pr[fewer than 2 calls in [t0 , y]]
= Pr[0 calls in [t0 , y]] + Pr[1 call in [t0 , y]]
= e−λy + yλe−λy
and therefore
λ2
pY2 (y) = λe−λy − λe−λy + yλ2 e−λy = ye−λy
Γ(2)
so Y2 ∼ Gam(2, 1/λ).
In general, the time Yn until the n’th call has the Gam(n, 1/λ) distribution.

The Chi-squared Distribution The Gamma distribution with β = 2 and

α = p/2 where p is a positive integer is called the chi-squared distribution with
p degrees of freedom. We write Y ∼ χ2p .
P 2
Theorem 5.15. Let Y1 , . . . , Yn ∼ i.i.d. N(0, 1). Define X = Yi . Then
X ∼ χ2n .
Proof. This theorem will be proved in Section 5.7.

5.6 The Beta Distribution

For positive numbers α and β, the Beta(α, β) distribution is a distribution for
a random variable Y on the unit interval. The density is
Γ(α + β) α−1
pY (y) = y (1 − y)β−1 for y ∈ [0, 1]
Γ(α)Γ(β)
The parameters are (α, β). We write Y ∼ Be(α, β). The mean and variance are
given by Theorem 5.16.
Theorem 5.16. Let Y ∼ Be(α, β). Then
α
E[Y ] =
α+β
αβ
Var(Y ) =
(α + β)2 (α + β + 1)
266 CHAPTER 5. SPECIAL DISTRIBUTIONS

Proof. See Exercise 25.

Figure 5.7 shows some Beta densities. Each panel shows four densities hav-
ing the same mean. It is evident from the Figure and the definition that the
parameter α (β) controls whether the density rises or falls at the left (right). If
both α > 1 and β > 1 then p(y) is unimodal. The Be(1, 1) is the same as the
U(0, 1) distribution.
The Beta distribution arises as the distribution of order statistics from the
U(0, 1) distribution. Let x1 , . . . , xn ∼ i.i.d. U(0, 1). What is the distribution of
x(1) , the first order statistic? Our strategy is first to find the cdf of x(1) , then
differentiate to get the pdf.

FX(1) (x) = P[X(1) ≤ x]

= 1 − P[all Xi ’s are greater than x]
= 1 − (1 − x)n

Therefore,

d Γ(n + 1)
pX(1) (x) = FX(1) (x) = n(1 − x)n−1 = (1 − x)n−1
dx Γ(1)Γ(n)

which is the Be(1, n) density. For the distribution of the largest order statistic
see Exercise 26.

5.7 The Normal and Related Distributions

5.7.1 The Univariate Normal Distribution
The histograms in Figure 1.12 on page 29

• are approximately unimodal,

• are approximately symmetric,

• have different means, and

• have different standard deviations.

Data with these properties is ubiquitous in nature. Statisticians and other scien-
tists often have to model similar looking data. One common probability density
for modelling such data is the Normal density, also known as the Gaussian
density. The Normal density is also important because of the Central Limit
Theorem.
For some constants µ ∈ R and σ > 0, the Normal density is
1 1 x−µ 2
p(x | µ, σ) = √ e− 2 ( σ ) . (5.4)
2πσ
5.7. NORMAL 267

8 a

(a,b) = (0.3,1.2)
(a,b) = (1,4)
6

(a,b) = (3,12)
p(y)

(a,b) = (10,40)
2
0

0.0 0.2 0.4 0.6 0.8 1.0

b
4

(a,b) = (0.3,0.3)
3

(a,b) = (1,1)
(a,b) = (3,3)
p(y)

(a,b) = (10,10)
1
0

0.0 0.2 0.4 0.6 0.8 1.0

c
8 10

(a,b) = (0.3,0.03)
(a,b) = (1,0.11)
p(y)

(a,b) = (3,0.33)
(a,b) = (10,1.11)
4
2
0

0.0 0.2 0.4 0.6 0.8 1.0

Figure 5.7: Beta densities — a: Beta densities with mean .2; b: Beta densities
with mean .5; c: Beta densities with mean .9;
268 CHAPTER 5. SPECIAL DISTRIBUTIONS

Example 5.4 (Ocean temperatures, continued)

To see a Normal density in more detail, Figure 5.8 reproduces the top right histogram
from Figure 1.12 redrawn with the Normal density overlaid, for the values µ ≈ 8.08
and σ ≈ 0.94. The vertical axis is drawn on the density scale. There are 112
temperature measurements that go into this histogram; they were recorded between
1949 and 1997; their latitudes are all between 44 ◦ and 46◦ ; their longitudes are all
between −21◦ and −19◦ .
0.4
density

0.2
0.0

4 6 8 10 12

temperature

Figure 5.8: Water temperatures (◦ C) at 1000m depth, 44 − 46◦ N latitude and

19 − 21◦ W longitude. The dashed curve is the N(8.08,0.94) density.

Visually, the Normal density appears to fit the data well. Randomly choosing
one of the 112 historical temperature measurements, or making a new measurement
near 45◦ N and 20◦ W at a randomly chosen time are like drawing a random variable
t from the N(8.08,0.94) distribution.
Look at temperatures between 8.5◦ and 9.0◦ C. The N(8.08, 0.94) density says
the probability that a randomly drawn temperature t is between 8.5 ◦ and 9.0◦ C is
Z 9.0
1 1 t−8.08 2
P [t ∈ (8.5, 9.0]] = √ e− 2 ( 0.94 ) dt ≈ 0.16. (5.5)
8.5 2π 0.94

The integral in Equation 5.5 is best done on a computer, not by hand. In R it can
be done with pnorm(9.0,8.08,.94 ) - pnorm(8.5,8.08,.94 ). A fancier way
to do it is diff(pnorm(c(8.5,9),8.08,.94)).
• When x is a vector, pnorm(x,mean,sd) returns a vector of pnorm’s.
5.7. NORMAL 269

• When x is a vector, diff(x) returns the vector of differences x[2]-x[1],

x[3]-x[2], ..., x[n]-x[n-1].

In fact, 19 of the 112 temperatures fell into that bin, and 19/112 ≈ 0.17, so
the N(8.08, 0.94) density seems to fit very well.
However, the N(8.08, 0.94) density doesn’t fit as well for temperatures between
7.5◦ and 8.0◦ C.
Z 8.0
1 1 t−8.08 2
P [t ∈ (7.5, 8.0]] = √ e− 2 ( 0.94 ) dt ≈ 0.20.
7.5 2π 0.94
In fact, 15 of the 112 temperatures fell into that bin; and 15/112 ≈ 0.13. Even so,
the N(8.08,0.94) density fits the data set very well.

Theorem 5.17. Let Y ∼ N(µ, σ). Then

σ 2 t2
+µt
MY (t) = e 2 .

Proof.
Z
1 1 2
MY (t) = ety √ e− 2σ2 (y−µ) dy
2πσ
Z
1
e− 2σ2 (y −(2µ−2σ t)y+µ ) dy
1 2 2 2
= √
2πσ
Z
µ2 1 2 (µ−σ2 t)2
e− 2σ2 (y−(µ−σ t)) + 2σ2 dy
1 2
− 2σ
=e 2 √
2πσ
−2µσ2 t+σ4 t2
=e 2σ2

σ 2 t2
+µt
=e 2 .

Theorem 5.18. Let Y ∼ N(µ, σ). Then

E[Y ] = µ and Var(Y ) = σ 2 .

Proof. For the mean,

σ 2 t2

E[Y ] = MY0 (0) = (tσ 2 + µ)e 2 +µt 
 = µ.
t=0

For the variance,

σ 2 t2 σ 2 t2

E[Y 2 ] = MY00 (0) = σ 2 e 2 +µt
+ (tσ 2 + µ)2 e 2 +µt 
 = σ 2 + µ2 .
t=0

So,
Var(Y ) = E[Y 2 ] − E[Y ]2 = σ 2 .
270 CHAPTER 5. SPECIAL DISTRIBUTIONS

The N(0, 1) distribution is called the standard Normal distribution. As Theo-

rem 5.19 shows, all Normal distributions are just shifted, rescaled versions of the
standard Normal distribution. The mean is a location parameter; the standard
deviation is a scale parameter. See Section 7.6.

Theorem 5.19.

1. If X ∼ N(0, 1) and Y = σX + µ then Y ∼ N(µ, σ).

2. If Y ∼ N(µ, σ) and X = (Y − µ)/σ then X ∼ N(0, 1).

Proof. 1. Let X ∼ N(0, 1) and Y = σX + µ. By Theorem 4.8

σ 2 t2
MY (t) = eµt MX (σt) = eµt e 2

2. Let Y ∼ N(µ, σ) and X = (Y − µ)/σ. Then

σ2 (t/σ)2 t2
MX (t) = e−µt/σ MY (t/σ) = e−µt/σ e 2 +µt/σ
=e2

Section 5.5 introduced the χ2 distribution, noting that it is a special case

of the Gamma distribution. The χ2 distribution arises in practice as a sum of
squares of standard Normals. Here we restate Theorem 5.15, then prove it.
P
Theorem 5.20. Let Y1 , . . . , Yn ∼ i.i.d. N(0, 1). Define X = Yi2 . Then
X ∼ χ2n .

Proof. Start with the case n = 1.

Z
tY12 2 1 y2
MX (t) = E[e ] = ety √ e− 2 dy
2π
Z
1 − 1 (1−2t)y2
= √ e 2 dy
2π
Z √
1 − 2t − 1 (1−2t)y2
= (1 − 2t)−1/2 √ e 2 dy
2π
= (1 − 2t)−1/2

So X ∼ Gam(1/2, 2) = χ21 .
If n > 1 then by Corollary 4.10

MX (t) = MY12 +···+Yn2 (t) = (1 − 2t)−n/2

So X ∼ Gam(n/2, 2) = χ2n .
5.7. NORMAL 271

5.7.2 The Multivariate Normal Distribution

Let X~ be an n-dimensional random vector with mean µ ~ and covariance matrix

X
~ has a multivariate Normal distribution if its joint density
ΣX~ . We say that X
is
1 − 12 (~
x−µX t −1
~) Σ (~
x−µX
~)
pX~ (~x) = 1 e (5.6)
(2π)n/2 |Σ| 2

where |Σ| refers to the determinant of the matrix Σ. We write X ~ ∼ N(µ, Σ).
Comparison of Equations 5.4 (page 266) and 5.6 shows that the latter is a
generalization of the former. The multivariate version has the covariance matrix
Σ in place of the scalar variance σ 2 .
To become more familiar with the multivariate Normal distribution, we begin
with the case where the covariance matrix is diagonal:
 
σ12 0 0 ···
0
 σ22 0 · · ·

Σ= .. .. 
0
 0 . . 

.. ..
. . · · · σn2

In this case the joint density is

1 1 t
pX~ (~x) = 1 e− 2 (~x−µX~ ) Σ−1 (~
x−µX
~)

(2π)n/2 |Σ| 2
 n Y n   1 Pn (xi −µi )2
1 1 − i=1 σ2
= √ e 2 i
2π i=1
σ i
n  ” 
xi −µi 2
1
“
Y − 21
= √ e σi
,
i=1
2πσi

the product of n separate one dimensional Normal densities, one for each di-
mension. Therefore the Xi ’s are independent and Normally distributed, with
Xi ∼ N(µi , σi ). Also see Exercise 28.
When σ1 = · · · = σn = 1, then Σ is the n-dimensional identity matrix In .
When, in addition, µ1 = · · · = µn = 0, then X~ ∼ N(0, In ) and X
~ is said to have
the standard n-dimensional Normal distribution.
Note: for two arbitrary random variables X1 and X2 , X1 ⊥ X2 implies
Cov(X1 , X2 ) = 0; but Cov(X1 , X2 ) = 0 does not imply X1 ⊥ X2 . However, if
X1 and X2 are jointly Normally distributed then the implication is true. I.e. if
(X1 , X2 ) ∼ N(µ, Σ) and Cov(X1 , X2 ) = 0, then X1 ⊥ X2 . In fact, something
stronger is true, as recorded in the next theorem.

~ = (X1 , . . . , Xn ) ∼ N(µ, Σ) where Σ has the so-called

Theorem 5.21. Let X
272 CHAPTER 5. SPECIAL DISTRIBUTIONS

block-diagonal form
 
Σ11 012 ··· 01m
 021
 Σ22 ··· 02m 

Σ= . .. .. .. 
 .. . . . 
0m1 ··· 0mm−1 Σmm
Pm
where Σii is an ni × ni matrix, 0ij is an ni × nj matrix of 0’s and 1 ni = n.
Partition X ~ to conform with Σ and define Y ~i ’s: Y~1 = (X1 , . . . , Xn1 ), Y
~2 =
~m = (Xn1 +···+nm−1 +1 , . . . , Xnm ) and νi ’s: ν1 =
(Xn1 +1 , . . . , Xn1 +n2 ), . . . , Y
(µ1 , . . . , µn1 ), ν2 = (µn1 +1 , . . . , µn1 +n2 ), . . . , νm = (µn1 +···+nm−1 +1 , . . . , µnm ).
Then
~i ’s are independent of each other, and
1. The Y
~i ∼ N(νi , Σii )
2. Y
~ → (Y
Proof. The transformation X ~1 , . . . , Y
~m ) is just the identity transforma-
tion, so

pY~1 ,...,Y~m (~y1 , . . . , ~ym )

1 1 t
= pX~ (~y1 , . . . , ~ym ) = 1 e− 2 (~y−µ) Σ−1 (~
y−µ)
(2π)n/2 |Σ| 2
1 1
Pm t
= Qm 1 e− 2 i=1 (~
yi −νi ) Σ−1
ii (~
yi −νi )
(2π) n/2 |Σii | 2
i=1
m
Y 1 1 t
= 1 e− 2 (~yi −νi ) Σ−1
ii (~
yi −νi )

i=1 (2π)ni /2 |Σ ii | 2

To learn more about the multivariate Normal density, look at the curves on
which pX~ is constant; i.e., {~x : pX~ (~x) = c} for some constant c. The density
depends on the xi ’s through the quadratic form (~x − µ)t Σ−1 (~x − µ), so pX~
is constant where this quadratic
Pn form is constant. But when Σ is diagonal,
(~x − µ)t Σ−1 (~x − µ) = 2 2
1 (xi − µi ) /σi so pX ~ (~
x) = c is the equation of an
ellipsoid centered at µ and with eccentricities determined by the ratios σi /σj .
What does this density look like? It is easiest to answer that question in two
dimensions. Figure 5.9 shows three bivariate Normal densities. The left-hand
column shows contour plots of the bivariate densities; the right-hand column
shows samples from the joint distributions. In all cases, E[X1 ] = E[X2 ] = 0.
In the top row, σX1 = σX2 = 1; in the second row, σX1 = 1; σX2 = 2; in the
third row, σX1 = 1/2; σX2 = 2. The standard deviation is a scale parameter,
so changing the SD just changes the scale of the random variable. That’s what
gives the second and third rows more vertical spread than the first, and makes
the third row more horizontally squashed than the first and second.
5.7. NORMAL 273

4 (a) (b)

4
2

2
x2

x2
0

0
−4

−4
−4 −2 0 2 4 −4 −2 0 2 4

x1 x1

(c) (d)
4

4
2

2
x2

x2
0

0
−4

−4

−4 −2 0 2 4 −4 −2 0 2 4

x1 x1

(e) (f)
4

4
2

2
x2

x2
0

0
−4

−4

−4 −2 0 2 4 −4 −2 0 2 4

x1 x1

Figure 5.9: Bivariate Normal density. E[X1 ] = E[X2 ] = 0.

(a), (b): σX1 = σX2 = 1.
(c), (d): σX1 = 1; σX2 = 2.
(e), (f ): σX1 = 1/2; σX2 = 2.
(a), (c), (e): contours of the joint density.
(b), (d), (f ): samples from the joint density.
274 CHAPTER 5. SPECIAL DISTRIBUTIONS

Figure 5.9 was produced with the following R code.

par ( mfrow=c(3,2) ) # a 3 by 2 array of plots

x1 <- seq(-5,5,length=60)
x2 <- seq(-5,5,length=60)
den.1 <- dnorm ( x1, 0, 1 )
den.2 <- dnorm ( x2, 0, 1 )
den.jt <- den.1 %o% den.2
contour ( x1, x2, den.jt, xlim=c(-5,5), ylim=c(-5,5), main="(a)",
xlab=expression(x[1]), ylab=expression(x[2]) )

samp.1 <- rnorm ( 300, 0, 1 )

samp.2 <- rnorm ( 300, 0, 1 )
plot ( samp.1, samp.2, xlim=c(-5,5), ylim=c(-5,5), main="(b)",
xlab=expression(x[1]), ylab=expression(x[2]), pch="." )

den.2 <- dnorm ( x2, 0, 2 )

den.jt <- den.1 %o% den.2
contour ( x1, x2, den.jt, xlim=c(-5,5), ylim=c(-5,5), main="(c)",
xlab=expression(x[1]), ylab=expression(x[2]), )

samp.2 <- rnorm ( 300, 0, 2 )

plot ( samp.1, samp.2, xlim=c(-5,5), ylim=c(-5,5), main="(d)",
xlab=expression(x[1]), ylab=expression(x[2]), pch="." )

den.1 <- dnorm ( x1, 0, .5 )

den.jt <- den.1 %o% den.2
contour ( x1, x2, den.jt, xlim=c(-5,5), ylim=c(-5,5), main="(e)",
xlab=expression(x[1]), ylab=expression(x[2]) )

samp.1 <- rnorm ( 300, 0, .5 )

plot ( samp.1, samp.2, xlim=c(-5,5), ylim=c(-5,5), main="(f)",
xlab=expression(x[1]), ylab=expression(x[2]), pch="." )

• The code makes heavy use of the fact that X1 and X2 are independent for
(a) calculating the joint density and (b) drawing random samples.
• den.1 %o% den.2 yields the outer product of den.1 and den.2. It is a
matrix whose ij’th entry is den.1[i] * den.2[j].

Now let’s see what happens when Σ is not diagonal. Let Y ∼ N(µY~ , ΣY~ ), so

1 1 t
pY~ (~y ) = 1 e− 2 (~y−µY~ ) Σ−1
~
Y
(~
y −µY
~)
,
(2π)n/2 |Σ ~
Y | 2
5.7. NORMAL 275

and let X ~ ∼ N(0, In ). X ~ is just a collection of independent N(0, 1) random

variables. Its curves of constant density are just (n − 1)-dimensional spheres
centered at the origin. Define Z ~ = Σ1/2 X ~ + µ. We will show that p ~ =
Z
~ ~
pY~ , therefore that Z and Y have the same distribution, and therefore that
any multivariate Normal random vector has the same distribution as a linear
transformation of a standard multivariate Normal random vector. To show
pZ~ = pY~ we apply Theorem 4.4. The Jacobian of the transformation from X ~ to
~ is |Σ| , the square root of the determinant of Σ. Therefore,
Z 1/2

 
pZ~ (~y ) = pX~ Σ−1/2 (~y − µ) |Σ|−1/2
1 y −µ))t (Σ−1/2 (~
−1/2(Σ−1/2 (~ y −µ))
=√ ne |Σ|−1/2
2π
1 1 t −1
= e− 2 (~y−µ) Σ (~y−µ)
(2π)n/2 |Σ|1/2
= pY~ (~y )

The preceding result says that any multivariate Normal random variable, Y ~ in
our notation above, has the same distribution as a linear transformation of a
standard Normal random variable.
To see what multivariate Normal densities look like it is easiest to look at 2
dimensions. Figure 5.10 shows three bivariate Normal densities. The left-hand
column shows contour plots of the bivariate densities; the right-hand column
shows samples from the joint distributions. In all cases, E[X1 ] = E[X2 ] = 0 and
σ1 = σ2 = 1. In the top row, σ1,2 = 0; in the second row, σ1,2 = .5; in the third
row, σ1,2 = −.8.

Figure 5.10 was produced with the following R code.

par ( mfrow=c(3,2) ) # a 3 by 2 array of plots
npts <- 60
sampsize <- 300

x1 <- seq(-5,5,length=npts)
x2 <- seq(-5,5,length=npts)

Sigma <- array ( NA, c(3,2,2) )

Sigma[1,,] <- c(1,0,0,1)
Sigma[2,,] <- c(1,.5,.5,1)
Sigma[3,,] <- c(1,-.8,-.8,1)

den.jt <- matrix ( NA, npts, npts )

for ( i in 1:3 )
Sig <- Sigma[i,,]
Siginv <- solve(Sig) # matrix inverse
276 CHAPTER 5. SPECIAL DISTRIBUTIONS

a b

4
2

2
x2

x2
0

0
−4

−4
−4 −2 0 2 4 −4 −2 0 2 4

x1 x1

c d
4

4
2

2
x2

x2
0

0
−4

−4

−4 −2 0 2 4 −4 −2 0 2 4

x1 x1

e f
4

4
2

2
x2

x2
0

0
−4

−4

−4 −2 0 2 4 −4 −2 0 2 4

x1 x1

Figure 5.10: Bivariate Normal density. E[X1 ] = E[X2 ] = 0; σ1 = σ2 = 1.

(a), (b): σ1,2 = 0.
(c), (d): σ1,2 = .5.
(e), (f ): σ1,2 = −.8.
(a), (c), (e): contours of the joint density.
(b), (d), (f ): samples from the joint density.
5.7. NORMAL 277

for ( j in 1:npts )
for ( k in 1:npts )
x <- c ( x1[j], x2[k] )
den.jt[j,k] <- ( 1 / sqrt(2*pi*det(Sig)) ) *
exp ( -.5 * t(x) %*% Siginv %*% x )

contour ( x1, x2, den.jt, xlim=c(-5,5), ylim=c(-5,5),

drawlabels=F,
xlab=expression(x[1]),
ylab=expression(x[2]), main=letters[2*i-1] )

samp <- matrix ( rnorm ( 2*sampsize ), 2, sampsize )

samp <- Sig %*% samp
plot ( samp[1,], samp[2,], pch=".",
xlim=c(-5,5), ylim=c(-5,5),
xlab=expression(x[1]),
ylab=expression(x[2]), main=letters[2*i] )

We conclude this section with some theorems about Normal random vari-
ables that will prove useful later.
Theorem 5.22. Let X ~ ∼ N(µ, Σ) be an n-dimensional Normal random vari-
able; let A be a full rank n by n matrix; and let Y = AX. Then Y ∼
N(Aµ, AΣAt ).
Proof. By Theorem 4.4 (pg. 231),

pY~ (~y ) = pX~ (A−1 ~y)|A−1 |

1 − 12 (A−1 ~y −µX ~) Σ
t −1
(A−1 ~y −µX~)
= 1 e
n/2
(2π) |A||Σ| 2
1 − 12 (A−1 (~y −AµX t −1
~ )) Σ (A−1 (~
y −AµX ~ ))
= 1 e
n/2
(2π) |A||Σ| 2

1 − 12 (~
y −AµX t
~ ) (A
−1 t −1 −1
) Σ A (~ y −AµX ~)
= 1 e
n/2
(2π) |A||Σ| 2

1 − 21 (~
y −AµX t
~ ) (AΣA )
t −1
(~
y −AµX ~)
= 1 e
n/2
(2π) |AΣA | t 2

which we recognize as the N(Aµ, AΣAt ) density.

Corollary 5.23. Let X ~ ∼ N(µ, Σ) be an n-dimensional Normal random vari-

able; let A be a full rank n by n matrix; let b be a vector of length n; and let
Y = AX + b. Then Y ∼ N(Aµ + b, AΣAt ).
Proof. See Exercise 29.
278 CHAPTER 5. SPECIAL DISTRIBUTIONS

Pn
Corollary 5.24. Let X1 , . . . , Xn ∼ N(µ, σ). Define S 2 ≡ i=1 (Xi − X̄)2 .
Then X̄ ⊥ S 2 .
~ = (Y1 , . . . , Yn )t by
Proof. Define the random vector Y

Y1 = X1 − X̄
Y2 = X2 − X̄
..
.
Yn−1 = Xn−1 − X̄
Yn = X̄

The proof follows these steps.

1. S 2 is a function only of (Y1 , . . . , Yn−1 )t ; i.e. not a function of Yn .

2. (Y1 , . . . , Yn−1 )t ⊥ Yn .

3. Therefore S 2 ⊥ Yn .
Pn Pn−1
1. i=1 (Xi − X̄) = 0. Therefore, (Xn − X̄) = − i=1 (Xi − X̄). And
therefore
n−1 n−1
!2 n−1 n−1
!2
X X X X
2 2
S = (Xi − X̄) + (Xi − X̄) = Yi2 + Yi
i=1 i=1 i=1 i=1

is a function of (Y1 , . . . , Yn−1 )t .

2.  
1 − n1 − n1 − n1 ··· − n1
 −1 1 − n1 − n1 ··· − n1 
 n 
~ = . .. .. .. ..  X~ ≡ AX
~
Y  .. . . . . 
 1 
 −
n − n1 ··· 1 − n1 − n
1
1 1 1 1
n n n ··· n

where the matrix A is defined by the preceding equation, so

Y~ ∼ N(Aµ, σ 2 AAt ). The first n − 1 rows of A are each orthogonal to the
last row. Therefore  
Σ ~0
AAt = ~11
0t 1/n

where Σ11 has dimension (n − 1) × (n − 1) and ~0 is the (n − 1)-dimensional

vector of 0’s. Thus, by Theorem 5.21, (Y1 , . . . , Yn−1 )t ⊥ Yn .

3. Follows immediately from 1 and 2.

5.8. T AND F 279

5.8 The t and F Distributions

5.8.1 The t distribution
The t distribution arises when making inference about the mean of a Normal
distribution.
Let X1 , . . . , Xn ∼ i.i.d. N(µ, σ) where both µ and σ are unknown, and sup-
pose our goal is to estimate √ µ. µ̂ = X̄ is a sensible estimator. Its sampling
distribution is X̄ ∼ N(µ, σ/ n) or, equivalently,

X̄ − µ
√ ∼ N(0, 1)
σ/ n

We would like to use this equation to tell us how accurately

√ we can estimate µ.
Apparently we can estimate µ to within about ±2σ/ n most of the time. But
that’s not an immediately useful statement because we don’t know σ. So we
P
1/2
estimate σ by σ̂ = n−1 (Xi − X̄)2 and say

X̄ − µ
√ ∼ N(0, 1),
σ̂/ n
√
approximately. This section derives the exact distribution of (X̄ − µ)/(σ̂/ n)
and assesses how good the Normal approximation is. We already know from
Corollary
P 5.24 that X̄ ⊥ σ̂. Theorem 5.26 gives the distribution of S 2 = nσ̂ 2 =
(Xi − X̄)2 . First we need a lemma.

Lemma 5.25. Let V = V1 + V2 and W = W1 + W2 where V1 ⊥ V2 and

W1 ⊥ W2 . If V and W have the same distribution, and if V1 and W1 have the
same distribution, then V2 and W2 have the same distribution.

Proof. Using moment generating functions,

MV2 (t) = MV (t)/MV1 (t)

= MW (t)/MW1 (t)
= MW2 (t)

Pn
Theorem 5.26. Let X1 , . . . , Xn ∼ i.i.d. N(µ, σ). Define S 2 = i=1 (Xi − X̄)
2
.
Then
S2
∼ χ2n−1 .
σ2
Proof. Let
n  2
X Xi − µ
V = .
i=1
σ
280 CHAPTER 5. SPECIAL DISTRIBUTIONS

Then V ∼ χ2n and

n  2
X (Xi − X̄) + (X̄ − µ)
V =
i=1
σ
X Xi − X̄ 2
n   
X̄ − µ
2 Xn
= +n + 2(X̄ − µ) (Xi − X̄)
i=1
σ σ i=1
n  2  2
X Xi − X̄ X̄ − µ
= + √
i=1
σ σ/ n
S2
≡ + V2
σ2
where S 2 /σ 2 ⊥ V2 and V2 ∼ N(0, 1). But also,
n−1
X 2  2
Xi − µ Xn − µ
V = + ≡ W1 + W2
i=1
σ σ

where W1 ⊥ W2 , W1 ∼ χ2n−1 and W2 ∼ N(0, 1). Now the conclusion follows by

Lemma 5.25.
Define r   √
n−1 X̄ − µ n(X̄ − µ)/σ
T ≡ √ =p .
n σ̂/ n S 2 /(n − 1)σ 2
Thenp by Corollary 5.24 and Theorem 5.26, T has the distribution of
U/ V /(n − 1) where U ∼ N(0, 1), V ∼ χ2n−1 , and U ⊥ V . This distribution
is called the t distribution with n − 1 degrees of freedom. We write T ∼ tn−1 .
Theorem 5.27 derives its density.
p
Theorem 5.27. Let U ∼ N(0, 1), V ∼ χ2p , and U ⊥ V . Then T ≡ U/ V /p ∼
χ2p has density
p  − p+1
Γ( p+1
2 )p
2
2

− p+1 Γ( p+1
2 ) t2 2

pT (t) = p √ t + p 2
= p √ 1+ .
Γ( 2 ) π Γ( 2 ) pπ p
Proof. Define
U
T =p and Y =V
V /p
We make the transformation (U, V ) → (T, Y ), find the joint density of (T, Y ),
and then the marginal density of T . The inverse transformation is
1
XY 2
U= √ and V =Y
p
The Jacobian is  dU   12 1

 dU   Y√ T Y√− 2  Y 21
dT dY = p 
 dV dV
2 p = √

dT dY
 
0 1  p
5.8. T AND F 281

The joint density of (U, V ) is

1 u2 1 p v
pU,V (u, v) = √ e− 2 p p v
2 −1 e− 2 .
2π Γ( 2 )2 2

Therefore the joint density of (T, Y ) is

1
1 t2 y 1 p y y2
pT,Y (t, y) = √ e− p p p y
2 −1 e− 2 √
2π Γ( 2 )2 2 p

and the marginal density of T is

Z
pT (t) = pT,Y (t, y) dy
Z ∞
1 p+1 y x2
=√ p√ y 2 −1 e− 2 ( p +1)
dy
2πΓ( p2 )2 2 p 0
  p+1
1 p+1 2p 2
=√ p+1 √ Γ( )
πΓ( p2 )2 2 p 2 t2 + p
Z ∞
1 p+1 − y
×   p+1 y 2 −1 e 2p/(t2 +p) dy
0 2
Γ( p+1 2p
2 ) t2 +p

Γ( p+1
2 )p
p/2
− p+1
= p √ t2 + p 2

Γ( 2 ) π
 − p+1
Γ( p+1
2 ) t2 2

= p √ 1+ .
Γ( 2 ) pπ p

Figure 5.8.1 shows the t density for 1, 4, 16, and 64 degrees of freedom, and
the N(0, 1) density. The two points to note are

1. The t densities are unimodal and symmetric about 0, but have less mass
in the middle and more mass in the tails than the N(0, 1) density.

2. In the limit, as p → ∞, the tp density appears to approach the N(0, 1)

density. (The appearance is correct. See Exercise 30.)

Figure 5.8.1 was produced with the following snippet.

x <- seq ( -5, 5, length=100 )
dens <- cbind ( dt(x,1), dt(x,4), dt(x,16), dt(x,64),
dnorm(x) )
matplot ( x, dens, type="l", ylab="density", xlab="t",
lty=c(2:5,1), col=1 )
282 CHAPTER 5. SPECIAL DISTRIBUTIONS

0.4

df = 1
df = 4
0.3

df = 16
df = 64
density

0.2

Normal
0.1
0.0

−4 −2 0 2 4

Figure 5.11: t densities for four degrees of freedom and the N(0, 1) density
5.9. EXERCISES 283

legend ( x=-5, y=.4, lty=c(2:5,1),

legend=c ( paste("df = ", c(1,4,16,64)),
"Normal" ) )

√
At the beginning of Section 5.8.1 we said the quantity n(X̄ − µ)/σ̂ had
a N(0, 1) distribution,
√ approximately. Theorem 5.27 derives the density of the
related quantity n − 1(X̄ − µ)/σ̂ which has a tn−1 distribution, exactly. Fig-
ure 5.8.1 shows how similar those distributions are. The t distribution has
slightly more spread than the N(0, 1) distribution, reflecting the fact that σ has
to be estimated. But when n is large, i.e. when σ is well estimated, then the
two distributions are nearly identical.
If T ∼ tp , then

 − p+1
Γ( p+1
Z ∞
2 ) t2 2
E[T ] = t p √ 1+ dt (5.7)
−∞ Γ( 2 ) pπ p

In the limit as t → ∞, the integrand behaves like t−p ; hence 5.7 is integrable
if and only if p > 1. Thus the t1 distribution, also known as the Cauchy
distribution, has no mean. When p > 1, E[T ] = 0, by symmetry. By a similar
argument, the tp distribution has a variance if and only if p > 2. When p > 2,
then Var(T ) = p/(p − 2). In general, T has a k-th moment (E[T k ] < ∞) if and
only if p > k.

5.8.2 The F distribution

5.9 Exercises
1. Prove Theorem 5.4 by moment generating functions.

2. Refer to Theorem 5.8.

(a) What was the point of the next to last step?

(b) Justify the last step.

3. Assume that all players on a basketball team are 70% free throw shooters
and that free throws are independent of each other.

(a) The team takes 40 free throws in a game. Write down a formula for
the probability that they make exactly 37 of them. You do not need
to evaluate the formula.
(b) The team takes 20 free throws the next game. Write down a formula
for the probability that they make exactly 9 of them.
284 CHAPTER 5. SPECIAL DISTRIBUTIONS

(c) Write down a formula for the probability that the team makes exactly
37 free throws in the first game and exactly 9 in the second game.
That is, write a formula for the probability that they accomplish both
feats.
4. Write down the distribution you would use to model each of the following
random variables. Be as specific as you can. I.e., instead of answering
“Poisson distribution”, answer “Poi(3)” or instead of answering “Bino-
mial”, answer “Bin(n, p) where n = 13 but p is unknown.”
(a) The temperature measured at a randomly selected point on the sur-
face of Mars.
(b) The number of car accidents in January at the corner of Broad Street
and Main Street.
(c) Out of 20 people in a post office, the number who, when exposed to
anthrax spores, actually develop anthrax.
(d) Out of 10,000 people given a smallpox vaccine, the number who de-
velop smallpox.
(e) The amount of Mercury in a fish caught in Lake Ontario.
5. A student types dpois(3,1.5) into R. R responds with 0.1255107.
(a) Write down in words what the student just calculated.
(b) Write down a mathematical formula for what the student just calcu-
lated.
6. Name the distribution. Your answers should be of the form Poi(λ) or
N(3, 22), etc. Use numbers when parameters are known, symbols when
they’re not.
You spend the evening at the roulette table in a casino. You bet on red
100 times. Each time the chance of winning is 18/38. If you win, you win
$1; if you lose, you lose $1. The average amount of time between bets is
90 seconds; the standard deviation is 5 seconds.
(a) the number of times you win
(b) the number of times you lose
(c) the number of bets until your third win
(d) the number of bets until your thirtieth loss
(e) the amount of time to play your first 40 bets
(f) the additional amount of time to play your next 60 bets
(g) the total amount of time to play your 100 bets
(h) your net profit at the end of the evening
(i) the amount of time until a stranger wearing a red carnation sits down
next to you
5.9. EXERCISES 285

(j) the number of times you are accidentally jostled by the person stand-
ing behind you
7. In a 1991 article (See Utts [1991] and discussants.) Jessica Utts reviews
some of the history of probability and statistics in ESP research. This
question concerns a particular series of autoganzfeld experiments in which
a sender looking at a picture tries to convey that picture telepathically to
a receiver. Utts explains:
“. . . ‘autoganzfeld’ experiments require four participants. The
first is the Receiver (R), who attempts to identify the target
material being observed by the Sender (S). The Experimenter
(E) prepares R for the task, elicits the response from R and
supervises R’s judging of the response against the four potential
targets. (Judging is double blind; E does not know which is the
correct target.) The fourth participant is the lab assistant (LA)
whose only task is to instruct the computer to randomly select
the target. No one involved in the experiment knows the identity
of the target.
“Both R and S are sequestered in sound-isolated, electrically
sheilded rooms. R is prepared as in earlier ganzfeld studies, with
white noise and a field of red light. In a nonadjacent room, S
watches the target material on a television and can hear R’s tar-
get description (‘mentation’) as it is being given. The mentation
is also tape recorded.
“The judging process takes place immediately after the 30-
minute sending period. On a TV monitor in the isolated room,
R views the four choices from the target pack that contains
the actual target. R is asked to rate each one according to
how closely it matches the ganzfeld mentation. The ratings are
converted to ranks and, if the correct target is ranked first, a
direct hit is scored. The entire process is automatically recorded
by the computer. The computer then displays the correct choice
to R as feedback.”
In the series of autoganzfeld experiments analyzed by Utts, there were a
total of 355 trials. Let X be the number of direct hits.
(a) What are the possible values of X?
(b) Assuming there is no ESP, and no cheating, what is the distribution
of X?
(c) Plot the pmf of the distribution in part (b).
(d) Find E[X] and SD(X).
(e) Add a Normal approximation to the plot in part (c).
(f) Judging from the plot in part (c), approximately what values of X
are consistent with the “no ESP, no cheating” hypothesis?
286 CHAPTER 5. SPECIAL DISTRIBUTIONS

(g) In fact, the total number of hits was x = 122. What do you conclude?
8. A golfer plays the same golf course daily for a period of many years. You
may assume that he does not get better or worse, that all holes are equally
difficult and that the results on one hole do not influence the results on
any other hole. On any one hole, he has probabilities .05, .5, and .45
of being under par, exactly par, and over par, respectively. Write down
what distribution best models each of the following random variables. Be
as specific as you can. I.e., instead of answering ”Poisson distribution”
answer ”Poi(3)” or ”Poi(λ) where λ is unknown.” For some parts the
correct answer might be ”I don’t know.”
(a) X, the number of holes over par on 17 September, 2002
(b) W, the number of holes over par in September, 2002
(c) Y, the number of rounds over par in September, 2002
(d) Z, the number of times he is hit by lightning in this decade
(e) H, the number of holes-in-one this decade
(f) T, the time, in years, until his next hole-in-one
9. During a CAT scan, a source (your brain) emits photons which are counted
by a detector (the machine). The detector is mounted at the end of a long
tube, so only photons that head straight down the tube are detected. In
other words, though the source emits photons in all directions, the only
ones detected are those that are emitted within the small range of angles
that lead down the tube to the detector.
Let X be the number of photons emitted by the source in 5 seconds.
Suppose the detector captures only 1% of the photons emitted by the
source. Let Y be the number of photons captured by the detector in those
same 5 seconds.
(a) What is a good model for the distribution of X?
(b) What is the conditional distribution of Y given X?
(c) What is the marginal distribution of Y?
Try to answer these questions from first principles, without doing any
calculations.
10. Prove Theorem 5.11 using moment generating functions.
11. (a) Prove Theorem 5.14 by finding E[Y 2 ] using the trick that was used
to prove Theorem 5.12.
(b) Prove Theorem 5.14 by finding E[Y 2 ] using moment generating func-
tions.
12. Case Study 4.2.3 in Larsen and Marx [add reference] claims that the
number of fumbles per team in a football game is well modelled by a
Poisson(2.55) distribution. For this quiz, assume that claim is correct.
5.9. EXERCISES 287

(a) What is the expected number of fumbles per team in a football game?
(b) What is the expected total number of fumbles by both teams?
(c) What is a good model for the total number of fumbles by both teams?
(d) In a game played in 2002, Duke fumbled 3 times and Navy fumbled 4
times. Write a formula (Don’t evaluate it.) for the probability that
Duke will fumble exactly 3 times in next week’s game.
(e) Write a formula (Don’t evaluate it.) for the probability that Duke
will fumble exactly three times given that they fumble at least once.
13. Clemson University, trying to maintain its superiority over Duke in ACC
football, recently added a new practice field by reclaiming a few acres of
swampland surrounding the campus. However, the coaches and players
refused to practice there in the evenings because of the overwhelming
number of mosquitos.
To solve the problem the Athletic Department installed 10 bug zappers
around the field. Each bug zapper, each hour, zaps a random number of
mosquitos that has a Poisson(25) distribution.
(a) What is the exact distribution of the number of mosquitos zapped by
10 zappers in an hour? What are its expected value and variance?
(b) What is a good approximation to the distribution of the number
of mosquitos zapped by 10 zappers during the course of a 4 hour
practice?
(c) Starting from your answer to the previous part, find a random vari-
able relevant to this problem that has approximately a N(0,1) distri-
bution.
14. Bob is a high school senior applying to Duke and wants something that
will make his application stand out from all the others. He figures his best
chance to impress the admissions office is to enter the Guinness Book of
World Records for the longest amount of time spent continuously brushing
one’s teeth with an electric toothbrush. (Time out for changing batteries
is permissible.) Batteries for Bob’s toothbrush last an average of 100
minutes each, with a variance of 100. To prepare for his assault on the
world record, Bob lays in a supply of 100 batteries.
The television cameras arrive along with representatives of the Guinness
company and the American Dental Association and Bob begins the quest
that he hopes will be the defining moment of his young life. Unfortunately
for Bob his quest ends in humiliation as his batteries run out before he
can reach the record which currently stands at 10,200 minutes.
Justice is well served however because, although Bob did take AP Statistics
in high school, he was not a very good student. Had he been a good
statistics student he would have calculated in advance the chance that his
batteries would run out in less than 10,200 minutes.
Calculate, approximately, that chance for Bob.
288 CHAPTER 5. SPECIAL DISTRIBUTIONS

15. A new article [add citation, Fall ’02 or spring ’03?] on statistical
fraud detection, when talking about records in a database, says:
”One of the difficulties with fraud detection is that typically there are
many legitimate records for each fraudulent one. A detection method
which correctly identifies 99% of the legitimate records as legitimate and
99% of the fraudulent records as fraudulent might be regarded as a highly
effective system. However, if only 1 in 1000 records is fraudulent, then,
on average, in every 100 that the system flags as fraudulent, only about 9
will in fact be so.”
QUESTION: Can you justify the ”about 9”?
16. [credit to FPP here, or change the question.] In 1988 men averaged
around 500 on the math SAT, the SD was around 100 and the histogram
followed the normal curve.
(a) Estimate the percentage of men getting over 600 on this test in 1988.
(b) One of the men who took the test in 1988 will be picked at random,
and you have to guess his test score. You will be given a dollar if you
guess it right to within 50 points.
i. What should you guess?
ii. What is your chance of winning?
17. Multiple choice.
(a) X ∼ Poi(λ). Pr[X ≤ 7] =
P7 −λ x
i. x=−∞ e λ /x!
P7 −λ x
ii. e λ /x!
Px=0
7 −λ x
iii. λ=0 e λ /x!
(b) X and Y are distributed uniformly on the unit square.
Pr[X ≤ .5|Y ≤ .25] =
i. .5
ii. .25
iii. can’t tell from the information given.
(c) X ∼ Normal(µ, σ 2 ). Pr[X > µ + σ]
i. is more than .5
ii. is less than .5
iii. can’t tell from the information given.
(d) X1 , . . . , X100 ∼ N(0, 1). X̄ ≡ (X1 + · · · + X100 )/100. Y ≡ (X1 + · · · +
X100 ). Calculate
i. Pr[−.2 ≤ X̄ ≤ .2]
ii. Pr[−.2 ≤ Xi ≤ .2]
iii. Pr[−.2 ≤ Y ≤ .2]
5.9. EXERCISES 289

Pr[−2 ≤ X̄ ≤ 2]
iv.
Pr[−2 ≤ Xi ≤ 2]
v.
Pr[−2 ≤ Y ≤ 2]
vi.
Pr[−20 ≤ X̄ ≤ 20]
vii.
Pr[−20 ≤ Xi ≤ 20]
viii.
Pr[−20 ≤ Y ≤ 20]
ix.
P100
(e) X ∼ Bin(100, θ). θ=0 f (x|θ) =
i. 1
ii. the question doesn’t make sense
iii. can’t tell from the information given.
(f) X and Y have joint density f (x, y) on the unit square. f (x) =
R1
i. 0 f (x, y) dx
R1
ii. 0 f (x, y) dy
Rx
iii. 0 f (x, y) dy
(g) X1 , . . . , Xn ∼ Gamma(r, λ) and are mutually independent.
f (x1 , . . . , xn ) =
Q P
i. [λr /(r − 1)!]( xi )r−1 e−λ xi
Q Q
ii. [λnr /((r − 1)!)n ]( xi )r−1 e−λ xi
Q P
iii. [λnr /((r − 1)!)n ]( xi )r−1 e−λ xi
18. In Figure 5.2, the plots look increasingly Normal as we go down each
column. Why? Hint: a well-known theorem is involved.
19. Prove Theorem 5.7.
20. Rongelap Island, Poisson distribution
21. seed rain, Poisson distribution
22. (a) Let Y ∼ U(1, n) where the parameter n is an unknown positive in-
teger. Suppose we observe Y = 6. Find the m.l.e. n̂. Hint: Equa-
tion 5.2 defines the pmf for y ∈ {1, 2, . . . , n}. What is p(y) when
y 6∈ {1, 2, . . . , n}?
(b) In World War II, when German tanks came from the factory they
had serial numbers labelled consecutively from 1. I.e., the numbers
were 1, 2, . . . . The Allies wanted to estimate T , the total number of
German tanks and had, as data, the serial numbers of the tanks they
had captured. Assuming that tanks were captured independently of
each other and that all tanks were equally likely to be captured find
the m.l.e. N̂ .
23. Let Y be a continuous random variable, Y ∼ U(a, b).
(a) Find E[Y ].
290 CHAPTER 5. SPECIAL DISTRIBUTIONS

(b) Find Var(Y ).

(c) Find MY (t).
24. (a) Is there a discrete distribution that is uniform on the positive inte-
gers? Why or why not? If there is such a distribution then we might
call it U(1, ∞).
(b) Is there a continuous distribution that is uniform on the real line?
Why or why not? If there is, then we might call it U(−∞, ∞).
25. Prove Theorem 5.16. Hint: Use the method of Theorem 5.12.
26. Let x1 , . . . , xn ∼ i.i.d. U(0, 1). Find the distribution of x(n) , the largest
order statistic.
27. In the R code to create Figure 2.18, why was it okay to use dgamma(...)
even though the data were Poisson?
28. Page 271 shows that the n-dimensional Normal density with a diagonal
covariance matrix is the product of n separate univariate Normal densities.
In this problem you are to work in the opposite direction. Let X1 , . . . ,
Xn be independent Normally distributed random variables with means
µ1 , . . . , µn and SD’s σ1 , . . . , σn .
(a) Write down the density of each Xi .
(b) Write down the joint density of X1 , . . . , Xn .
(c) Show that the joint density can be written in the form of Equa-
tion 5.6.
(d) Derive the mean vector and covariance matrix of X1 , . . . , Xn .
29. Prove Corollary 5.23.
30. Show, for every x ∈ R,
1 x2
lim ptp (x) = √ e− 2
p→∞ 2π
where ptp (x) is the t density with p degrees of freedom, evaluated at x.
Hint: use Sterling’s formula. (This problem is Exercise 5.18(c) in Statis-
tical Inference, 2nd ed. by Casella and Berger.)
 Chapter 6

More Models

6.1 Hierarchical Models

It is often useful to think of populations as having subpopulations, and those as
having subsubpopulations, and so on. One example comes from [cite Worsley et
al] who describe fMRI (functional magnetic resonance imaging) experiments. A
subject is placed in an MRI machine and subjected to several stimuli while the
machine measures the amount of oxygen flowing to various parts of the brain.
Different stimuli affect different parts of the brain, allowing scientists to build
up a picture of how the brain works. Let the generic parameter θ be the change
in blood flow to a particular region of the brain under a particular stimulus. θ
is called an effect. As citation explain, θ may vary from subject to subject, from
session to session even for the same patient, and from run to run even within
the same session. To describe the situation fully we need three subscripts, so
let θijk be the effect in subject i, session j, run k. For a single subject i and
session j there will be an overall average effect; call it µij . The set {θijk }k
will fall around µij with a bit of variation for each run k. Assuming Normal
distributions, we would write

{θijk }k | µij , σk ∼ i.i.d. N(µij , σk )

Likewise, for a single subject i there will be an overall average effect; call it
µi . The set {µij }j will fall around µi with a bit of variation for each session
j. Further, each µi is associated with a different subject so they are like draws
from a population with a mean and standard deviation, say µ and σi . Thus the
whole model can be written

{θijk }k | µij , σk ∼ i.i.d. N(µij , σk )

{µij }j | µi , σj ∼ i.i.d. N(µi , σj )
{µi }i | µ, σi ∼ i.i.d. N(µ, σi )

Figure 6.1 is a graphical representation of this model.

291
292 CHAPTER 6. MORE MODELS

pretty picture here

Figure 6.1: Graphical representation of hierarchical model for fMRI

More examples: metaanalysis, Jackie Mohan’s germination records, Chan-

tal’s arabadopsis, CO2 uptake from R and Pinheiro and Bates, FACE growth
rates by tree—ring—treatment
Is a sample from one population or several? Mixtures of Normals. Extra
variation in Binomials, Poisson, etc. Hierarchical and random effects models.
Discrete populations: medical trials, different species, locations, subjects, treat-
ments.

6.2 Markov Chains

time series, dlm’s, population matrix models

Example 6.1 (FACE)

Richter’s throughfall data

Example 6.2 (hydrology)

Jagdish’s data. Is this example too complicated?

6.3 Contingency Tables

loglinear models? Simpson’s paradox? Census tables as examples?

6.4 Survival analysis

6.5 The Poisson process

Example 6.3
Earthquakes

Example 6.4
Neurons firing

6.6 Change point models

Example 6.5
Everglades
6.7. EVALUATING AND ENHANCING MODELS 293

6.7 Evaluating and enhancing models

residuals, RSS, pred. vs. obs., QQ plots, chi-square, AIC, BIC, DIC, SE’s of
coefficients, testing nested models, others?
294 CHAPTER 6. MORE MODELS
 Chapter 7

Mathematical Statistics

7.1 Properties of Statistics

7.1.1 Sufficiency
Consider the following two facts.

1. Let Y1 , . . . , Yn ∼Pi.i.d. Poi(λ). Chapter 2, Exercise 6 showed that `(λ)

depends only on Yi and not on the specific values of the individual Yi ’s.

2. Let Y1 , . . . , Yn ∼P
i.i.d. Exp(λ). Chapter 2, Exercise 19 showed that `(λ)
depends only on Yi and not on the specific values of the individual Yi ’s.

Further, since `(λ) quantifies how strongly the data support each value of λ,
other aspects of y are irrelevant. For
P inference about λ it suffices to know `(λ),
and therefore it suffices
P to know Yi . We don’t need to know the individual
Yi ’s. We say that Yi is a sufficient statistic for λ.
Section 7.1.1 examines the general concept of sufficiency. We work in the
context of a parametric family. The idea of sufficiency is formalized in Defini-
tion 7.1.

Definition 7.1. Let {p(· | θ)} be a family of probability densities indexed by a

parameter θ. Let y = (y1 , . . . , yn ) be a sample from p(· | θ) for some unknown
θ. Let T (y) be a statistic such that the joint distribution factors as
Y
p(yi | θ) = g(T (y), θ)h(y).

for some functions g and h. Then T is called a sufficient statistic for θ.

The idea is that once the data

Q have been observed, h(y) is a constant that
does not depend of θ, so `(θ) ∝ p(yi | θ) = g(T, θ)h(y) ∝ g(T, θ). Therefore,
in order to know the likelihood function and make inference about θ, we only
need to know T (y), not anything
P else about y. For our Poisson and Exponential
examples we can take T (y) = yi .

295
296 CHAPTER 7. MATHEMATICAL STATISTICS

For a more detailed look at sufficiency, think of generating three Bern(θ)

trials y ≡ (y1 , y2 , y3 ). y can be generated, obviously, by generating y1 , y2 , y3
sequentially. The possible outcomes and their probabilities are

(0, 0, 0) (1 − θ)3

(1, 0, 0)
(0, 1, 0) θ(1 − θ)2
(0, 0, 1)

(1, 1, 0)
(1, 0, 1) θ2 (1 − θ)
(0, 1, 1)

(1, 1, 1) θ3
But y can also be generated by a two-step procedure:
P
1. Generate yi = 0, 1, 2, 3 with probabilities (1−θ)3 , 3θ(1−θ)2 , 3θ2 (1−θ),
θ3 , respectively.
P
2. (a) If yi = 0, generate (0, 0, 0)
P
(b) If yi = 1, generate (1, 0, 0), (0, 1, 0), or (0, 0, 1) each with proba-
bility 1/3.
P
(c) If yi = 2, generate (1, 1, 0), (1, 0, 1), or (0, 1, 1) each with proba-
bility 1/3.
P
(d) If yi = 3, generate (1, 1, 1)

It is easy to check that the two-step procedure generates each of the 8 possible
outcomes with the same probabilities as the obvious sequential procedure. For
generating y the two procedures are equivalent. But in the two-step procedure,
only the first step depends on θ. So if we want to use the data to learn about θ,
we only need know the outcome P of the first step. P The second step is irrelevant.
I.e., we only need to know yi . In other words, yi is sufficient.
For an example of another type, let y1 , . . . , yn ∼ i.i.d. U(0, θ). What is a
sufficient statistic?
(
1
n if yi < θ for i = 1, . . . , n
p(y | θ) = θ
0 otherwise
1
= 1(0,θ) (y(n) )
θn
shows that y(n) , the maximum of the yi ’s, is a one dimensional sufficient statistic.

Example 7.1
In World War II, when German tanks came from the factory they had serial numbers
labelled consecutively from 1. I.e., the numbers were 1, 2, . . . . The Allies wanted to
7.1. PROPERTIES OF STATISTICS 297

estimate T , the total number of German tanks and had, as data, the serial numbers
of captured tanks. See Exercise 22 in Chapter 5. Assume that tanks were captured
independently of each other and that all tanks were equally likely to be captured.
Let x1 , . . . , xn be the serial numbers of the captured tanks. Then x(n) is a sufficient
statistic. Inference about the total number of German tanks should be based on
x(n) and not on any other aspect of the data.

Sometimes sufficient statistics are higher dimensional. Let y1 , . . . , yn ∼

i.i.d. Gam(α, β). Then
  n Y 
Y Y 1 α−1 −yi /β 1 α−1 P
p(yi | α, β) = α
y i e = α
yi e− yi /β
Γ(α)β Γ(α)β
Q P
so T (y) = ( yi , yi ) is a two dimensional sufficient statistic.
Sufficient statistics are not unique. If T = T (y) is a sufficient statistic, and if
f is a 1-1 function, then f (T ) is Palso sufficient. So in theP Poisson, Exponential,
and Bernoulli examples where yi was sufficient, ȳ = yi /n is also sufficient.
But the lack of uniqueness is even more severe. The whole data set T (y) = (y)
is an n-dimensional sufficient statistic because
Y
p(yi | θ) = g(T (y), θ)h(y)

where g(T (y), θ) = p(y | θ) and h(y) = 1. Also, the order statistic T (y) =
(y(1) , . . . , y(n) ) is another n-dimensional sufficient statistic. Also, if T is any
sufficient one dimensional statistic then T2 = (y1 , T ) is a two dimensional suf-
ficient statistic. But it is intuitively clear that these sufficient statistics are
higher-dimensional than necessary. They can be reduced to lower dimensional
statistics while retaining sufficiency, that is, without losing information.
The key idea in the preceding paragraph is that the high dimensional suffi-
cient statistics can be transformed into the low dimensional ones, but not vice
versa. E.g., ȳ is a function of (y(1) , . . . , y(n) ) but (y(1) , . . . , y(n) ) is not a function
of ȳ. Definition 7.2 is for statistics that have been reduced as much as possible
without losing sufficiency.
Definition 7.2. A sufficient statistic T (y) is called minimal sufficient if, for
every other sufficient statistic T2 , T (y) is a function of T2 (y).
This book does not delve into methods for finding minimal sufficient statis-
tics. In most cases the user can recognize whether a statistic is minimal suffi-
cient.
Does the theory of sufficiency imply that statisticians need look only at suf-
ficient statistics and not at other aspects of the data? Not quite. Let y1 , . . . , yn
be binary random variables and suppose we adopt the model P y 1 , . . . , yn ∼
i.i.d. Bern(θ). Then for estimating θ we need look only at yi . But suppose
(y1 , . . . , yn ) turn out to be

0 · · · 0} 1
| 0 {z · · · 1},
| 1 {z
many 0’s many 1’s
298 CHAPTER 7. MATHEMATICAL STATISTICS

i.e., many 0’s followed by many 1’s. Such a dataset would cast doubt on the
assumption that the yi ’s are independent. Judging from this dataset, it looks
much more likely that the yi ’s come in streaks. So statisticians should look at
all the data, not just sufficient statistics, because looking at all the data can
help us create and critique models. But once a model has been adopted, then
inference should be based on sufficient statistics.

7.1.2 Consistency, Bias, and Mean-squared Error

Consistency Heuristically speaking, as we collect ever more data we should
be able to learn the truth ever more accurately. This heuristic is captured
formally, at least for parameter estimation, by the notion of consistency. To say
whether an estimator is consistent we have to define it for every sample size.
To that end, let Y1 , Y2 , · · · ∼ i.i.d. f for some unknown density f having finite
mean µ and SD σ. For each n ∈ N let Tn : Rn → R. I.e. Tn is a real-valued
function of (y1P, . . . , yn ). For example, if we’re trying to estimate µ we might
take Tn = n−1 n1 yi .

Definition 7.3. The sequence of estimators T1 , T2 , . . . is said to be consistent

for the parameter θ if for every θ and for every  > 0.

lim P[|Tn − θ| < ] = 1.

n→∞

For example, the Law of Large

Pn Numbers, Theorem 1.12, says the sequence
−1
of sample
P means {T n = n 1 yi } is consistent for µ. Similarly, let Sn =
N −1 i (yi − Tn )2 be the sample variance. Then {Sn } is consistent for σ 2 .
More generally, m.l.e.’s are consistent.

Theorem 7.1. Let Y1 , Y2 , · · · ∼ i.i.d. pY (y | θ) and let θ̂n be the m.l.e. from the
sample (y1 , . . . , yn ). Further, let g be a continuous function of θ. Then, subject
to regularity conditions, {g(θ̂n )} is a consistent sequence of estimators for g(θ).

Proof. The proof requires regularity conditions relating to differentiability and

the interchange of integral and derivative. It is beyond the scope of this book.

Consistency is a good property; one should be wary of an inconsistent estima-

tor. On the other hand, consistency alone does not guarantee that a sequence
of estimators is optimal, or even sensible. For example, let Rn (y1 , . . . , yn ) =
(bn/2c)−1 (y1 + · · · + ybn/2c ), the mean of the first half of the observations. (bwc
is the floor of w, the largest integer not greater than w.) The sequence {R n } is
consistent for µ but is not as good as the sequence of sample means.

Bias It seems natural to want the sampling distribution of an estimator to be

centered around the parameter being estimated. This desideratum is captured
formally, at least for centering in the sense of expectation, by the notion of bias.
7.2. TRANSFORMATIONS OF PARAMETERS 299

Definition 7.4. Let θ̂ = θ̂(y1 , . . . , yn ) be an estimator of a parameter θ. The

quantity E[θ̂] − θ is called the bias of θ̂. An estimator whose bias is 0 is called
unbiased.

Here are some examples.

An unbiased estimator Let y1 , . . . , yn ∼ i.i.d. N(µ, σ) and consider µ̂ = ȳ as

an estimate of µ. Because E[ȳ] = µ, ȳ is an unbiased estimate of µ.

A biased estimator Let y1 , . . . , yn ∼ i.i.d. N(µ, σ) and consider

X
σ̂ 2 = n−1 (yi − ȳ)2

as an estimate of σ 2 .
X X
E[n−1 (yi − ȳ)2 ] = n−1 E[ (yi − µ + µ − ȳ)2 ]
n X X
= n−1 E[ (yi − µ)2 ] + 2E[ (yi − µ)(µ − ȳ)]
X o
+ E[ (µ − ȳ)2 ]

= n−1 nσ 2 − 2σ 2 + σ 2
= σ 2 − n−1 σ 2
n−1 2
= σ
n

Therefore σ̂ 2 is a biased estimator of σ 2 . It’s bias is −σ 2 /n.PSome statis-

ticians prefer to use the unbiased estimator σ̃ 2 = (n − 1)−1 (yi − ȳ)2 .

A biased estimator Let x1 , . . . , xn ∼ i.i.d. U(0, θ) and consider θ̂ = x(n) as

an estimate of θ. (θ̂ is the m.l.e.; see Section 5.4.) But x(n) < θ; therefore
E[x(n) ] < θ; therefore x(n) is a biased estimator of θ.

Mean Squared Error

7.1.3 Efficiency
7.1.4 Asymptotic Normality
7.1.5 Robustness

7.2 Transformations of Parameters

Equivalent parameterizations, especially ANOVA’s, etc. Invariance of MLEs.
300 CHAPTER 7. MATHEMATICAL STATISTICS

7.3 Information

7.4 More Hypothesis Testing

[a more formal presentation here?]

7.4.1 p values
7.4.2 The Likelihood Ratio Test
7.4.3 The Chi Square Test
7.4.4 Power

7.5 Exponential families

7.6 Location and Scale Families

Location/scale families

7.7 Functionals
functionals

7.8 Invariance
Invariance

7.9 Asymptotics
In real life, data sets are finite: (y1 , . . . , yn ). Yet we often appeal to the Law
of Large Numbers or the Central Limit Theorem, Theorems 1.12, 1.13, and
1.14, which concern the limit of a sequence of random variables as n → ∞.
The hope is that when n is large those theorems will tell us something, at least
approximately, about the distribution of the sample mean. But we’re faced with
the questions “How large is large?” and “How close is the approximation?”
To take an example, we might want to apply the Law of Large Numbers or
the Central Limit Theorem to a sequence Y1 , Y2 , . . . of random variables from a
distribution with mean µ and SD σ. Here are a few instances of the first several
elements of such a sequence.
7.9. ASYMPTOTICS 301

0.70 0.29 0.09 -0.23 -0.30 -0.79 -0.72 -0.35 1.79 ···
-0.23 -0.24 0.29 -0.16 0.37 -0.01 -0.48 -0.59 0.39 ···
-1.10 -0.91 -0.34 0.22 1.07 -1.51 -0.41 -0.65 0.07 ···
.. .. .. .. .. .. .. .. .. ..
. . . . . . . . . .
Each sequence occupies one row of the array. The “· · · ” indicates that the
.
sequence continues infinitely. The “..” indicates that there are infinitely many
such sequences. The numbers were generated by
y <- matrix ( NA, 3, 9 )
for ( i in 1:3 ) {
y[i,] <- rnorm(9)
print ( round ( y[i,], 2 ) )
}
• I chose to generate Yi ’s from the N(0, 1) distribution, so I used rnorm, and
so, for this example, µ = 0 and σ = 1. Those are arbitrary choices. I
could have used any values of µ and σ and any distribution for which I
know how to generate random variables on the computer.
• round does rounding. In this case we’re printing each number with two
decimal places.
Because there are multiple sequences, each with multiple elements, we need two
subscripts to keep track of things properly. Let Yij be the j’th element of the
i’th sequence. For the i’th sequence of random variables, we’re interested in the
sequence of means Ȳi1 , Ȳi2 , . . . where Ȳin = (Yi1 + · ·√
· + Yin )/n. And we’re also
interested in the sequence Zi1 , Zi2 , . . . where Zin = n(Ȳin − µ). For the three
instances above, the Ȳin ’s and Zin ’s can be printed with
for ( i in 1:3 ) {
print ( round ( cumsum(y[i,]) / 1:9, 2) )
print ( round ( cumsum(y[i,]) / (sqrt(1:9)), 2) )
}
• cumsum computes a cumulative sum; so cumsum(y[1,]) yields the vector
y[1,1], y[1,1]+y[1,2], ..., y[1,1]+...+y[1,9]. (Print out
cumsum(y[1,]) if you’re not sure what it is.) Therefore,
cumsum(y[i,])/1:9 is the sequence of Ȳin ’s.
• sqrt computes the square root. So the second print statement prints the
sequence of Zin ’s.
The results for the Ȳin ’s are
0.70 0.49 0.36 0.21 0.11 -0.04 -0.14 -0.16 0.05 ···
-0.23 -0.23 -0.06 -0.08 0.01 0.00 -0.07 -0.13 -0.07 ···
-1.10 -1.01 -0.78 -0.53 -0.21 -0.43 -0.43 -0.45 -0.40 ···
.. .. .. .. .. .. .. .. .. ..
. . . . . . . . . .
302 CHAPTER 7. MATHEMATICAL STATISTICS

and for the Zin ’s are

0.70 0.91 0.96 0.84 0.71 0.39 0.11 -0.01 0.59 ···
-0.23 -0.40 -0.23 -0.31 -0.15 -0.15 -0.33 -0.54 -0.41 ···
-1.10 -1.74 -1.94 -1.83 -1.35 -1.97 -2.12 -2.35 -2.33 ···
.. .. .. .. .. .. .. .. .. ..
. . . . . . . . . .
We’re interested in the following questions.
1. Will every sequence of Ȳi ’s or Zi ’s converge? This is a question about the
limit along each row of the array.
2. If they converge, do they all have the same limit?
3. If not every sequence converges, what fraction of them converge; or what is
the probability that a randomly chosen sequence of Ȳi ’s or Zi ’s converges?
4. For a fixed n, what is the distribution of Ȳn or Zn . This is a question
about the distribution along columns of the array.
5. Does the distribution of Ȳn or Zn depend on n?
6. Is there a limiting distribution as n → ∞?
Some simple examples and the Strong Law of Large Numbers, Theorem 1.13,
answer questions 1, 2, and 3 for the sequences of Ȳi ’s. The Central Limit
Theorem, Theorem 1.14, answers question 6 for the sequences of Zi ’s.
1. Will every sequence of Ȳi ’s converge? No. Suppose the sequence of Yi ’s is
1, 2, 3, . . . . Then {Ȳi } increases without limit and does not converge.
2. If they converge, do they have the same limit? No. Here are two sequences
of Yi ’s.

1 1 1 ···
-1 -1 -1 ···

The corresponding sequences {Ȳi } converge to different limits.

3. What is the probability of convergence? The probability of convergence
is 1. That’s the Strong Law of Large Numbers. In particular, the probabil-
ity of randomly getting a sequence that doesn’t converge, like 1, 2, 3, . . . ,
is 0. But the Strong Law of Large Numbers says even more. It says

P[ lim Ȳn = µ] = 1.
n→∞

So the probability of getting sequences such as 1, 1, 1, . . . or −1, −1, −1, . . .

that converge to something other than µ is 0.
4. What is the distribution of Zn ? We cannot say in general. It depends on
the distribution of the individual Yi ’s.
7.9. ASYMPTOTICS 303

5. Does the distribution of Zn depend on n? Yes, except in the special case

where Yi ∼ N(0, 1) for all i.

6. Is there a limiting distribution? Yes. That’s the Central Limit Theorem.

Regardless of the distribution of the Yi ’s, as long as Var(Yi ) < ∞, the
limit, as n → ∞, of the distribution of Zn is N(0, 1).

The Law of Large Numbers and the Central Limit Theorem are theorems
about the limit as n → ∞. When we use those theorems in practice we hope that
our sample size n is large enough that Ȳin ≈ µ and Zin ∼ N(0, 1), approximately.
But how large should n be before relying on these theorems, and how good is
the approximation? The answer is, “It depends on the distribution of the Yi ’s”.
That’s what we look at next.
To illustrate, we generate sequences of Yij ’s from two distributions, compute
Ȳin ’s and Zin ’s for several values of n, and compare. One distribution is U(0, 1);
the other is a recentered and rescaled version of Be(.39, .01).
The Be(.39, .01) density, shown in Figure 7.1, was chosen for its asymmetry.
It has a mean of .39/.40 = .975 and a variance of (.39)(.01)/((.40) 2 (1.40)) ≈
.017. It was recentered and rescaled to have a mean of .5 and variance of 1/12,
the same as the U(0, 1) distribution.
Densities of the Ȳin ’s are in Figure 7.2. As the sample size increases from
n = 10 to n = 270, the Ȳin ’s from both distributions get closer to their expected
value of 0.5. That’s the Law of Large Numbers at work. The amount by which
they’re off their mean goes from about ±.2 to about ±.04. That’s Corollary 1.10
at work. And finally, as n → ∞, the densities get more Normal. That’s the
Central Limit Theorem at work.
Note that the density of the Ȳin ’s derived from the U(0, 1) distribution is
close to Normal even for the smallest sample size, while the density of the Ȳin ’s
derived from the Be(.39, .01) distribution is way off. That’s because U(0, 1) is
symmetric and unimodal, and therefore close to Normal to begin with, while
Be(.39, .01) is far from symmetric and unimodal, and therefore far from Normal,
to begin with. So Be(.39, .01) needs a larger n to make the Central Limit
Theorem work; i.e., to be a good approximation.
Figure 7.3 is for the Zin ’s. It’s the same as Figure 7.2 except that each
density has been recentered and rescaled to have mean 0 and variance 1. When
put on the same scale we can see that all densities are converging to N(0, 1).

Figure 7.1 was produced by

x <- seq ( .01, .99, length=80 )
plot ( x, dbeta(x,.39,.01), type="l", ylab="", xlab="" )

Figure 7.2 was generated by the following R code.

304 CHAPTER 7. MATHEMATICAL STATISTICS

8
6
4
2
0

0.0 0.2 0.4 0.6 0.8 1.0

Figure 7.1: The Be(.39, .01) density

samp.size <- c ( 10, 30, 90, 270 )

n.reps <- 500
Y.1 <- matrix ( NA, n.reps, max(samp.size) )
Y.2 <- matrix ( NA, n.reps, max(samp.size) )
for ( i in 1:n.reps ) {
Y.1[i,] <- runif ( max(samp.size), 0, 1 )
Y.2[i,] <- ( rbeta ( max(samp.size), 0.39, .01 ) - .975 ) *
sqrt( .4^2*1.4 / (.39*.01*12 ) ) + .5
}
par ( mfrow=c(2,2) )
for ( n in 1:length(samp.size) ) {
Ybar.1 <- apply ( Y.1[,1:samp.size[n]], 1, mean )
Ybar.2 <- apply ( Y.2[,1:samp.size[n]], 1, mean )
sd <- sqrt ( 1 / ( 12 * samp.size[n] ) )
x <- seq ( .5-3*sd, .5+3*sd, length=60 )
y <- dnorm ( x, .5, sd )
den1 <- density ( Ybar.1 )
den2 <- density ( Ybar.2 )
ymax <- max ( y, den1$y, den2$y )
plot ( x, y, ylim=c(0,ymax), type="l", lty=3, ylab="",
xlab="", main=paste("n =", samp.size[n]) )
lines ( den1, lty=2 )
lines ( den2, lty=4 )
}

• The manipulations in the line Y.2[i,] <- ... are so Y.2 will have mean
1/2 and variance 1/12.
7.9. ASYMPTOTICS 305

30 n = 10 n = 30

30
25

25
20

20
15

15
10

10
5

5
0

0.0 0.4 0.8 0.0 0.4 0.8

n = 90 n = 270
30

30
25

25
20

20
15

15
10

10
5

5
0

0.0 0.4 0.8 0.0 0.4 0.8

Figure 7.2: Densities of Ȳin for the U(0, 1) (dashed), modified Be(.39, .01) (dash
and dot), and Normal (dotted) distributions.
306 CHAPTER 7. MATHEMATICAL STATISTICS

7.10 Exercises
1. Let Y1 , . . . , Yn be a sample from N(µ, σ).

(a) Suppose µ is unknown but σ is known. Find a one dimensional

sufficient statistic.
(b) Suppose µ is known but σ is unknown. Find a one dimensional
sufficient statistic.
(c) Suppose µ and σ are both unknown. Find a two dimensional sufficient
statistic.
2. Let Y1 , . . . , Yn be a sample from Be(α, β). Find a two dimensional suffi-
cient statistic.
3. Let Y1 , . . . , Yn ∼ i.i.d. U(−θ, θ). Find a low dimensional sufficient statistic.
7.10. EXERCISES 307

1.0 n = 10 n = 30

1.0
0.8

0.8
0.6

0.6
0.4

0.4
0.2

0.2
0.0

0.0

−3 −1 1 3 −3 −1 1 3

n = 90 n = 270
1.0

1.0
0.8

0.8
0.6

0.6
0.4

0.4
0.2

0.2
0.0

0.0

−3 −1 1 3 −3 −1 1 3

Figure 7.3: Densities of Zin for the U(0, 1) (dashed), modified Be(.39, .01) (dash
and dot), and Normal (dotted) distributions.
308 CHAPTER 7. MATHEMATICAL STATISTICS
 Bibliography

Consumer Reports, June:366–367, 1986.

D. F. Andrews and A. M. Herzberg. Data. Springer-Verlag, New York, 1985.

H. Bateman. On the probability distribution of α particles. Philosophical Mag-

azine Series 6, 20:704–705, 1910.

Paul Brodeur. Annals of radiation, the cancer at slater school. The New Yorker,
Dec. 7, 1992.

Jason C. Buchan, Susan C. Alberts, Joan B. Silk, and Jeanne Altmann. True
paternal care in a multi-male primate society. Nature, 425:179–181, 2003.

George Casella and Roger L. Berger. Statistical Inference. Duxbury, Pacific

Grove, second edition, 2002.

Lorraine Denby and Daryl Pregibon. An example of the use of graphics in

regression. The American Statistician, 41:33–38, 1987.

A. J. Dobson. An Introduction to Statistical Modelling. Chapman and Hall,

London, 1983.

D. Freedman, R. Pisani, and R. Purves. Statistics. W. W. Norton and Company,

New York, 1998.

Michael Lavine. What is Bayesian statistics and why everything else is wrong.
The Journal of Undergraduate Mathematics and Its Applications, 20:165–174,
1999.

Michael Lavine, Brian Beckage, and James S. Clark. Statistical modelling of

seedling mortality. Journal of Agricultural, Biological and Environmental
Statistics, 7:21–41, 2002.

E. Rutherford and H. Geiger. The probability variations in the distribution of

α particles. Philosophical Magazine Series 6, 20:698–704, 1910.

Mark J. Schervish. Theory of Statistics. Springer-Verlag, New York, 1995.

309
310 BIBLIOGRAPHY

T.S. Tsou and R.M. Royall. Robust likelihoods. Journal of the American Sta-
tistical Association, 90:316–320, 1995.
Jessica Utts. Replication and meta-analysis in parapsychology. Statistical Sci-
ence, 4:363–403, 1991.
Sanford Weisberg. Applied Linear Regression. John Wiley & Sons, New York,
second edition, 1985.
 Index

α particle, 253 Cauchy, 283

Exponential, 18
bandwidth, 90 Gamma, 259
bias, 298 Multinomial, 250
Negative binomial, 245
case, 186 Normal, 20, 266
cdf, see cumulative distribution func- Poisson, 16, 251
tion standard multivariate Normal, 271
Central Limit Theorem, 69 standard Normal, 26
change of variables, 11 Uniform, 256
characteristic functions, 237
Chebychev’s Inequality, 68 errors, 188
chi-squared distribution, 265 estimate, 130
consistency, 298 expected value, 30
coplots, 104 explanatory variable, 186
correlation, 46
covariance, 44 fitted values, 192, 213
covariance matrix, 229 fitting, 193
covariate, 185 floor, 298
cross tabulation, 99 formula, 193
cumulative distribution function, 233
gamma function, 259
DASL, see Data and Story Library, see Gaussian density, 266
Data and Story Library, 175 generalized moment, 33
Data and Story Library, 89, 120 genotype, 250
density
probability, 227 half-life, 264
density estimation, 89 histogram, 89
dependence, 47
distribution, 2 independence, 47
Distributions joint, 229
Bernoulli, 242 mutual, 229
Beta, 265 indicator function, 47
Binomial, 13, 241 indicator variable, 47

311
312 INDEX

Jacobian, 231 [[]], see subscript

[], see subscript
Laplace transform, 236 #, 4
Law of Large Numbers, 68, 69 %o%, 274
likelihood function, 112, 130 ,̃ 193
likelihood set, 132 abline, 125
linear model, 188 apply, 55
linear predictor, 207 array, 58
location parameter, 270 as.factor, 212
logistic regression, 207 assignment, 4, 5
logit, 207 +<-+,4
boxplot, 53, 94
marginal likelihood, 119 c, 8
maximum likelihood estimate, 131 contour, 125
mean, 27, 30 coplot, 109
median, 81 cor, 46
Mediterranean tongue, 27 cov, 44
mgf, see moment generating function cumsum, 301
minimal sufficient, 297 data.frame, 212
moment, 32 dbinom, 14, 243
moment generating function, 236 density, 11
mosaic plot, 100 dexp, 20
multinomial coefficient, 251 diff, 269
multivariate dim, 63
change of variables, 231 dmultinom, 251
dnbinom, 248
order statistic, 82, 297
dnorm, 21
outer product, 274
dpois, 17
parameter, 13, 112, 241 expression, 114
parametric family, 13, 112, 241 fitted, 216
pdf, see probability density, see prob- for, 4, 5
ability density glm, 211
physics, 7 hist, 11
Poisson process, 264 if, 5
predicted values, 213 is.na, 125
probability legend, 20
continuous, 1, 6 length, 64
density, 6 lines, 11
discrete, 1, 6 lm, 193
log10, 130
QQ plots, 95 lowess, 176
quantile, 81 matplot, 20
matrix, 20, 53
R commands mean, 11
!, 51 median, 81
==, 4 mosaicplot, 100
INDEX 313

names, 63 response variable, 185

pairs, 44
par, 9 sampling distribution, 135
paste, 20 scale parameter, 270
pbinom, 243 scatterplot smoothers, 176
plot, 8, 17 standard deviation, 31
pnbinom, 248 standard error, 139
print, 4 standard Normal distribution, 270
qbinom, 243 standard units, 24
qnbinom, 248 StatLib, 63
qqnorm, 95 Strong Law of Large Numbers, 69
quantile, 82 sufficient statistic, 295
rbinom, 33, 243
read.table, 63 variance, 31
rep, 5
Weak Law of Large Numbers, 68
rmultinom, 251
rnbinom, 248
rnorm, 24
round, 301
sample, 3
segments, 234
seq, 17
sqrt, 301
stripchart, 92
subscript, 63, 64
sum, 4, 5
supsmu, 176
unique, 85
var, 11
while, 51
R data sets
airquality, 173
attenu, 173
co2, 55
discoveries, 9
faithful, 103, 173
iris, 44
mtcars, 173, 197, 216
PlantGrowth, 184
Tooth Growth, 84, 157
UCBAdmissions, 98
random vector, 228
regression function, 179
regressor, 186
residual, 180
residual plots, 199
 Index of Examples

1970 draft lottery, 171

baboons, 156

CEO salary, 116

craps, 2, 3, 15, 42, 49, 51, 129

FACE, 55, 117, 198

hot dogs, 86, 177, 182, 189

Ice Cream Consumption, 185

O-rings, 198
ocean temperatures, 26, 258

quiz scores, 90, 120

Rutherford and Geiger, 244

seedlings, 16, 36, 42, 48, 111, 122, 139,

172, 207
Slater school, 112, 115, 161

314