Lent 2000 version of March 8, 2001 Richard Weber

C11: STATISTICS

Contents
Aims of this course . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv
Schedules . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iv
Recommended books . . . . . . . . . . . . . . . . . . . . . . . . . . . iv
Keywords . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . v
Notation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . vii

1 Parameter estimation 1
1.1 What is Statistics? . . . . . . . . . . . . . . . . . . . . . . . . . . . . 1
1.2 RVs with values in R n or Zn . . . . . . . . . . . . . . . . . . . . . . . 2
1.3 Some important random variables . . . . . . . . . . . . . . . . . . . . 4
1.4 Independent and IID RVs . . . . . . . . . . . . . . . . . . . . . . . . 5
1.5 Indicating dependence on parameters . . . . . . . . . . . . . . . . . . 5
1.6 The notion of a statistic . . . . . . . . . . . . . . . . . . . . . . . . . 5
1.7 Unbiased estimators . . . . . . . . . . . . . . . . . . . . . . . . . . . 6
1.8 Sums of independent RVs . . . . . . . . . . . . . . . . . . . . . . . . 6
1.9 More important random variables . . . . . . . . . . . . . . . . . . . . 7
1.10 Laws of large numbers . . . . . . . . . . . . . . . . . . . . . . . . . . 7
1.11 The Central Limit Theorem . . . . . . . . . . . . . . . . . . . . . . . 8
1.12 Poisson process of rate λ . . . . . . . . . . . . . . . . . . . . . . . . . 8

2 Maximum likelihood estimation 9

2.1 Maximum likelihood estimation . . . . . . . . . . . . . . . . . . . . . 9
2.2 Sufficient statistics . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11

3 The Rao-Blackwell theorem 13

3.1 Mean squared error . . . . . . . . . . . . . . . . . . . . . . . . . . . . 13
3.2 The Rao-Blackwell theorem . . . . . . . . . . . . . . . . . . . . . . . 14
3.3 Consistency and asymptotic efficiency∗ . . . . . . . . . . . . . . . . . 16
3.4 Maximum likelihood and decision-making . . . . . . . . . . . . . . . . 16

4 Confidence intervals 17
4.1 Interval estimation . . . . . . . . . . . . . . . . . . . . . . . . . . . . 17
4.2 Opinion polls . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 18
4.3 Constructing confidence intervals . . . . . . . . . . . . . . . . . . . . 19
4.4 A shortcoming of confidence intervals* . . . . . . . . . . . . . . . . . 20

i
5 Bayesian estimation 21
5.1 Prior and posterior distributions . . . . . . . . . . . . . . . . . . . . . 21
5.2 Conditional pdfs . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 22
5.3 Estimation within Bayesian statistics . . . . . . . . . . . . . . . . . . 24

6 Hypothesis testing 25
6.1 The Neyman–Pearson framework . . . . . . . . . . . . . . . . . . . . 25
6.2 Terminology . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 26
6.3 Likelihood ratio tests . . . . . . . . . . . . . . . . . . . . . . . . . . . 27
6.4 Single sample: testing a given mean, simple alternative, known vari-
ance (z-test) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 28

7 Further aspects of hypothesis testing 29

7.1 The p-value of an observation . . . . . . . . . . . . . . . . . . . . . . 29
7.2 The power of a test . . . . . . . . . . . . . . . . . . . . . . . . . . . . 29
7.3 Uniformly most powerful tests . . . . . . . . . . . . . . . . . . . . . . 30
7.4 Confidence intervals and hypothesis tests . . . . . . . . . . . . . . . . 31
7.5 The Bayesian perspective on hypothesis testing . . . . . . . . . . . . 32

8 Generalized likelihood ratio tests 33

8.1 The χ2 distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . 33
8.2 Generalised likelihood ratio tests . . . . . . . . . . . . . . . . . . . . 33
8.3 Single sample: testing a given mean, known variance (z-test) . . . . . 34
8.4 Single sample: testing a given variance, known mean (χ2-test) . . . . 35
8.5 Two samples: testing equality of means, known common variance (z-
test) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
8.6 Goodness-of-fit tests . . . . . . . . . . . . . . . . . . . . . . . . . . . 36

9 Chi-squared tests of categorical data 37

9.1 Pearson’s chi-squared statistic . . . . . . . . . . . . . . . . . . . . . . 37
9.2 χ2 test of homogeneity . . . . . . . . . . . . . . . . . . . . . . . . . . 38
9.3 χ2 test of row and column independence . . . . . . . . . . . . . . . . 40

10 Distributions of the sample mean and variance 41

10.1 Simpson’s paradox . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
10.2 Transformation of variables . . . . . . . . . . . . . . . . . . . . . . . 41
10.3 Orthogonal transformations of normal variates . . . . . . . . . . . . . 42
10.4 The distributions of X̄ and SXX . . . . . . . . . . . . . . . . . . . . . 43
10.5 Student’s t-distribution . . . . . . . . . . . . . . . . . . . . . . . . . . 44

ii
11 The t-test 45
11.1 Confidence interval for the mean, unknown variance . . . . . . . . . . 45
11.2 Single sample: testing a given mean, unknown variance (t-test) . . . . 46
11.3 Two samples: testing equality of means, unknown common variance
(t-test) . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 47
11.4 Single sample: testing a given variance, unknown mean (χ2-test) . . . 48

12 The F -test and analysis of variance 49

12.1 F -distribution . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
12.2 Two samples: comparison of variances (F -test) . . . . . . . . . . . . . 49
12.3 Non-central χ2 . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 50
12.4 One way analysis of variance . . . . . . . . . . . . . . . . . . . . . . . 50

13 Linear regression and least squares 53

13.1 Regression models . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
13.2 Least squares/MLE . . . . . . . . . . . . . . . . . . . . . . . . . . . . 53
13.3 Practical usage . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 54
13.4 Data sets with the same summary statistics . . . . . . . . . . . . . . 55
13.5 Other aspects of least squares . . . . . . . . . . . . . . . . . . . . . . 56

14 Hypothesis tests in regression models 57

14.1 Distributions of the least squares estimators . . . . . . . . . . . . . . 57
14.2 Tests and confidence intervals . . . . . . . . . . . . . . . . . . . . . . 58
14.3 The correlation coefficient . . . . . . . . . . . . . . . . . . . . . . . . 58
14.4 Testing linearity . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
14.5 Analysis of variance in regression models . . . . . . . . . . . . . . . . 59

15 Computational methods 61
15.1 Analysis of residuals from a regression . . . . . . . . . . . . . . . . . 61
15.2 Discriminant analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . 62
15.3 Principal components / factor analysis . . . . . . . . . . . . . . . . . 62
15.4 Bootstrap estimators . . . . . . . . . . . . . . . . . . . . . . . . . . . 64

16 Decision theory 65
16.1 The ideas of decision theory . . . . . . . . . . . . . . . . . . . . . . . 65
16.2 Posterior analysis . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 67
16.3 Hypothesis testing as decision making . . . . . . . . . . . . . . . . . . 68
16.4 The classical and subjective points of view . . . . . . . . . . . . . . . 68

iii
Aims of the course
The aim of this course is to aquaint you with the basics of mathematical statistics:
the ideas of estimation, hypothesis testing and statistical modelling.
After studying this material you should be familiar with
1. the notation and keywords listed on the following pages;
2. the definitions, theorems, lemmas and proofs in these notes;
3. examples in notes and examples sheets that illustrate important issues concerned
with topics mentioned in the schedules.

Schedules
Estimation
Review of distribution and density functions, parametric families, sufficiency, Rao-
Blackwell theorem, factorization criterion, and examples; binomial, Poisson, gamma.
Maximum likelihood estimation. Confidence intervals. Use of prior distributions and
Bayesian inference.
Hypothesis Testing
Simple examples of hypothesis testing, null and alternative hypothesis, critical re-
gion, size, power, type I and type II errors, Neyman-Pearson lemma. Significance
level of outcome. Uniformly most powerful tests. Likelihood ratio, and the use of
likelihood ratio to construct test statistics for composite hypotheses. Generalized
likelihood-ratio test. Goodness-of-fit and contingency tables.
Linear normal models
The χ2 , t and F distribution, joint distribution of sample mean and variance, Stu-
dent’s t-test, F -test for equality of two variances. One-way analysis of variance.
Linear regression and least squares
Simple examples, *Use of software*.

Recommended books
M. H. De Groot, Probability and Statistics, 2nd edition, Addison-Wesley, 1986.
J. A. Rice, Mathematical Statistics and Data Analysis, 2nd edition, Duxbury Press,
1994.
G. Casella and J. O. Berger, Statistical Inference, Brooks Cole, 1990.
D. A. Berry and B. W. Lindgren, Statistics, Theory and Methods, Brooks Cole, 1990
(out of print).

iv
Keywords

absolute error loss, 24 geometric distribution, 7

acceptance region, 31 goodness-of-fit, 26, 36
alternative hypothesis, 25 goodness-of-fit test, 37
analysis of variance, 50
hypothesis testing, 25
asymptotically efficient, 16
asymptotically unbiased, 12, 14 IID, 5
independent, 5
Bayesian inference, 21–24, 32 interval estimate, estimator, 17
beta distribution, 7
between samples sum of squares, 51 Jacobian, 42
biased, 10 least squares estimators, 53
binomial distribution, 4 likelihood, 9, 27
bootstrap estimate, 64 likelihood ratio, 27
Central Limit theorem, 8 likelihood ratio test, 27
chi-squared distribution, 33 location parameter, 19
log-likelihood, 9
χ2 test of homogeneity, 38
loss function, 24
χ2 test of independence, 40
composite hypothesis, 26 maximum likelihood estimator
confidence interval, 17–20, 31–32 (MLE), 9
consistent, 16 mean squared error (MSE), 13
contingency table, 40 multinomial distribution, 36
critical region, 26
Neyman–Pearson lemma, 27
decision-making, 2 non-central chi-squared, 50
degrees of freedom, 33, 38 nuisance parameter, 48
discriminant analysis, 62 null hypothesis, 25
distribution function, 3 one-tailed test, 28
estimate, estimator, 6 outlier, 56
expectation, 3 p-value, 29
exponential distribution, 7 paired samples t-test, 47
F -distribution, 49 parameter estimation, 1
parametric family, 5, 21, 25
factor analysis, 62
Pearson chi-squared statistic, 37
factorization criterion, 11
point estimate, 17
gamma distribution, 7 Poisson distribution, 4
generalised likelihood ratio test, 33 Poisson process, 8

v
posterior, 21 Simpson’s paradox, 41
posterior mean, 24 size, 26
posterior median, 24 standard error, 45
power function, 29 standard normal, 4
predictive confidence interval, 58 standardized, 8
principal components, 62 standardized residuals, 61
prior, 21 statistic, 5
probability density function, 3
strong law of large numbers, 8
probability mass function, 3
sufficient statistic, 11
quadratic error loss, 24
t-distribution, 18, 44
Rao–Blackwell theorem, 14 t-test, 46
Rao–Blackwellization, 15 two-tailed test, 28, 34
regression through the origin, 56 type I error, 26
residual sum of squares, 57 type II error, 26
RV, 2
unbiased estimator, 6
sample correlation coefficient, 58
uniform distribution, 4
scale parameter, 19
significance level of a test, 26, 29 uniformly most powerful, 30
significance level of an observation, variance, 3
29
simple hypothesis, 26 weak law of large numbers, 8
simple linear regression model, 53 within samples sum of squares, 51

vi
Notation
X a scalar or vector random variable, X = (X1, . . . , Xn)
X∼ X has the distribution . . .
E X, var(X) mean and variance of X
µ, σ 2 mean and variance as typically used for N(µ, σ 2)
RV, IID ‘random variable’, ‘independent and identically distributed’
beta(m, n) beta distribution
B(n, p) binomial distribution
χ2n chi-squared distribution with n d.f.
E(λ) exponential distribution
Fm,n F distribution with m and n d.f.
gamma(n, λ) gamma distribution
N(µ, σ 2) normal (Gaussian) distribution
P (λ) Poisson distribution
U [a, b] uniform distribution
tn Student’s t distribution with n d.f.
Φ distribution function of N(0, 1)
φ density function of N(0, 1)
(n) (m,n)
zα , tα , Fα upper α points of N(0, 1), tn and Fm,n distributions
θ a parameter of a distribution
θ̂(X), θ̂(x) an estimator of θ, a estimate of θ.
MLE ‘maximum likelihood estimator’
FX (x | θ) distribution function of X depending on a parameter θ
fX (x | θ) density function of X depending on a parameter θ
fθ (x) density function depending on a parameter θ
fX|Y conditional density of X given Y
p(θ | x) posterior density of θ given data x
x 1 , . . . , xn n
Pobserved P data values
P
xi·, x·j , x·· j xij , i xij and ij xij
T (x) a statistic computed from x1, . . . , xn

vii
 H0 , H1 null and alternative hypotheses
f0 , f1 null and alternative density functions
Lx (H0), Lx(H1) likelihoods of H0 and H1 given data x
Lx (H0, H1) likelihood ratio Lx(H1)/Lx(H0)
(n) (m,n)
tα , Fα points to the right of which lie α100% of Tn and Fm,n
C critical region: reject H0 if T (x) ∈ C.
W (θ) power function, W (θ) = P(X ∈ C | θ)
α,β probabilities of Type I and Type II error
intercept and gradient of a regression line, Yi = α + βwi + i
oi , ei , δi observed and expected counts; δi = oi − ei
X̄ mean of X1, . .P
P . , Xn P
SXX , SY Y , SXY (Xi − X̄) , (Yi − Ȳ )2, (Xi − X̄)(Yi − Ȳ )
2

s.e. ‘standard error’,

square root of an unbiased estimator of a variance.
R residual sum of square in a regression model
s2 unbiased estimate of the variance, s2 = SXX /(n − 1).
d(X) decision function, d(X) = a.
L(θ, a) loss function when taking action a.
R(θ, d) risk function, R(θ, d) = E [L(θ, d(X))].
B(d) Bayes risk, E [R(θ, d)].

WWW site
There is a web page for this course, with copies of the lecture notes, examples sheets,
corrections, past tripos questions, statistical tables and other additional material. It
can be accessed as http://www.statslab.cam.ac.uk/~rrw1/stats/

viii
1 Parameter estimation
Statisticians do it when it counts.

1.1 What is Statistics?

Statistics is a collection of procedures and principles for gaining and processing
information in order to make decisions when faced with uncertainty.
This course is concerned with “Mathematical Statistics”, i.e., mathematical ideas
and tools that are used by statisticians to analyse data. We will study techniques
for estimating parameters, fitting models, and testing hypotheses. However, as we
study these techniques we should not lose sight of the fact that a practicing statisti-
cian needs more than simply a knowledge of mathematical techniques. The collection
and interpretation of data is a subtle art. It requires common sense. It can some-
times raise philosophical questions. Although this course is primarily concerned with
mathematical techniques, I will try — by means of examples and digressions — also
to introduce you to some of the non-mathematical aspects of Statistics.
Statistics is concerned with data analysis: using data to make inferences. It is
concerned with questions like ‘what is this data telling me?’ and ‘what does this
data suggest it reasonable to believe?’ Two of its principal concerns are parameter
estimation and hypothesis testing.

Example 1.1 Suppose we wish to estimate the proportion p of students in Cambridge

who have not showered or bathed for over a day.
This is poses a number of questions. Who do we mean by students? Suppose
time is limited and we can only interview 20 students in the street. Is it important
that our survey be ‘random’ ? How can we ensure this? Will students we question
be embarrassed to admit if they have not bathed? And even if we can get truthful
answers, will we be happy with our estimate if that random sample turns out to
include no women, or if it includes only computer scientists?
Suppose we find that 5 have not bathed for over a day. We might estimate p by
p̂ = 5/20 = 0.25. But how large an error might we expect p̂ to have?
Many families of probability distributions depend on a small number of parame-
ters; for example, the Poisson family depends on a single parameter λ and the Normal
family on two parameters µ and σ. Unless the values of the parameters are known
in advance, they must be estimated from data. One major themes of mathematical
statistics is the theory of parameter estimation and its use in fitting probability
distributions to data. A second major theme of Statistics is hypothesis testing.

Example 1.2 A famous study investigated the effects upon heart attacks of taking
an aspirin every other day. The results after 5 years were

1
 Condition Heart attack No heart attack Attacks per 1000
Aspirin 104 10,933 9.42
Placebo 189 10,845 17.13

What can make of this data? Is it evidence for the hypothesis that aspirin prevents
heart attacks?
The aspirin study is an example of a controlled experiment. The subjects were
doctors aged 40 to 84 and none knew whether they were taking the aspirin or the
placebo. Statistics is also concerned with analysing data from observational stud-
ies. For example, most of us make an intuitive statistical analysis when we use our
previous experience to help us choose the shortest checkout line at a supermarket.
The data analysis of observational studies and experiments is a central component
of decision-making, in science, medicine, business and government.

By the way: data is a plural noun referring to a collection of numbers or other

pieces of information to which meaning has been attached.
The numbers 1.1, 3.0, 6.5 are not necessarily data. They become so when we are
told that they are the muscle weight gains in kg of three athletes who have been
trying a new diet.

1.2 RVs with values in R n or Zn

In Statistics, our data are modelled by a vector of random variables

X = (X1, X2 , . . . , Xn )

where Xi takes values in Z or R .

To succeed in this course you should brush up on your knowledge of basic proba-
bility: of key distributions and how to make calculations with random variables. Let
us review a few facts.
When our sample space Ω (a set of possible outcomes) is discrete (finite or count-
ably infinite) we have a random variable (RV) X with values in Z:

X : Ω → Z.

RVs can also take values in R rather than in Z and the sample space Ω can be
uncountable.

X : Ω → R.

Since the outcome ω, ω ∈ Ω, is random, X is a function whose value, X(ω), is

also random. E.g., to model the experiment of tossing a coin twice we might take
Ω = {hh, ht, th, th}. Then X might be the total number of heads.

2
 In both cases the distribution function FX of X is defined as:
X
FX (x) := P(X ≤ x) = P(ω).
{ω : X(ω)≤x}

In the discrete case the probability mass function (pmf) fX of X is

fX (k) := P(X = k), k ∈ Z.

So
X
P(X ∈ A) = fX (x), A ⊆ Z.
x∈A

In the continuous case we have the probability density function (pdf) fX of X.

In all cases we shall meet, X will have a piecewise smooth pdf such that
Z
P(X ∈ A) = fX (x) dx, for nice (measurable) subsets A ⊆ R .
x∈A

Expectation of X: In the discrete case

X X
E (X) := X(ω)P(ω) = k P(X = k),
ω∈Ω k∈Z

the first formula being the real definition. In the continuous case the calculation
Z
E(X) = X(ω) P(dω)
Ω

needs measure theory. However,

d
fX (x) = FX (x) except perhaps for finitely many x.
dx
Measure theory shows that for any nice function h on R ,
Z
E h(X) = h(x)fX (x) dx .
R

Variance of X: If E (X) = µ, then

var(X) = E (X − µ)2 = E (X 2 ) − µ2 .

3
1.3 Some important random variables
(a) We say that X has the binomial distribution B(n, p), and write X ∼ B(n, p),
if
(

k p (1 − p) if k ∈ {0, . . . , n},
n k n−k
P(X = k) =
0 otherwise.

Then E (X) = np, var(X) = np(1 − p). This is the distribution of the number of
successes in n independent trials, each of which has probability of success p.
(b) We say that X has the Poisson distribution with parameter λ, and write
X ∼ P (λ), if
(
e−λλk /k! if k ∈ {0, 1, 2, . . . },
P(X = k) =
0 otherwise.

Then E (X) = var(X) = λ. The Poisson is the limiting distribution of B(n, p) as

n → ∞ and p → 0 with λ = np.
(c) We say that X is standard normal, and write X ∼ N(0, 1), if
1
fX (x) = ϕ(x) := √ exp(−x2/2), −∞ ≤ x ≤ ∞.
2π
Then
Z x Z x
FX (x) = fX (y) dy = Φ(x) := ϕ(y) dy.
−∞ −∞

Then E (X) = 0, var(X) = 1. Φ and ϕ are standard notations.

(d) We say that X is Normal with mean µ and variance σ 2 and write X ∼ N(µ, σ 2)
if
 
1 (x − µ)2
fX (x) = √ exp − , −∞ ≤ x ≤ ∞.
σ 2π 2σ 2

Then E (X) = µ, var(X) = σ 2 .

(e) We say that X is uniform on [a, b], and write X ∼ U [a, b], if
1
fX (x) = , x ∈ [a, b].
b−a

Then E (X) = 12 (a + b), var(X) = 1

12 (b − a)2 .

4
1.4 Independent and IID RVs
Random variables X1 , . . . , Xn are called independent if for all x1, . . . , xn
P(X1 ≤ x1; . . . ; Xn ≤ xn) = P(X1 ≤ x1) · · · P(Xn ≤ xn).
IID stands for independent identically distributed. Thus if X1 , X2, . . . , Xn are
IID RVs, then they all have the same distribution function and hence the same mean
and same variance.
We work with the probability mass function (pmf) of X in Zn or probability
density function (pdf) of X in R n : In most cases, X1 , . . . , Xn are independent, so
that if x = (x1, . . . , xn) ∈ R n , then
fX (x) = fX1 (x1) · · · fXn (xn).

1.5 Indicating dependence on parameters

If X ∼ N(µ, σ 2 ), then we indicate the dependence of the pdf of X on µ and σ 2 by
writing it as
 
1 (x − µ)2
f (x | µ, σ ) =
2
exp −
(2πσ 2)1/2 2σ 2
Or if X = (X1 , . . . , Xn ), where X1 , . . . , Xn are IID N(µ, σ 2), then we would have
 
1 kx − µ1k 2
f (x | µ, σ 2) = exp −
(2πσ 2)n/2 2σ 2
where µ1 denotes the vector (µ, µ, . . . , µ)>.
In general, we write f (x | θ) to indicate that the pdf depends on a parameter θ.
θ may be a vector of parameters. In the above θ = (µ, σ 2)>. An alternative notation
we will sometimes employ is fθ (x).
The set of distributions with densities fθ (x), θ ∈ Θ, is called a parametric
family. E.g.,, there is a parametric family of normal distributions, parameterised
by values of µ, σ 2 . Similarly, there is a parametric family of Poisson distributions,
parameterised by values of λ.

1.6 The notion of a statistic

A statistic, T (x), is any function of the data. E.g., given the data x = (x1, . . . , xn),
four possible statistics are
1 x1 + x 3
(x1 + · · · + xn ), max xi, log x4, 1997 + 10 min xi.
n i xn i

Clearly, some statistics are more natural and useful than others. The first of these
would be useful for estimating µ if the data are samples from a N(µ, 1) distribution.
The second would be useful for estimating θ if the data are samples from U [0, θ].

5
1.7 Unbiased estimators
An estimator of a parameter θ is a function T = T (X) which we use to estimate θ
from an observation of X. T is said to be unbiased if

E (T ) = θ.

The expectation above is taken over X. Once the actual data x is observed, t = T (x)
is the estimate of θ obtained via the estimator T .

Example 1.3 Suppose X1 , . . . , Xn arePIID B(1, p) and p is unknown. Consider the

estimator for p given by p̂(X) = X̄ = i Xi /n. Then p̂ is unbiased, since
 
1 1 1
E p̂(X) = E (X1 + · · · + Xn ) = (E X1 + · · · + E Xn ) = np = p .
n n n

Another possible unbiased estimator for p is p̃ = 13 (X1 + 2X2) (i.e., we ignore most
of the data.) It is also unbiased since
 
1 1 1
E p̃(X) = E (X1 + 2X2) = (E X1 + 2E X2 ) = (p + 2p) = p .
3 3 3
Intuitively, the first estimator seems preferable.

1.8 Sums of independent RVs

In the above calculations we have used the fact the expectation of a sum of random
variables is the sum of their expectations. It is always true (even when X1 , . . . , Xn
are not independent) that

E (X1 + · · · + Xn ) = E (X1 ) + · · · + E (Xn ),

and for linear combinations of RVs

E (a1 X1 + · · · + an Xn ) = a1 E (X1 ) + · · · + an E (Xn ).

If X1, X2 , . . . , Xn are independent, then

E (X1 X2 . . . Xn ) = E (X1 )E (X2 ) · · · E (Xn ),

var(X1 + · · · + Xn ) = var(X1) + · · · + var(Xn ),

and for linear combinations of independent RVs

var(a1X1 + · · · + an Xn ) = a21 var(X1) + · · · + a2n var(Xn ).

6
1.9 More important random variables
(a) We say that X is geometric with parameter p, if
(
p(1 − p)k−1 if k ∈ {1, 2, . . . },
P(X = k) =
0 otherwise.
Then E(X) = 1/p and var(X) = (1 − p)/p2. X is the number of the toss on which
we first observe a head if we toss a coin which shows heads with probability p.
(b) We say that X is exponential with rate λ, and write X ∼ E(λ), if
(
λe−λx if x > 0,
fX (x) =
0 otherwise.
Then E (X) = λ−1, var(X) = λ−2 .
The geometric and exponential distributions are discrete and continuous ana-
logues. They are the unique ‘memoryless’ distributions, in the sense that P(X ≥
t + s | X ≥ t) = P(X ≥ s). The exponential is the distribution of the time between
successive events of a Poisson process.
(c) We say that X is gamma(n, λ) if
(
λn xn−1e−λx /(n − 1)! if x > 0,
fX (x) =
0 otherwise.
X has the distribution of the sum of n IID RVs that have distribution E(λ). So
E(λ) = gamma(1, λ). E (X) = nλ−1 and var(X) = nλ−2.
This also makes sense for real n > 0 (and λ > 0), if we interpret (n − 1)! as Γ(n),
R ∞ n−1
where Γ(n) = 0 x e−x dx.
(d) We say that X is beta(a, b) if
(
1
B(a,b) xa−1(1 − x)b−1 if 0 < x < 1,
fX (x) =
0 otherwise.
Here B(a, b) = Γ(a)Γ(b)/Γ(a + b). Then
a ab
E (X) = , var(X) = .
a+b (a + b + 1)(a + b)2

1.10 Laws of large numbers

Suppose X1, X2 , . . . is a sequence of IID RVs, each having finite mean µ and variance
σ 2 . Let
Sn := X1 + X2 + · · · + Xn , so that E (Sn ) = nµ, var(Sn ) = nσ 2.

7
 The weak law of large numbers is that for  > 0,

P(|Sn /n − µ| > ) → 0, as n → ∞ .

The strong law of large numbers is that

P(Sn /n → µ) = 1 .

1.11 The Central Limit Theorem

Suppose X1 , X2, . . . are as above. Define the standardized version Sn∗ of Sn as
Sn − nµ
Sn∗ = √ , so that E (Sn∗ ) = 0, var(Sn∗ ) = 1.
σ n
Then for large n, Sn∗ is approximately standard Normal: for a < b,
√ √

lim P(a ≤ Sn∗ ≤ b) = Φ(b) − Φ(a) = lim P nµ + aσ n ≤ Sn ≤ nµ + bσ n .
n→∞ n→∞

In particular, for large n,

√
P(|Sn − nµ| < 1.96σ n) + 95%
since Φ(1.96) = 0.0975 and Φ(−1.96) = 0.025.

1.12 Poisson process of rate λ

The Poisson process is used to model a process of arrivals: of people to a supermarket
checkout, calls at telephone exchange, etc.
Arrivals happen at times

T1 , T 1 + T 2 , T 1 + T 2 + T 3 , . . .

where T1 , T2, . . . are independent and each exponentially distributed with parameter
λ. Numbers of arrivals in disjoint intervals are independent RVs, and the number of
arrivals in any interval of length t has the P (λt) distribution. The time

Sn = T1 + T2 + · · · + Tn

of the nth arrival has the gamma(n, λ) distribution, and 2λSn ∼ X2n
2
.

8
2 Maximum likelihood estimation
When it is not in our power to follow what is true, we ought
to follow what is most probable. (Descartes)

2.1 Maximum likelihood estimation

Suppose that the random variable X has probability density function f (x | θ). Given
the observed value x of X, the likelihood of θ is defined by

lik(θ) = f (x | θ) .

Thus we are considering the density as a function of θ, for a fixed x. In the case
of multiple observations, i.e., when x = (x1, . . . , xn) is a vector of observed values
of X1 , . . . , Xn, we assume, unless otherwise stated, that X1 , . . . , Xn are IID; in this
case f (x1, . . . , xn | θ) is the product of the marginals,
Y
n
lik(θ) = f (x1, . . . , xn | θ) = f (xi | θ) .
i=1

It makes intuitive sense to estimate θ by whatever value gives greatest likelihood

to the observed data. Thus the maximum likelihood estimate θ̂(x) of θ is defined
as the value of θ that maximizes the likelihood. Then θ̂(X) is called the maximum
likelihood estimator (MLE) of θ.
Of course, the maximum likelihood estimator need not exist, but in many examples
it does. In practice, we usually find the MLE by maximizing log f (x | θ), which is
known as the log-likelihood.

Examples 2.1

(a) Smarties are sweets which come in k equally frequent colours. Suppose we do
not know k. We sequentially examine 3 Smarties and they are red, green, red. The
likelihood of this data, x = the second Smartie differs in colour from the first but the
third Smartie matches the colour of the first, is
 
k−1 1
lik(k) = p(x | k) = P(2nd differs from 1st)P(3rd matches 1st) =
k k
= (k − 1)/k ,
2

which equals 1/4, 2/9, 3/16 for k = 2, 3, 4, and continues to decrease for greater k.
Hence the maximum likelihood estimate is k̂ = 2.
Suppose a fourth Smartie is drawn and it is orange. Now

lik(k) = (k − 1)(k − 2)/k3 ,

9
which equals 2/27, 3/32, 12/125, 5/54 for k = 3, 4, 5, 6, and decreases thereafter.
Hence the likelihood estimate is k̂ = 5. Note that although we have seen only 3
colours the maximum likelihood estimate is that there are 2 colours we have not yet
seen.
(b) X ∼ B(n, p), n known, p to be estimated.
Here
 
n x
log p(x | n, p) = log p (1 − p)n−x = · · · + x log p + (n − x) log(1 − p) .
x
This is maximized where
x n−x
− = 0,
p̂ 1 − p̂
so the MLE of p is p̂ = X/n. Since E [X/n] = p the MLE is unbiased.
(c) X ∼ B(n, p), p known, n to be estimated.
Now we want to maximize
 
n x
p(x | n, p) = p (1 − p)n−x
x
with respect to n, n ∈ {x, x + 1, . . . }. To do this we look at the ratio


x
p (1 − p)
n+1 x
p(x | n + 1, p) n+1−x
(1 − p)(n + 1)
= = .
p(x | n, p) x p (1 − p)
n x n−x n+1−x
This is monotone decreasing in n. Thus p(x | n, p) is maximized by the least n for
which the above expression is ≤ 1, i.e., the least n such that
(1 − p)(n + 1) ≤ n + 1 − x ⇐⇒ n + 1 ≥ x/p ,
giving a MLE of n̂ = [X/p]. Note that if x/p happens to be an integer then both
n = x/p − 1 and n = x/p maximize p(x | n, p). Thus the MLE need not be unique.
(d) X1 , . . . , Xn ∼ geometric(p), p to be estimated.
Because the Xi are IID their joint density is the product of the marginals, so
!
Y
n X n
log f (x1, . . . , xn | p) = log (1 − p)xi −1p = xi − n log(1 − p) + n log p .
i=1 i=1
with a maximum where
P
−n n
i xi
− + =0.
1 − p̂ p̂
So the MLE is p̂ = X̄ −1 . This MLE is biased. For example, in the case n = 1,
X∞
1 p
E [1/X1 ] = (1 − p)x−1p = − log p > p .
x=1
x 1 − p
Note that E [1/X1 ] does not equal 1/E X1 .

10
2.2 Sufficient statistics
The MLE, if it exists, is always a function of a sufficient statistic. The informal no-
tion of a sufficient statistic T = T (X1, . . . , Xn ) is that it summarises all information
in {X1, . . . , Xn } which is relevant to inference about θ.
Formally, the statistic
T = T (X) is said to be sufficient for θ ∈ Θ if, for each
t, Pθ X ∈ · | T (X) = t does not depend on θ. I.e., the conditional distribution of
X1 , . . . , Xn given T (X) = t does not involve θ. Thus to know more about x than
that T (x) = t is of no additional help in making any inference about θ.

Theorem 2.2 The statistic T is sufficient for θ if and only if f (x | θ) can be ex-
pressed as


f (x | θ) = g T (x), θ h(x).

This is called the factorization criterion.

Proof. We give a proof for the case that the sample space is discrete. A contin-
uous sample space needs measure theory. Suppose f (x | θ) = Pθ (X = x) has the
factorization above and T (x) = t. Then



Pθ (X = x) g T (x), θ h(x)
Pθ X = x | T (X) = t =
=P

Pθ T (X) = t x:T (x)=t g T (x), θ h(x)
g(t, θ)h(x) h(x)
=P =P
x:T (x)=t g(t, θ)h(x) x:T (x)=t h(x)

which does not depend on θ. Conversely, if T is sufficient and T (x) = t,

Pθ (X = x) = Pθ T (X) = t Pθ X = x | T (X) = t

where by sufficiency the second factor does not depend on θ. So we identify the first
and second terms on the r.h.s. as g(t, θ) and h(x) respectively.

Examples 2.3
(a) X1 , . . . , Xn ∼ P (λ), λ to be estimated.

Y
n P .Y
n
xi −λ −nλ
f (x | λ) = {λ e /xi!} = λ i e
xi
xi ! .
i=1 i=1

P Q P
So g T (x), λ = λ i xi e−nλ and h(x) = 1 / i xi!. A sufficient statistic is t = i xi .
Note that the sufficient statistic is not unique. If T (X) is a sufficient statistic,
then so are statistics like T (X)/n and log T (X).

11
 The MLE is found by maximizing f (x | λ), and so
 P
d  xi
log f (x | λ) = i −n=0.
dλ λ̂
λ=λ̂

Hence λ̂ = X̄. It is easy to check that λ̂ is unbiased.

Note that the MLE is always a function of the sufficient statistic. This is because


the MLE is the value of θ which maximizes f (x | θ),
and f (x | θ) = g T (x), θ h(x).
Thus the MLE is the θ which maximizes g T (x), θ , and hence a function of T (x).
(b) X1 , . . . , Xn ∼ N(µ, σ 2), θ = (µ, σ 2) to be estimated.
Y
n
1 1 P
e−(xi −µ) /2σ = − i (xi −µ)2 /2σ2
2 2
f (x | µ, σ ) =
2
√ e
i=1 2πσ 2 (2πσ 2)n/2
1 P
−[ i (xi −x̄)2 +n(x̄−µ)2 )]/2σ2
= e
(2πσ 2)n/2


Thus, with g T (x), θ taken P as the whole
r.h.s. and h(x) = 1, the sufficient statistic
for θ = (µ, σ ) is T (x) = x̄, i (xi − x̄)2 .
2

Note that sometimes the sufficient statistic

is not just a single number, but as
here, a vector T (X) = T1 (X), . . . , Tr (X) . This usually occurs when the parameter
is a vector, as θ = (µ, σ 2).
In this example, if σ 2 had
P been known, then x̄ would have been sufficient for µ.
If µ had been known, then i (xi − µ)2 would have been sufficient for σ 2 .
(c) X1 , . . . , Xk ∼ U [0, θ], θ to be estimated.
Y
n
1 1
f (x | θ) = 1{0 ≤ xi ≤ θ} = 1{0 ≤ max xi ≤ θ} n ,
i=1
θ i θ


where 1{condition} = 1 or 0 as ‘condition’ is true or false. Thus g T (x), θ = 1{0 ≤
maxi xi ≤ θ}/θn, h(x) = 1 and T (x) = maxi xi is sufficient for θ. The MLE is
θ̂ = maxi Xi .
To find E θ̂ we must find the distribution function of maxi xi. This is
F (t) = P(max xi ≤ t) = P(x1 ≤ t) · · · P(xn ≤ t) = (t/θ)n .
i

By differentiation, f (t) = ntn−1/θn , and hence

Z θ
ntn−1 n
E max xi = t n dt = θ.
i 0 θ n+1
So θ̂ is biased.
However, E θ̂ → θ as n → ∞. We say that θ̂ is asymptotically unbiased.
Under some mild assumptions, MLEs are always asymptotically unbiased. This is
one reason why we like the maximum likelihood estimation procedure.

12
3 The Rao-Blackwell theorem
Variance is what any two statisticians are at.

3.1 Mean squared error

A good estimator should take values close to the true value of the parameter it is
attempting to estimate. If θ̂ is an unbiased estimator of θ then E (θ̂ − θ)2 is the
variance of θ̂. If θ̂ is a biased estimator of θ then E (θ̂ − θ)2 is no longer the variance
of θ̂, but it is still useful as a measure of the mean squared error (MSE) of θ̂.

Example 3.1 Consider the estimators in Example 1.3. Each is unbiased, so its MSE
is just its variance.
 
1 var(X1) · · · + var(Xn) np(1 − p) p(1 − p)
var(p̂) = var (X1 + · · · + Xn ) = = =
n n2 n2 n
 
1 var(X1 ) + 4 var(X2) 5p(1 − p)
var(p̃) = var (X1 + 2X2 ) = =
3 9 9
Not surprisingly, var(p̂) < var(p̃). In fact, var(p̂)/ var(p̃) → 0, as n → ∞.
Note that p̂ is the MLE of p. Another possible unbiased estimator would be
1
p∗ = 1 (X1 + 2X2 + · · · + nXn )
2 n(n + 1)

with variance
1
2(2n + 1)
var(p∗) =  2 1 + 2 + · · · + n p(1 − p) =
2 2
p(1 − p) .
1
n(n + 1) 3n(n + 1)
2

Here var(p̂)/ var(p∗) → 3/4.

The next example shows that neither a MLE or an unbiased estimator necessarily
minimizes the mean square error.

Example 3.2 Suppose X1, . . . , Xn ∼ N(µ, σ 2), µ and σ 2 unknown and to be esti-
mated. To find the MLEs we consider
Y
n
1 n 1 X
n
−(xi −µ)2 /2σ2
log f (x | µ, σ ) = log
2
√ e = − log(2πσ 2) − 2 (xi − µ)2 .
i=1 2πσ 2 2 2σ i=1

This is maximized where ∂(log f )/∂µ = 0 and ∂(log f )/∂σ 2 = 0. So

X
n
n 1 X
n
2
(1/σ̂ ) (xi − µ̂) = 0, and − 2+ 4 (xi − µ̂)2 = 0,
i=1
2σ̂ 2σ̂ i=1

13
and the MLEs are
1X 1X
n n
1
µ̂ = X̄ = Xi , 2
σ̂ = SXX := (Xi − X̄)2.
n i=1 n n i=1

It is easy to check that µ̂ is unbiased. As regards σ̂ 2 note that

" n # " n # " n #
X X X
E (Xi − X̄)2 = E (Xi − µ + µ − X̄)2 = E (Xi − µ)2 − nE (µ − X̄)2
i=1 i=1 i=1
= nσ − n(σ /n) = (n − 1)σ
2 2 2

so σ̂ 2 is biased. An unbiased estimator is s2 = SXX /(n − 1).

Let us consider an estimator of the form λSXX . Above we see SXX has mean
(n − 1)σ 2 and later we will see that its variance is 2(n − 1)σ 4. So
 2  
E λSXX − σ 2 = 2(n − 1)σ 4 + (n − 1)2σ 4 λ2 − 2(n − 1)σ 4λ + σ 4 .

This is minimized by λ = 1/(n + 1). Thus the estimator which minimizes the mean
squared error is SXX /(n + 1) and this is neither the MLE nor unbiased. Of course
there is little difference between any of these estimators when n is large.
Note that E [σ̂ 2 ] → σ 2 as n → ∞. So again the MLE is asymptotically unbiased.

3.2 The Rao-Blackwell theorem

The following theorem says that if we want an estimator with small MSE we can
confine our search to estimators which are functions of the sufficient statistic.

Theorem 3.3 (Rao-Blackwell Theorem) Let θ̂ be an estimator of θ with E (θ̂2 ) <

∞ for all θ. Suppose that T is sufficient for θ, and let θ∗ = E (θ̂ | T ). Then for all θ,

E (θ∗ − θ)2 ≤ E (θ̂ − θ)2.

The inequality is strict unless θ̂ is a function of T .

Proof.

E [θ∗ − θ]2
h
i2 h
i2 h
i
= E E θ̂ | T − θ = E E θ̂ − θ | T ≤ E E (θ̂ − θ) | T
2
= E (θ̂ − θ)2

The outer expectation is being taken with respect to T . The inequality follows from
the fact that for any RV, W , var(W ) = E W 2 − (E W )2 ≥ 0. We put W = (θ̂ − θ | T )
and note that there is equality only if var(W ) = 0, i.e., θ̂ − θ can take just one value
for each value of T , or in other words, θ̂ is a function of T .

14
 Note that if θ̂ is unbiased then θ∗ is also unbiased, since
h i
∗
E θ = E E (θ̂ | T ) = E θ̂ = θ .
We now have a quantitative rationale for basing estimators on sufficient statistics:
if an estimator is not a function of a sufficient statistic, then there is another estimator
which is a function of the sufficient statistic and which is at least as good, in the
sense of mean squared error of estimation.
Examples 3.4
(a) X1 , . . . , Xn ∼ P (λ), λ to be estimated. P
In Example 2.3 (a) we saw that a sufficient statistic is i xi. Suppose we start
with the unbiased estimator λ̃ = X1 . Then ‘Rao–Blackwellization’ gives
P
λ∗ = E [X1 | i Xi = t] .
But
X  P  P P 
E Xi | X
i i = t = E X
i i | X
i i = t =t.
i
By the fact that X1, . . . , Xn are IID, every term within the sum on the l.h.s. must
be the same, and hence equal to t/n. Thus we recover the estimator λ∗ = λ̂ = X̄.
(b) X1 , . . . , Xn ∼ P (λ), θ = e−λ to be estimated.
Now θ = P(X1 = 0). So a simple unbiased estimator is θ̂ = 1{X1 = 0}. Then
" # !
X n Xn
 
θ∗ = E 1{X1 = 0}  Xi = t = P X1 = 0  Xi = t
i=1
! ! i=1
X
n . X
n
=P X1 = 0; Xi = t P Xi = t
i=2 i=1
((n − 1)λ) et −(n−1)λ . (nλ)t e−nλ  n − 1 t
= e−λ =
t! t! n
Since θ̂ is unbiased, so is θ∗ . As it should be, θ∗ is only a function of t. If you do
Rao-Blackwellization and you do not get just a function of t then you have made a
mistake.
(c) X1 , . . . , Xn ∼ U [0, θ], θ to be estimated.
In Example 2.3 (c) we saw that a sufficient statistic is maxi xi . Suppose we start
with the unbiased estimator θ̃ = 2X1. Rao–Blackwellization gives
 
1 n − 1 n+1
θ∗ = E [2X1 | maxi Xi = t] = 2 t+ (t/2) = t.
n n n
This is an unbiased estimator of θ. In the above calculation we use the idea that
X1 = maxi Xi with probability 1/n, and if X1 is not the maximum then its expected
value is half the maximum. Note that the MLE θ̂ = maxi Xi is biased.

15
3.3 Consistency and asymptotic efficiency∗
Two further properties of maximum likelihood estimators are consistency and asymp-
totic efficiency. Suppose θ̂ is the MLE of θ.
To say that θ̂ is consistent means that

P(|θ̂ − θ| > ) → 0 as n → ∞ .

In Example 3.1 this is just the weak law of large numbers:

  
 X1 + · · · + Xn 
P  − p >  → 0 .
n

It can be shown that var(θ̃) ≥ 1/nI(θ) for any unbiased estimate θ̃, where 1/nI(θ)
is called the Cramer-Rao lower bound. To say that θ̂ is asymptotically efficient
means that

lim var(θ̂)/[1/nI(θ)] = 1 .
n→∞

The MLE is asymptotically efficient and so asymptotically of minimum variance.

3.4 Maximum likelihood and decision-making

We have seen that the MLE is a function of the sufficient statistic, asymptotically
unbiased, consistent and asymptotically efficient. These are nice properties. But
consider the following example.

Example 3.5 You and a friend have agreed to meet sometime just after 12 noon.
You have arrived at noon, have waited 5 minutes and your friend has not shown
up. You believe that either your friend will arrive at X minutes past 12, where you
believe X is exponentially distributed with an unknown parameter λ, λ > 0, or that
she has completely forgotten and will not show up at all. We can associate the later
event with the parameter value λ = 0. Then
Z ∞
P(data | λ) = P(you wait at least 5 minutes | λ) = λe−λt dt = e−5λ .
5

Thus the maximum likelihood estimator for λ is λ̂ = 0. If you base your decision as
to whether or not you should wait a bit longer only upon the maximum likelihood
estimator of λ, then you will estimate that your friend will never arrive and decide
not to wait. This argument holds even if you have only waited 1 second.
The above analysis is unsatisfactory because we have not modelled the costs of
either waiting in vain, or deciding not to wait but then having the friend turn up.

16
4 Confidence intervals
Statisticians do it with 95% confidence.

4.1 Interval estimation

Let a(X) and b(X) be two statistics satisfying a(X) ≤ b(X) for all X. Suppose that
on seeing the data X = x we make the inference a(x) ≤ θ ≤ b(x). Here [a(x), b(x)]
is called an interval estimate and [a(X), b(X)] is called an interval estimator.
Previous lectures were concerned with making a point estimate for θ. Now we
are being less precise. By giving up precision in our assertion about the value of θ
we gain confidence that our assertion is correct. Suppose
 
Pθ a(X) ≤ θ ≤ b(X) = γ,
 
where γ does not depend on θ. Then the random interval a(X), b(X) is called a
100γ% confidence interval for θ. Typically γ is 0.95 or 0.99, so that the probability
the interval contains θ is close to 1.
Given data x, we would call [a(x), b(x)] a ‘100γ% confidence interval for θ’. Notice
however, that θ is fixed, and therefore the interval either does or does not contain
the true θ. However, if we repeat the procedure of sampling and constructing a a
confidence interval many times, then our confidence interval will contain the true θ
100γ% of the time. The point to understand here is that it is the endpoints of the
confidence interval that are random variables, not the parameter θ.

Examples 4.1

(a) If X1 , . . . , Xn ∼ N(µ, σ 2) independently, with µ unknown and σ 2 known, then

√
X̄ ∼ N(µ, σ 2/n) and hence n(X̄ − µ)/σ ∼ N(0, 1) .

So if ξ and η are such that P(ξ ≤ N(0, 1) ≤ η) = γ, we have

 √ 
n(X̄ − µ)
P(µ,σ2 ) ξ ≤ ≤η =γ,
σ
which can be rewritten as
 
ησ ξσ
P(µ,σ2 ) X̄ − √ ≤ µ ≤ X̄ − √ .
n n
Note that the choice of ξ and η is not unique. However, it is natural to try to make
the length of the confidence interval as small as possible, so the symmetry of the
normal distribution implies that we should take ξ and η symmetric about 0.

17
 Hence for a 95% confidence interval we would take −ξ = η = 1.96, as Φ(1.96) =
0.975. The 95% confidence interval is
 
1.96σ 1.96σ
X̄ − √ , X̄ + √
n n

For a 99% confidence interval, 1.96 would be replaced by 2.58, as Φ(2.58) = 0.995.
(b) If X1 , . . . , Xn ∼ N(µ, σ 2) independently, with µ and σ 2 both unknown, then
√
n(X̄ − µ)
p ∼ tn−1,
SXX /(n − 1)

where tn−1 denotes the ‘Student’s t-distribution on n−1 degrees of freedom’ which
will be studied later. So if ξ and η are such that P(ξ ≤ tn−1 ≤ η) = γ, we have
√ !
n(X̄ − µ)
P(µ,σ2 ) ξ ≤ p ≤ η = γ,
{SXX /(n − 1)}

which can be rewritten as

 p p 
P(µ,σ2 ) X̄ − η SXX /n(n − 1) ≤ µ ≤ X̄ − ξ SXX /n(n − 1) = γ.

Again the choice of ξ and η is not unique, but it is natural to try to make the length
of the confidence interval as small as possible. The symmetry of the t-distribution
implies that we should choose ξ and η symmetrically about 0.

4.2 Opinion polls

Opinion polls are typically quoted as being accurate to ±3%. What does this mean
and how many people must be polled to attain this accuracy?
Suppose we are trying to estimate p, the proportion of people who support the
Labour party amongst a very large population. We interview n people and estimate
p from p̂ = n1 (X1 + · · · + Xn ), where Xi = 1 if the ith person supports Labour and
Xi = 0 otherwise. Then
p(1 − p) 1
E p̂ = p and var p̂ = ≤ ,
n 4n
where the inequality follows from the fact that p(1 − p) is maximized
by p = 12 .
Let us approximate the distribution of p̂(X) by N p, p(1 − p)/n . This is very
good for n more than about 20. Then we have that approximately
p
(p̂ − p)/ p(1 − p)/n ∼ N(0, 1).

18
So

P(p̂ − 0.03 ≤ p ≤ p̂ + 0.03)

!
0.03 p̂ − p 0.03
= P −p ≤p ≤p
p(1 − p)/n p(1 − p)/n p(1 − p)/n
 p   p 
≈ Φ 0.03 n/p(1 − p) − Φ −0.03 n/p(1 − p)
√ √
≥ Φ(0.03 4n) − Φ(−0.03 4n)
√
For this to be at least 0.95, we need 0.03 4n ≥ 1.96, or n ≥ 1068.
Opinion polls typically use a sample size of about 1,100.

Example 4.2 U.S. News and World Report (Dec 19, 1994) reported on a telephone
survey of 1,000 Americans, in which 59% said they believed the world would come
to an end, and of these 33% believed it would happen within a few years or decades.
Let us find a confidence interval for the proportion of Americans who believe the
end of the world in imminent. Firstly, p̂ = 0.59(0.33) = 0.195. The variance of
p̂ is p(1 − p)/590 which we estimate by (0.195)(0.805)/590
√ = 0.000266. Thus an
approximate 95% confidence interval is 0.195 ± 0.00266(1.96), or [0.163, 0.226].
Note that this is only approximately a 95% confidence interval. We have used the
normal approximation, and we have approximated p(1 − p) by p̂(1 − p̂). These are
both good approximations and this is therefore a very commonly used analysis.

Sampling from a small population*

For small populations the formula for the variance of p̂ depends on the total popula-
tion size N. E.g., if we are trying to estimate the proportion p of N = 200 students
in a lecture who support the Labour party and we take n = 200, so we sample them
all, then clearly var(p̂) = 0. If n = 190 the variance will be close to 0. In fact,
 
N − n p(1 − p)
var(p̂) = .
N −1 n

4.3 Constructing confidence intervals

The technique we have used in these examples is based upon finding some statistic
whose distribution does not depend on the unknown parameter θ. This can be done
when θ is a location parameter or scale parameter. In section 4.1 µ is an example
of a location parameter and σ is an example of a scale parameter. We saw that the
√
distribution of n(X̄ − µ)/σ does not depend on µ or σ.
In the following example we consider a scale parameter.

19
Example 4.3 Suppose that X1 , . . . , Xn are IID E(θ). Then
Yn P
f (x | θ) = θe−θxi = θn e−θ i xi
i=1
P
so T (X) = i Xi is sufficient for θ. Also, T ∼ gamma(n, θ) with pdf
fT (t) = θn tn−1e−θt /(n − 1)!, t > 0.
Consider S = 2θT . Now P(S ≤ s) = P(T ≤ s/2θ), so by differentiation with respect
to s, we find the density of S to be
1 θn (s/2θ)n−1e−θ(s/2θ) 1 sn−1(1/2)ne−s/2
fS (s) = fT (s/2θ) = = , s > 0.
2θ (n − 1)! 2θ (n − 1)!


So S = 2θT ∼ gamma n, 12 ≡ χ22n .
Suppose we want a 95% confidence interval for the mean, 1/θ. We can write
P(ξ ≤ 2T θ ≤ η) = P (2T /η ≤ 1/θ ≤ 2T /ξ) = F2n(ξ) − F2n (η) ,
where F2n is the cdf of a χ22n RV.
For example, if n = 10 we refer to tables for the χ220 distribution and pick ξ = 34.17
and η = 9.59, so that F20(ξ) = 0.975, F20(η) = 0.025 and F20(ξ) − F20(η) = 0.95.
Then a 95% confidence interval for 1/θ is
[2t/34.17 , 2t/9.59 ] .
Along the same lines, a confidence interval for σ can be constructed in the cir-
cumstances of Example 4.1 (b) by using fact that SXX /σ 2 ∼ χ2n−1 . E.g., if n = 21 a
95% confidence interval would be
hp p i
Sxx /34.17 , Sxx /9.59 .

4.4 A shortcoming of confidence intervals*

Confidence intervals are widely used, e..g, in reporting the results of surveys and
medical experiments. However, the procedure has the problem that it sometimes
fails to make the best interpretation of the data.


Example 4.4 Suppose X1 , X2 are two IID samples from U θ − 12 , θ + 12 . Then
11 11 1
P(min xi ≤ θ ≤ max xi ) = P(X1 ≤ θ ≤ X2) + P(X2 ≤ θ ≤ X1 ) = + = .
i i 22 22 2
So (mini xi, maxi xi ) is a 50% confidence interval for θ.
But suppose the data is x = (7.4, 8.0). Then we know θ > 8.0 − 0.5 = 7.5 and
θ < 7.4 + 0.5 = 7.9. Thus with certainty, θ ∈ (7.5, 7.9) ⊂ (7.4, 8.0), so we can be
100% certain, not 50% certain, that our confidence interval has captured θ. This
happens whenever maxi xi − mini xi > 12 .

20
5 Bayesian estimation
Bayesians probably do it.

5.1 Prior and posterior distributions

Bayesian statistics, (named after the Rev. Thomas Bayes, an amateur 18th century
mathematician), represents a different approach to statistical inference. Data are still
assumed to come from a distribution belonging to a known parametric family. How-
ever, whereas classical statistics considers the parameters to be fixed but unknown,
the Bayesian approach treats them as random variables in their own right. Prior
beliefs about θ are represented by the prior distribution, with a prior probability
density (or mass) function, p(θ). The posterior distribution has posterior density
(or mass) function, p(θ | x1 , . . . , xn), and captures our beliefs about θ after they have
been modified in the light of the observed data.
By Bayes’ celebrated formula,

f (x1, . . . , xn | θ)p(θ)
p(θ | x1, . . . , xn) = R .
f (x1, . . . , xn | φ)p(φ) dφ

The denominator of the above equation does not involve θ and so in practice is
usually not calculated. Bayes’ rule is often just written,

p(θ | x1, . . . , xn) ∝ p(θ)f (x1, . . . , xn | θ).

Example 5.1 Consider the Smarties example addressed in Example 2.1 (a) and
suppose our prior belief is that the number of colours is either 5, 6, 7 or 8, with prior
probabilities 1/10, 3/10, 3/10 and 3/10 respectively. On seeing the data x =‘red,
green, red’ we have f (x | k) = (k − 1)/k 2. Similarly, if the fourth Smartie is orange,
f (x | k) = (k − 1)(k − 2)/k 3. Then

x = ‘red, green, red’ x = ‘red, green, red, orange’

k p(k) f (x | k) p(k)f (x | k) p(k | x) k p(k) f (x | k) p(k)f (x | k) p(k | x)
5 .1 .160 .016 .13 5 .1 .096 .010 .11
6 .3 .139 .042 .33 6 .3 .093 .028 .31
7 .3 .122 .037 .29 7 .3 .087 .026 .30
8 .3 .109 .033 .26 8 .3 .082 .025 .28

There is very little modification of the prior. This analysis reveals, in a way that
the maximum likelihood approach did not, that the data obtained from looking at
just 4 Smarties is not very informative. However, as we sample more Smarties the
posterior distribution will come to concentrate on the true value of k.

21
5.2 Conditional pdfs
The discrete case

Thus Bayesians statistics relies on calculation of conditional distributions. For two

events A and B (measurable subsets of the sample space) with P(B) 6= 0, we define
P(A | B) := P(A ∩ B)/P(B).

We can harmlessly agree to define P(A | B) := 0 if P(B) = 0.

If X and Y are RVs with values in Z, and if fX,Y is their joint pmf:
P(X = x; Y = y) = fX,Y (x, y),

then we define
P(X = x; Y = y) fX,Y (x, y)
fX|Y (x | y) := P(X = x | Y = y) = =
P(Y = y) fY (y)
if fY (y) 6= 0. We can safely define fX|Y (x | y) := 0 if fY (y) = 0. Of course,
X X
fY (y) = P(Y = y) = P(X = x; Y = y) = fX,Y (x, y).
x x

Example 5.2 Suppose that X and R are independent RVs, where X is Poisson with
parameter λ and R is Poisson with parameter µ. Let Y = X + R.
Then
 X
λx e−λ µy−x e−µ λxe−λ µr e−µ
fX|Y (x | y) =
x! (y − x)! x,r:x+r=y x! r!
 X
y! x (y−x) y! x r
= λ µ λ µ
x!(y − x)! x,r:x+r=y
x!r!
  x  y−x
y λ µ
= .
x λ+µ λ+µ
Hence (X | Y = y) ∼ B(y, p), where p = λ/(λ + µ).

The continuous case

Let Z = (X, Y ) be a RV with values in R m+n , X having values in R m and Y values

in R n . Assume that Z has nice pdf fZ (z) and write
fZ (z) = fX,Y (x, y), (z = (x, y), x ∈ R m , y ∈ R n ).
Then the pdf of Y is given by
Z
fY (y) = fX,Y (x, y) dx.
Rm
22
We define fX|Y , the conditional pdf of X given Y , by
(
fX,Y (x, y)/fY (y) if fY (y) 6= 0,
fX|Y (x | y) :=
0 if fY (y) = 0.

The intuitive idea is: P(X ∈ dx | Y ∈ dy) = P(X ∈ dx; Y ∈ dy)/P(Y ∈ dy).

Examples 5.3
(a) A biased coin is tossed n times. Let xi be 1 or 0 as the ith toss is or is not a
head. Suppose we have no idea how biased the coin is, so we place a uniform prior
distribution on θ, to give a so-called ‘noninformative prior’ of

p(θ) = 1, 0 ≤ θ ≤ 1.

Let t be the number of heads. Then the posterior distribution of θ is

Z 1
p(θ | x1, . . . , xn) = θ (1 − θ) × 1
t n−t
φt (1 − φ)n−t × 1 dφ .
0

We would usually not bother with the denominator and just write

p(θ | x) ∝ θt (1 − θ)n−t .

By inspection we recognise that if the appropriate constant of of proportionality

is inserted on the r.h.s. then we have the density of beta(t + 1, n − t + 1), so this is
the posterior distribution of θ given x.
(b) Suppose X1 , . . . , Xn ∼ N(µ, 1), p(µ) ∼ N(0, τ −2) for known τ −2. Then
( !)
1 X n
p(µ | x1, . . . , xn) ∝ exp − (xi − µ)2 + µ2 τ 2
2 i=1
(  P 2 )
1 xi
∝ exp − (n + τ 2) µ −
2 n + τ2
 P 
xi 1
µ | x 1 , . . . , xn ∼ N ,
n + τ2 n + τ2
Note that as τ → 0 the prior distribution becomes less informative.
(c) Suppose X1 , . . . , Xn ∼ IID E(λ), and the prior for λ is given by λ ∼ E(µ), for
fixed and known µ. Then
Y Pn
p(λ | x1, . . . , xn) ∝ µe−λµ λe−λxi = λn e−λ(µ+ i=1 xi ) ,
i
P
i.e., gamma(n + 1, µ + xi ).

23
5.3 Estimation within Bayesian statistics
The Bayesian approach to the parameter estimation problem is to use a loss func-
tion L(θ, a) to measure the loss incurred by estimating the value of a parameter to
be a when its true value is θ. Then θ̂ is chosen to minimize E [L(θ, θ̂)], where this
expectation is taken over θ with respect to the posterior distribution p(θ | x).

Loss functions for quadratic and absolute error loss

(a) L(θ, a) = (a − θ)2 is the quadratic error loss function.
Z Z
E [L(θ, a)] = L(θ, a)p(θ | x1, . . . , xn) dθ = (a − θ)2p(θ | x1 , . . . , xn) dθ .

Differentiating with respect to a we get

Z Z
2 (a − θ)p(θ | x1 , . . . , xn) dθ = 0 =⇒ a = θp(θ | x1, . . . , xn) dθ .

Therefore quadratic error loss is minimized by taking θ̂ to be the posterior mean.

(b) L(θ, a) = |a − θ| is the absolute error loss function.
Z
E [L(θ, a)] = L(θ, a)p(θ | x1 , . . . , xn) dθ
Z a Z ∞
= (a − θ)p(θ | x1, . . . , xn) dθ + (θ − a)p(θ | x1, . . . , xn) dθ .
θ=−∞ a
Differentiating with respect to a we find that the minimum is where
Z a Z ∞
p(θ | x1, . . . , xn) dθ − p(θ | x1, . . . , xn) dθ = 0 .
−∞ a

The minimum is achieved when both integrals are equal to 12 , i.e., by taking θ̂ to be
the posterior median.
Example 5.4 Let X1 , . . . , Xn ∼ P (λ), λ ∼ E(1) so that p(λ) = e−λ, λ ≥ 0.
The posterior distribution is
Yn
e−λλxi P
−λ −λ(n+1)
p(λ | x1, . . . , xn) = e ∝e λ xi ,
i=1
xi !
P

i.e., gamma xi + 1, (n + 1) . So under quadratic error loss,
Pn
xi + 1
θ̂ = posterior mean = i=1 .
n+1
Under absolute error loss, θ̂ solves
Z θ̂ −λ(n+1) P xi P
e λ (n + 1) xi +1 1
P dλ = .
0 ( xi)! 2

24
6 Hypothesis testing
Statistics is the only profession which demands the right to make mistakes 5 per
cent of the time – Thomas Huxley.

6.1 The Neyman–Pearson framework

The second major area of statistical inference is hypothesis testing. A statistical
hypothesis is an assertion or conjecture about the distribution of one or more ran-
dom variables, and a test of a statistical hypothesis is a rule or procedure for deciding
whether to reject that assertion.

Example 6.1 It has been suggested that dying people may be able to postpone
their death until after an important occasion. In a study of 1919 people with Jewish
surnames it was found that 922 occurred in the week before Passover and 997 in the
week after. Is there any evidence in this data to reject the hypothesis that a person
is as likely to die in the week before as in the week after Passover?

Example 6.2 In one of his experiments, Mendel crossed 556 smooth, yellow male
peas with wrinkled, green female peas. Here is what he obtained and its comparison
with predictions based on genetic theory.
type observed predicted expected
count frequency count
smooth yellow 315 9/16 312.75
smooth green 108 3/16 104.25
wrinkled yellow 102 3/16 104.25
wrinkled green 31 1/16 34.75
Is there any evidence in this data to reject the hypothesis that theory is correct?
We follow here an approach developed by Neyman and Pearson. Suppose we have
data x = (x1, x2, . . . , xn) from a density f . We have two hypotheses about f . On
the basis of the data one is accepted, the other rejected. The two hypotheses have
different philosophical status. The first, called the null hypothesis, and denoted
by H0, is a conservative hypothesis, not to be rejected unless evidence is clear. The
second, the alternative hypothesis, denoted by H1 , specifies the kind of departure
from the null hypothesis of interest to us.
It is often assumed that f belongs to a specified parametric family f (· | θ) indexed
by a parameter θ ∈ Θ (e.g. N(θ, 1), B(n, θ)). We might then want to test a parametric
hypothesis

H0 : θ ∈ Θ0 against H1 : θ ∈ Θ1

25
with Θ0 ∩ Θ1 = ∅. We may, or may not, have Θ0 ∪ Θ1 = Θ.
We will usually be concerned with testing a parametric hypothesis of this kind,
but alternatively, we may wish to test

H0 : f = f0 against H1 : f 6= f0

where f0 is a specified density. This is a ‘goodness-of-fit’ test.

A third alternative is that we wish to test

H0 : f = f0 against H1 : f = f1

where f0 and f1 are specified, but do not necessarily belong to the same family.

6.2 Terminology
A hypothesis which specifies f completely is called simple, e.g., θ = θ0. Otherwise,
a hypothesis is composite, e.g., θ > θ0.
Suppose we wish to test H0 against H1 . A test is defined by a critical region C.
We write C̄ for the complement of C.
(
C then H0 is rejected, and
If x = (x1, x2, . . . , xn) ∈
C̄ then H0 is accepted (not rejected).

Note that when x ∈ C̄ we might sometimes prefer to say ‘not rejected’, rather
than ‘accepted’. This is a minor point which need not worry us, except to note
that sometimes ‘not rejected’ does more accurately express what we are doing: i.e.,
looking to see if the data provides any evidence to reject the null hypothesis. If it
does not, then we might want to consider other things before finally ‘accepting H0’.
There are two possible types of error we might make:
H0 might be rejected when it is true (a type I error), or
H0 might be accepted when it is false (a type II error).
Since H0 is conservative, a type I error is generally considered to be ‘more serious’
that a type II error. For example, the jury in a murder trial should take as its null
hypothesis that the accused is innocent, since the type I error (that an innocent
person is convicted and the true murderer is never caught) is more serious than the
type II error (that a murderer is acquitted).
Hence, we fix (an upper bound on) the probability of type I error, e.g., 0.05 or
0.01, and define the critical region C by minimizing the type II error subject to this.
If H0 is simple, Θ0 = {θ0}, the probability of a type I error is called the size,
or significance level, of the test. If H0 is composite, the size of the test is α =
supθ∈Θ0 P(X ∈ C | θ). .

26
 The likelihood of a simple hypothesis H : θ = θ∗ given data x is
Lx (H) = fX (x | θ = θ∗ ).
If H is composite, H : θ ∈ Θ, we define
Lx(H) = sup fX (x | θ).
θ∈Θ
The likelihood ratio for two hypotheses H0 , H1 is
Lx(H1 )
Lx(H0, H1) = .
Lx(H0 )
Notice that if T (x) is a sufficient statistic for θ then by the factorization criterion
Lx (H0, H1) is simply a function of T (x).

6.3 Likelihood ratio tests

A test given by a critical region C of the form C = {x : Lx(H0, H1) > k}, for some
constant k, is called a likelihood ratio test. The value of k is determined by fixing
the size α of the test, so that P(X ∈ C | H0 ) = α.
Likelihood ratio tests are optimal for simple hypotheses. Most standard tests are
likelihood ratio tests, though tests can be built from others statistics.
Lemma 6.3 (Neyman–Pearson Lemma) H0 : f = f0 is to be tested against H1 :
f = f1 . Assume that f0 and f1 are > 0 on the same regions and continuous.
Then, among all tests of size ≤ α, the test with smallest probability of type II error
is given by C = {x : f1(x)/f0(x) > k}, where k is determined by
Z
α = P(X ∈ C | H0 ) = f0(x) dx.
C
Proof. Consider any test with size ≤ α, i.e., with a critical region D such that
P(X ∈ D | H0) ≤ α. Define

1 ∈
φD (x) = as x D
0 6∈
and let C and k be defined as above. Note that



0 ≤ φC (x) − φD (x) f1(x) − kf0(x) , for all x.
since this is always the product of two terms with the same sign. Hence
Z



0≤ φC (x) − φD (x) f1(x) − kf0(x) dx
x  
= P(X ∈ C | H1 ) − P(X ∈ D | H1) − k P(X ∈ C | H0 ) − P(X ∈ D | H0)
 
= P(X ∈ C | H1 ) − P(X ∈ D | H1) − k α − P(X ∈ D | H0 )
≤ P(X ∈ C | H1 ) − P(X ∈ D | H1)
This implies P(X 6∈ C | H1 ) ≤ P(X 6∈ D | H1 ) as required.

27
6.4 Single sample: testing a given mean, simple alternative, known vari-
ance (z-test)
Let X1 , . . . , Xn be IID N(µ, σ 2 ), where σ 2 is known. We wish to test H0 : µ = µ0
against H1 : µ = µ1 , where µ1 > µ0 .
The Neyman–Pearson test is to reject H0 if the likelihood ratio is large (i.e., greater
than some k). The likelihood ratio is
 Pn 
f (x | µ1 , σ 2) (2πσ 2)−n/2 exp − i=1(xi − µ1 )2/2σ 2
= P
f (x | µ0 , σ 2) (2πσ 2 )−n/2 exp [− ni=1(xi − µ0 )2/2σ 2]
" n #
X
= exp (xi − µ0 )2 − (xi − µ1 )2 /2σ 2
 i=1
 
= exp n 2x̄(µ1 − µ0 ) + (µ20 − µ21 ) /2σ 2
It often turns out, as here, that the likelihood ratio is a monotone function of a
sufficient statistic and we can immediately rephrase the critical region in more con-
venient terms. We notice that the likelihood ratio above is increasing in the sufficient
statistic x̄ (since µ1 − µ0 > 0). So the Neyman–Pearson test is equivalent to ‘reject
H0 if x̄ > c’, where we choose c so that P(X̄ > c | H0) = α. There is no need to try
to write c in terms of k.
However, under H0 the distribution of X̄ is N(µ0 , σ 2/n). This means that
√
Z = n(X̄ − µ0 )/σ ∼ N(0, 1) .
It is now convenient to rephrase the test in terms of Z, so that a test of size α is to
reject H0 if z > zα , where zα = Φ−1(1 − α) is the ‘upper α point of N(0, 1)’ i.e., the
point such that P(N(0, 1) > zα ) = α. E.g., for α = 0.05 we would take zα = 1.645,
since 5% of the standard normal distribution lies to the right of 1.645.
Because we reject H0 only if z is in the upper tail of the normal distribution we
call this a one-tailed test. We shall see other tests in which H0 is rejected if the
test statistic lies in either of two tails. Such a test is called a two-tailed test.
Example 6.4 Suppose X1 , . . . , Xn are IID N(µ, σ 2) as above, and we want a test
of size 0.05 of H0 : µ = 5 against H1 : µ = 6, with σ 2 = 1. Suppose the data is
x = (5.1, 5.5, 4.9, 5.3). Then x̄ = 5.2 and z = 2(5.2 − 5)/1 = 0.4. Since this is less
than 1.645 we do not reject µ = 5.
Suppose the hypotheses are reversed, so that we test H0 : µ = 6 against H1 : µ = 5.
The test statistic is now z = 2(5.2 − 6)/1 = −1.6 and we should reject H0 for values
of Z less than −1.645. Since z is more than −1.645, it is not significant and we do
not reject µ = 6.
This example demonstrates the preferential position given to H0 and therefore
that it is important to choose H0 in a way that makes sense in the context of the
decision problem with which the statistical analysis is concerned.

28
7 Further aspects of hypothesis testing

Statisticians do it with only a 5% chance of being rejected.

7.1 The p-value of an observation

The significance level of a test is another name for α, the size of the test. For a
composite hypothesis H0 : θ ∈ Θ0 and rejection region C this is

α = sup P(X ∈ C | θ) .
θ∈Θ0

For a likelihood ratio test the p-value of an observation x is defined to be

p∗ = sup Pθ LX (H0, H1) ≥ Lx(H0 , H1) .
θ∈Θ0

The p-value of x is the probability under H0 of seeing x or something at least as

‘extreme’, in the sense of containing at least as much evidence against H0 . E.g., in
Example 6.4, the p-value is p∗ where

p∗ = P(Z > z | µ = µ0 ) = 1 − Φ(z) .

A test of size α rejects H0 if and only if α ≥ p∗ . Thus the p-value of x is the

smallest value of α for which H0 would be rejected on the basis of seeing x. It is often
more informative to report the p-value than merely to report whether or not the null
hypothesis has been rejected.
The p-value is also sometimes called the significance level of x. Historically, the
term arises from the practice of ‘significance testing’. Here, we begin with an H0
and a test statistic T , which need not be the likelihood ratio statistic, for which, say,
large positive values suggest falsity of H0 . We observe value t0 for the statistic: the
significance level is P(T ≥ t0 | H0). If this probability is small, H0 is rejected.

7.2 The power of a test

For a parametric hypothesis about θ ∈ Θ, we define the power function of the test
specified by the critical region C as

W (θ) = P(X ∈ C | θ).

Notice that α = supθ∈Θ0 W (θ), and 1 − W (θ) = P(X ∈ C̄ | θ) = P(type II error | θ)

for θ ∈ Θ1 .

29
7.3 Uniformly most powerful tests
For the test of H0 : µ = µ0 against H1 : µ = µ1 , µ1 > µ0 described in section 6.4 the
critical region turned out to be
 √
C(µ0) = x : n(x̄ − µ0 )/σ > αz .
This depends on µ0 but not on the specific value of µ1 . The test with this critical
region would be optimal for any alternative H1 : µ = µ1 , provided µ1 > µ0 . This is
the idea of a uniformly most powerful (UMP) test.
We can find the power function of this test.
W (µ) = P(Z > αz | µ)
√

= P n(X̄ − µ0 )/σ > αz | µ
√ √

= P n(X̄ − µ)/σ + n(µ − µ0 /σ > αz | µ)
√

= 1 − Φ αz − n(µ − µ0 )/σ
Note that W (µ) increases from 0 to 1 as µ goes from −∞ to ∞ and W (µ0) = α.
More generally, suppose H0 : θ ∈ Θ0 is to be tested against H1 : θ ∈ Θ1, where
Θ0 ∩ Θ1 = ∅. Suppose H1 is composite. H0 can be simple or composite. We want a
test of size α. So we require W (θ) ≤ α for all θ ∈ Θ0 , W (θ0) = α for some θ0 ∈ Θ0 .
A uniformly most powerful (UMP) test of size α satisfies (i) it is of size α, (ii)
W (θ) is as large as possible for every θ ∈ Θ1 .
UMP tests may not exist. However, likelihood ratio tests are often UMP.
Example 7.1 Let X1, . . . , Xn be IID N(µ, σ 2 ), where µ is known. Suppose H0 :
σ 2 ≤ 1 is to be tested against H1 : σ 2 > 1.
We begin by finding the most powerful test for testing H00 : σ 2 = σ02 against
H10 : σ 2 = σ12, where σ02 ≤ 1 < σ12. The Neyman–Pearson test rejects H00 for large
values of the likelihood ratio:


−n/2  P 
f x | µ, σ12 2πσ12 exp − ni=1(xi − µ)2 /2σ12

=
−n/2  P 
f x | µ, σ02 2πσ02 exp − ni=1(xi − µ)2 /2σ02
" X n
#
1 1
= (σ0/σ1)n exp 2 − 2 (xi − µ)2
2σ0 2σ1 i=1
P
which is large when i (xi − µ)2 is large. If σ 2 = 1 then
X
n
(xi − µ)2 ∼ χ2n .
i=1
P (n) (n)
So a test of the form ‘reject H0 if T :=
 i (x i − µ) 2
> F
α ’, has size α where Fα
(n)
is the upper α point of χn . That is, P T > Fα | σ ≤ 1 ≤ α, for all σ 2 ≤ 1, with
2 2

30
equality for σ 2 = 1. But this test doesn’t depend on the value of σ12 , and hence is
the UMP test of H0 against H1.

Example 7.2 Consider Example 6.1. Let p be the probability that if a death occurs
in one of the two weeks either side of Passover it actually occurs in the week after
the Passover. Let us test H0 : p = 0.5 vs. H1 : p > 0.5.
The distribution of the number of deaths in the week after
Passover, say X,
is B(n, p), which we can approximate by N np, np(1 − p) ; under
H0 this is
√
N(0.5n, 0.25n). √ So a size 0.05 test is to reject H0 if z = n x̄ − µ0 /σ > 1.645,


where here z = 1919 997/1919 − 0.5 /0.5 = 1.712. So the data is just significant
at the 5% level. We reject the hypothesis that death p = 1/2.
It is important to realise that this does not say anything about why this might
be. It might be because people really are able to postpone their deaths to enjoy
the holiday. But it might also be that deaths increase after Passover because of
over-eating or stress during the holiday.

7.4 Confidence intervals and hypothesis tests

There is an interesting duality between confidence intervals and hypothesis tests.
In the following, we speak of the acceptance region, i.e., the complement of the
critical (or rejection) region C.

Theorem 7.3

(i) Suppose that for every θ0 there is a size α test of H0 : θ = θ0 against some
alternative. Denote the acceptance region by A(θ0). Then I(X) = {θ : X ∈
A(θ)} is a 100(1 − α)% confidence interval for θ.
(ii) Conversely, if I(X) is a 100(1−α)% confidence interval for θ then an acceptance
region for a size α test of H0 : θ = θ0 is A(θ0) = {X : θ0 ∈ I(X)}.

Proof. The definitions in the theorem statement give

P X ∈ A(θ0) | θ = θ0 = P θ ∈ I(X) | θ = θ0 .

By assumption the l.h.s. is 1 − α in case (i) and the r.h.s. is 1 − α in case (ii).
This duality can be useful. In some circumstances it can be easier to see what
is the form of a hypothesis test and then work out a confidence interval. In other
circumstances it may be easier to see the form of the confidence interval.
In Example 4.1 we saw that a 95% confidence interval for µ based upon X1 , . . . , Xn
being IID samples from N(µ, σ 2), σ 2 known, is
 
1.96σ 1.96σ
X̄ − √ , X̄ + √ .
n n
31
Thus H0 : µ = µ0 is rejected in a 5% level test against H1 : µ 6= µ0 if and only if µ0
is not in this interval; i.e., if and only if
√  
nX̄ − µ0 /σ > 1.96 .

7.5 The Bayesian perspective on hypothesis testing

Suppose a statistician wishes to test the hypothesis that a coin is fair (H0 : p = 1/2)
against the alternative that it biased towards heads with probability p, (H1 : p =
p1 > 1/2). He decides to toss it five times. It is easy to see that the best test is to
reject H0 if the total number of heads, say T , is large. Here T ∼ B(5, p). Suppose
he observes H,H,H,H,T, so T = 4. The p-value (or significance level) of this result is
the probability that under the null hypothesis he should see a result which is equally
or more extreme, i.e.,
P(T = 4 or 5) = 5(0.54)(0.5) + 0.55 = 6/32 = 0.1875 .
Another statistician wishes also wishes to test H0 against H1, but with a different
experiment. He plans to toss the coin until he gets a tail. Now the best test is
to reject H0 if the number of tosses, say N, is large. Here N ∼ geometric(1 − p).
Suppose this statistician also observes H,H,H,H,T, so N = 5. He figures the p-value
as the probability of seeing this result or one that is even more extreme and obtains
P(N ≥ 5) = 0.54 = 1/16 = 0.0625 .
Thus the two statisticians come to different conclusions about the significance
of what they have seen. This is disturbing! The coin knew nothing about the
experimental procedure. Maybe there were four tosses simply because that was the
point at which the experimenter spilt his coffee and so decided to stop tossing the
coin. What then is the significance level?
The Bayesian perspective on hypothesis testing avoids this type of problem. Sup-
pose the experimenter places prior probabilities of P(H0 ) and P(H1 ) on the truth of
two mutually exclusive hypotheses. Having observed the data x, these are modified
into posterior probabilities in the usual way, i.e.,
P(H1 | x) P(x | H1 ) P(H1 ) P(H1 )
= = Lx(H0 , H1) .
P(H0 | x) P(x | H0 ) P(H0 ) P(H0 )
Thus the ratio of the posterior probabilities is just the ratio of the prior probabilities
multiplied by the likelihood ratio.
Under both procedures above x = {H,H,H,H,T}, and the likelihood ratios is
p1 p1 p1 p1 (1 − p1 ) p41 (1 − p1)
Lx(H0, H1 ) = × × × × = .
0.5 0.5 0.5 0.5 0.5 0.55
In general, the Bayesian analysis does not depend on how the data was obtained.

32
8 Generalized likelihood ratio tests
An approximate answer to the right problem is worth a good deal more than an
exact answer to an approximate problem.

8.1 The χ2 distribution

This distribution plays a huge role in Statistics. For n a positive integer, the χ2n
distribution is the distribution of
X12 + · · · + Xn2 , where X1, . . . , Xn are IID each N(0, 1).
It is not difficult to show that


χ2n = gamma 1 1
2 n, 2 ,
with pdf
.
f (t) = ( 12 )n/2tn/2−1e−t/2 Γ(n/2), t > 0.

We speak of the ‘chi-squared distribution with (or on) n degrees of freedom’. If

X ∼ χ2n , E (X) = n, var(X) = 2n.

8.2 Generalised likelihood ratio tests

Tests with critical regions of the form C = {x : Lx(H0 , H1) > k} are intuitively
sensible, and in certain cases optimal. But such a critical region may not reduce
to a region depending only on the value of a simple function T (x). Even if it does,
the distribution of the statistic T (X), required to determine k, may not be simple.
However, the following theorem allows us to use a likelihood ratio test even in this
case. We must take care in describing the circumstances for which the theorem is
valid.
So far we have considered disjoint alternatives, but if our real interest is in testing
H0 and we are not interested in any specific alternative it is simpler to take (in the
parametric framework) Θ1 = Θ, rather than Θ1 = Θ \ Θ0 .
So we now suppose we are testing H0 : θ ∈ Θ0 against H1 : θ ∈ Θ, i.e., a null
hypothesis which restricts θ against a general alternative.
Suppose that with this formulation Θ0 imposes p independent restrictions on θ,
so that, for example, we have Θ = {θ : θ = (θ1, . . . , θk )} and
H0 : θi1 = α1 , . . . , θip = αp , for given αj ; or
H0 : Aθ = b, for given Ap×k , bp×1 ; or
H0 : θi = θi(φ1 , . . . , φk−p), i = 1, . . . , k, for given θ1 (·), . . . , θk (·)
and φ1 , . . . , φk−p to be estimated.

33
Θ1 has k free parameters and Θ0 has k − p free parameters. We write |Θ1 | = k and
|Θ0 | = k − p. Then we have the following theorem (not to be proved.)
Theorem 8.1 Suppose Θ0 ⊂ Θ1 and |Θ1 |−|Θ0 | = p. Then under certain conditions,
as n → ∞ with X = (X1, . . . , Xn ) and Xi IID,
2 log LX (H0, H1) ∼ χ2p ,
if H0 is true. If H0 is not

true, 2 log LX tends to be larger. We reject H0 if 2 log Lx >
c, where α = P χp > c to give a test of size approximately α.
2

We say that 2 log LX (H0, H1) is asymptotically distributed as χ2p . The conditions
required by the theorem hold in all the circumstances we shall meet in this course.
Lemma 8.2 Suppose X1 , . . . , Xn are IID N(µ, σ 2). Then
 P 
(i) maxµ f (x | µ, σ 2) = (2πσ 2 )−n/2 exp − i (xi − x̄)2 /2σ 2 .
h P i
2 −n/2
i (xi −µ)
(ii) maxσ2 f (x | µ, σ ) = 2π
2
n exp [−n/2].
h P i−n/2
i (xi −x̄)
2
(iii) maxµ,σ2 f (x | µ, σ ) = 2π
2
n exp [−n/2].

8.3 Single sample: testing a given mean, known variance (z-test)

Let X1 , . . . , Xn be IID N(µ, σ 2 ), where σ 2 is known. We wish to test H0 : µ = µ0
against H1 : µ 6= µ0 . The generalized likelihood ratio test suggests that we should
reject H0 if Lx(H0 , H1) is large, where


supµ f x | µ, σ 2
Lx (H0, H1) =

f x | µ0 , σ 2

−n/2  P 
2πσ 2 exp − i (xi − x̄)2/2σ 2
=
−n/2  P 
2πσ 2 exp − i (xi − µ0 )2/2σ 2
" #
X n

= exp (1/2σ 2) (xi − µ0 )2 − (xi − x̄)2
 i=1

= exp (1/2σ )n(x̄ − µ0 )2
2

That is, we should reject H0 if (x̄ − µ0 )2 is large.

This is no surprise. For under H0 , X̄ ∼ N(µ0 , σ 2/n), so that
√
Z = n(X̄ − µ0 )/σ ∼ N(0, 1),
and a test of size α is to reject H0 if z > zα/2 or if z < −zα/2 , where zα/2 is the ‘upper
α/2 point of N(0, 1)’ i.e., the point such that P(N(0, 1) > zα/2 ) = α/2. This is an
example of a two-tailed test.

34
 Note that 2 log LX (H0, H1) = Z 2 ∼ χ21 . In this example H0 imposes p = 1
constraint on the parameter space and the approximation in Theorem 8.1 is exact.

8.4 Single sample: testing a given variance, known mean (χ2 -test)
As above, let X1, . . . , Xn be IID N(µ, σ 2), where µ is known. We wish to test
H0 : σ 2 = σ02 against H1 : σ 2 6= σ02 . The generalized likelihood ratio test suggests
that we should reject H0 if Lx (H0, H1) is large, where
h P i−n/2

2π i (xi −µ)2
supσ2 f x | µ, σ 2
n
exp [−n/2]
Lx (H0, H1) =
=  P .
f x | µ, σ02 (2πσ02)−n/2 exp − i (xi − µ)2/2σ02
P
If we let t = i(xi − µ)2 /nσ02 we find
2 log Lx (H0, H1) = n(t − 1 − log t) ,
which increases as t increases from 1 and t decreases from 1. Thus we should reject
H0 when the difference of t and 1 is large.
Again, this is not surprising, for under H0,
Xn
T = (Xi − µ)2 /σ02 ∼ χ2n .
i=1
(n) (n)
So a test of size α is the two-tailed test which rejects H0 if t > Fα/2 or t < F1−α/2
(n) (n)
where F1−α/2 and Fα/2 are the lower and upper α/2 points of χ2n , i.e., the points such
   
(n) (n)
that P χn < F1−α/2 = P χn > Fα/2 = α/2.
2 2

8.5 Two samples: testing equality of means, known common variance

(z-test)
Let X1, . . . , Xm be IID N(µ1 , σ 2) and let Y1, . . . , Yn be IID N(µ2, σ 2), and suppose
that the two samples are independent. It is required to test H0 : µ1 = µ2 against
H1 : µ1 6= µ2 . The likelihood ratio test is based on



supµ1 ,µ2 f x | µ1 , σ 2 f y | µ2 , σ 2
Lx(H0, H1) =


supµ f x | µ, σ 2 f y | µ, σ 2
 P   P 
(2πσ 2 )−(m+n)/2 exp − i (xi − x̄)2/2σ 2 exp − i(yi − ȳ)2/2σ 2
= h P
i h P
i
mx̄+nȳ 2 mx̄+nȳ 2
(2πσ ) 2 −(m+n)/2 exp − i xi − m+n /2σ exp − i yi − m+n /2σ
2 2
"  2  2 #
m mx̄ + nȳ n mx̄ + nȳ
= exp x̄ − + ȳ −
2σ 2 m+n 2σ 2 m+n
 
1 mn
= exp 2
(x̄ − ȳ)2
2σ (m + n)

35
 So we should reject H0 if |x̄ − ȳ| is large. Now, X̄ ∼ N(µ1 , σ 2/m) and Ȳ ∼
N(µ2 , σ 2/n), and the samples are independent, so that, on H0 ,
  
1 1
X̄ − Ȳ ∼ N 0, σ 2 +
m n
or
 − 1
1 1 2 1
Z = (X̄ − Ȳ ) + ∼ N(0, 1).
m n σ
A size α test is the two-tailed test which rejects H0 if z > zα/2 or if z < −zα/2 , where
zα/2 is, as in 8.3 the upper α/2 point of N(0, 1). Note that 2 log LX (H0 , H1) = Z 2 ∼
χ21 , so that for this case the approximation in Theorem 8.1 is again exact.

8.6 Goodness-of-fit tests

Suppose we observe n independent trials and note the numbersP of times that each
of k possible outcomes occurs, i.e., (x1, . . . , xk ), with j xj = n. Let pi be the
probability of outcome i. On the basis of this data we want to test the hypothesis
that p1, . . . , pk take particular values. We allow that these values might depend on
some unknown parameter θ (or parameters if θ is a vector). I.e., we want to test
H0 : pi = pi(θ) for θ ∈ Θ0 against H1 : pi are unrestricted.
For example, H0 might be the hypothesis that the trials are samples
from ak−ibinomial
distribution B(k, θ), so that under H0 we would have pi (θ) = i θ (1 − θ) . k i

This is called a goodness-of-fit test, because we are testing whether our data
fit a particular distribution (in the above example the binomial distribution).
The distribution of (x1, . . . , xk ) is the multinomial distribution
n!
P(x1 , . . . , xk | p) = px1 1 · · · pxkk ,
x1 ! · · · xk !
Pk
for (x1, . . . , xk ) s.t. xi ∈ {0, . . . , n} and i=1 xi = n. Then we have
( k  )
X  Xk

sup log f (x) = const + sup xi log pi  0 ≤ pi ≤ 1, pi = 1 .
H1 
i=1 i=1
P P
Now, i xi log pi may be maximised subject to i pi = 1 by a Lagrangian technique
and we get p̂i = xi /n. Likewise,
( k )
X
sup log f (x) = const + sup xi log pi(θ) .
H0 θ i=1

The generalized likelihood tells us to reject H0 if 2Lx(H0, H1) is large compared to

the chi-squared distribution with d.f. |Θ1 | − |Θ0 |. Here |Θ1 | = k − 1 and |Θ0 | is the
number of independent parameters to be estimated under H0 .

36
9 Chi-squared tests of categorical data
A statistician is someone who refuses to play the national lottery,
but who does eat British beef. (anonymous)

9.1 Pearson’s chi-squared statistic

Suppose, as in Section 8.6, that we observe x1 , . . . , xk , the numbers of times that
each of k possible outcomes occurs in n independent trials, and seek to make the
goodness-of-fit test of

H0 : pi = pi(θ) for θ ∈ Θ0 against H1 : pi are unrestricted.

Recall
X
k X
k X
k


2 log Lx(H0 , H1) = 2 xi log p̂i − 2 xi log pi (θ̂) = 2 xi log p̂i /pi(θ̂) ,
i=1 i=1 i=1

where p̂i = xi/n and θ̂ is the MLE of θ under H0. Let oi = xi denote the number
of time that outcome i occurred and let ei = npi (θ̂) denote the expected number of
times it would occur under H0 . It is usual to display the data in k cells, writing oi
in cell i. Let δi = oi − ei . Then
X
k


2 log Lx (H0, H1) = 2 xi log (xi/n)/pi(θ̂)
i=1
Xk
=2 oi log(oi /ei)
i=1
Xk
=2 (δi + ei ) log(1 + δi /ei)
i=1
Xk
=2 (δi + ei )(δi/ei − δi2 /2e2i + · · ·)
i=1
X
k
+ δi2 /ei
i=1
Xk
(oi − ei )2
= (1)
i=1
ei

This is called the Pearson chi-squared statistic.

For H0 we have to choose θ. Suppose the optimization over θ has p degrees of
freedom. For H1 we have k − 1 parameters to choose. So the difference of these

37
degrees of freedom is k − p − 1. Thus, if H0 is true the statistic (1) ∼ χ2k−p−1
approximately. A mnemonic for the d.f. is
d.f. = #(cells) − #(parameters estimated) −1. (2)
Note that
X
k k  2
X  X
k X
k
(oi − ei )2 o o2 o2
= i
− 2oi + ei = i
− 2n + n = i
− n. (3)
i=1
ei i=1
ei i=1
ei i=1
ei

Sometimes (3) is easier to compute than (1).

Example 9.1 For the data from Mendel’s experiment, the test statistic has the value
0.618. This is to be compared to χ23 , for which the 10% and 95% points are 0.584
and 7.81. Thus we certainly do not reject the theoretical model. Indeed, we would
expect the observed counts to show even greater disparity from the theoretical model
about 90% of the time.
Similar analysis has been made of many of Mendel’s other experiments. The data
and theory turn out to be too close for comfort. Current thinking is that Mendel’s
theory is right but that his data were massaged by somebody (Fisher thought it was
Mendel’s gardening assistant) to improve its agreement with the theory.

9.2 χ2 test of homogeneity

Suppose we have a rectangular array of cells with m rows and n columns, with Xij
items in the (i, j) th cell of the array. Denote the row, column and overall sums by
X
n X
m X
m X
n
Xi· = Xij , X·j = Xij , X·· = Xij .
j=1 i=1 i=1 j=1

Suppose the row sums are fixed and the distribution of (Xi1, . . . , Xin ) in row i is
multinomial with probabilities (pi1, . . . , pin), independently of the other rows. We
want to test the hypothesis that the distribution in each row is the same, i.e., H0 : pij
is the same for all i, (= pj ) say, for each j = 1, . . . , n. The alternative hypothesis is
H1 : pij are unrestricted. We have
XX
log f (x) = const + xij log pij , so that
i j
(  )
X
m X
n  X
n

sup log f (x) = const + sup xij log pij  0 ≤ pij ≤ 1, pij = 1 ∀i
H1 
i=1 j=1 j=1
P P
Now, x log p may be maximized subject to j pij = 1 by a Lagrangian tech-
j ij ij
P  P
nique. The maximum of j xij log pij + λ 1 − j pij occurs when xij /pij = λ,

38
 P
∀j. Then P the constraints give λ = j xij and the corresponding maximizing pij is
p̂ij = xij / j xij = xij /xi·. Hence,
X
m X
n
sup log f (x) = const + xij log(xij /xi·).
H1 i=1 j=1

Likewise,
(  )
XX  X

sup logf (x) = const + sup xij log pj  0 ≤ pj ≤ 1, pj = 1 ,
H0 
i j j
XX
= const + xij log(x·j /x··).
i j

Here p̂j = x·j /x··. Let oij = xij and write eij = p̂j xi· = (x·j /x··)xi· for the expected
number of items in position (i, j) under H0 . As before, let δij = oij − eij . Then,

XX
2 log Lx(H0, H1) = 2 xij log(xij x·· /xi·x·j )
i j
XX
=2 oij log(oij /eij )
i j
XX
=2 (δij + eij ) log(1 + δij /eij )
i j
XX
+ δij2 /eij
i j
XX
= (oij − eij )2/eij . (4)
i j

For H0, we have (n−1) parameters to choose, for H1 we have m(n−1) parameters
to choose, so the degrees of freedom is (n − 1)(m − 1). Thus, if H0 is true the
statistic (4) ∼ χ2(n−1)(m−1) approximately.

Example 9.2 The observed (and expected) counts for the study about aspirin and
heart attacks described in Example 1.2 are
Heart attack No heart attack Total
Aspirin 104 (146.52) 10,933 (10890.5) 11,037
Placebo 189 (146.48) 10,845 (10887.5) 11,034
Total 293 21,778 22,071
293


E.g., e11 = 22071 11037 = 146.52. The χ2 statistic is
(104−146.52)2 (189−146.48)2 2
(10845−10887.5)2
146.52 + 46.48 + (10933−10890.5)
10890.5 + 10887.5 = 25.01 .

39
The 95% point of χ21 is 3.84. Since 25.01 > 3.84, we reject the hypothesis that heart
attack rate is independent of whether the subject did or did not take aspirin.
Note that if there had been only a tenth as many subjects, but the same percent-
ages in each in cell, the statistic would have been 2.501 and not significant.

9.3 χ2 test of row and column independence

This χ2 test is similar to that of Section 9.2, but the hypotheses are different. Again,
observations are classified into a m × n rectangular array of cells, commonly called a
contingency table. The null hypothesis is that the row into which an observation
falls is independent of the column into which it falls.
Example 9.3 A researcher pretended to drop pencils in a lift and observed whether
the other occupant helped to pick them up.
Helped Did not help Total
Men 370 (337.171) 950 (982.829) 1,320
Women 300 (332.829) 1,003 (970.171) 1,303
Total 670 1,953 2,623
To test the independence of rows and columns we take
X X
H0 : pij = piqj with 0 ≤ pi , qj ≤ 1, pi = 1, qj = 1 ;
i j
X
H1 : pij arbitrary s.t. 0 ≤ pij ≤ 1, pij = 1 .
i,j

The same approach as previously gives MLEs under H0 and H1 of

p̂i = xi·/x··, q̂j = x·j /x··,
eij = p̂i q̂j x·· = (xi·x·j /x··), and p̂ij = xij /x·· .
P
The test statistic can again be show to be about ij (oij − eij )2/eij . The eij are

670

shown in parentheses in the table. E.g., e11 = p̂1q̂1 n = 1320 2623 2623 2623 = 337.171.
The number of free parameters under H1 and H0 are mn − 1 and (m − 1) + (n − 1)
respectively. The difference of these is (m−1)(n−1), so the statistic is to be compared
to χ2(m−1)(n−1). For the data above this is 8.642, which is significant compared to χ21 .

We have now seen Pearson χ2 tests in three different settings. Such a test is
appropriate whenever the data can be viewed as numbers of times that certain out-
comes have occurred and we wish to test a hypothesis H0 about the probabilities
with which they occur. Any unknown parameter is estimated by maximizing the
likelihood function that pertains under H0 and ei is computed as the expected num-
ber of times outcome i occurs if that parameter is replaced by this MLE value. The
statistic is (1), where the sum is computed over all cells. The d.f. is given by (2).

40
10 Distributions of the sample mean and variance

Statisticians do it. After all, it’s only normal.

10.1 Simpson’s paradox

Example 10.1 These are some Cambridge admissions statistics for 1996.

Women Men
applied accepted % applied accepted %
Computer Science 26 7 27 228 58 25
Economics 240 63 26 512 112 22
Engineering 164 52 32 972 252 26
Medicine 416 99 24 578 140 24
Veterinary medicine 338 53 16 180 22 12
Total 1184 274 23 2470 584 24

In all five subjects women have an equal or better success rate in applications than
do men. However, taken overall, 24% of men are successful but only 23% of women
are successful! This is called Simpson’s paradox (though it was actually discovered
by Yule 50 years earlier). It can often be found in real data. Of course it is not a
paradox. The explanation here is that women are more successful in each subject,
but tend to apply more for subjects that are hardest to get into (e.g., Veterinary
medicine). This example should be taken as a warning that pooling contingency
tables can produce spurious associations. The correct interpretation of this data is
that, for these five subjects, women are significantly more successful in gaining entry
than are men.
In order to produce an example of Simpson’s paradox I carefully selected five
subjects from tables of 1996 admissions statistics. Such ‘data snooping’ is cheating; a
similar table that reversed the roles of men and women could probably be constructed
by picking different subjects.

10.2 Transformation of variables

The rest of this lecture is aimed Pat proving2 some important facts about distribution
of the statistics X̄ and SXX = i(Xi − X̄) , when X1 , . . . , Xn are IID N(µ, σ 2). We
begin by reviewing some ideas about transforming random variables.
Suppose the joint density of X1, . . . , Xn is fX , and there is a 1–1 mapping between
X1 , . . . , Xn and Y1, . . . , Yn such that Xi = xi (Y1, . . . , Yn). Then the joint density of

41
Y1 , . . . , Yn is
 ∂x (y) 
 1 ∂x1 (y) 
 ∂y1 · · · ∂yn 
 .. 
fY (y1 , . . . , yn ) = fX (x1(y), . . . , xn(y))  ... . 
 ∂xn (y) ∂xn (y) 
 ···
∂y1 ∂yn

where the Jacobian := J(y1, . . . , yn) is the absolute value of the determinant of the
matrix (∂xi(y)/∂yj ).
The following example is an important one, which also tells us more about the
beta distribution.
Example 10.2 Let X1 ∼ gamma(n1, λ) and X2 ∼ gamma(n2, λ), independently.
Let Y1 = X1 /(X1 + X2 ), Y2 = X1 + X2 . Since X1 and X2 are independent we
multiply their pdfs to get
λn1 xn1 1 −1 −λx1 λn2 xn2 2−1 −λx2
fX (x) = e × e .
(n1 − 1)! (n2 − 1)!
Then x1 = y1y2 , x2 = y2 − y1 y2 , so
 
 ∂x1 (y) ∂x1 (y)   

 ∂y1 ∂y2  y y
J(y1, y2) =  ∂x2 (y) ∂x2 (y)  =  2 1  = y2
 ∂y1
 ∂y2
−y2 1 − y1 
Hence making the appropriate substitutions and arranging terms we get
(n1 + n2 − 1)! n1 −1 n2 −1 λn1 +n2 y2n1 +n2−1 −λy2
fY (y) = y (1 − y1 ) × e
(n1 − 1)!(n2 − 1)! 1 (n1 + n2 − 1)!
from which it follows that Y1 and Y2 are independent RVs (since their joint den-
sity function factors into marginal density functions) and Y1 ∼ beta(n1, n2), Y2 ∼
gamma(n1 + n2 , λ).

10.3 Orthogonal transformations of normal variates

Lemma 10.3 Let X1 , . . . , Xn , be independently distributed with distributions
N(µi , σ 2) respectively. Let A = (aij ) be an orthogonal matrix, so that A>A =
AA> = I. Then
the elements of Y = >AX are independently distributed, and
Yi ∼ N (Aµ)i, σ , where µ = (µ1 , . . . , µn ) .
2

Proof. The joint density of X1 , . . . , Xn is

Y
fX (x1, . . . , xn | µ, σ ) =
2
fXi (xi | µi , σ 2)
i
1 P
− i (xi −µi )2 /2σ2
= e
(2πσ 2)n/2
1 >
e−(x−µ) (x−µ)/2σ
2
= 2 n/2
(2πσ )

42
Since x = A>y, we have ∂xi/∂yj = aji and hence J(y1, . . . , yn) = | det(A>)| = 1.
Thus
1  > > >

fY (y1, . . . , yn | µ, σ 2) = exp −(A y − µ) (A y − µ)/2σ 2
(2πσ 2 )n/2
1  > > > > >

= exp −(A y − A Aµ) (A y − A Aµ)/2σ 2
(2πσ 2 )n/2
1  > >

= exp −(y − Aµ) AA (y − Aµ)/2σ 2
(2πσ 2 )n/2
1  >

= exp −(y − Aµ) (y − Aµ)/2σ 2
(2πσ 2 )n/2

Remark. An alternative proof can be given using moment generating functions.

For θ ∈ R n , the mgf of the joint distribution is
   
E exp θ>Y = E exp θ>AX
 
= E exp (A>θ)>X
 
= exp (A>θ)>µ + 12 σ 2 (A>θ)>(A>θ)
 
= exp θ> Aµ + 12 σ 2 θ>θ


which we recognise as the mgf of independent RVs with distributions N (Aµ)i, σ 2 .

10.4 The distributions of X̄ and SXX

Pn
Lemma 10.4 Let X1 , . . . , Xn be IID N(µ, σ 2) and let X̄ = n−1
P i=1 Xi , SXX =
n
i=1 (Xi − X̄) . Then:
2

(i) X̄ ∼ N(µ, σ 2/n) and n(X̄ − µ)2 ∼ σ 2 χ21 .

P
(ii) Xi − µ ∼ N(0, σ 2), so ni=1(Xi − µ)2 ∼ σ 2χ2n .
P
(iii) ni=1(Xi − µ)2 = SXX + n(X̄ − µ)2.

(iv) SXX /(n − 1) is an unbiased estimator of σ 2 .

(v) X̄ and SXX are independent.

(vi) SXX ∼ σ 2 χ2n−1.

Proof.
(i) and (ii) are immediate from the fact that linear combinations of normal RVs
are normally distributed and the definition of χ2n . To prove (iii) and (iv) we note

43
that
X
n X
n

2
(Xi − µ) =
2
[Xi − X̄] + [X̄ − µ]
i=1 i=1
Xn


= [Xi − X̄]2 + 2[Xi − X̄][X̄ − µ] + [X̄ − µ]2
i=1
= SXX + n[X̄ − µ]2
Let A be an orthogonal matrix such that
√

Y = A X − µ1) = n(X̄ − µ), Y2, . . . , Yn .
I.e., we take
√ √ √ 
1/ n 1/ n · · · 1/ n
A =  ... ..
.
.. 
.
· · ··· ·
where the rows below the first are chosen to make the matrix orthogonal. P Then
√ n
Y1 = n(X̄ − µ) ∼ N(0, σ 2) and Y1 is independent of Y2, . . . , Yn. Since i=1 Yi2 =
P
i (Xi − µ) , we must have
2

X
n X
n
Yi2 = (Xi − µ)2 − n(X̄ − µ)2 = SXX .
i=2 i=1

Hence SXX and Y1 (and equivalently P

SXX and X̄) are independent. This gives (v).
Finally, (vi) follows from SXX = ni=2 Yi2 and the fact that Y2, . . . , Yn are IID
N(0, σ 2).

10.5 Student’s t-distribution

If X ∼ N(0, 1), Y ∼ χ2n , independently of X, then
1
Z = X/(Y /n) 2 ∼ tn ,
where tn is the Student’s t-distribution with (or on) n degrees of freedom. Like the
normal distribution, this distribution is symmetric about 0, and bell-shaped, but has
more probability in its tails, i.e., for all t > 0, P(Z > t) > P(X > t).
√
From Lemma 10.4 we have n(X̄ − µ) ∼ σN(0, 1) and SXX ∼ σ 2 χ2n−1, indepen-
dently. So from these and the definition of the t-distribution follows the important
fact that if X1 , . . . , Xn are IID N(µ, σ 2 ), then
√
n(X̄ − µ)
p ∼ tn−1 .
SXX /(n − 1)

44
11 The t-test
Statisticians do it with two-tail T tests.

11.1 Confidence interval for the mean, unknown variance

Suppose X1 , . . . , Xn IID N(µ, σ 2), but now σ 2 is unknown. Recall
√
n(X̄ − µ)
T = ∼ tn−1 .
σ̂
where σ̂ 2 = SXX /(n − 1). A 100(1 − α)% confidence interval for µ follows from
 √   
(n−1) n( X̄ − µ) (n−1) tα/2 σ̂ tα/2 σ̂
1 − α = P −tα/2 ≤ ≤ tα/2 = P X̄ − √ ≤ µ ≤ X̄ + √
σ̂ n n
(n−1)
where tα/2 is the ‘upper α/2 point of a t-distribution on n − 1 degrees of freedom’,
 
(n−1)
i.e., such that P T > tα/2 = α/2.

Example 11.1 In ‘Sexual activity and the lifespan of male fruitflies’, Nature, 1981,
Partridge and Farquhar report experiments which examined the cost of increased re-
production in terms of reduced longevity for male fruitflies. They kept numbers of
male flies under different conditions. 25 males in one group were each kept with 1
receptive virgin female. 25 males in another group were each kept with 1 female who
had recently mated. Such females will refuse to remate for several days. These served
as a control for any effect of competition with the male for food or space. The groups
were treated identically in number of anaesthetizations (using CO2) and provision of
fresh food.
To verify ‘compliance’ two days per week throughout the life of each experimental
male, the females that had been supplied as virgins to that male were kept and exam-
ined for fertile eggs. The insemination rate declined from approximately 1 per day at
age one week to about 0.6 per day at age eight weeks.
The data was as follows

Groups of 25 mean life s.e.

males kept with (days)
1 uninterested female 64.80 15.6525
1 interested female 56.76 14.9284

p Here s.e. is an abbreviation for standard error, i.e. the value of σ̂ =

Sxx /(n − 1). Here n = 25. The mean life, x̄ and the s.e., σ̂, are sufficient statistics
for (µ, σ 2), so there is nothing else we need to know about the individual values of

45
the the longevities of these 50 flies in order to compute confidence intervals or test
statistics.
From these summary statistics we can compute 95% confidence intervals for the
mean lives of the control and test groups to be
√ √
[64.80 − 2.06(15.6526)/ 25, 64.80 + 2.06(15.6526)/ 25] = [58.35, 71.25]
√ √
[56.76 − 2.06(14.9284)/ 25, 56.76 + 2.06(14.9284)/ 25] = [50.61, 62.91]
It is interesting to look at the data, and doing so helps us check that lifespan is
normally distributed about a mean. The longevities for control and test groups were
42 42 46 46 46 48 50 56 58 58 63 65 65 70 70 70 70 72 72 76 76 80 90 92 97
21 36 40 40 44 48 48 48 48 53 54 56 56 60 60 60 60 65 68 68 68 75 81 81 81

0 10 20 30 40 50 60 70 80 90 100

11.2 Single sample: testing a given mean, unknown variance (t-test)

Suppose that with the same assumptions as above it is required to test H0 : µ = µ0
against H1 : µ 6= µ0 .
Adopting the paradigm of the generalized likelihood ratio test we consider


maxµ,σ2 f x | µ, σ 2
Lx(H0, H1 ) =

maxσ2 f x | µ0 , σ 2
 P −n/2
2π i (xi − x̄)2 /n exp [−n/2]
= P
[2π i (xi − µ0 )2/n]−n/2 exp [−n/2]
P 
2 n/2
(x − µ )
= Pi
i 0

i (xi − x̄)
2
P n/2
i (xi −P
x̄)2 + n(x̄ − µ0 )2
=
i (xi − x̄)
2
 n/2
n(x̄ − µ0 )2
= 1+ P .
i (xi − x̄)
2
P
This is large when T 2 := n(n − 1)(x̄ − µ0 )2 i (xi − x̄) is large, equivalently when
2

|T | is large. Under H0 we have T ∼ tn−1. So a size α test is the two-tailed test which
(n−1) (n−1)
rejects H0 if t > tα/2 or if t < −tα/2 .
Example 11.2 Does jogging lead to a reduction in pulse rate? Eight non-jogging
volunteers engaged in a one-month jogging programme. Their pulses were taken before
and after the programme.

46
 pulse rate before 74 86 98 102 78 84 79 70
pulse rate after 70 85 90 110 71 80 69 74
decrease 4 1 8 -8 7 4 10 -4
Although there are two sets of data it is really just the changes that matter. Let the
decreases in pulse rates be x1, . . . , x8 and assume these are samples from N(µ, σ 2)
for some unknown σ 2 . To test H0 : µ = 0 against H1 : µ 6= 0 we compute
X X X
xi = 22, x̄ = 2.75, 2
xi = 326, Sxx = x2i − 8x̄2 = 265.5.

Hence the test statistic is

√
8(2.75 − 0)
t= p = 1.263,
265.5/(8 − 1)
(7)
which is to be compared to t0.025 = 2.365. Hence the data is not sufficient to reject
H0 at the 5% level. This may surprise you since 6 of the 8 subjects had lowered pulse
rates. This sort of test is called a paired samples t-test.

11.3 Two samples: testing equality of means, unknown common variance

(t-test)
We have the same samples as in 8.5, i.e., X1 , . . . , Xm are IID N(µ1 , σ 2) and Y1, . . . , Yn
are IID N(µ2 , σ 2). These two samples are independent. It is required to test H0 :
µ1 = µ2 against H1 : µ1 6= µ2 , but now σ 2 is unknown. Note how this differs from
the paired t-test above: the samples are not paired, and can be of unequal sizes.
As above, a maximum likelihood approach could convince us that the test should
be of the form ‘reject H0 if (x̄ − ȳ)2/(Sxx + Syy ) is large.
As in 8.5 we have that under H0 ,
 − 1
1 1 2 1
(X̄ − Ȳ ) + ∼ N(0, 1) .
m n σ
P Pn
If SXX = m i=1 (Xi − X̄) 2
, and S Y Y = j=1 (Yj − Ȳ ) , then
2

(SXX + SY Y )/σ 2 ∼ χ2m+n−2

so that (since X̄ is independent of SXX , Ȳ is independent of SY Y and the two samples

are independent)
s  
1 1 SXX + SY Y
T = (X̄ − Ȳ ) + ∼ tm+n−2.
m n m+n−2
(m+n−2) (m+n−2)
A test of size α rejects H0 if t > tα/2 or if t < −tα/2 .

47
Example 11.3 For the fruitfly data we might test H0 : that mean longevity is the
same for males living with 8 interested females as with 8 uninterested females. The
test statistic is
s  
1 1 24(15.6525) + 24(14.9284)
t = (64.80 − 56.76) + = 1.859
25 25 25 + 25 − 2
(48)
which can be compared to t0.025 = 2.01, and therefore is not significant at the 0.05%
(48)
level. H0 is not rejected. (It is however, significant at the 10% level, since t0.05 =
1.68).
Similarly, we can give a 95% confidence interval for the difference of the means.
This has endpoints
s  
1 1 24(15.6525) + 24(14.9284)
(64.80 − 56.76) ± 2.01 +
25 25 25 + 25 − 2
= 8.04 ± 8.695.
I.e., a 95% confidence interval for the extra longevity of celibate males is
[−0.655, 16.735] days. Notice again that finding we cannot reject µ1 − µ2 = 0 at
the 5% level is equivalent to finding that the 95% confidence interval for the differ-
ence of the means contains 0.
In making the above test we have assumed that the variances for the two popula-
tions are the same. In the next lecture we will see how we might test that hypothesis.

11.4 Single sample: testing a given variance, unknown mean (χ2 -test)
Let X1 , . . . , Xn be IID N(µ, σ 2), and suppose we wish to test H0 : σ 2 = σ02 against
H1 : σ 2 6= σ02, where µ is unknown, and therefore a ‘nuisance parameter’.
Following Theorem 8.1, the likelihood ratio is
h P i
2 −n/2

i (xi −x̄)
maxµ,σ2 f x | µ, σ 2 2π n
exp [−n/2]
Lx(H0, H1) =
= −n/2 P
maxµ f x | µ, σ02 (2πσ02 ) exp [−(1/2σ02) i(xi − x̄)2]
P
As in Section 8.4 this is large when i (xi − x̄)/nσ02 (= Sxx /nσ02) differs substantially
from 1.
Under H0, SXX /σ02 ∼ χ2n−1. Given the required size of test α, let a1 , a2 be such
that
P(SXX /σ02 < a1 ) + P(SXX /σ02 > a2 ) = α
under H0 . Then a size α test is to reject H0 if Sxx /σ02 < a1 or if Sxx /σ02 > a2 .
−1 −1
Usually we would take a1 = Fn−1 (α/2), a2 = Fn−1 (1 − α/2), where Fn−1 is the
2
distribution function of a χn−1 random variable.

48
12 The F -test and analysis of variance
The statistician’s attitude to variation is like that of the evangelist to sin;
he sees it everywhere to a greater or lesser extent.

12.1 F -distribution
If X ∼ χ2m and Y ∼ χ2n , independently of X, then

Z = (X/m)/(Y /n) ∼ Fm,n,

has the F -distribution with (or on) m and n degrees of freedom.

Note that if T ∼ Fm,n then 1/T ∼ Fn,m . Tables for the F -distribution usually only
give the upper percentage points. If we want to know x such that P(T < x) = 0.05
we can use Fn,m tables to find 1/x such that P(1/T > 1/x) = 0.05.
Note that if X ∼ tn then X 2 ∼ F1,n. It is always nice to recognise connections
between distributions.

12.2 Two samples: comparison of variances (F -test)

Suppose X1 , . . . , Xm are IID N(µ1 , σ12) and Y1 , . . . , Yn are IID N(µ2 , σ22), with the
two samples independent. It is required to test H0 : σ12 = σ22 against H1 : σ12 > σ22 ,
with µ1 , µ2 unknown nuisance parameters.
Now, by either the generalized likelihood ratio test, or common sense, we are led
to consider the statistic
σ̂12 SXX /(m − 1) σ12 χ2m−1/(m − 1) σ12
F = 2= ∼ 2 2 = 2 Fm−1,n−1.
σ̂2 SY Y /(n − 1) σ2 χn−1/(n − 1) σ2

Thus, under H0 , F ∼ Fm−1,n−1.

If H1 is true, F will tend to be greater than when H0 is true, so we reject H0 if
(m−1,n−1)
this ratio is large. A size α test is to reject H0 if f > Fα , the upper α point
of Fm−1,n−1.

Example 12.1 Suppose we wish to test the hypothesis that the variance of longevity
is the same for male fruitflies kept with 1 interested or 1 uninterested female, i.e.,
H0 : σ12 = σ22 against H0 : σ12 6= σ22 .The test statistic is

f = (15.6525)2/(14.9284)2 = 1.099,
(24,24)
which, as F0.05 = 1.98, is not significant at the 10% level (the test is two-tailed).
Notice that in order to use F tables we put the larger of σ̂12 and σ̂22 in the numerator.

49
12.3 Non-central χ2
Pk 2
If X1 , . . . , Xk are independent N(µi , 1) then Z = i=1 Xi has thePnon-central
chi-squared distribution, χ2k (λ), with non-centrality parameter λ = ki=1 µ2i . Note
that EW = k + λ; thus a non-central χ2k tends to be larger than a central χ2k .
To see that it is only the value of λ matters, let A be an orthogonal matrix
such
Pk that Aµ
P = (λ1/2, 0, . . . , 0)>, so (Aµ)>P
(Aµ) = µ> µ = λ. Let Y = AX; then
2 k 2 2 2 k 2 2
i=1 Xi = i=1 Yi , with Y1 = χ1 (λ) and i=2 Yi = χk−1 .

12.4 One way analysis of variance

Analysis of variance (ANOVA) is a technique for testing hypotheses about means
by looking at sample variances. We consider here the question of testing equality of
the means of k > 2 groups. The mathematical model is:
Xij = µi + ij , j = 1, . . . , ni,
i = 1, . . . , k.
P
Thus there are ni observations in the ith group. Let ki=1 ni = N.
It is assumed that the ij are IID N(0, σ 2), and that our data consists of observa-
tions xij which are realisations of random variables Xij satisfying the model.
One-way ANOVA is used to test the null hypothesis H0 : µ1 = . . . = µk . The
alternative hypothesis H1 is ‘H0 is not true’. Application of the generalized likelihood
ratio test gives
 P 
maxµ1 ,··· ,µk ,σ2 (2πσ 2)−N/2 exp − ij (xij − µi )2/2σ 2
Lx (H0, H1) =  P 
maxµ,σ2 (2πσ 2)−N/2 exp − ij (xij − µ)2 /2σ 2
 N/2 X X
s0
= , where s0 := (xij − x̄·· ) and s1 :=
2
(xij − x̄i·)2.
s1 ij ij
P P
Here, x̄·· = ij xij /N = i ni x̄i·/N is the overall mean (and the MLE of µ under
P i
H0 ). Similarly, x̄i· = nj=1 xij /ni is the mean within the ith group (and the MLE of
µi under H1 ).
Thus we are led to consider rejecting H0 when s0 /s1 is large. Now
X
k X
ni
s0 = (xij − x̄i· + x̄i· − x̄·· )2
i=1 j=1
XX 
= (xij − x̄i·)2 + 2(xij − x̄i·)(x̄i· − x̄··) + (x̄i· − x̄·· )2
i j
X X
= (xij − x̄i· ) +
2
ni(x̄i· − x̄·· )2
ij i
= s1 + s2 ,

50
 P
where s2 := i ni (x̄i· − x̄·· )2, and thus s0 /s1 is large when s2 /s1 is large.
s1 is called the within samples sum of squares and s2 is called the between
samples sum of squares. P
Now, whether or not H0 is true, j (Xij − X̄i· )2 ∼ σ 2χ2ni −1, since E(Xij ) depends
only on i.PHence, S1 ∼ σ 2χ2N −k , since samples for different i are independent.
Also, j (Xij − X̄i· )2 is independent of X̄i· , so that S1 is independent of S2 . If H0
is true S2 ∼ σ 2 χ2k−1, and if H0 is not true, S2 ∼ σ 2χ2k−1 (λ), where
P P
E(S2) = (k − 1)σ 2 + λ, λ = ki=1 ni (µi − µ̄)2, µ̄ = i niµi /N.
Intuitively, if H0 is not true S2 tends to be inflated.
So, if H0 is true then Q = {S2/(k − 1)}/{S1/(N − k)} ∼ Fk−1,N −k , while if H0 is
(k−1,N −k)
not true, Q tends to be larger. So for a size α test we reject H0 if q > Fα .
P An interpretation of this is that the variability in the total data set is s0 =
(x − x̄·· )2. Under H1 we expect xij to be about x̄i· and so a variability of s2 =
Pij ij
ij (x̄i· − x̄·· ) is ‘explained’ by H1 . Statisticians say that H1 ‘explains (s2 /s0 )100%
2

of the variation in the data’, (where since s0 = s1 + s2 , we must have s2/s0 ≤ 1.) If
s2 /s0 is near 1, or equivalently if s2 /s1 is large, then H1 does much better than H0
in explaining why the data has the variability it does.

Example 12.2 Partridge and Farquhar did experiments with five different groups of
25 male fruitflies. In addition to the groups kept with 1 interested or 1 uninteresed
female, 25 males were each kept with no companions, and groups of 25 were each
kept with 8 uninterested or 8 interested females. The ‘compliance’ of the males who
were supplied with 8 virgin females per day varied from 7 inseminations per day at
age one week to just under 2 per day at age eight weeks.
Groups of 25 mean life s.e.
males kept with (days)
no companions 63.56 16.4522
1 uninterested female 64.80 15.6525
1 interested female 56.76 14.9284
8 uninterested females 63.36 14.5398
8 interested females 38.72 12.1021

Suppose we wish to test equality of means in the three control groups, i.e., those
kept with either no companions, or 1 or 8 uninterested females (rows 1, 2 and 4).
First we reconstruct the sums of squares,
P25
j=1 (x1j − x̄1 ) = 24(16.4522 ) = 6496.16
2 2
P25
j=1 (x2j − x̄2 ) = 24(15.6525 ) = 5880.00
2 2
P25
j=1 (x4j − x̄4 ) = 24(14.5398 ) = 5073.76
2 2

51
then we calculate the within and between sums of squares,
x̄ = (63.56 + 64.80 + 63.36)/3 = 63.91
s1 = 6496.16 + 5880.00 + 5073.76 = 17449.92
X X
s2 = 25(x̄i· − x̄·· )2 = 25x̄2i· − 75x̄2·· = 30.427
i=1,2,4 i=1,2,4

and finally we compute the test statistic,

30.427/(3 − 1)
q= = 0.0628.
17449.92/(75 − 3)
It is usual to display this data in an ANOVA table of the following form.
Source of Degrees of Sum of Mean square F statistic
variation freedom squares
s2 /(k−1)
Between k−1 2 s2 30.427 s2 /(k − 1) 15.213 s1 /(N −k) 0.0628
groups
Within N −k 72 s1 17449.92 s1 /(N − k) 242.36
groups

Total N −1 74 s0 17480.35
(2,72)
The value of 0.0628 is not significant compared to F0.05 = 3.12 and hence we do
not reject the hypothesis of equal means.
A similar test for equality of all five group means gives a statistic with value 507.5,
(4,120)
to be compared to F0.05 = 2.45. Clearly we reject the hypothesis of equal means.
It does seem that sexual activity is associated with reduced longevity.
ANOVA can be carried out for many other experimental designs. We might want
to investigate more than one treatment possibility, or combinations of treatments.
(E.g., in the fruitfly experiments each male fly was kept separate from other males;
we might want to do experiments in which males are kept with different numbers
of interested females and/or competing males.) If there are k possible treatments
which can be applied or not applied, then 2k different combinations are possible and
this may be more than is realistic. The subject of ‘experimental design’ has to do
with deciding how to arrange the treatments so as to gather as much information
as possible from the fewest observations. The data is to be analysed to compare
treatment effects and this typically involves some sort of ANOVA. The methodology
is the same as for the one-way ANOVA considered above; we consider a normalised
quotient, such as q above, between the reduction in the residual sums of squares that
is obtained when moving from H0 to H1 (e.g., s0 − s1 ) and the value of the residual
sum of squares under H1 (e.g., s1 ). In subsequent lectures we will see further examples
of this idea in the context of regression models.

52
13 Linear regression and least squares
Numbers are like people; torture them enough and they’ll tell you anything.

13.1 Regression models

One of the most widely used examples of estimation in statistics is provided by linear
regression. For example, if Yi is the number of unemployed in UK in the ith month
after some date, we might make the hypothesis that
Yi = a + βi + i , i = 1, . . . , n,
where 1 , . . . , n are IID N(0, σ 2) and a, β are some unknown constants. The business
of estimating β is to do with detecting a trend in unemployment. A related problem
is that of testing H0 : β = 0, a test of whether there is any trend in unemployment.
The model above is a special case of the simple linear regression model in
which, with the same assumptions on {i }, a, β, σ 2 ,
Yi = a + βxi + i , i = 1, . . . , n,
where the xi are known constants. In the case above xi = i.
A multiple regression model has more that one explanatory variable on the
right hand side, e.g.,
Yi = a + β1 log i + β2 zi−5 + i , i = 1, . . . , n,
where perhaps zi−5 is the number of unemployed people who were in training pro-
grammes five months earlier. The estimation of a, β1 and β2 and associated tests are
similar to what we find for simple linear regression.

13.2 Least squares/MLE

Suppose Y1, . . . , Yn are independent and Yi = α + βwi + i , where i ∼ N(0, σ 2),
or equivalently that Yi ∼ N(α + βwi, σ 2), and wherePα, β and σ 2 are unknown
parameters and the wi are known constants such that i wi = 0.
Theorem 13.1 The MLEs of α and β are obtained by minimizing
X
n

2 X
n
S= Yi − E (Yi ) = (Yi − α − βwi)2
i=1 i=1

with respect to α and β. These are called the least squares estimators and are
given by:

α̂ = Ȳ and β̂ = SwY /Sww

P 2
P
where Sww = i wi , and SwY = i wi Yi .

53
Proof. Since Yi ∼ N(α + βwi, σ 2) the likelihood of of y1 , . . . , yn is
!
1 1 Xn
1 −S/2σ2
fY (y | µ, σ 2) = exp − (yi − α − βwi ) 2
= e .
(2πσ 2)n/2 2σ 2 i=1 (2πσ 2)n/2
The maximum likelihood estimator minimizes S, and so at a minimum,
 
∂S  Xn
∂S  Xn
= −2 (yi − α̂ − β̂wi ) = 0 , α=α̂ = −2 wi (yi − α̂ − β̂wi ) = 0.
∂α α=α̂ i=1
∂β i=1
β=β̂ β=β̂

Hence
X
n X
n X
n
Yi − nα̂ = 0 and wi Yi − β̂ wi2 = 0,
i=1 i=1 i=1
from which the answers follow.

13.3 Practical usage

Given a linear regression model
Yi = a + βxi + i ,
P
in which i xi 6= 0 we make the transformation wi = xi − x̄ and consider
Yi = α + βwi + i ,
P
where x̄ = ni=1 xi/n and α = a + β x̄. This gives the situation described in 13.1,
and we can use results of 13.2 to estimate the regression and the results in 14.1 to
perform tests. Making the necessary transformations we have
â = Ȳ − β̂ x̄ and β̂ = SxY /Sxx
P P
where Sxx = w> w = i (xi − x̄)2 and SxY = i (xi − x̄)(Yi − Ȳ ).
We speak of ‘regressing y on x’. A package such as MINITAB will return the
estimated regression line in the form
y = â + β̂x.
Note that the point (x̄, Ȳ ) always lies on the regression line, i.e., Ȳ = â + β̂ x̄.
Example 13.2 The following data for 40 nations has been extracted from a 1993
almanac.
The correlations of life expectancy with people/television and people/doctor are
−0.606 and −0.666 respectively. Scatter plots suggests that a better fit might be
obtained by a regression of life expectancy on either the logarithm of people/television
or logarithm of people/doctor. When this is done the correlations are respectively
−0.855 and −0.832.

54
 country mean life people per people per
expectancy, y television, u doctor, v
Argentina 70.5 4.0 370
Bangladesh 53.5 315.0 6166
Brazil 65.0 4.0 684
.. ..
. .
United Kingdom 76.0 3.0 611
United States 75.5 1.3 404
Venezuela 74.5 5.6 576
Vietnam 65.0 29.0 3096
Zaire 54.0 * 23193

80 80
life expectancy

life expectancy
o o

60 60

40 40

0 100 200 300 400 500 600 0 1 2 3

people per television log people per television

Let xi = log10 ui and consider fitting a regression of y against x. There is data for
38 countries (as television data for Zaire and Tanzania is missing). We compute the
following summary statistics

ȳ = 67.76, x̄ = 1.0322, Syy = 2252.37, Sxx = 17.120, Sxy = −167.917.

These give
1
β̂ = Sxy /Sxx = −9.808, α̂ = ȳ − β̂ x̄ = 77.887, r = Sxy /(SxxSyy ) 2 = −0.855.

Although people/television appears to be a useful predictor of a country’s life

expectancy, we don’t really expect that sending shiploads of televisions to countries
with short life expectancies would cause their people to live longer. This points up
the obvious, but sometimes forgotten fact, that there may be correlation between
two variables without causation.

13.4 Data sets with the same summary statistics

Tha application of regression analysis requires care. The following data sets have
nearly the same value of the sufficient statistics x̄, ȳ, Sxx and Sxy . The regression
line is about y = 300 + 50 x in each case. However, a simple linear regression is
only appropriate in the first case. In the second case a quadratic would be more

55
appropriate. The third case is affected by the presence of an outlier and the fourth
case is really no more than a straight line fit through 2 points. The lesson is: plot
the data!
1200 1200 1200 1200

1000 1000 1000 1000

800 800 800 800

600 600 600 600

400 400 400 400

200 200 200 200

0
0 2 4 6 8 10 12 14 16 18 20 0
0 2 4 6 8 10 12 14 16 18 20
0
0 2 4 6 8 10 12 14 16 18 20
0
0 2 4 6 8 10 12 14 16 18 20

10 804 8 695 13 758 9 881 11 833 14 996 6 724 4 426 12 1084 7 482 5 568
10 914 8 814 13 874 9 877 11 926 14 810 6 613 4 310 12 913 7 726 5 474
10 746 8 677 13 1274 9 711 11 781 14 884 6 608 4 539 12 815 7 642 5 573
8 658 8 576 8 771 8 884 8 847 8 704 8 525 19 1250 8 556 8 791 8 689

13.5 Other aspects of least squares

Least squares can be used to fit other models. For example, to fit a regression
through the origin we would minimize
P
S = i (Yi − βxi)2
P P
and get β̂ = i xiYi / i x2i . To fit a multiple regrssion model, such as
Yi = β0 + β1 xi1 + · · · + βp xip + i ,
in which y is predicted from p variables, x1 , . . . , xp, we would minimize
P
i (Yi − β0 + β1 xi1 + · · · + βp xip )
2

with respect to β0 , . . . , βp . Stationarity conditions give p + 1 simultaneous linear

equations in β̂0 , . . . , β̂p , which can be solved for these estimators.
Least squares estimators have many nice properties, one of which is that they
are best linear unbiased estimators. Consider simple linear regression, in which
Yi = a + βxi + i , where the i are independent with common mean 0 and variance
σ 2 (but are now not necessarily normal RVs). Suppose we want to estimate β by
P
a linear function of the observations, i.e., β̂ = i ci Yi . If this estimator is to be
unbiased then we need
P
E β̂ = i ci (a + βxi) = β, for all a, β.
P P
Hence we need i ci = 0 and i ci xi = 1. Now the variance of the estimator is
P P
var(β̂) = var ( i ci Yi ) = σ 2 i c2i .
So we have the constrained optimization problem:
P 2 P P
minimize c
i i subject to i ci = 0 and i ci xi = 1.
Using Lagrangian methods it is easy to find the solution: ci = (xi − x̄)/Sxx. This
gives the usual LSE β̂ = Sxy /Sxx . A similar analysis shows that the best linear
unbiased estimator of a is also the usual LSE of a, i.e., â = Ȳ − β̂ x̄.

56
14 Hypothesis tests in regression models
Statisticians do it with a little deviance.

14.1 Distributions of the least squares estimators

Suppose as before that Y1, . . . , Yn are independent and Yi = α + βwi + i , where
i ∼ N(0, σ 2), equivalently Yi ∼ N(α + βwi , σ 2), and where
P
α, β and σ 2 are unknown
parameters and the wi are known constants such that i wi = 0.

Theorem 14.1
(i) α̂ = Ȳ is distributed as N(α, σ 2/n);


(ii) β̂ is distributed as N β, (w>w)−1σ 2 independently of α̂;
(iii) the residual sum of squares R, the minimised value of S, is distributed as
σ 2 χ2n−2 independently of α̂ and β̂, and is equal to
X
R= Yj2 − nȲ 2 − (w>w)β̂ 2;

(iv) σ̂ 2 = R/(n − 2) is an unbiased estimator of σ 2 .

Proof. Let
 √ √ 
1/ n ··· 1/ n
 (w> w)−1/2w1 · · · (w> w)−1/2wn 
 
A= .. .. 
 . . 
· ··· ·

be an orthogonal matrix by appropriate choice of rows

3, . . . , n. Then Z1 , . . . , Zn
are independent with Z = AY ∼ N A(α1 + βw), σ I , and
2

√ √

Z1 = nα̂ ∼ N nα, σ 2


Z2 = (w> w)1/2β̂ ∼ N (w>w)1/2
β, σ 2
Z3 = · ∼ N 0, σ 2
.. ..
. .

Zn = · ∼ N 0, σ 2

from which all the statements in (i) and (ii) follow.

(iii) and (iv) follow from
X
n X
n
>
Zi2 2
= nȲ + (w w)β̂ +2
Zi2
i=1 i=3

57
and
X
n X
n
Zi2 = Yi2
i=1 i=1
= k(Y − α̂1 − β̂w) + α̂1 + β̂wk2
= kY − α̂1 − β̂wk2 + nα̂2 + β̂ 2kwk2
(since all cross-product terms vanish)
= R + nȲ 2 + (w>w)β̂ 2
Pn
So R = 3 Zi2 ∼ σ 2 χ2n−2 and is independent of Z1 and Z2, i.e., of α̂ and β̂.

14.2 Tests and confidence intervals

(a) A t-statistic may be constructed for testing the hypothesis that β takes a partic-
ular value β0 , since if β0 is the true value of β:
√ √
(β̂ − β0 ) w>w (β̂ − β0 ) w> w
T0 = p = ∼ tn−2.
R/(n − 2) σ̂

Therefore, to test H0 : β = β0 against H1 : β 6= β0, we compute t0 and reject H0

(n−2) (n−2)
in a test of size α if t0 > tα/2 or t0 < −tα/2 .
(b) A (1 − α)100% confidence interval may be found for β. Starting from the distri-
butional result in (a) above, we find similarly as in Section 11.1,
 √ √ 
(n−2) > (n−2) >
P β̂ − tα/2 σ̂/ w w < β < β̂ + tα/2 σ̂/ w w = 1 − α.

(c) We predict the value of Y that would be observed

at a given w0 by Ŷ = α̂ + β̂w0 .
−1
Then Y − Ŷ ∼ N 0, σ (1 + 1/n + w0 (w w) ) . Hence a (1 − α)100% predictive
2 2 2

confidence interval for Y at w0 is

 q q 
(n−2) (n−2)
Ŷ − tα/2 σ̂ 1 + 1/n + w02(w2w)−1 , Ŷ + tα/2 σ̂ 1 + 1/n + w02(w2w)−1

14.3 The correlation coefficient

1
The sample correlation coefficient of x and y is defined as r = Sxy /(SxxSyy ) 2 .
Suppose Yi ∼ N(a + βxi, σ 2), independently for each i. The hypothesis that Y does
not vary with x is H0 : β = 0. The test statistic in (a) can be rewritten as follows:
√ √ p √
β̂ Sxx (Sxy /Sxx) Sxx (n − 2) n−2r
t0 = p = q = √ ,
R/(n − 2) Syy − Sxy
2 /S
xx
1 − r 2

58
and so H0 should be rejected if r2 is near 1. P
Note that the variation in the data is Syy = j (yj − ȳ)2. The regression model
P P
‘explains’ variation of j (ŷj − ȳ)2 where ŷi = â + β̂xi . One can check that j (ŷj −
ȳ)2 = Sxy
2
/Sxx and so the ratio of these is r2 . We say that ‘the regression explains
100r2% of the variation in the data’.

14.4 Testing linearity

If we are fitting a linear regression, how can we know how good the fit of the estimated
regression line is? In general we cannot be sure: a bad fit could quite well be caused
by a large value of σ 2, which is unknown. We can test linearity if we are able to
replicate the readings, so that, say, we take m readings at each value xi, and get,

Yij = a + βxi + ij , j = 1, . . . , m,

for each i = 1, . . . , n. Then averaging over j for fixed i we have

Ȳi = a + βxi + ηi = α + β(xi − x̄) + ηi

P
where the ηi are IID N(0, σ 2/m), independently of m j=1 (Yij − Ȳi ) , which are IID
2

σ 2 χ2m−1. Now, if we do a linear regression of Ȳi on xi , the residual sum of squares is

X
n

2 σ2 2
Ȳi − α̂ − β̂(xi − x̄) ∼ χn−2,
i=1
m

if the means are indeed linearly related. Thus to test linearity we consider
P
2
m ni=1 Ȳi − α̂ − β̂(xi − x̄) /(n − 2)
F = Pn Pm ∼ Fn−2,n(m−1),
j=1 (Yij − Ȳi ) /n(m − 1)
2
i=1

(n−2),n(m−1)
if the model of linearity holds. We reject the hypothesis if f > Fα .

14.5 Analysis of variance in regression models

Example 14.2 This tables shows other data for male fruitflies.
Groups of 25 mean life s.e. length s.e. sleep s.e.
males kept with (days) (mm) (%/day)
no companions 63.56 16.4522 0.8360 0.084261 21.56 12.4569
1 uninterested female 64.80 15.6525 0.8256 0.069886 24.08 16.6881
1 interested female 56.76 14.9284 0.8376 0.070550 25.76 18.4465
8 uninterested females 63.36 14.5398 0.8056 0.081552 25.16 19.8257
8 interested females 38.72 12.1021 0.8000 0.078316 20.76 10.7443

59
‘Length’ is the length of the fruitfly’s thorax. It turns out that longevity (y) is
positively correlated to thorax size (x) (as plots of the data show).
Suppose we consider only the data for rows 2 and 3 and adopt a model that for
i = 2, 3,
yij = ai + βxij , j = 1 . . . , 25.
Let ā = 12 (a2 + a3 ). Our model ‘explains’ the observed variation in longevity within
group i in terms of the sum of two effects: firstly, an effect due to thorax size, ā+βxij ;
secondly, an effect specific to group i, ai − ā. We would like to test
H0 : a2 = a3 against H1 : a2 6= a3 .
To do this we need to fit the appropriate regression models under the two hy-
potheses by minimizing the residual sum of squares
X
25 X
25
S= (y2j − a2 − βx2j ) +
2
(y3j − a3 − βx3j )2 .
j=1 j=1

Under H1 we minimize freely over a2 , a3 , β and get â2 = −46.04, â3 = −55.69,
β̂ = 134.25, with residual sum of squares R1 = 6962.90.
Under H0 we minimize subject to a2 = a3 and get â2 = â3 = −45.82, β̂ = 128.18,
with residual sum of squares R0 = 8118.39. We can write
R0 = (R0 − R1) + R1 .
The degrees of freedom of H0 and H1 are 2 and 3 respectively. It can be shown that
R1 ∼ σ 2 χ250−3, whether or not H0 is true. Also R1 and R0 − R1 are independent. If
H0 is true, then R0 − R1 ∼ σ 2χ23−2 . If H0 is not true then R0 − R1 is inflated.
As we have done previously for ANOVA in Section 12.4, we compute an F statistic
(R0 − R1 )/(3 − 2)
f= = 7.80,
R1 /(50 − 3)
(1,47)
which upon comparison to F0.05 = 4.21 leads us to reject H0 ; there is indeed a
significant difference between the longevities of the males in the two groups. This is
the opposite to what we found with a t-test for equality of means in Example 11.3.
The explanation is that the mean thorax size happens to be greater within the group
of the males exposed to interested females. This is usually associated with greater
longevity. When we take into account the fact that this group did not show the
greater longevity that would be appropriate to its greater mean thorax size then we
do find a difference in longevities between males in this group and those in the group
that were kept with a nonreceptive female.
Thus we see that the analysis in Example 11.3 was deficient. There is a lesson in
this example, which might be compared to that in Simpson’s paradox.

60
15 Computational methods
Computers have freed statisticians from the grip of mathematical tractability.

15.1 Analysis of residuals from a regression

Lacking powerful computers, statisticians could once only analyse data in ways that
were not computationally too difficult, or using pre-calculated tables. Modern com-
puting power, high resolution screens and computational packages such as MINITAB
and SPLUS make it easy to display and analyse data in useful ways.
In our regression models we hypothesised that errors are IID N(0, σ 2). It is worth
checking this by an analysis of the residuals. We estimate the errors by
ˆi = Yi − Ŷi = Yi − â − β̂xi .
It can be shown by a calculation, which we omit here, that


var(Yi − Ŷi ) = 1 − 1/n − (xi − x̄)2/Sxx σ 2 .
Recall that an estimate of σ 2 is obtained from the residual sum of squares, R, as
σ̂ 2 = R/(n − 2). So we can calculate standardized residuals,
Yi − Ŷi ˆi
ˆs,i = q = p ,
σ̂ var(Yi − Ŷi )/σ 2 σ̂ 1 − 1/n − (x i − x̄) 2 /S
xx

These should be distributed approximately as N(0, 1).

Example 15.1 Standardized residuals for (a) life expectancy on log people per televi-
sion, and (b) fruitfly longevity on thorax length (for the 25 kept with no companions).
3 3
standardized residuals
standardized residuals

2 2

1 1

0 0

-1 -1

-2 -2
-3 -3
(a) 0 1 2 3 (b) 0.6 0.8 1
log people per television thorax length (mm)

We draw lines at ±1.96, the values between which samples from a N(0, 1) will
lie 95% of the time. In (a) the pattern of residuals is consistent with samples from
N(0, 1). In (b) it looks as though the magnitude of the errors might be increasing
with thorax length. This is known as ‘heteroscedasticity’. Perhaps a better model
would be i ∼ N(0, σ 2xi). This would suggest we try fitting, with ηi ∼ N(0, σ 2):
√ √ √
yi / xi = a/ xi + β xi + ηi .

61
15.2 Discriminant analysis
A technique which would be impossible in practice without computer assistance is
the credit-scoring used by banks and others to screen potential customers.
Suppose a set of individuals {1, 2, . . . , n} can be divided into two disjoint sets, A
and B, of sizes nA and nB respectively. Those in set A are known good credit risks
and those in set B are known bad credit risks. For each individual we have measured
p variables which we believe to be related to credit risk. These might be years at
present address, annual income, age, etc. For the ith individual these are xi1, . . . , xip.
The question is: given measurements for a new individual, say x01, . . . , x0p, is that
individual more likely to be a good or bad credit risk? Is he more similar to the
people in group A or to those in group B?
One approach to this problem is to use least squares to fit a model

yi = β0 + β1 xi1 + · · · + βp xip + i

where yi is defined to be 1 or −1 as i ∈ A or i ∈ B. Then the ‘discriminant function’

ŷ0 = β̂0 + β̂1x01 + · · · + β̂p x0p

is used to P
classify the new individual
P as being in group A or group B as ŷ0 is closer
to (1/nA) i∈A ŷi or to (1/nB ) i∈B ŷi . We do not go any further with the theory
here. The point is that this is a practically important application of statistics, but
a lot of calculation is required to find the discriminant function. Of course a mail
order company will experiment with building its discriminant function upon different
variables and doing this research is also computer-intensive.
Other uses of discriminant analysis, (and related ideas of ‘cluster analysis’ when
there are more than two groups), include algorithms used in speech recognition and
in finance to pick investments for a portfolio.

15.3 Principal components / factor analysis

Suppose we have measured a large number of variables, say xi1, . . . , xip, for individ-
uals, i = 1, . . . , n. Maybe these are answers to p questions on a psychological test,
such as the Myers–Briggs. The question is: can we find a much smaller number of
variables (factors) which explain most of the variation? In the Myers–Briggs test
subjects answer a large number of questions of the sort ‘when the telephone rings,
are you pleased?’ and the answers are converted to scores on 4 factors measuring
strengths of extroversion, intuition, thinking and judging. How might these four
factors have been identified from the data?
To keep things simple, we explain an approach via ‘principal components analysis’.
True factor analysis involves some further ideas that we skip over here. We begin by

62
finding that linear function of the variables with the greatest variance, i.e.,
X
n
 2 X
p
maximize (β1 xi1 + · · · + βp xip) − (β1x̄1 + · · · + βp x̄p ) subject to βi2 = 1
i=1 i=1

where x̄i is the mean of the ith variable within the population. Equivalently,
maximize β >Gβ subject to β >β = 1,
P
where G is the p × p matrix with Gjk = ni=1(xij − x̄j )(xik − x̄k ). By Lagrangian
methods we find that the maximum equals the largest eigenvalue of G, say λ1 , and is
achieved when β is the corresponding right hand eigenvector, say β 1 = (β11, . . . , βp1)> .
We call β 1 the ‘first principal component’. Similarly, we can find the eigenvector
β 2 of G corresponding to the second largest eigenvalue, λ2 . Continuing, we find
an orthogonal set of eigenvectors β 1, . . . , β m, m < p, such that the proportion of
variance explained, i.e.,
, p n
Xm X n
 j 2
XX
(β1 xi1 + · · · + βpj xip) − (β1j x̄1 + · · · + βpj x̄p) (xij − x̄j )2
j=1 i=1 j=1 i=1
Pm Pp
is near 1. This amounts to
Pp the same thing as j=1 λj / j=1 λj ; indeed the denomina-
tor above is trace(G) = j=1 λj . The above ratio is also the proportion of variation
explained by using least squares to fit
xij = α1j zi1 + · · · + αm
j
zim + ij ,
when we take zij = β1j xi1 + · · · βpj xip. Here zij is the ‘score of individual i on factor j’.
The final step is to try to give some natural interpretation to the factors,
z1 , . . . , zm . For example, if we observe that the components of β 1 which are large in
magnitude seem to match up with components of x which have something to do with
whether or not an individual is extroverted, and other components of β 1 are near 0,
then we might interpret factor 1 as an ‘extroversion factor’. Then if zi1, the score of
individual i on this factor, is large and positive we could say that i is extroverted,
and if large and negative that i is introverted.
To be fair, we should say that things are rarely so simple in practice and that
many statisticians are dubious about the value of factor analysis. For one thing, the
factors depend on the relative units in which the variables are measured.
Nevertheless, here is a simple illustration for p = 2, m = 1, n = 8. Suppose 8
students are scored on two tests, one consisting of verbal puzzles and the other of
maths puzzles; the ith student scores (xi1, xi2). The first principal component is a
line through the data which minimizes the sum of squared differences between the
data points and their orthogonal projections onto this line. A reasonable name for
this component might be ‘IQ’. The ‘IQ’ of student i is zi1 = β11xi1 + β21 xi2.

63
 90
student math verbal IQ mathmo
1
verbal score 80 score score factor factor
3 1 85 80 116.1 12.1
70 6 2 77 62 97.2 17.8
4 3 75 75 105.8 7.8
7 2 4 70 65 94.9 10.5
60
5 67 50 81.6 18.1
6 63 69 93.4 2.6
8 5
50 7 60 62 86.1 4.9
50 60 70 80 90
8 55 49 73.0 9.6
maths score

15.4 Bootstrap estimators

Suppose students are scored on two tests, and we wish to reduce their scores to single
‘IQ’ scores. Let x = (x1, . . . , x8) P be the vector of test scores, where xi = (xi1, xi2).
Define the statistic t(x) = λ1 (x)/ 2j=1 λj (x), i.e., the proportion of variation that is
explained by a single factor corresponding to the first principal component of G(x).
Suppose we are interested in θ = E t(X), a measure of how well we can do on average
when using this procedure to summarise 8 pairs of test scores in 8 single ‘IQ’ scores.
We can estimate θ by θ̂ = t(x). But to assess the accuracy of θ̂ we need to know its
variance. This depends on the distribution from which our IID samples X1 , . . . , X8
have been drawn, say F . It is no surprise that there is not a nice formula for the
variance of t(X), nor that percentage points of the distribution of t(X) have not been
tabulated; (that would require some assumption about F , e.g., that it is bivariate
normal).
A modern method of estimating the variance of θ̂ is the bootstrap estimate.
The idea is to approximate F by the empirical distribution F̂ , a sample from which is
equally likely to take any of the values x1, . . . , x8. We take a sample of 8 pairs of tests
scores from F̂ ; this corresponds to randomly choosing 8 out of the set {x1, . . . , x8},
with replacement. Perhaps we get x∗ = (x3, x8, x1, x2, x3, x3, x5, x1). From this sample
we calculate a value of the estimator, θ̂∗ = t(x∗). We repeat this procedure B times,
to get θ̂1∗ , . . . , θ̂B
∗
. Of course we use a computer to do the random sampling, the
calculation of G and of λ1 . The bootstrap estimate of the variance of θ̂ = t(X) under
F is then the estimate of the variance of t(X) under F̂ given by
!2
1 XB
1 X
B
σ̂θ̂2 = θ̂i∗ − θ̂k∗ .
B − 1 i=1 B
k=1

For the data above, z1 = 0.653x1 + 0.757x2. The proportion of variation explained
is θ̂ = t(x) = 0.86. A bootstrap estimate with B = 240 gives σ̂θ̂ = 0.094.
Formalisation of the bootstrap method dates from 1979; the study of its use for
constructing estimators, tests and confidence intervals is an active area of research.

64
16 Decision theory
To guess is cheap, to guess wrongly is expensive. (Old Chinese proverb)

16.1 The ideas of decision theory

We began the course with the following definition of Statistics:
a collection of procedures and principles for gaining and processing in-
formation in order to make decisions when faced with uncertainty.
We have studied various ways to process data and to draw inferences from it: e.g.,
point estimation, interval estimation, hypothesis testing and regression modelling.
There have been some key concepts, such as unbiasedness, the Neyman-Pearson
lemma, and the fact that least squares estimators are the best linear unbiased esti-
mators of regression parameters. But there are things that we have done which may
seem to have been rather ad hoc, and which beg unanswered questions: e.g., do we
always prefer unbiased estimators to biased ones? Do we care about estimators being
linear? Sometimes we have have done things simply so that we can get an answer.
Decision theory attempts to provide Statistics with a satisfying foundation by
placing everything within a unifying framework. In this framework the act of de-
cision making is made central and ideas of optimality are introduced so that one
can properly speak about making the ‘best’ inference. The conclusions are often the
same as those reached by other means, but can also lead in new directions.
The decision theoretic approach begins with a careful definition of all the elements
of a decision problem. It is imagined that there is a decision-maker who is to choose an
action a from a set A. He is to do this based upon observation of a random variable,
or data X. This X (typically a vector X1 , . . . , Xn ) has a probability distribution
which depends on an unknown parameter θ. Here θ denotes a state of nature. The
set of all possible values of θ is the parameter space Θ.
The decision is to be made by a statistical decision function (or rule) d; this is
a function which specifies d(x) as the action to be taken when the observed data is
X = x. On taking action a = d(X) the decision-maker incurs a loss of L(θ, a). A
good decision function is one that has a small value of the risk function
 
R(θ, d) = E L(θ, d(X)) ,
where this expectation is taken over X.
Clearly if R(θ, d1) ≤ R(θ, d2) for all θ and R(θ, d1) < R(θ, d2) for some θ then
we would never want to use rule d2, since d1 can always do as well and sometimes
better. We say d2 is inadmissible.
Decision theory requires several lectures or a whole course to cover fully. Here we
just give the flavour of some of the ideas.

65
Example 16.1 In Nature (29 August, 1996, p. 766) Matthews gives the following
table for various outcomes of Meteorological Office forecasts of weather covering 1000
one-hour walks in London.
Rain No rain Sum
Forecast of rain 66 156 222
Forecast of no rain 14 764 778
Sum 80 920 1000
Should one pay any attention to weather forecasts when deciding whether or not
to carry an umbrella?
To analyse this question in a decision-theoretic way, let W , F and U be respec-
tively the events that it is going to rain (be wet), that rain has been forecast, and
that we carry an umbrella. The possible states of nature are W and W c . The data is
X = F or X = F c . Possible actions are chosen from the set A = {U, U c }. We might
present the loss function as
Wc W
c
U L00 L01
U L10 L11
For example, we might take L01 = 4, L11 = 2, L10 = 1, L00 = 0. Of course these
are subjective choices, but most people would probably rank the four outcomes this
way.
One possible decision function is given by d1 (X) = U c , i.e., never carry an um-
brella. It’s risk function is
R(W c, d1) = L00; R(W, d1) = L01 .
Another possible decision function is given by d2 (F ) = U and d2(F c ) = U c , i.e., carry
an umbrella if and only if rain is forecast. The risk function is
R(W c , d2) = (764/920)L00 + (156/920)L10; R(W, d2) = (66/80)L11 + (14/80)L01 .
We see that if θ = W c then d1 is better, but if θ = W then d2 is better. Thus neither
rule is uniformly better for both states of nature. Both d1 and d2 are admissible. By
averaging over the states of nature we have the so-called Bayes risk, defined as
B(d) = E [R(θ, d)],
where the expected value is now taken over θ. For example, in our problem, P(W ) =
0.08 and P(W c ) = 0.92, so B(d) = 0.08R(W, d) + 0.92R(W c, d).
The Bayes rule is defined as the rule d which minimizes the Bayes risk. Thus to
find the Bayes rule for our problem, we must compare
B(d1) = .08L01 + .92L00

66
to
   
B(d2) = .08 (66/80)L11 + (14/80)L01 + .92 (764/920)L00 + (156/920)L10
= .066L11 + .014L01 + .764L00 + .156L10.

It follows that it is better to ignore weather forecasts and simply go for walks without
an umbrella, if

L01 − L11 .156

B(d1) < B(d2) ⇐⇒ ∆ := < = 2.364,
L10 − L00 .066

which can hold for reasonable values of the loss function, such as those given above,
for which ∆ = 2. It all depends how you feel about getting wet versus the inconve-
nience of carrying an umbrella. Similar analysis shows that the commonly followed
rule of always carrying an umbrella is better than doing so only if rain is forecast
only if one is very adverse to getting wet, i.e., if ∆ > 764/14 + 53.

16.2 Posterior analysis

In Lecture 5 we considered a decision theoretic approach to the point estimation

problem. We used a loss function L(θ, a) to measure the loss incurred by estimating
the value of a parameter to be a when its true value is θ. Then θ̂ was chosen to
minimize E [L(θ, θ̂)], where this expectation is over θ with respect to the posterior
distribution p(θ | x).
Another way to think about the decision problem above is similar. We consider
the expected loss under the posterior distribution. The posterior distribution for rain,
given the data that there has been a forecast of rain, is P(W | F ) = 66/222 + 0.30.
(Note that this is less than 0.50!) Hence, given a forecast of rain, the expected loss
if we carry an umbrella is

B(U | F ) = (66/222)L11 + (156/222)L10,

whereas if we don’t carry an umbrella the expected loss is

B(U c | F ) = (66/222)L01 + (156/222)L00.

Not surprisingly, this leads to exactly the same criterion for choosing between d1 and
d2 as we have already found above.
This is a general principle: the Bayes rule, d, can be determined as the action a
which minimizes E θ|X [R(θ, a)], this expectation being taken over θ with respect to
the posterior distribution p(θ | x).

67
16.3 Hypothesis testing as decision making
We conclude by elucidating a decision theoretic approach to hypothesis testing. Con-
sider the problem of testing a simple null hypothesis H0 : θ = θ0 against a simple
alternative hypothesis H1 : θ = θ1 . On the basis of an observation X we must decide
in favour of H0 (i.e., take action a0 ) or decide in favour of H1 (i.e., take action a1 ).
For the case of so-called 0–1 loss we take L(θ0, a0 ) = L(θ1, a1 ) = 0 and L(θ0, a1) =
L(θ1, a0) = 1. I.e., there is unit loss if and only if we make the wrong decision.
The risk function is then simply the probability of making the wrong decision, so
R(θ0, d) = P(d(X) = a1 | H0) and R(θ1, d) = P(d(X) = a0 | H1).
Suppose we have prior probabilities on H0 and H1 of p0 and p1 respectively. This
gives Bayes risk of

B(d) = p0R(θ0, d) + p1 R(θ1, d).

As we have seen in the previous section the Bayes rule minimizes the posterior
losses, so we should choose d(X) to be a1 or a0 as
B(a0 | x) P(H1 | x) p1 P(x | H1 ) p1 f (x | θ1 )
= = =
B(a1 | x) P(H0 | x) p0 P(x | H0 ) p0 f (x | θ0 )
is greater or less than 1.
This is of course simply a likelihood ratio test. Observe, however, that we have
reached this form of test by a rather different route than in Lecture 6.

16.4 The classical and subjective points of view

The decision theoretic approach to statistical inference is appealing for the way it
directly addresses issues such as loss, risk, admissibility, etc. These have intuitive
interpretations in terms of the economics of decision making.
Decision theory also has the philosophical merit or dismerit, depending on your
point of view, that it incorporates the Bayesian notions of prior and posterior beliefs.
In the analysis of the hypothesis test above, we had to introduce a prior distribution
on H0 and H1 , as given by the probabilities p0 and p1. Some statisticians argue that
this is fine; people always come to decision problems armed with prior beliefs, if only
an uninformed belief expressed as p0 = p1 = 1/2. Others take the ‘classical’ line
that statistical procedures should not depend upon the introduction of subjective
prior beliefs on the part of the person analysing the data. They argue that only
the data should matter: two people should automatically come to the exactly the
same conclusion when presented with the same data. Most practising statisticians
are happy to take the best of both viewpoints, letting the actual question under
consideration decide which concepts and procedures are most helpful.

68