Introduction to Statistical Thought

Michael Lavine

February 21, 2013

 i

Copyright c 2005 by Michael Lavine

 Contents

List of Figures v

List of Tables x

Preface xi

1 Probability 1
1.1 Basic Probability . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 Probability Densities . . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.3 Parametric Families of Distributions . . . . . . . . . . . . . . . . . . . 14
1.3.1 The Binomial Distribution . . . . . . . . . . . . . . . . . . . . . 14
1.3.2 The Poisson Distribution . . . . . . . . . . . . . . . . . . . . . 17
1.3.3 The Exponential Distribution . . . . . . . . . . . . . . . . . . . 20
1.3.4 The Normal Distribution . . . . . . . . . . . . . . . . . . . . . 22
1.4 Centers, Spreads, Means, and Moments . . . . . . . . . . . . . . . . . . 29
1.5 Joint, Marginal and Conditional Probability . . . . . . . . . . . . . . . 41
1.6 Association, Dependence, Independence . . . . . . . . . . . . . . . . . . 50
1.7 Simulation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 56
1.7.1 Calculating Probabilities . . . . . . . . . . . . . . . . . . . . . 57
1.7.2 Evaluating Statistical Procedures . . . . . . . . . . . . . . . . . 61
1.8 R . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 72
1.9 Some Results for Large Samples . . . . . . . . . . . . . . . . . . . . . . 76
1.10 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 80

2 Modes of Inference 94
2.1 Data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
2.2 Data Description . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
2.2.1 Summary Statistics . . . . . . . . . . . . . . . . . . . . . . . . . 95

ii
CONTENTS iii

2.2.2 Displaying Distributions . . . . . . . . . . . . . . . . . . . . . . 101

2.2.3 Exploring Relationships . . . . . . . . . . . . . . . . . . . . . . 114
2.3 Likelihood . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
2.3.1 The Likelihood Function . . . . . . . . . . . . . . . . . . . . . . 133
2.3.2 Likelihoods from the Central Limit Theorem . . . . . . . . . . 140
2.3.3 Likelihoods for several parameters . . . . . . . . . . . . . . . . 145
2.4 Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 154
2.4.1 The Maximum Likelihood Estimate . . . . . . . . . . . . . . . . 154
2.4.2 Accuracy of Estimation . . . . . . . . . . . . . . . . . . . . . . 156
2.4.3 The sampling distribution of an estimator . . . . . . . . . . . . . 159
2.5 Bayesian Inference . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 164
2.6 Prediction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
2.7 Hypothesis Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 178
2.8 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 192

3 Regression 202
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 202
3.2 Normal Linear Models . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
3.2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . 210
3.2.2 Inference for Linear Models . . . . . . . . . . . . . . . . . . . 222
3.3 Generalized Linear Models . . . . . . . . . . . . . . . . . . . . . . . . 235
3.3.1 Logistic Regression . . . . . . . . . . . . . . . . . . . . . . . . 235
3.3.2 Poisson Regression . . . . . . . . . . . . . . . . . . . . . . . . . 244
3.4 Predictions from Regression . . . . . . . . . . . . . . . . . . . . . . . . 248
3.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 252

4 More Probability 263

4.1 More Probability Density . . . . . . . . . . . . . . . . . . . . . . . . . 263
4.2 Random Vectors . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 264
4.2.1 Densities of Random Vectors . . . . . . . . . . . . . . . . . . . 264
4.2.2 Moments of Random Vectors . . . . . . . . . . . . . . . . . . . 266
4.2.3 Functions of Random Vectors . . . . . . . . . . . . . . . . . . . 266
4.3 Representing Distributions . . . . . . . . . . . . . . . . . . . . . . . . . 270
4.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 275

5 Special Distributions 279

5.1 Binomial and Negative Binomial . . . . . . . . . . . . . . . . . . . . . . 279
5.2 Multinomial . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 289
5.3 Poisson . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 291
CONTENTS iv

5.4 Uniform . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 304

5.5 Gamma, Exponential, Chi Square . . . . . . . . . . . . . . . . . . . . . 305
5.6 Beta . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 312
5.7 Normal . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 315
5.7.1 The Univariate Normal Distribution . . . . . . . . . . . . . . . 315
5.7.2 The Multivariate Normal Distribution . . . . . . . . . . . . . . 320
5.7.3 Marginal, Conditional, and Related Distributions . . . . . . . . 327
5.8 The t Distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 330
5.9 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 335

6 Bayesian Statistics 346

6.1 Multidimensional Bayesian Analysis . . . . . . . . . . . . . . . . . . . . 346
6.2 Metropolis, Metropolis-Hastings, and Gibbs . . . . . . . . . . . . . . . 358
6.3 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 377

7 More Models 382

7.1 Random Effects . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 382
7.2 Time Series and Markov Chains . . . . . . . . . . . . . . . . . . . . . . 396
7.3 Survival analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 409
7.4 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 416

8 Mathematical Statistics 420

8.1 Properties of Statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . 420
8.1.1 Sufficiency . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 420
8.1.2 Consistency, Bias, and Mean-squared Error . . . . . . . . . . . 423
8.2 Information . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 428
8.3 Exponential families . . . . . . . . . . . . . . . . . . . . . . . . . . . . 431
8.4 Asymptotics . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 433
8.4.1 Modes of Convergence . . . . . . . . . . . . . . . . . . . . . . . 439
8.4.2 The δ-method . . . . . . . . . . . . . . . . . . . . . . . . . . . . 443
8.4.3 The Asymptotic Behavior of Estimators . . . . . . . . . . . . . 446
8.5 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 451

Bibliography 456
 List of Figures

1.1 pdf for time on hold at Help Line . . . . . . . . . . . . . . . . . . . . . 7

1.2 pY for the outcome of a spinner . . . . . . . . . . . . . . . . . . . . . . 9
1.3 (a): Ocean temperatures; (b): Important discoveries . . . . . . . . . . . 11
1.4 Change of variables . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
1.5 Binomial probabilities . . . . . . . . . . . . . . . . . . . . . . . . . . . 16
1.6 P[X = 3 | λ] as a function of λ . . . . . . . . . . . . . . . . . . . . . . . 19
1.7 Exponential densities . . . . . . . . . . . . . . . . . . . . . . . . . . . . 21
1.8 Normal densities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
1.9 Ocean temperatures at 45◦ N, 30◦ W, 1000m depth . . . . . . . . . . . . . 25
1.10 Normal samples and Normal densities . . . . . . . . . . . . . . . . . . . 27
1.11 hydrographic stations off the coast of Europe and Africa . . . . . . . . 31
1.12 Water temperatures . . . . . . . . . . . . . . . . . . . . . . . . . . . . 32
1.13 Water temperatures with standard deviations . . . . . . . . . . . . . . . 36
1.14 Two pdf’s with ±1 and ±2 SD’s. . . . . . . . . . . . . . . . . . . . . . . 38
1.15 Permissible values of N and X . . . . . . . . . . . . . . . . . . . . . . . 44
1.16 Features of the joint distribution of (X, Y) . . . . . . . . . . . . . . . . 48
1.17 Lengths and widths of sepals and petals of 150 iris plants . . . . . . . . 52
1.18 correlations . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
1.19 1000 simulations of θ̂ for n.sim = 50, 200, 1000 . . . . . . . . . . . . . 60
1.20 1000 simulations of θ̂ under three procedures . . . . . . . . . . . . . . . 64
1.21 Monthly concentrations of CO2 at Mauna Loa . . . . . . . . . . . . . . 65
1.22 1000 simulations of a FACE experiment . . . . . . . . . . . . . . . . . . 69
1.23 Histograms of craps simulations . . . . . . . . . . . . . . . . . . . . . . 81

2.1 quantiles . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
2.2 Histograms of tooth growth . . . . . . . . . . . . . . . . . . . . . . . . 102
2.3 Histograms of tooth growth . . . . . . . . . . . . . . . . . . . . . . . . 103

v
LIST OF FIGURES vi

2.4 Histograms of tooth growth . . . . . . . . . . . . . . . . . . . . . . . . 104

2.5 calorie contents of beef hot dogs . . . . . . . . . . . . . . . . . . . . . 108
2.6 Strip chart of tooth growth . . . . . . . . . . . . . . . . . . . . . . . . 111
2.7 Quiz scores from Statistics 103 . . . . . . . . . . . . . . . . . . . . . . 113
2.8 QQ plots of water temperatures (◦ C) at 1000m depth . . . . . . . . . . . 115
2.9 Mosaic plot of UCBAdmissions . . . . . . . . . . . . . . . . . . . . . . 119
2.10 Mosaic plot of UCBAdmissions . . . . . . . . . . . . . . . . . . . . . . 120
2.11 Old Faithful data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
2.12 Waiting time versus duration in the Old Faithful dataset . . . . . . . . 124
2.13 Time series of duration and waiting time at Old Faithful . . . . . . . . . 125
2.14 Time series of duration and waiting time at Old Faithful . . . . . . . . . 126
2.15 Temperature versus latitude for different values of longitude . . . . . 129
2.16 Temperature versus longitude for different values of latitude . . . . . 130
2.17 Spike train from a neuron during a taste experiment. The dots show the
times at which the neuron fired. The solid lines show times at which the
rat received a drop of a .3 M solution of NaCl. . . . . . . . . . . . . . . 131
2.18 Likelihood function for the proportion of red cars . . . . . . . . . . . . 135
`(θ) after yi = 40 in 60 quadrats. . . . . . . . . . . . . . . . . . . . .
P
2.19 137
2.20 Likelihood for Slater School . . . . . . . . . . . . . . . . . . . . . . . 139
2.21 Marginal and exact likelihoods for Slater School . . . . . . . . . . . . 142
2.22 Marginal likelihood for mean CEO salary . . . . . . . . . . . . . . . . 144
2.23 FACE Experiment: data and likelihood . . . . . . . . . . . . . . . . . . 147
2.24 Likelihood function for Quiz Scores . . . . . . . . . . . . . . . . . . . 149
2.25 Log of the likelihood function for (λ, θ f ) in Example 2.13 . . . . . . . . 153
2.26 Likelihood function for the probability of winning craps . . . . . . . . 158
2.27 Sampling distribution of the sample mean and median . . . . . . . . . . . 160
2.28 Histograms of the sample mean for samples from Bin(n, .1) . . . . . . . . 162
2.29 Prior, likelihood and posterior in the seedlings example . . . . . . . . . 169
2.30 Prior, likelihood and posterior densities for λ with n = 1, 4, 16 . . . . . 171
2.31 Prior, likelihood and posterior densities for λ with n = 60 . . . . . . . 172
2.32 Prior, likelihood and posterior density for Slater School . . . . . . . . 173
2.33 Plug-in predictive distribution for seedlings . . . . . . . . . . . . . . . 175
2.34 Predictive distributions for seedlings after n = 0, 1, 60 . . . . . . . . . 179
2.35 pdf of the Bin(100, .5) distribution . . . . . . . . . . . . . . . . . . . . . 183
2.36 pdfs of the Bin(100, .5) (dots) and N(50, 5) (line) distributions . . . . . . 184
2.37 Approximate density of summary statistic w . . . . . . . . . . . . . . . . 186
2.38 Number of times baboon father helps own child . . . . . . . . . . . . . . 190
2.39 Histogram of simulated values of w.tot . . . . . . . . . . . . . . . . . . 191
LIST OF FIGURES vii

3.1 Four regression examples . . . . . . . . . . . . . . . . . . . . . . . . . 203

3.2 1970 draft lottery. Draft number vs. day of year . . . . . . . . . . . . 206
3.3 Draft number vs. day of year with smoothers . . . . . . . . . . . . . . . 207
3.4 Total number of New seedlings 1993 – 1997, by quadrat. . . . . . . . . . 209
3.5 Calorie content of hot dogs . . . . . . . . . . . . . . . . . . . . . . . . 211
3.6 Density estimates of calorie contents of hot dogs . . . . . . . . . . . . 213
3.7 The PlantGrowth data . . . . . . . . . . . . . . . . . . . . . . . . . . . 215
3.8 Ice cream consumption versus mean temperature . . . . . . . . . . . . . 221
3.9 Likelihood functions for (µ, δ M , δP ) in the Hot Dog example. . . . . . . 227
3.10 pairs plot of the mtcars data . . . . . . . . . . . . . . . . . . . . . . . 230
3.11 mtcars — various plots . . . . . . . . . . . . . . . . . . . . . . . . . . 232
3.12 likelihood functions for β1 , γ1 , δ1 and δ2 in the mtcars example. . . . . 234
3.13 Pine cones and O-rings . . . . . . . . . . . . . . . . . . . . . . . . . . . 238
3.14 Pine cones and O-rings with regression curves . . . . . . . . . . . . . . 239
3.15 Likelihood function for the pine cone data . . . . . . . . . . . . . . . . 242
3.16 Actual vs. fitted and residuals vs. fitted for the seedling data . . . . . 247
3.17 Diagnostic plots for the seedling data . . . . . . . . . . . . . . . . . . . 249
3.18 Actual mpg and fitted values from three models . . . . . . . . . . . . . 251
3.19 Happiness Quotient of bankers and poets . . . . . . . . . . . . . . . . . 255

4.1 The (X1 , X2 ) plane and the (Y1 , Y2 ) plane . . . . . . . . . . . . . . . . . . 269

4.2 pmf’s, pdf’s, and cdf’s . . . . . . . . . . . . . . . . . . . . . . . . . . . . 272

5.1 The Binomial pmf . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 285

5.2 The Negative Binomial pmf . . . . . . . . . . . . . . . . . . . . . . . . . 288
5.3 Poisson pmf for λ = 1, 4, 16, 64 . . . . . . . . . . . . . . . . . . . . . . . 294
5.4 Rutherford and Geiger’s Figure 1 . . . . . . . . . . . . . . . . . . . . . 299
5.5 Numbers of firings of a neuron in 150 msec after five different tastants.
Tastants: 1=MSG .1M; 2=MSG .3M; 3=NaCl .1M; 4=NaCl .3M; 5=wa-
ter. Panels: A: A stripchart. Each circle represents one delivery of a
tastant. B: A mosaic plot. C: Each line represents one tastant. D: Like-
lihood functions. Each line represents one tastant. . . . . . . . . . . . 301
5.6 The line shows Poisson probabilities for λ = 0.2; the circles show the
fraction of times the neuron responded with 0, 1, . . . , 5 spikes for each
of the five tastants. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 303
5.7 Gamma densities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 306
5.8 Exponential densities . . . . . . . . . . . . . . . . . . . . . . . . . . . . 310
5.9 Beta densities . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 314
5.10 Water temperatures (◦ C) at 1000m depth . . . . . . . . . . . . . . . . . 316
LIST OF FIGURES viii

5.11 Bivariate Normal density . . . . . . . . . . . . . . . . . . . . . . . . . 323

5.12 Bivariate Normal density . . . . . . . . . . . . . . . . . . . . . . . . . 325
5.13 t densities for four degrees of freedom and the N(0, 1) density . . . . . 334

6.1 Posterior densities of β0 and β1 in the ice cream example using the prior
from Equation 6.4. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 350
6.2 Numbers of pine cones in 1998 as a function of dbh . . . . . . . . . . . . 354
6.3 Numbers of pine cones in 1999 as a function of dbh . . . . . . . . . . . . 355
6.4 Numbers of pine cones in 2000 as a function of dbh . . . . . . . . . . . . 356
6.5 10,000 MCMC samples of the Be(5, 2) density. Top panel: histogram of
samples from the Metropolis-Hastings algorithm and the Be(5, 2) den-
sity. Middle panel: θi plotted against i. Bottom panel: p(θi ) plotted
against i. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 361
6.6 10,000 MCMC samples of the Be(5, 2) density. Left column: (θ∗ | θ) =
U(θ − 100, θ + 100); Right column: (θ∗ | θ) = U(θ − .00001, θ + .00001).
Top: histogram of samples from the Metropolis-Hastings algorithm and
the Be(5, 2) density. Middle: θi plotted against i. Bottom: p(θi ) plotted
against i. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 363
6.7 Trace plots of MCMC output from the pine cone code on page 365. . . . 367
6.8 Trace plots of MCMC output from the pine cone code with a smaller
proposal radius. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 368
6.9 Trace plots of MCMC output from the pine cone code with a smaller
proposal radius and 100,000 iterations. The plots show every 10’th it-
eration. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 370
6.10 Trace plots of MCMC output from the pine cone code with proposal
function g.one and 100,000 iterations. The plots show every 10’th it-
eration. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 371
6.11 Pairs plots of MCMC output from the pine cones example. . . . . . . . . 373
6.12 Trace plots of MCMC output from the pine cone code with proposal
function g.group and 100,000 iterations. The plots show every 10’th
iteration. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 375
6.13 Pairs plots of MCMC output from the pine cones example with proposal
g.group. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 376
6.14 Posterior density of β2 and γ2 from Example 6.3. . . . . . . . . . . . . . 377

7.1 Plots of the Orthodont data: distance as a function of age, grouped by

Subject, separated by Sex. . . . . . . . . . . . . . . . . . . . . . . . . . 384
7.2 Plots of the Orthodont data: distance as a function of age, separated by
Subject. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 387
LIST OF FIGURES ix

7.3 Percent body fat of major (blue) and minor (purple) Pheidole morrisi
ants at three sites in two seasons. . . . . . . . . . . . . . . . . . . . . . 391
7.4 Residuals from Model 7.4. Each point represents one colony. There is
an upward trend, indicating the possible presence of colony effects. . . . 393
7.5 Some time series . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 398
7.6 Yt+1 vs. Yt for the Beaver and Presidents data sets . . . . . . . . . . . . 399
7.7 Yt+k vs. Yt for the Beaver data set and lags 0–5 . . . . . . . . . . . . . . 400
7.8 coplot of Yt+1 ∼ Yt−1 | Yt for the Beaver data set . . . . . . . . . . . . . 402
7.9 Fit of CO2 data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 405
7.10 DAX closing prices . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 407
7.11 DAX returns . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 408
7.12 Survival curve for bladder cancer. Solid line for placebo; dashed line
for thiotepa. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 412
7.13 Cumulative hazard and log(hazard) curves for bladder cancer. Solid
line for thiotepa; dashed line for placebo. . . . . . . . . . . . . . . . . 415

8.1 Mean Squared Error for estimating Binomial θ. Sample size = 5, 20,
100, 1000. α = β = 0: solid line. α = β = 0.5: dashed line. α = β = 1:
dotted line. α = β = 4: dash–dotted line. . . . . . . . . . . . . . . . . . 427
8.2 The Be(.39, .01) density . . . . . . . . . . . . . . . . . . . . . . . . . . . 437
8.3 Densities of Ȳin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 438
8.4 Densities of Zin . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 440
8.5 The δ-method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 444
8.6 Top panel: asymptotic standard deviations of δn and δ0n for Pr[X ≤ a].
The solid line shows the actual relationship. The dotted line is the line
of equality. Bottom panel: the ratio of asymptotic standard deviations. . 448
 List of Tables

1.1 Party Affiliation and Referendum Support . . . . . . . . . . . . . . . . 42

1.2 Steroid Use and Test Results . . . . . . . . . . . . . . . . . . . . . . . 43

2.1 New and Old seedlings in quadrat 6 in 1992 and 1993 . . . . . . . . . . 151

3.1 Correspondence between Models 3.3 and 3.4 . . . . . . . . . . . . . . . 214

3.2 β’s for Figure 3.14 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 237

5.1 Rutherford and Geiger’s data . . . . . . . . . . . . . . . . . . . . . . . 297

6.1 The numbers of pine cones on trees in the FACE experiment, 1998–2000. 352

7.1 Fat as a percentage of body weight in ant colonies. Three sites, two
seasons, two castes. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 395

x
 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 emphasis” is meant that the likelihood function and
likelihood principle are unifying ideas throughout the text. Another unusual aspect is the
use of statistical software 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 Development Core Team [2006]). 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.

xi
 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

1
1.1. BASIC PROBABILITY 2

die to be fair then we would also assign µ({1}) = µ({2}) = · · · = µ({6}) = 1/6. The laws of
probability then imply 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).

1.1. BASIC PROBABILITY 3

Using the language of page 1, what is X in this case? Let d1 denote the number
showing on the first die and d2 denote the number showing on the second die. d1 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, Example 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 ran-
domly 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
1.1. BASIC PROBABILITY 4

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.
> 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 as-
signed to a variable called x. The assignment symbol is two characters: < fol-
lowed 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.
1.1. BASIC PROBABILITY 5

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] 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
1.2. PROBABILITY DENSITIES 6

# 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 Ex-
ample 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.

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, longitude and depth
1.2. PROBABILITY DENSITIES 7

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 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.
1.2. PROBABILITY DENSITIES 8

R b first property holds because, if p(y) < 0 on the interval (a, b) then P[Y ∈ (a, b)] =
The
a
p(y) dy < 0; and we can’t have probabilities less than 0. The second property holds
R∞
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.

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:
1.2. PROBABILITY DENSITIES 9

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

• 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. Figure 1.3 shows how that
works. The upper panel of the figure is a histogram of 112 measurements of ocean tem-
perature at a depth of 1000 meters in the North Atlantic near 45◦ North latitude and 20◦
degrees West longitude. Example 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
1.2. PROBABILITY DENSITIES 10

temperature at that location is like randomly drawing a value from f . The 112 measure-
ments, 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 oceanog-
raphers 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 dis-
coveries 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 approxi-
mately. 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 figures in a 2 by 1 layout.
1.2. PROBABILITY DENSITIES 11

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 12

• 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 latitude 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 differentiable, mono-
tonic, invertible function and define Z = g(X). Then the pdf of Z is

d g−1 (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 ∈ (g−1 (a), g−1 (b)]]
db db
Z g−1 (b)
d d g−1 (t)
= pX (x) dx = × pX (g−1 (b))
db g−1 (a) dt b

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

1.2. PROBABILITY DENSITIES 13

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,

d g−1 (z) 2 1 2
pZ (z) = pX (g−1 (z)) = − 2 = 3
dz z z z

And the possible values of Z are from 1 to ∞. So pZ (z) = 2/z3 on the interval (1, ∞). As a
partial check, we can verify that the integral is 1.
Z ∞
2 1 ∞
3
dz = − = 1.
1 z z2 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 I x . It gets mapped by g into a
little interval around z; call it Iz . The density is

P[Z ∈ Iz ] P[X ∈ I x ] length(I x )

pZ (z) ≈ = ≈ pX (x) |h0 (z)| (1.3)
length(Iz ) length(I x ) length(Iz )

The approximations in Equation 1.3 are exact as the lengths of I x and Iz decrease 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 14

1.3 Parametric Families of Distributions

Probabilities often depend on one or more unknown numerical constants. Suppose, 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 numerical constant. In this case we have a
family of probability measures, one 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.
1.3. PARAMETRIC FAMILIES OF DISTRIBUTIONS 15

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 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 Nk is called a binomial coefficient and is read “N choose k”. Nk = k!(N−k)! N!
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.

1.3. PARAMETRIC FAMILIES OF DISTRIBUTIONS 16

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

1.3. PARAMETRIC FAMILIES OF DISTRIBUTIONS 17

Let X be the number of successes. 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 Binomial prob-
abilities. 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.

1.3.2 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 mathematician 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 genome.

1.3. PARAMETRIC FAMILIES OF DISTRIBUTIONS 18

Traffic flow Cars arrive at an intersection.

Customer service Customers arrive at a service counter.

Neurobiology Neurons fire.

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 in-
stance 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 seed disper-
sal, 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 previous 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
1.3. PARAMETRIC FAMILIES OF DISTRIBUTIONS 19

P[X = 3 | λ] for different values of λ. For example,

13 e−1
P[X = 3 | λ = 1] = ≈ 0.06
3!
3 −2
2e
P[X = 3 | λ = 2] = ≈ 0.18
3!
33 e−3
P[X = 3 | λ = 3] = ≈ 0.22
3!
3 −4
4e
P[X = 3 | λ = 4] = ≈ 0.20
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.
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

1.3. PARAMETRIC FAMILIES OF DISTRIBUTIONS 20

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.

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.
1.3. PARAMETRIC FAMILIES OF DISTRIBUTIONS 21

A useful pdf for such situations is the Exponential density

1 − λx
p(x) = e for x > 0. (1.7)
λ
We say X has an exponential distribution with parameter λ and write X ∼ Exp(λ). Fig-
ure 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

x
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
1.3. PARAMETRIC FAMILIES OF DISTRIBUTIONS 22

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 ),
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 coordi-
nates 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
1.3. PARAMETRIC FAMILIES OF DISTRIBUTIONS 23

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 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 y−µ 2
e− 2 (
1
p(y) = √ 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.
1.3. PARAMETRIC FAMILIES OF DISTRIBUTIONS 24

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 25

As a further illustration, Figure 1.9 shows a histogram of 105 ocean temperatures (◦ 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.
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.

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 )
1.3. PARAMETRIC FAMILIES OF DISTRIBUTIONS 26

lines ( t, dnorm ( t, mean(y), sd(y) ) )

• 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 distribu-

tion. n is the sample size; mu is the mean; and sig is the standard deviation. To demon-
strate, 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,
1.3. PARAMETRIC FAMILIES OF DISTRIBUTIONS 27

(a)

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

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) density.
(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.
1.3. PARAMETRIC FAMILIES OF DISTRIBUTIONS 28

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 temperatures 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 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 standard 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
1.4. CENTERS, SPREADS, MEANS, AND MOMENTS 29

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) distribution is
called the standard Normal distribution.

1.4 Centers, Spreads, Means, and Moments

Recall Figure 1.3 (pg. 11). 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.
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, pres-
sure, oxygen concentration, and potential vorticity of the world’s oceans. Data about
the oceans’ surface can be collected by satellites’ bouncing signals off the surface. 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 instruments 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 measure-
ments 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
1.4. CENTERS, SPREADS, MEANS, AND MOMENTS 30

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 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 Figure 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 F X and
pdf pX . Then the mean of X (equivalently, the mean of F X ) is
P
R i i P[X = i] if X is discrete

E(X) = 

 x p (x) dx if X is continuous (1.10)
X

1.4. CENTERS, SPREADS, MEANS, AND MOMENTS 31

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 32

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
1.4. CENTERS, SPREADS, MEANS, AND MOMENTS 33

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 p j (1 − p)n−1− j
j=0
j!(n − 1 − j)!
= np

The first five equalities are just algebra. The sixth is worth remembering. The sum
Pn−1
j=0 · · · is the sum of the probabilities of the Bin(n − 1, p) 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 19.

Exponential If X ∼ Exp(λ) then

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

Use integration by parts.

Normal If X ∼ N(µ, σ) then E(X) = µ. The derivation is left as Exercise 19.

1.4. CENTERS, SPREADS, MEANS, AND MOMENTS 34

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 σ2Y (σY ). The variances of
common distributions will be derived later in the book.
Caution: for reasons which we don’t go into here, many books define the variance of a
sample as Var(y1 , . . . , yn ) = (n − 1)−1 (yi − ȳ)2 . For large n there is no practical difference
P
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 interpretation
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 .

1.4. CENTERS, SPREADS, MEANS, AND MOMENTS 35

Proof.

Var(Y) = E((Y − EY)2 )

= E(Y 2 − 2YEY + (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 Fig-
ure 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 sym-
metric, 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.
To illustrate further, we’ll calculate the SD of a few mound-shaped random variables
and compare the SD’s to the pdf’s.
1.4. CENTERS, SPREADS, MEANS, AND MOMENTS 36

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
1.4. CENTERS, SPREADS, MEANS, AND MOMENTS 37

Binomial Let Y ∼ Bin(30, .5).

Var(Y) = E(Y 2 ) − (EY)2

30 !
2 30
X
= y .530 − 152
y=0
y
30
X 30!
= y .530 − 152
y=1
(y − 1)!(30 − y)!
29
X 29!
= 15 (v + 1) .529 − 152 (1.12)
v=0
v!(29 − v)!
 29 29

X 29! X 29!
= 15  .529 + .529  − 152
 
v
v=0
v!(29 − v)! v=0
v!(29 − v)!
!
29
= 15 + 1 − 15
2
15
=
2
√
and therefore SD(Y) = 15/2 ≈ 2.7. (See Exercises 20 and 21.)

Normal Let Y ∼ N(0, 1).

Var(Y) = E(Y 2 )
Z ∞ 2
y y2
= √ e− 2 dy
−∞ 2π
Z ∞ 2
y y2
=2 √ e− 2 dy (1.13)
Z0 ∞ 2π 2
1 y
=2 √ e− 2 dy
0 2π
=1

and therefore SD(Y) = 1. (See Exercises 20 and 21.)

Figure 1.14 shows the comparison. The top panel shows the pdf of the Bin(30, .5) distri-
bution; the bottom panel shows the N(0, 1) distribution.
1.4. CENTERS, SPREADS, MEANS, AND MOMENTS 38

0.15
0.10
p(y)

0.05

+/− 1 SD
0.00

+/− 2 SD’s

0 5 10 15 20 25 30

y
0.0 0.1 0.2 0.3 0.4
p(y)

+/− 1 SD

+/− 2 SD’s

−3 −2 −1 0 1 2 3

Figure 1.14: Two pdf’s with ±1 and ±2 SD’s. top panel: Bin(30, .5); bottom panel: N(0, 1).
1.4. CENTERS, SPREADS, MEANS, AND MOMENTS 39

Figure 1.14 was produced by the following R code.

par ( mfrow=c(2,1) )

y <- 0:30
sd <- sqrt ( 15 / 2 )
plot ( y, dbinom(y,30,.5), ylab="p(y)" )
arrows ( 15-2*sd, 0, 15+2*sd, 0, angle=60, length=.1,
code=3, lwd=2 )
text ( 15, .008, "+/- 2 SD’s" )
arrows ( 15-sd, .03, 15+sd, .03, angle=60, length=.1,
code=3, lwd=2 )
text ( 15, .04, "+/- 1 SD" )

y <- seq(-3,3,length=60)
plot ( y, dnorm(y,0,1), ylab="p(y)", type="l" )
arrows ( -2, .02, 2, .02, angle=60, length=.1, code=3, lwd=2 )
text ( 0, .04, "+/- 2 SD’s" )
arrows ( -1, .15, 1, .15, angle=60, length=.1, code=3, lwd=2 )
text ( 0, .17, "+/- 1 SD" )

• arrows(x0, y0, x1, y1, length, angle, code, ...) adds arrows to a
plot. See the documentation for the meaning of the arguments.

• text adds text to a plot. See the documentation for the meaning of the arguments.

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 applicability.
R has built-in functions to compute means and variances and can compute other mo-
ments easily. Note that R uses the divisor n − 1 in its definition of variance.
# Use R to calculate moments of the Bin(100,.5) distribution
1.4. CENTERS, SPREADS, MEANS, AND MOMENTS 40

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.

Let h be a function. Then E[h(Y)] = h(y)p(y) dy ( h(y)p(y) in the discrete case) is

R P
the expected value of h(Y) and is called a generalized moment. There are sometimes two
ways to evaluate E[h(Y)]. One is Rto evaluate the integral. 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] = xp x (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 y fy (y) dy

= a + bEY

1.5. JOINT, MARGINAL AND CONDITIONAL PROBABILITY 41

There is a corresponding theorem for variance.

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

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[b2 (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 Republicans 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 probabil-
ity 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 summa-
rizes the calculations. The quantities .60, .40, .62, and .38 are called marginal probabili-
ties. The name derives from historical reasons, because they were written in the margins
of the table. Marginal probabilities are probabilities for one variable alone, the ordinary
probabilities that we’ve been talking about all along.
1.5. JOINT, MARGINAL AND CONDITIONAL PROBABILITY 42

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

Table 1.1: Party Affiliation and Referendum Support

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).
These calculations illustrate a general principle: To get a marginal probability 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) (1.14)
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 conditionals; some-
times 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
1.5. JOINT, MARGINAL AND CONDITIONAL PROBABILITY 43

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

Table 1.2: Steroid Use and Test Results

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.

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 fa-
miliar with the concepts and notation in terms of marginal, conditional and joint distribu-
tions, and not to rely too heavily on the tabular representation because in more complicated
problems there is no convenient tabular representation.
Example 1.6 is a further illustration of joint, conditional, and marginal distributions.
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
1.5. JOINT, MARGINAL AND CONDITIONAL PROBABILITY 44

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 1.15 shows the possible values
of the pair (N, X). The probabilities associated with each of the points in Figure 1.15
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.

...
6
4
X

...
2
0

0 1 2 3 4 5 6 7

Figure 1.15: Permissible values of N and X, the number of new seedlings and the number
that survive.

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
The e−λ λ3 /6 is a marginal probability like the 60% in the affiliation/support problem.
The binomial probabilities above are conditional probabilities like the 80% and 20%;
1.5. JOINT, MARGINAL AND CONDITIONAL PROBABILITY 45

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 1.15 along the horizontal line corresponding to X = x. To get
fX (x) ≡ P[X = x], we must add up all the probabilities on that line.
∞
e−λ λn n x
X 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!
The last equality follows since z · · · = 1 because it is the sum of probabilities from the
P
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 remembering.

For continuous random variables, conditional and joint densities are written pX | Y (x | y)
and pX,Y (x, y) respectively and, analogously to Equation 1.14 we have

pX,Y (x, y) = pX (x)pY | X (y | x) = pY (y)pX | Y (x | y) (1.15)

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.
1.5. JOINT, MARGINAL AND CONDITIONAL PROBABILITY 46

1. First X = x occurs then Y = y occurs. The “probability” of that happening 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 integralsR of the density. If A is a
region in the (x, y) plane, P[(X, Y) ∈ A] = A p(x, y) dx dy, where A . . . indicates a double
R

integral over the region A.

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 that must be integrated to calculate a probability”, on one hand,
P[X ∈ B] = B pX (x) dx. On the other hand,
R

Z Z !
P[X ∈ B] = P[(X, Y) ∈ B × R] = pX,Y (x, y) dy dx
B R

which implies pX (x) = R pX,Y (x, y) dy.

R

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)
1.5. JOINT, MARGINAL AND CONDITIONAL PROBABILITY 47

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

p(x, y)
p(y | x) = = e x−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 ∞
d x+w

= −e−y dx
dw 0 x
Z ∞
d
= e−x (1 − e−w ) dx = e−w
dw 0
Figure 1.16 illustrates the Help Line calculations. For questions 1 and 2, the answer
comes from using Equations 1.15. The only part deserving comment is the limits of in-
tegration. 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 (e) 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
"
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)



(1.16)
Z Z !
E(X) = xpX | Y (x | y) dx pY (y) dy = E (E(X | Y)) . (1.17)
1.5. JOINT, MARGINAL AND CONDITIONAL PROBABILITY 48

(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=
6
0.8

x=
p(y|x)

y!x=w
4

x=
y
0.4

y!x=0
0.0

0 2 4 6 8 10 0 1 2 3 4

y x

Figure 1.16: (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 49

The notation deserves some explanation. For any R number x, pX | Y (x | y) is a function of y,

say g(y). The middle term in Equation 1.16 is g(y)p(y)dy, which equals E(g(y)), which
is the right hand term.
R Similarly, E(X | Y) is a function of Y, say h(Y). The middle term in
Equation 1.17 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 forest
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 1.17 shows that E(X) can be computed as

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

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 loses

Let X = 

1 if shooter wins


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.
1.6. ASSOCIATION, DEPENDENCE, INDEPENDENCE 50

Equation 1.17 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.

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 sepals and petals of 150 iris plants. The first several lines of iris are
1.6. ASSOCIATION, DEPENDENCE, INDEPENDENCE 51

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 1.17 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 1.17 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 correlation.

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 covariances.
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 52

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 1.17: Lengths and widths of sepals and petals of 150 iris plants
1.6. ASSOCIATION, DEPENDENCE, INDEPENDENCE 53

> 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 mea-
sured. A scale-free measure of association would also be useful. Correlation 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 visual 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 measurement
scale.
1.6. ASSOCIATION, DEPENDENCE, INDEPENDENCE 54

Theorem 1.6. Let X and Y be random variables. Then Cor(aX + b, cY + d) = Cor(X, Y).
Proof. See Exercise 47. 
Correlation doesn’t measure all types of association; it only measures clustering around
a straight line. The first two columns of Figure 1.18 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 correla-
tion, the tighter the points cluster 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 pattern 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 random variables, X and Y, are said to be independent 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.
1.6. ASSOCIATION, DEPENDENCE, INDEPENDENCE 55

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 1.18: correlations

1.7. SIMULATION 56

For an example, recall the Help Line story on page 46. 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?
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
e−λ λni e−kλ λ ni
P
Y Y
p(n1 , . . . , nk | λ) = p(ni | λ) = = Q
i=1 i=1
ni ! 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 years had different λ’s, we chose a year at random or
if we thought of Nature as randomly choosing λ for our particular year. In that case the
data from early years, N1 , . . . , Nm , say, yield information about λ and therefore about
likely values of Nm+1 , . . . , Nk , so the Ni ’s are dependent. In fact,
e−kλ λ ni
Z Z P

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 Ni ’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: cal-
1.7. SIMULATION 57

culating 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 easily calculated by
hand, but composed of smaller events that are easily mimicked 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 S 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 ) {

made <- NA
while ( is.na(made) ) { # roll until outcome is determined
1.7. SIMULATION 58

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

if ( roll == point ) {
made <- TRUE
} else if ( roll == 7 ) {
made <- FALSE
}
} # end while
return ( made )
} # end makepoint

sim.craps <- function () {

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 neces-
sary as long as !determined is T.

Try the example code a few times. See whether you get about 49% as Example 1.8 sug-
gests.
Along with the estimate itself, it is useful to estimate the accuracy of µ̂g as an estimate
of µg . If the simulations are independent then Var(µ̂g ) = n−1 Var(g(Y)); if there are many
of them then Var(g(Y)) can be well estimated by n−1 ((g(y) − µ̂g )2 ) and SD(g(Y)) can be
P
1.7. SIMULATION 59

well estimated by
((g(y) − µ̂g )2 ). Because SD’s decrease in proportion to n−1/2 (See the Central
p
−1/2 P
n
Limit Theorem.), 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)
SD(X/n.sim) = (θ)(1 − θ)/n.sim
p

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 1.19. In Figure 1.19 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 1.19 for a further description of boxplots.

Here is the R code for the simulations and Figure 1.19.

1.7. SIMULATION 60

0.7
0.5
0.3

50 200 1000

n.sim

Figure 1.19: 1000 simulations of θ̂ for n.sim = 50, 200, 1000

N <- 1000

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

1.7. SIMULATION 61

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 behaves 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 worth-
while 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.
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
1.7. SIMULATION 62

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 calculation, 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 )
}

sim.2 <- function() {

n.g <- rbinom ( 1, sampsize, prop.g )
1.7. SIMULATION 63

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 1.20 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, repro-
duced here as Figure 1.21. Typing help(co2) gives a brief explanation. The data
are the concentrations of CO2 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
1.7. SIMULATION 64

0.35 0.50 0.65

1 2 3

Figure 1.20: 1000 simulations of θ̂ under three possible procedures for conducting a poll

is a nutrient needed by plants. It is possible that an increase in CO2 will cause an

increase in plant growth which in turn will partly absorb some of the extra carbon.
To learn about plant growth under elevated CO2 , ecologists began by conducting
experiments in greenhouses. In a greenhouse, two sets of plants could be grown un-
der conditions that are identical except for the amount of CO2 in the atmosphere. But
the controlled environment of a greenhouse is quite unlike the uncontrolled natural en-
vironment, 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. CO2 can be added to the air inside
the chamber. Because the chamber is mostly enclosed, not much CO2 will escape,
and more can be added as needed. Some plants can be grown in chambers 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 Research
in the U.S. Department of Energy (DOE) began supporting research using a technol-
ogy called FACE, or Free Air CO2 Enrichment developed at the Brookhaven National
Laboratory. As the lab’s webpage http://www.bnl.gov/face/ explains

“FACE provides a technology by which the microclimate around grow-

1.7. SIMULATION 65

(a) (b)

4
360

2
resids
340
co2

0
!2
320

!4
1960 1980 1960 1980

Time Time

(c) (d)
3

360
2
1
cycle

340
co2
0
!3 !2 !1

320

2 4 6 8 12 1960 1980

Index Time

Figure 1.21: Monthly concentrations of CO2 at Mauna Loa

1.7. SIMULATION 66

ing 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, elevated
[CO2] simulating climate change conditions.
“No containment is required with FACE equipment and there is no sig-
nificant 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 concentra-
tions of atmospheric CO2 expected in the mid-twenty-first century.”
See the website for pictures and more information. In a FACE experiment, CO2 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 atmo-
sphere 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), Switzerland
(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 experiment in a for-
est, 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 partic-
ular, they wanted to demonstrate that the experiment would 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 demon-
stration, 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 exper-
iment 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.
1.7. SIMULATION 67

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, ] <- ...

For subsequent years the biomass in year i is the biomass in year i-1 multiplied
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
1.7. SIMULATION 68

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.
For our purposes here, we have replaced the t-test with a plot, Figure 1.22,
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 )
1.7. SIMULATION 69

2.0
growth rate

1.5
1.0

2 4 6 8 10

year

Figure 1.22: 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 70

M.observed.control [,1,i] <- temp [1:npairs]

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, ]
1.7. SIMULATION 71

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, ] <-
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 ) )
1.8. R 72

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 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
1.8. R 73

• 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 argument, 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 dimen-
sion. 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 in-
tended.

Individual elements of matrices can be accessed by two-dimensional subscripts 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.
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 relationships are
among the variables. Try pairs(bodyfat).
1.8. R 74

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) ) # need just 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) )
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 variables, let’s divide the men into groups accord-
ing 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.
1.8. R 75

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

groups <- cut ( abd, cut.pts, include.lowest=T, 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
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 )

• tapply means "apply to every element of a table." In this case, the table is fat,
grouped according to groups.

• cbind means "bind together in columns". There is an analgous command rbind.

Finally, let’s make a figure similar to Figure 1.12.

1.9. SOME RESULTS FOR LARGE SAMPLES 76

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 cir-
cumference 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 circumference even approximately then (1) we can estimate
his percent body fat more accurately and (2) the typical estimation error is smaller.

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 8.4.

Definition 1.9 (Random Sample). A collection y1 , . . . , yn of random variables 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

2. the yi ’s are mutually independent, i.e. f (y1 , . . . , yn ) =

Qn
i=1 f (yi ).

The collection (y1 , . . . , yn ) is called a data set. We write

y1 , . . . , yn ∼ i.i.d. f

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

Many introductory statistics texts describe how to collect random samples, many pit-
falls that await, and many pratfalls taken in the attempt. We omit that discussion here
1.9. SOME RESULTS FOR LARGE SAMPLES 77

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 theR Current Population Survey.
Suppose y1 , y2 , . . . , ∼ i.i.d. f . Let µ = y f (y) dy and σ2 = (y − µ)2 f (y) dy be the
R

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.
"
E[x1 + x2 ] = (x1 + x2 ) f (x1 , x2 ) dx1 dx2
" "
= 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 ] = σ21 + · · · + σ2n .
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 + 2 (E(X1 X2 ) − µ1 µ2 )
= σ21 + σ22 + 2 (E(X1 X2 ) − µ1 µ2 ) .
1.9. SOME RESULTS FOR LARGE SAMPLES 78

But if X1 ⊥ X2 then
"
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 ) = σ21 + 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 σ2 . Then Var(ȳn ) =
σ2 /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| ≥ ] ≤ σ2 / 2 .
Proof.
Z
σ =
2
(x − µ)2 f (x) dx
Z µ− Z µ+ Z ∞
= (x − µ) f (x) dx +
2
(x − µ) f (x) dx +
2
(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
f (x) dx +  2
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 77.
Theorem 1.12 (Weak Law of Large Numbers). Let y1 , . . . , yn be a random sample from a
distribution with mean µ and variance σ2 . Then for any  > 0,
lim P[|ȳn − µ| < ] = 1. (1.18)
n→∞
1.9. SOME RESULTS FOR LARGE SAMPLES 79

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

lim P[|ȳn − µ| < ] = lim 1 − P[|ȳn − µ| ≥ ] ≥ lim 1 − σ2 /n 2 = 1.

n→∞ n→∞ n→∞


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 random sample from
a distribution with mean µ and variance σ2 . 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 8.4.
Theorem 1.14 (Central Limit Theorem). Let y√ 1 , . . . , yn be a random sample from f with
mean µ and variance σ2 . Let zn = (ȳn − µ)/(σ/ n). Then, for any numbers a < b,
Z b
1
√ e−w /2 dw.
2
lim P[zn ∈ [a, b]] =
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),

1.10. EXERCISES 80

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 1.23. 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 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 1.23. 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.
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 1.23.
The Central Limit Theorem makes three statements about the distribution of ȳn (zn ) in
large samples:
1. E[ȳn ] = µ (E[zn ] = 0),
√
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
1.10. EXERCISES 81

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 1.23: Histograms of craps simulations. Solid curves are Normal approximations
according to the Central Limit Theorem.
1.10. EXERCISES 82

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 than 1030 1’s? Run an R
simulation to estimate the answer.

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) = ky2 for some constant, k.
1.10. EXERCISES 83

(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.
(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. Consider a random variable Y with probability density function (pdf) given by

f (y) = ce−y /2 ,
2
−∞<y<∞

(a) Find c.
(b) Find the expected value and variance of Y.
(c) What is the pdf of Y 2 ?

9. 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 student 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].

10. Sue is studying the Bin(25,.4) distribution. In R she types

1.10. EXERCISES 84

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?

11. 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).
(h) Find E[X | Y = .5].

12. X and Y are uniformly distributed in the unit disk. I.e., the joint density p(x, y) is
constant on the region of R2 such that x2 + y2 ≤ 1.

(a) Find p(x, y).

(b) Are X and Y independent?
(c) Find the marginal densities p(x) and p(y).
(d) Find the conditional densities p(x | y) and p(y | x).
(e) Find E[X], E[X | Y = .5], and E[X | Y = −.5].

13. Verify the claim in Example 1.4 that argmaxλ P[x = 3 | λ] = 3. Hint: differentiate
Equation 1.6.
R
14. (a) p is the pdf of a continuous random variable w. Find R p(s) ds.
1.10. EXERCISES 85
R
(b) Find R
p(s) ds for the pdf in Equation 1.7.

15. 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.

16.
1 1 2
p(y) = √ e− 2 y
2π
R0
is the pdf of a continuous random variable y. Find −∞
p(s) ds.

17. When spun, an unbiased spinner points to some number y ∈ (0, 1]. What is p(y)?

18. Some exercises on the densities of tranformed variables. One of them should illus-
trate the need for the absolute value of the Jacobian.

19. (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.

20. (a) Prove: if X ∼ Bin(n, p) then Var(X) = np(1 − p). Hint: use Theorem 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 Theorem 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.

21. (a) Justify each step of Equation 1.12.

(b) Justify each step of Equation 1.13. (Hint: integrate by parts.)

22. 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?

23. 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.

24. 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 succes-
sive interarrival times are independent of each other. Let T 1 be the arrival time of the
next customer and T 2 be the additional time until the arrival of the second customer.
1.10. EXERCISES 86

(a) What is the joint density of (T 1 , T 2 )?

(b) Let S = T 1 +T 2 , 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 )?

25. 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?

26. If human DNA contains xxx bases, and if each base mutates with probability 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?

27. 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.

28. 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.)

29. 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 some-
where 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?
1.10. EXERCISES 87

30. Let (x, y) have!joint density function f x,y . There are two ways to find Ey. One way
is to evaluate y f x,y (x, y) dxdy. The other
R is to start with the joint density f x,y , find
the marginal density fy , then evaluate y fy (y) dy. Show that these two methods give
the same answer.

31. Prove Theorem 1.3 (pg. 40) in the discrete case.

32. Prove Theorem 1.7 (pg. 77) in the continuous case.

33. 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 x j are independent

(b) xi and y j are independent
(c) yi and y j are independent
(d) xi and yi are independent

34. 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?

35. A blood test is 99 percent effective in detecting a certain disease when the disease
is present. However, the test also yields a false-positive result for 2 percent of the
healthy patients tested, who have no such disease. Suppose 0.5 percent of the popu-
lation has the disease. Find the conditional probability that a randomly tested indi-
vidual actually has the disease given that her test result is positive.

36. A doctor suspects a patient has the rare medical condition DS, or disstaticularia,
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 probability that
the patient has DS.

37. Ecologists are studying salamanders in a forest. There are two types of forest. Type
A is conducive to salamanders while type B is not. They are studying one forest but
don’t know which type it is. Types A and B are equally likely.
1.10. EXERCISES 88

During the study, they randomly sample quadrats. (A quadrat is a square-meter

plot.) In each quadrat they count the number of salamanders. Some quadrats have
poor salamander habitat. In those quadrats the number of salamanders is 0. Other
quadrats have good salamander habitat. In those quadrats the number of salamanders
is either 0, 1, 2, or 3, with probabilities 0.1, 0.3, 0.4, and 0.2, respectively. (Yes,
there might be no salamanders in a quadrat with good habitat.) In a type A forest,
the probability that a quadrat is good is 0.8 and the probability that it is poor is 0.2.
In a type B forest the probability that a quadrat is good is 0.3 and the probability that
it is poor is 0.7.

(a) On average, what is the probability that a quadrat is good?

(b) On average, what is the probability that a quadrat has 0 salamanders, 1 sala-
mander, 2 salamanders, 3 salamanders?
(c) The ecologists sample the first quadrat. It has 0 salamanders. What is the
probability that the quadrat is good?
(d) Given that the quadrat had 0 salamanders, what is the probability that the forest
is type A?
(e) Now the ecologists prepare to sample the second quadrat. Given the results
from the first quadrat, what is the probability that the second quadrat is good?
(f) Given the results from the first quadrat, what is the probability that they find
no salamanders in the second quadrat?

38. Let α ∼ Bern(1/4), β ∼ Bern(1/3), and α ⊥ β. Define λ = α + β. Let y1 and y2 be

conditionally independent given (α, β), each with distribution Poi(λ).

(a) Find p(α, β, y1 ).

(b) Find p(y1 ).
(c) Does y1 have a Poisson distribution?
(d) Find p(y1 , y2 ).
(e) Are y1 and y2 independent?
(f) Find p(y1 , y2 | α).
(g) Are y1 and y2 independent given α?
(h) Find p(y1 , y2 | β).
(i) Are y1 and y2 independent given β?
1.10. EXERCISES 89

Now we observe y1 = 0.

(j) Find p(α, β | y1 = 0).

(k) Are α and β independent, given y1 = 0?

39. Let X and Y be distributed in the triangle with corners (0, 0), (1, 0), and (1, 1). Let
their joint density be p(x, y) = k(x + y) for some constant k.

(a) Find k.
(b) Find p(x) and p(y).
(c) Are X and Y independent? If so, prove. If not, find a pair (x, y) such that
p(x, y) , p(x)p(y).
(d) Find p(x | y) and p(y | x).
(e) Find E[X] and E[Y].
(f) Find E[X | Y = y] and E[Y | X = x].
(g) Plot E[X | Y = y] as a function of y and E[Y | X = x] as a function of x.

40. Let A ∼ N(0, 1). Given A = a, let B and C be independent, each with distribution
N(a, 1). Given A = a, B = b, and C = c, let D ∼ N(b, 1), E ∼ N(c, 1) and D and E
be independent. Prove: D and E are conditionally independent given A = a.

41. Let λ ∼ Exp(1). Given λ, let X and Y be independent, each with distribution
Exp(1/λ).

(a) Find p(λ, x, y), p(x, y), and p(x).

(b) Prove: Z ∞
p(y | x) = p(y | λ)p(λ | x) dλ
0

42. For various reasons, researchers often want to know the number of people 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.
1.10. EXERCISES 90

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

43. 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 target 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.”
1.10. EXERCISES 91

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?
(g) In fact, the total number of hits was x = 122. What do you conclude?

44. 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?

45. 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
1.10. EXERCISES 92

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.

46. The so-called Monty Hall or Let’s Make a Deal problem has caused much conster-
nation over the years. It is named for an old television program. A contestant is
presented with three doors. Behind one door is a fabulous prize; behind the other
two doors are virtually worthless prizes. The contestant chooses a door. The host
of the show, Monty Hall, then opens one of the remaining two doors, revealing one
of the worthless prizes. Because Monty 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
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 cor-
rect.

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

47. Prove Theorem 1.6 (pg. 54).

48. Theorem 1.11, Chebychev’s Inequality, gives an upper bound for how far a random
variable can vary from its mean. Specifically, the theorem states that for any random
variable X, with mean µ and SD σ and for any  > 0, P[|X − µ| ≥ ] ≤ σ2 / 2 . This
exercise investigates the conditions under which the upper bound is tight, i.e., the
conditions under which P[|X − µ| ≥ ] = σ2 / 2 .
1.10. EXERCISES 93

(a) Let µ = 0 and σ = 1. Consider the value  = 1. Are there any random variables
X with mean 0 and SD 1 such that P[|X| ≥ 1] = 1? If not, explain why. If so,
construct such a distribution.
(b) Now consider values of  other than 1. Are there any random variables X with
mean 0 and SD 1, and values of  , 1 such that P[|X − µ| ≥ ] = σ2 / 2 ?
(c) Now let µ and σ be arbitrary and let  = σ. Are there any continuous random
variables X with mean µ and SD σ such that P[|X − µ| ≥ ] = σ2 / 2 ? If not,
explain. If so, show how to construct one.
(d) Are there any discrete random variables X with mean µ and SD σ such that
P[|X − µ| ≥ ] = σ2 / 2 ? If not, explain. If so, show how to construct one.
(e) Let µ and σ be arbitrary and let  , σ. Are there any random variables X with
mean µ and SD σ such that P[|X − µ| ≥ ] = σ2 / 2 ? If not, explain. If so, show
how to construct one.

49. (a) State carefully the Central Limit Theorem for a sequence of i.i.d.random vari-
ables.
(b) Suppose X1 , . . . , X100 ∼ iid U(0, 1). What is the standard deviation of X̄, the
mean of X1 , . . . , X100 ?
(c) Use the Central Limit Theorem to find approximately the probability that the
average of the 100 numbers chosen exceeds 0.56.
 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 ex-
periment 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 R such as
“Since (y1 , . . . , yn ) turned out to be . . . it seems that f is likely to be . . . ”, or “ y f (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 populations 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 anything
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 numeri-
cal, through calculating sample means, SD’s, etc. Displaying a few simple features

94
2.2. DATA DESCRIPTION 95

of the data y1 , . . . , yn can allow us to visualize those same features of f . Data de-
scription requires few a priori assumptions about f .

Likelihood In likelihood inference we assume that f is a member of a parametric 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 mea-
sures 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.

Hypothesis Testing Sometimes we want to test hypotheses like Head Start is good for
kids or lower taxes are good for the economy or the new treatment is better than the
old.

Decision Making Often, decisions have to be made on the basis of what we have learned
about f . In addition, making good decisions requires accounting 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 de-
scription 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 measurements of
2.2. DATA DESCRIPTION 96

temperatures from each of 9 locations. The measurements from each location were sum-
marized by the sample mean ȳ = n−1 yi ; comparisons of the 9 sample means helped
P
oceanographers deduce the presence of the Mediterranean tongue. Similarly, the essen-
tial 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. Important ex-
amples of location statistics include

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

P

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 109). 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
2.2. DATA DESCRIPTION 97

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 q p (y1 , . . . , yn ), or simply as q p if the data set is clear from
the context. With only a finite sized sample q p (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−1i ≡ y(i+1) .
i
4. For p ∈ ( n−1 , n−1
i+1
) let q p be any number in the interval (q n−1i , q n−1
i+1 ).

If p is a “nice” number then q p 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
R can compute quantiles. When faced with p ∈ ( n−1 , n−1
i+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.

Figure 2.1 was created with the following R code.

par ( mfrow=c(2,2) )
quant <- c ( .05, .5, .9 )
nquant <- length(quant)

y <- rgamma ( 100, 3 )

stripchart( y, method="jitter", pch=1, xlim=c(0,10),
xlab="y", main="a")
abline ( v=quantile(y,quant) )
points ( x=quantile(y,quant), y=rep(.5,nquant), pch=19 )
2.2. DATA DESCRIPTION 98

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. The open squares
are the numbers .05, .5, and .9 on the vertical axis. Panels a and b are for a sample; panels c
and d are for a distribution.
2.2. DATA DESCRIPTION 99

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 )

• plot.ecdf plots the empirical cumulative distribution function. Here the

word “empirical” means that the cdf comes from a sample, as opposed to the-
oretical probability calculations.

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
2.2. DATA DESCRIPTION 100

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
(yi − ȳ)2
P
2
s ≡
n
Note: some statisticians prefer
(yi − ȳ)2
P
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 s2 .
interquartile range The interquartile range is q.75 − q.25
Presenting a low dimensional statistic is useful if we believe that the statistic is repre-
sentative 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 temperatures, one temperature for each moment in time. That popula-
tion 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 sample means are estimates of the nine climato-
logical 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 statistician 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 displaying 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. DATA DESCRIPTION 101

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 use histograms to display the full distribution 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 they
are. This example uses 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. Column 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] )
2.2. DATA DESCRIPTION 102

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 103

VC, 0.5 VC, 1 VC, 2

3.0

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).
2.2. DATA DESCRIPTION 104

VC, 0.5 VC, 1 VC, 2

3.0

2.0
4
3
1.5

1.0
2
1
0.0

0.0
0 10 25 0 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 105

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 horizon-
tally.) 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 2.4 repeats Figure 2.3 but us-
ing 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 de-
livery method less so. A further analysis could try to be more quantitative: what 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
2.2. DATA DESCRIPTION 106

“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 com-
pare 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 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.

2.2. DATA DESCRIPTION 107

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 ∞
1
fˆ(x) dx = 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 differs
R from g0 by a rescaling
of the horizontal axis; the factor h compensates to make g = 1. Now define the
−1

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.

Figure 2.5 was produced with

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

2.2. DATA DESCRIPTION 108

(a) (b)

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.
2.2. DATA DESCRIPTION 109

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 g0 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 illustrate stripcharts,
also called dotplots, an alternative to histograms. Panel (a) has three rows of points corre-
sponding 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
2.2. DATA DESCRIPTION 110

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?

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 simultaneously, 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 Univer-
sity. 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
2.2. DATA DESCRIPTION 111

(a)

2
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
2.2. DATA DESCRIPTION 112

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 class
had quiz averages between about 5 and 9 but that 4 averages were much lower.

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. But 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 left panel which was
2.2. DATA DESCRIPTION 113

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 114

constructed from the n = 213 points x1 , . . . , x213 taken near (45, -40). Those points are
sorted, from smallest to largest, to create the order 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 approximately 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 )

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 between several random variables that are 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 admission to
UC Berkeley graduate school.
2.2. DATA DESCRIPTION 115

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

2.2. DATA DESCRIPTION 116

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 graduate 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.

, , 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
2.2. DATA DESCRIPTION 117

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 displayed.
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 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
2.2. DATA DESCRIPTION 118

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. (A beautiful example of a mosaic plot is
on the cover of CHANCE magazine in Spring, 2007.) 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 equiv-
alent 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 116 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 rough
equality except for department A which admitted women at a higher rate than men.
2.2. DATA DESCRIPTION 119

Student admissions at UC Berkeley

Admitted Rejected

Male
Gender
Female

Admit

Figure 2.9: Mosaic plot of UCBAdmissions

2.2. DATA DESCRIPTION 120

Student admissions at UC Berkeley

Male Female

Admitted
Admit
Rejected

Gender

Figure 2.10: Mosaic plot of UCBAdmissions

2.2. DATA DESCRIPTION 121

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 expla-
nations; 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 stripcharts and density estimates of durations
and intervals. These are shown in Figure 2.11. The figure suggests bimodal distri-
butions. For duration there seems to be one bunch of data around two minutes and
2.2. DATA DESCRIPTION 122

another around four or five minutes. For interval, the modes are around 50 minutes
and 80 minutes. A plot of interval versus duration, Figure 2.12, suggests that the bi-
modality is present in the joint distribution of the two variables. Because the data were
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 po-
tentially 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 ( mfcol=c(2,2) )
stripchart ( eruptions, method="jitter", pch=1, xlim=c(1,6),
xlab="duration (min)", main="(a)" )
plot ( density ( eruptions ), type="l", xlim=c(1,6),
xlab="duration (min)", main="(b)" )
stripchart ( waiting, method="jitter", pch=1, xlim=c(40,100),
xlab="waiting (min)", main="(c)" )
plot ( density ( waiting ), type="l", xlim=c(40,100),
xlab="waiting (min)", main="(d)" )

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" )
2.2. DATA DESCRIPTION 123

(a) (c)

1 2 3 4 5 6 40 60 80 100

duration (min) waiting (min)

(b) (d)
0.5

0.03
0.4
0.3
Density

Density

0.02
0.2

0.01
0.1

0.00
0.0

1 2 3 4 5 6 40 60 80 100

duration (min) waiting (min)

Figure 2.11: Old Faithful data: duration of eruptions and waiting time between eruptions.
Stripcharts: (a) and (c). Density estimates: (b) and (d).
2.2. DATA DESCRIPTION 124

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 125

4.5
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
2.2. DATA DESCRIPTION 126

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
2.2. DATA DESCRIPTION 127

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 relationship among
three variables. They represent the ocean temperature data from Example 1.5. In Fig-
ure 2.15 there are six panels in which temperature is plotted against latitude. Each panel is
made from the points in a restricted range of longitude. 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 temperature
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 left, 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.
2.2. DATA DESCRIPTION 128

Figure 2.16 reverses the roles of latitude and longitude. The impression is that temper-
ature 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 )

Example 2.6 shows one way to display the relationship between two sequences of
events.
Example 2.6 (Neurobiology)
To learn how the brain works, neurobiologists implant electrodes into animal brains.
These electrodes are fine enough to record the firing times of individual neurons. A
sequence of firing times of a neuron is called a spike train. Figure 2.17 shows the spike
train from one neuron in the gustatory cortex of a rat while the rat was in an experiment
on taste. This particular rat was in the experiment for a little over 80 minutes. Those
minutes are marked on the y-axis. The x-axis is marked in seconds. Each dot on the
plot shows a time at which the neuron fired. We can see, for example, that this neuron
fired about nine times in the first five seconds, then was silent for about the next ten
seconds. We can also see, for example, that this neuron undergoes some episodes of
very rapid firing lasting up to about 10 seconds.
Since this neuron is in the gustatory cortex — the part of the brain responsible for
taste — it is of interest to see how the neuron responds to various tastes. During the
experiment the rat was licking a tube that sometimes delivered a drop of water and
sometimes delivered a drop of water in which a chemical, or tastant, was dissolved.
The 55 short vertical lines on the plot show the times at which the rat received a drop
of 300 millimolar (.3 M) solution of NaCl. We can examine the plot for relationships
between deliveries of NaCl and activity of the neuron.

Figure 2.17 was produced by

datadir <- "~/research/neuro/data/stapleton/"

spikes <- list (
2.2. DATA DESCRIPTION 129

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

2.2. DATA DESCRIPTION 130

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

2.2. DATA DESCRIPTION 131

a spike train
13 18 24 29 34 39 44 50 55 60 65 70 76 81

| | | ||
|||||

|| | | |
minutes

| |||

| ||||

||

||||
||| || | | || |
8

|||||
4

| | | || | | || |
0

0 10 20 30 40 50 60

seconds

Figure 2.17: Spike train from a neuron during a taste experiment. The dots show the times
at which the neuron fired. The solid lines show times at which the rat received a drop of a
.3 M solution of NaCl.
2.2. DATA DESCRIPTION 132

sig002a = scan ( paste ( datadir, "sig002a.txt", sep="" ) ),

sig002b = scan ( paste ( datadir, "sig002b.txt", sep="" ) ),
sig002c = scan ( paste ( datadir, "sig002c.txt", sep="" ) ),
sig003a = scan ( paste ( datadir, "sig003a.txt", sep="" ) ),
sig003b = scan ( paste ( datadir, "sig003b.txt", sep="" ) ),
sig004a = scan ( paste ( datadir, "sig004a.txt", sep="" ) ),
sig008a = scan ( paste ( datadir, "sig008a.txt", sep="" ) ),
sig014a = scan ( paste ( datadir, "sig014a.txt", sep="" ) ),
sig014b = scan ( paste ( datadir, "sig014b.txt", sep="" ) ),
sig017a = scan ( paste ( datadir, "sig017a.txt", sep="" ) )
)
tastants <- list
(
MSG100 = scan
( paste ( datadir, "MSG100.txt", sep="" ) ),
MSG300 = scan
( paste ( datadir, "MSG300.txt", sep="" ) ),
NaCl100 = scan
( paste ( datadir, "NaCl100.txt", sep="" ) ),
NaCl300 = scan
( paste ( datadir, "NaCl300.txt", sep="" ) ),
water = scan
( paste ( datadir, "water.txt", sep="" ) )
)
stripchart ( spikes[[8]] %% 60 ~ spikes[[8]] %/% 60, pch=".",
main="a spike train", xlab="seconds", ylab="minutes" )
points ( tastants$NaCl300 %% 60, tastants$NaCl300 %/% 60 + 1,
pch="|" )

• The line datadir <- ... stores the name of the directory in which I keep the
neuro data. When used in paste it identifies individual files.

• The command list() creates a list. The elements of a list can be anything. In this
case the list named spikes has ten elements whose names are sig002a, sig002b,
. . . , and sig017a. The list named tastants has five elements whose names are
MSG100, MSG300, NaCl100, NaCl300, and water. Lists are useful for keeping
related objects together, esepecially when those objects aren’t all of the same type.

• Each element of the list is the result of a scan(). scan() reads a file and stores the
result in a vector. So spikes is a list of ten vectors. Each vector contains the firing
times, or a spike train, of one neuron. tastants is a list of five vectors. Each vector
contains the times at which a particular tastant was delivered.
2.3. LIKELIHOOD 133

• There are two ways to refer an element of a list. For example, spikes[[8]] refers
to the eighth element of spikes while tastants$NaCl300 refers to the element
named NaCl300.

• Lists are useful for keeping related objects together, especially when those objects
are not the same type. In this example spikes$sig002a is a vector whose length is
the number of times neuron 002a fired, while the length of spikes$sig002b is the
number of times neuron 002b fired. Since those lengths are not the same, the data
don’t fit neatly into a matrix, so we use a list instead.

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 prob-
ability distributions 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 family, 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 (Exam-
ples 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. 19) is a plot of the likelihood function.
2.3. LIKELIHOOD 134

In a typical problem we know the data come from a parametric family indexed 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 func-
tion 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 exam-
ple, `(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
NR, R, NR, NR, NR, R, NR, NR, NR, R
and records X = 3. She did a Binomial
  experiment; her statistical model is X ∼ Bin(10, θ);
her likelihood function is `A (θ) = 10
3
θ3 (1 − θ)7 . It is plotted in Figure 2.18. Because only
ratios matter, the likelihood function can be rescaled by any arbitrary positive constant. In
Figure 2.18 it has been rescaled so the maximum is 1. The interpretation of Figure 2.18
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.

Figure 2.18 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).
2.3. LIKELIHOOD 135

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 2.18: Likelihood function `(θ) for the proportion θ of red cars on Campus Drive
2.3. LIKELIHOOD 136

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 / 103 . But since multiplicative 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 2.18 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 = θ3 (1 − θ)7

`C is proportional to `A and `B and hence contains exactly the same information and looks
exactly like Figure 2.18. 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 2.7 (Seedlings, cont.)

Examples 1.4, 1.6, 1.7, and 1.9 reported data from a single quadrat on the number
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
2.3. LIKELIHOOD 137

total of 40 seedlings so the likelihood function was

`(λ) ≡ p(Data | λ)
= p(y1 , . . . , y60 | λ)
Y60
= p(yi | λ)
1
60
Y e−λ λyi
=
1
yi !
∝e λ
−60λ 40

Note that yi ! is a multiplicative factor that does not depend on λ and so is irrelevant
Q
to `(λ). Note also that `(λ) depends only on yi , not on the individual yi ’s. I.e., we
P
only need to know yi = 40; we don’t need to know the individual yi ’s. `(λ) is plotted
P
in Figure 2.19. Compare to Figure 1.6 (pg. 19). Figure 2.19 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

Figure 2.19: `(θ) after yi = 40 in 60 quadrats.

P
2.3. LIKELIHOOD 138

Figure 2.19 was created with

lam <- seq ( 0, 2, length=50 )

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 2.8 (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 members,
. . . , [and] the number of person-years in terms of National Cancer Institute
statistics showing the annual rate of invasive cancer in American 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 2.20. 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.
2.3. LIKELIHOOD 139

likelihood

0.00 0.05 0.10 0.15 0.20

Figure 2.20: Likelihood for Slater School

Figure 2.20 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" )

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 2.7 and 2.8 show how likelihood functions are used. They reveal which val-
ues 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 hypoth-
2.3. LIKELIHOOD 140

esis 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 under-
standing 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 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 around 8 is like tossing three consecutive
Heads; a likelihood ratio around 1000 is like tossing ten consecutive Heads.
In Example 2.8 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. Assume 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) = f (Xi | µ, σ). But we don’t know f and can’t calculate `(µ, σ).
Q
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
2.3. LIKELIHOOD 141

3. We can estimate σ2 from the data by

X
σ̂2 = s2 = δ2i /n

4. And therefore the function

 1 µ − X̄ 2 
 !
` M (µ) ∝ exp − √  (2.4)
2 σ̂/ n 

is a good approximation to the likelihood function.

In the preceding reasoning we separated the data into two parts — X̄ and {δi }; used {δi } to
estimate σ; and used X̄ to find a likelihood function for µ. We cannot, in general, justify
such a separation mathematically. We justified it if and when our main interest is in µ and
we believe {δi } tell us little about µ.
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 2.9 (Slater School, continued)
We redo the Slater School example (Example 2.8) to illustrate the marginal likelihood
and see how it compares to the exact likelihood. In that example the Xi ’s were 1’s and
0’s indicating which teachers got cancer. There were 8 1’s out of 145 teachers, so
X̄ = 8/145 ≈ .055. Also, σ̂2 = (8(137/145)2 + 137(8/145)2 ))/145 ≈ .052, so σ̂ ≈ .23.
We get
 1 µ − .055 2 
 !
` M (µ) ∝ exp − √  (2.5)
2 .23/ 145
Figure 2.21 shows the marginal and exact likelihood functions. The marginal likelihood
is a reasonably good approximation to the exact likelihood.

Figure 2.21 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) )
lik <- lik/max(lik)
2.3. LIKELIHOOD 142

0.8
marginal
likelihood

0.4
0.0 exact

0.00 0.05 0.10 0.15 0.20

Figure 2.21: Marginal and exact likelihoods for Slater School

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 2.10 (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.”
2.3. LIKELIHOOD 143

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 2.22(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 2.22(b)
displays the marginal likelihood function ` M (µ).

Figure 2.22 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)",
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 miss-
ing data point.
2.3. LIKELIHOOD 144

(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 2.22: Marginal likelihood for mean CEO salary

2.3. LIKELIHOOD 145

• 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 2.22(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 2.11 (FACE, continued)
This example continues Example 1.12 about a FACE experiment in Duke Forest.
There were six rings; three were treated with excess CO2 . The dominant canopy
tree in the FACE experiment is pinus taeda, or loblolly pine. Figure 2.23a is a his-
togram 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 2.23a?
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 2.23b 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
2.3. LIKELIHOOD 146

(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 2.23c shows slices through the likelihood
function at σ = .09, .10, and.11, the locations of the dashed lines in Figure 2.23b. 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 2.23c 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 2.23 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 )
lik.09 <- lik[,13] / max(lik[,13])
lik.10 <- lik[,26] / max(lik[,26])
lik.11 <- lik[,38] / max(lik[,38])
2.3. LIKELIHOOD 147

(a) (b)

0.12
5
4

0.11
3

0.10
!
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 2.23: 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 148

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 val-
ues 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 con-
tour 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 2.12 (Quiz Scores, continued)

This example continues Example 2.3 about scores in Statistics 103. Figure 2.7 shows
2.3. LIKELIHOOD 149

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 2.24 shows the likelihood function. The data support values of µ from 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
!

0.9
0.8
0.7

6.8 7.0 7.2 7.4 7.6

µ
Figure 2.24: Likelihood function for Quiz Scores
2.3. LIKELIHOOD 150

Figure 2.24 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 )
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 2.11 and 2.12 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 2.13 illustrates the problem.

Example 2.13 (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.
How shall we model the data? Let NiT be the true number of New seedlings in
2.3. LIKELIHOOD 151

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.

year i, i.e., including those that emerge after the census; and let NiO be the observed
number of seedlings in year i, i.e., those that are counted in the census. As in Exam-
ple 1.4 we model NiT ∼ Poi(λ). Furthermore, each seedling has some chance θ f of
being found in the census. (Nominally θ f is the proportion of seedlings that emerge
before the census, but in fact it may also include a component accounting for the fail-
ure 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 , θ f ). The
data are the NiO ’s; the NiT ’s are not observed. What do the data tell us about the two
parameters (λ, θ f )?
Ignore the Old seedlings for now and just look at 1992 data N1992O
= 0. Dropping
the subscript 1992, the likelihood function is

`(λ, θ f ) = P[N O = 0 | λ, θ f ]
X∞
= P[N O = 0, N T = n | λ, θ f ]
n=0
∞
X
= P[N T = n | λ] P[N O = 0 | N T = n, θ f ]
n=0
∞ (2.6)
X e−λ λn
= (1 − θ f ) n

n=0
n!
 n
∞ e−λ(1−θ f ) λ(1 − θ )
X f
=
n=0
eλθ f n!
=e −λθ f

Figure 2.25a plots log10 `(λ, θ f ). (We plotted log10 ` instead of ` for variety.) The
contour lines are not circular. To see what that means, focus on the curve log10 `(λ, θ f ) =
−1 which runs from about (λ, θ f ) = (2.5, 1) to about (λ, θ f ) = (6, .4). Points (λ, θ f ) 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.
2.3. LIKELIHOOD 152

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 θ f might be as small as .4. To investigate further, Figure 2.25b is similar to 2.25a
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 θ f . In fact, arbitrarily large values of λ coupled with sufficiently small
values of θ f 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
θ f 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 θ f are linked, and the data alone does not tell us much about either
parameter individually.

Figure 2.25(a) was produced with the following snippet.

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

th <- seq ( 0, 1, by=.02 )
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 2.25(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] )
2.3. LIKELIHOOD 153

(a)

0.8
"f

0.4
0.0

0 1 2 3 4 5 6

(b)
0.8
"f

0.4
0.0

0 200 400 600 800 1000

Figure 2.25: Log of the likelihood function for (λ, θ f ) in Example 2.13
2.4. ESTIMATION 154

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 (2.11 and 2.12) in which likelihood contours are
roughly circular and one (2.13) in which they’re not. By far the most common and impor-
tant case is similar to Example 2.11 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 statis-
tic 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 compute a single best guess at a
parameter θ. An informed guess at the value of θ is called an estimate and denoted θ̂.
One way to estimate θ is to find θ̂ ≡ argmax `(θ), the value of θ for which `(θ) is largest
and hence the value of θ that best explains the data. 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 2.8 and Figure 2.20 θ was the rate of cancer occurence and
we calculated `(θ) based on y = 8 cancers in 145 people. Figure 2.20 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 2.8 the likelihood was `(θ) ∝ θ8 (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]
2.4. ESTIMATION 155

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 − y 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 func-
tion, argmax `(θ) = argmax log(`(θ)), so the m.l.e. can 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
Q P
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
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. ESTIMATION 156

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 2.19 and 2.20. 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 2.9 (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 2.20 to see whether this interval makes sense.
The data in Example 2.9 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
2.4. ESTIMATION 157

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 2.19. Example 2.14 illustrates the
point further.
Example 2.14 (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 2.26. 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 2.26 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])
}

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

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

In Figure 2.26 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
2.4. ESTIMATION 158

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 2.26: 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 159

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 1.19 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 esti-
mate θ ≡ E[Y]. We consider two potential estimators, the sample mean θ̂1 = (1/n) yi and
P
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 2.27 shows the result. The figure
suggests that the sampling distributions of both θ̂1 and θ̂2 are centered 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 2.27 was produced by the following snippet.

sampsize <- c ( 4, 16, 64, 256 )
n.sim <- 1000

par ( mfrow=c(2,2) )
for ( i in seq(along=sampsize) ) {
y <- matrix ( rnorm ( n.sim*sampsize[i], 0, 1 ),
2.4. ESTIMATION 160

(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 2.27: 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 161

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 four-fold.

1. An estimator θ̂ is a random variable and has a distribution.

2. Fθ̂ is a guide to estimation accuracy.

3. Statisticians study conditions under which one estimator is better than another.

4. 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 simulations 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 2.28.
2.4. ESTIMATION 162

12 (a) (b)

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

!^ !^

(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

!^ !^

Figure 2.28: Histograms of θ̂, the sample mean, for samples from Bin(n, .1). Dashed line:
kernel density estimate. Dotted line: Normal approximation. (a): n=5; (b): n=10; (c):
n=25; (d): n=100
2.4. ESTIMATION 163

Figure 2.28 was produced by the following snippet.

sampsize <- c ( 5, 10, 25, 100 )

n.sim <- 1000
p.true <- .1
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 approximation for n = 5. But for large n, the
Normal approximation is quite good. That’s the Central Limit Theorem kicking in. For
2.5. BAYESIAN INFERENCE 164

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 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 hap-
pens 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
2
cases when a marginal likelihood argument applies, then `(θ) ≈ exp − 12 σ̂/θ−√Ȳn


(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 dif-
ferent 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
2.5. BAYESIAN INFERENCE 165

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 probability
distributions of stocks they may buy.

• When making business decisions, firms must assess the future probability distribu-
tions of outcomes.

• 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 outcomes under
several alternative treatments.

• At the Slater School, Example 2.8, 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 infor-
mation 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 in-
formation 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 distribu-
tion p(y1 , . . . , yn , θ). Finally, the joint distribution yields p(θ | y1 , . . . , yn ) the conditional
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.
2.5. BAYESIAN INFERENCE 166

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% ac-
curate 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 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]
= (2.9)
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

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

2.5. BAYESIAN INFERENCE 167

Here’s another way to look at the medical screening problem, one that highlights the mul-
tiplicative 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]
.95
! !
1
=
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 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 distributions 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
where c = p(θ)`(θ) dθ is a constant that does not depend on θ. (An integral with respect
R

to θ does not depend on θ; after integration it does not contain θ.) The effect of the constant
2.5. BAYESIAN INFERENCE 168

c is to rescale the function in the numerator so that it integrates to 1. I.e., p(θ | y) dθ = 1.

R

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 c through Equa-

tion 2.10. But it is usually much easier to find c by using Equation 2.11, then setting
c = [ p(θ)`(θ) dθ]−1 and that’s the approach we will adopt throughout most of this book.
R

The next two examples show Bayesian statistics with real data.

Example 2.15 (Seedlings, continued)

Recall the Seedlings examples (1.4, 1.6, 1.7, 1.9, 2.7, and 2.13) 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 2.29 which
is p(λ) = 4λ2 e−2λ . (This is the Gam(3, 1/2) density; see Section 5.5.) Figure 2.29 also
displays the likelihood function p(y | λ) ∝ λ3 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 propor-
tionality. Therefore c in this example must be the constant that appears in the Gamma
density: c = 1/[5! × (1/3)6 ].

In Figure 2.29 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
`(λ) = p(yi | λ) = ∝ e−nλ λ yi
P

yi !
To see what this means in practical terms, Figure 2.30 shows (a): the same prior we used
in Example 2.15, (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.
2.5. BAYESIAN INFERENCE 169

0.5

prior
likelihood
posterior
0.4
0.3
0.2
0.1
0.0

0 1 2 3 4 5
Figure 2.29: Prior, likelihood and posterior densities for λ in the seedlings example after
the single observation y = 3 !
2.5. BAYESIAN INFERENCE 170

2. As n increases the posterior density becomes increasingly peaked and becomes in-
creasingly like `(λ). That’s because as n increases, the amount of information in
the data increases and the likelihood function becomes increasingly peaked. Mean-
while, the prior density remains as it was. Eventually 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 likelihood, looks increasingly like the likelihood.
Another way to look at it is through the loglikelihood log `(λ) = c + log p(λ) +
1 log p(yi | λ). As n → ∞ there is an increasing number of terms in the sum, so the
Pn
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 2.31. Compare to Figure 2.29.

Example 2.16 shows Bayesian statistics at work for the Slater School. See Lavine
[1999] for further analysis.
Example 2.16 (Slater School, cont.)
At the time of the analysis reported in Brodeur [1992] there were two other lines of ev-
idence regarding the effect of power lines on cancer. First, there were some epidemi-
ological 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 size of mag-
netic 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 ap-
preciable 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 138 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
2.5. BAYESIAN INFERENCE 171

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

! !

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

2.5. BAYESIAN INFERENCE 172

prior
likelihood
posterior
3
density

2
1
0

0.0 0.5 1.0 1.5 2.0

Figure 2.31: Prior, likelihood and posterior densities for λ with n = 60, yi = 40.
P
2.5. BAYESIAN INFERENCE 173

below θ ≈ .06. For the sake of argument let’s adopt the prior depicted in Figure 2.32.
Its 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 2.20). 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 2.32.
40

prior
likelihoo
30

posterio
20
10
0

0.00 0.05 0.10 0.15 0.20

Figure 2.32: Prior, likelihood and posterior density for Slater School

Examples 2.15 and 2.16 have the convenient feature that the prior density had the same
form — λa e−bλ in one case and θa (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
2.6. PREDICTION 174

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 observations.
Let y1 , . . . , yn , y f be a sample from p(· | θ). We observe y1 , . . . , yn but not y f , and want a
prediction for y f . There are three common forms that predictions take.
point predictions A point prediction is a single guess for y f . It might be a predictive
mean, predictive median, predictive mode, or any other type of point prediction that
seems sensible.

interval predictions An interval prediction or predictive interval, is an interval of plau-

sible values for y f . A predictive interval is accompanied by a probability. For ex-
ample, we might say that “The interval (0, 5) is a 90% predictive interval for y f ”
which would mean Pr[y f ∈ (0, 5)] = .90. In a given problem there are, for two rea-
sons, 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
probability. 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 distribution for y f . 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 pretend that we do know θ.
Specifically, let’s pretend that we know y1 , . . . , yn , y f ∼ 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 y f that is not already contained in the knowledge of θ. In other words, y1 , . . . , yn and
y f are conditionally independent given θ. In symbols:

p(y f | θ, y1 , . . . , yn ) = p(y f | θ).

Therefore, our prediction should be based on the knowledge of θ alone, not on any aspect
of y1 , . . . , yn .
2.6. PREDICTION 175

A sensible point prediction for y f is ŷ f = −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 y f might take, a middle set of values, or the
highest values. And, of course, the predictive distribution of y f is N(−2, 1). It completely
describes the extent of our knowledge and ability to predict y f .
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(y f | 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 2.33. ŷ f = 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.
0.4
probability

0.2
0.0

0 2 4 6 8 10

Figure 2.33: Plug-in predictive distribution y f ∼ Poi(λ = 2/3) for the seedlings example

There are two sources of uncertainty in making predictions. First, because y f is ran-
dom, 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 more important source of uncer-
2.6. PREDICTION 176

tainty. 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 ig-
nored. 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 , y f given θ provide the full joint distribution of y1 , . . . , yn , y f , θ,
which in turn provides the conditional distribution of y f given y1 , . . . , yn . Specifically,

Z
p(y f | y1 , . . . , yn ) = p(y f , θ | y1 , . . . , yn ) dθ
Z
= p(θ | y1 , . . . , yn )p(y f | θ, y1 , . . . , yn ) dθ (2.14)
Z
= p(θ | y1 , . . . , yn )p(y f | θ) dθ

Equation 2.14 is just the y f marginal density derived from the joint density of (θ, y f ),
all densities being conditional R on the dataRobserved so far. To say it another way, the
predictive density p(y f ) is p(θ, y f ) dθ = p(θ)p(y f | 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 dispersed 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 2.17 (Seedlings, cont.)

Refer to Examples 1.4 and 2.15 about Y , the number of new seedlings emerging
each year in a forest quadrat. Our model is Y ∼ Poi(λ). The prior (page 168) was
p(λ) = 4λ2 e−2λ . Before collecting any data our predictive distribution would be based
2.6. PREDICTION 177

on that prior. For any number y we could calculate

Z
pY f (y) ≡ P[Y f = y] = pY f | Λ (y | λ)pΛ (λ) dλ
λ e 2 2 −2λ
Z y −λ 3
= λ e dλ
y! Γ(3)
23
Z
= λy+2 e−3λ dλ (2.15)
y!Γ(3)
23 Γ(y + 3) 3y+3 y+2 −3λ
Z
= λ e dλ
y!Γ(3)3y+3 Γ(y + 3)
! !3 !y
y+2 2 1
= ,
y 3 3

(We will see in Chapter 5 that this is a Negative Binomial distribution.) Thus for exam-
ple, according to our prior,

!3
2 8
Pr[Y f = 0] = =
3 27
!3
2 1 8
Pr[Y f = 1] = 3 =
3 3 27
etc.

Figure 2.34 displays these probabilities.

In the first quadrat we found y1 = 3 and the posterior distribution (Example 2.15,
pg. 168)
36 5 −3λ
p(λ | y1 = 3) = λe .
5!

So, by calculations similar to Equation 2.15, the predictive distribution after observing
y1 = 3 is
Z
pY f | Y1 (y | y1 = 3) =pY f | Λ (y | λ)pΛ | Y1 (λ | y1 = 3) dλ
! !6 !y (2.16)
y+5 3 1
=
y 4 4
2.7. HYPOTHESIS TESTING 178

So, for example,

!6
3
Pr[Y f = 0 | y1 = 3] =
4
!6
3 1
Pr[Y f = 1 | y1 = 3] = 6
4 4
etc.

Figure 2.34 displays these probabilities.

Finally, when we collected data from 60 quadrats, we found

6243 42 −62λ
p(λ | y1 , . . . , y60 ) = λ e (2.17)
42!
Therefore , by calculations similar to Equation 2.15, the predictive distribution is
!43 !y
y + 42 62
!
1
Pr[Y f = y | y1 , . . . , y60 ] = (2.18)
y 63 63

Figure 2.34 displays these probabilities.

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 noticably 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.

2.7 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 hypothesis 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.

2.7. HYPOTHESIS TESTING 179

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 2.34: Predictive distributions of y f in the seedlings example after samples of size
n = 0, 1, 60, and the plug-in predictive
2.7. HYPOTHESIS TESTING 180

• 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.

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 diver-
sity 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.
2.7. HYPOTHESIS TESTING 181

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 H0 . 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 four 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 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 calculated 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 .
2.7. HYPOTHESIS TESTING 182

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 medication. 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 pressures
are 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

YC,1 , . . . YC,n ∼ i.i.d. fC ; E[YC,i ] = µC ; Var(YC,i ) = σC2

YT,1 , . . . YT,n ∼ i.i.d. fT ; E[YT,i ] = µT ; Var(YT,i ) = σ2T

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

H0 : µT = µC
Ha : µT , 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 w ∼ N(0, σ2w ) under
H0 with σ2w = (σ2T + σC2 )/n. The mean of 0 comes from H0 . The variance σ2w comes from
adding variances of independent random variables. σ2T and σC2 and therefore σ2w 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 H0 .
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 relevant medical character-
istics. 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
2.7. HYPOTHESIS TESTING 183

other receives the old (control). After a prespecified amount of time their blood pressures
are 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 statistical terms is

H0 : p = .5
Ha : p , .5

Let w = Xi . Under H0 , w ∼ Bin(n, .5). To test H0 we plot the Bin(n, .5) distribution and
P
see where w falls on the plot . Figure 2.35 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.
0.08
0.04
0.00

30 40 50 60 70

Figure 2.35: pdf of the Bin(100, .5) distribution

2.7. HYPOTHESIS TESTING 184

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 2.36 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 calculated 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 .
0.08
0.04
0.00

30 40 50 60 70

Figure 2.36: pdfs of the Bin(100, .5) (dots) and N(50, 5) (line) distributions

The following example shows hypothesis testing at work.

2.7. HYPOTHESIS TESTING 185

Example 2.18 (Tooth Growth, continued)

This continues Example 2.1 (pg. 101). Let’s concentrate on a particular dosage, say
dose = 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 , µOJ

Those are the hypotheses in statistical terms.

Because we’re testing a difference in means, we choose our one dimensional sum-
mary statistic to be w = | x̄ − ȳ|. Small values of w support H0 ; 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 ,

 s 
 σ 2
VC + σ 2 
OJ  ,
w ∼ N 0,

 n 

approximately. The statistic w can be calculated, its SD estimated, and its approximate
density plotted as in Figure 2.37. We can see from the Figure, or from the fact that
w/σw ≈ 3.2 that the observed value of w is moderately far from its expected value
under H0 . The data provide moderately strong evidence against H0 .
2.7. HYPOTHESIS TESTING 186

0.15
0.00

!6 !4 !2 0 2 4 6

Figure 2.37: Approximate density of summary statistic w. The black dot is the value of w
observed in the data.

Figure 2.37 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 ]
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.
2.7. HYPOTHESIS TESTING 187

Example 2.19 (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. Ba-
boons live in social groups comprised of several adult males, several adult females,
and juveniles. Researchers followed several groups of baboons periodically over a pe-
riod 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 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 juvenile
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
1
We have slightly modified the data to avoid some irrelevant complications.
2.7. HYPOTHESIS TESTING 188

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 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 ]
2.7. HYPOTHESIS TESTING 189

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 2.38 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 H0 . 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 H0
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 H0 , 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 2.39. The dot in the figure is at 84, the actual number of instances in the full
data set. Figure 2.39 suggests that it is almost impossible that 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 2.38 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 )
2.7. HYPOTHESIS TESTING 190

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 2.38: Number of times baboon father helps own child in Example 2.19. Histograms
are simulated according to H0 . Dots are observed data.
2.7. HYPOTHESIS TESTING 191

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 )
}
200
100
0

50 60 70 80

w.tot

Figure 2.39: Histogram of simulated values of w.tot. The dot is the value observed in the
baboon data set.

Figure 2.39 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] )
2.8. EXERCISES 192

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 on page 107.
(b) Show that the function g(x) defined just after Equation 2.1 is a probability
density. I.e., show that it integrates to 1.

2. This exercise uses the ToothGrowth data from Examples 2.1 and 2.18.

(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. See Example 2.5. 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. Figure 2.12 suggests that the duration of an eruption is a useful
predictor of the waiting time until the next eruption. Figure 2.14 suggests that the
previous waiting time is another potentially useful predictor of the next waiting time.
This exercise explores what can be gained by using the pair (duration, previous
waiting time) to predict the next waiting time.

(a) From Figure 2.11, estimate, visually, how accurately one could predict waiting
time if one knew nothing about the previous duration or the previous waiting
time.
2.8. EXERCISES 193

(b) From Figure 2.12, estimate, visually, how accurately one could predict waiting
time if one knew the duration of the previous eruption.
(c) Plot waiting time as a function of the previous waiting time. You can use a
command such as plot(waiting[-1]˜waiting[-272]). Describe the re-
sult. Estimate, visually, how accurately one could predict waiting time if one
knew the the previous waiting time. Is the previous waiting time a useful pre-
dictor of the next waiting time?
(d) Make a coplot of waiting time as a function of duration, given the previous
waiting time. Does the relationship between waiting time and duration depend
on the length of the previous waiting time? What does that say about the joint
distribution of (waitingi−1 , waitingi , durationi )?
(e) Make a coplot of the next waiting time as a function of the previous waiting
time, given duration. Given duration, are the previous waiting time and the next
waiting time strongly related? What does that say about the joint distribution
of (waitingi−1 , waitingi , durationi )?
(f) If we’re already using duration to predict the next waiting time, will there be
much benefit also to use the previous waiting time to improve the prediction?

4. This exercise relies on data from the neurobiology experiment described in Exam-
ple 2.6.

(a) Download the data from the book’s website.

(b) Reproduce Figure 2.17.
(c) Make a plot similar to Figure 2.17 but for a different neuron and different
tastant.
(d) Write an R function that accepts a neuron and tastant as input and produces a
plot like Figure 2.17.
(e) Use the function from the previous part to look for neurons that respond to
particular tastants. Describe your results.

5. This exercise relies on Example 2.8 about the Slater school. There were 8 cancers
among 145 teachers. Figure 2.20 shows the likelihood function. Suppose the same
incidence rate had been found among more teachers. How would that affect `(θ)?
Make a plot similar to Figure 2.20, but pretending that there had been 80 cancers
among 1450 teachers. Compare to Figure 2.20. What is the result? Does it make
sense? Try other numbers if it helps you see what is going on.
2.8. EXERCISES 194

6. This exercise continues Exercise 42 in Chapter 1. Let p be the fraction of the popu-
lation 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 randomized
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?

7. 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 .

(b) Show that `(λ) depends only on xi and not on the specific values of the
P
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.

8. 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 Yi j be the number of
deaths in year i, corps j, for i = 1875, . . . , 1894 and j = 1, . . . , 14. The Yi j 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 195

(e) Is there any evidence that different corps had different death rates? How would
you investigate that possibility?

9. Use the data from Example 2.8. Find the m.l.e. for θ.

10. John is a runner and frequently runs from his home to his office. He wants to measure
the distance, so once a week he drives to work and measures the distance on his car’s
odometer. Unfortunately, the odometer records distance only to the nearest mile.
(John’s odometer changes abruptly from one digit to the next. When the odometer
displays a digit d, it is not possible to infer how close it is to becoming d + 1.) John
drives the route ten times and records the following data from the odometer: 3, 3, 3,
4, 3, 4, 3, 3, 4, 3. Find the m.l.e. of the exact distance. You may make reasonable
assumptions as needed.

11. X1 , . . . , Xn ∼ Normal(µ, 1). Multiple choice: The m.l.e. µ̂ is found from the equation

(a) d d
dµ dx
f (x1 , . . . , xn |µ) = 0
(b) d
dµ
f (x1 , . . . , xn |µ) = 0
(c) d
dx
f (x1 , . . . , xn |µ) = 0

12. 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 .

(b) Show that `(µ) depends only on yi and not on the specific values of the
P
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.

13. Let y1 , . . . , yn be a sample from N(µ, 1). Show that µ̂ = ȳ is the m.l.e.

14. Let y1 , . . . , yn be a sample from N(µ, σ) where µ is known. Show that σ̂2 = n−1
P
(yi −
µ)2 is the m.l.e.
2.8. EXERCISES 196

15. Recall the discoveries data from page 10 on the number of great discoveries 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?

16. Justify each step of Equation 2.6.

17. Page 159 discusses a simulation experiment comparing the sample mean and sample
median as estimators of a population mean. Figure 2.27 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?

18. In the medical screening example on page 166, find the probability that the patient
has the disease given that the test is negative.

19. 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 in-
spectors. 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. (2.19)

No weapons are found. Find the probability that B has weapons. I.e., find

Pr[B has weapons|no weapons are found].

20. 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?
(e) Show that `(λ) depends only on ti and not on the values of the individual ti ’s.
P
2.8. EXERCISES 197

21. There are two coins. One is fair; the other is two-headed. You randomly choose a
coin and toss it.

(a) What is the probability the coin lands Heads?

(b) What is the probability the coin is two-headed given that it landed Heads?
(c) What is the probability the coin is two-headed given that it landed Tails? Give
a formal proof, not intuition.
(d) You are about to toss the coin a second time. What is the probability that the
second toss lands Heads given that the first toss landed Heads?

22. There are two coins. For coin A, P[H] = 1/4; for coin B, P[H] = 2/3. You randomly
choose a coin and toss it.

(a) What is the probability the coin lands Heads?

(b) What is the probability the coin is A given that it landed Heads? What is the
probability the coin is A given that it landed Tails?
(c) You are about to toss the coin a second time. What is the probability the second
toss lands Heads given that the first toss landed Heads?

23. At Dupont College (apologies to Tom Wolfe) Math SAT scores among math majors
are distributed N(700, 50) while Math SAT scores among non-math majors are dis-
tributed N(600, 50). 5% of the students are math majors. A randomly chosen student
has a math SAT score of 720. Find the probability that the student is a math major.

24. 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].
2.8. EXERCISES 198

(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 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 calculate a prob-
ability 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.

25. 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 approxi-
mately represents your prior opinion for w.

26. Let t be the amount of time between the moment when the sun first touches the hori-
zon in the afternoon and the moment when it sinks completely below the horizon.
Without making any observations, assess your distribution for t.

27. 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. Give the M&M’s to your professor, so he or she can check your work.

28. (a) Let y ∼ N(θ, 1) and let the prior distribution for θ be θ ∼ N(0, 1).
i. When y has been observed, what is the posterior density of θ?
ii. Show that the density in part i. is a Normal density.
iii. Find its mean and SD.
2.8. EXERCISES 199

(b) Let y ∼ N(θ, σy ) and let the prior distribution for θ be θ ∼ N(m, σ). Suppose
that σy , m, and σ are known constants.
i. When y has been observed, what is the posterior density of θ?
ii. Show that the density in part i. is a Normal density.
iii. Find its mean and SD.
(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). Suppose that σy , m, and σ are known constants.
i. When y1 , . . . , yn have been observed, what is the posterior density of θ?
ii. Show that the density in part i. is a Normal density.
iii. Find its mean and SD.

29. Verify Equations 2.16, 2.17, and 2.18.

30. Refer to the discussion of predictive intervals on page 175. Justify the claim that
(−∞, −.72), (−3.65, −0.36), and (−3.28, ∞) are 90% prediction intervals. Find the
corresponding 80% prediction intervals.

31. (a) Following Example 2.17 (pg. 176), find Pr[y f = k | y1 , . . . yn ] for k = 1, 2, 3, 4.
(b) Using the results from part (a), make a plot analagous to Figure 2.33 (pg. 175).

32. 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 con-
sider first testing whether R generates 0 with the right frequency, then repeating the
analysis for each digit.

33. (a) Repeat the analysis of Example 2.18 (pg. 184), 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?

34. Continue Exercise 43 from Chapter 1. The autoganzfeld trials resulted in X = 122.

(a) What is the parameter in this problem?

(b) Plot the likelihood function.
2.8. EXERCISES 200

(c) Test the “no ESP, no cheating” hypothesis.

(d) Adopt and plot a reasonable and mathematically tractable prior distribution for
the parameter. Compute and plot the posterior distribution.
(e) Find the probability of a match on the next trial given X = 122.
(f) What do you conclude?
35. Three biologists named Asiago, Brie, and Cheshire are studying a mutation in morn-
ing 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 mutation, 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.
36. 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 faculty in
the 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 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 laboratory 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 distribution. 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 = 100. (λw
is E(xi ).)
Chris Stats writes the following R code.
2.8. EXERCISES 201

y <- rexp(15,.01)
m <- mean(y)
s <- sqrt( var(y ) / 15)
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?

37. Refer to the R code in Example 2.1 (pg. 101). 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 )?
 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 earthquakes, in-
cluding 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 transmissions than for cars with manual
transmissions.

4. The data set faithful contains data about eruptions of the Old Faithful 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.

202
3.1. INTRODUCTION 203

(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 204

Figure 3.1 was produced by the following R snippet.

par ( mfrow=c(2,2) )

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 continuous. 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 transmissions 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 between 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)

3.1. INTRODUCTION 205

The result of the 1970 draft lottery is available at DASL . The website explains:

“In 1970, Congress instituted a random selection process for the mili-
tary 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.”

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 scat-
terplot 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. 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 ) )
3.1. INTRODUCTION 206

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
3.1. INTRODUCTION 207

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 208

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 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.

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 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
3.1. INTRODUCTION 209

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.

3.2. NORMAL LINEAR MODELS 210

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 customary 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 populations
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 different 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 211

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
3.2. NORMAL LINEAR MODELS 212

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 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(µ M , 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; µ M ≈ 160; µP ≈ 120; σ ≈ 30.
An equivalent formulation is
B1 , . . . , B20 ∼ i.i.d. N(µ, σ)
M1 , . . . , M17 ∼ i.i.d. N(µ + δ M , 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 ??. 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" )
3.2. NORMAL LINEAR MODELS 213

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

3.2. NORMAL LINEAR MODELS 214

• hotdogs$C and hotdogs$T illustrate a convenient feature of R, that components

of a structure can be abbreviated. Instead of typing hotdogs$Calories and
hotdogs$Type we can use the abbreviations. The same thing applies to argu-
ments 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.

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
µ M − µB δM between Beef and Meat 10
hotdogs
mean calorie difference
µP − µ B δP between Beef and Poul- -30
try hotdogs
SD of calorie content
σ σ within a single type of 30
hot 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

3.2. NORMAL LINEAR MODELS 215

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?
trt2
trt1
ctrl

3.5 4.0 4.5 5.0 5.5 6.0

weight

Figure 3.7: The PlantGrowth data

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, a look at the data sug-
gests that the weights in each group are clustered around a central value, approximately
3.2. NORMAL LINEAR MODELS 216

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 different 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 statistical 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, regressor,
or explanatory variable. Covariates are often called X .
The data consists of multiple data points, or cases. We write Yi 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.)

3.2. NORMAL LINEAR MODELS 217

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 + δ M 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, δ M , δP )t ,
E = (1 , . . . , 54 )t ,
(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
3.2. NORMAL LINEAR MODELS 218

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 214) 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
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)
3.2. NORMAL LINEAR MODELS 219

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 + δ M 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 differences 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 + δ M X1,i + δP X2,i is a linear function of (µ, δ M , 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 deter-
mine 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

The variables are

3.2. NORMAL LINEAR MODELS 220

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 consump-
tion 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.
3.2. NORMAL LINEAR MODELS 221

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 222

3.2.2 Inference for Linear Models

Section 3.2.1 showed some graphical displays of data that were eventually described 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 covari-
ates, say p. Equation 3.12 is shorthand for the collection of univariate equations

Yi = β0 + β1 X1,i + · · · + β p X p,i + i (3.13)

or equivalently,

Yi ∼ N(µi , σ)

for i = 1, . . . , n where µi = β0 + j β j X j,i and the i ’s are i.i.d. N(0, σ). There are p + 2
P
parameters: (β0 , . . . , β p , σ). The likelihood function is

n
Y
`(β0 , . . . , β p , σ) = p(yi | β0 , . . . , β p , σ)
i=1
n
Y 1 yi −µi 2
e− 2 (
1
= √ σ )
i=1 2πσ
n  y −(β
0 + β j X j,i )
P 2
Y 1 − 12 i
= e√ σ

i=1 2πσ
− n 2
= 2πσ2 2 e− 2σ2 i (yi −(β0 + β j X j,i ))
 1 P P
(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 derivatives 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 , . . . , β p , σ) = C − n log σ − y
 i
 − (β0 + β X )
j i, j 

2σ2 i  j

3.2. NORMAL LINEAR MODELS 223

for some irrelevant constant C, so we get the system of equations

1 X X
β̂ + β̂ =0


yi − ( 0 j X i, j )
σ̂2 i j
1 X X
β̂ + β̂ Xi,1 = 0


y i − ( 0 j X i, j )
σ̂2 i j
..
. (3.15)
1 X X
β̂ + β̂ Xi,p = 0


yi − ( 0 j X i, j )
σ̂2 i j
n 1 X X
2
− + 3 yi − (β̂0 + β̂ j Xi, j ) = 0
σ̂ σ̂ i j

Note the hat notation to indicate estimates. The m.l.e.’s (β̂0 , . . . , β̂ p , σ̂) are the values of
the parameters that make the derivatives equal to 0 and therefore satisfy Equations 3.15.
The first p + 1 of these equations can be multiplied by σ2 , 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 + · · · + x pi β̂ p .

The ŷi ’s are called fitted values. The residuals are

ri = yi − ŷi
= yi − β̂0 + x1i β̂1 + · · · + x pi β̂ p

and are estimates of the errors i . Finally, referring to the last line of Equation 3.15, the
m.l.e. σ̂ is found from
n 1 X X
2
0=− + 3 yi − (β0 + β j Xi, j )
σ̂ σ̂ i j
n 1 X
=− + 3 r2
σ̂ σ̂ i i
3.2. NORMAL LINEAR MODELS 224

so
1X 2
σ̂2 = ri
n
and
! 12
ri2
P
σ̂ = (3.16)
n
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 sin-
gle 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 param-
eters.

• 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 )

3.2. NORMAL LINEAR MODELS 225

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)

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
3.2. NORMAL LINEAR MODELS 226

where X1 and X2 are indicator variables for the type of hot dog. The model implies

Yi = β0 + i for beef hotdogs

Yi = β0 + β1 + i for meat hotdogs
Yi = β0 + β2 + 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 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 β0 is likely some-
where around 157, plus or minus about 10 or so; β1 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 ??.) 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.
3.2. NORMAL LINEAR MODELS 227

likelihood

140 150 160 170 likelihood !20 0 10

µ !M
likelihood

!60 !40 !20

!P

Figure 3.9: Likelihood functions for (µ, δ M , δP ) in the Hot Dog example.
3.2. NORMAL LINEAR MODELS 228

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.

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,
1
R, like most statistical software, does not report the m.l.e. but reports instead σ̂ ≡ ( ri2 /(n − p − 1))1/2 .
P
Compare to Equation 3.16 for the m.l.e. in which the denominator is n. The situation is similar to the sample
SD on page 99. When n  p there is little difference between the two estimates.
3.2. NORMAL LINEAR MODELS 229

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 per-
formance 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 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 Equation 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 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
3.2. NORMAL LINEAR MODELS 230

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 231

should show no obvious patterns. Patterns in residual plots are clues to model in-
adequacy and ways to improve models. Figure 3.11(c) and (d) are residual 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 -4 to about +6, 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) )
3.2. NORMAL LINEAR MODELS 232

(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 233

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 0 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 )
3.2. NORMAL LINEAR MODELS 234

!8 !6 !4 !2 !0.10 !0.04 0.00

!1 "1

!8 !6 !4 !2 !0.10 !0.04 0.00

#1 #2

Figure 3.12: likelihood functions for β1 , γ1 , δ1 and δ2 in the mtcars example.

3.3. GENERALIZED LINEAR MODELS 235

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" )
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 203. 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 2.11 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
3.3. GENERALIZED LINEAR MODELS 236

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 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.
3.3. GENERALIZED LINEAR MODELS 237

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 according to
Equation 3.20. The values of β0 and β1 are in Table ??.

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

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)" )
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 )
3.3. GENERALIZED LINEAR MODELS 238

(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
3.3. GENERALIZED LINEAR MODELS 239

(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 240

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

eφi
θi = .
1 + eφi

The logistic regression model is

φi = β0 + β1 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 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, indeed, all generalized linear models differ from linear regres-
sion 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.
3.3. GENERALIZED LINEAR MODELS 241

The likelihood function is derived from first principles.

Y
p(Y1 , . . . , Yn | x1 , . . . , xn , β0 , β1 ) = p(Yi | xi , β0 , β1 )
i
Y
= θiyi (1 − θi )1−yi
i
Y Y
= θi (1 − θi )
i:yi =1 i:yi =0
Y eβ0 +β1 xi Y 1
=
i:y =1
1+e β 0 +β 1 xi
i:y =0
1 + eβ0 +β1 xi
i 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.
• The lines b0 <- ... and b1 <- ... set some values of (β0 , β1 ) at which to eval-
uate the likelihood. They were chosen after looking at the output from fitting the
logistic regression model.
3.3. GENERALIZED LINEAR MODELS 242

0.50
0.45
0.40
0.35
!1

0.30
0.25
0.20
0.15

!11 !10 !9 !8 !7 !6 !5 !4

!0

Figure 3.15: Likelihood function for the pine cone data

3.3. GENERALIZED LINEAR MODELS 243

• lik <- ... creates a matrix to hold values of the likelihood function.

• linpred is the linear predictor. Because cones$dbh[ring1] is a vector, linpred

is also a vector. Therefore theta is also a vector, as is thetaˆmature *
(1-theta)ˆ(1-mature). It will help your understanding of R to understand what
these vectors are.

One notable feature of Figure 3.15 is the diagonal slope of the contour ellipses. 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 neighborhood 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
3.3. GENERALIZED LINEAR MODELS 244

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

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 argu-
ment 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
3.3. GENERALIZED LINEAR MODELS 245

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 Yi j 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)
)
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 struc-
ture 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 num-
ber 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 indicator
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 covariates.

• The formula count ˜ quadrat says to fit a model in which quadrat is a covari-
ate. Of course that’s really 59 new covariates, indicator variables for 59 of the 60
quadrats.
3.3. GENERALIZED LINEAR MODELS 246

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 )

Panels (b) and (c) show diagnostic plots for fit2. Compare to similar panels in Fig-
ure 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)),
3.3. GENERALIZED LINEAR MODELS 247

(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 248

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.
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 continue
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)
3.4. PREDICTIONS FROM REGRESSION 249

(a) (b)

15
3
2

actual values

10
residuals

1
0
!1

5
!3

0
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.
3.4. PREDICTIONS FROM REGRESSION 250

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 = β0 + β1 x +  (3.22)

where x was mean temperature during the week and y was ice cream consumption 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 predictions are. Let x f be the predicted mean
temperature for some future week and y f be consumption. x f is known; y f is not. Our
model says

y f ∼ N(µ f , σ)

where

µ f = β0 + β1 x f

µ f is unknown because β0 and β1 are unknown. But (β0 , β1 ) can be estimated from the
data, so we can form an estimate

µ̂ f = β̂0 + β̂1 x f
3.4. PREDICTIONS FROM REGRESSION 251

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
3.5. EXERCISES 252

How accurate is µ̂ f as an estimate of µ f ? The answer depends on how accurate (β̂0 , β̂1 ) are
as estimates of (β0 , β1 ). Advanced theory about Normal distributions, beyond the scope of
this book, tells us
µ̂ f ∼ N(µ f , σfit )
for some σfit which may depend on x f ; we have omitted the dependency from the notation.
µ f is the average ice cream consumption in all weeks whose mean temperature is x f .
So µ̂ f is also an estimator of y f . But in any particular week the actual consumption won’t
exactly equal µ f . Our model says
yf = µf + 
where  ∼ N(0, σ). So in any given week y f will differ from µ f by an amount up to about
±2σ or so.
Thus the uncertainty σfit in estimating y f has two components: (1) the uncertainty
of µ f which comes because we don’t know (β0 , β1 ) and (2) the variability σ due to .
We can’t say in advance which component will dominate. Sometimes it will be the first,
sometimes the second. What we can say is that as we collect more and more data, we
learn (β0 , β1 ) more accurately, so the first component becomes negligible and the second
component dominates. When that happens, we won’t go far wrong by simply ignoring the
first component.

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 relationship 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 tun-
ing 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 generated uni-
formly? What would H0 be? What would be a good test statistic w? How would
estimate the distribution of w under H0 ?
3.5. EXERCISES 253

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.

(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 Figure 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 214. 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,
3.5. EXERCISES 254

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.

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 football 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 214 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 University 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 vol-
umes 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 ma-
jors. 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 leave Duke’s campus and go up Chapel Hill to interview
poets and investment bankers at UNC-CH. In all of Chapel Hill there are 90 po-
ets 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 Fig-
ure 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 .

3.5. EXERCISES 255

p
b

39 40 41 42 43 44 45

HQ

Figure 3.19: Happiness Quotient of bankers and poets

3.5. EXERCISES 256

(b) Express β1 and β2 in terms of α0 and α1 .

(c) In their data set J&J find HQ = 43 among poets, HQ = 44 among bankers and
σ̂2 = 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

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 relation-
ship 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.
(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 assuming the program was
successful, can you figure out which school it was?
3.5. EXERCISES 257

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. (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 258

if ( abs(fit$coef[2]-1) <= .44 ) in1 <- in1 + 1

}

Make an intelligent guess of in0 and in1 after Jane ran this code.

21. 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 distance from the howitzer (measured horizon-
tally along the ground) at times t1 , t2 , . . . , t100 . The yi ’s and zi ’s are measured by the
laser. The measurements are not perfect; there is some measurement error. In an-
swering 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

yi = β0 + β1 ti + i

should fit the data well.

(b) True or False: The equation

yi = β0 + β1 ti + β2 ti2 + i (3.23)

should fit the data well.

(c) True or False: The equation

zi = β0 + β1 ti + i (3.24)

should fit the data well.

(d) True or False: The equation

zi = β0 + β1 ti + β2 ti2 + i (3.25)

should fit the data well.

(e) True or False: The equation

yi = β0 + β1 zi + i (3.26)

should fit the data well.

3.5. EXERCISES 259

(f) True or False: The equation

yi = β0 + β1 zi + β2 z2i + i (3.27)

should fit the data well.

(g) Approximately what value did the Army find for β̂0 in Part (b)?
(h) Approximately what value did the Army find for β̂2 in Part (d)?

22. 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?

23. 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.
3.5. EXERCISES 260

(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 ) )

What does z1 estimate? What does z2 estimate? Why did the code writer write
sqrt ( var ( w + 2 ) ) instead of sqrt ( var ( w ) )? Does it matter?

24. 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 = β∗0 + β∗1 wi + i∗ . Note that (β∗0 , β∗1 ) 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, 0, 3 )
fit <- lm ( y ~ w )
val[k] <- fit$coef[2]
3.5. EXERCISES 261

}
m <- mean ( val )
sd <- sqrt ( var ( val ) )
print ( c ( m, sd ) )
}

What is she trying to do? The last time through the loop, the print statement yields
[1] 9.0805986 0.5857724. What does this show?

25. The purpose of this exercise is to familiarize yourself with plotting logistic regres-
sion 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.
(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.

26. 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 temperature of the Chal-
lenger launch. How relevant should such an analysis have been to the decision of
whether to postpone the launch?

27. 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 panels (a) and (b)?
(c) If there were no jittering, how many distinct values would there be on the
abscissa of panels (c) and (d)?
3.5. EXERCISES 262

(d) Download the seedling data. Fit a model in which year is a predictor 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 F X . Equation 1.2 on page 8 implies
that fX (x) = dbd
F X (b) b=x and therefore that we can define the pdf by fX (x) ≡ F X0 (x) =
d
F (b) b=x . In fact, this definition is a little too restrictive. The key property of pdf’s is
db X
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
∗
ButR if f isR a function that 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 intervals A,
Z
P[X ∈ A] = f (x) dx
A

263
4.2. RANDOM VECTORS 264

is called a probability density function, or pdf, for the random variable X. Any such func-
tion may be denoted fX .
Definition 4.1 can be used in an alternate proof of Theorem 1.1 on page 12. 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 differentiable,
monotonic, invertible function and define Z = g(X). Then the pdf of Z is

d g−1 (t)
pZ (t) = pX (g−1 (t))
dt

Proof. For any set A, P[Z ∈ g(A)] = P[X ∈ A] = A pX (x) dx. Let z = g(x) and change
R

variables in the integral to get

Z
dx
P[Z ∈ g(A)] = pX (g−1 (z)) dz
g(A) dz

I.e., P[Z ∈ g(A)] = g(A) something dz. Therefore something must be pZ (z). Hence,
R

pZ (z) = pX (g−1 (z))|dx/dz|. 

4.2 Random Vectors

It is often useful, even essential, to talk about several random variables simultaneously.
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.
~ = (X1 , . . . , Xn )
Let X1 , . . . , Xn be a set of n random variables. The n-dimensional vector X
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 ).
4.2. RANDOM VECTORS 265

As in the univariate case, the pdf is any function whose integral yields probabilities. That
is, if A is a region in Rn then
(
~
P[X ∈ A] = pX~ (x1 , · · · , xn ) dx1 . . . dxn
A

For example, let X1 ∼ Exp(1); X2 ∼ Exp(1/2); X1 ⊥ X2 ; and X ~ = (X1 , X2 ) and suppose

we want to find P[|X1 − X2 | ≤ 1]. Our plan for solving this 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.
"
P[|X1 − X2 | ≤ 1] = pX~ (x1 , x2 ) dx1 dx2
A
Z 1Z x1 +1 Z ∞ Z x1 +1
1 −x1 −x2 /2 1
= e e 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 inde-
pendent 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 , X j ) 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. RANDOM VECTORS 266

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 i j’th entry of the covariance matrix is Cov(Xi , X j ). The notation is

 σ1 σ12 · · · σ1n 

 2 
σ σ2 · · · σ2n 
 
~ ≡ Σ ~ =  .12 .2
Cov(X) .. . 
X
 .. .. . .. 
σ1n σ2n · · · σ2n
 

where σi j = Cov(Xi , X j ) and σ2i = Var(Xi ). Sometimes σ2i 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
~ to R then
X (
~
E[g(X)] = g(x1 , . . . , xn )pX~ (x1 , . . . , xn ) dx1 · · · dxn

~ When g is a linear function

but it’s hard to say much in general about the variance of g(X).
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 Y = ~at X
P
ai Xi . Then,

1. E[Y] = E[ ai Xi ] = ai E[Xi ]
P P

2. Var(Y) = a2i Var(Xi ) + 2 n−1 j=i+1 ai a j Cov(Xi , X j ) = ~

at ΣX~ ~a
P P Pn
i=1
4.2. RANDOM VECTORS 267

Proof. Use Lemma 4.1 and Theorems 1.3 (pg. 40) and 1.4 (pg. 41). See Exercise 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 = ai j X j = ~ati X
j

~ = (Y1 , . . . , Yk ). In
where the ai j ’s are arbitrary constants and ~ai = (ai1 , . . . , ain ). Let Y
matrix notation,
~ = AX
Y ~
where A is the k × n matrix of elements ai j . Covariances of the Yi ’s are given by
~ ~atj X)
Cov(Yi , Y j ) = Cov(~ati X, ~
Xn X n
= Cov(aik Xk , a j` X j )
k=1 `=1
Xn n−1 X
X n
= aik a jk σ2k + (aik a j` + a jk ai` )σk`
k=1 k=1 `=k+1
= ~ati ΣX~ ~a j
Combining the previous result with Theorem 4.2 yields Theorem 4.3.
~ be a random vector of dimension n with mean E[X]
Theorem 4.3. Let X ~ = µ and covari-
~ = Σ; let A be a k × n matrix of rank k; and let Y
ance matrix Cov(X) ~ = AX.
~ Then
~ = Aµ, and
1. E[Y]
~ = AΣA0
2. Cov(Y)
Finally, we take up the question of multivariate transformations, extending the univari-
ate version, Theorem 1.1 (pg. 12). Let X ~ = (X1 , . . . , Xn ) be an n-dimensional continuous
random vector with pdf fX~ . Define a new n-dimensional random vector Y ~ = (Y1 , . . . , Yn ) =
~ . . . , gn (X))
(g1 (X), ~ where the gi ’s are differentiable functions and where the the transfor-
~
mation g : X 7→ Y ~ is invertible. What is f ~ , the pdf of Y? ~
Y
Let J be the so-called Jacobian matrix of partial derivatives.
 ∂Y1 ∂Y1 
 ∂X1 · · · ∂X n

 ∂Y2 ∂Y2 
 ∂X · · · ∂Xn 
J =  . 1 . .. 
 .. .. . 
 ∂Y ∂Yn 
∂X1
n
· · · ∂Xn
and |J| be the absolute value of the determinant of J.
4.2. RANDOM VECTORS 268

Theorem 4.4.
fY~ (~y) = fX~ (g−1 (~y))|J|−1

Proof. The proof follows the'

alternate proof of Theorem 1.1 on page 264. For any set A,
~ ~
P[Y ∈ g(A)] = P[X ∈ A] = p ~ (~x) dx1 · · · dxn . Let ~y = g(~x) and change variables in
A X
the integral to get
(
~ ∈ g(A)] =
P[Y pX~ (g−1 (~y)) |J|−1 dy1 · · · dyn
g(A)

'
~ ∈ g(A)] =
I.e., P[Y something dy1 · · · dyn . Therefore something must be pY~ (~y).
g(A)
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 265, 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 the joint density of Y ~ = (Y1 , Y2 ), then integrate over the region
where |Y1 | ≤ 1. 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 . The precise
definition of Y2 is unimportant, as long as the transformation from X ~ to Y
~ is differentiable
and invertible. For convenience, we define Y2 = X2 . With these definitions,
 ∂Y1 ∂Y1  !
1 −1
J =  ∂X ∂X2 
=
 
1
∂Y2 ∂Y2 
∂X1 ∂X2
0 1
|J| = 1
X1 = Y1 + Y2

and

X2 = Y2

From the solution on page 265 we know pX~ (x1 , x2 ) = e−x1 × 12 e−x2 /2 , so pY~ (y1 , y2 ) =
4.2. RANDOM VECTORS 269

e−(y1 +y2 ) × 12 e−y2 /2 = 12 e−y1 e−3y2 /2 . Figure 4.1 shows the region over which to integrate.
"
P[|X1 − X2 | ≤ 1] = P[|Y1 | ≤ 1] = pY~ (y1 , y2 ) dy1 dy2
A
∞
0
1 1 −y1 ∞ −3y2 /2
Z Z Z Z
1 −3y2 /2
= e −y1
e dy2 dy1 + e e dy2 dy1
2 −1 −y1 2 0 0

1 0 −y1 h −3y2 /2 i∞ 1 1 −y1 h −3y2 /2 i∞

Z Z
= e −e dy1 + e −e dy1
3 −1 −y1 3 0 0

1 0 y1 /2 1 1 −y1
Z Z
= e dy1 + e dy1
3 −1 3 0
2 0 1 1
= ey1 /2 −1 − e−y1 0
3 3
2h i 1h i
= 1−e −1/2
+ 1 − e−1
3 3
≈ 0.47 (4.3)

X2 Y2

X1 Y1

~
Figure 4.1: The (X1 , X2 ) plane and the (Y1 , Y2 ) plane. The light gray regions are where X
~ live. The dark gray regions are where |X1 − X2 | ≤ 1.
and Y
4.3. REPRESENTING DISTRIBUTIONS 270

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
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 the method, not the answer. 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
4.3. REPRESENTING DISTRIBUTIONS 271

be described by its cumulative distribution function, or cdf, FY which is defined by

P
R y=−∞ P[Y = y] if Y is discrete
 c
FY (c) ≡ P[Y ≤ c] = 

(4.4)
 c 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)" )
4.3. REPRESENTING DISTRIBUTIONS 272

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

4.3. REPRESENTING DISTRIBUTIONS 273

• segments ( x0, y0, x1, y1) draws line segments. The line segments run from
(x0,y0) to (x1,y1). The arguments may be vectors.

The other alternative representation for Y is its moment generating function or mgf MY .
The moment generating function is defined as
P
R y ety pY (y)
 if Y is discrete
MY (t) = E[etY ] = 

(4.6)
 ety 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 12.) 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

dn
E[Y n ] = MY(n) (0) ≡ MY (t)
dtn 0

Proof. We provide the proof for the case n = 1. The proof for larger values of n is similar.
Z
d d
MY (t) = ety pY (y) dy
dt 0 dt
Z
0
d ty
= e pY (y) dy
Z dt 0
= yety pY (y) dy
0
Z
= ypY (y) dy

= E[Y]


4.3. REPRESENTING DISTRIBUTIONS 274

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.
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
F X = FY ; i.e., X and Y have the same distribution.
Theorem 4.7. Let Y1 , . . . be a sequence of random variables with moment generating
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 F such that
1.
F(y) = lim FYn (y)
n→∞

for all y where F is continuous and

2. M is the mgf of F.
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 characteristic
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 omitted but may be found in more advanced books.
Two more useful results are theorems 4.8 and 4.9.
4.4. EXERCISES 275

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
h i
= 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 eYt ] = E[eXt ]E[eYt ] = 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

MX (t) = [MY (t)]n

4.4 Exercises
1. Refer to Equation 4.1 on page 265.

(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 inte-
gration 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.
4.4. EXERCISES 276

2. Refer to Example 1.6 on page 43 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 disk. 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 265 that mutual independence implies pairwise inde-
pendence.

(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 265. 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 266.

9. X and Y are uniformly distributed in the rectangle whose corners are (1, 0), (0, 1),
(−1, 0), and (0, −1).

(a) i. Find p(x, y).

ii. Are X and Y independent?
iii. Find the marginal densities p(x) and p(y).
iv. Find the conditional densities p(x | y) and p(y | x).
v. Find E[X], E[X | Y = .5], and E[X | Y = −.5].
(b) Let U = X + Y and V = X − Y.
i. Find the region where U and V live.
ii. Find the joint density p(u, v).
4.4. EXERCISES 277

iii. Are U and V independent?

iv. Find the marginal densities p(u) and p(v).
v. Find the conditional densities p(u | v) and p(v | u).
vi. Find E[U], E[U | V = .5], and E[U | V = −.5].
10. Let the random vector (U, V) be distributed uniformly on the unit square. Let X =
UV 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].
11. (a) Let (X1 , X2 ) be distributed uniformly on the disk where X12 + X22 ≤ 1. Let
q
R = X12 + X22 and Θ = arctan(X1 /X2 ). Hint: it may help to draw a picture.
What is the joint density p(x1 , x2 )?
i.
ii.
Are X1 and X2 independent? Explain.
iii.Find the joint density p(r, θ).
iv.Are R and Θ independent? Explain.
q
(b) Let (X1 , X2 ) be i.i.d. N(0,1). Let R = X12 + X22 and Θ = arctan(X1 /X2 ).
i. What is the joint density p(x1 , x2 )?
ii. Find the joint density p(r, θ).
iii. Are R and Θ independent? Explain.
iv. Find the marginal density p(r).
v. Let V = R2 . Find the density p(v).
(c) Let (X1 , X2 ) be distributed uniformly on
q the square whose corners are (1, 1),
(−1, 1), (−1, −1), and (1, −1). Let R = X12 + X22 and Θ = arctan(X1 /X2 ).
4.4. EXERCISES 278

i. What is the joint density p(x1 , x2 )?

ii. Are X1 and X2 independent? Explain.
iii. Are R and Θ independent? Explain.

12. Just below Equation 4.6 is the statement “the mgf is always defined at t = 0.” For
any random variable Y, find MY (0).

13. Provide the proof of Theorem 4.5 for the case n = 2.

14. Refer to Theorem 4.9. Where in the proof is the assumption X ⊥ Y used?
 Chapter 5

Special Distributions

Statisticians often make use of standard parametric families of probability distributions. A

parametric family is a collection of probability distributions distinguished 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 Nor-
mal (Section 1.3.4) distributions. Chapter 5 examines these in more detail and introduces
several others.

5.1 The Binomial and Negative Binomial Distributions

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

279
5.1. BINOMIAL AND NEGATIVE BINOMIAL 280

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), 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 S x be the subset of S consisting of
sequences with x 1’s and n − x 0’s. If s ∈ S x then Pr(s) = θ x (1 − θ)n−x . In particular, all s’s
in S x have the same probability. Therefore,
pX (x) = P(X = x) = P(S x )
= (size of S x ) · θ x (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) = θ x (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.
5.1. BINOMIAL AND NEGATIVE BINOMIAL 281

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).

Now let X =
Pn
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)

n  n
= θet + (1 − θ) 1 θet + (1 − θ) 2

n +n
= θet + (1 − θ) 1 2


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).

√
3. SD(X) = nθ(1 − θ).

Proof. The proof for E[X] was given earlier. If X ∼ Bin(n, θ), then X = ni=1 Xi where
P
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.
Exercise 1 asks you to prove Theorem 5.4 by moment generating functions. 
5.1. BINOMIAL AND NEGATIVE BINOMIAL 282

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:

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.
5.1. BINOMIAL AND NEGATIVE BINOMIAL 283

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>

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.
5.1. BINOMIAL AND NEGATIVE BINOMIAL 284

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 ∼ i.i.d. Bern(p). Then the distribution of X is the same as
P P
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 number 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 ran-
dom variable and is said to have the Negative Binomial distribution with parameters (r, θ),
written N ∼ NegBin(r, θ). (Warning: some authors say that the total number of trials,
N + r, has the Negative Binomial distribution.) One example is a gambler who decides to
play the daily lottery until she wins. The prespecified number of successes is r = 1. The
number of failures N until she wins is random. In this case, and whenever r = 1, N is
said to have a Geometric 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
5.1. BINOMIAL AND NEGATIVE BINOMIAL 285

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 286

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 independent of each
other. Then one can imagine a sequence of trials of length (Ni +ri ) having ri successes.
P P
N1 is the number of failures before the r1 ’th success; . . . ; N1 + · · · + Nt is the number of
failures before the r1 + · · · + rt ’th success. It is evident that N ≡ Ni is the number of
P
failures before the r ≡ ri ’th success occurs and therefore that N ∼ NegBin(r, θ).
P

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 the
foregoing argument and 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
∞
d
X
= −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
5.1. BINOMIAL AND NEGATIVE BINOMIAL 287

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 (1 − θ)n
θ n=1
θ2
∞
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 + θ 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) ) {
5.1. BINOMIAL AND NEGATIVE BINOMIAL 288

r = 1, ! = 0.1 r = 1, ! = 0.5 r = 1, ! = 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, ! = 0.1 r = 5, ! = 0.5 r = 5, ! = 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, ! = 0.1 r = 30, ! = 0.5 r = 30, ! = 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

5.2. MULTINOMIAL 289

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 = substitute ( list ( r == a, theta == b),
list(a=i,b=j) ) )
}

• 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 following 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, usually an ex-
perimental treatment or a standard, control treatment. Later, each patient may be
scored as either success, failure, or censored. Censoring 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.
5.2. MULTINOMIAL 290

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 20.) (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.
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 yi !
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

pyi i . Every sequence with y1 a’s, . . . , yk k’s

Q
The probability of this particular sequence is
has the same probability. So
!Y
Y n
fY (y1 , . . . , yk ) = (number of such sequences) × pyi i = pyi i .
y1 · · · yk

5.3. POISSON 291

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 deter-
mined 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.

• The occurence of an event at any location `1 is independent of the occurence 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 19a. It is

E[Y] = λ.

Theorem 5.8. Let Y ∼ Poi(λ). Then

MY (t) = eλ(e −1)

t
5.3. POISSON 292

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!
= eλ(e −1)
t

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.
∞
X e−λ λy
E[Y ] =
2
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!
=λ +λ2

So Var(Y) = E[Y 2 ] − (E[Y])2 = λ.

5.3. POISSON 293

Proof 2.

d2
E[Y 2 ] = MY (t)
dt2 t=0
2
d
= 2 eλ(e −1)
t

dt t=0
d t λ(et −1)
= λe e
hdt t=0
t λ(et −1)
+ λ2 e2t eλ(e −1)
t
i
= λe e
t=0
= λ + λ2

So Var(Y) = E[Y 2 ] − (E[Y])2 = λ. 

Theorem 5.10. Let Yi ∼ Poi(λi ) for i = 1, . . . , n and let the Yi s be mutually independent.
Let Y = n1 Yi and λ = n1 λi . Then Y ∼ Poi(λ).
P P

Proof. Using Theorems 4.9 and 5.8 we have

Y Y
eλi (e −1) = eλ(e −1)
t t
MY (t) = MYi (t) =

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 ; Yi ∼

Poi(λi ). Suppose the Di ’s are disjoint and the Yi ’s are independent. Let Y = Yi be the
P
number of events arising 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
P
indeed have the Poi(λ) distribution. Another way to put it: If Y ∼ Poi(λ), and if the
individual events that Y counts are randomly divided into two types Y1 and Y2 according
to a binomial distributuion with parameter θ, then (1) Y1 ∼ Poi(λθ) and Y2 ∼ Poi(λ(1 − θ))
and (2) Y1 ⊥ Y2 .
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 Theo-
rem. When Y ∼ Poi(λ) Theorem 5.10 tells us we can think of Y as Y = λi=1 Yi where each
P
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.

Figure 5.3 was produced with the following snippet.

5.3. POISSON 294

!=1 !=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

! = 16 ! = 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 295

y <- 0:7
plot ( y, dpois(y,1), xlab="y", ylab=expression(p[Y](y)),
main=expression(lambda==1) )
y <- 0:10
plot ( y, dpois(y,4), xlab="y", ylab=expression(p[Y](y)),
main=expression(lambda==4) )
y <- 6:26
plot ( y, dpois(y,16), xlab="y", ylab=expression(p[Y](y)),
main=expression(lambda==16) )
y <- 44:84
plot ( y, dpois(y,64), xlab="y", ylab=expression(p[Y](y)),
main=expression(lambda==64) )

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 20th
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 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 de-
rive 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
5.3. POISSON 296

screen. The scintillations 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 num-
bers, given in the horizontal lines, correspond to the number of scintilla-
tions 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 that n α particles are observed
n
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 Rutherford
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.
“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 particles 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 particles were
observed was 525. The number calculated from the equation was the
same, viz. 525.”
 5.3. POISSON

Number 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 Number Number Average

of α of α of inter- number
particles particles vals
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

Table 5.1: Rutherford and Geiger’s data

297
5.3. POISSON 298

Finally, how did Rutherford and Geiger compare their actual and theoretical distribu-
tions? 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 excellent
accord. . . . We may consequently conclude that the distribution of α par-
ticles 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 test-
ing the laws of probability by observing the variations in quantities involved
in a spontaneous material process.”

Example 5.2 (neurobiology)

This example continues Example 2.6. We would like to know whether this neuron
responds differently to different tastants and, if so, how. To that end, we’ll see how often
the neuron fires in a short period of time after receiving a tastant and we’ll compare
the results for different tastants. Specifically, we’ll count the number of spikes in the
150 milliseconds (150 msec = .15 s) immediately following the delivery of each tastant.
(150 msec is about the rate at which rats can lick and is thought by neurobiologists to
be about the right interval of time.) Let Yi j be the number of spikes in the 150 msec
following the j’th delivery of tastant i. Because we’re counting the number of events in
a fixed period of time we’ll adopt a Poisson model:
Yi j ∼ Poi(λi )
where λi is the average firing rate of this neuron to tastant i.
We begin by making a list to hold the data. There should be one element for each
tastant. That element should be a vector whose length is the number of times that
tastant was delivered. Here is the R code to do it. (Refer to Example 2.6 for reading in
the data.)
nspikes <- list(
MSG100 = rep ( NA, length(tastants$MSG100) ),
MSG300 = rep ( NA, length(tastants$MSG300) ),
NaCl100 = rep ( NA, length(tastants$NaCl100) ),
NaCl300 = rep ( NA, length(tastants$NaCl300) ),
water = rep ( NA, length(tastants$water) )
)

Now we fill in each element by counting the number of neuron firings in the time
interval.
5.3. POISSON 299

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.
5.3. POISSON 300

for ( i in seq(along=nspikes) )
for ( j in seq(along=nspikes[[i]]) )
nspikes[[i]][j] <- sum ( spikes[[8]] > tastants[[i]][j]
& spikes[[8]] <= tastants[[i]][j] + .15
)

Now we can see how many times the neuron fired after each delivery of, say, MSG100
by typing nspikes$MSG100.
Figure 5.5 compares the five tastants graphically. Panel A is a stripchart. It has
five tick marks on the x-axis for the five tastants. Above each tick mark is a collection
of circles. Each circle represents one delivery of the tastant and shows how many
times the neuron fired in the 150 msec following that delivery. Panel B shows much
the same information in a mosaic plot. The heights of the boxes show how often that
tastant produced 0, 1, . . . , 5 spikes. The width of each column shows how often that
tastant was delivered. Panel C shows much the same information in yet a different
way. It has one line for each tastant; that line shows how often the neuron responded
with 0, 1, . . . , 5 spikes. Panel D compares likelihood functions. The five curves are the
likelihood functions for λ1 , . . . , λ5 .
There does not seem to be much difference in the response of this neuron to
different tastants. Although we can compute the m.l.e. λ̂i ’s with

lapply ( nspikes, mean )

and find that they range from a low of λ̂3 ≈ 0.08 for .1 M NaCl to a high of λ̂1 ≈ 0.4 for
.1M MSG, panel D suggests the plausibility of λ1 = · · · = λ5 ≈ .2.

Figure 5.5 was produced with the following snippet.

spiketable <- matrix ( NA, length(nspikes), 6,

dimnames = list ( tastant = 1:5,
counts = 0:5 )
)
for ( i in seq(along=nspikes) )
spiketable[i,] <- hist ( nspikes[[i]], seq(-.5,5.5,by=1),
plot=F )$counts
freqtable <- apply ( spiketable, 1, function(x)x/sum(x) )
5.3. POISSON 301

A Neuron’s Responses
to 5 Tastants
A B

1 2 3 4 5
5
number of firings

0
counts
3
2

1
1

543 2
0

1 2 3 4 5 tastant

tastant

C D
0.8

0.8
likelihood
fraction

0.4

0.4
0.0

0.0

0 1 2 3 4 5 0.0 0.2 0.4 0.6 0.8 1.0

number of firings lambda

Figure 5.5: Numbers of firings of a neuron in 150 msec after five different tastants. Tas-
tants: 1=MSG .1M; 2=MSG .3M; 3=NaCl .1M; 4=NaCl .3M; 5=water. Panels: A: A
stripchart. Each circle represents one delivery of a tastant. B: A mosaic plot. C: Each line
represents one tastant. D: Likelihood functions. Each line represents one tastant.
5.3. POISSON 302

• The line spiketable <- ... creates a matrix to hold the data and illustrates
the use of dimnames to name the dimensions. Some plotting commands use
those names for labelling axes.

• The line spiketable[i,] <- ... shows an interesting use of the hist com-
mand. Instead of plotting a histogram it can simply return the counts.

• The line freqtable <- ... divides each row of the matrix by its sum, turning
counts into proportions.

But let’s investigate a little further. Do the data really follow a Poisson distribution?
Figure 5.6 shows the Poi(.2) distribution while the circles show the actual fractions of
firings. There is apparently good agreement. But numbers close to zero can be de-
ceiving. The R command dpois ( 0:5, .2 ) reveals that the probability of getting
5 spikes is less than 0.00001, assuming λ ≈ 0.2. So either the λi ’s are not all ap-
proximately .2, neuron spiking does not really follow a Poisson distribution, or we have
witnessed a very unusual event.

Figure 5.6 was produced with the following snippet.

matplot ( 0:5, freqtable, pch=1, col=1,

xlab="number of firings", ylab="fraction" )
lines ( 0:5, dpois ( 0:5, 0.2 ) )

Let us examine one aspect of Example 5.1 more closely. Rutherford and Geiger are
counting α particle emissions from a polonium source and find that the number of emis-
sions in a fixed interval of time has a Poisson distribution. But they could have reasoned
as follows: at the beginning of the experiment there is a fixed number of polonium atoms;
each atom either decays or not; atoms decay independently of each other; therefore the
number of decays, and hence the number of α particles, has a Binomial distribution where
n is the number of atoms and p is the probability that a given atom decays within the time
interval.
Why did Rutherford and Geiger end up with the Poisson distribution and not the Bino-
mial; are Rutherford and Geiger wrong? The answer is that a binomial distribution with
large n and small p is extremely well approximated by a Poisson distribution with λ = np,
5.3. POISSON 303

0.8
0.6
fraction

0.4
0.2
0.0

0 1 2 3 4 5

number of firings

Figure 5.6: The line shows Poisson probabilities for λ = 0.2; the circles show the fraction
of times the neuron responded with 0, 1, . . . , 5 spikes for each of the five tastants.
5.4. UNIFORM 304

so Rutherford and Geiger are correct, to a very high degree of accuracy. For a precise
statement of this result we consider a sequence of random variables X1 ∼ Bin(n1 , p1 ), X2 ∼
Bin(n2 , p2 ), . . . such that limi→∞ ni = ∞ and limi→∞ pi = 0. But if the sequence were some-
thing like X1 ∼ Bin(100, .01), X2 ∼ Bin(1000, .005), X3 ∼ Bin(10000, .0025), . . . then the
Xi ’s would have expected values E[X1 ] = 1, E[X2 ] = 5, E[X3 ] = 25, . . . and their distribu-
tions would not converge. To get convergence we need, at a minimum, for the expected
values to converge. So let M be a fixed number and consider the sequence of random
variables X1 ∼ Bin(n1 , p1 ), X2 ∼ Bin(n2 , p2 ), . . . where pi = M/ni . That way, E[Xi ] = M,
for all i.

Theorem 5.11. Let X1 , X2 , . . . be a sequence of random variables such that Xi ∼ Bin(ni , pi )

where for some constant M, pi = M/ni . Then for any integer k ≥ 0

exp−M M k
lim P[Xi = k] = .
i→∞ k!
I.e., the limit of the distributions of the Xi ’s is the Poi(M) distribution.

Proof. See Exercise 22. 

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 23.

The Continuous Uniform Distribution The continuous uniform distribution is the dis-
tribution whose pdf is flat over the interval [a, b]. We write Y ∼ U(a, b). Although the
notation might be confused with the discrete uniform, 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 24.
5.5. GAMMA, EXPONENTIAL, CHI SQUARE 305

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
 if a ≤ y(1) and b ≥ y(n)
p(y1 , . . . , yn ) = 

 b−a
0
 otherwise
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 Distribu-

tions
“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 > 0
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) = yα−1 e−y/β for y ≥ 0 (5.3)
Γ(α)βα
We write Y ∼ Gam(α, β).
Figure 5.7 shows Gamma densities for four values of α and four values of β.
• In each panel of Figure 5.7 the curves for different α’s have different shapes. Some-
times 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 5.12.
5.5. GAMMA, EXPONENTIAL, CHI SQUARE 306

! = 0.5 !=1

3.0

1.5
" = 0.5 " = 0.5
"=1 "=1
"=2 "=2
2.0

1.0
"=4 "=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

!=2 !=4
0.8

0.4

" = 0.5 " = 0.5

"=1 "=1
0.6

0.3

"=2 "=2
"=4 "=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 5.7: Gamma densities for various values of α and β.

5.5. GAMMA, EXPONENTIAL, CHI SQUARE 307

Figure 5.7 was produced by the following snippet.

par ( mfrow=c(2,2) )

shape <- c ( .5, 1, 2, 4 )

scale <- c ( .5, 1, 2, 4 )
leg <- expression ( alpha == .5, alpha == 1,
alpha == 2, alpha == 4 )

for ( i in seq(along=scale) ) {
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]), legend = leg )
}

Theorem 5.12. 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)βα
Since Y = cX, x = y/c and dx/dy = 1/c, so

1
pY (y) = α
(y/c)α−1 e−y/cβ
cΓ(α)β
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 5.13. Let Y ∼ Gam(α, β) Then E[Y] = αβ.

5.5. GAMMA, EXPONENTIAL, CHI SQUARE 308

Proof.
Z ∞
1
E[Y] = y yα−1 e−y/β dy
0 Γ(α)βα

Γ(α + 1)β ∞
Z
1
= yα 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. Also see Exercise 10. 
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 5.14. Let Y ∼ Gam(α, β). Then the moment generating function is MY (t) =
(1 − tβ)−α for t < 1/β.

Proof.
Z ∞
1
MY (t) = ety yα−1 e−y/β dy
0 Γ(α)βα
β α Z ∞
( 1−tβ ) 1 1−tβ
α−1 −y β
= y e dy
βα 0 Γ(α)(
β
)α1−tβ
= (1 − tβ) −α

Theorem 5.15. Let Y ∼ Gam(α, β). Then

Var(Y) = αβ2

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

5.5. GAMMA, EXPONENTIAL, CHI SQUARE 309

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 exponen-
tial 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 5.8 (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, SD, and mgf are given by Theorems 5.13 – 5.15.
Each exponential density has its maximum at y = 0 and decreases monotonically. 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 small values of λ; large values of y are evidence for
large values of λ.

Example 5.3 (Radioactive Decay)

It is well known that some chemical elements are radioactive. Every atom of a ra-
dioactive 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.)
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. You will be asked to verify this claim in Exercise 31.
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 property. 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 remaining time on hold; i.e., S = T − t. What is the
5.5. GAMMA, EXPONENTIAL, CHI SQUARE 310

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 5.8: Exponential densities

5.5. GAMMA, EXPONENTIAL, CHI SQUARE 311

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] e−(t+r)/λ
= = −t/λ = e−r/λ .
P[T ≥ t] e
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 Exponential 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 .

The Poisson process There is a close relationship between Exponential, Gamma, and
Poisson distributions. For illustration consider a company’s customer 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 distri-
bution 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.
Suppose we start monitoring calls at time t0 . Let T 1 be the time of the first call after t0
and Y1 = T 1 − t0 , the time until the first call. T 1 and Y1 are random variables. What is the
distribution of Y1 ? For any positive number y,

Pr[Y1 > y] = Pr[no calls in [t0 , t0 + y]] = e−λy

where the second equality follows by the Poisson assumption. But

Pr[Y1 > y] = e−λy ⇒ Pr[Y ≤ y] = 1 − e−λy ⇒ pY (y) = λe−λy ⇒ Y1 ∼ Exp(1/λ)

What about the time to the second call? Let T 2 be the time of the second call after t0 and
Y2 = T 2 − 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
5.6. BETA 312

and therefore
λ2 −λy
pY2 (y) = λe−λy − λe−λy + yλ2 e−λy = ye
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. This fact is
an example of the following theorem.
Theorem 5.16. Let Y1 , . . . , Yn be mutually independent and let Yi ∼ Gam(αi , β). Then
X
Y≡ Yi ∼ Gam(α, β)
where α ≡ αi .
P

Proof. See Exercise 32. 

In Theorem 5.16 note that the Yi ’s must all have the same β even though they may have
different αi ’s.

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 .
Theorem 5.17. Let Y1 , . . . , Yn ∼ i.i.d. N(0, 1). Define X = Yi2 . Then X ∼ χ2n .
P

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 5.18.
Theorem 5.18. Let Y ∼ Be(α, β). Then
α
E[Y] =
α+β
αβ
Var(Y) =
(α + β)2 (α
+ β + 1)
5.6. BETA 313

Proof. See Exercise 28. 

Figure 5.9 shows some Beta densities. Each panel shows four densities having 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) dis-
tribution. 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.

F X(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) = F X(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 Exer-
cise 29.

Figure 5.9 was produced by the following R snippet.

par ( mfrow=c(3,1) )
y <- seq ( 0, 1, length=100 )
mean <- c ( .2, .5, .9 )
alpha <- c ( .3, 1, 3, 10 )
for ( i in 1:3 ) {
beta <- (alpha - mean[i]*alpha) / mean[i]
den <- NULL
for ( j in 1:length(beta) )
den <- cbind ( den, dbeta(y,alpha[j],beta[j]) )
matplot ( y, den, type="l", main=letters[i], ylab="p(y)" )
if ( i == 1 )
legend ( .6, 8, paste ( "(a,b) = (", round(alpha,2), ",",
round(beta,2), ")", sep="" ), lty=1:4 )
else if ( i == 2 )
legend ( .1, 4, paste ( "(a,b) = (", round(alpha,2), ",",
round(beta,2), ")", sep="" ), lty=1:4 )
5.6. BETA 314

8
(a,b) = (0.3,1.2)
(a,b) = (1,4)

6
(a,b) = (3,12)
p(y) (a,b) = (10,40)

4
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 5.9: Beta densities — a: Beta densities with mean .2; b: Beta densities with mean
.5; c: Beta densities with mean .9;
5.7. NORMAL 315

else if ( i == 3 )
legend ( .1, 10, paste ( "(a,b) = (", round(alpha,2), ",",
round(beta,2), ")", sep="" ), lty=1:4 )

The Beta density is closely related to the Gamma density by the following theorem.

Theorem 5.19. Let X1 ∼ Gam(α1 , β); X2 ∼ Gam(α2 , β); and X1 ⊥ X2 . Then

X1
Y≡ ∼ Be(α1 , α2 )
X1 + X2
Proof. See Exercise 33. 
Note that Theorem 5.19 requires X1 and X2 both to have the same value of β, but the
result doesn’t depend on what that value is.

5.7 The Normal and Related Distributions

5.7.1 The Univariate Normal Distribution
The histograms in Figure 1.12 on page 32

• 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 scientists 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 316

Example 5.4 (Ocean temperatures, continued)

To see a Normal density in more detail, Figure 5.10 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 5.10: 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.

Figure 5.10 was produced by the following R snippet.

good <- abs ( med.1000$lon - lons[3] ) < 1 &

abs ( med.1000$lat - lats[1] ) < 1
temps <- med.1000$temp[good]
hist ( temps, xlim=c(4,12), breaks=seq(4,12,by=.5),
freq=F, xlab="temperature", ylab="density", main = "")
mu <- mean ( temps )
sig <- sqrt ( var ( temps ) )
x <- seq ( 4, 12, length=60 )
5.7. NORMAL 317

lines ( x, dnorm(x,mu,sig), lty=2 )

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.

• 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 t−8.08 2
e− 2 (
1
P [t ∈ (7.5, 8.0]] = √ 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 5.20. Let Y ∼ N(µ, σ). Then

σ2 t 2
MY (t) = e 2 +µt .
5.7. NORMAL 318

Proof.
Z
1 1 2
MY (t) = ety √ e− 2σ2 (y−µ) dy
Z 2πσ
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
= e− 2σ2 √
2πσ
2µσ2 t+σ4 t2
=e 2σ2

σ2 t 2
=e 2 +µt .

The technique used in the proof of Theorem 5.20 is worth remembering, so let’s look
at it more abstractly. Apart from multiplicative constants, the first integral in the proof is
Z Z
ty − 2σ1 2 (y−µ)2
e− 2σ2 (y−µ) +ty dy
1 2
e e dy =

The exponent is quadratic in y and therefore, for some values of a, b, c, d, e, and f can be
written !2
1 1 y−d
− 2 (y − µ) + ty = ay + by + c = −
2 2
+f
2σ 2 e
This last expression has the form of a Normal distribution with mean d and SD e. So the
integral can be evaluated by putting it in this form and manipulating the constants so it
becomes the integral of a pdf and therefore equal to 1. It’s a technique that is often useful
when working with integrals arising from Normal distributions.
Theorem 5.21. Let Y ∼ N(µ, σ). Then
E[Y] = µ and Var(Y) = σ2 .
Proof. For the mean,
σ2 t 2
E[Y] = MY0 (0) = (tσ2 + µ)e 2 +µt = µ.
t=0
For the variance,
σ2 t 2 σ2 t 2
E[Y 2 ] = MY00 (0) = σ2 e 2 +µt + (tσ2 + µ)2 e 2 +µt = σ2 + µ2 .
t=0
So,
Var(Y) = E[Y 2 ] − E[Y]2 = σ2 .

5.7. NORMAL 319

The N(0, 1) distribution is called the standard Normal distribution. As Theorem 5.22
shows, all Normal distributions are just shifted, rescaled versions of the standard Nor-
mal distribution. The mean is a location parameter and the standard deviation is a scale
parameter, as shown by Theorem 5.22.
Theorem 5.22.
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
σ2 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
+µt/σ
MX (t) = e−µt/σ MY (t/σ) = e−µt/σ e 2 = 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 5.17, then prove it.
Theorem 5.23. Let Y1 , . . . , Yn ∼ i.i.d. N(0, 1). Define X = Yi2 . Then X ∼ χ2n .
P

Proof. Start with the case n = 1.

Z
tY12 2 1 y2
MX (t) = E[e ] = ety √ e− 2 dy
Z 2π
1 − 12 (1−2t)y2
= √ e dy
2π
Z √
1 − 2t − 12 (1−2t)y2
= (1 − 2t)−1/2 √ e 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 320

5.7.2 The Multivariate Normal Distribution

Let X~ be an n-dimensional random vector with mean µ ~ and covariance matrix Σ ~ . We
X X
~ has a multivariate Normal distribution if its joint density is
say that X
1
e− 2 (~x−µX~ ) Σ
1 t −1 (~
pX~ (~x) = 1
x−µ ~)
X (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 315) 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 σ2 .
To become more familiar with the multivariate Normal distribution, we begin with the
case where the covariance matrix is diagonal:

σ1 0 0 · · ·
 2 
 0 σ2 0 · · ·
2
Σ =  . . . .. 
 
 0 0 . 
 . .
.. .. · · · σ2 

n

In this case the joint density is

1
e− 2 (~x−µX~ ) Σ
1 t −1 (~
pX~ (~x) = 1
x−µ ~)
X

(2π)n/2 |Σ| 2
!n Y n 2
1 − 21 ni=1 (xi σ−µ2i )
!
1
P
= √ e i

2π i=1 σi
n 
x −µ 2
 !
Y 1 − 21 iσ i
= √ e i ,
i=1 2πσ i

the product of n separate one dimensional Normal densities, one for each dimension.
Therefore the Xi ’s are independent and Normally distributed, with Xi ∼ N(µi , σi ). Also
see Exercise 34.
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.
5.7. NORMAL 321

~ = (X1 , . . . , Xn ) ∼
Theorem 5.24. Let X N(µ, Σ) where Σ has the so-called block-diagonal
form
 Σ11
 
012 ··· 01m 
 021 Σ22 ··· 02m 
 
Σ =  .. .. .. .. 
 . . . . 
0m1 · · · 0mm−1 Σmm
 

where Σii is an ni × ni matrix, 0i j is an ni × n j matrix of 0’s and m1 ni = n. Partition

P
~ to conform with Σ and define Y
X ~i ’s: Y
~1 = (X1 , . . . , Xn1 ), Y
~2 = (Xn1 +1 , . . . , Xn1 +n2 ), . . . ,
~m = (Xn1 +···+nm−1 +1 , . . . , Xnm ) and νi ’s: ν1 = (µ1 , . . . , µn1 ), ν2 = (µn1 +1 , . . . , µn1 +n2 ), . . . ,
Y
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 transformation, so

pY~1 ,...,Y~m (~y1 , . . . , ~ym )

1
e− 2 (~y−µ) Σ
1 t −1 (~
= pX~ (~y1 , . . . , ~ym ) = 1
y−µ)
(2π)n/2 |Σ| 2

1 − 21 m yi −νi )t Σ−1
=
P
i=1 (~ yi −νi )
ii (~
Q m 1
e
n/2
(2π) i=1 |Σii |
2

m
Y 1
e− 2 (~yi −νi ) Σii (~yi −νi )
1 t −1
= 1
n /2
i |Σ | 2
i=1 (2π) ii


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 form
is constant. But when Σ is diagonal, (~x − µ)t Σ−1 (~x − µ) = n1 (xi − µi )2 /σ2i so pX~ (~x) = c is
P
the equation of an ellipsoid centered at µ and with eccentricities determined by the ratios
σi /σ j .
What does this density look like? It is easiest to answer that question in two dimen-
sions. Figure 5.11 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
5.7. NORMAL 322

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.

Figure 5.11 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="." )
5.7. NORMAL 323

(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 5.11: 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.
5.7. NORMAL 324

• The code makes heavy use of the fact that X1 and X2 are independent for (a) calcu-
lating 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 i j’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 − 12 (~y−µY~ )t Σ−1 (~y−µY~ )

pY~ (~y) = 1
e Y~ ,
(2π)n/2 |ΣY~ | 2

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 ~ = p ~ , therefore that Z~ and Y
Z 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 Z~
is |Σ|1/2 , the square root of the determinant of Σ. Therefore,
 
pZ~ (~y) = pX~ Σ−1/2 (~y − µ) |Σ|−1/2
1 −1/2 t −1/2
= √ e−1/2(Σ (~y−µ)) (Σ (~y−µ)) |Σ|−1/2
2πn
1 − 21 (~y−µ)t Σ−1 (~y−µ)
= e
(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 dimen-
sions. Figure 5.12 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.
5.7. NORMAL 325

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 5.12: 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 326

Figure 5.12 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
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] )
5.7. NORMAL 327

5.7.3 Marginal, Conditional, and Related Distributions

We conclude this section with some theorems about Normal random variables that will
prove useful later.
Theorem 5.25. Let X ~ ∼ N(µ, Σ) be an n-dimensional Normal random variable; 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. 268),

pY~ (~y) = pX~ (A−1~y)|A−1 |

1
e− 2 (A ~y−µX~ ) Σ (A ~y−µX~ )
1 −1 t −1 −1
= 1
(2π)n/2 |A||Σ| 2
1
e− 2 (A (~y−AµX~ )) Σ (A (~y−AµX~ ))
1 −1 t −1 −1
= 1
(2π)n/2 |A||Σ| 2
1
e− 2 (~y−AµX~ ) (A ) Σ A (~y−AµX~ )
1 t −1 t −1 −1
= 1
(2π)n/2 |A||Σ| 2
1 1 t t −1
= 1
e− 2 (~y−AµX~ ) (AΣA ) (~y−AµX~ )
n/2
(2π) |AΣA | 2 t

which we recognize as the N(Aµ, AΣAt ) density. 

Corollary 5.26. Let X ~ ∼ N(µ, Σ) be an n-dimensional Normal random variable; 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 35. 
~
!
Let X~ ∼ N(µ, Σ). Divide X ~ into two parts: X ~ = X1 where X ~ 1 consists of the first
X~2
several components of X ~ and X ~ 2 consists of the last several components. Then the mean
µ1 Σ1,1 Σ1,2
! !
and covariance matrix can be divided conformably: µ = ; Σ = where
µ2 Σ2,1 Σ2,2
Σ2,1 = Σt1,2 . We often need to know the marginal distribution of X ~ 1 or the conditional
~ 2 given X
distribution of X ~ 1 . Those are given by Theorems 5.27 and 5.28. Our proofs
follow Anderson [1984].
5.7. NORMAL 328

~ 1 ∼ N(µ1 , Σ1,1 ).
Theorem 5.27. X

Proof. Let B = −Σ2,1 Σ−1

1,1 and define

~1 ~1
! ! !
~≡ Y I 0 X
Y ~2 = B I X
Y ~2 .

~ has a Normal distribution. The mean is

Theorem 5.25 says Y

µ µ
! ! !
~ = I 0
E(Y) 1
= 1
.
B I µ2 Bµ1 + µ2

The covariance matrix is

!t
Σ Σ
! !
~ = I 0 1,1 1,2 I 0
Cov(Y)
B I Σ2,1 Σ2,2 B I
Σ1,1 Σ1,2
! !
I Bt
=
BΣ1,1 + Σ2,1 BΣ1,2 + Σ2,2 0 I
Σ1,1 Σ1,1 Bt + Σ1,2
!
=
BΣ11 + Σ2,1 BΣ1,1 Bt + Σ2,1 Bt + BΣ1,2 + Σ2,2
Σ1,1
!
0
= .
0 Σ2,2 − Σ2,1 Σ−1
1,1 Σ1,2

Then Theorem 5.24 says Y ~1 ∼ N(µ1 , Σ1,1 ). It also says Y ~2 ∼ N(Bµ1 + µ2 , Σ2,2 − Σ2,1 Σ−1 Σ1,2 ),
1,1
a fact we shall use in the proof of Theorem 5.28 
~2 | X
~ 1 ∼ N µ2 + Σ2,1 Σ−1 (X ~ 1 − µ1 ), Σ2,2 − Σ2,1 Σ−1 Σ1,2 .
 
Theorem 5.28. X 1,1 1,1

Proof. Let B = −Σ2,1 Σ−1

1,1 and define

~1 ~1
! ! !
~≡ Y I 0 X
Y ~2 = B I X
Y ~2 .

The distribution of Y~ was worked out in the proof of Theorem 5.27. Here we work back-
~1
!
X ~ 1 to get
wards and derive the joint density of ~ , then divide by the marginal density of X
X2
the density ~2 | X
of X ~ 1 . This may seem strange, because we already know the joint density
~
!
X
of ~ 1 . However, in this proof we shall write the joint density in a different way.
X2
5.7. NORMAL 329

Because the Jacobian of the transformation is 1, the joint density pX~ 1 ,X~ 2 (~x1 , ~x2 ) is found
by writing the joint density pY~1 ,Y~2 (~y1 , ~y2 ) = pY~1 (~y1 )· pY~2 | Y~1 (~y2 | ~y1 ) = pY~1 (~y1 )· pY~2 (~y2 ) (because
~1 ⊥ Y
Y ~2 ), then substituting ~x1 for ~y1 and B~x1 + ~x2 for ~y2 . Let n2 be the length of X ~ 2 and
Σ2,2 | 1 = Σ2,2 − Σ2,1 Σ1,1 Σ1,2 . Then,
−1

pX~ 1 ,X~ 2 (~x1 , ~x2 )

pX~ 2 | X~ 1 (~x2 | ~x1 ) =
pX~ 1 (~x1 )
pY~1 ,Y~2 (~x1 , B~x1 + ~x2 )
=
pY~1 (~x1 )
pY~1 (~x1 )pY~2 (B~x1 + ~x2 )
=
pY~1 (~x1 )
= pY~2 (B~x1 + ~x2 )
1 h
− 12 (B~x1 +~x2 −(Bµ1 +µ2 ))t Σ−1 x1 +~x2 −(Bµ1 +µ2 ))
i
= n2 1
e 2,2 | 1 (B~

(2π) 2 |Σ2,2 | 1 | 2
1 h t
− 12 (~x2 −(µ2 −B(~x1 −µ1 ))) Σ−1 ~
2,2 | 1 ( x2 −(µ2 −B(~x1 −µ1 )))
i
= n2 1
e ,
(2π) 2 |Σ2,2 | 1 | 2

which is the Normal density with mean E[X ~2 | X

~ 1 ] = µ2 − B(~x1 − µ1 ) = µ2 + Σ2,1 Σ−1 (~x1 − µ1 )
1,1
~ ~
and covariance matrix Cov(X2 | X1 ) = Σ2,2 | 1 = Σ2,2 − Σ2,1 Σ1,1 Σ1,2 .
−1


Theorem 5.29. Let X1 , . . . , Xn ∼ i.i.d. N(µ, σ). Define S 2 ≡ ni=1 (Xi − X̄)2 . Then X̄ ⊥ S 2 .
P

~ = (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 .
5.8. THE T DISTRIBUTION 330

− X̄) = 0. Therefore, (Xn − X̄) = −

Pn Pn−1
1. i=1 (Xi i=1 (Xi − X̄). And therefore

n−1
 n−1 2 n−1  n−1 2
X X X X 
S = (Xi − X̄) +  (Xi − X̄) = Yi +  Yi 
2 2
 2

i=1 i=1 i=1 i=1

is a function of (Y1 , . . . , Yn−1 )t .

2.
1 − n1 − 1n − 1n · · · − 1n 
 
 − 1 1 − 1 − 1 · · · − 1 
 n n n n
~  .. .
. .
. .
. .
.

~ ~
Y =  . . . . .  ≡ AX
 X
 1
 − n − 1n · · · 1 − 1n − 1n 
 1 1 1 1 
n n n
··· n

where the matrix A is defined by the preceding equation, so

~ ∼ N(Aµ, σ2 AAt ). The first n − 1 rows of A are each orthogonal to the last row.
Y
Therefore
Σ11 ~0 
 
AA =  ~ t
t 
0 1/n

where Σ11 has dimension (n − 1) × (n − 1) and ~0 is the (n − 1)-dimensional vector of

0’s. Thus, by Theorem 5.24, (Y1 , . . . , Yn−1 )t ⊥ Yn .

3. Follows immediately from 1 and 2.

5.8 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 suppose 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
5.8. THE T DISTRIBUTION 331
 P 1/2
useful statement because we don’t know σ. So we 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 5.29 that X̄ ⊥ σ̂.
Theorem 5.31 gives the distribution of S 2 = nσ̂2 = (Xi − X̄)2 . First we need a lemma.
P

Lemma 5.30. 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)

Theorem 5.31. Let X1 , . . . , Xn ∼ i.i.d. N(µ, σ). Define S 2 =
Pn
i=1 (Xi − X̄)2 . Then

S2
∼ χ2n−1 .
σ 2

Proof. Let
n 
X Xi − µ  2
V= .
i=1
σ
Then V ∼ χ2n and
n !2
X (Xi − X̄) + (X̄ − µ)
V=
i=1
σ
n 2 !2 n
X̄ − µ
!
X Xi − X̄ X
= +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
5.8. THE T DISTRIBUTION 332

where S 2 /σ2 ⊥ V2 and V2 ∼ χ21 . But also,

n−1 
X Xi − µ  2  X − µ 2
n
V= + ≡ W1 + W2
i=1
σ σ

where W1 ⊥ W2 , W1 ∼ χ2n−1 and W2 ∼ χ21 . Now the conclusion follows by Lemma 5.30. 

Define r √
n − 1 X̄ − µ n(X̄ − µ)/σ
!
T≡ √ = p .
n σ̂/ n S 2 /(n − 1)σ2
Then√ by Corollary 5.29 and Theorem 5.31,2 T has the distribution of
U/ V/(n − 1) where U ∼ N(0, 1), V ∼ χn−1 , and U ⊥ V. This distribution is called the t
distribution with n − 1 degrees of freedom. We write T ∼ tn−1 . Theorem 5.32 derives its
density.

Theorem 5.32. Let U ∼ N(0, 1), V ∼ χ2p , and U ⊥ V. Then T ≡ U/ V/p has density
p

p p+1
Γ( p+1 Γ( p+1
!−
)p 2  2 − p+1 ) t2 2
pT (t) = 2
√ t +p
2
= p √ 2
1+ .
Γ( 2p ) 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
TY 2
U= √ and V=Y
p

The Jacobian is
1 1 1
dU dU Y√2 T Y√− 2 Y2
dT
dV
dY
dV = p 2 p = √
dT dY 0 1 p

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
5.8. THE T DISTRIBUTION 333

Therefore the joint density of (T, Y) is

1
1 t2 y 1 p y y2
pT,Y (t, y) = √ e− 2p 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 t2
= √ p √ y 2 −1 e− 2 ( p +1) dy
2πΓ( 2p )2 2 p 0
! p+1
1 p+1 2p 2
= √ p+1 √ Γ( 2 ) t2 + p
πΓ( 2p )2 2 p
Z ∞
1 p+1 − y
2 −1 e 2p/(t2 +p) dy
×  2p  p+1 y
0
Γ( p+1
2
) 2
t2 +p

Γ( p+1 )p p/2   − p+1

= 2
√ t2 + p 2

Γ( 2p ) π
!− p+1
Γ( p+1 ) t2 2
= p √
2
1+ .
Γ( 2 ) pπ p

We formalize Theorem 5.32 with a definition. A random variable T with density
p p+1
Γ( p+1 Γ( p+1
!−
)p 2  2 − p+1 ) t2 2
pT (t) = 2
√ t +p
2
= p √ 2
1+ .
Γ( 2p ) π Γ( 2 ) pπ p
is said to have a t distribution, or a Student’s t distribution, with p degrees of freedom. The
distribution is named for William Sealy Gosset who used the pseudonym Student when he
wrote about the t distribution in 1908.
Figure ?? 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 t p density appears to approach the N(0, 1) density. (The
appearance is correct. See Exercise 36.)
5.8. THE T DISTRIBUTION 334

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 5.13: t densities for four degrees of freedom and the N(0, 1) density
5.9. EXERCISES 335

Figure ?? 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 )
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 we said the quantity n(X̄ − µ)/σ̂ had a N(0, 1)
distribution,
√ approximately. Theorem 5.32 derives the density of the related quantity
n − 1(X̄ − µ)/σ̂ which has a tn−1 distribution, exactly. Figure ?? shows how similar
those distributions are. The t distribution has slightly more spread than the N(0, 1) distri-
bution, 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 ∼ t p , then
! p+1
Γ( p+1 2 − 2
Z ∞
) t
E[T ] = t p √ 2
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 t p 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. Because the t distribution does not have
moments of all order, it does not have a moment generating function.

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.
5.9. EXERCISES 336

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 prob-
ability 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.
(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. Explain the role of qnbinom(...) in the R code for Figure 5.2.
5. 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 distribu-
tion”, answer “Poi(3)” or instead of answering “Binomial”, answer “Bin(n, p) where
n = 13 but p is unknown.”
(a) The temperature measured at a randomly selected point on the surface 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 develop
smallpox.
(e) The amount of Mercury in a fish caught in Lake Ontario.
6. 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 calculated.
7. 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.
5.9. EXERCISES 337

(a) the number of times you win

(b) the number of times you lose
(c) the number of losses until your third win (If you don’t win three times in 100
tries, you keep playing.)
(d) the number of wins until your thirtieth loss (If you don’t lose thirty times in
100 tries, you keep playing.)
(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
(j) the number of times you are accidentally jostled by the person standing behind
you

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 PET 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
5.9. EXERCISES 338

source emits photons in all directions, the only ones detected are those that are emit-
ted 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. (a) Prove Theorem 5.12 using moment generating functions.
(b) Prove Theorem 5.13 using moment generating functions.
11. (a) Prove Theorem 5.15 by finding E[Y 2 ] using the trick that was used to prove
Theorem 5.13.
(b) Prove Theorem 5.15 by finding E[Y 2 ] using moment generating functions.
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.
(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 sur-
rounding 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.
5.9. EXERCISES 339

(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 variable relevant
to this problem that has approximately a N(0,1) distribution.

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.

15. An article on statistical fraud detection (Bolton and Hand [1992]), when talking
about records in a database, says:
"One of the difficulties with fraud detection is that typically there are many legiti-
mate 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. In 1988 men averaged around 500 on the math SAT, the SD was around 100 and the
histogram followed the normal curve.
5.9. EXERCISES 340

(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?

This question was inspired by Freedman et al. [1998].

17. Multiple choice.

(a) X ∼ Poi(λ). Pr[X ≤ 7] =

i. 7x=−∞ e−λ λ x /x!
P

ii. 7x=0 e−λ λ x /x!

P

iii. 7λ=0 e−λ λ x /x!

P

(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). 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]
iv. Pr[−2 ≤ X̄ ≤ 2]
v. Pr[−2 ≤ Xi ≤ 2]
vi. Pr[−2 ≤ Y ≤ 2]
vii. Pr[−20 ≤ X̄ ≤ 20]
viii. Pr[−20 ≤ Xi ≤ 20]
5.9. EXERCISES 341

ix. Pr[−20 ≤ Y ≤ 20]

(e) X ∼ Bin(100, θ). 100
θ=0 f (x|θ) =
P

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, 1/λ) and are mutually independent.
f (x1 , . . . , xn ) =
i. [λr /(r − 1)!]( xi )r−1 e−λ xi
Q P

ii. [λnr /((r − 1)!)n ]( xi )r−1 e−λ xi

Q Q

iii. [λnr /((r − 1)!)n ]( xi )r−1 e−λ xi

Q P

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. Prove a version of Equation 5.1 on page 290. Let k = 3. Start from the joint pmf of
Y1 and Y2 (Use Theorem 5.7.), derive the marginal pmf of Y1 , and identify it.
21. Ecologists who study forests have a concept of seed rain. The seed rain in an area is
the number of seeds that fall on that area. Seed rain is a useful concept in studying
how forests rejuvenate and grow. After falling to earth, some seeds germinate and
become seedlings; others do not. For a particular square-meter quadrat, let Y1 be
the number of seeds that fall to earth and germinate and Y2 be the number of seeds
that fall to earth but do not germinate. Let Y = Y1 + Y2 . Adopt the statistical model
Y1 ∼ Poi(λ1 ); Y2 ∼ Poi(λ2 ) and Y1 ⊥ Y2 . Theorem 5.10 says Y ∼ Poi(λ) where
λ = λ1 + λ2 . Find the distribution of Y1 given Y. I.e., find P[Y1 = y1 | Y = y].
22. Prove Theorem 5.11. Hint: use moment generating functions and the fact that
limn→∞ (1 + 1/n)n = e.
23. (a) Let Y ∼ U(1, n) where the parameter n is an unknown positive integer. Suppose
we observe Y = 6. Find the m.l.e. n̂. Hint: Equation 5.2 defines the pmf for
y ∈ {1, 2, . . . , n}. What is p(y) when y < {1, 2, . . . , n}?
5.9. EXERCISES 342

(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. T̂ .

24. Let Y be a continuous random variable, Y ∼ U(a, b).

(a) Find E[Y].

(b) Find Var(Y).
(c) Find MY (t).

25. (a) Is there a discrete distribution that is uniform on the positive integers? 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(−∞, ∞).

26. Ecologists are studying pitcher plants in a bog. To a good approximation, pitcher
plants follow a Poisson process with parameter λ. That is, (i) in a region of area A,
the number of pitcher plants has a Poisson distribution with parameter Aλ; and (ii)
the numbers of plants in two disjoint regions are conditionally independent given λ.
The ecologists want to make an inference about λ.

(a) This section of the problem describes a simplified version of the ecologists’
sampling plan. They choose an arbitrary site s near the middle of the bog.
Then they look for pitcher plants near s and record the location of the nearest
one. Let r be the distance from s to the nearest pitcher plant and let q = r2 .
i. Find the density of q.
ii. Find the density of r.
iii. Find the m.l.e., λ̂.
iv. Answer the previous questions under the supposition that the ecologists
had recorded the distance to the k’th nearest pitcher plant. (The previous
questions are for k = 1.)
(b) This section of the problem describes the ecologists’ sampling plan more ac-
curately. First, they choose some sites s1 , . . . , sn in the bog. Then they go to
each site, find the nearest pitcher plant and record its location. The data are
these locations L1 , . . . , Ln . Some sites may share a nearest plant, so some of
5.9. EXERCISES 343

the Li ’s may be referring to the same plant. Let Di be the distance from si to Li
and Di, j be the distance from si to L j .
i. Let D1 be the distance from s1 to L1 . Find the density p(d1 |λ). You may
assume that pitcher plants arise according to a homogeneous Poisson pro-
cess with rate λ. Hint: use the relationship between a Poisson process and
the exponential distribution.
ii. When the ecologists go to s2 they discover that the nearest plant is the
same plant they already found nearest to s1 . In other words, they discover
that D2 = D2,1 . Find Pr[D2 = D2,1 |λ].
iii. Find the likelihood function `(λ) ≡ p[L1 , . . . , Ln |λ].
iv. Find the m.l.e. λ̂ ≡ argmaxλ `(λ).
v. Suppose the prior distribution for λ is Gam(α, β). Find the posterior dis-
tribution for λ.

27. Let x ∼ Gam(α, β) and let y = 1/x. Find the pdf of y. We say that y has an inverse
Gamma distribution with parameters α and β and write y ∼ invGam(α, β).

28. Prove Theorem 5.18. Hint: Use the method of Theorem 5.13.

29. Let x1 , . . . , xn ∼ i.i.d. U(0, 1). Find the distribution of x(n) , the largest order statistic.

30. In the R code to create Figure 2.19, explain how to use dgamma(...) instead of
dpois(...).

31. Prove the claim on page 309 that the half-life of a radioactive isotope is m = λ log 2.

32. Prove Theorem 5.16.

33. Prove Theorem 5.19.

34. Page 320 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 Equation 5.6.
(d) Derive the mean vector and covariance matrix of X1 , . . . , Xn .
5.9. EXERCISES 344

35. Prove Corollary 5.26.

36. Show, for every x ∈ R,
1 x2
lim pt p (x) = √ e− 2
p→∞ 2π
where pt p (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 Statistical Inference, 2nd
ed. by Casella and Berger.)
37. It turns out that the t distribution with p degrees of freedom can be written as a
mixture of Normal distributions, a fact that is sometimes useful p in statistical calcu-
lations. Let τ ∼ Gam(p/2, 2) and, conditional on τ, y ∼ N(0, p/τ). Show that the
marginal distribution of y is the t distribution with p degrees of freedom.
38. This exercise provides a generalization of Theorem 5.25. In that theorem, X ~ ∼
N(µ, Σ) and A is a full-rank (invertible) n by n matrix. In this exercise we let A be an
m by n matrix where m < n and A has rank m. Let Y = AX. What is the distribution
of Y?
(a) Explain why A can be augmented with n − m additional rows, which we think
of as a matrix B, to create an n by n full-rank matrix
!
A
A =
∗
B
You may cite results from linear algebra.
(b) Use Theorem 5.25 to derive the distribution of A∗ .
(c) Use Theorem 5.27 to derive the distribution of A.
(d) In Section 3.2.2 we saw that the m.l.e. of the coefficients in a Normal linear
model, under the assumption Y ∼ N(XB, σ2 I) is B̂ = (X t X)−1 X t Y. Since Y is a
random variable, and B̂ is a transformation of Y, B̂ is also a random variable.
Find its distribution.
39. This exercise will be useful in the Bayesian analysis of Normal distributions. It is a
generalization of Exercise 28. Let τ ∼ Gam(α, β) and, conditional on τ, let µ have a
Normal distribution with mean m and variance v/τ. All of α, β, m and v are known.
(a) Find the joint density of (µ, τ).
(b) Find the marginal density of µ. Show that µ is a linear transformation of a
random variable that has a t distribution. Find its degrees of freedom.
5.9. EXERCISES 345

Given (µ, τ), let Y1 , . . . , Yn be a random sample from the Normal distribution with
mean µ and precision τ. Our goal is to find the posterior distribution of (µ, τ).

(c) Write an expression that is proportional to p(µ, τ | y1 , . . . , yn ).

(d) Show that the conditional posterior p(µ | τ, y1 , . . . , yn ) is Normal. Find its mean
and variance.
(e) Show that the marginal posterior p(τ | y1 , . . . , yn ) is Gamma. Find its parame-
ters.
(f) Find the marginal posterior density p(µ | y1 , . . . , yn ). Show that it represents a
linear transformation of a random variable that has a t distribution. Find its
degrees of freedom.
 Chapter 6

Bayesian Statistics

6.1 Multidimensional Bayesian Analysis

This chapter takes up Bayesian statistics. Modern Bayesian statistics relies heavily on
computers, computation, programming, and algorithms, so that will be the major focus of
this chapter. We cannot give a complete treatment here, but there are several good books
that cover these topics in more depth. See, for example, Gelman et al. [2004], Liu [2004],
Marin and Robert [2007], or Robert and Casella [1997].
Recall the framework of Bayesian inference from Section 2.5.
• We posit a parametric family of distributions {p(y | θ)}.
• We express our old knowledge of θ through a prior probability density p(θ).
• The previous two items combine to yield p(y, θ) and, ultimately, p(θ | y).
• The posterior density p(θ | y) represents our new state of knowledge about θ.
The posterior density is
p(θ)p(y | θ)
p(θ | y) = R ∝ p(θ)p(y | θ). (6.1)
p(θ)p(y | θ) dθ
So far, so good. But in many interesting applications, θ is multi-dimensional and
problems arise when we want to examine the posterior. Equation 6.1 tells us how to
evaluate the posterior at any value of θ, but that’s not always sufficient for getting a sense
of which values of θ are most likely, somewhat likely, unlikely, etc. One way to develop a
feeling for a multidimensional posterior is to examine a marginal posterior density, say
Z Z
p(θ1 | y) = · · · p(θ1 , . . . , θk | y) dθ2 . . . dθk . (6.2)

346
6.1. MULTIDIMENSIONAL BAYESIAN ANALYSIS 347

Unfortunately, the integral in Equation 6.2 is often not analytically tractable and must be
integrated numerically. Standard numerical integration techniques such as quadrature may
work well in low dimensions, but in Bayesian statistics Equation 6.2 is often sufficiently
high dimensional that standard techniques are unreliable. Therefore, new numerical inte-
gration techniques are needed. The most important of these is called Markov chain Monte
Carlo integration, or MCMC. Other techniques can be found in the references at the be-
ginning of the chapter. For the purposes of this book, we investigate MCMC. But first,
to get a feel for Bayesian analysis, we explore posteriors in low dimensional, numerically
tractable examples.
The general situation is that there are multiple parameters θ1 , . . . , θk , and data y1 , . . . ,
yn . We may be interested in marginal, conditional, or joint distributions of the parameters
either a priori or a posteriori. Some examples:
• p(θ1 , . . . , θk ), the joint prior

• p(θ1 , . . . , θk | y1 , . . . , yn ), the joint posterior

• p(θ1 | y1 , . . . , yn ) = · · · p(θ1 , . . . , θn | y1 , . . . , yn ) dθ2 · · · dθk , the marginal posterior

R R

of θ1

p(θ2 , . . . , θk | θ1 , y1 , . . . , yn ) = p(θ1 , . . . , θk | y1 , . . . , yn )/p(θ1 | y1 , . . . , yn )

∝ p(θ1 , . . . , θk | y1 , . . . , yn ),

the conditional joint posterior density of (θ2 , . . . , θk ) given θ1 , where the “∝” means
that we substitute θ1 into the numerator and treat the denominator as a constant.
The examples in this section illustrate the ideas.
Example 6.1 (Ice Cream Consumption, cont.)
This example continues Example 3.5 in which weekly ice cream consumption is mod-
elled as a function of temperature and possibly other variables. We begin with the
model in Equation 3.10:

consumption = β0 + β1 temperature + error

or
yi = β0 + β1 xi + i
where 1 , . . . , 30 ∼ i.i.d.N(0, σ). For now, let us suppose that σ is known. Later we’ll
drop that assumption. Because σ is known, there are only two parameters: β0 and β1 .
6.1. MULTIDIMENSIONAL BAYESIAN ANALYSIS 348

For a Bayesian analysis we need a prior distribution for them; then we can compute
the posterior distribution. For now we adopt the following prior without comment. Later
we will see why we chose this prior and examine its consequences.

β0 ∼ N(µ0 , σ0 )
β1 ∼ N(µ1 , σ1 )
β0 ⊥ β1

i. e.

β0 µ0 σ20 0
! ! !!
∼N ,
β1 µ1 0 σ21
for some choice of (µ0 , µ1 , σ0 , σ1 ). The likelihood function is

`(β0 , β1 ) = p(y1 , . . . , y30 | β0 , β1 )

Y30
= p(yi | β0 , β1 )
i=1
30 
y −(β +β x ) 2

Y 1 − 12 i 0σ 1 i
= (√ )e
i=1 2πσ

β0 µ0
! !
To find the posterior density we will use matrix notation. Let β = , µ = ,
β1 µ1
σ2 0 ~
!
Σ= 0 , Y = (Y1 , . . . , Yn )t and
0 σ21
 
1 temperature1 
1 temperature 
2
X =  .. ..  .
 . . 

1 temperature30


~ has a 30-dimensional Normal distribution with mean Xβ and co-

Conditional on β, Y
variance matrix σ I30 . The posterior density is proportional to the prior times the likeli-
2

hood.
~ ~
~ ∝ e− 2 (β−µ) Σ
p(β | Y)
1 t −1 (β−µ)− 1 (Y−Xβ)
2
t (Y−Xβ)/σ2
(6.3)

At this point we observe that the exponent is a quadratic form in β. Therefore the
posterior density will be a two-dimensional Normal distribution for β and we just have
6.1. MULTIDIMENSIONAL BAYESIAN ANALYSIS 349

to complete the square to find the mean vector and covariance matrix. The exponent
is, apart from a factor of − 12 and some irrelevant constants involving the yi ’s,
~ 2 ] + · · · = (β − µ∗ )t (Σ∗ )−1 (β − µ∗ ) + · · ·
βt [Σ−1 + X t X/σ2 ]β − 2βt [Σ−1 µ + X t Y/σ
where Σ∗ = (Σ−1 + X t X/σ2 )−1 and µ∗ = Σ∗ (Σ−1 µ + X t Y/σ ~ 2 ). Therefore, the posterior
distribution of β given Y ~ is Normal with mean µ and covariance matrix Σ∗ . It is worth
∗

noting (1) that the posterior precision matrix (Σ∗ )−1 is the sum of the prior precision
matrix Σ−1 and a part that comes from the data, X t X/σ2 and (2) that the posterior mean
~ 2 ) = Σ∗ (Σ−1 µ + (X t X/σ2 )(X t X)−1 X t Y)
is Σ∗ (Σ−1 µ + X t Y/σ ~ , a matrix weighted average of
the prior mean µ and the least-squares estimate (X X)−1 X t Y
t ~ where the weights are the
two precisions Σ and X X/σ .
−1 t 2

The derivation of the posterior distribution does not depend on any particular choice
of (µ, Σ), but it does depend on the fact that the prior distribution was Normal because
that’s what gives us the quadratic form in the exponent. That’s one reason we took the
prior distribution for β to be Normal: it made the calculations easy.
Now let’s look at the posterior more closely, see what it implies for (β0 , β1 ), and
see how sensitive the conclusions are to the choice of (µ, Σ). We’re also treating σ as
known, so we’ll need a value. Let’s start with the choice
µ0 σ20 0
! ! ! !
0 106 0
µ= = ; Σ= = ; and σ = 0.05. (6.4)
µ1 0 0 σ21 0 106
The large diagonal entries in Σ say that we have very little a priori knowledge of β. We
can use R to calculate the posterior mean and covariance.
ic <- read.table ( "data/ice cream.txt", header=T )

mu <- c ( 0, 0 )
Sig <- diag ( rep ( 10^6, 2 ) )
sig <- 0.05
x <- cbind ( 1, ic$temp )
y <- ic$IC

Sigstar <- solve ( solve(Sig) + t(x) %*% x / (sig^2) )

mustar <- Sigstar %*% ( solve(Sig) %*% mu + t(x) %*% y / (sig^2) )

The result is
.207
! !
8.54 × 10−4 −1.570 × 10−5
µ =
∗
and Σ =
∗
. (6.5)
.003 −1.570 × 10−5 3.20 × 10−7
6.1. MULTIDIMENSIONAL BAYESIAN ANALYSIS 350

600
12
8

300
4
0

0
0.10 0.25 0.001 0.004

β0 β1

Figure 6.1: Posterior densities of β0 and β1 in the ice cream example using the prior from
Equation 6.4.

Note:

• solve performs matrix inversions and solves systems of linear equations.

• t(x) is the transpose of x.

• %*% performs matrix multiplication.

• The matrix product t(x) %*% y is so common it has a special shortcut, crossprod
( x, y ). We could have used crossprod to find Σ∗ and µ∗ .

Figure 6.1 shows the posterior densities. The posterior density of β0 is not very mean-
ingful because it pertains to ice cream consumption when the temperature is 0. Since
our data was collected at temperatures between about 25 and 75, extrapolating to tem-
peratures around 0 would be dangerous. And because β0 is not meaningful, neither
is the joint density of (β0 , β1 ). On the other hand, our inference for β1 is meaningful. It
says that ice cream consumption goes up about .003 (±.001 or so) pints per person
for every degree increase in temperature. You can verify whether that’s sensible by
looking at Figure 3.8.
6.1. MULTIDIMENSIONAL BAYESIAN ANALYSIS 351

Figure 6.1 was produced by the following snippet.

opar <- par ( mfrow=c(1,2) )

m <- mustar[1]
sd <- sqrt ( Sigstar[1,1] )
x <- seq ( m-4*sd, m+4*sd, length=60 )
plot ( x, dnorm ( x, m, sd ), type="l",
xlab=expression(beta[0]), ylab="" )
m <- mustar[2]
sd <- sqrt ( Sigstar[2,2] )
x <- seq ( m-4*sd, m+4*sd, length=60 )
plot ( x, dnorm ( x, m, sd ), type="l",
xlab=expression(beta[1]), ylab="" )

Now we’d like to investigate the role of the

√ prior density. We notice that the prior
SD of β1 was 103 while the posterior SD is 3.2 × 10−7 ≈ 5.7 × 10−4 . In other words,
the data has reduced the uncertainty by a huge amount. Because there’s so much
information in the data, we expect the prior to have little influence. We can verify that
by considering priors with different SD’s and comparing their posteriors. To that end,
consider the prior
µ0 σ20 0
! ! ! !
0 1 0
µ= = ; Σ= = ; and σ = 0.05. (6.6)
µ1 0 0 σ21 0 1
With prior 6.6, the posterior would be
.207
! !
8.53 × 10−4 −1.568 × 10−5
µ =
∗
and Σ =
∗
,
.003 −1.568 × 10−5 3.19 × 10−7
nearly identical to posterior 6.5. That’s because the data contain much more infor-
mation than the prior, so the prior plays a negligible role in determining the posterior
distribution. The posterior precision matrix
! (inverse covariance matrix) is Σ−1 + X t X/σ2 .
30 1473
In our case, X t X/σ2 = , which has entries much bigger than those in Σ−1 ,
1473 80145
whichever prior we use. Even if we did a careful job of assessing our prior, it would be
influential only if our prior precision matrix had entries of the same order of magnitude
as X t X/σ2 . Since that’s unlikely — our true a priori variances are probably not as small
as 1/30 — there’s little to be gained by choosing the prior carefully and much effort to
be saved by using an arbitrary prior, as long as it has reasonably large variances.
6.1. MULTIDIMENSIONAL BAYESIAN ANALYSIS 352

ring ID xcoor ycoor spec dbh 1998 1999 2000

1 11003 0.71 0.53 pita 19.4 0 0 0
1 11004 1.26 2.36 pita 14.1 0 0 4
1 11011 1.44 6.16 pita 19.4 0 6 0
1 11013 3.56 5.84 pita 21.6 0 0 0
1 11017 3.75 8.08 pita 10.8 0 0 0
..
.
6 68053 0.82 10.73 pita 14.4 0 0 0
6 68055 -2.24 13.34 pita 11 0 0 0
6 68057 -0.78 14.21 pita 8 0 0 0
6 68058 0.76 14.55 pita 10.6 0 0 0
6 68059 1.48 13 pita 21.2 0 5 10

Table 6.1: The numbers of pine cones on trees in the FACE experiment, 1998–2000.

Example 6.2 (Pine Cones)

One possible result of increased CO2 in the atmosphere is that plants will use some of
the excess carbon for reproduction, instead of growth. They may, for example produce
more seeds, produce bigger seeds, produce seeds earlier in life, or produce seeds
when they, the plants, are smaller. To investigate this possibility in the Duke FACE
experiment (See Example 1.12 and its sequels.) a graduate student went to the FACE
site each year and counted the number of pine cones on pine trees in the control and
treatment plots (LaDeau and Clark [2001] and Example 3.8).
The data are in Table 6.1. The first column is ring. Rings 1, 5, and 6 were control;
2, 3, and 4 were treatment. The next column, ID, identifies each tree uniquely; xcoor
and ycoor give the location of the tree. The next column, spec, gives the species;
pita stands for pinus taeda, or loblolly pine, the dominant canopy tree in the FACE
experiment. The column dbh gives diameter at breast height, a common way for
foresters and ecologists to measure the size of a tree. The final three columns show
the number of pine cones in 1998, 1999, and 2000. We investigate the relationship
between dbh and the number of pine cones, and whether that relationship is the same
in the control and treatment plots.
Figures 6.2, 6.3, and 6.4 plot the numbers of pine cones as a function of dbh in
the years 1998–2000. In 1998, very few trees had pine cones and those that did had
very few. But by 1999, many more trees had cones and had them in greater number.
There does not appear to be a substantial difference between 1999 and 2000. As a
quick check of our visual impression we can count the fraction of pine trees having
pine cones each year, by ring. The following R code does the job.
6.1. MULTIDIMENSIONAL BAYESIAN ANALYSIS 353

for ( i in 1:6 ) {
good <- cones$ring == i
print ( c ( sum ( cones$X1998[good] > 0 ) / sum(good),
sum ( cones$X1999[good] > 0 ) / sum(good),
sum ( cones$X2000[good] > 0 ) / sum(good) ) )
}
[1] 0.0000000 0.1562500 0.2083333
[1] 0.05633803 0.36619718 0.32394366
[1] 0.01834862 0.21100917 0.27522936
[1] 0.05982906 0.39316239 0.37606838
[1] 0.01923077 0.10576923 0.22115385
[1] 0.04081633 0.19727891 0.18367347

Since there’s not much action in 1998 we will ignore the data from that year. The data
show a greater contrast between treatment (rings 2, 3, 4) and control (rings 1, 5, 6) in
1999 than in 2000. So for the purpose of this example we’ll use the data from 1999. A
good scientific investigation, though, would use data from all years.
We’re looking for a model with two features: (1) the probability of cones is an
increasing function of dbh and of the treatment and (2) given that a tree has cones,
the number of cones is an increasing function of dbh and treatment. Here we describe
a simple model with these features. The idea is (1) a logistic regression with covariates
dbh and treatment for the probability that a tree is sexually mature and (2) a Poisson
regression with covariates dbh and treatment for the number of cones given that a tree
is sexually mature. Let Yi be the number of cones on the i’th tree. Our model is

1 if the i’th tree had extra CO2

xi = 

0 otherwise


1 if the i’th tree is sexually mature

θi = 

0 otherwise

(6.7)
exp(β0 + β1 dbhi + β2 xi )
πi = P[θi = 1] =
1 + exp(β0 + β1 dbhi + β2 xi )
φi = exp(γ0 + γ1 dbhi + γ2 xi )
Yi ∼ Poi(θi φi )

This model is called a zero-inflated Poisson model. There are six unknown parame-
ters: β0 , β1 , β2 , γ0 , γ1 , γ2 . We must assign prior distributions and compute posterior
6.1. MULTIDIMENSIONAL BAYESIAN ANALYSIS 354

ring

6
5
4
3
2
1

5 10 15 20 25 5 10 15 20 25

100
80
60
40
20
cones 1998

●
●● ● ● ● ● ●
● ●●●● ●
●●●●●
●●●
●
●
●●●
●●
●●
●
●●●
●
●●●
●
●●
●●
●●
●●●
●
●
●●●
●●
●●
●●
●
●●●
●● ● ● ●● ● ●●●●●●
●
●●●●●
●●●
●●
●●●
●●●
●●●●
●●
●●●
●
●●●
●●●
●●●
●
●●
●●●
●
●●●
●
●●
●●●● ● ●●
●
●●
●
●●●
●
●●●
●●
●●
●●
●●
●●
●●
●●
●●
●
●●
●●●●
●● ●
●●●
●
●●
●●●
●●●
●●
●
●●
●●
●
●●
●●●
●●
● ●
●●●
●●
● ●

0
100
80
60
40
20

● ●●
●● ●●●
●●
●
●●●●●● ● ●
●●●●
●●●● ●
●●●
●
●●
●
●●●●
●
●●
●●●
●
●●●●
●●●
●●● ●●● ●● ● ●●
●● ●●●
●● ● ●
●●
●●●●● ●
●● ●●●●●●●
●●
● ●
●●
●
●●
●●●
●
●●● ●● ●●
● ● ● ●● ●● ●●● ●
●●
●●
●
●●●
●●
●●●
●
●●●●●
●
●●
●●
●
●●●
●
●●●
●●●●
●
●●
●●●
●●
●●
●●●
●●●● ●
0

5 10 15 20 25

dbh

Figure 6.2: Numbers of pine cones in 1998 as a function of dbh

6.1. MULTIDIMENSIONAL BAYESIAN ANALYSIS 355

ring

6
5
4
3
2
1

5 10 15 20 25 5 10 15 20 25

100
80
60
●

40
● ●
●

● ●
●

20
●
●
cones 1999

●
●●● ● ● ● ●●
● ● ●● ●●● ●
●●●
●●●● ●
●
●●
● ●● ●●
● ●
● ● ● ●● ●●● ●● ●● ● ●
● ●●●● ● ●
● ●●●●
●●●
●
● ●
●●●
●
●●
● ●
●●●●●●●●
●
● ●
●●
●
●● ●●● ● ●
●●●
● ● ●● ● ●●●●●●
●
●●●●●
●●●
●●
●●●
●●●
●●●
● ●●
●●
●●
●●
● ●●●
●●●● ●
●●
●●●
●
●●●●
●●● ● ●●
●
●●
●
●●●
●
●●●
●●
●●●●
●●●
●●●
●
●●●
●
●●
●●●●●
●
●●●●●
●
●●●●
●●
●
●●
●●
●
●
●●
●●●
●●
● ● ●●●
●

0
100
80
60

●
40

●
● ● ●
● ●
20

●● ●
● ● ●
●● ●● ●
● ●
●
● ●
● ● ●● ● ● ●●
●
●
●●●●● ●●● ● ●
● ●
● ●
●● ●●●
●●
●
●●●●●● ● ●
●●●●
●●●● ●
●●●
●
●●
●
●●●●●●●●●●●
●●●
●●● ●● ● ●● ●● ● ●
●● ●●● ●●●●● ●
●● ●● ●●●●●●
●●
● ●●●●
●● ●
●●
●● ●● ● ● ●● ●● ●●● ●
●●
●●
●
●●●
●●
●●●
●
●●●●●
●
●●
●●●
●
● ●
●●
●●●●●
●●●
●
●●
●
●●●
●
●●● ●
0

5 10 15 20 25

dbh

Figure 6.3: Numbers of pine cones in 1999 as a function of dbh

6.1. MULTIDIMENSIONAL BAYESIAN ANALYSIS 356

ring

6
5
4
3
2
1

5 10 15 20 25 5 10 15 20 25

100
●

80
60
40
● ● ●
● ● ●
●

20
●● ● ● ●
cones 2000

● ● ● ●
● ●
● ● ●●
●●● ●●● ● ● ● ● ●
● ● ●● ● ●● ●
● ●● ● ●●●●●●●●●●
● ●●●●● ●●● ●●●
●●●● ● ● ● ●● ●
●●●● ●●●● ●
● ●●●● ●
●●●●●●●
●
●
●●●
●●
●●
●●●
●
●●●
●
●●
●●●●
●●●
●●●●●
●
● ● ● ● ●●
● ●● ●
●●●●●
●●●
●●
●●●
●●●
●●●●
●●
●●●
●
●●●
●●●
●●●
●● ●●
●●●●●● ● ● ●●
●
●●
●
●●●
●
●●●
●●
●●
●●
●●●●
●●●●
●●●
●●●●●
●●
● ●●●
●
●●
●●
●
●●●●
●●
●
●●
●●
●●
●
●●●
● ● ●●●
●

0
100
80
60

● ●
40

●
●
●● ●●
● ●
●
20

● ● ●
● ● ● ●
● ● ●
● ● ● ●
● ● ● ● ● ● ●●
● ●
●
● ●●● ●●●● ● ●●●● ●●●●●● ●●●●● ●●●● ●
●● ●●●
●●
●
●●●●●● ● ● ●●●● ●
●●●● ●●●
●
●●
●
●●●●●
●
●●●●●●
●●●● ●● ● ●●
●● ●●●
●● ● ●
●● ●● ●●●●●●●● ●
●●●● ● ●● ●●● ●● ●●
● ● ● ●● ●● ●●● ●
●●
●●
●
●●●
●●
●●●
●
●● ●
●●●●
●●
●
●●
●●●●●●●●●●●●●
●●
●● ●
0

5 10 15 20 25

dbh

Figure 6.4: Numbers of pine cones in 2000 as a function of dbh

6.1. MULTIDIMENSIONAL BAYESIAN ANALYSIS 357

distributions of these parameters. In addition, each tree has an indicator θi and we will
be able to calculate the posterior probabilities P[θi = 1 | y1 , . . . , yn ] for i = 1, . . . , n.
We start with the priors β0 , β1 , β2 , γ0 , γ1 , γ2 ∼ i.i.d.U(−100, 100). This prior distribu-
tion is not, obviously, based on any substantive prior knowledge. Instead of arguing
that this is a sensible prior, we will later check the robustness of conclusions to spec-
ification of the prior. If the conclusions are robust, then we will argue that almost any
sensible prior would lead to roughly the same conclusions.
To begin the analysis we write down the joint distribution of parameters and data.

p(y1 , . . . , yn , β0 , β1 , β2 , γ0 , γ1 , γ2 )
= p(β0 , β1 , β2 , γ0 , γ1 , γ2 ) × p(y1 , . . . , yn | β0 , β1 , β2 , γ0 , γ1 , γ2 )
!6
1
= 1(−100,100) (β0 ) × 1(−100,100) (β1 ) × 1(−100,100) (β2 ) × 1(−100,100) (γ0 )
200
× 1(−100,100) (γ1 ) × 1(−100,100) (γ2 )
Y exp(β0 + β1 dbhi + β2 xi )
×
i:y >0
1 + exp(β0 + β1 dbhi + β2 xi )
i

exp(− exp(γ0 + γ1 dbhi + γ2 xi )) exp(γ0 + γ1 dbhi + γ2 xi )yi

!
×
yi !
Y 1
× +
i:y =0
1 + exp(β 0 + β 1 dbhi + β2 xi )
i

exp(β0 + β1 dbhi + β2 xi )
!
exp(− exp(γ0 + γ1 dbhi + γ2 xi )) (6.8)
1 + exp(β0 + β1 dbhi + β2 xi )
Q
In Equation 6.8 each term in the product i:yi >0 is

P[i’th tree is sexually mature] × p(yi | i’th tree is sexually mature)

Q
while each term in i:yi =0 is

P[i’th tree is immature] + P[i’th tree is mature but produces no cones].

The posterior p(β0 , β1 , β2 , γ0 , γ1 , γ2 | y1 , . . . , yn ) is proportional, as a function of

(β0 , β1 , β2 , γ0 , γ1 , γ2 ), to Equation 6.8. Similarly, conditional posteriors such as
p(β0 | β1 , β2 , γ0 , γ1 , γ2 , y1 , . . . , yn ) are proportional, as a function of β0 , to Equation 6.8.
But that doesn’t allow for much simplification; it allows us to ignore only the factorials
in the denominator.
6.2. METROPOLIS, METROPOLIS-HASTINGS, AND GIBBS 358

To learn about the posterior in, say, Equation 6.8 it is easy to write an R function
that accepts (β0 , β1 , β2 , γ0 , γ1 , γ2 ) as input and returns 6.8 as output. But that’s quite a
complicated function of (β0 , β1 , β2 , γ0 , γ1 , γ2 ) and it’s not obvious how to use the function
or what it says about any of the six parameters or the θi ’s. Therefore, in Section 6.2 we
present an algorithm that is very powerful for evaluating the integrals that often arise in
multivariate Bayesian analyses.

6.2 The Metropolis, Metropolis-Hastings, and Gibbs Sam-

pling Algorithms
In “Markov chain Monte Carlo”, the term “Monte Carlo” refers to evaluating an integral by
using many random draws from a distribution. To fix ideas, suppose we want to evaluate
Equation 6.2. Let ~θ = (θ1 , . . . , θk ). If we could generate many samples ~θ1 , . . . , ~θM of
~θ (where ~θi = (θi,1 , . . . , θi,k )) from its posterior distribution then we could approximate
Equation 6.2 by

1. discarding θi,2 , . . . , θi,k from each iteration,

2. retaining θ1,1 , . . . , θ M,1 ,

3. using θ1,1 , . . . , θ M,1 and standard density estimation techniques (page 105) to estimate
p(θ1 | y), or

4. for any set A, using

number of θi,1 ’s in A
M
as an estimate of P[θ1 ∈ A | y].

That’s the idea behind Monte Carlo integration.

The term “Markov chain” refers to how the samples ~θ1 , . . . , ~θM are produced. In a
Markov chain there is a transition density or transition kernel k(~θi | ~θi−1 ) which is a density
for generating ~θi given ~θi−1 . We first choose ~θ1 almost arbitrarily, then generate (~θ2 | ~θ1 ),
(~θ3 | ~θ2 ), and so on, in succession, for as many steps as we like. Each ~θi has a density
pi ≡ p(~θi ) which depends on ~θ1 and the transition kernel. But,

1. under some fairly benign conditions (See the references at the beginning of the chap-
ter for details.) the sequence p1 , p2 , . . . converges to a limit p, the stationary distri-
bution, that does not depend on ~θ1 ;
6.2. METROPOLIS, METROPOLIS-HASTINGS, AND GIBBS 359

2. the transition density k(~θi | ~θi−1 ) can be chosen so that the stationary distribution p is
equal to p(~θ | y);

3. we can find an m such that i > m ⇒ pi ≈ p = p(~θ | y);

4. then ~θm+1 , . . . , ~θM are, approximately, a sample from p(~θ | y).

The Metropolis-Hastings algorithm [Metropolis et al., 1953, Hastings, 1970] is one

way to construct an MCMC algorithm whose stationary distribution is p(~θ | y). It works
according to the following steps.

1. Choose a proposal density g(~θ∗ | ~θ).

2. Choose ~θ1 .

3. For i = 2, 3, . . .

• Generate a proposal ~θ∗ from g(~θ∗ | ~θi−1 ).

• Set
~ ~ ~
 
 p(θ ∗
| y)g( θi−1 | θ ∗ 
)
.
 
r ≡ min 1, (6.9)
 
 p(~θi−1 | y)g(~θ∗ | ~θi−1 ) 
 


• Set
~θ∗

~θi = 
 with probability r,
~θi−1

 with probability 1 − r.

Step 3 define the transition kernel k. In many MCMC chains, the acceptance probability r
may be strictly less than one, so the kernel k is a mixture of two parts: one that generates
a new value of ~θi+1 , ~θi and one that sets ~θi+1 = ~θi .
To illustrate MCMC, suppose we want to generate a sample θ1 , . . . , θ10,000 from the
Be(5, 2) distribution. We arbitrarily choose a proposal density g(θ∗ | θ) = U(θ − .1, θ + .1)
and arbitrarily choose θ1 = 0.5. The following R code draws the sample.

samp <- rep ( NA, 10000 )

samp[1] <- 0.5
for ( i in 2:10000 ) {
prev <- samp[i-1]
thetastar <- runif ( 1, prev - .1, prev + .1 )
r <- min ( 1, dbeta(thetastar,5,2) / dbeta(prev,5,2) )
if ( rbinom ( 1, 1, r ) == 1 )
new <- thetastar
6.2. METROPOLIS, METROPOLIS-HASTINGS, AND GIBBS 360

else
new <- prev
samp[i] <- new
}
The top panel of Figure 6.5 shows the result. The solid curve is the Be(5, 2) density and the
histogram is made from the Metropolis-Hastings samples. They match closely, showing
that the algorithm performed well.

Figure 6.5 was produced by

par ( mfrow=c(3,1) )
hist ( samp[-(1:1000)], prob=TRUE, xlab=expression(theta),
ylab="", main="" )
x <- seq(0,1,length=100)
lines ( x, dbeta(x,5,2) )
plot ( samp, pch=".", ylab=expression(theta) )
plot ( dbeta(samp,5,2), pch=".", ylab=expression(p(theta)) )

The code samp[-(1:1000)] discards the first 1000 draws in the hope that the sampler
will have converged to its stationary distribution after 1000 iterations.

Assuming that convergence conditions have been met and that the algorithm is well-
constructed, MCMC chains are guaranteed eventually to converge and deliver samples
from the desired distribution. But the guarantee is asymptotic and in practice the output
from the chain should be checked to diagnose potential problems that might arise in finite
samples.
The main thing to check is mixing. An MCMC algorithm operates in the space of ~θ.
At each iteration of the chain, i.e., for each value of i, there is a current location ~θi . At the
next iteration the chain moves to a new location ~θi . In this way the chain explores the ~θ
space. While it is exploring it also evaluates p(~θi ). In theory, the chain should spend many
iterations at values of ~θ where p(~θ) is large — and hence deliver many samples of ~θ’s with
large posterior density — and few iterations at values where p(~θ) is small. For the chain to
do its job it must find the mode or modes of p(~θ), it must move around in their vicinity, and
it must move between them. The process of moving from one part of the space to another
is called mixing.
The middle and bottom panels of Figure 6.5 illustrate mixing. The middle panel plots
θi vs. i. It shows that the chain spends most of its iterations in values of θ between about 0.6
6.2. METROPOLIS, METROPOLIS-HASTINGS, AND GIBBS 361

2.0
1.0
0.0

0.2 0.4 0.6 0.8 1.0

θ
1.0
0.6
θ

0.2

0 2000 4000 6000 8000 10000

Index
2.0
p (θ )

1.0
0.0

0 2000 4000 6000 8000 10000

Index

Figure 6.5: 10,000 MCMC samples of the Be(5, 2) density. Top panel: histogram of
samples from the Metropolis-Hastings algorithm and the Be(5, 2) density. Middle panel:
θi plotted against i. Bottom panel: p(θi ) plotted against i.
6.2. METROPOLIS, METROPOLIS-HASTINGS, AND GIBBS 362

and 0.9 but makes occasional excursions down to 0.4 or 0.2 or so. After each excursion
it comes back to the mode around 0.8. The chain has taken many excursions, so it has
explored the space well. The bottom panel plots p(θi ) vs. i. It shows that the chain spent
most of its time near the mode where p(θ) ≈ 2.4 but made multiple excursions down to
places where p(θ) is around 0.5, or even less. This chain mixed well.
To illustrate poor mixing we’ll use the same MCMC algorithm but with different pro-
posal kernels. First we’ll use (θ∗ | θ) = U(θ − 100, θ + 100) and change the corresponding
line of code to
thetastar <- runif ( 1, prev - 100, prev + 100 ). Then we’ll use (θ∗ | θ) =
U(θ − .00001, θ + .00001) and change the corresponding line of code to
thetastar <- runif ( 1, prev - .00001, prev + .00001 ). Figure 6.6 shows
the result. The left-hand side of the figure is for (θ∗ | θ) = U(θ − 100, θ + 100). The top
panel shows a very much rougher histogram than Figure 6.5; the middle and bottom pan-
els show why. The proposal radius is so large that most proposals are rejected; therefore,
θi+1 = θi for many iterations; therefore we get the flat spots in the middle and bottom pan-
els. The plots reveal that the sampler explored fewer than 30 separate values of θ. That’s
too few; the sampler has not mixed well. In contrast, the right-hand side of the figure —
for (θ∗ | θ) = U(θ − .00001, θ + .00001) — shows that θ has drifted steadily downward, but
over a very small range. There are no flat spots, so the sampler is accepting most propos-
als, but the proposal radius is so small that the sampler hasn’t yet explored most of the
space. It too has not mixed well.
Plots such as the middle and bottom plots of Figure 6.6 are called trace plots because
they trace the path of the sampler.
In this problem, good mixing depends on getting the proposal radius not too large and
not too small, but just right (Hassall [1909]). To be sure, if we run the MCMC chain
long enough, all three samplers would yield good samples from Be(5, 2). But the first
sampler mixed well with only 10,000 iterations while the others would require many more
iterations to yield a good sample. In practice, one must examine the output of one’s MCMC
chain to diagnose mixing problems. No diagnostics are fool proof, but not diagnosing is
foolhardy.
Several special cases of the Metropolis-Hastings algorithm deserve separate mention.

Metropolis algorithm It is often convenient to choose the proposal density g(~θ∗ | ~θ) to
be symmetric; i.e., so that g(~θ∗ | ~θ) = g(~θ | ~θ∗ ). In this case the Metropolis ratio
p(~θ∗ | y)g(~θi−1 | ~θ∗ )/p(~θi−1 | y)g(~θ∗ | ~θi−1 ) simplifies to p(~θ∗ | y)/p(~θi−1 | y). That’s what
happened in the Be(5, 2) illustration and why the line
r <- min ( 1, dbeta(thetastar,5,2) / dbeta(prev,5,2) ) doesn’t involve
g.
6.2. METROPOLIS, METROPOLIS-HASTINGS, AND GIBBS 363

4
3

500 1000
2
1
0

0
0.2 0.4 0.6 0.8 0.4990 0.4994 0.4998 0.5002

θ θ
0.9

0.4998
0.7
θ

θ
0.5

0.4992
0.3

0 2000 6000 10000 0 2000 6000 10000

Index Index
0.938
2.0
p (θ )

p (θ )

0.935
1.0

0.932
0.0

0 2000 6000 10000 0 2000 6000 10000

Index Index

Figure 6.6: 10,000 MCMC samples of the Be(5, 2) density. Left column: (θ∗ | θ) = U(θ −
100, θ + 100); Right column: (θ∗ | θ) = U(θ − .00001, θ + .00001). Top: histogram of
samples from the Metropolis-Hastings algorithm and the Be(5, 2) density. Middle: θi
plotted against i. Bottom: p(θi ) plotted against i.
6.2. METROPOLIS, METROPOLIS-HASTINGS, AND GIBBS 364

Independence sampler It may be convenient to choose g(~θ∗ | ~θ) = g(~θ∗ ) not dependent
on ~θ. For example, we could have used thetastar <- runif(1) in the Be(5, 2)
illustration.

Multiple transition kernels We may construct multiple transition kernels, say g1 , . . . , gm .

Then for each iteration of the MCMC chain we can randomly choose j ∈ {1, . . . , m}
and make a proposal according to g j . We would do this either for convenience or to
improve the convergence rate and mixing properties of the chain.

Gibbs sampler [Geman and Geman, 1984] In many practical examples, the so-called full
conditionals or complete conditionals p(θ j | ~θ(− j) , y) are known and easy to sample
for all j, where θ(− j) = (θ1 , . . . , θ j−1 , θ j+1 , . . . , θk ). In this case we may sample θi, j
from
p(θ j | θi,1 , . . . , θi, j−1 , θi−1, j+1 , . . . , θi−1,k ) for j = 1, . . . , k and set ~θi = (θi,1 , . . . , θi,k ). We
would do this for convenience.

The next example illustrates several MCMC algorithms on the pine cone data of Ex-
ample 6.2.
Example 6.3 (Pine Cones, cont)
In this example we try several MCMC algorithms to evaluate and display the posterior
distribution in Equation 6.8. Throughout this example, we shall, for compactness, refer
to the posterior density as p(~θ) instead of p(~θ | y1 , . . . , yn ).
First we need functions to return the prior density and the likelihood function.

dprior <- function ( params, log=FALSE ) {

logprior <- ( dunif ( params["b0"], -100, 100, log=TRUE )
+ dunif ( params["b1"], -100, 100, log=TRUE )
+ dunif ( params["b2"], -100, 100, log=TRUE )
+ dunif ( params["g0"], -100, 100, log=TRUE )
+ dunif ( params["g1"], -100, 100, log=TRUE )
+ dunif ( params["g2"], -100, 100, log=TRUE )
)
if (log) return (logprior)
else return (exp(logprior))
}

lik <- function ( params, n.cones=cones$X2000, dbh=cones$dbh,

trt=cones$trt, log=FALSE ) {
zero <- n.cones == 0
6.2. METROPOLIS, METROPOLIS-HASTINGS, AND GIBBS 365

tmp1 <- params["b0"] + params["b1"] * dbh + params["b2"] * trt

tmp2 <- params["g0"] + params["g1"] * dbh + params["g2"] * trt
etmp1 <- exp(tmp1)
etmp2 <- exp(tmp2)

loglik <- ( sum ( tmp1[!zero] )

- sum ( etmp2[!zero] )
+ sum ( n.cones[!zero] * tmp2[!zero] )
+ sum ( log ( 1 + etmp1[zero] * exp ( -etmp2[zero] ) ) )
- sum ( log ( 1 + etmp1 ) )
)
if (log) return (loglik)
else return (exp(loglik))
}

Now we write a proposal function. This ones makes (~θ∗ | ~θ) ∼ N(~θ, .1I6 ), where I6 is the
6 × 6 identity matrix.
g.all <- function ( params ) {
sig <- c(.1,.1,.1,.1,.1,.1)
proposed <- mvrnorm ( 1, mu=params, Sigma=diag(sig) )
return ( list ( proposed=proposed, ratio=1 ) )
}

Finally we write the main part of the code. Try to understand it; you may have to write
something similar. Notice an interesting feature of R: assigning names to the compo-
nents of params allows us to refer to the components by name in the lik function.

# initial values
params <- c ( "b0"=0, "b1"=0, "b2"=0, "g0"=0, "g1"=0, "g2"=0 )

# number of iterations
mc <- 10000

# storage for output

mcmc.out <- matrix ( NA, mc, length(params)+1 )
6.2. METROPOLIS, METROPOLIS-HASTINGS, AND GIBBS 366

# the main loop

for ( i in 1:mc ) {
prop <- g.all ( params )
new <- prop$proposed
log.accept.ratio <- ( dprior ( new, log=TRUE )
- dprior ( params, log=TRUE )
+ lik ( new, log=TRUE )
- lik ( params, log=TRUE )
- log ( prop$ratio )
)
accept.ratio <- min ( 1, exp(log.accept.ratio) )

if ( as.logical ( rbinom(1,1,accept.ratio) ) )
params <- new

mcmc.out[i,] <- c ( params, lik ( params, log=TRUE ) )

}

Figure 6.7 shows trace plots of the output. The plots show that the sampler did not
move very often; it did not mix well and did not explore the space effectively.

Figure 6.7 was produced by the following snippet.

par ( mfrow=c(4,2), mar=c(4,4,1,1)+.1 )

for ( i in 1:6 )
plot ( mcmc.out[,i], ylab=names(params)[i], pch="." )
plot ( mcmc.out[,7], ylab=expression(p(theta)), pch="." )

When samplers get stuck, sometimes it’s because the proposal radius is too large.
So next we try a smaller radius: sig <- rep(.01,6). Figure 6.8 shows the result.
The sampler is still not mixing well. The parameter β0 travelled from its starting point of
β0 = 0 to about β0 ≈ −1.4 or so, then seemed to get stuck; other parameters behaved
similarly. Let’s try running the chain for more iterations: mc <- 100000. Figure 6.9
shows the result. Again, the sampler does not appear to have mixed well. Parameters
β0 and β1 , for example, have not yet settled into any sort of steady-state behavior and
6.2. METROPOLIS, METROPOLIS-HASTINGS, AND GIBBS 367

0.1
−0.2 0.2

0.0
b0

b1

−0.2
−0.8

0 2000 6000 10000 0 2000 6000 10000

Index Index

0.0
−0.2

−0.5
b2

g0
−0.6

−1.5
−1.0

0 2000 6000 10000 0 2000 6000 10000

Index Index
0.2 0.4
0.10
g1

g2

−0.2
0.00

0 2000 6000 10000 0 2000 6000 10000

Index Index
1500
p(θ)

500
−500

0 2000 6000 10000

Index

Figure 6.7: Trace plots of MCMC output from the pine cone code on page 365.
6.2. METROPOLIS, METROPOLIS-HASTINGS, AND GIBBS 368

0.10
−0.2

0.00
−0.8
b0

b1

−0.15
−1.4

0 2000 6000 10000 0 2000 6000 10000

Index Index

0.2
0.1
−0.1

−0.4
b2

g0
−0.3

0 2000 6000 10000 −1.0 0 2000 6000 10000

Index Index
0.30
0.14

0.15
0.10
g1

g2
0.06

0.00

0 2000 6000 10000 0 2000 6000 10000

Index Index
1400
p(θ)

1000

0 2000 6000 10000

Index

Figure 6.8: Trace plots of MCMC output from the pine cone code with a smaller proposal
radius.
6.2. METROPOLIS, METROPOLIS-HASTINGS, AND GIBBS 369

p(~θ) seems to be steadily increasing, indicating that the sampler may not yet have
found the posterior mode.

It is not always necessary to plot every iteration of an MCMC sampler. Figure 6.9
plots every 10’th iteration; plots of every iteration look similar. The figure was produced
by the following snippet.

par ( mfrow=c(4,2), mar=c(4,4,1,1)+.1 )

plotem <- seq ( 1, 100000, by=10 )
for ( i in 1:6 )
plot ( mcmc.out[plotem,i], ylab=names(params)[i], pch="." )
plot ( mcmc.out[plotem,7], ylab=expression(p(theta)), pch="." )

The sampler isn’t mixing well. To write a better one we should try to understand
why this one is failing. It could be that proposing a change in all parameters simulta-
neously is too dramatic, that once the sampler reaches a location where p(~θ) is large,
changing all the parameters at once is likely to result in a location where p(~θ) is small,
therefore the acceptance ratio will be small, and the proposal will likely be rejected. To
ameliorate the problem we’ll try proposing a change to only one parameter at a time.
The new proposal function is

g.one <- function ( params ) {

sig <- c ( "b0"=.1, "b1"=.1, "b2"=.1, "g0"=.1, "g1"=.1, "g2"=.1 )
which <- sample ( names(params), 1 )
proposed <- params
proposed[which] <- rnorm ( 1, mean=params[which], sd=sig[which] )
return ( list ( proposed=proposed, ratio=1 ) )
}

which randomly chooses one of the six parameters and proposes to update that pa-
rameter only. Naturally, we edit the main loop to use g.one instead of g.all. Fig-
ure 6.10 shows the result. This is starting to look better. Parameters β2 and γ2 are ex-
hibiting steady-state behavior; so are β0 and β1 , after iteration 10,000 or so ( x = 1000
in the plots). Still, γ0 and γ1 do not look like they have converged.
Figure 6.11 illuminates some of the problems. In particular, β0 and β1 seem to be
linearly related, as do γ0 and γ1 . This is often the case in regression problems; and we
6.2. METROPOLIS, METROPOLIS-HASTINGS, AND GIBBS 370

−2.0

0.12
b0

b1
−3.0

0.06
−4.0

0 2000 6000 10000 0 2000 6000 10000

Index Index
0.7

−1.2
0.5

−1.6
b2

g0
0.3
0.1

0 2000 6000 10000 −2.0 0 2000 6000 10000

Index Index
0.19

0.45
0.17
g1

g2

0.35
0.25
0.15

0 2000 6000 10000 0 2000 6000 10000

Index Index
1770
1755
p(θ)

1740

0 2000 6000 10000

Index

Figure 6.9: Trace plots of MCMC output from the pine cone code with a smaller proposal
radius and 100,000 iterations. The plots show every 10’th iteration.
6.2. METROPOLIS, METROPOLIS-HASTINGS, AND GIBBS 371

0.25
−1
−3
b0

b1

0.10
−5

−0.05
−7

0 2000 6000 10000 0 2000 6000 10000

Index Index

−0.5 0.0
1.2
0.8
b2

g0
0.4
0.0

0 2000 6000 10000 −1.5 0 2000 6000 10000

Index Index
0.4
0.10
g1

g2

0.2
0.00

0.0

0 2000 6000 10000 0 2000 6000 10000

Index Index
1500
p(θ)

500
−500

0 2000 6000 10000

Index

Figure 6.10: Trace plots of MCMC output from the pine cone code with proposal function
g.one and 100,000 iterations. The plots show every 10’th iteration.
6.2. METROPOLIS, METROPOLIS-HASTINGS, AND GIBBS 372

have seen it before for the pine cones in Figure 3.15. In the current setting it means
that p(~θ | y1 , . . . , yn ) has ridges: one along a line in the (β0 , β1 ) plane and another along
a line in the (γ0 , γ1 ) plane.

Figure 6.11 was produced by the following snippet.

plotem <- seq ( 10000, 100000, by=10 )

pairs ( mcmc.out[plotem,], pch=".",
labels=c(names(params),"density") )

As Figure 6.10 shows, it took the first 10,000 iterations or so for β0 and β1 to reach a
roughly steady state and for p(~θ) to climb to a reasonably large value. If those iterations
were included in Figure 6.11, the points after iteration 10,000 would be squashed
together in a small region. Therefore we made plotem <- seq ( 10,000, 100000,
by=10 ) to drop the first 9999 iterations from the plots.

If our MCMC algorithm proposes a move along the ridge, the proposal is likely to be
accepted. But if the algorithm proposes a move that takes us off the ridge, the proposal
is likely to be rejected because p would be small and therefore the acceptance ratio
would be small. But that’s not happening here: our MCMC algorithm seems not to be
stuck, so we surmise that it is proposing moves that are small compared to the widths
of the ridges. However, because the proposals are small, the chain does not explore
the space quickly. That’s why γ0 and γ1 appear not to have reached a steady state.
We could improve the algorithm by proposing moves that are roughly parallel to the
ridges. And we can do that by making multivariate Normal proposals with a covariance
matrix that approximates the posterior covariance of the parameters. We’ll do that by
finding the covariance of the samples we’ve generated and using it as the covariance
matrix of our proposal distribution. The R code is

Sig <- cov ( mcmc.out[10000:100000,-7])

g.group <- function ( params ) {
proposed <- mvrnorm ( 1, mu=params, Sigma=Sig )
return ( list ( proposed=proposed, ratio=1 ) )
}

We drop the first 9999 iterations because they seem not to reflect p(~θ) accurately. Then
we calculate the covariance matrix of the samples from the previous MCMC sampler.
6.2. METROPOLIS, METROPOLIS-HASTINGS, AND GIBBS 373

0.16 0.30 −1.6 −0.4 0.20 0.55

−6.5 −5.0
b0
0.24

b1
0.16

1.2
0.6
b2

0.0
−0.4
−1.0

g0
−1.6

0.15
g1

0.12
0.40

g2
0.20

1768

density
1760

−6.5 0.0 1.2 0.12 0.17 1760

Figure 6.11: Pairs plots of MCMC output from the pine cones example.
6.2. METROPOLIS, METROPOLIS-HASTINGS, AND GIBBS 374

That covariance matrix is used in the proposal function. The results are shown in
Figures 6.12 and 6.13. Figure 6.12 shows that the sampler seems to have converged
after the first several thousand iterations. The posterior density has risen to a high
level and is hovering there; all six variables appear to be mixing well. Figure 6.13
confirms our earlier impression that the posterior density seems to be approxiately
Normal — at least, it has Normal-looking two dimensional marginals — with β0 and
β1 highly correlated with each other, γ1 and γ2 highly correlated with each other, and
no other large correlations. The sampler seems to have found one mode and to be
exploring it well.

Figures 6.12 and 6.13 were produced with the following snippet.

plotem <- seq ( 1, 100000, by=10 )

par ( mfrow=c(4,2), mar=c(4,4,1,1)+.1 )
for ( i in 1:6 )
plot ( mcmc.out[plotem,i], ylab=names(params)[i], pch="." )
plot ( mcmc.out[plotem,7], ylab=expression(p(theta)), pch="." )

plotem <- seq ( 1000, 100000, by=10 )

pairs ( mcmc.out[plotem,], pch=".",
labels=c(names(params),"density") )

Now that we have a good set of samples from the posterior, we can use it to
answer substantive questions. For instance, we might want to know whether the extra
atmospheric CO2 has allowed pine trees to reach sexual maturity at an earlier age
or to produce more pine cones. This is a question of whether β2 and γ2 are positive,
negative, or approximately zero. Figure 6.14 shows the answer by plotting the posterior
densities of β2 and γ2 . Both densities put almost all their mass on positive values,
indicating that P[β2 > 0] and P[γ2 > 0] are both very large, and therefore that pines
trees with excess CO2 mature earlier and produce more cones than pine trees grown
under normal conditions.

Figure 6.14 was produced by the following snippet.

par ( mfrow=c(1,2) )
plot ( density ( mcmc.out[10000:100000,"b2"] ),
6.2. METROPOLIS, METROPOLIS-HASTINGS, AND GIBBS 375

0.4
−2
−4
b0

b1

0.2
−6
−8

0.0
0 2000 6000 10000 0 2000 6000 10000

Index Index
1.0

0.0
b2

g0

−1.5
0.0

−3.0
−1.0

0 2000 6000 10000 0 2000 6000 10000

Index Index
0.6
0.20

0.4
0.10
g1

g2

0.2
0.0
0.00

0 2000 6000 10000 0 2000 6000 10000

Index Index
1500
p(θ)

500
−500

0 2000 6000 10000

Index

Figure 6.12: Trace plots of MCMC output from the pine cone code with proposal function
g.group and 100,000 iterations. The plots show every 10’th iteration.
6.2. METROPOLIS, METROPOLIS-HASTINGS, AND GIBBS 376

0.15 0.35 −2.0 0.2 0.6

−5
b0

−7
0.15 0.25 0.35

b1

1.0
b2

0.0
−1.0

g0
−2.0

0.16
g1

0.12
0.6
0.4

g2
0.2

1770

density
1760

−7 −4 0.0 0.12 0.19 1760

Figure 6.13: Pairs plots of MCMC output from the pine cones example with proposal
g.group.
6.3. EXERCISES 377

2.0

6
p(β2)

p(γ2)
1.0

4
2
0.0

0
0.0 1.0 0.2 0.4 0.6

β2 γ2

Figure 6.14: Posterior density of β2 and γ2 from Example 6.3.

xlab=expression(beta[2]),
ylab=expression(p(beta[2])), main="" )
plot ( density ( mcmc.out[10000:100000,"g2"] ),
xlab=expression(gamma[2]),
ylab=expression(p(gamma[2])), main="" )

6.3 Exercises
1. This exercise follows from Example 6.1.
(a) Find the posterior density for C50 , the expected amount of ice cream consumed
when the temperature is 50 degrees, by writing C50 as a linear function of
(β0 , β1 ) and using the posterior from Example 6.1.
(b) Find the posterior density for C50 by reparameterizing. Instead of working with
parameters (β0 , β1 ), work with parameters (C50 , β1 ). Write the equation for Yi
as a linear function of (C50 , β1 ) and find the new X matrix. Use that new matrix
and a convenient prior to calculate the posterior density of (C50 , β1 ).
6.3. EXERCISES 378

(c) Do parts (a) and (b) agree about the posterior distribution of C50 . Does part
(b) agree with Example 6.1 about the posterior distribution of β1 ? Should they
agree?

2. This exercise asks you to enhance the code for the Be(5, 2) example on page 359.

(a) How many samples is enough? Instead of 10,000, try different numbers. How
few samples can you get away with and still have an adequate approximation
to the Be(5, 2) distribution? You must decide what "adequate" means; you can
use either a firm or fuzzy definition. Illustrate your results with figures similar
to 6.5.
(b) Try an independence sampler in the Be(5, 2) example on page 359. Replace
the proposal kernel with θ∗ ∼ U(0, 1). Run the sampler, make a figure similar
to Figure 6.5 and describe the result.
(c) Does the proposal distribution matter? Instead of proposing with a radius of
0.1, try different numbers. How much does the proposal radius matter? Does
the proposal radius change your answer to part 2a? Illustrate your results with
figures similar to 6.5.
(d) Try a non-symmetric proposal. For example, you might try a proposal distri-
bution of Be(5, 1), or a distribution that puts 2/3 of its mass on (xi−1 − .1, xi−1 )
and 1/3 of its mass on (xi−1 , xi−1 +.1). Illustrate your results with figures similar
to 6.5.
(e) What would happen if your proposal distribution were Be(5, 2)? How would
the algorithm simplify?

3. (a) Some researchers are interested in θ, the proportion of students who ever cheat
on college exams. They randomly sample 100 students and ask “Have you
ever cheated on a college exam?” Naturally, some students lie. Let φ1 be the
proportion of non-cheaters who lie and φ2 be the proportion of cheaters who
lie. Let X be the number of students who answer “yes” and suppose X = 40.
i. Create a prior distribution for θ, φ1 , and φ2 . Use your knowledge guided
by experience. Write a formula for your prior and plot the marginal prior
density of each parameter.
ii. Write a formula for the likelihood function `(θ, φ1 , φ2 ).
iii. Find the m.l.e..
iv. Write a formula for the joint posterior density p(θ, φ1 , φ2 | X = 40).
v. Write a formula for the marginal posterior density p(θ | X = 40).
6.3. EXERCISES 379

vi. Write an MCMC sampler to sample from the joint posterior.

vii. Use the sampler to find p(θ | X = 40). Summarize your results. Include
information on how you assessed mixing and on what you learned about
p(θ | X = 40).
viii. Assess the sensitivity of your posterior, p(θ | X = 40), to your prior for φ1
and φ2 .
(b) Randomized response This part of the exercise uses ideas from Exercises 42
in Chapter 1 and 6 in Chapter 2. As explained there, researchers will sometimes
instruct subjects as follows.
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.
(a) Does your telephone number end in an even digit?
(b) Have you ever cheated on an exam in college?
The idea of the randomization is, of course, to reduce the incentive to lie.
Nonetheless, students may still lie.
i. If about 40 students answered ‘yes’ in part (a), about how many do you
think will answer ‘yes’ under the conditions of part (b)?
ii. Repeat part (a) under the conditions of part (b) and with your best guess
about what X will be under these conditions.
iii. Assess whether researchers who are interested in θ are better off using the
conditions of part (a) or part (b).

4. The EPA conducts occasional reviews of its standards for airborne asbestos. During
a review, the EPA examines data from several studies. Different studies keep track
of different groups of people; different groups have different exposures to asbestos.
Let ni be the number of people in the i’th study, let xi be their asbestos exposure,
and let yi be the number who develop lung cancer. The EPA’s model is yi ∼ Poi(λi )
where λi = ni xi λ and where λ is the typical rate at which asbestos causes cancer.
The ni ’s and xi ’s are known constants; the yi ’s are random variables. The EPA wants
a posterior distribution for λ.

(a) Write down the likelihood function for λ.

P
(b) Show that the likelihood function depends only on yi , not on any other aspect
of the yi ’s.
(c) What family of distributions would be convenient prior distributions for λ?
6.3. EXERCISES 380

(d) If the EPA used the prior λ ∼ Gam(a, b), what would be the EPA’s posterior?
5. Figures 6.12 and 6.13 suggest that the MCMC sampler has found one mode of the
posterior density. Might there be others? Use the lik function and R’s optim func-
tion to find out. Either design or randomly generate some starting values (You must
decide on good choices for either the design or the randomization.) and use optim
to find a mode of the likelihood function. Summarize and report your results.
6. Example 6.3 shows that β2 and γ2 are very likely positive, and therefore that pine
trees with extra CO2 mature earlier and produce more cones. But how much earlier
and how many more?
(a) Find the posterior means E[β2 | y1 , . . . , yn ] and E[γ2 | y1 , . . . , yn ] approximately,
from the Figures in the text.
(b) Suppose there are three trees in the control plots that have probabilities 0.1, 0.5,
and 0.9 of being sexually mature. Plugging in E[β2 | y1 , . . . , yn ] from the pre-
vious question, estimate their probabilities of being mature if they had grown
with excess CO2 .
(c) Is the plug-in estimate from the previous question correct? I.e., does it correctly
calculate the probability that those trees would be sexually mature? Explain
why or why not. If it’s not correct, explain how to calculate the probabilities
correctly.
7. In the context of Equation 6.7 in Example 6.3 we might want to investigate whether
the coefficient of dbh should be the same for control trees and for treated trees.
(a) Write down a model enhancing that on page 353 to allow for the possibility of
different coefficients for different treatments.
(b) What parts of the R code have to change?
(c) Write the new code.
(d) Run it.
(e) Summarize and report results. Report any difficulties with modifying and run-
ning the code. Say how many iterations you ran and how you checked mixing.
Also report conclusions: does it look like different treatments need different
coefficients? How can you tell?
8. You go to the bus stop to catch a bus. You know that buses arrive every 15 minutes,
but you don’t know when the next is due. Let T be the time elapsed, in hours, since
the previous bus. Adopt the prior distribution T ∼ Unif(0, 1/4).
6.3. EXERCISES 381

(a) Find E[T ].

Passengers, apart from yourself, arrive at the bus stop according to a Poisson process
with rate λ = 2 people per hour; i.e., in any interval of length `, the number of
arrivals has a Poisson distribution with parameter 2` and, if two intervals are disjoint,
then their numbers of arrivals are independent. Let X be the number of passengers,
other than yourself, waiting at the bus stop when you arrive.

(b) Suppose X = 0. Write an intuitive argument for whether that should increase
or decrease your expected value for T . I.e., is E[T |X = 0] greater than, less
than, or the same as E[T ]?

(c) Find the density of T given X = 0, up to a constant of proportionality. It is a

truncated version of a familiar density. What is the familiar density?

(d) Suppose X = 1. Write an intuitive argument for whether that should increase
or decrease your expected value for T . I.e., is E[T |X = 1] greater than, less
than, or the same as E[T ]?

(e) Find the density of T given X = 1, up to a constant of proportionality. It is a

truncated version of a familiar density. What is the familiar density?

(f) Find the density of T given X = x, up to a constant of proportionality. It is a

truncated version of a familiar density. What is the familiar density?
 Chapter 7

More Models

This chapter takes up a wide variety of statistical models. It is beyond the scope of this
book to give a full treatment of any one of them. But we hope to introduce each model
enough so the reader can see in what situations it might be useful, what it’s primary char-
acteristcs are, and how a simple analysis might be carried out in R. A more thorough
treatment of many of these models can be found in Venables and Ripley [2002].
Specialized models call for specialized methods. Because not all methods are built
in to R, many people have contributed packages of specialized methods. Their packages
can be downloaded from http://probability.ca/cran/. To use a package, you must
first install it. Then, in each session of R in which you want to use it, you must load
it. For example, to use the survival package for survival analysis you must first type
install.packages("survival"). R should respond either by installing the package or
by asking you to choose a mirror, then installing the package. Next, whenever you want
to use the package, you must type library("survival").

7.1 Random Effects

To explain the idea of random effects it is best to look at an example. The nlme package
comes with a dataset Orthodont. As the help file explains,

“Investigators at the University of North Carolina Dental School followed

the growth of 27 children (16 males, 11 females) from age 8 until age 14. Ev-
ery two years they measured the distance between the pituitary and the ptery-
gomaxillary fissure, two points that are easily identified on x-ray exposures of
the side of the head. . . .
“[The] data frame contains the following columns:

382
7.1. RANDOM EFFECTS 383

distance a numeric vector of distances from the pituitary to the pterygomax-

illary fissure (mm). These distances are measured on x-ray images of the
skull.
age a numeric vector of ages of the subject (yr).
Subject an ordered factor indicating the subject on which the measurement
was made. The levels are labelled M01 to M16 for the males and F01
to F11 for the females. The ordering is by increasing average distance
within sex.
Sex a factor with levels Male and Female”

Each child was measured four times, so there are 108 lines in the data set. The first several
lines of data are

Grouped Data: distance ~ age | Subject

distance age Subject Sex
1 26.0 8 M01 Male
2 25.0 10 M01 Male
3 29.0 12 M01 Male
4 31.0 14 M01 Male
5 21.5 8 M02 Male
6 22.5 10 M02 Male

Figure 7.1 shows the data and was produced by xyplot ( distance ˜ age | Sex,
data = Orthodont, groups = Subject, type="o", col=1 ). The function xyplot
is part of the lattice package. As illustrated in the figure, it produces plots of Y versus
X in which points can be grouped by one variable and separated into different panels by
another variable.
The figure shows that for most subjects, distance is an approximately linear function
of age, that males tend to be larger than females, and that different subjects have different
intercepts. Therefore, to fit this data we need a model with the following features: (1) a
parameter for the average intercept of males (or females); (2) a parameter for the average
difference in intercepts for males and females; (3) one parameter per subject for how that
subject’s intercept differs from its group average; and (4) one parameter for the slope.
Model 7.1 has those features.

Yi,t = distance for subject i at time t

Yi,t ∼ N(µi,t , σ) (7.1)
µi,t = β0 + β1 Sexi + δi + β2 t
7.1. RANDOM EFFECTS 384

8 9 10 11 12 13 14

Male Female
● ●
●
30 ●
●
●
● ●
● ● ●
● ● ●
distance
●
● ●
● ●
●
● ● ●
● ●
● ●
●
25 ●
●
●
● ●
●
● ● ●
●
●
●
●
●
●
●
● ●
● ●
● ●
● ● ●
● ●
●
●
● ●
● ● ● ●
● ●
● ● ●
● ●
● ●
● ●
● ●
20 ● ● ●
● ● ●

● ●

8 9 10 11 12 13 14

age

Figure 7.1: Plots of the Orthodont data: distance as a function of age, grouped by Subject,
separated by Sex.

Here is the key feature: we want β0 to represent the average intercept of all males and
β0 + β1 to represent the average intercept of all females, not just those in this study; and
we want to think of the δi ’s in this study as random draws from a population of δ’s rep-
resenting all children. The terminology is that β0 , β1 , and β2 are fixed effects because
they describe the general population, not the particular individuals in the sample, while
δ1 , . . . , δn are random effects because they are random draws from the general population.
It is customary to add a term to the model to describe the distribution of random effects.
In our case we shall model the δi ’s as N(0, σran. eff. ). Thus Model 7.1 becomes

Yi,t = distance for subject i at time t

Yi,t ∼ N(µi,t , σ)
(7.2)
µi,t = β0 + β1 Sexi + δi + β2 t
δ1 , . . . , δn ∼ i.i.d. N(0, σran. eff. )

Modelling the random effects as Normal is an arbitrary choice. We could have used a
different distribution, but the Normal is convenient and not obviously contraindicated by
Figure 7.1. Choosing the mean of the Normal distribution to be 0 is not arbitrary. It’s
the result of thinking of the δi ’s as draws from a larger population. In any population, the
average departure from the mean must be 0.
7.1. RANDOM EFFECTS 385

Model 7.2 is called a mixed effects model because it contains both fixed and random
effects. Often, in mixed effects models, we are interested in σran. eff. . Whether we are
interested in the individual δi ’s depends on the purpose of the investigation.
We can fit Model 7.2 by the following R code.

ortho.fit1 <- lme ( distance ˜ age + Sex, random = ˜ 1,

data = Orthodont )

• lme stands for “linear mixed effects model.” It is in the nlme package.

• The formula distance ˜ age + Sex is just like formulas for linear models.

• random = ˜ 1 specifies the random effects. In this case, the random effects are
intercepts, or coefficients of 1. If we thought that each child had his or her own slope,
then we would have said random = ˜ age - 1 to say that the random effects are
coefficients of age but not of 1.

Printing ortho.fit1 yields

Linear mixed-effects model fit by REML

Data: Orthodont
Log-restricted-likelihood: -218.7563
Fixed: distance ~ age + Sex
(Intercept) age SexFemale
17.7067130 0.6601852 -2.3210227

Random effects:
Formula: ~1 | Subject
(Intercept) Residual
StdDev: 1.807425 1.431592

Number of Observations: 108

Number of Groups: 27

The fixed effects part of the output is just like the fixed effects part of linear models. The
estimates of (β0 , β1 , β2 ) are around (17.7, −2.32, 0.66). The random effects part of the
output shows that the estimate of σ is about 1.43 while the estimate of σran. eff. is about
1.81. In other words, most of the intercepts are within about 3.6 or so of their mean and
the differences from child to child are about the same size as any unexplained variation
with each child’s data. We can see whether that’s sensible by inspecting Figure 7.1. It also
7.1. RANDOM EFFECTS 386

appears, from Figure 7.1, that one of the males doesn’t fit the general pattern. Because that
subject’s data fluctuates wildly from the pattern, it may have inflated the estimate of σ. We
might want to remove that subject’s data and refit the model, just to see how influential it
really is. That’s a topic we don’t pursue here.
We can see estimates of δ1 , . . . , δ27 by typing random.effects(ortho.fit1), which
yields

(Intercept)
M16 -1.70183357
M05 -1.70183357
M02 -1.37767479
M11 -1.16156894
M07 -1.05351602
M08 -0.94546309
M03 -0.62130432
M12 -0.62130432
M13 -0.62130432
M14 -0.08103969
M09 0.13506616
M15 0.78338371
M06 1.21559540
M04 1.43170125
M01 2.40417758
M10 3.91691853
F10 -3.58539251
F09 -1.31628108
F06 -1.31628108
F01 -1.10017523
F05 -0.01964599
F07 0.30451279
F02 0.30451279
F08 0.62867156
F03 0.95283034
F04 1.92530666
F11 3.22194176

The reasonableness of these estimates can be judged by comparison to Figure 7.2, where
we can also see that the strange male is M09.
7.1. RANDOM EFFECTS 387

8 9 11 13

F04 F11
● ●
30
●
● ● ● ● ● 25
20

F05 F07 F02 F08 F03

30
● ●
25 ● ● ●
● ● ● ● ● ● ● ●
● ● ● ● ●
● ●
20

M10 F10 F09 F06 F01

● ●
●
30
●

●
25
● ● ● ●
● ● ● ●
●● ● ● 20
distance

● ●
●
M09 M15 M06 M04 M01
● ●
30 ●
● ●
● ● ● ●●
● ● ● ●
25 ● ●
● ● ●

20 ●

M08 M03 M12 M13 M14

●
● 30
●
● ● ● ● ●
● ●
● ● ● ● ● 25
● ● ● ●
20
●

M16 M05 M02 M11 M07

30
● ● ●
25 ●
●
● ● ● ● ●
● ●
● ● ● ● ● ● ●
20 ●

8 9 11 13 8 9 11 13 8 9 11 13

age

Figure 7.2: Plots of the Orthodont data: distance as a function of age, separated by
Subject.
7.1. RANDOM EFFECTS 388

To see what is gained by fitting the mixed effects model, we compare it to two other
models that have fixed effects only.

ortho.fit2 <- lm ( distance ~ age + Sex, data = Orthodont )

ortho.fit3 <- lm ( distance ~ age+Sex+Subject, data=Orthodont )

The summary of ortho.fit2 is

...
Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 17.70671 1.11221 15.920 < 2e-16 ***
age 0.66019 0.09776 6.753 8.25e-10 ***
SexFemale -2.32102 0.44489 -5.217 9.20e-07 ***
...
Residual standard error: 2.272 on 105 degrees of freedom
...

The estimates of the coefficients are exactly the same as in ortho.fit1 But, since ortho.fit2
ignores differences between subjects, it attributes those differences to the observation SD
σ. That’s why the estimate of σ from ortho.fit1, 1.43, is smaller than the estimate from
ortho.fit2, 2.27.
The summary of ortho.fit3 is

Coefficients: (1 not defined because of singularities)

Estimate Std. Error t value Pr(>|t|)
(Intercept) 23.19192 1.85453 12.506 < 2e-16 ***
age 0.66019 0.06161 10.716 < 2e-16 ***
SexFemale -15.78472 4.22373 -3.737 0.000348 ***
Subject.L 33.91970 9.21165 3.682 0.000418 ***
Subject.Q 6.64753 2.31508 2.871 0.005228 **
Subject.C -5.80525 2.99918 -1.936 0.056448 .
Subject^4 -3.18380 2.17427 -1.464 0.147028
Subject^5 3.26565 1.39346 2.344 0.021584 *
...
Subject^25 -0.14291 0.89189 -0.160 0.873100
Subject^26 NA NA NA NA

We see that the estimates of β0 and β1 have changed. That’s because, in the fixed effects
model, those estimates are too dependent on the particular subjects in the study. The
7.1. RANDOM EFFECTS 389

estimates from ortho.fit1 are preferred, as they come from a model that describes the
data more correctly.
Finally, in Figure 7.1 there is a slight suggestion that the slope for males is slightly
larger than the slope for females. Model 7.3 allows for that possibility: β3 is the difference
in slope between males and females.

Yi,t = distance for subject i at time t

Yi,t ∼ N(µi,t , σ)
(7.3)
µi,t = β0 + β1 Sexi + δi + β2 t + β3 t × Sexi
δ1 , . . . , δn ∼ i.i.d.N(0, σran. eff. )

The model can be fit by ortho.fit4 <- lme ( distance ˜ age * Sex, random =
˜ 1, data = Orthodont ) and yields

...
Random effects:
Formula: ~1 | Subject
(Intercept) Residual
StdDev: 1.816214 1.386382

Fixed effects: distance ~ age * Sex

Value Std.Error DF t-value p-value
(Intercept) 16.340625 0.9813122 79 16.651810 0.0000
age 0.784375 0.0775011 79 10.120823 0.0000
SexFemale 1.032102 1.5374208 25 0.671321 0.5082
age:SexFemale -0.304830 0.1214209 79 -2.510520 0.0141
...

This model suggests that the slope for males is about 0.78 while the slope for females is
about 0.3 less, plus or minus about 0.25 or so. The evidence is not overwhelming in either
direction whether males and females have the same slope. But if they differ, it could be by
an amount (roughly, .30 ± 2 × .12) that is large compared to the slope for males. For its
extra complexity, this model has reduced the residual SD, or our accuracy of prediction,
from about 1.43 to about 1.39. At this point, we don’t have a strong preference for either
model and, if we were investigating further, would keep both of them in mind.
The next example comes from a study of how ants store nutrients over the winter.
Example 7.1 (Ant Fat)
The data in this example were collected to examine the strategy that ants of the
7.1. RANDOM EFFECTS 390

species Pheidole morrisi employ to store nutrients, specifically fat, over the winter;
the study was reported by Yang [2006]. Many animals need to store nutrients over
the winter when food is scarce. For most species, individual animals store their own
nutrients. But for some species, nutrients can be stored collectively. As Yang explains,
"Among ants, a common mechanism of colony fat storage is for workers of both castes
(majors and minors) to uniformly increase the amount of fat they hold . . . The goal of
this study is to better understand the specific mechanisms by which ants use divi-
sion of labor to store colony fat . . . ." The need to store fat varies with the severity of
winter, which in turn varies with latitude. So Yang studied ants at three sites, one in
Florida, one in North Carolina, and one in New York. At each site he dug up several
ant colonies in the Spring and another several colonies in the Fall. From each colony,
the ants were separated into their two castes, majors and minors, and the fat content
of each caste was measured. The first several lines of data look like this.
colony season site caste fat
1 1 Spring New York minors 15.819209
2 2 Spring New York minors 10.526316
3 3 Spring New York minors 18.534483
4 4 Spring New York minors 21.467098
5 5 Spring New York minors 9.784946
6 6 Spring New York minors 20.138289
...

There are 108 lines in all. Figure 7.3 shows the data along with kernel density esti-
mates. It was produced by densityplot ( ˜ fat | site+season, groups=caste,
data=ants, adjust=1.5 ). Purple lines are for minors; blue lines for majors. In New
York and North Carolina, it appears that minors have, on average, less fat than ma-
jors and less variability. There is also some suggestion that fat content increases with
increasing latitude. The main question is how the average percent fat differs by site,
season, and caste. Because the predictors are categorical, we need to define indicator
variables: NYi = 1 if the i’th observation was from NY; NCi = 1 if the i’th observation
was from NC; Springi = 1 if the i’th observation was from Spring; and minorsi = 1 if
the i’th observation was on the minor caste. With those conventions, we begin with
the following linear model.

yi = β0 + β1 NYi + β2 NCi + β3 Springi + β4 minorsi + β5 NYi Springi

+ β6 NCi Springi + β7 NYi minorsi + β8 NCi minorsi + β9 Springi minorsi
+ β10 NYi Springi minorsi + β11 NCi Springi minorsi + i . (7.4)
7.1. RANDOM EFFECTS 391

0 20 40 60

Spring Spring Spring

Florida New York North Carolina

0.15

0.10

0.05

●
●●●
●
●●●
● ● ●
●●●
● ●
●
●●●●●●● ●● ●
●●
●●●●●
● 0.00
Density

Fall Fall Fall

Florida New York North Carolina

0.15

0.10

0.05

0.00 ●
● ●
●
●●
●
●●
● ●●
●
●● ● ●●
●●
●●●
●
●●●●●●●●
●● ●
●●
●●
●●
●
●●●
● ●
●

0 20 40 60 0 20 40 60

fat

Figure 7.3: Percent body fat of major (blue) and minor (purple) Pheidole morrisi ants at
three sites in two seasons.
7.1. RANDOM EFFECTS 392

In Model 7.4, β0 is the average fat content of major, Florida ants in the Fall; β1 is the
difference between New York majors and Florida majors in the Fall; β2 is the difference
between North Carolina majors and Florida majors in the Fall; . . . ; β9 is the difference
between Florida majors in the Fall and Florida minors in the Spring; . . . ; etc. Model 7.4
can be fit by

fit1 <- lm ( fat ~ site * season * caste, data=ants )

The parameter estimates and their SD’s, given by summary ( fit1 ), are

(Intercept) 16.8857 2.0016

siteNew York 13.2084 2.8307
siteNorth Carolina 8.6182 2.8307
seasonSpring -6.0555 3.0024
casteminors -1.5489 2.8307
siteNew York:seasonSpring 4.2587 4.2460
siteNorth Carolina:seasonSpring -3.7789 4.2460
siteNew York:casteminors -4.2622 4.0032
siteNorth Carolina:casteminors -1.6258 4.0032
seasonSpring:casteminors 0.6997 4.2460
siteNew York:seasonSpring:casteminors -5.2305 6.0048
siteNorth Carolina:seasonSpring:casteminors -2.2812 6.0048

We see that New York majors are much fatter, by about 13.2 percentage points on
average, than Florida majors in the Fall, while North Carolina majors are in between.
That seems consistent with the theory that New York ants need to store more fat in
the Fall because they are going to face a longer, harder winter than Florida ants. In
the Spring, Florida majors have about 16.9% − 6.1% ≈ 13.8% body fat while New York
majors have about 16.9% + 13.2% − 6.1% + 4.3% ≈ 28.2% body fat. It appears that
New York majors did not lose as much fat over the winter as Florida majors. Perhaps
they didn’t need to store all that fat after all.
Before addressing that question more thoroughly, we want to examine one possible
inadequacy of the model. The data sets contains two data points from each colony —
one for majors, one for minors — and it’s possible that those two data points are not
independent. In particular, there might be colony effects; one colony might be fatter,
on average, than another from the same site and season, and that extra fatness might
apply to both castes. To see whether that’s true, we’ll examine residuals from fit1.
Specifically, we’ll plot residuals for minors on the abscissa and residuals for majors
7.1. RANDOM EFFECTS 393

residuals

10 15
●

● ●
● ● ● ●
● ●
● ●●●
●
5
majors

● ●
● ●
● ● ●●●
● ●
0

●● ●
● ● ●●●
●●
●
●● ●
● ● ●
● ●
−10

●● ●
●
●●
●

−10 −5 0 5 10 15

minors

Figure 7.4: Residuals from Model 7.4. Each point represents one colony. There is an
upward trend, indicating the possible presence of colony effects.

on the ordinate. Figure 7.4 is the plot. There is a clear, upward, approximately linear
trend, indicating that majors and minors from one colony tend to be thin or fat together.
We can capture that tendency by including random colony effects in the model. That
leads to Model 7.5:

yi = β0 + β1 NYi + β2 NCi + β3 Springi + β4 minorsi + β5 NYi Springi

+ β6 NCi Springi + β7 NYi minorsi + β8 NCi minorsi + β9 Springi minorsi
+ β10 NYi Springi minorsi + β11 NCi Springi minorsi + δcol(i) + i (7.5)

which differs from Model 7.4 by the presence of the δ’s, which are assumed to be
distributed i.i.d.N(0, σran. eff. ). Model 7.5 can be fit by

fit2 <- lme ( fat˜site*season*caste, data=ants, random = ˜1|colony )

7.1. RANDOM EFFECTS 394

The R code (...random = ˜1|colony) says to give each colony its own, random,
intercept. The summary of fit2 is, in part,

Linear mixed-effects model fit by REML

Data: ants
AIC BIC logLik
671.0577 706.9585 -321.5288

Random effects:
Formula: ~1 | colony
(Intercept) Residual
StdDev: 4.164122 4.76696

Fixed effects: fat ~ site * season * caste

Value Std.Error
(Intercept) 16.885710 2.001595
siteNew York 13.208390 2.830683
siteNorth Carolina 8.618221 2.830683
seasonSpring -6.055542 3.002392
casteminors -1.548911 2.131849
siteNew York:seasonSpring 4.258693 4.246024
siteNorth Carolina:seasonSpring -3.778862 4.246024
siteNew York:casteminors -4.262151 3.014890
siteNorth Carolina:casteminors -1.625755 3.014890
seasonSpring:casteminors 0.699721 3.197774
siteNew York:seasonSpring:casteminors -5.230472 4.522335
siteNorth Carolina:seasonSpring:casteminors -2.281162 4.522335
...
Number of Observations: 108
Number of Groups: 54

There are several things to note about fit2. First, σ̂ ≈ 4.77 is somewhat smaller than
the estimate from fit1. That’s because some of the variability that fit1 attributes to
the  ’s, fit2 attributes to the δ’s. Second, σ̂ran. eff. ≈ 4.16 is about the same size as
σ̂, indicating that colony effects explain a sizable portion of the variation in the data.
Third, estimates of the β’s are unchanged. And fourth, some of the estimates of SD’s
of the β’s are changed while others are not. Specifically, all the SD’s involving caste
effects have decreased. That’s because some of the variability that fit1 attributes to
7.1. RANDOM EFFECTS 395

Majors Minors
Fall Spring Fall Spring
Florida 16.9 10.8 15.3 10.1
North Carolina 25.0 15.7 24.0 11.0
New York 30.1 28.3 24.3 18.0

Table 7.1: Fat as a percentage of body weight in ant colonies. Three sites, two seasons,
two castes.

the  ’s, fit2 attributes to the δ’s.

With fit2 in hand, we can turn our attention to the question of interest: what are
the fat storage strategies and how do they differ from New York to Florida? Table 7.1
summarizes the analysis; the numbers come from adding the appropriate β̂’s. The
main points are

• Both castes in all locations stored more fat in the Fall than in the Spring.

• The amount of fat stored increases with increasing latitude.

• Majors store more fat than minors, in some cases by a lot, in other cases by a
little.

• Moving from Florida to North Carolina to New York, majors increase their fat
content in both Fall and Spring. But moving from Florida to North Carolina,
minors increase their fat content only in the Fall while from North Carolina to
New York, they increase only in the Spring.

The summary in Table 7.1 and the preceding bullets could have been carried out with
no formal statistical analysis. That’s the way it goes, sometimes. In this example,
we could use the statistical analysis to be more precise about how accurately we can
estimate each of the effects noted in the bullets. Here, there seems little need to do
that.
Yang [2006] carries the analysis further. For one thing, he is more formal about
statistics. In addition, he notes that majors can be further divided into repletes and non-
repletes. According to Yang, “[R]epletes carry and store a disproportionate amount of
nutrients relative to other individuals in a colony and provide it to other colony members
through trophallaxis in times of food scarcity." And according to Wikipedia, "Trophal-
laxis is the transfer of food or other fluids among members of a community through
mouth-to-mouth (stomodeal) or anus-to-mouth (proctodeal) feeding. It is most highly
7.2. TIME SERIES AND MARKOV CHAINS 396

developed in social insects such as ants, termites, wasps and bees." Yang finds in-
teresting differences in fat storage among repletes and other majors, differences that
vary according to season and location.

7.2 Time Series and Markov Chains

Figure 7.5 shows some data sets that come with R. The following descriptions are taken
from the R help pages.
Beaver The data are a small part of a study of the long-term temperature dynamics of
beaver Castor canadensis in north-central Wisconsin. Body temperature was mea-
sured by telemetry every 10 minutes for four females, but data from one period of
less than a day is shown here.

Mauna Loa Monthly atmospheric concentrations of CO2 are expressed in parts per mil-
lion (ppm) and reported in the preliminary 1997 SIO manometric mole fraction
scale.

DAX The data are the daily closing prices of Germany’s DAX stock index. The data are
sampled in business time; i.e., weekends and holidays are omitted.

UK Lung Disease The data are monthly deaths from bronchitis, emphysema and asthma
in the UK, 1974 – 1979.

Canadian Lynx The data are annual numbers of lynx trappings for 1821 – 1934 in Canada.

Presidents The data are (approximately) quarterly approval rating for the President of the
United states from the first quarter of 1945 to the last quarter of 1974.

UK drivers The data are monthly totals of car drivers in Great Britain killed or seriously
injured Jan 1969 to Dec 1984. Compulsory wearing of seat belts was introduced on
31 Jan 1983.

Sun Spots The data are monthly numbers of sunspots. They come from the World Data
Center-C1 For Sunspot Index Royal Observatory of Belgium, Av. Circulaire, 3, B-
1180 BRUSSELS http://www.oma.be/KSB-ORB/SIDC/sidc_txt.html.
What these data sets have in common is that they were all collected sequentially in time.
Such data are known as time series data. Because each data point is related to the ones
before and the ones after, they usually cannot be treated as independent random variables.
Methods for analyzing data of this type are called time series methods. More formally,
7.2. TIME SERIES AND MARKOV CHAINS 397

a time series is a sequence Y1 , . . . , YT of random variables indexed by time. The generic

element of the series is usually denoted Yt .

Figure 7.5 was produced by the following snippet.

par ( mfrow=c(4,2), mar=c(3,4,1,1)+.1 )

plot.ts ( beaver1$temp, main="Beaver", xlab="Time",
ylab="Temperature" )
plot.ts ( co2, main="Mauna Loa", ylab="CO2 (ppm)" )
plot.ts ( EuStockMarkets[,1], main="DAX",
ylab="Closing Price" )
plot.ts ( ldeaths, main="UK Lung Disease",
ylab="monthly deaths" )
plot.ts ( lynx, main="Canadian Lynx", ylab="trappings" )
plot.ts ( presidents, main="Presidents", ylab="approval" )
plot.ts ( Seatbelts[,"DriversKilled"], main="UK drivers",
ylab="deaths" )
plot.ts ( sunspot.month, main="Sun Spots",
ylab="number of sunspots" )

• plot.ts is the command for plotting time series.

• The mar argument in the par command says how many lines are in the margins of
each plot. Those lines are used for titles and axis labels. The command is used here
to decrease the default so there is less white space between the plots, hence more
room for each plot.

The data sets in Figure 7.5 exhibit a feature common to many time series: if one data
point is large, the next tends to be large, and if one data point is small, the next tends to
be small; i.e., Yt and Yt+1 are dependent. The dependence can be seen in Figure 7.6 which
plots Yt+1 vs. Yt , for the Beaver and President datasets. The upward trend in each panel
shows the dependence. Time series analysts typically use the term autocorrelation — the
prefix auto refers to the fact that the time series is correlated with itself — even though
they mean dependence. R has the built-in function acf for computing autocorrelations.
The following snippet shows how it works.
> acf ( beaver1$temp, plot=F, lag.max=5 )
7.2. TIME SERIES AND MARKOV CHAINS 398
Beaver Mauna Loa

360
Temperature

CO2 (ppm)
37.0

340
320
36.4
0 20 40 60 80 100 1960 1970 1980 1990

Time
DAX Time
UK Lung Disease
6000

monthly deaths
Closing Price

3000
4000
2000

1500
1992 1994 1996 1998 1974 1976 1978 1980

Time Lynx
Canadian Time
Presidents
5000

70
trappings

approval

50
0 2000

30

1820 1860 1900 1945 1955 1965 1975

UK Time
drivers SunTime
Spots
number of sunspots

200
160
deaths

100
60 100

1970 1975 1980 1985 1750 1850 1950

Figure 7.5: Time series. Beaver: Body temperature of a beaver, recorded every 10 min-
utes; Mauna Loa: Atmospheric concentration of CO2 ; DAX: Daily closing prices of the
DAX stock exchange in Germany; UK Lung Disease: monthly deaths from bronchitis,
emphysema and asthma; Canadian Lynx: annual number of trappings; Presidents: quar-
terly approval ratings; UK drivers: deaths of car drivers; Sun Spots: monthly sunspot
numbers. In all cases the abscissa is time.
7.2. TIME SERIES AND MARKOV CHAINS 399

Autocorrelations of series ’beaver1$temp’, by lag

0 1 2 3 4 5
1.000 0.826 0.686 0.580 0.458 0.342
The six numbers in the bottom line are Cor(Yt , Yt ), Cor(Yt , Yt+1 ), . . . , Cor(Yt , Yt+5 ) and are
referred to as autocorrelations of lag 0, lag 1, . . . , lag 5. Those autocorrelations can, as
usual, be visualized with plots as in Figure 7.7.

Beaver Presidents

30 50 70
37.0
yt+1

yt+1
36.4

36.4 37.0 30 60

yt yt

Figure 7.6: Yt+1 plotted against Yt for the Beaver and Presidents data sets

Figure 7.6 was produced by the following snippet.

dim ( beaver1 )
plot ( beaver1$temp[-114], beaver1$temp[-1], main="Beaver",
xlab=expression(y[t]), ylab=expression(y[t+1]) )
length ( presidents )
plot ( presidents[-120], presidents[-1], main="Presidents",
xlab=expression(y[t]), ylab=expression(y[t+1]) )
7.2. TIME SERIES AND MARKOV CHAINS 400

lag = 0 lag = 1

37.2

37.2
Yt+k

Yt+k
36.8

36.8
36.4

36.4
36.4 36.8 37.2 36.4 36.8 37.2

Yt Yt

lag = 2 lag = 3
37.2

37.2
Yt+k

Yt+k
36.8

36.8
36.4

36.4

36.4 36.8 37.2 36.4 36.8 37.2

Yt Yt

lag = 4 lag = 5
37.4

37.4
37.0

37.0
Yt+k

Yt+k
36.6

36.6

36.4 36.8 37.2 36.4 36.8 37.2

Yt Yt

Figure 7.7: Yt+k plotted against Yt for the Beaver data set and lags k = 0, . . . , 5
7.2. TIME SERIES AND MARKOV CHAINS 401

Figure 7.7 was produced by the following snippet.

par ( mfrow=c(3,2) )
temp <- beaver1$temp
n <- length(temp)
for ( k in 0:5 ) {
x <- temp[1:(n-k)]
y <- temp[(1+k):n]
plot ( x, y, xlab=expression(Y[t]),
ylab=expression(Y[t+k]), main=paste("lag =", k) )
}

Because time series data cannot usually be treated as independent, we need special
methods to deal with them. It is beyond the scope of this book to present the major the-
oretical developments of time series methods. As Figure 7.5 shows, there can be a wide
variety of structure in time series data. In particular, the Beaver, and Presidents data sets
have no structure readily apparent to the eye; DAX has seemingly minor fluctuations im-
posed on a general increasing trend; UK Lung Disease and UK drivers have an annual
cycle; Mauna Loa has an annual cycle imposed on a general increasing trend; and Cana-
dian Lynx and Sun Spots are cyclic, but for no obvious reason and with no obvious length
of the cycle. In the remainder of this section we will show, by analyzing some of the data
sets in Figure 7.5, some of the possibilities.

Beaver Our goal is to develop a more complete picture of the probabilistic structure of
the {Yt }’s. To that end, consider the following question. If we’re trying to predict Yt+1 , and
if we already know Yt , does it help us also to know Yt−1 ? I.e., are Yt−1 and Yt+1 conditionally
independent given Yt ? That question can be answered visually with a coplot (Figures 2.15
and 2.16). Figure 7.8 shows the coplot for the Beaver data.

Figure 7.8 was produced by the following snippet.

temp <- beaver1$temp

n <- length ( temp )
coplot ( temp[3:n] ~ temp[1:(n-2)] | temp[2:(n-1)],
xlab=c (expression(Y[t-1]), expression(Y[t])),
7.2. TIME SERIES AND MARKOV CHAINS 402

Yt
36.4 36.6 36.8 37.0 37.2 37.4

36.4 36.8 37.2 36.4 36.8 37.2

37.2
36.8
36.4
Yt+1

37.2
36.8
36.4

36.4 36.8 37.2

Yt!1

Figure 7.8: coplot of Yt+1 as a function of Yt−1 given Yt for the Beaver data set
7.2. TIME SERIES AND MARKOV CHAINS 403

ylab=expression(Y[t+1]) )

The figure is ambiguous. In the first, second, and sixth panels, Yt+1 and Yt−1 seem
to be linearly related given Yt , while in the third, fourth, and fifth panels, Yt+1 and Yt−1
seem to be independent given Yt . We can examine the question numerically with the
partial autocorrelation, the conditional correlation of Yt+1 and Yt−1 given Yt . The following
snippet shows how to compute partial autocorrelations in R using the function pacf.
> pacf ( temp, lag.max=5, plot=F )

Partial autocorrelations of series ’temp’, by lag

1 2 3 4 5
0.826 0.014 0.031 -0.101 -0.063
The numbers in the bottom row are Cor(Yt , Yt+k | Yt+1 , . . . , Yt+k−1 ). Except for the first,
they’re small. Figure 7.8 and the partial autocorrelations suggest that a model in which
Yt+1 ⊥ Yt−1 | Yt would fit the data well. And the second panel in Figure 7.7 suggests that
a model of the form Yt+1 = β0 + β1 Yt + t+1 might fit well. Such a model is called an
autoregression. R has a function ar for fitting them. Here’s how it works with the Beaver
data.
> fit <- ar ( beaver1$temp, order.max=1 )
> fit # see what we’ve got

Call:
ar(x = beaver1$temp, order.max = 1)

Coefficients:
1
0.8258

Order selected 1 sigma^2 estimated as 0.01201

The 0.8258 means that the fitted model is Yt+1 = β0 + 0.8258Yt + t+1 . You can see
whether the 0.8258 makes sense by examing the second panel of Figure 7.7. The t ’s
have an estimated variance of 0.012. fit$x.mean shows that β̂0 = 36.86. Finally,
qqnorm(fit$resid) (Try it.) √ shows a nearly linear plot, except for one point, indicating
that Yt+1 ∼ N(36.86 + .8258Yt , .012) is a reasonably good model, except for one outlier.
7.2. TIME SERIES AND MARKOV CHAINS 404

Mauna Loa The Mauna Loa data look like an annual cycle superimposed on a steadily
increasing long term trend. Our goal is to estimate both components and decompose the
data as
Yt = long term trend + annual cycle + unexplained variation.
Our strategy, because it seems easiest, is to estimate the long term trend first, then use
deviations from the long term trend to estimate the annual cycle. A sensible estimate of
the long term trend at time t is the average of a year’s CO2 readings, for a year centered at
t. Thus, let
.5yt−6 + yt−5 + · · · + yt+5 + .5yt+6
ĝ(t) = (7.6)
12
where g(t) represents the long term trend at time t. R has the built-in command filter to
compute ĝ. The result is shown in Figure 7.9 (a) which also shows how to use filter.
Deviations from ĝ are co2 - g.hat. See Figure 7.9 (b). The deviations can be grouped
by month, then averaged. The average of the January deviations, for example, is a good
estimate of how much the January CO2 deviates from the long term trend, and likewise for
other months. See Figure 7.9 (c). Finally, Figure 7.9 (d) shows the data, ĝ, and the fitted
values ĝ + monthly effects. The fit is good: the fitted values differ very little from the data.

Figure 7.9 was produced by the following snippet.

filt <- c ( .5, rep(1,11), .5 ) / 12

g.hat <- filter ( co2, filt )
par ( mfrow=c(2,2) )
plot.ts ( co2, main="(a)" )
lines ( g.hat )
resids <- co2 - g.hat
plot.ts ( resids, main="(b)" )
resids <- matrix ( resids, nrow=12 )
cycle <- apply ( resids, 1, mean, na.rm=T )
plot ( cycle, type="b", main="(c)" )
plot.ts ( co2, type="p", pch=".", main="(d)" )
lines ( g.hat )
lines ( g.hat + cycle )

DAX Yt is the closing price of the German stock exchange DAX on day t. Investors
often care about the rate of return Yt∗ = Yt+1 /Yt , so we’ll have to consider whether to
7.2. TIME SERIES AND MARKOV CHAINS 405

(a) (b)

4
360

2
resids
340
co2

0
!2
320

!4
1960 1980 1960 1980

Time Time

(c) (d)
3

360
2
1
cycle

340
co2
0
!3 !2 !1

320

2 4 6 8 12 1960 1980

Index Time

Figure 7.9: (a): CO2 and ĝ; (b): residuals; (c): residuals averaged by month; (d): data, ĝ,
and fitted values
7.2. TIME SERIES AND MARKOV CHAINS 406

analyze the Yt ’s directly or convert them to Yt∗ ’s first. Figure 7.10 is for the DAX prices
directly. Panel (a) shows the Yt ’s. It seems to show minor fluctuations around a steadily
increasing trend. Panel (b) shows the time series of Yt − Yt−1 . It seems to show a series
of fluctuations approximately centered around 0, with no apparent pattern, and with larger
fluctuations occuring later in the series. Panel (c) shows Yt versus Yt−1 . It shows a strong
linear relationship between Yt and Yt−1 . Two lines are drawn on the plot: the lines Yt = β0 +
β1 Yt−1 for (β0 , β1 ) = (0, 1) and for (β0 , β1 ) set equal to the ordinary regression coefficients
found by lm. The two lines are indistinguishable, suggesting that Yt ≈ Yt−1 is a good model
for the data. Panel (d) is a Q-Q plot of Yt − Yt−1 . It is not approximately linear, suggesting
that Yt ∼ N(Yt−1 , σ) is not a good model for the data.

Figure 7.10 was produced by the following snippet.

par ( mfrow=c(2,2) )
plot.ts ( DAX, main="(a)" )
plot.ts ( diff(DAX), ylab = expression ( DAX[t] - DAX[t-1] ),
main="(b)" )
plot ( DAX[-n], DAX[-1], xlab = expression ( DAX[t-1] ),
ylab = expression ( DAX[t] ), main="(c)", pch="." )
abline ( 0, 1 )
abline ( lm ( DAX[-1] ~ DAX[-n] )$coef, lty=2 )
qqnorm ( diff(DAX), main="(d)", pch="." )

• The R command diff is for taking differences, typically of time series. diff(y)
yields y[2]-y[1], y[3]-y[2], ... which could also be accomplished eas-
ily enough without using diff: y[-1] - y[-n]. But additional arguments, as
in diff ( y, lag, differences ) make it much more useful. For exam-
ple, diff(y,lag=2) yields y[3]-y[1], y[4]-y[2], ... while diff ( y,
differences=2 ) is the same as diff ( diff(y) ). The latter is a construct
very useful in time series analysis.

Figure 7.11 is for the Yt∗ ’s. Panel (a) shows the time series. It shows a seemingly
patternless set of data centered around 1. Panel (b) shows the time series of Yt∗ − Yt−1 ∗
,a
∗ ∗
seemingly patternless set of data centered at 0. Panel (c) shows Yt versus Yt−1 . It shows
no apparent relationship between Yt∗ and Yt−1 ∗
, suggesting that Yt∗ ⊥ Yt−1
∗
is a good model
for the data. Panel (d) is a Q-Q plot of Yt . It is approximately linear, suggesting that Yt∗ ∼
∗

N(µ, σ) is a good model for the data, with a few outliers on both the high and low ends.
7.2. TIME SERIES AND MARKOV CHAINS 407

(a) (b)

DAXt − DAXt−−1
5000
DAX

0
2000

−200
1992 1995 1998 1992 1995 1998

Time Time

(c) (d)
Sample Quantiles
5000
DAXt

0
2000

−200

2000 4000 6000 −3 −1 1 3

DAXt−−1 Theoretical Quantiles

Figure 7.10: DAX closing prices. (a): the time series of Yt ’s; (b): Yt − Yt−1 ; (c): Yt versus
Yt−1 ; (d): QQ plot of Yt − Yt−1 .
7.2. TIME SERIES AND MARKOV CHAINS 408

The mean and SD of the Y ∗ ’s are about 1.000705 and 0.01028; so Y ∗ ∼ N(1.0007, 0.01)
might be a good model.

(a) (b)

ratet − ratet−−1

0.05
1.00
rate

−0.05
0.92

0 500 1500 0 500 1500

Time Time

(c) (d)
Sample Quantiles

●●●
●
●●
●
●●
●
●
●
●●
●
●
●
●●
●
●●
●
●
●
1.00

1.00

●
●
●
●●
●
●
●●
●
●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●●
ratet

●●
●
●
●●
●
●
●●
●
●
●
●●
●
●
●●
●
●
●●
●
●
●●
●
●●
●
●●
●
●
●
●
●
0.92

0.92

0.92 0.98 1.04 −3 −1 1 3

ratet−−1 Theoretical Quantiles

Figure 7.11: DAX returns. (a): the time series of Yt∗ ’s; (b): Yt∗ − Yt−1
∗
; (c): Yt∗ versus Yt−1
∗
;
∗
(d): QQ plot of Yt .

Figure 7.11 was produced by the following snippet.

par ( mfrow=c(2,2) )
plot.ts ( rate, main="(a)" )
7.3. SURVIVAL ANALYSIS 409

plot.ts ( diff(rate), ylab = expression ( rate[t] - rate[t-1] ),

main="(b)" )
plot ( rate[-n2], rate[-1], xlab = expression ( rate[t-1] ),
ylab = expression ( rate[t] ), main="(c)", pch="." )
qqnorm ( rate, main="(d)" )

We now have two possible models for the DAX data: Yt ≈ Yt−1 with a still to be
determined distribution and Y ∗ ∼ N(1.0007, 0.01) with the Yt∗ ’s mutually independent.
Both seem plausible on statistical grounds. (But see Exercise 8 for further development.)
It is not necessary to choose one or the other. Having several ways of describing a data
set is useful. Each model gives us another way to view the data. Economists and investors
might prefer one or the other at different times or for different purposes. There might
even be other useful models that we haven’t yet considered. Those would be beyond the
scope of this book, but could be covered in texts on time series, financial mathematics,
econometrics, or similar topics.

7.3 Survival analysis

In many studies, the random variable is the time at which an event occurs. For example,

medicine The time until a patient dies.

neurobiology The time until a neuron fires.

quality control The time until a computer crashes.

higher education The time until an Associated Professor is promoted to Full Professor.

Such data are called survival data. For the i’th person, neuron, computer, etc., there is a
random variable
yi = time of event on i’th unit.
We usually call yi the lifetime, even though the event is not necessarily death. It is often the
case with survival data that some measurements are censored. For example, if we study
a university’s records to see how long it takes to get promoted from Associate to Full
Professor, we will find some Associate Professors leave the university — either through
retirement or by taking another job — before they get promoted while others are still
Associate Professors at the time of our study. For these people we don’t know their time
7.3. SURVIVAL ANALYSIS 410

of promotion. If either (a) person i left the university after five years, or (b) person i
became Associate Professor five years prior to our study, then we don’t know yi exactly.
All we know is yi > 5. This form of censoring is called right censoring. In some data sets
there may also be left censoring or interval censoring. Survival analysis typically requires
specialized statistical techniques. R has a package of functions for this purpose; the name
of the package is survival. The survival package is automatically distributed with
R. To load it into your R session, type library(survival). The package comes with
functions for survival analysis and also with some example data sets. Our next example
uses one of those data sets.
Example 7.2 (Bladder Tumors)
This example comes from a study of bladder tumors, originally published in Byar [1980]
and later reanalyzed in Wei et al. [1989]. Patients had bladder tumors. The tumors
were removed and the patients were randomly assigned to one of three treatment
groups (placebo, thiotepa, pyridoxine). Then the patients were followed through time
to see whether and when bladder tumors would recur. R’s survival package has the
data for the first two treatment groups, placebo and thiotepa. Type bladder to see it.
(Remember to load the survival package first.) The last several lines look like this.
id rx number size stop event enum
341 83 2 3 4 54 0 1
342 83 2 3 4 54 0 2
343 83 2 3 4 54 0 3
344 83 2 3 4 54 0 4
345 84 2 2 1 38 1 1
346 84 2 2 1 54 0 2
347 84 2 2 1 54 0 3
348 84 2 2 1 54 0 4
349 85 2 1 3 59 0 1
350 85 2 1 3 59 0 2
351 85 2 1 3 59 0 3
352 85 2 1 3 59 0 4

• id is the patient’s id number. Note that each patient has four lines of data. That’s
to record up to four recurrences of tumor.

• rx is the treatment: 1 for placebo; 2 for thiotepa.

• number is the number of tumors the patient had at the initial exam when the
patient joined the study.
7.3. SURVIVAL ANALYSIS 411

• size is the size (cm) of the largest initial tumor.

• stop is the time (months) of the observation.

• event is 1 if there’s a tumor; 0 if not.

• enum line 1, 2, 3, or 4 for each patient

For example, patient 83 was followed for 54 months and had no tumor recurrences;
patient 85 was followed for 59 months and also had no recurrences. But patient 84,
who was also followed for 54 months, had a tumor recurrence at month 38 and no fur-
ther recurrences after that. Our analysis will look at the time until the first recurrence,
so we want bladder[bladder$enum==1,], the last several lines of which are

id rx number size stop event enum

329 80 2 3 3 49 0 1
333 81 2 1 1 50 0 1
337 82 2 4 1 4 1 1
341 83 2 3 4 54 0 1
345 84 2 2 1 38 1 1
349 85 2 1 3 59 0 1

Patients 80, 81, 83, and 85 had no tumors for as long as they were followed; their
data is right-censored. The data for patients 82 and 84 is not censored; it is observed
exactly.
Figure 7.12 is a plot of the data. The solid line is for placebo; the dashed line for
thiotepa. The abscissa is in months. The ordinate shows the fraction of patients who
have survived without a recurrence of bladder tumors. The plot shows, for example,
that at 30 months, the survival rate without recurrence is about 50% for thiotepa pa-
tients compared to a little under 40% for placebo patients. The circles on the plot show
censoring. I.e., the four circles on the solid curve between 30 and 40 months repre-
sent four placebo patients whose data was right-censored. There is a circle at every
censoring time that is not also the time of a recurrence (for a different patient).

Figure 7.12 was produced with the snippet

event.first <- bladder[,"enum"] == 1

blad.surv <- Surv ( bladder[event.first,"stop",],
7.3. SURVIVAL ANALYSIS 412

1.0
fraction without recurrence

0.8

●
●● ●
0.6

●
●
●
● ●●● ●● ● ●
0.4

●
●
●●
●● ● ● ●
0.2
0.0

0 10 20 30 40 50 60

months

Figure 7.12: Survival curve for bladder cancer. Solid line for placebo; dashed line for
thiotepa.
7.3. SURVIVAL ANALYSIS 413

bladder[event.first,"event"] )
blad.fit <- survfit ( blad.surv ~ bladder[event.first,"rx"] )
plot ( blad.fit, conf.int=FALSE, mark=1, xlab="months",
ylab="fraction without recurrence", lty=1:2 )

• Surv is R’s function for creating a survival object. You can type
print(blad.surv) and summary(blad.surv) to learn more about survival ob-
jects.

• survfit computes an estimate of a survival curve.

In survival analysis we think of y1 , . . . , yn as a sample from a distribution F, with

density f , of lifetimes. In survival analysis, statisticians often work with the survivor
function S (t) = 1 − F(t) = P[yi > t], the probability that a unit survives beyond time t.
The lines in Figure 7.12 are the so-called Kaplan-Meier estimates of S for patients in the
thiotepa and placebo groups, which arise from the following argument. Partition R+ into
intervals (0, t1 ], (t1 , t2 ], . . . , and let pi = P[y > ti |y > ti−1 ]. Then for each i, S (ti ) = ij=1 pi .
Q
The pi ’s can be estimated from data as p̂i = (r(ti ) − di )/r(ti ) where r(ti ) is the number of
people at risk (in the study but not yet dead) at time ti−1 and di is the number of deaths in
the interval (ti−1 , ti ]. Thus, Ŝ (ti ) = ij=1 (r(ti ) − di )/r(ti ). As the partition becomes finer,
Q
most terms in the product are equal to one; only those intervals with a death contribute a
term that is not one. The limit yields the Kaplan-Meier estimate
Y r(yi ) − di
Ŝ (t) = .
i:y <t
r(yi )
i

This estimate is reasonably accurate so long as r(t) is reasonably large; Ŝ (t) is more accu-
rate for small values of t than for large values of t; and there is no information at all for
estimating S (t) for t > max{yi }.
Survival data is often modelled in terms of the hazard function
P[y ∈ [t, t + h) | y ≥ t] P[y ∈ [t, t + h)] f (t)
h(t) = lim = lim = . (7.7)
h→0 h h→0 h P[y ≥ t] S (t)
The interpretation of h(t) is the fraction, among people who have survived to time t, of
those who will die soon thereafter. There are several parametric families of distributions
for lifetimes in use for survival analysis. The most basic is the exponential — f (y) =
7.3. SURVIVAL ANALYSIS 414

(1/λ) exp−y/λ — which has hazard function h(y) = 1/λ, a constant. A constant hazard
function says, for example, that young people are just as likely to die as old people, or that
new air conditioners are just as likely to fail as old air conditioners. For many applications
that assumption is unreasonable, so statisticians may work with other parametric families
for lifetimes, especially the Weibull, which has h(y) = λα(λy)α−1 , an increasing function
of y if α > 1. We will not dwell further on parametric models; the interested reader should
refer to a more specialized source.
However, the goal of survival analysis is not usually to estimate S and h, but to compare
the survivor and hazard functions for two groups such as treatment and placebo or to see
how the survivor and hazard functions vary as functions of some covariates. Therefore it
is not usually necessary to estimate S and h well, as long as we can estimate how S and
h differ between groups, or as a function of the covariates. For this purpose it has become
common to adopt a proportional hazards model:

h(y) = h0 (y) exp(β0 x) (7.8)

where h0 is the baseline hazard function that is adjusted according to x, a vector of covari-
ates and β, a vector of coefficients. Equation 7.8 is known as the Cox proportional hazards
model. The goal is usually to estimate β. R’s survival package has a function for fitting
Equation 7.8 to data.

Example 7.3 (Bladder Tumors, cont.)

This continues Example 7.2. Here we adopt the Cox proportional hazards model and
see how well we can estimate the effect of the treatment thiotepa compared to placebo
in preventing recurrence of bladder tumors. We will also examine the effects of other
potential covariates.
We’d like to fit the Cox proportional hazards model h(y) = h0 (y) exp(βtrt · trt) where
trt is an indicator variable that is 1 for patients on thiotepa and 0 for patients on
placebo; but first we check whether such a model looks plausible; i.e. whether the
hazards look R y proportional. Starting from Equation 7.7 we can integrate both sides to
get H(y) ≡ 0 h(z) dz = − log S (y). H(y) is called the cumulative hazard function. Thus,
if the two groups have proportional hazards, they also have proportional cumulative
hazard functions and log survivor functions. Figure 7.13 plots the estimated cumula-
tive hazard and log (cumulative hazard) functions for the bladder tumor data. The log
(cumulative hazard) functions look parallel, so the proportional hazards assumption
looks reasonable.
7.3. SURVIVAL ANALYSIS 415

●
●● ● ● ●●● ● ●
1.2

log(cumulative hazard)
●●

0
●● ●●
●
●●
●●●
cumulative hazard

●● ● ●
●

−1
● ●● ●
0.8

● ●●
●●●
●● ● ●
●
●
●
−2
0.4

●
●●
●
−3
0.0

−4

0 10 30 50 20 50 200

months months

Figure 7.13: Cumulative hazard and log(hazard) curves for bladder cancer. Solid line for
thiotepa; dashed line for placebo.
7.4. EXERCISES 416

Figure 7.13 was produced with the snippet

plot ( blad.fit, conf.int=FALSE, mark=1, xlab="months",

ylab="cumulative hazard", lty=1:2, fun="cumhaz" )
plot ( blad.fit, conf.int=FALSE, mark=1, xlab="months",
ylab="log(cumulative hazard)", lty=1:2, fun="cloglog" )

• The fun argument allows transformations of the survival curve. fun="cumhaz"

plots the cumulative hazard function and fun="cloglog" plots the log (cumula-
tive hazard) function.

Since the proportional hazards model looks reasonable, we fit it:

blad.cox <- coxph ( blad.surv ˜ bladder[event.first,"rx"] ).
Printing blad.cox yields
Call:
coxph(formula = blad.surv ˜ bladder[event.first, "rx"])

coef exp(coef) se(coef) z p

bladder[event.first, "rx"] -0.371 0.69 0.303 -1.22 0.22

Likelihood ratio test=1.54 on 1 df, p=0.215 n= 85

The estimated coefficient is β̂trt = −0.371. Thus the hazard function for thiotepa pa-
tients is estimated to be exp(−0.371) = 0.69 times that for placebo patients. The
standard error of β̂trt = −0.371 is about 0.3; so β̂trt is accurate to about ±0.6 or so.

7.4 Exercises
1. In the orthodont data, M09 doesn’t follow the general pattern. Maybe there’s an
error in the data or maybe that person really did grow in an unusual way. Fit a model
similar to ortho.fit1 but excluding the data from M09. Does anything change in
an important way?
2. In the orthodont data, it seems clear that different individuals have different inter-
cepts. It’s not as clear whether they have different slopes. Conduct an analysis to
7.4. EXERCISES 417

investigate that possibility. Your analysis must address at least two questions. First,
do different individuals seem to have different slopes? Second, is the apparent differ-
ence in Male and Female slopes found by ortho.fit4 really a difference between
males and females on average, or is it due to the particular males and females who
happened to be chosen for this study?
3. Carry out a Bayesian analysis of the model in Equation 7.2.
4. (a) Make plots analagous to Figures 7.6 and 7.7, compute autocorrelations, and
interpret for the other datasets in Figure 7.5.
(b) Make plots analagous to Figure 7.8, compute partial autocorrelations, and in-
terpret for the other data sets in Figure 7.5.
5. Create and fit a good model for body temperatures of the second beaver. Use the
dataset beaver2.
6. (a) Why does Equation 7.6 average over a year? Why isn’t it, for example,
.5yt−k + yt−k+1 + · · · + yt+k−1 + .5yt+k
ĝ(t) =
2k
for some k , 6?
(b) Examine ĝ in Equation 7.6. Use R if necessary. Why are some of the entries
NA?
7. The R code for Figure 7.9 contains the lines

resids <- matrix ( resids, nrow=12 )

cycle <- apply ( resids, 1, mean, na.rm=T )

Would the following lines work instead?

resids <- matrix ( resids, ncol=12 )

cycle <- apply ( resids, 2, mean, na.rm=T )

Why or why not?

8. Figure 7.10 and the accompanying text suggest that Yt ≈ Yt−1 is a good model for the
DAX data. But that doesn’t square with the observation that the Yt ’s have a generally
increasing trend.
7.4. EXERCISES 418

(a) Find a quantitative way to show the trend.

(b) Say why the DAX analysis missed the trend.
(c) Improve the analysis so it’s consistent with the trend.

9. Figures 7.10 and 7.11 and the accompanying text analyze the DAX time series as
though it has the same structure throughout the entire time. Does that make sense?
Think of and implement some way of investigating whether the structure of the
series changes from early to late.

10. Choose one or more of the other EU Stock Markets that come with the DAX data.
Investigate whether it has the same structure as the DAX.

11. (a) Make a plausible analysis of the UK Lung Disease data.

(b) R has the three data sets ldeaths, fdeaths, and mdeaths which are the total
deaths, the deaths of females, and the deaths of males. Do the deaths of females
and males follow similar distributional patterns? Justify your answer.

12. Make a plausible analysis of the Presidents approval ratings.

13. (a) Make a plausible analysis of the UK drivers deaths.

(b) According to R, “Compulsory wearing of seat belts was introduced on 31 Jan
1983.” Did that effect the number of deaths? Justify your answer.
(c) Is the number of deaths related to the number of kilometers driven? (Use the
variable kms in the Seatbelts data set.) Justify your answer.

14. This question follows up Example 7.3. In the example we analyzed the data to learn
the effect of thiotepa on the recurrence of bladder tumors. But the data set has two
other variables that might be important covariates: the number of initial tumors and
the size of the largest initial tumor.

(a) Find the distribution of the numbers of initial tumors. How many patients had
1 initial tumor, how many had 2, etc?
(b) Divide patients, in a sensible way, into groups according to the number of
initial tumors. You must decide how many groups there should be and what
the group boundaries should be.
(c) Make plots similar to Figures 7.12 and 7.13 to see whether a proportional haz-
ard model looks sensible for number of initial tumors.
(d) Fit a proportional hazard model and report the results.
7.4. EXERCISES 419

(e) Repeat the previous analysis, but for size of largest initial tumor.
(f) Fit a proportional hazard model with three covariates: treatment, number of
initial tumors, size of largest initial tumor. Report the results.
 Chapter 8

Mathematical Statistics

8.1 Properties of Statistics

8.1.1 Sufficiency
Consider the following two facts.

1. Let Y1 , . . . , Yn ∼ i.i.d. Poi(λ). Chapter 2, Exercise 7 showed that `(λ) depends only
P
on Yi and not on the specific values of the individual Yi ’s.

2. Let Y1 , . . . , Yn ∼ i.i.d. Exp(λ). Chapter 2, Exercise 20 showed that `(λ) depends only
P
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. Thus, for inference about λ it suffices to know `(λ), and therefore,
P
for Poisson and Exponential data, it suffices to know Yi . We don’t need to know the
individual Yi ’s. We say that Yi is a sufficient statistic for λ.
P
Section 8.1.1 examines the general concept of sufficiency. We work in the context of a
parametric family. The idea of sufficiency is formalized in Definition 8.1.

Definition 8.1. 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.

420
8.1. PROPERTIES OF STATISTICS 421

The idea is that once the data 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
Q
likelihood function and make inference about θ, we need only know T (y), not anything
else about y. For our Poisson and Exponential examples we can take T (y) = yi .
P

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 possi-
ble 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:

1. Generate yi = 0, 1, 2, 3 with probabilities (1 − θ)3 , 3θ(1 − θ)2 , 3θ2 (1 − θ), θ3 ,

P
respectively.

yi = 0, generate (0, 0, 0)
P
2. (a) If
yi = 1, generate (1, 0, 0), (0, 1, 0), or (0, 0, 1) each with probability 1/3.
P
(b) If
yi = 2, generate (1, 1, 0), (1, 0, 1), or (0, 1, 1) each with probability 1/3.
P
(c) If
yi = 3, generate (1, 1, 1)
P
(d) If

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 need only know the outcome of the first
P P
step. The second step is irrelevant. I.e., we need only know yi . In other words, yi is
sufficient.
8.1. PROPERTIES OF STATISTICS 422

For an example of another type, let y1 , . . . , yn ∼ i.i.d. U(0, θ). What is a sufficient
statistic for θ?

 θ1n if yi < θ for i = 1, . . . , n

p(y | θ) = 

0 otherwise

1
= n 1(0,θ) (y(n) )
θ
shows that y(n) , the maximum of the yi ’s, is a one dimensional sufficient statistic for θ.
Example 8.1
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
estimate T , the total number of German tanks and had, as data, the serial numbers
of captured tanks. See Exercise 23 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.

If y is a random variable whose values are in a space Y, then y is a random variable

whose values are in Yn . For any statistic T we can divide Yn into subsets indexed by T .
I.e., for each value t, we define the subset

Ytn = {y ∈ Yn : T (y) = t}

Then T is a sufficient statistic if and only if

p(y | y ∈ Ytn )

does not depend on θ.

Sometimes sufficient statistics are higher dimensional. For example, let y1 , . . . , yn ∼
i.i.d. Gam(α, β). Then
!n Y 
Y Y 1 1 α−1
α−1 −yi /β − yi /β
p(yi | α, β) = =
P
y e y i e
Γ(α)βα i Γ(α)βα
so T (y) = ( yi , yi ) is a two dimensional sufficient statistic.
Q P
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 also sufficient. So in the Poisson, Exponential, and Bernoulli exam-
ples where yi was sufficient, ȳ = yi /n is also sufficient. But the lack of uniqueness is
P P
8.1. PROPERTIES OF STATISTICS 423

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. 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 T 2 = (y1 , T ) is a two dimensional sufficient 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 sufficient 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 8.2 is for statistics that
have been reduced as much as possible without losing sufficiency.

Definition 8.2. A sufficient statistic T (y) is called minimal sufficient if, for any other suf-
ficient statistic T 2 , T (y) is a function of T 2 (y).

This book does not delve into methods for finding minimal sufficient statistics. In most
cases the user can recognize whether a statistic is minimal sufficient.
Does the theory of sufficiency imply that statisticians need look only at sufficient statis-
tics and not at other aspects of the data? Not quite. Let y1 , . . . , yn be binary random vari-
ables and suppose we adopt the model y1 , . . . , yn ∼ i.i.d. Bern(θ). Then for estimating θ
we need look only at yi . But suppose (y1 , . . . , yn ) turn out to be
P

· · · }1,
· · · }0 1| 1{z
0| 0{z
many 0’s many 1’s

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 or that θ is increasing over time. 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.

8.1.2 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 pa-
rameter estimation, by the notion of consistency. To say whether an estimator is consistent
8.1. PROPERTIES OF STATISTICS 424

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 T n : Rn → R. I.e.
T n is a real-valued function of (y1 , . . . , yn ). For example, if we’re trying to estimate µ we
might take T n = n−1 n1 yi .
P

Definition 8.3. The sequence of estimators T 1 , T 2 , . . . is said to be consistent for the pa-
rameter θ if for every θ and for every  > 0.

lim P[|T n − θ| < ] = 1.

n→∞

For example, the Law of Large Numbers, Theorem 1.12, says the sequence of sample
means {T n = n−1 n1 yi } is consistent for µ. Similarly, let S n = n−1 i (yi − T n )2 be the
P P
sample variance. Then {S n } is consistent for σ2 . More generally, m.l.e.’s are consistent.

Theorem 8.1. 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 condi-
tions, {g(θ̂n )} is a consistent sequence of estimators for g(θ).

Proof. The proof requires regularity conditions relating to differentiability and the inter-
change of integral and derivative. It is beyond the scope of this book. 

Consistency is a good property; one should be wary of an inconsistent estimator. On

the other hand, consistency alone does not guarantee that a sequence of estimators is op-
timal, 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 {Rn } 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.

Definition 8.4. 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.
8.1. PROPERTIES OF STATISTICS 425

A biased estimator Let y1 , . . . , yn ∼ i.i.d. N(µ, σ) and consider

X
σ̂2 = n−1 (yi − ȳ)2

as an estimate of σ2 . Theorem 5.31 says that (yi − ȳ)2 /σ2 ∼ χ2n−1 and therefore
P
E[σ̂2 ] = n−1n
σ2 . I.e., σ̂2 is a biased estimator of σ2 . It’s bias is −σ2 /n. Some
statisticians prefer to use the unbiased estimator σ̃2 = (n − 1)−1 (yi − ȳ)2 .
P

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 If θ̂ is an estimate of θ, then the mean squared error (MSE) of θ̂
is E[(θ̂ − θ)2 ]. MSE is a combination of variance and bias:
h i h i
MSE(θ̂) = E (θ̂ − θ)2 = E (θ̂ − Eθ̂ + Eθ̂ − θ)2
h i
= E (θ̂ − Eθ̂)2 + 2(θ̂ − Eθ̂)(Eθ̂ − θ) + (Eθ̂ − θ)2
h i
= E (θ̂ − Eθ̂)2 + (Eθ̂ − θ)2 = Var(θ̂) + [bias(θ̂)]2 .

Whenever possible, one would like to have an estimator with minimum bias and min-
imum variance; those two properties would also imply minimum MSE. But it is not al-
ways possible to achieve both desiderata simultaneously. For example, let Y1 , . . . , Yn ∼
i.i.d. N(µ, σ) and suppose we want an estimate of σ2 . We have looked at two estima-
tors so far: the mle σ̂2 = n−1 (yi − ȳ)2 and the unbiased estimator σ̃2 = (n/(n − 1))σ̂2 .
P
We already know the bias of σ̃2 is 0 and the bias of σ̂2 is −σ2 /n. We also know that
(yi − ȳ)2 /σ2 ∼ χ2n−1 . Therefore Var( (yi − ȳ)2 /σ2 ) = 2(n − 1); Var(σ̂2 ) = 2(n − 1)σ4 /n2
P P
and Var(σ̃2 ) = 2σ4 /(n − 1). Thus,

2(n − 1)σ4 σ4 (2n − 1)σ4 (2n2 − 3n + 1)σ4

MSE(σ̂2 ) = + 2 = =
n2 n n2 n2 (n − 1)
2n2 σ4 2σ4
< 2 = = MSE(σ̃2 ),
n (n − 1) n − 1
showing that the unbiased estimator does not minimize MSE. In choosing an estimator for
σ2 one must trade bias for variance. That is a common problem in statistics to which there
is no universal solution.
We now investigate the variance-bias tradeoff for estimating θ in a Binomial distribu-
tion. Let X1 , . . . , Xn ∼ i.i.d. Bin(n, θ). A sensible estimator of θ is X̄ ≡ Xi /n. We know
P
that X̄ is unbiased, and its variance is θ(1 − θ)/n, so its MSE is also θ(1 − θ)/n.
8.1. PROPERTIES OF STATISTICS 426

But we could take a Bayesian approach to the problem instead and use the posterior
mean, E[θ | X1 , . . . , Xn ], as an estimator of θ. For convenience, adopt the prior distribution
θ ∼ Be(α, β). Then the posterior is given by [θ | X1 , . . . , Xn ] ∼ Be(α + X, β + n − X) and
the posterior mean is θ̃ ≡ E[θ | X1 , . . . , Xn ] = (α + X)/(α + β + n). We want the MSE of θ̃,
which we shall calculate from its bias and variance.
X+α nθ + α
" #
E[θ̃] = E =
α+β+n α+β+n
X+α nθ(1 − θ)
" #
Var[θ̃] = Var = .
α+β+n (α + β + n)2

Therefore
!2
nθ(1 − θ) nθ + α
MSE(θ̃) = + −θ .
(α + β + n)2 α+β+n

We want to compare MSE(X̄) to MSE(θ̃), but the comparison might depend on n, α,

and β. Figure 8.1 shows the comparison for four values of n and four values of (α, β),
always keeping α = β. Note: the Bayes estimator when α = β = 0 is the same as X̄. Using
large values of (α, β) looks advantageous when θ ≈ .5; small values are advantageous
when θ is near 0 or 1. There is no general rule for choosing which estimator to use or
for choosing (α, β) if we decide to use the Bayes estimator. We might make the choice
according to a prior guess of whether θ is likely to be near 0, .5, or 1.

Figure 8.1 was produced with the following snippet.

theta <- seq ( 0, 1, length=60 )

mse <- function ( a, b, n, theta ) {

( n*theta*(1-theta ) ) / ( a+b+n )^2
+ ( (n*theta + a)/(a+b+n) - theta )^2
}

mse.out <- data.frame ( ab=NULL, n=NULL, mse=NULL )

for ( ab in c ( 0, .5, 1, 4 ) )
for ( n in c ( 5, 20, 100, 1000 ) ) {
mse.out <- rbind ( mse.out, cbind ( a=ab, b=ab, n=n,
theta=theta, mse=mse(ab,ab,n,theta) ) )
8.1. PROPERTIES OF STATISTICS 427

n = 100 n = 1000

0.00000 0.00015
0.0015
0.0000
mse

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0
n=5 n = 20
0.000 0.010 0.020
0.08
0.04
0.00

0.0 0.2 0.4 0.6 0.8 1.0 0.0 0.2 0.4 0.6 0.8 1.0

theta

Figure 8.1: Mean Squared Error for estimating Binomial θ. Sample size = 5, 20, 100,
1000. α = β = 0: solid line. α = β = 0.5: dashed line. α = β = 1: dotted line. α = β = 4:
dash–dotted line.
8.2. INFORMATION 428

}
mse.out$n <- factor ( paste("n =", mse.out$n),
levels = paste("n =", c(5,20,100,1000) ) )

xyplot ( mse ˜theta | n, groups=a, data=mse.out, type="l",

scales = list ( relation="free" ), lty=1:4 )

8.2 Information
We have talked loosely about the amount of information in a sample and we now want to
make the notion more precise. This is an important idea in statistics that can be found in
many texts. One devoted wholly to information is Kullback [1968].
Suppose that Y1 , . . . , Yn is a random sample from an unknown density. And suppose
there are only two densities under consideration, p1 and p2 . We know that the Yi ’s are a
sample from one of them, but we don’t know which. Now consider the first observation,
Y1 . How can we be precise about the amount of information that Y1 gives us for distin-
guishing between p1 and p2 ? We have already seen that the strength of evidence provided
by Y1 = y1 is the likelihood ratio p1 (y1 )/p2 (y1 ). The strength of evidence provided by the
Q
whole sample is the product i [p1 (yi )/p2 (yi )] of the evidence provided by the individual
observations. If we transform to a log scale, then the evidence is additive:
p1 (y1 , . . . , yn ) X p1 (yi )
log = log .
p2 (y1 , . . . , yn ) p2 (yi )
This seems a suitable quantity to call information. Therefore we make the following defi-
nition.

Definition 8.5. The information in a datum y for distinguishing between densities p1 and
(y)
p2 is log pp21 (y) .

The information in a random sample is the sum of the information in the individual el-
ements. The i’th term in the sum is log pp12 (y i)
(yi )
. That term is a function of yi and is therefore
a random variable whose distribution is determined by either p1 or p2 , whichever is the
true density generating the
R sample. If p1 is the true distribution, then the expected value of
(y)
that random variable is log pp21 (y) p1 (y) dy and is the expected amount of information from
a single observation. The Law of Large Numbers says that a large sample will contain
8.2. INFORMATION 429

approximately n times that amount of information. This view of information was origi-
nally discussed thoroughly in Kullback and Leibler [1951]. Accordingly, we make the
following definition.
Definition 8.6. I(p1 , p2 ) ≡ log pp21 (y)
R (y)
p1 (y) dy is called the Kullback-Leibler divergence
from p1 to p2 .
For example, suppose that we’re tossing a coin, and we know the probability of Heads
is either 0.4 or 0.8. Over a sequence of many tosses, how will information accumulate
to distinguish the two possibilities? First, if we’re tossing the 0.4 coin, then the likeli-
hood ratio on a single toss is either 0.4/0.8 (if the coin lands Heads) or 0.6/0.2 (if the
coin lands Tails). The two possibilities have probabilities 0.4 and 0.6, respectively. So
I (Bern(.4), Bern(.8)) = 0.4 log(.5) + 0.6 log(3) ≈ .382. But for tossing the 0.8 coin,
I (Bern(.8), Bern(.4)) = 0.8 log(2) + 0.2 log(1/3) ≈ .335. Notice that the divergence
is not symmetric. The calculation shows that information accumulates slightly faster if
we’re tossing the .4 coin than if we’re tossing the .8 coin. To carry the example to an
extreme, suppose one coin is fair but the other is two-headed. For tossing the fair coin,
I (Bern(.5), Bern(1)) = .5 log(.5) + .5 log(∞) = ∞; but I (Bern(1), Bern(.5)) = log(2) ≈
.693. The asymmetry and the infinity make sense: if we’re tossing the fair coin then even-
tually we will toss a Tail and we’ll know for certain which coin it is. But if we’re tossing
the two-headed coin, we’ll always toss Heads and never know for certain which coin it is.
Theorem 8.2. The Kullback-Leibler divergence between two distributions is non-negative.

Proof. We prove the theorem in the continuous case; the discrete case is similar. Our proof
follows Theorem 3.1 in Kullback [1968]. Let f1 and f2 be the two densities. We want
Z ! Z
f1 (x)
I( f1 , f2 ) = log f1 (x) dx = g(x) log g(x) f2 (x) dx
f2 (x)
where g(x) = f1 (x)/ f2 (x). Let φ(t) = t log(t) and use Taylor’s Theorem to get
1
φ(g(x)) = φ(1) + [g(x) − 1]φ0 (1) + [g(x) − 1]2 φ00 (h(x))
2
where for every x, h(x) ∈ [1, g(x)]. Integrate to get
Z
I( f1 , f2 ) = φ(g(x)) f2 (x) dx =
Z Z Z
1
φ(1) f2 (x) dx + [g(x) − 1]φ (1) f2 (x) dx +
0
[g(x) − 1]2 φ00 (h(x)) f2 (x) dx
2Z
1
=0+0+ [g(x) − 1]2 φ00 (h(x)) f2 (x) dx
2
8.2. INFORMATION 430

But g(x) ≥ 0 and h(x) ∈ [1, g(x)] so h(x) ≥ 0 and therefore I( f1 , f2 ) ≥ 0. 

Theorem 8.2 shows that Kullback-Leibler divergences are non-negative. See Exer-
cise 6 for conditions under which they are zero.
While the Kullback-Leibler divergence measures the ability to discriminate between
two arbitrary densities, the usual situation in statistics is that we are working with a para-
metric family of distributions, so we want to look at the ability to discriminate just among
distributions in the family. And when the sample size is large, we can usually focus atten-
tion on a small neighborhood of θ’s. To that end, note first that

d2
" #
d 1 d
log f (x | θ) = f (x | θ)
dθ2 dθ f (x | θ) dθ
!2
1 d 1 d2
=− f (x | θ) + f (x | θ).
f (x | θ)2 dθ f (x | θ) dθ2

Now we want to find I( f (x | θ), f (x | θ + δ)). Expand in a Taylor’s series around f (x | θ) to

get
Z
I( f (x | θ), f (x | θ + δ)) = f (x | θ) log f (x | θ) − log f (x | θ + δ) dx


δ δ2 d2
Z ( " #)
d
≈ f (x | θ) log f (x | θ) − log f (x | θ) + f (x | θ) + log f (x | θ) dx
f (x | θ) dθ 2 dθ2
δ2 d 2
Z Z
d
= −δ f (x | θ) − f (x | θ) log f (x | θ) dx.
dθ 2 dθ2
Assume we can differentiate under the integral so that
Z Z
d d d
f (x | θ) dx = f (x | θ) dx = 1 = 0
dθ dθ dθ
and the first term in the previous expression vanishes to yield

δ2 d2
Z
I( f (x | θ), f (x | θ + δ)) ≈ − f (x | θ) 2 log f (x | θ) dx
2 dθ
 !2 
δ 2 Z  1 d 1 d2 
=− f (x | θ)  θ) + θ)
 
− f (x | f (x | dx
 
2 f (x | θ) dθ
2 f (x | θ) dθ 2

 


#2 #2
δ2 δ2
Z " "
1 d d
= f (x | θ) f (x | θ) dx = E log f (x | θ) dx.
2 f (x | θ) dθ 2 dθ
8.3. EXPONENTIAL FAMILIES 431

Definition 8.7. The expectation on the right in the previous expression,

" #2
d
E log p(x | θ)
dθ

is called the Fisher Information for sampling from the family f (x | θ) and is denoted I(θ).

Fisher Information is the most studied form of information in statistics and is relevant
for inference in parametric families when the sample size is large. The penultimate integral
2
in the previous derivation shows that I(θ) is also equal to −E[ dθd 2 log f (x | θ)]. The word
information is justified because the Fisher Information also tells us the maximum precision
(minimum variance) with which we can estimate parameters, in a limiting, asymptotic
sense to be made precise in Section 8.4.3. Next we calculate the information in some
common parametric families. Others are in the exercises.
Let X ∼ N(µ, σ) where σ is fixed and µ is the unknown parameter.
#2 #2 #2
1 (x − µ)
" " "
d 1 d 1
I(µ) = E log f (x | µ) = E f (x | µ) = E f (x | µ) = 2 .
dµ f (x | µ) dµ f (x | µ) σ2 σ

Let X ∼ Poi(λ).
" #2 " #2
d d
I(λ) = E log f (x | λ) = E (−λ + x log λ − log x!)
dλ dλ
 x 2  x − λ 2 Var(x | λ) 1
= E −1 + =E = = .
λ λ λ2 λ

8.3 Exponential families

Section 8.3 examines a structure that is shared by many common parametric families of
distributions, including the Normal, Gamma, Beta, Binomial, and Poisson. We illustrate
with the Normal and Poisson. If X ∼ Poi(λ) then

e−λ λ x e−λ e x log λ

p(x | λ) = = ,
x! x!
which we can write as h(x)c(λ) exp(w(λ)x) for some functions h, c, and w. If X ∼ N(µ, σ)
then
1 1 1 µ2 x2 µx
e− 2σ2 e− 2σ2 + σ2
1 2 1 2 2
p(x | µ, σ) = √ e− 2σ2 (x−µ) = √ e− 2σ2 (x −2µx+µ ) = √
2πσ 2πσ 2πσ
8.3. EXPONENTIAL FAMILIES 432

which we can write as h(x)c(µ, σ) exp( wi (µ, σ)ti (x)) for some functions h, c, wi , and ti .
P
These two examples reveal the structure we’re looking for.
A parametric family of distributions p(x | θ) is called an exponential family if its pdf’s
can be written as
Pd
p(x | θ) = h(x)c(θ)e i=1 wi (θ)ti (x) .
We study exponential families because they have many features in common. A good
source for an advanced treatment of exponential families is Brown [1986].
The normalizing constant c(θ) can be absorbedP into the exponent, so it is common to
see exponential families written as p(x | θ) = h(x)e wi (θ)ti (x)−c (θ) . We have just seen that the
∗

family of Poisson distributions is an exponential family with d = 1. The family of Normal

distributions is an exponential family with d = 2. See Exercise 10 for other examples of
exponential families. We may rewrite the parameters as ηi = wi (θ). The ηi ’s are called
the natural parameters of the family. For Poisson distributions, the natural parameter is
η = log λ. For Normal distributions the natural parameters are (− 2σ1 2 , σµ2 ).
A function of the form p(x | η) ∝ h(x)e ηi ti (x) can be made a probability density by
P

dividing by its integral, provided that its integral is finite. The set of η’s for which
ηi ti (x)
< ∞ is called the natural parameter space and denoted H. For every η ∈ H,
R P
h(x)e
there is a number c(η) such that p(x | η) = h(x)c(η)e ηi ti (x) is a probability density.
P

Suppose X1 , . . . , Xn is a sample from an exponential family. Then

n
 n 
Y Y  Pd Pn
p(x1 , . . . , xn | θ) = p(x j | θ) =  h(x j ) (c(θ))n e i=1 wi (θ) j=1 ti (x j )
j=1 j=1

which shows that the collection T 1 (x) = j t1 (x j ), . . . , T d (x) = j td (x j ) is a d-dimensional

P P
sufficient statistic. hR i
In derivations — as in Section 8.4 — we sometimes need to study forms like dyd g(x, y) dx .
hR i R 
For arbitrary functions g, dyd g(x, y) dx , dyd g(x, y) dx. But exponential families

obey the following theorem, which we state without proof. The statement comes from
Lehmann [1983].

Theorem 8.3. For any integrable function g and any η in the interior of H, the integral
Z
ηi ti (x)
P
g(x)h(x)c(η)e dx

is continuous and has derivatives of all orders with respect to the η’s, and these can be
obtained by differentiating under the integral sign.
8.4. ASYMPTOTICS 433

For example, take g(x) = 1. Write a one-parameter exponential family in the form
p(x | η) = h(x)eηt(x)−c (η) . The integral of the density is, as always, 1, so taking derivatives
∗

yields
Z Z Z
d ηt(x)−c∗ (η) ηt(x)−c∗ (η)
dx − c∗ h(x)eηt(x)−c (η) dx,
0 ∗
0= h(x)e dx = t(x)h(x)e (8.1)
dη

or E[t(x)] = c∗ .
0

It is sometimes useful and natural to consider the random variable T = t(X). We have
just derived its expectation.

8.4 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.

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 con-
.
tinues 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 ) )
}
8.4. ASYMPTOTICS 434

• 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 to two decimal places.
Because there are multiple sequences, each with multiple elements, we need two subscripts
to keep track of things properly. Let Yi j be the j’th element of the i’th sequence. A real
data set X1 , . . . , Xn is analogous to the first n observations, {Yi, j }nj=1 , along one row of the
array. For each sequence of random variables (each row of the array), we’re interested in
things like the behavior as n → ∞ of the sequence of sample means Ȳi1 , Ȳi2 , . . . where
Ȳin = (Yi1 + · · · + Yin )/n. And for √the Central Limit Theorem, 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 · · ·
.. .. .. .. .. .. .. .. .. . .
. . . . . . . . . .
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 · · ·
.. .. .. .. .. .. .. .. .. . .
. . . . . . . . . .
8.4. ASYMPTOTICS 435

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, the collection of random variables {Ȳi,n }∞

i=1 are a set of i.i.d. draws from
some distribution. What is it? And what is the analogous distribution for {Zi,n }∞ i=1 ?
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 probability of randomly getting
a sequence like 1, 2, 3, . . . that doesn’t converge 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 like 1, 1, 1, . . . or −1, −1, −1, . . . that con-
verges to something other than µ is 0.
8.4. ASYMPTOTICS 436

4. What is the distribution of Zn ? We cannot say in general. It depends on the distribu-

tion of the individual Yi ’s.

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 j ’s, as long as Var(Yi j ) < ∞, 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 j ’s”. That’s what we look at next.
To illustrate, we generate sequences of Yi j ’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 8.2, 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 8.3. 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 Nor-
mal 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 8.4 is for the Zin ’s. It’s the same as Figure 8.3 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 8.2 was produced by

8.4. ASYMPTOTICS 437

8
6
4
2
0

0.0 0.2 0.4 0.6 0.8 1.0

Figure 8.2: The Be(.39, .01) density

x <- seq ( .01, .99, length=80 )

plot ( x, dbeta(x,.39,.01), type="l", ylab="", xlab="" )

Figure 8.3 was generated by the following R code.

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) )
8.4. ASYMPTOTICS 438

n = 10 n = 30

30

30
25

25
20

20
15

15
10

10
5

5
0

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 8.3: Densities of Ȳin for the U(0, 1) (dashed), modified Be(.39, .01) (dash and dot),
and Normal (dotted) distributions.
8.4. ASYMPTOTICS 439

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.

8.4.1 Modes of Convergence

There are at least three different meanings for convergence of a sequence of random vari-
ables. Section 8.4.1 describes convergence in distribution, convergence in probability, and
convergence almost surely.

Definition 8.8 (Convergence in Distribution). Let Y be a random variable and Y1 , Y2 , . . . be

a sequence of random variables. The sequence Y1 , Y2 , . . . is said to converge in distribution
to Y if, at every point y where FY is continuous,

lim FYn (y) = FY (y).

n→∞

Definition 8.9 (Convergence in Probability). The sequence Y1 , Y2 , . . . is said to converge

in probability to Y if, for every  > 0,

lim P[|Yn − Y| < ] = 1.

n→∞
8.4. ASYMPTOTICS 440

n = 10 n = 30

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

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 8.4: Densities of Zin for the U(0, 1) (dashed), modified Be(.39, .01) (dash and dot),
and Normal (dotted) distributions.
8.4. ASYMPTOTICS 441

Definition 8.10 (Convergence almost surely). The sequence Y1 , Y2 , . . . is said to converge

almost surely to Y if, for every  > 0,

P[ lim |Yn − Y| < ] = 1.

n→∞

Almost sure (a.s.) convergence is also called convergence almost everywhere (a.e.) and
convergence with probability 1 (w.p.1).

The three modes of convergence are related by the following Theorems, stated here
without proof.

Theorem 8.4. If Yn → Y almost surely, then Yn → Y in probability.

Theorem 8.5. If Yn → Y in probability, then Yn → Y in distribution.

One result we will need later is that if g is a function, continuous at c, and if Y1 , Y2 , . . .

converges to c in probability, then g(Y1 ), g(Y2 ), . . . converges to g(c) in probability. The
proof is left as an exercise.
The converses to Theorems 8.4 and 8.5 are not true, as the following examples demon-
strate.

Convergence in Distribution

1. Toss a fair penny. For all i = 0, 1, . . . , let Xi = 1 if the penny lands heads and Xi = 0
if the penny lands tails. (All Xi ’s are equal to each other.) Toss a fair dime. Let
Y = 1 if the dime lands heads and Y = 0 if the dime lands tails. Then the sequence
X1 , X2 , . . . converges to X0 in distribution, in probability and almost surely. The
cdf’s of X1 , X2 , . . . and Y are all the same, so the Xi ’s converge to Y in distribution.
But the Xi ’s do not converge to Y in probability or almost surely.

2. See the Exercises.

Convergence in Probability See Exercise 21 and others.

Convergence Almost Surely See Exercise 20 and others.

One reason that convergence in distribution is sometimes useful is the following theo-
rem.

Theorem 8.6. The sequence Xn → X in distribution if and only if E[g(Xn )] → E[g(X)] for
every bounded, continuous, real-valued function g.
8.4. ASYMPTOTICS 442

The proof is beyond the scope of this book, but may be found in advanced texts on
probability.
Sometimes we know that a sequence Xn → X in distribution. But we may be interested
in Y = g(X) for some know function g. Does g(Xn ) → g(X)? Theorems 8.7 and 8.8
provide answers in some important special cases.
Theorem 8.7 (Slutsky). Let Xn → X in distribution and Yn → c in probability, where
|c| < ∞. Then,
1. Xn + Yn → X + c in distribution;

2. Xn Yn → cX in distribution; and

3. Xn /Yn → X/c in distribution, if c , 0.

Proof. This proof follows Serfling [1980]. Choose t such that F X is continuous at t − c
and let  > 0 be such that F X is continuous at t − c −  and t − c + . (Such a t and  can
always be found if F X is continuous except at finitely many points.) Then

F Xn +Yn (t) = Pr[Xn + Yn ] ≤ t]

= Pr[Xn + Yn ≤ t; |Yn − c| ≤ ] + Pr[Xn + Yn ≤ t; |Yn − c| > ]
≤ Pr[Xn + Yn ≤ t; |Yn − c| ≤ ] + Pr[|Yn − c| > ]
≤ Pr[Xn ≤ t − c + ] + Pr[|Yn − c| > ].
So

lim sup F Xn +Yn (t) ≤ lim sup Pr[Xn ≤ t − c + ] + Pr[|Yn − c| > ]

≤ lim sup Pr[Xn ≤ t − c + ] = F X (t − c + ).

Letting  ↓ 0 gives
lim sup F Xn +Yn (t) ≤ F X (t − c) = F X+c (t). (8.2)
Similarly,

Pr[Xn ≤ t − c − ] = Pr[Xn ≤ t − c − ; |Yn − c| < ] + Pr[Xn ≤ t − c − ; |Yn − c| ≥ ]

≤ Pr[Xn ≤ t − c − ; |Yn − c| < ] + Pr[|Yn − c| ≥ ]
= Pr[Xn + Yn ≤ t − c −  + Yn ; |Yn − c| < ] + Pr[|Yn − c| ≥ ]
≤ Pr[Xn + Yn ≤ t − c −  + (c + ); |Yn − c| < ] + Pr[|Yn − c| ≥ ]
= Pr[Xn + Yn ≤ t; |Yn − c| < ] + Pr[|Yn − c| ≥ ]
≤ Pr[Xn + Yn ≤ t] + Pr[|Yn − c| ≥ ].
8.4. ASYMPTOTICS 443

So,
lim inf Pr[Xn + Yn ≤ t] + Pr[|Yn − c| ≥ ] ≥ lim inf Pr[Xn ≤ t − c − ] = F X (t − c − ).
Letting  ↓ 0 gives
lim inf F Xn +Yn (t) ≥ F X (t − c) = F X+c (t). (8.3)
Combining 8.2 and 8.3 yields
lim F Xn +Yn (t) = F X+c (t),
showing that Xn + Yn → X + c in distribution.
We leave the proofs of parts 2 and 3 as exercises. 
One application of Slutsky’s Theorem is the following, which comes from Casella
and Berger [2002]. Let X1 , X2 , . . . be a sample from a distribution whose mean µ and SD
σ are finite. The Central Limit Theorem says
√
n(X̄n − µ)
→ N(0, 1)
σ
in distribution. But in almost all practical problems we don’t know σ. However, we can
estimate σ by either the m.l.e. σ̂ or the unbiased estimator σ̃. Both of them are consistent,
meaning they converge to σ in probability. Thus,
√ √
n(X̄n − µ) σ n(X̄n − µ)
= → N(0, 1) in distribution
σ̂ σ̂ σ
by Slutsky’s Theorem.

8.4.2 The δ-method

√
For inferences about µ, the Central Limit Theorem tells us that n(X̄n − µ)/σ → N(0, 1)
in distribution. But sometimes we’re interested in a function of µ, say g(µ). That’s where
the δ-method comes in. It translates the Central Limit Theorem into a result about g(Xn )
and g(µ). Before stating and proving the δ-method, let’s try to get some intuition.
Suppose X ∼ N(µ, σ). Figure 8.5 shows p(x), a smooth function g(x), and the tangent
line at x = µ. Now consider a sequence of random variables Xn ∼ N(µ, σ/n). They
would have pictures similar to 8.5, but the densities would become more concentrated
around µ. As the densities become more concentrated, the relevant portion of g would
become more concentrated around the point of tangency and g(X) would be more like the
transformation represented by the tangent line `(X) = g(µ) + g0 (µ)(X − µ). In the limit,
g(X) ∼ N(g(µ), σg0 (µ)). The δ-method formalizes this intuition.
8.4. ASYMPTOTICS 444

tangent line

g(x) p(x)

Figure 8.5: The δ-method

8.4. ASYMPTOTICS 445
√
Theorem 8.8 (The δ-method). Let n(Xn − µ)/σ → N(0, 1) in distribution. Let g be a
function with a continuous derivative such that g0 (µ) is finite and not zero. Then
√ g(Xn ) − g(µ)
n → N(0, 1).
σg0 (µ)

Proof. Expand g in a Taylor’s series about µ:

g(Xn ) = g(µ) + g0 (µ̃)(Xn − µ)

for some µ̃ ∈ [µ, Xn ] and therefore

√ g(Xn ) − g(µ) √ g0 (µ̃)(Xn − µ)
n = n → N(0, g0 (µ))
σ σ
by Slutsky’s theorem and because g0 (µ̃) → g0 (µ). Now use Slutsky’s theorem again to get
√ g(Xn ) − g(µ)
n → N(0, 1).
σg0 (µ)


To illustrate the δ-method, suppose we’re testing a new medical treatment to learn its
probability of success, θ. Subjects, n of them, are recruited into a trial and we observe
Xn , the number of successes. √ We could√ use the m.l.e. θ̂n = Xn /n as an estimator of θ.
And we already know that n(θ̂n − θ)/ θ(1 − θ) → N(0, 1). But researchers sometimes
want to phrase their results in terms of the odds of success, θ/(1 − θ), or the log odds,
log(θ/(1 − θ)). We could use θ̂n /(1 − θ̂n ) to estimate the odds of success, but how would
we know its accuracy; what do we know about its distribution? The δ-method gives the
answer, at least approximately for large n.
Let g(x) = x/(1 − x). Then g0 (θ) = (1 − θ)−2 . Therefore,
θ̂n θ
 
√
"
g(θ̂n ) − g(θ)
#
√  1−θ̂n
− 1−θ

n √ = n  √  → N(0, 1),
θ(1 − θ)g (θ)
0 θ(1 − θ)(1 − θ) 
−2 

or equivalently,
θ̂n θ θ1/2
" # !
√
n − → N 0, .
1 − θ̂n 1 − θ (1 − θ)3/2
8.4. ASYMPTOTICS 446

8.4.3 The Asymptotic Behavior of Estimators

Suppose that X1 , X2 , . . . is an i.i.d. sequence of draws from a distribution p(x | θ) and we
are trying to estimate θ. From X1 , . . . , Xn we construct an estimate δn . (We use the notation
δn rather than θ̂n so as not to imply that δn is the m.l.e.) In most cases, we would choose a
consistent estimator:
√ δn → θ in probability. Typically, the Central Limit Theorem would
also apply: n(δn − θn ) → N(0, σ) in distribution.
Since the sequence δn is consistent, then for any d ≥ 0, Pr[|δn − θ| > d] → 0. In √the
Central Limit Theorem, we multiply √ the sequences of differences by the sequence n,
which magnifies it just enough so n(δn − θn ) doesn’t go to 0, doesn’t go to infinity, but
goes to a distribution, in this case N(0, σ). But more general behavior is possible. Suppose
there is a sequence of constants, c1 , . . . and a cdf G such that for all x,

Pr[cn (δn − θ) ≤ x] → G(x).

The sequence of constants tells us the rate of convergence. We say the errors (δn − θ)
converge to zero at the rate 1/cn . For example, in cases
√ where the Central Limit Theorem
applies, (δn − θ) →√0 in distribution at the rate of 1/ n. The δ-method is another example.
When it applies,
√ n(g(δn ) − g(θ)) → N(0, σg0 (θ)), so the errors (g(δn ) − g(θ)) → 0 at the
rate of 1/ n.
When there are competing estimators, say δn and δ0n , we might want to choose the one
whose errors converge√to zero at a faster rate. But the √ situation can be more subtle than
that. It might be that n(δn − θ) → N(0, σ) while n(δn − θ) → N(0, σ0 ). In that case,
0

the two estimators converge at the same rate and we might want to know which has the
smaller asymptotic SD.
This section of the book is about rates of convergence and comparison of asymptotic
SD’s. It will guide us in choosing estimators. Of course, with any finite data set, rates
do not tell the whole story and can be only a guide. A more advanced treatment of these
topics can be found in many texts on mathematical statistics. Lehmann [1983] is a good
example.
First, we have to see whether the concept of rate is well-defined. The following theo-
rem gives that assurance under some mild conditions.

Theorem 8.9. Suppose that for two sequences of constants, cn and c0n , cn (δn − θ) → G
and c0n (δn − θ) → G0 in distribution. Then, under conditions given in Lehmann [1983]
(pg. 347), there exists a constant k such that (a) c0n /cn → k and (b) G0 (x) = G(x/k) for all
x.

The proof is beyond our scope but can be found in Lehmann. Theorem 8.9 says that if
two sequences of constants lead to convergence then (a) the sequences themselves differ
8.4. ASYMPTOTICS 447

asymptotically by only a scale factor and (b), the corresponding limits differ by only that
same scale factor.
Now let’s √ look at an illustration of two estimators δn and δ0n that both converge at
the rate of 1/ n. The illustration comes from Example 2.1 in Chapter 5 of Lehmann.
Suppose X1 , X2 , . . . are a sample from an unknown distribution F. We want to estimate
θ = F(a) = Pr[Xi ≤ a]. If we think that F is approximately N(µ, σ), and if we know or can
estimate σ, then we might takeR x X̄1n as 1an2 estimator of µ and δn = Φ((a− X̄n )/σ), where Φ(x)
is the N(0, 1) cdf: Φ(x) = −∞ √2π e− 2 u du. But if we don’t know that F is approximately
Normal, then we might instead use the estimator δ0n = n−1 (# of Xi ≤ a).
If F is not Normal, then δn is likely to be a bad estimator; it won’t even be consistent.
But what if F is Normal; how can we compare δn to δ0n ? Since both are consistent, we
compare their asymptotic standard deviations.
√ 
First, the Central Limit Theorem says n X̄nσ−µ → N(0, 1). Then, since δn is a known
transformation of X̄n , we can get its asymptotic distribution by Theorem 8.7:
 a−X̄   a−µ 
√ Φ σn −Φ σ
n   → N(0, 1)
σφ a−µσ
1
σ

where φ = Φ0 is the standard Normal pdf. Equivalently,

√   a − µ 
n (δn − θ) → N 0, φ . (8.4)
σ
On the other hand, δ0n is a transformed version of a Binomial random variable. Let Yn =
(# of Xi ≤ a). Then Yn ∼ Bin(n, θ). The Central Limit Theorem says

√ Yn
−θ
n√n → N(0, 1)
θ(1 − θ)
or, equivalently,
√ 0  p 
n(δn − θ) → N 0, θ(1 − θ) . (8.5)
Since there
 a−µ  is a 1-to-1 correspondence between values of θ and values of a−µ σ
, we can
√
plot φ σ against θ(1 − θ). Figure 8.6 shows the plot. The top panel, a plot of the
asymptotic SD of δn as a function of the asymptotic SD of δ0n , shows that δn has the smaller
SD. The bottom panel, a plot of the ratio of the SD’s, shows that the advantage of δn over
δ0n grows as θ goes to either 0 or 1. Therefore, we would normally prefer δn , especially for
extreme values of a, at least when F is Normal and the sample size is large enough that
the asymptotics are reliable.
8.4. ASYMPTOTICS 448

0.4
0.3
)

actual SD's
a−µ

0.2
σ

equal SD's
φ(

0.1
0.0

0.1 0.2 0.3 0.4 0.5

θ(1 − θ)
0.7
SD ratio

0.5

ratio of SD's
0.3
0.1

0.1 0.2 0.3 0.4 0.5

θ(1 − θ)

Figure 8.6: Top panel: asymptotic standard deviations of δn and δ0n for Pr[X ≤ a]. The
solid line shows the actual relationship. The dotted line is the line of equality. Bottom
panel: the ratio of asymptotic standard deviations.
8.4. ASYMPTOTICS 449

When there are competing estimators, there may be one or more that have faster rate,
or smaller asymptotic SD, than the others. If so, those estimators would be the best. Our
next task is to show that under some mild regularity conditions there is, in fact, a lower
bound on the asymptotic SD, find what that bound is, and show that the sequence θ̂n of
m.l.e.’s achieves that lower bound. Thus the sequence of m.l.e.’s is, in this sense, a good
sequence of estimators. There may be other estimators that achieve the same bound, but
no sequence can do better. The main theorem is the following, which we state without
proof.
Theorem 8.10. Under suitable regularity conditions concerning continuity, differentia-
bility, support of the parametric family, and interchanging the order of differentiation
√ integration (See Lehmann, pg. 406.), if δn is a sequence
and of estimators satisfying
n(δn − θ) → N(0, SD(θ)) in distribution, then SD(θ) ≥ I(θ) , except possibly for a
−1/2

set of θ’s of Lebesgue measure 0.

Theorem 8.10 shows the sense in which Fisher Information quantifies the ability to
estimate a parameter, at least in large samples. Theorem 8.10 provides only a bound.
It is quite typical that there are sequences of estimators that achieve the bound. Such
estimators are called asymptotically efficient. Our next task is to show that under suitable
conditions, the sequence θ̂n of m.l.e.’s is asymptotically efficient. The material here follows
the development in Chapter 6, Section 2 of Lehmann.
Suppose that X1 , . . . is an i.i.d. sequence from f (x | θ0 ) and that f (x | θ0 ) is a member of
a parametric family f (x | θ) that satisfies suitable regularity conditions. (We will not men-
tion the conditions further. A precise statement can be found in Lehmann and elsewhere.)
Theorem 8.11. For any θ , θ0 ,
 n n

Y Y
f (x | θ0 ) > f (x | θ) → 1 as n → ∞.

Pr 
i=1 i=1

Proof.
n n n
Y Y f (x | θ0 ) X
f (x | θ0 ) > f (x | θ) ⇐⇒ n −1
> 0.log
i=1 i=1 i=1
f (x | θ)
f (x | θ0 )
h i
The Law of Large Numbers says that the right-hand side goes to E log f (x | θ)
, which
Theorem 8.2 and the exercises say is positive. 
Theorem 8.11 says that for any fixed θ, `(θ0 ) > `(θ) with high probability. It does
not say that argmax `(θ) = θ0 with high probability. In fact, argmax `(θ) ≡ θ̂ is a random
variable with a continuous distribution, so Pr[argmax `(θ) = θ0 ] = 0.
8.4. ASYMPTOTICS 450

Theorem 8.12. The sequence of maximum likelihood estimators, θ̂n , is consistent; i.e.
θ̂n → θ0 in probability.
Proof. Choose  > 0, let ~xn = x1 , . . . , xn , and let

S n = ~xn : p(~xn | θ0 ) ≥ p(~xn | θ0 − ) and p(~xn | θ0 ) ≥ p(~xn | θ0 + ) .



For any ~x ∈ S n , the likelihood function p(~xn | θ) has a local maximum in [θ0 − , θ0 + ].
And by Theorem 8.11, Pr[S n ] → 1. Therefore Pr[|θ̂n − θ0 | < ] → 1. 
Theorem 8.13. If there exists a number c and a function M(x) such that

d3
log p(x | θ) ≤ M(x)
dθ3
for all x and for all θ ∈ (θ0 −c, θ0 +c), then the sequence of maximum likelihood estimators,
θ̂n , is asymptotically efficient; i.e.
!
√   1
n θ̂n − θ → N 0, √ in distribution.
I(θ0 )

Proof. Let g(θ) be the derivative of the log-likelihood function:

d X p0 (xi | θ)
g(θ) = log p(~xn | θ) = .
dθ p(xi | θ)

We want to examine g(θ̂n ); we know g(θ̂n ) is close to g(θ̂0 ), so expand in a Taylor’s series.
1
g(θ̂n ) = g(θ0 ) + (θ̂n − θ0 )g0 (θ0 ) + (θ̂n − θ0 )2 g(2) (θ∗ )
2
for some θ∗ ∈ [θ̂n , θ0 ]. By assumption, the left-hand side is zero. Rearrange terms to get
g(θ0 )
(θ̂n − θ0 ) = −
g0 (θ0 ) + 12 (θ̂n − θ0 )g(2) (θ∗ )
or
√ √1 g(θ0 )
n
n(θ̂n − θ0 ) =
− 1n g0 (θ0 ) − 1
(θ̂
2n n
− θ0 )g(2) (θ∗ )
Now we’re going to analyze the numerator and the two terms in the denominator sepa-
rately, then use Slutsky’s Theorem (Thm 8.7).
8.5. EXERCISES 451

1. g(θ0 ) is the sum of terms like p0 (xi | θ)/p(xi | θ). Because xi is a random variable,
each term is a random variable. Because the xi ’s are i.i.d., the terms are i.i.d. The
expected value of each term is

p (xi | θ)
" 0 # Z Z
d d
E = p(xi | θ) dθ = p(xi | θ) dθ = 0.
p(xi | θ) dθ dθ
√
The Central Limit Theorem applies so, by the definition of I(θ), √1 g(θ0 ) → N(0, I(θ))
n
in probability.

2. g0 (θ0 ) is also the sum of i.i.d.summands. Specifically,

1 0 1 X p(2) (xi | θ) [p0 (xi | θ)]2

g (θ) = − .
n n p(xi | θ)2 p(xi | θ)2

The expectation of the first term is 0; the expectation of the second is I(θ). Therefore
1 0
n
g (θ) → −I(θ) by the Law of Large Numbers.

3. By assumption of the theorem, g(2) (θ∗ ) ≤ M(x) for all θ’s in an interval around θ0 .
And θ∗ → θ0 in probability, so 2n1 (θ̂n − θ0 )g(2) (θ∗ ) → 0 in probability.

Now Slutsky’s Theorem gives the desired result. 

The condition of Theorem 8.13 may seem awkward, but it is often easy to check. The
following corollary is an example.

Corollary 8.14. If {p(x | θ)} is a one-dimensional exponential family, then θ̂n is asymptoti-
cally efficient.

Proof. If {p(x | θ)} is a one-dimensional exponential family, then

d3 d3
log p(x | θ) = log h(x) + log c(θ) + w(θ)t(x)
dθ3 dθ3

which satisfies the conditions of the theorem, at least when c, w, and t are continuous
functions. 

8.5 Exercises
1. Let Y1 , . . . , Yn be a sample from N(µ, σ).
8.5. EXERCISES 452

(a) Suppose µ is unknown but σ is known. Find a one dimensional sufficient

statistic for µ.
(b) Suppose µ is known but σ is unknown. Find a one dimensional sufficient
statistic for σ.
(c) Suppose µ and σ are both unknown. Find a two dimensional sufficient statistic
for (µ, σ).

2. Let Y1 , . . . , Yn be a sample from Be(α, β). Find a two dimensional sufficient statistic
for (α, β).

3. Let Y1 , . . . , Yn ∼ i.i.d. U(−θ, θ). Find a low dimensional sufficient statistic for θ.

4. Let Y1 , . . . Yn ∼ i.i.d. U(0, θ).

(a) Find the m.l.e. θ̂.

(b) Find the distribution of θ̂.
(c) Find E[θ̂].
(d) Find Var[θ̂].
(e) Find MSE(θ̂).
(f) Because E[Ȳ] = θ/2, we might consider whether θ̃ ≡ 2Ȳ as an estimator of θ.
i. Is θ̃ a function of a sufficient statistic?
ii. Is θ̃ unbiased?
iii. Find MSE(θ̃). Compare to MSE(θ̂).

5. Refer to Exercise 26 in Chapter 5. When ecologists sample the nearest pitcher plants
to, say, M points, the total number of plants found is some number M ∗ , which is less
than or equal to M. The total region they search, say A, is the union of the circles of
radii Di , centered at the si .

(a) Show that (M ∗ , |A|) is a sufficient statistic for λ. (|A| is the area of A.)
(b) If any of the circles intersect, then |A| is difficult to find analytically. But |A|
can be estimated numerically by the following method.
i. Construct a rectangle B that contains A.
ii. Generate N (a large integer) points at random, uniformly, in B.
iii. Let X be the number of generated points that lie within A.
iv. Use X/N as an estimate of |A|/B and |B|X/N as an estimate of |A|.
8.5. EXERCISES 453

The user can choose B, as long as it’s big enough to contain A. If the goal is to
estimate |A| as accurately as possible, what advice would you give the user for
choosing the size of B? Should |B| be large, small, or somewhere in between?
Justify your answer. Hint: think about Binomial distributions.

6. Theorem 8.2 shows that Kullback-Leibler divergences are non-negative. This exer-
cise investigates conditions under which they are zero.

(a) Let F be a distribution. Find the Kullback-Leibler divergence from F to itself.

(b) Look at the proof of Theorem 8.2 to see what conditions are necessary for
I( f1 , f2 ) to be zero. Prove that if f1 and f2 differ on any interval, then I( f1 , f2 ) >
0.

7. (a) Find the Kullback-Leibler divergence from Bern(p1 ) to Bern(p2 ) and from
Bern(p2 ) to Bern(p1 ).
(b) Find the Kullback-Leibler divergence from Bin(n, p1 ) to Bin(n, p2 ) and from
Bin(n, p2 ) to Bin(n, p1 ).

8. (a) Let X ∼ N(µ, σ) where µ is fixed. Find I(σ).

(b) Let X ∼ Bin(n, θ). Find I(θ).

9. (a) Let X ∼ Poi(λ). We know I(λ) = λ−1 . But we may be interested in λ∗ ≡ log λ.
Find I(λ∗ ).
(b) Let X ∼ f (x | θ). Let φ = h(θ). Show I(φ) = ( dφ
dθ 2
) I(θ).

10. Show that the following are exponential families of distributions. In each case,
identify the functions h, c, wi , and ti and find the natural parameters.

(a) Bin(n, θ) where n is known and θ is the parameter.

(b) Gam(α, β).
(c) Be(α, β).

11. Verify that Equation 8.1 gives the correct value for the means of the following dis-
tributions.

(a) Poi(λ).
(b) Exp(θ).
(c) Bin(n, θ).
8.5. EXERCISES 454

12. Differentiate Equation 8.1 to show Var(t(x)) = c∗(2) .

13. Derive the two-parameter version of Equation 8.1.

14. In a one-parameter exponential family, it is sometimes natural and useful to consider

the random variable T = t(x). Equation 8.1 gives E[T ].

(a) Use the method of transformations to find p(t | η). Show that it is an exponential
family.
(b) Find the moment generating function MT (s) of T .

15. Prove that if g is a function continuous at a number c, and if {Yn } → c in probability,

then {g(Yn )} → c in probability.

16. Prove the claims in item 1 on page 441 that Xn → X0 in distribution, in probability,
and almost surely, but Xn → Y in distribution only.
√
17. Let Xn ∼ N(0, 1/ n). Does the sequence {Xn }∞ n=1 converge? Explain why or why
not. If yes, also explain in what sense it converges — distribution, probability or
almost sure — and find its limit.

18. Let X1 , X2 , · · · ∼ i.i.d. N(µ, σ) and let X̄n = n−1 ni=1 Xi . Does the sequence {X̄n }∞
P
n=1
converge? In what sense? To what limit? Justify your answer.

19. Let X1 , X2 , . . . be√an i.i.d. random sample from a distribution F with mean µ and SD
σ and let Zn = n(X̄n − µ)/σ. A well-known theorem says that {Zn }∞ n=1 converges
in distribution to a well-known distribution. What is the theorem and what is the
distribution?

20. Let U ∼ U(0, 1). Now define the sequence of random variables X1 , . . . in terms of
U by 
1 if U ≤ n−1

Xn = 

0 otherwise.


(a) What is the distribution of Xn ?

(b) Find the limit, in distribution, of the Xn ’s.
(c) Show that the Xn ’s converge to that limit in probability.
(d) Show that the Xn ’s converge to that limit almost surely.
(e) Find the sequence of numbers EX1 , EX2 , . . . .
8.5. EXERCISES 455

(f) Does the sequence EXn converge to EX?

21. This exercise is similar to Exercise 20 but with a subtle difference. Let U ∼ U(0, 1).
Now define the sequence of constants c0 = 0, c1 = 1 and, in general, cn = cn−1 + 1/n.
In defining the ci ’s, addition is carried out modulo 1; so c2 = (1+1/2) mod 1 = 1/2,
etc. Now define the sequence of random variables X1 , . . . in terms of U by

1 if Xn ∈ [cn−1 , cn ]

Xn = 

0 otherwise.


where intervals are understood to wrap around the unit interval. For example,
[c3 , c4 ] = [5/6, 13/12] = [5/6, 1/12] is understood to be the union [5/6, 1]∪[0, 1/12].
(It may help to draw a picture.)

(a) What is the distribution of Xn ?

(b) Find the limit, in distribution, of the Xn ’s.
(c) Find the limit, in probability, of the Xn ’s.
(d) Show that the Xn ’s do not converge to that limit almost surely.

22. (a) Prove part 2 of Slutsky’s theorem (8.7).

(b) Prove part 3 of Slutsky’s theorem (8.7).

23. Let Xn ∼ Bin(n, θ) and let θ̂n = Xn /n. Use the δ-method to find the asymptotic
distribution of the log-odds, log(θ/(1 − θ)).
  √
24. In Figure 8.6, show that the ratio of asymptotic SD’s, φ a−µ
σ
/ θ(1 − θ), goes to
infinity as θ goes to 0 and also as θ goes to 1.

√ Theorem 8.10, show that if ηn = h(δn ) is a sequence

25. Starting from
0
√ of estimators
satisfying n(ηn − h(θ)) → N(0, SD(h(θ))), then SD(h(θ)) ≥ h (θ)/ I(θ).

26. Page 447 compares the asymptotic variances of two estimators, δn and δ0n , when the
underlying distribution F is Normal. Why is Normality needed?
 Bibliography

Consumer Reports, June:366–367, 1986.

T. W. Anderson. An Introduction to Multivariate Statistical Analysis. John Wiley, New

York, 2nd edition, 1984.

D. F. Andrews and A. M. Herzberg. Data. Springer-Verlag, New York, 1985.

H. Bateman. On the probability distribution of α particles. Philosophical Magazine Series

6, 20:704–705, 1910.

Richard J. Bolton and David J. Hand. Statistical fraud detection: A review. Statistical
Science, 17:235–255, 1992.

Paul Brodeur. Annals of radiation, the cancer at Slater school. The New Yorker, Dec. 7,
1992.

Lawrence D. Brown. Fundamentals of statistical exponential families with applications

in statistical decision theory. Institute of Mathematical Statistics Lecture Notes—
Monograph Series, 9. Institute of Mathematical Statistics, Hayward, CA, 1986. ISBN
0-940600-10-2.

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.

D. P. Byar. The veterans administration study of chemoprophylaxis for recurrent stage

I bladder tumors: Comparisons of placebo, pyridoxine, and topical thiotepa. In
M. Pavone-Macaluso, P. H. Smith, and F. Edsmyn, editors, Bladder Tumors and Other
Topics in Urological Oncology, pages 363–370. Plenum, New York, 1980.

George Casella and Roger L. Berger. Statistical Inference. Duxbury, Pacific Grove, second
edition, 2002.

456
BIBLIOGRAPHY 457

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,
4th edition, 1998.

Andrew Gelman, John B. Carlin, Hal S. Stern, and Donald B. Rubin. Bayesian Data
Analysis. Chapman and Hall, Boca Raton, 2nd edition, 2004.

S. Geman and D. Geman. Stochastic relaxation, Gibbs distributions, and the Bayesian
restoration of images. IEEE Transactions on Pattern Analysis and Machine Intelligence,
6:721–741, 1984.

John Hassall. The Old Nursery Stories and Rhymes. Blackie and Son Limited, London,
1909.

W. K. Hastings. Monte Carlo sampling methods using Markov chains and their applica-
tions. Biometrika, 57:97–109, 1970.

S. Kullback and R. A. Leibler. On information and sufficiency. Annals of Mathematical

Statistics, 22:79–86, 1951.

Solomon Kullback. Information Theory and Statistics. Dover Publications, Inc., 1968.

Shannon LaDeau and James Clark. Rising co2 levels and the fecundity of forest trees.
Science, 292(5514):95–98, 2001.

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. L. Lehmann. Theory of Point Estimation. John Wiley, New York, 1983.

Jun S. Liu. Monte Carlo Strategies in Scientific Computing. Springer-Verlag, New York,
2004.

Jean-Michel Marin and Christian P. Robert. Bayesian Core: A Practical Approach to

Computational Bayesian Statistics. Springer-Verlag, New York, 2007.
BIBLIOGRAPHY 458

N. Metropolis, A. W. Rosenbluth, M. N. Rosenbluth, A. H. Teller, and E. Teller. Equation

of state calculations by fast computing machines. Journal of Chemical Physics, 21:
1087–1092, 1953.

R Development Core Team. R: A Language and Environment for Statistical Computing.

R Foundation for Statistical Computing, Vienna, Austria, 2006. URL http://www.
R-project.org. ISBN 3-900051-07-0.

Christian P. Robert and George Casella. Monte Carlo Statistical Methods. Springer-Verlag,
New York, 1997.

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.

Robert J. Serfling. Approximation Theorems of Mathematical Statistics. John Wiley, New

York, 1980.

T.S. Tsou and R.M. Royall. Robust likelihoods. Journal of the American Statistical Asso-
ciation, 90:316–320, 1995.

Jessica Utts. Replication and meta-analysis in parapsychology. Statistical Science, 4:

363–403, 1991.

W. N. Venables and B. D. Ripley. Modern Applied Statistics with S. Springer, New York,
fourth edition, 2002.

L. J. Wei, D. Y. Lin, and L. Weissfeld. Regression analysis of multivariate incomplete

failure time data by modeling marginal distributions. Journal of the American Statistical
Association, 84:1065–1073, 1989.

Sanford Weisberg. Applied Linear Regression. John Wiley & Sons, New York, second
edition, 1985.

A.S. Yang. Seasonality, division of labor, and dynamics of colony-level nutrient storage
in the ant Pheidole morrisi. Insectes Sociaux, 53:456–452, 2006.
 Index

α particle, 293 DASL, see Data and Story Library, see

Data and Story Library, 203
autocorrelation, 393 Data and Story Library, 105, 141
autoregression, 399 density
probability, 263
bandwidth, 107 density estimation, 106
bias, 419 dependence, 53
distribution, 2
case, 215 Distributions
cdf, see cumulative distribution function Bernoulli, 278
Central Limit Theorem, 79 Beta, 310
change of variables, 12 Binomial, 14, 277
characteristic functions, 273 Cauchy, 333
Chebychev’s Inequality, 78 Exponential, 20
chi-squared distribution, 310 Gamma, 303
consistency, 418 inverse Gamma, 341
Convergence Multinomial, 287
almost surely, 434 Negative binomial, 282
in distribution, 434 Normal, 22, 313
in probability, 434 Poisson, 17, 289
coplots, 127 standard multivariate Normal, 318
correlation, 52 standard Normal, 29
covariance, 50 Uniform, 302
covariance matrix, 265
covariate, 215 errors, 218
cross tabulation, 117 estimate, 153
cumulative distribution function, 269 expected value, 32
cumulative hazard function, 410 explanatory variable, 215

459
INDEX 460

fitted values, 223, 247 Mediterranean tongue, 30

fitting, 223 mgf, see moment generating function
floor, 419 minimal sufficient, 418
formula, 224 moment, 39
moment generating function, 272
gamma function, 303 mosaic plot, 117
Gaussian density, 313 multinomial coefficient, 288
generalized moment, 39 multivariate
genotype, 287 change of variables, 266
half-life, 307 order statistic, 97, 418
histogram, 105 outer product, 322
independence, 53 parameter, 14, 133, 277
joint, 264 parametric family, 14, 133, 277
mutual, 264 partial autocorrelation, 399
indicator function, 55 pdf, see probability density, see probabil-
indicator variable, 55 ity density
information physics, 8
Fisher, 425 Poisson process, 309
Jacobian, 266 predicted values, 247
probability
Kaplan-Meier estimate, 409 continuous, 1, 7
density, 7
Laplace transform, 272 discrete, 1, 6
Law of Large Numbers, 78 proportional hazards model, 410
likelihood function, 133
likelihood set, 155 QQ plots, 112
linear model, 218 quantile, 96
linear predictor, 239
location parameter, 317 R commands
logistic regression, 239 !, 58
logit, 239 ==, 4
[[]], see subscript
marginal likelihood, 140 [], see subscript
Markov chain Monte Carlo, 344 #, 4
maximum likelihood estimate, 153 %*%, 347
mean, 32 %o%, 322
mean squared error, 420 ,̃ 224
median, 96 abline, 147
INDEX 461

acf, 393 lines, 12

apply, 62 list, 130
ar, 399 lm, 224
array, 66 log10, 151
arrows, 37 lowess, 207
as.factor, 244 matplot, 22
assignment, 4, 6 matrix, 22, 60
+<-+,4 mean, 12
boxplot, 60, 111 median, 96
c, 9 mosaicplot, 117
cbind, 75 names, 72
contour, 147 pacf, 399
coplot, 127 pairs, 50
cor, 53 par, 10
cov, 52 paste, 22
crossprod, 347 pbinom, 280
cumsum, 429 plot, 9, 20
data.frame, 244 plot.ecdf, 99
dbinom, 17, 280 plot.ts, 393
density, 12 pnbinom, 285
dexp, 22 print, 5
diff, 315, 402 qbinom, 280
dim, 72 qnbinom, 285
dimnames, 300 qqnorm, 114
dmultinom, 289 quantile, 97
dnbinom, 285 rbinom, 39, 280
dnorm, 23 read.table, 72
dpois, 20 rep, 5, 6
expression, 135 rmultinom, 289
filter, 400 rnbinom, 285
fitted, 249 rnorm, 26
for, 5 round, 429
glm, 243 sample, 3
hist, 12 scan, 130
if, 6 segments, 270
is.na, 147 seq, 20
legend, 22 solve, 347
length, 73 sqrt, 429
INDEX 462

stripchart, 109 regressor, 215

subscript, 72, 73 residual, 207
sum, 4, 5 residual plots, 230
supsmu, 207 response variable, 215
Surv, 409
survfit, 409 sampling distribution, 158
t, 347 scale parameter, 317
scatterplot smoothers, 204
tapply, 75
standard deviation, 34
text, 39
standard error, 162
unique, 105
standard Normal distribution, 317
var, 12
standard units, 26
while, 58
stationary distribution, 355
xyplot, 379
StatLib, 71
R data sets
Strong Law of Large Numbers, 78
airquality, 201
sufficient statistic, 415
attenu, 201
beaver1, 393, 395 trace plot, 359
co2, 63, 393
discoveries, 10 variance, 34
EuStockMarkets, 393
Weak Law of Large Numbers, 78
faithful, 121, 201
iris, 50
ldeaths, 393
lynx, 393
mtcars, 201, 228, 247
PlantGrowth, 213
presidents, 393, 395
Seatbelts, 393
sunspot.month, 393
Tooth Growth, 101, 183
UCBAdmissions, 114
R datasets
Orthodont, 378
R packages
lattice, 379
nlme, 378
random vector, 263
regression function, 207
 Index of Examples

1970 draft lottery, 204

baboons, 187
bladder cancer, 359

CEO salary, 141

craps, 2, 3, 15, 49, 58, 59, 157

FACE, 64, 144, 236

hot dogs, 105, 210, 216, 225

Ice Cream Consumption, 220

neurobiology, 127, 298

O-rings, 236
ocean temperatures, 29, 316

quiz scores, 109, 148

Rutherford and Geiger, 296

seedlings, 18, 43, 49, 56, 136, 148, 167,

208, 245
Slater school, 136, 140, 193

463