All Ex Sol

Introduction to Generalized Linear Modelling,
Example Sheets 1, 2 and 3, with solutions

P.M.E.Altham, Statistical Laboratory, University of Cambridge.
Contents
1 Example Sheet 1 3
1.1 Example Sheet 1: questions . . . . . . . . . . . . . . . . . . . . . 3
1.2 Solutions to Example Sheet 1 . . . . . . . . . . . . . . . . . . . . 5
2 Example Sheet 2 11
3 Example Sheet 3 24
1
Preface
These three Examples Sheets, and their solutions, have been built up from
1992, when I rst gave this lecture course, to 2005, when I gave it for the
last time. I am grateful to many users, particularly undergraduates, for their
comments and questions. There is some overlap between these examples and
the examples in my lecture notes. An expanded version of these notes may be
seen at http://www.statslab.cam.ac.uk/~pat/All.pdf
2
Chapter 1
Example Sheet 1
1.1 Example Sheet 1: questions
(This is meant to be a very easy sheet, to get you started.)
1. Suppose X
1
, . . . , X
n
are i.i.d. Poisson random variables with parameter .
Show that = X
i
/n, and var ( ) = /n.
What is
E(
2
L
2
)?
What is the exact distribution of (n )? What is the asymptotic distribution of
?
2. Suppose we have n independent trials, and the outcome of each trial is
Red with probability
1
,
or White with probability
2
,
or Blue with probability
3
,
where
1
+
2
+
3
= 1.
Let (X, Y, Z) be the total number of (Red, White, Blue) trials in the sequence
of n; write X =
n
1
X
i
, Y =
n
1
Y
i
for suitably dened (X
i
, Y
i
).
Find E(X), var (X), and show that
cov (X, Y ) = n
1
2
.
Find
_
2
_
, and nd the mean vector, and covariance matrix, of its asymptotic
distribution (which is of course bivariate normal).
3. Suppose Y
i
independent Poisson, mean
i
, and our model is
H : log(
i
) = +x
i
where (x
i
) are given.
Write down the log likelihood log f(y|, ) and hence nd
3
P.M.E.Altham 4
(i) the sucient statistics for (, );
(ii) equations for ( ,

), the maximum likelihood estimator (mle), and
(iii) an expression for
max
=0
f(y|, ).
Show how you would use this, together with Wilks theorem, to test
H
0
: = 0.
4. Suppose
f(y
i
|) = (1/2) exp |y
i
|, < y
i
< .
Let L() =
n
1
log f(y
i
|). Sketch this as a function of , and hence nd an ex-
pression for

, the mle, given that the ordered observations are y
(1)
< . . . < y
(n)
.
What is
L
?
5. Suppose f(y
i
|) = e
yi
, y
i
> 0, 1 i n.
Show that this is an exponential family form distribution, with natural param-
eter = . Find the sucient statistic and its distribution, and nd the mle
for each of , .
6. Suppose f(y
i
|, ) is the pdf of N( + x
i
,
2
), where
2
is known, and
x
1
, . . . , x
n
are known. Show that the loglikelihood function L
n
(, ) is a con-
cave function of (, ).
7. Suppose
f(y
i
|) =
1
2
exp y
2
i
/2,
for i = 1, . . . , n. Show that the mle of is

= 1/n
n
1
y
2
i
, and that
n

2
n
.
Find the exact mean and variance of

, and compare these with the asymptotic
mean and variance obtained from general likelihood theory.
8. Suppose Y
i
Bi(1,
i
) independent, (i = 1, . . . , n) (i.e. P(Y
i
= 1) =
i
,
P(Y
i
= 0) = 1
i
). Suppose also log(
i
/(1
i
)) = +x
i
, where x
1
, . . . , x
n
are given.
Write down L
n
(, ), the loglikelihood function, and nd equations for ( ,

)
the mle of (, ).
9. Suppose Y
i
NID(
1
+
2
x
i
+
3
P
2
(x
i
),
2
), 1 i n, where x
i
= 0 and
P
2
(x
i
) is a given quadratic function of x
i
such that
P
2
(x
i
) = x
i
P
2
(x
i
) = 0.
Find
,

2
T
and E

2
T
.
Find

3
and var (
3
), and show how you would test H
0
:
3
= 0, when
2
is
known.
If P
2
(x
i
) = x
2
i
+ ax
i
+ b, nd expressions for a, b in terms of (x
i
). Can you see
any advantages in tting the model with
E(Y
i
) =
1
+
2
x
i
+
3
P
2
(x
i
)
P.M.E.Altham 5
as above rather than the model written as
E(Y
i
) =
1
+
2
x
i
+x
2
i
say?
(The purpose of this question is to introduce you to ORTHOGONAL polyno-
mials.)
10. Suppose Y
i
Bi(n
i
, p
i
), 1 i k, independent, and g(p
i
) = +x
i
where
g() is a known link function. Write down
.
Write down the sucient statistics for (, )
(i) if g(p) = log(p/(1 p)) (link=logit)
(ii) if g(p) = log(log(1 p)) (link= cloglog).
1.2 Solutions to Example Sheet 1
1. The likelihood is
f(x|) = e
xi
/x
i
!
giving the loglikelihood as
L = log(f(x|)) = n + x
i
log() +constant.
Hence
L
= n + x
i
/
and
2
L
2
= x
i
/
2
( < 0)
so that L is maximised at
= x
i
/n.
Clearly E(X
i
) = = var(X
i
). Thus E( ) =
and var( ) = /n, and
E
2
L
2
= n/.
The exact distribution of n = X
i
is Po(n).
The asymptotic distribution of , by the Central Limit Theorem, is N(, /n).
P.M.E.Altham 6
2. (This is the 3-cell multinomial distribution). With
f(x, y, z|) = n!
x
1
y
2
z
3
x!y!z!
for x, y, z = 0, 1, 2..., and x + y + z = n , we have X = X
i
say, where X
i
=1
if ith trial results in a Red, and X
i
=0, otherwise, Y = Y
i
, and Y
i
= 1 if ith
trial results in a White, Y
i
=0 otherwise.
Clearly, P(X
i
= 1) =
1
, and X is Bi(n,
1
)
so that var(X)= n
1
(1
1
), E(X) = n
1
.
Further
cov(X, Y ) = cov(X
i
Y
i
) = n(E(X
1
Y
1
) E(X
1
)E(Y
1
)).
Clearly, E(X
1
Y
1
) = 0, so cov(X, Y ) = n
1
2
.
Now
L = logf(x, y|) = xlog(
1
) +ylog(
2
) +zlog(
3
) + constant
which is maximised subject to
1
+
2
+
3
=1 (use a Lagrange multiplier)
by

1
= x/n,

2
= y/n,

3
= z/n.
Hence E(
i
) =
i
for i = 1, 2, 3 . Now
L()
1
= (x/
1
) (z/
3
)
L()
2
= (y/
2
) (z/
3
).
Hence minus the matrix of 2nd derivatives of L is
_
x/
2
1
+z/
2
3
z/
2
3
z/
2
3
y/
2
2
+z/
2
3
_
Substituting for E(x), E(y), E(z), we see that the expectation of the above matrix
is
_
n(1
2
)/
1
3
n/
3
n/
3
n(1
1
)/
2
3
_
.
It now remains for you to check that the inverse of this 2 2 matrix is
_

1
(1
1
)/n
1
2
/n
2
/n
2
(1
2
)/n
_
.
This is what the general formula for the asymptotic covariance matrix gives
us. In this case, it agrees exactly with the exact covariance matrix.
3.(i)
With f(y
i
|
i
) proportional to e
i
i
yi
and
i
= exp( +x
i
) we see that the likelihood for (, ) is proportional to
[exp e
+xi
]exp [ t
1
+ t
2
]
P.M.E.Altham 7
where t
1
is dened as y
i
, and t
2
as x
i
y
i
.
Hence, by the factorisation theorem, (t
1
, t
2
) are sucient for (, ).
The log likelihood is
L(, ) = e
+xi
+ t
1
+ t
2
+constant.
ii)Thus
L
= 0 for t
1
= e
+xi
L
= 0 for t
2
= x
i
e
+xi
.
These are the equations for ( ,

). To verify that this is indeed the max-
imum, we should check that (minus the matrix of 2nd derivatives) is
positive- denite at ( ,

).
The equations for ( ,

) do not have an explicit solution, but we could solve
them iteratively to nd ( ,

), and hence we could evaluate the maximum of L.
iii) Now, if = 0, L(, ) = e
+ t
1
. It is easily seen that this is max-
imised with respect to by
say, where
= log(t
1
/n).
We know, by Wilks theorem,that to test H
0
: = 0 against H
1
: arbitrary,
we should refer
2[L( ,

) L(
, 0)] to
2
1
.
4.(An example of how things can be tricky when we are not in the glm family).
logf(y|) =
|y
i
| + constant = g() + constant say.
Dening y
(1)
, ..., y
(n)
as the ordered sample values, as instructed in the question,
we see that
g() =
(y
(i)
) for < y
(1)
(this is a straight line of slope n)
g() = (y
(1)
) +
2
n
(y
(i)
) for y
(1)
< < y
(2)
(this is a straight line of slope (n + 2))
and so on....
Finally,
g() =
1
n
( y
(i)
), for > y
(n)
This is a straight line of slope n. Thus, sketching g() we see that the log-
likelihood function is concave in shape, consisting of straight line segments, and
if n is odd, say n = 2m+ 1, then

= y
(m+1)
,
if n is even, say n = 2m, then

is anywhere between y
(m)
and y
(m+1)
.
L/ is of course not dened at y
(1)
, ..., y
(n)
.
P.M.E.Altham 8
Otherwise it is the slope of the appropriate linear segment.
In this example we could nd the asymptotic distribution of

by going back
to rst principles (it would make for quite a tough exercise). But we CANNOT
nd it by quoting the general theorem for the asymptotic distribution of mles:
the appropriate regularity conditions do not hold.
5. f(y|) =
n
exp y
i
which is of exponential family form, with
t(y) = y
i
as our sucient statistic, having gamma(n, ) distribution,
and loglikelihood L = n log y
i
. Thus
L/ = n/ t(y).
Hence

= n/t(y), and = n/t(y) .
6. logf(y|, ) = (y
i
x
i
)
2
/2
2
+constant = L
n
(, ) say.
Now nd [minus the matrix of 2nd derivatives] : show that it is positive-denite.
Hence L
n
(, ) is a concave function.
7. L
n
() = (n/2)log() y
2
i
/(2) +constant.
Thus
L/ = (n/2) + y
2
i
/(2
2
)
Hence
= y
2
i
/n
and
n
/ = y
2
i
/,
where y
i
/
are NID(0, 1).

Hence n
/ is distd as
2
n
hence has mean n, variance 2n.
Thus E(
) = , as we would hope.
Now,
2
L
2
= n/(2
2
) (y
2
i
/
3
)
giving
E(
2
L
2
) = n/(2
2
).
So the asymptotic variance of

is (2
2
)/n, which is the same as the exact
variance.
8. With Y
i
independent Bi(1,
i
) as given, we see that
f(y|, ) =
i
yi
(1
i
)
1yi
P.M.E.Altham 9
Hence,
L
n
(, ) = y
i
log(
i
/(1
i
)) log(1
i
)
Substituting for (
i
) gives
L
n
(, ) = ( + x
i
)y
i
log(1 +e
+ xi
)
thus,
L
n
(, ) = y
+
+ x
i
y
i
log(1 +e
+ xi
)
Hence we can write down the equations
L
n
/ = 0, L
n
/ = 0.
Observe that there is no closed form solution to these equations. We can only
nd the mles by an iterative solution.
9. With Y
i
distributed as NID(
1
+
2
x
i
+
3
P
2
(x
i
),
2
)
where x
i
= 0, P
2
(x
i
) = 0, x
i
P
2
(x
i
) = 0
we see that the loglikelihood is l() + constant, where
l() = (y
i
2
x
i
3
P
2
(x
i
))
2
/2
2
.
This gives
l/
1
= (y
i
2
x
i
3
P
2
(x
i
))/
2
= (y
i
1
)/
2
.
Similarly,
l/
2
= x
i
(y
i
2
x
i
)/
2
l/
3
= P
2
(x
i
) (y
i
3
P
2
(x
i
))/
2
.
Hence
2
l
T
= (1/
2
) diag(n, x
2
i
, (P
2
(x
i
))
2
)
= E(
2
l/
T
).
Solving l/ = 0, gives

, in particular
3
= y
i
P
2
(x
i
)/(P
2
(x
i
))
2
.
Now, since E(Y
i
) =
1
+
2
x
i
+
3
P
2
(x
i
), we can see that
E(
3
) =
3
.
Further, Y
i
are NID, each with variance
2
, hence
3
has variance
2
/(P
2
(x
i
))
2
, and is normally distributed.
We can test H
0
:
3
= 0 , by referring
3
/
_
(its variance) to N(0, 1) .
P.M.E.Altham 10
Given P
2
(x
i
) = x
i
2
+a x
i
+b, we nd a, b by solving the pair of equations
(x
i
2
+a x
i
+b) = 0, (x
i
2
+a x
i
+b)x
i
= 0
giving
x
2
i
+n b = 0, x
3
i
+a x
2
i
= 0.
The advantage of parametrising the model in ths way is that the parameters
1
,
2
,
3
are orthogonal. Thus for example, if we want to t
E(Y
i
) =
1
+
2
x
i
we nd that the least squares estimators for
1
,
2
are the same as in the full
model, ie the same as they were when we included the term
3
.
10. We may write
l(, ) = y
i
log(p
i
/(1 p
i
)) + n
i
log(1 p
i
)
from which we may nd the expressions for the partial derivatives; for general
link function g( ) these do not simplify.
i) The sucient statistics for (, ) for the logit link are
(y
i
, y
i
x
i
).
The logit link is of course the canonical link for the binomial distribution.
ii) For the complementary log log link, there is no reduction in dimensionality
from n for the sucient statistics: we will still need the original data (y
i
, x
i
) to
construct the likelihood function.
Chapter 2
Example Sheet 2
1. If Y
i
are independent Poisson, means exp
T
x
i
, 1 i n, how would you
evaluate

and its asymptotic covariance matrix? What does scale parameter
taken as 1.000 mean in the corresponding glm output?
2 * If the loglikelihood can be written
() = (
T
t ())/ where > 0,
and t = t(y) is a p-dimensional vector, show that the covariance matrix of t(y)
is
_
+
2
T
_
and hence that () is a strictly concave function of . What is the practical
application of this result in estimation of ? Illustrate your answer for either
the binomial or the Poisson distribution.
3. We can say that an asymptotic 90 % condence region for

is derived from
(
)
T
(V (
))
1
(
)
2
p
(approximately).
Show that if p = 2, the resulting region is an ellipse, and nds its equation in
the case where
2
0 .
Why is the remark
_
c
e
x
dx = e
c
relevant in this context?
4. In the least-squares t of the model
y
ij
= +
i
+
ij
, 1 j n
i
, 1 i t
with
1
= 0 (i.e. the glm constraint), and the usual assumption that
ij
are
NID(0,
2
), show that
= y
1+
/n
1
,

i
= y
1+
/n
1
+y
i+
/n
i
(i = 1)
11
P.M.E.Altham 12
and
var (
i
) =
2
_
1
n
1
+
1
n
i
_
(i = 1).
Find cov (
i
,

) for i = , (i, = 1). (Here y

i+
is dened as

ni
j=1
y
ij
).
Show that the tted value of y
ij
is y
i+
/n
i
, for i = 1, . . . , t.
Hint: its easier to do the algebra with the sum to zero constraint,
and then transform back to the corner-point constraint.
5. In the least-squares t of the model
y
ijk
= +
i
+
j
+
ijk
, 1 k r, 1 i I, 1 j J
with
1
= 0,
1
= 0, and the usual assumption that
ijk
are NID(0,
2
), show
that, for i = 1,

i
=
1
Jr
(
j,k
y
ijk
j,k
y
1jk
),
with the corresponding expression for

j
. How is your answer aected if the
condition 1 k r is replaced by the condition 1 k r
ij
?
6. Why would you expect, in an experiment with IJ observations, that the
model
y
ij
= +
i
+
j
+
ij
+
ij
,
where
ij
NID(0,
2
), 1 i I, 1 j J
gives a perfect t to the data (i.e. zero deviance)?
7. In the usual model
Y = X +
with
i
NID(0,
2
) explain why displaying the estimated covariance matrix
of

is a method of nding out about (X
T
X)
1
.
8. The data-set below is taken from the Minitab student Handbook(1976) by
Ryan,T., Joiner, B. and Ryan, B. and is also discussed in the book by Aitkin et
al. (See Chapter 3 of the book.) For the 31 cherry trees, the table below shows
d, h, v. These are dened by d is the diameter (in inches), at a height of 4.5 feet
from the ground, h is the height (in feet) of the trees, v is the volume of useable
wood, in cubic feet.
Reminder: 1 foot= 12 inches.
This is one of the datasets already in R: try
data(trees); attach(trees); trees[1,]
(but you need to take Girth as d, which is confusing, I know).
The order is (d, h, v).
8.3 70 10.3, 8.6 65 10.3, 8.8 63 10.2, 10.5 72 16.4,
10.7 81 18.8, 10.8 83 19.7, 11.0 66 15.6, 11.0 75 18.2,
11.1 80 22.6, 11.2 75 19.9, 11.3 79 24.2, 11.4 76 21.0,
P.M.E.Altham 13
11.4 76 21.4, 11.7 69 21.3, 12.0 75 19.1, 12.9 74 22.2,
12.9 85 33.8, 13.3 86 27.4, 13.7 71 25.7, 13.8 64 24.9,
14.0 78 34.5, 14.2 80 31.7, 14.5 74 36.3, 16.0 72 38.3,
16.3 77 42.6, 17.3 81 55.4, 17.5 82 55.7, 17.9 80 58.3,
18.0 80 51.5, 18.0 80 51.0, 20.6 87 77.0.
Take lv = log(v), with ld, lh dened similarly.
Verify that the following models give the estimates (with ses) and deviances
below, and discuss the t of these models.
M
0
: E(lv) =
for which
= 3.273(0.0945)
and deviance = 8.3087(df = 30);
M
1
: E(lv) = + ld
for which
= 2.353(0.2307),

= 2.200(0.0898)
and deviance= 0.38324(df = 29);
M
2
: E(lv) = + ld + lh
for which
= 6.632(0.7998),

= 1.983(0.0750), = 1.117(0.2044)
and deviance= 0.18546(df = 28);
M
3
: E(lv) = + lh
for which
= 13.96(3.755), = 3.982(0.8677)
and deviance= 4.8130(df = 29).
What are the numerical consequences of the non-orthogonality of the parameters
, ?
The volume of a cylinder of length , diameter d, is (d
2
)/4, and the volume
of a cone of height and base diameter d is (d
2
)/12. Are these cherry trees
more like cylinders than cones?
9. The Independent (21/12/88) gave the League Table of football-related
arrests, printed in the table below. This list details a total of 6147 football-
related arrests in the 1987-8 season, and is compiled by the Association of Chief
Constables. It does not dierentiate between Home-fans and Away-fans.
There are 4 Divisions (with Division 1 containing the best clubs) and these have
21, 23, 24 and 24 clubs respectively, each Division corresponding to a pair of
columns in the Table below.
P.M.E.Altham 14
The columns below give (a, n) where
a= attendance, in thousands, and
n= number of arrests,
for each of the 4 soccer divisions, with the order (reading across the rows) being
(a, n) for Division 1, (a, n) for Division 2, (a, n) for Division 3, (a, n) for Division
4.
YOU ARE NOT INTENDED TO TYPE IN THIS DATA: ASK ME TO EMAIL
THE SET TO YOU.
a1 n1 a2 n2 a3 n3 a4 n4
325 282 404 308 116 99 71 145
409 271 286 197 401 80 227 132
291 208 443 184 105 72 145 90
350 194 169 149 77 66 56 83
598 153 222 132 63 62 77 53
420 149 150 126 145 50 74 46
396 149 321 110 84 47 102 43
385 130 189 101 128 47 39 38
219 105 258 99 71 39 40 35
266 91 223 81 97 36 45 32
396 90 211 79 205 34 53 29
343 86 215 78 106 32 51 28
518 74 108 68 43 28 51 27
160 49 210 67 59 22 115 21
291 43 224 60 88 22 52 21
783 38 211 57 226 21 67 21
792 33 168 55 61 21 52 17
314 32 185 44 91 21 52 17
556 24 158 38 140 20 72 15
174 14 429 35 85 18 49 12
162 1 226 29 127 11 101 10
NA NA 150 20 59 5 90 8
NA NA 148 19 87 4 50 5
NA NA NA NA 79 3 41 0
Let (n
ij
, a
ij
) be the observations for the ith club of the jth Division, for j =
1, ..., 4.
Making the standard assumption that the errors are NID(0,
2
) consider how
to t the following models in R or Splus, and sketch graphs to show what these
models represent:
a) E(n
ij
) = ,
b) E(n
ij
) = + a
ij
,
c) E(n
ij
) = +
j
+ a
ij
,
d) E(n
ij
) = +
j
+
j
a
ij
.
You will nd that the deviances for these 4 models are, respectively
371056(91df), 296672(90df), 272973(87df), 241225(84df).
Now plot n against a. What do you conclude about your assumption of constant
P.M.E.Altham 15
error variances?
Now repeat the model-tting exercise with n, a replaced by log(n), log(a).
Can you now think of a way of identifying certain clubs as rogues, or indeed
as saints within their Division?
10. The data-set below which is also discussed by Agresti (1995, p101) is based
on a study of British doctors by R.Doll and A.B.Hill(1966) and gives the number
of coronary deaths for smokers and non-smokers, for each of 5 dierent age-
groups, with the corresponding person-years, ie the total time at risk. Thus
for example, in the youngest age-group, in the non-smoking category, the total
time at risk was 18793 years, and during this time there were 2 coronary deaths
in this particular class. The age-groups are 35-44, 45-54, 55-64, 65-74, 75-84
years.
Dene d
ij
as the number of deaths in age group i, smoking group j, where
i = 1, . . . , 5, j = 1, 2 with j = 1 for non-smokers, j = 2 for smokers. Assume
that (d
ij
) are distributed as independent Poisson variables, with E(d
ij
) =
ij
and
ij
=
ij
p
ij
where p
ij
= total person-years at risk, for age i, smoking group
j. Hence log(
ij
) = log(
ij
) + log(p
ij
) for all i, j. This is why we take log(p
ij
)
as the oset in the glm analysis below. We model log(
ij
), the parameter of
interest.
Verify and interpret the results from the models tted below.
pyears <- scan()
18793 52407
10673 43248
5710 28612
2585 12663
1462 5317
# BLANK LINE
deaths <- scan()
2 32
12 104
28 206
28 186
31 102
# BLANK LINE
sm <-rep(c(1,2),times=5)
age <- c(1,1,2,2,3,3,4,4,5,5)
#these are crude but easy-to-understand ways of setting up the factor levels
sm <- factor(sm) ; age <- factor(age)
prop <- deaths/pyears ; tapply(prop,list(age,sm),mean) <- log(pyears)
summary(glm(deaths ~ age + offset(l),poisson),cor=F)
summary(glm(deaths ~ age + sm + offset(l),poisson),cor=F)
It should now be obvious to you that
(i) smoking is bad for you and
(ii) so is getting old.
But you will also see that this nal model does not t well (its deviance is
P.M.E.Altham 16
12.134, with 4 df). Can you suggest a way of improving the t?
11. With y
1
, . . . , y
n
independent observations, and y
i
having pdf from the usual
glm family, with g(
i
) =
T
x
i
as usual, nd the expectation of the second
derivative of the log-likelihod function with respect to , , and hence show
that

,

are asymptotically independent.
12. New for 2005: introduction to the inverse Gaussian distribution.
Consider the density
f(y|, ) = exp[(y b())/ +c(y, )], y > 0,
where
=
2
, b() = (2)
1/2
,
and
2c(y, ) = log(2y
3
) + (1/y).
Show that E(Y ) = (2)
1/2
= , say. Check that 1/
2
is the canonical link
function for this glm, and that var(Y ) =
3
2
. nb, no fancy integration
required, at all.
1. Here Y
i
are distributed as independent Po(
i
), with log(
i
) =
T
x
i
. Thus
f(y
i
|) (exp
i
) (
i
)
yi
so that
log f(y|) =
n
1
[exp(
T
x
i
) +y
i
T
x
i
] +constant
ie L
n
=
n
1
[exp(
T
x
i
)] +
T

n
1
y
i
x
i
+constant.
Thus
L
n
=
n
i
[x
i
exp(
T
x
i
) +y
i
x
i
]
and minus the matrix of second derivatives of L
n
is say J, where
J = +
n
1
x
i
x
T
i
exp(
T
x
i
).
This is +
n
1
x
i
x
T
i

i
, where
i
> 0.
Hence J is a positive denite matrix, and so

(the mle) is the solution of
L
n
= 0,
which we may write as
x
i
i
=
x
i
y
i
P.M.E.Altham 17
ie the observed and the expected values of the sucient statistics

x
i
y
i
agree
exactly at =

.
The equation
Ln
= 0 has to be solved iteratively. The asymptotic covariance

matrix of

is the inverse of E(J), which here is
n
i
x
i
x
T
i
.
(There is no simple formula for the inverse.)
Scale parameter taken as 1.000 means that in writing the pdf in glm formu-
lation, ie as
f(y
i
|
i
, ) = exp [(y
i
i
b(
i
))/ +c(y
i
, )]
then we take as 1.
2. The loglikelihood is
l() = (
T
t ())/ where > 0.
Thus
l() = (t

())/
and as always
E(
l
) = 0.
so that E(t) =

(). Further the matrix of 2nd derivatives of the loglikelihood

is -
1
times the matrix of 2nd derivatives of and, using
_
f(y|)dy = 1 for all
again, we have
E(matrix of 2nd derivatives of l) = E(UU
T
), where U =

l
Thus covariance matrix of t(y) is
matrix of 2nd derivatives of
which must therefore be a positive denite matrix, so that l() is a STRICTLY
CONCAVE function. Thus any solution

say, of

l = 0 must be THE MAX-

IMUM of l(). Thus, eg if we solve U = 0 by the Newton-Raphson algorithm,
we will unerringly home in on the right solution, rather than something nasty
and irrelevant like a local minimum.
P.M.E.Altham 18
3. We know (using to mean approximately distributed as)
N(, V ())
and so
(
)
T
(V ())
1
(
)
2
p
and hence
(
)
T
(V (
))
1
(
)
2
p
.
Hence from
2
tables we can choose c, given , such that
Pr[(
)
T
(V (
))
1
(
) c] 1 .
Write V = V (
)) for short; our (1 ) condence region is

(
)V
1
(
) c
which (because V
1
is positive-denite) is an ELLIPSE, centred on

. For
2
= 0
we have V a diagonal matrix, with diagonal entries v
1
, v
2
say and the ellipse is
(
1
1
)
2
/v
1
+ (
2
2
)
2
/v
2
= c.
The relevance of the nal remark is that
2
p
has, for p = 2 the probability density
function
(1/2)exp(x/2), x > 0.
4. Given
y
ij
= +
i
+
ij
,
for i = 1 . . . t, j = 1 . . . n
i
with
1
= 0
equivalently
y
ij
=
i
+
ij
with
1
= , and
i
= +
i
for i > 1. We see that
(y
ij

i
)
2
is minimised, equivalently (y
ij

i
)
2
is minimised with respect to (
i
) by

i
=
y
i+
n
i
.
giving
=
y
1+
n
1
and
P.M.E.Altham 19
i
=
y
1+
n
1
+
y
i+
n
i
for i > 1,
as required. Clearly, since
ij
are NID(0,
2
),
var(
y
i+
n
i
) =

2
n
i
and
cov(y
1+
, y
i+
) = 0 for i > 1
hence var(
i
) is as given, and
cov(
i
,

l
) = var(
y
1+
n
1
) for i = l.
The tted value of y
ij
is
i
, ie y
i+
/n
i
.
5.
y
ijk
= +
i
+
j
+
ijk
, k = 1 . . . r, i = 1 . . . I, j = 1, . . . J.
We can reparametrise this as
y
ijk
= m+a
i
+b
j
+
ijk
,
with a
i
= 0, b
j
= 0. Then = m+a
1
+b
1
, +
i
= m+a
i
+b
1
, +
j
=
m+a
1
+b
j
, so that
i
= a
i
a
1
, and
j
= b
j
b
1
.
Straightforward minimisation of
(y
ijk
ma
i
b
j
)
2
subject to the constraints a
i
= 0 , b
j
= 0 gives
m = y, a
i
=
y
i++
n
i
y , etc.
Hence, returning to the original model, we see that

i
=
y
i++
n
i
y
1++
n
1
.
If now y
ijk
= +
i
+
j
+
ijk
, with
k = 1 . . . r
ij
, i = 1, . . . , I, j = 1, . . . , J
then we nd that
i
is no longer given by the above simple formula .
This is best seen by the following simple example. Suppose the observations
(y
ijk
) are
23.9 , 7.2 for (i, j) = (1, 1),
P.M.E.Altham 20
3.6 for (i, j) = (1, 2),
10.4 for (i, j) = (2, 1),
29.7 for (i, j) = (2, 2).
Note that you will dierent estimates for, say
i
, depending on whether or
not
j
is in the model, because the design is unbalanced.
6. Here the model is
y
ij
= +
i
+
j
+ ()
ij
+
ij
for i = 1 . . . I, j = 1 . . . J, which is equivalent to
y
ij
=
ij
+
ij
,
and the lse(=mle) of (
ij
) are obtained by minimising
(y
ij

ij
)
2
with respect to
ij
. This gives
ij
= y
ij
so that the minimised residual sum of
squares is 0, trivially.
Further, we have no degrees of freedom left to estimate
2
. This is called a
saturated model: it is saturated with parameters.
7.
Y = X +,
with distributed as N(0,
2
I), gives
(Y X)
T
(Y X)
is minimised by

such that X
T
X
= X
T
Y
ie

= (X
T
X)
1
X
T
Y.
Hence

is distributed as N(0,
2
(X
T
X)
1
).
If we display the the estimated covariance matrix of

it is of course
s
2
(X
T
X)
1
where s
2
= deviance/df . Since s
2
is therefore known, we can evaluate (X
T
X)
1
.
8. Using an obvious notation, you need to try
lm(lv ~ 1) ; lm(lv ~ lh) ; lm(lv ~ ld) ; lm(lv~lh + ld)
Observe that we never expected M
0
to t: we always start by tting it as our
baseline. Observe also that M
1
ts much better than M
0
, as we would expect,
and that

is clearly signicant. (The ratio 2.2/.0898 is much greater than 2).
Observe that the estimate of changes between M
1
and M
2
because and
are not mutually orthogonal. [There is a non-zero correlation between the
vectors (ld),(lh)].
The coecients

and in M
2
are clearly both signicant.
P.M.E.Altham 21
Observe: the reduction in deviance in moving from M
0
to M
1
the ss due to
is dierent from the reduction in deviance in moving from M
1
to M
2
, the ss
due to , allowing for .
This is another consequence of the non-orthogonality of these 2 parameters.
Our nal model is thus
lv
i
= 6.632(.7998) + 1.117(.2044) lh
i
+ 1.983(.07501) ld
i
Observe that both a cylinder & a cone would give
lv
i
= C + 1 log(h
i
) + 2 log(d
i
/12)
where C = log(/4) for a cylinder, log(/12) for a cone. [warning : d
i
is actually
measured in inches. Do todays metric students know about feet and inches? ]
The rest of the solution is up to you. Not surprisingly, trees do turn out to be
more like cones than cylinders.
9. Note that the linear regression of n, the number of arrests on a, the atten-
dance, although having a signicant positive slope (as we would expect) is really
rather a poor t: it has R
2
, the proportion of the total deviance explained by
the regression as only 0.2005. (By denition, R
2
lies between 0 and 1, with
R
2
=1 for a perfect t.)
If we plot the residuals = (observedfitted values) against the corresponding
fitted values, we get a very pronounced fanning-out eect, suggesting that
var(n
i
) increases as E(n
i
) increases, which is what we would expect, since n
i
might well be Poisson-like in its behaviour. Thus the analysis presented, which
fails to START by doing some simple plots, is not very sensible. Its big draw-
back is that it relies on the assumption that var(n
i
) is constant over i.
This is what the analysis does. First t
n
ij
= +
ij
, for j = 1 . . . 4, i = 1 . . . n
j
.
with the usual assumption that
ij
is NID(0,
2
).
Then t
n
ij
= + a
ij
+
ij
ie the same line for all 4 divisions. Then t
n
ij
= +
j
+ a
ij
+
ij
ie parallel lines for all 4 divisions. Lastly try
n
ij
= +
j
+
j
a
ij
+
ij
.
There isnt any appreciable improvement in t after the second model.
And, as we have already said, the assumption of homogeneity of error variance
looks very implausible.
However, if we try
log(n
ij
) = + log(a
ij
) +
ij
P.M.E.Altham 22
it does now seem reasonable to assume var(
ij
) is constant. It seems that the
same line will t all 4 divisions, namely
log(n
ij
) = 0.01365(.6475) + 0.7506(.1285)log(a
ij
)
which has R
2
= .2748 (just a little better than before). Indeed the constant can
be dropped, to show that a straight line through the origin ts quite well.
The question about rogues and saints clearly relates to picking out those clubs
with high (positive or negative ) standardised residuals.
10. Let p
ij
=pyears = person-years at risk, and let d
ij
= the number of deaths,
for i = 1, . . . , 5, where i corresponds to age, j = 1(for non-smoker), j = 2 (for
smoker).
We assume d
ij
is distributed as indep. Po(
ij
), with the default link for ,
which is log( ). Further, we take
ij
=
ij
p
ij
, so that
log(
ij
) = log(
ij
) +log(p
ij
).
This is the interpretation of log(p
ij
) as the oset. We try various models for
log(
ij
).
The table suggests that
ij
increases with i, for each xed j.
The rst model tted is in eect
log(
ij
) = +
i
, for i = 1 . . . 5, j = 1, 2
with
1
=0. This model includes an age eect but not a smoking eect. This
model has deviance= 23.99(df = 5), but the t is not good enough.
But clearly this model shows that
i
increases with i: this makes sense.
Next we try the model
log(
ij
) = +
i
+
j
.
In this model, both smoking and age eects are present, but they operate
additively (thus the dierence between smokers and non-smokers is constant
over all 5 ages). The t is now much improved, and shows a clear eect of
smoking(

2
/(its se) > 0 ), but the t is still not acceptable (refer 12.13 to
2
4
).
The model corresponding to
glm(deaths ~ age:sm + offset(l), poisson)
would give a perfect t (since it is the saturated model) so therefore to improve
the t, we include an interaction term in a weaker way as our nal model. We
have
log(
ij
) = +
i
for sm =1, ie non-smoker,
log(
ij
) = +
i
+
j
+ i for sm=2, ie smoker.
This model ts very well (compare the deviance of 1.54 with
2
3
). It shows that
although smoking is bad for you (compute 1.445/.3729), there is an interaction
between smoking and age: the discrepancy between the mortality rates for
P.M.E.Altham 23
smokers and non-smokers DECLINES with age.
11. Write down (, ), and nd
.
It is easy to show that this has expectation zero.
12. Use the glm facts that E(Y ) = b
() = and var(Y ) = b
(). (Note added

November 2007: Wikipedia is also aware of this example.)
Chapter 3
Example Sheet 3
(Warning: somehow most people nd questions 2, 3 to be hard. There is a
practical point behind these 2 questions and the others too!)
1. The observed waiting times t
1
, . . . , t
n
are independent, with T
i
having pdf
f(t
i
| , ) = (t
i
/
i
)
1
e
ti/i
1
i
1
()
,
for t
i
> 0, where is a known parameter, and
i
depends linearly on a known
covariate x
i
through the following link function:
1
i
= +x
i
, for
i
> 0.
Let a
1
=
i
, a
2
= x
i
i
. Show that (a
1
, a
2
) are sucient for (, ).
The observations t
1
, . . . , t
m
are times between consecutive earthquakes in Mex-
ico City, and the observations t
m+1
, . . . , t
n
are the times between consecutive
earthquakes in Turkey. Take x
1
= . . . = x
m
= 0 and x
m+1
= . . . = x
n
= 1, and
discuss the estimation of when m, n are large, quoting any general asymptotic
likelihood results needed for your solution.
This introduces you to another error distribution available in glm: the gamma.
2. Your client gives you data
(y
ij
, x
ij
, 1 j n
i
, 1 i t)
and asks you to t the model
y
ij
= +x
ij
+
ij
,
24
P.M.E.Altham 25
with
ij
NID(0,
2
),
2
unknown. He has arranged the experiment so that
x
ij
= x
i
, 1 j n
i
, 1 i t.
Find expressions for ( ,

), the least squares estimators.
Your client now observes that a consequence of the model above is that
E(y
i
) = +x
i
, 1 i t,
where y
i
=
j
y
ij
/n
i
).
He suggests that some of your (highly paid) time could be saved by reading in
the data as the t pairs (y
i
, x
i
), 1 i t instead of the original (n
1
+ + n
t
)
pairs of points. How do you advise him? Give reasons for your answer.
[Hint: write down the likelihood given by
a) the full set of data (y
ij
, x
ij
), and
b) the reduced set of data (y
i
, x
i
).
Show that the maximum likelihood estimates of (, ) are the same for a)
and for b).]
3. Another client gives you data for binary regression, consisting of observations
(y, x) with the following structure:
(0, x
1
), (1, x
1
), (0, x
1
), (1, x
2
), (1, x
2
),
(1, x
2
), (0, x
2
), (1, x
3
), (1, x
3
), (0, x
3
),
(1, x
4
), (1, x
4
), (1, x
4
), (1, x
4
), (0, x
4
).
Thus there are 15 independent observations, with the rst digit of each pair
being 1 or 0 with probabilities p(x), q(x) respectively. She asks you to t the
model:
log(p(x)/q(x)) = +x
and to use the appropriate dierence in deviances to test the hypothesis = 0.
You observe that the data can in fact be compressed and read in as four inde-
pendent values
(1, 3, x
1
), (3, 4, x
2
), (2, 3, x
3
), (4, 5, x
4
).
(eg (1, 3, x
1
) means that of the 3 readings at x = x
1
, exactly 1 has value 1.)
Would this approach result in misleading your client? Give reasons for your
answer.
[Hint: think LIKELIHOODS, as in question 2 above.]
4. Suppose
Y = X +, N(0,
2
I)
where Y is n 1, X is n p and of rank p, and is the unknown vector of
parameters.
(a) Show that

= (X
T
X)
1
X
T
Y is the lse of .
P.M.E.Altham 26
(b) Show that

Y = X
= HY say, where H = H
T
and HH = H (
Y is the
vector of tted values).
(c) Suppose e = Y

Y (the vector of residuals). Show that
e N(0,
2
(I H))
and hence that
e
i
N(0,
2
(1 h
i
)), 1 i n
where h
i
is the i
th
diagonal element of H.
(d) Show that if is an eigenvalue of H, then is either 1 or 0. Hence show
n
1
h
i
= p. (The quantity h
i
is called the inuence or leverage of the i
th
data
point x
i
where y
i
=
T
x
i
+
i
, 1 i n.)
(e) Show that 0 h
i
1, and nd h
i
for simple linear regression, i.e. for
y
i
=
1
+
2
(x
i
x) +
i
.
(f) How do you interpret the statement Values of x
i
corresponding to large
leverage exert a pronounced eect on the t of the linear model at (x
i
, y
i
)?
(g) Use R to construct leverages (also called inuence values) and qqplots for
simple (y, x) data-sets.
5. (a) Suppose (y
i
) Mn(n, (p
i
)),

k
1
p
i
= 1. Consider testing
H
0
: log p
i
= +x
i
, 1 i k
where (x
i
) is given, and is such that p
i
= 1, thus
e
= 1/e
xi
.
Show that

is the solution of
() y
i
x
i
= nx
j
e
xj
/e
xj
.
Let e
i
= np
i
(
) (expected values under the null hypothesis).

(b) Now suppose instead that
y
i
independent Poisson(
i
), with
i
0.
Consider testing
HP
0
: log
i
=
+x
i
, 1 i k.
Find (,
) the loglikelihood function, and hence show that the mle for is
given by

, as in equation above. Show also that
k
i
(
) = y
+
(i.e. the observed and expected values of y
+
agree exactly at the mle). Com-
ment on the glm application of the Poisson distribution for problem (a).
P.M.E.Altham 27
S=1 D=1 89 2 4 1
S=1 D=2 8 4 3 1
S=2 D=1 70 6 2 0
S=2 D=2 1 0 1 1
A=1 A=2 A=1 A=2
B=1 B=1 B=2 B=2
Table 3.1: A 4-way contingency table
6. A random sample of 193 individuals, classied according to four 2-level fac-
tors, S, D, A, B respectively, gave the following 2 2 2 2 contingency table
as Table 3.1.
Let p
ijk
denote the corresponding underlying cell probabilities where i, j, k,
correspond to the factors S, D, B and A respectively. With the Poisson dis-
tribution for n, the cell frequencies, and log link, glm (in S-Plus) nds, in the
following order,
(a)n S D A B gives deviance = 0, df = 0,
(b)n (S +D +A+B) 3 gives deviance 2.72, df = 1,
(c) n (S +D+A+B) 3 S : A : BD : A : B gives deviance 3.42, df = 3,
(d) n (S +A+B) D gives deviance 8.48, df = 8.
Write down the model for p
ijk
for each of (a), (b), (c) and (d), give an inter-
pretation of the deviance at each stage, and give an interpretation in terms of
conditional independence for the model in (d).
How would you check the t of the model S, D, A, B mutually independent?
7. Which? (August 1980, p. 436) gives the data in Table 3.2 on lager (available
to you on catam stats). The columns are price per half pint, o.g. (original grav-
ity), percent alcohol, calories per half pint, and experts rating. The original
gravity is described by Which? as another guide to strength; its a measure
of what has gone into the beer besides water, and is used to calculate the duty
payable. The experts rating column contains entries 3, 4, 5 for lagers actually
tasted (5 being the most liked and 3 the least liked), and an entry 0 for lagers
not actually tasted. Reading the data provided on le, use
lm()
to answer the following questions:
(a) Does the price depend on o.g., percent alcohol, and calories per half pint,
and if so, how?
(b) Is the price of those lagers tasted signicantly dierent from the price of
those not tasted?
8. In a study on the possible relationship between the psychological well-being
of mothers and that of their children, the psychiatrist Prof I. Goodyer collected
data, some of which is summarised below. Table 3.3 shows, for the 200 children
in the study, how many in each of eight categories were cases, i.e. anxious or
P.M.E.Altham 28
Price og percent cal rating
17.5 1031 3.1 76 3
20.5 1032 3.2 79 3
22.5 1031 3.3 76 3
22.5 1032 3.3 78 0
18.5 1035 3.4 83 4
22.5 1033 3.5 78 3
23 1031 3.6 76 3
22 1033 3.6 82 3
19.5 1033 3.6 81 0
18.5 1033 3.6 81 0
22.5 1036 3.7 88 0
22 1034 3.7 83 3
24 1036 3.8 87 5
24 1036 3.8 87 0
18.5 1037 3.8 91 4
19.5 1038 3.9 91 5
25 1041 4.0 100 3
26 1036 4.0 84 3
20 1037 4.0 90 5
27 1037 4.0 91 5
22.5 1037 4.1 89 3
20.5 1038 4.1 92 0
43.5 1045 4.7 110 4
27.5 1045 4.8 109 4
29 1046 4.8 110 4
26.5 1047 4.9 116 3
29 1046 4.9 112 0
24.5 1047 4.9 116 0
18.5 1034 3.5 81 3
31 1045 5.0 109 0
32 1047 5.0 117 3
24 1046 5.0 111 4
29 1046 5.1 111 3
33 1046 5.1 110 0
29 1048 5.2 119 0
33.5 1050 5.4 121 0
26 1051 5.5 125 5
43 1058 6.0 146 0
31.5 1079 8.9 197 0
31.5 1081 8.9 204 0
Table 3.2: Lager data table
P.M.E.Altham 29
depressed, and the total number in each category. Those who are not cases
are controls, assumed well. The eight categories are dened by three binary
factors:
(a) rmq, which corresponds to a particular measure of the mothers psycholog-
ical well-being;
(b) mcr, which indicates whether or not the mother has good conding rela-
tions with other adults;
(c) events, which indicates whether or not the child has experienced recent
stressful life events in the 12 months prior to the study.
In each case a value of 1 for the factor corresponds to its status being good or
normal, and a value of 2 corresponds to its being poor.
Assuming that the number of cases in a given category is binomial, given the
total in that category, nd a model relating the probability that a child is a
case rather than a control to the three factors given. Discuss carefully how to
interpret your best-tting model, and its estimates, to the psychiatrist.
Table 3.4 shows the result of separating the cases into two categories, anx-
ious or depressed (dened to be mutually exclusive), and the eight further
categories are dened by the same three factors as before. Does the particular
diagnosis of a case (i.e. anxious rather than depressed) depend at all on any of
the factors?
[This illustrates the use of binomial regression.]
Case Total rmq mcr events
19 81 1 1 1
5 9 2 1 1
1 4 1 2 1
4 4 2 2 1
38 66 1 1 2
14 15 2 1 2
12 13 1 2 2
7 8 2 2 2
Table 3.3: Prof I.Goodyers rst dataset
anxious depressed Case rmq mcr events
13 6 19 1 1 1
5 0 5 2 1 1
0 1 1 1 2 1
2 2 4 2 2 1
28 10 38 1 1 2
7 7 14 2 1 2
9 3 12 1 2 2
4 3 7 2 2 2
Table 3.4: Prof I.Goodyers second dataset
P.M.E.Altham 30
9. Agresti (1990) Categorical Data Analysis, p. 377, gives the Table 3.5 below,
relating mothers education to fathers education for a sample of eminent black
Americans (dened as persons having a biographical sketch in the publication
Whos Who Among Black Americans). Here, for education,
Mother Father=1 Father=2 Father=3 Father=4
1 81 3 9 11
2 14 8 9 6
3 43 7 43 18
4 21 6 24 87
Table 3.5: Eminent Black Americans: educational levels of Mothers and Fathers
1 = 8th grade or less, 2 = Part High School, 3 = High School, 4 = College.
Let p
ij
= Pr(mothers education is i, fathers education is j), 1 i, j 4.
Consider the model
: p
ij
=
i
+ (1 )
i
j
, i = j
p
ij
= (1 )
i
j
, i = j.
where
i
= 1,
i
= 1,
j
= 1, and 0 < < 1.
Can you interpret to a sociologist?
Show that, under ,
log p
ij
= a
i
+b
j
, i = j,
for suitably dened a
i
, b
j
.
With n, the cell frequencies, declared as Poisson variables, with the default link
function, and with the two factors M.Ed and F.Ed each with the 4 given values,
glm nds that
n M.Ed +F.Ed has deviance 159.25, with 9 df.
Why should you expect this deviance to be so large?
But if we omit the 4 diagonal entries of the table, and t
n M.Ed +F.Ed, we nd that the deviance is 4.6199, with 5 df. How do you
interpret this?
Find the tted frequencies for this latter model.
10. You see below the results of using glm to analyse data from Agresti(1996,
p247) on tennis matches between 5 top women tennis players (1989-90). Let (r
ij
)
be the number of wins of player i against player j, and let n
ij
be the total number
of matches of i against j, for 1 i < j 5. Thus we have 10 observations, which
we will assume are independent binomial, with E(r
ij
) = n
ij
p
ij
.
The model we will t is
log(p
ij
/(1 p
ij
) =
i
j
, with
5
= 0.
equivalently
p
ij
=
i
/(
i
+
j
)
P.M.E.Altham 31
where
i
= e
i
.
The data can be read in (read.table( , header=T)) as the table
wins tot sel graf saba navr sanc
2 5 1 -1 0 0 0
1 1 1 0 -1 0 0
3 6 1 0 0 -1 0
2 2 1 0 0 0 -1
6 9 0 1 -1 0 0
3 3 0 1 0 -1 0
7 8 0 1 0 0 -1
1 3 0 0 1 -1 0
3 5 0 0 1 0 -1
3 4 0 0 0 1 -1
Thus for example, the rst row of numbers tells us that sel played graf for a
total of 5 matches, and sel won 2 of these.
The result of
glm(wins/tot~ sel+graf+saba+navr-1, binomial, weights=tot)
is deviance 4.6493, with 6 df.
The parameter estimates are
sel = 1.533(0.7870),
graf = 1.933(0.6783),
saba = 0.731(0.6770),
navr = 1.087( 0.7236),
with sanc =0 by assumption.
(i) Why do we impose a constraint on
1
,
2
,
3
,
4
,
5
?
(ii) Can we condently say that Graf is better than Sanchez?
(iii) Can we condently say that Graf is better than Seles?
(iv) What is your estimate of the probability that Sabatini beats Sanchez, in a
single match? (Answer: 0.6750)
(v) Table 3.6 gives corresponding data for 5 top men tennis players during 1989
90, taken from Agresti (1996, p255). Analyse them, tting a model of the same
form as above.
11. The purpose of this example is to introduce you to the topic of overdispersion
in the context of the Poisson distribution.
Suppose now that the observations Y
1
, ..., Y
n
are independent, with
E(Y
i
) =
i
, and var(Y
i
) =
i
, and log(
i
) = x
i
,
for some unknown and unknown scalar parameter .
Let
0
be the true value of this unknown parameter.
Our aim is to estimate , but is an unknown dispersion parameter. Clearly
P.M.E.Altham 32
Loser
Winner Edberg Lendl Agassi Sampras Becker
Edberg 5 3 2 4
Lendl 4 3 1 2
Agassi 2 0 1 3
Sampras 0 1 2 0
Becker 6 4 2 1
Table 3.6: The 5 top men tennis players in 1989-90
> 1 will correspond to over-dispersion relative to the Poisson. In the absence
of knowledge of , we choose our estimator

to maximise the function l
p
(),
where
l
p
() =
i
+x
i
y
i
+ constant.
(Thus l
p
() is in general not the correct loglikelihood function: we work out
below whether this is a serious problem.)
By expanding
l
p
()
evaluated at

, about
0
, show that (
0
) is approximately equal to (I(
0
))
1
U(
0
),
where
U() =
l
p
()
,
and
I() = x
2
i
exp x
i
,
and hence show that, approximately,
E(
) =
0
, and var(
) = (I(
0
))
1
.
Thus if > 1, the true variance of the

will be greater than the value given by
software which assumes the Poisson distribution.
R allows you to make this simple correction for overdispersion, by
summary(glm(y~x, poisson), dispersion =3.21))
where, for example, the dispersion has been estimated as (deviance/df), here
3.21. (A better estimate for may be obtained by replacing the deviance in
this formula by the chi-squared statistic, (o e)
2
/e, following the standard
notation.) Experiment with this.
Generalise your result to the case where , x
i
are vectors of dimension p.
1. Clearly log(f(t
i
|, )) = log(
i
) t
i
/
i
+ constant .
P.M.E.Altham 33
Hence loglikelihood is
L(, ) = log( +x
i
) t
i
x
i
t
i
+constant
where runs from i = 1 to i = n.
Hence, by the factorisation theorem, (a
1
, a
2
) is sucient for (, ).
Take x
i
= 0 for i=1,...,m, and x
i
=1 for i = m+ 1, ..., n.
Then
L(, ) = mlog() +(n m)log( +) (T
1
+T
2
) T
2
+constant
where T
1
= t
1
+... +t
m
, T
2
= t
m+1
+... +t
n
.
To estimate , solve L/ = 0, L/ = 0 for ( ,

).
We know that the asymptotic distribution of ( ,

) is bivariate normal, with
mean (, ), and covariance matrix V , say, where V
1
is the expectation of
the matrix of 2nd partial derivatives of L.
Now
L/ = (m/) +(n m)/( +) (T
1
+T
2
)
L/ = (n m)/( +) +T
2
Hence, after evaluating the matrix of 2nd derivatives of L, (whose elements turn
out not to be random variables), we see that V
1
is of the form
_
a b
b c
_
(where you will nd the values of a, b, c).
Hence, inverting V shows us that

is approximately N(, a/(ac b
2
)).
Thus the approximate condence interval for will be

2
_
(a/(ac b
2
)).
2. With the given model, we see that the mles of (, ) minimise
(y
ij
x
ij
)
2
where in , j = 1, ..., n
i
, i = 1, ..., t.
But x
ij
= x
i
, for all i, j so that the mles of (, ) minimise
(y
ij
x
i
)
2
giving

= S
xy
/Sxx where S
xy
= (y
ij
y)x
i
and S
xx
is dened similarly.
Now it is true that
y
i
= +x
i
+
i
say
where
i
is distd as NID(0,
2
/n
i
).
If we choose (, ) to minimise ( y
i
x
i
)
2
, we will obtain less accurate
estimators of (, ) than ( ,

). You can check that the estimators thus obtained
P.M.E.Altham 34
will still be unbiased, but will have larger variances.
However, if we choose (, ) to minimise
( y
i
x
i
)
2
n
i
then we will obtain ( ,

) as above, ie the correct estimators. So using only the
reduced data set (x
i
, y
i
, n
i
) in this way will be ne, as long as
2
is known.
If
2
is unknown, as is almost always the case in practice, then we will get a
more accurate estimator of it by using the ORIGINAL data set(and estimating
it as usual by (residual ss)/df) than by using the corresponding expression
when we have condensed the data down to (x
i
, y
i
, n
i
), namely
1
(t 2)
( y
i

x
i
)
2
n
i
.
Furthermore, if we only have the condensed dataset, we are much less likely
to be able to do an adequate job of checking the validity of the original linear
model with constant error variance (Sketch some simple examples with t = 3).
3. Write Y
ij
for the (i, j)th binary observation. We will assume that Y
ij
are
distributed independently as Bi(1, p(x
i
)) for j = 1, ..., n
i
, i = 1, ..., t.
Thus the loglikelihood of the data is
L(, ) = [y
ij
logp(x
i
) + (1 y
ij
)log(1 p(x
i
))]
where log(p(x
i
)/(1 p(x
i
)) = logit(p(x
i
)) = +x
i
. Hence
L(, ) = (y
i+
log(p(x
i
)) + (n
i
y
i+
)log(1 p(x
i
)).
Clearly, the loglikelihood is the same whether we enter the data as
(0, x
1
), (1, x
1
), (0, x
1
), .... ie 15 separate cases,
or as (1, 3, x
1
), .... ie 4 separate cases.
Thus, since ( ,

), and its asymptotic covariance matrix, are obtained purely
from the loglikelihood function, we will get the same answers for these whichever
of the 2 ways we choose to enter the data.
(But try a simple numerical example. Why do you get dierent expressions for
the deviances and their dfs ?)
4.a) Minimising (Y X)
T
(Y X) in gives
(X
T
X)
= X
T
Y
Hence, if X is of full rank,

= (X
T
X)
1
X
T
Y so that
b)

Y = X
= HY say, where
H = X(X
T
X)
1
X
T
.
It is easy to check that H = H
T
and HH = H,(ie H is a projection matrix).
P.M.E.Altham 35
c) e = Y X
= (I H)Y = (I H).
Now, distd as N(0,
2
I). Thus, using (I H)(I H)
T
= (I H) we see that
e is distributed as N(0,
2
(I H))
giving e
i
as N(0,
2
(1 h
i
)) as required.
d) Suppose u is an eigen vector of H corresponding to eigen value .
Then Hu = u, thus HHu = Hu thus Hu = Hu thus
Hu = 0 or = 1, so that = 0 or 1.
Hence h
1
+ ... + h
n
= trace(H) =
1
+
2
+ ... +
n
where (
i
) are the eigen
values of H. Hence
(h
1
+... +h
n
) = rank(H) = rank(X) = p.
e) Clearly, var(e
i
) = (1 h
i
)
2
, hence h
i
1.
Further,

Y = HY = H(X +),
hence var(
Y
i
) = h
i
2
hence h
i
0.
For the case of simple linear regression, the rst column of X is 1 (the vector of
1s) and the second column of X is say x
, where x
i
= x
i
x .
Hence X
T
X is diag(n, S
xx
) where S
xx
= (x
i
x)
2
.
Evaluating H from this shows that
h
i
= 1/n + (x
i
x)
2
/S
xx
f) We see from e) above that x
i
distant from x will give rise to relatively large
h
i
.(Sketch graphs of various possible congurations to get a feel for (h
i
)) .
5. The solution to this is in the lecture notes.
The practical importance is of course that there is no need for a multinomial
distribution within glm(): so long as we are tting log-linear models we can
use the Poisson distribution and log-linear models, provided that we include the
intercept term.
6. This example arose from some psychiatry data provided by Professor I.Goodyer
of Cambridge University. The 4 factors were
S = 1, 2 for girls,boys
D = 1, 2 for depression no or yes
A = 1, 2 for anxiety symptoms no or yes
B = 1, 2 for behavioural symptoms no or yes.
Thus the table is in fact rather sparse for the large sample theory to be realistic,
but we give this analysis as an illustrative example of the glm( ) modelling.
Sat : log(p
ijkl
) = +
i
+... + ()
ij
+... + ()
ijk
+... + ()
ijkl
is the saturated model (we assume the usual glm() constraints
1
= 0 etc.
a) ts the saturated model (ie 2
4
parameters for 2
4
observations) so that we
expect to get deviance 0 with df = 0.
P.M.E.Altham 36
However a quick glance at the parameter estimates and their ses for this model
will usually suggest to us which of the high-order interactions(if any) can be
dropped. Our object is to t the simplest possible model which is consistent
with the data, and of course we want to interpret this model to the scientist
who provided the data.
b) ts Sat with
H
0
: ()
ijkl
= 0
ie no 4th order interaction. Refer the increase in deviance (2.72) to
2
1
to test
H
0
(applying Wilks theorem). The result is non-signicant, so we accept H
0
.
c) ts Sat with H
0
and H
1
: ()
ikl
= 0, ()
jkl
= 0.
Refer the resulting increase in deviance (ie 3.422.72 ) to
2
2
to test H
1
assuming
H
0
true. You nd that you accept H
1
.
d) ts
H
2
: log(p
ijkl
) = +
i
+
j
+
k
+
l
+ ()
ij
+ ()
jk
+ ()
jl
so we have dropped yet more parameters in moving from c) to d). The increase
in deviance is (8.48 3.42) which is non-signicant when referred to
2
5
so
dropping these extra parameters was permissible.
Furthermore, we can assess the t of the model H
2
by referring 8.48 to
2
8
. In
fact 8.48 is almost the same as E(
2
8
), so you see that H
2
is a good t. You
could check that no further parameters can safely be dropped.
The nal model is therefore p
ijkl
= a
ij
b
jk
c
jl
for some a, b, c. We will rewrite
this in a more enlightening way as
Pr(A, B, D|S) = Pr(A|D)Pr(B|D)Pr(D|S).
(We choose to write it in this asymmetric form this since S is clearly not a
response variable.) We can express this in words as
D depends directly on S, but B and A only depend on S through D. Further-
more, conditional on the level of D, the variables B and A are independent.
This is a simple example of a graphical model of conditional independence.
Draw a graph in which the 4 points S, D, B, A are connected only by the arcs
SD, DB, DA, as in Figure 3.1.
Finally, the model in which S, D, B, A are independent corresponds to a graph
with no links at all between the 4 points, and in glm terms it is
H
I
: log(p
ijkl
) = +
i
+
j
+
k
+
l
.
7. a) As with most regression problems, it is best to START by doing some sim-
ple plots: in this case all possible pairwise plots of the 4 variable concerned.This
reveals that price is positively related to each of og, percent and calories, but
also, as a student doing this problem in an examination wrote,
as any experienced lager-drinker knows, the 3 variables og, percent and calo-
ries are all measuring more or less the same thing, so we could not reasonably
P.M.E.Altham 37
S D
A
B
Figure 3.1: The graph showing the interdependence of the S, D, A, B data
expect 2 or 3 of them to give a much better prediction of price than just one of
them.(Were teetotal students at a disadvantage in this examination?)
Linear regression of price on these 3 explanatory variables bears out this ob-
servation. In fact, og is the best single predictor. Inspection of the residuals
reveals that some lagers are priced at a ridiculously high level, but perhaps they
are connoisseurs lagers, ie not aimed at those interested only in the alcohol
content.
b) This question is asking the student to set up a 2-level factor (tasted/not
tasted) and to do the regression of price on og, percent, calories and this new
2-level factor.
The SPlus6 output and corresponding graph are given at the end of these solu-
tions. (You could use R to the same eect.)
8. Setting up the 3 factors rmq, mcr, events, each with 2 levels (level 2 being
the bad one ) with binomial regression of case on these 3, and the logistic link
function, and total as the binomial denominator, gives
log(P(case)/P(control))
= 1.248(.253)+1.716(.514)rmq(2)+1.257(.561)mcr(2)+1.620(.330)events(2)
with deviance 5.053, df = 4.
Thus this model ts well, and none of the 3 factors can be dropped from the
model. Either compare each estimate with its se, or compare the increase in
P.M.E.Altham 38
deviance with
2
1
when each factor is dropped from the model in turn.
Note: having rmq at level 2 rather than level 1 increases the odds in favour of
being a case by a factor of exp(1.716). Having all of rmq, mcr, events at the
bad level rather than the goodlevel increases the odds of being a case by
a factor of exp(1.716 + 1.257 + 1.620). We could use the covariance matrix of
these estimates to nd the corresponding condence interval for this odds ratio.
Table 2 is analysed similarly.
9. Let M be the Mothers educational level and F the Fathers. Under ,
(M, F) = (M
1
, F
1
) with probability , where M
1
= F
1
and P(M
1
= i) =
i
.
(M, F) = (M
2
, F
2
) with probability 1 , where M
2
, F
2
are independent, with
P(M
2
= i) =
i
, P(F
2
= j) =
j
,
ie, with probability , the mothers education MUST be identical to the fathers,
and with probability 1 , they are independent (in which case there is still a
chance that they are identical).
It is easily seen that is equivalent to the model
log p
ij
= a
i
+b
j
for i = j.
If we set up ma, pa as the 4-level factors corresponding respectively to mothers,
fathers educational level, then using the Poisson family and log link function the
glm of n(the frequency) on (ma+pa), gives a deviance of 159.25,which is clearly
hugely signicant compared with
2
9
. This tests the hypothesis of independence
of the factors ma, pa, so common-sense suggests that this hypothesis is unlikely
to hold: in any case, a glance at the 4 4 contingency table shows that the
diagonal entries are much too large for the independence hypothesis to be plau-
sible.
Now omit the entries of the table for which i = j (use
subset= (i!=j)
in glm() command). Now the deviance is much the same as its expected value,
showing that is a good t. From the table of tted values we could nd

,
and so on.
This model is what sociologists call a mover-stayer model; it is also an example
of a quasi-independence model for a contingency table.
10. These data on womens and mens tennis matches are taken from An intro-
duction to categorical data analysis by Alan Agresti (1996) and their analysis
is discussed in Agrestis Chapter 9.
i) The model being tted is:
w
ij
(the number of wins) is distd as independent Bi(t
ij
, p
ij
) for1 i < j 5
with logit(p
12
) = +sel graf, logit(p
13
) = +sel sab etc
but the import of the 1 term in the glm statement is that we set = 0.
Hence our model is logit(p
ij
) =
i
j
for 1 i < j 5.
But note that with this model we could replace eg,
5
by
5
+ 13, and then
P.M.E.Altham 39
replace
4
by
4
+ 13 and so on, without changing the formula for logit(p
ij
).
So we impose a constraint, without loss of generality
5
= 0, to ensure param-
eter identiability .
(If we include a term + sanc in our tting, glm( ) will estimate sel, graf, saba,
navr as given, but will obligingly tell us, in eect, that sanc is aliased, mean-
ing that it cannot be estimated if the previous 4 parameters are already in the
model.)
This is an example of the Bradley-Terry model for paired comparisons.
ii) Can we condently say that Graf is better than Sanchez?
YES, because the model ts well (refer its deviance of 4.65 to
2
6
) and
2
/se(
2
)
= 2.854: refer this to N(0, 1).
iii) Can we condently say that Graf is better than Seles?
Now
2

1
= 1.933 1.533.
and the estimated variance of (
2

1
) = (.6687)
2
(Use the parameter estimate correlation matrix)
Referring (
2

1
)/.6687 = 1.886 to N(0, 1), we see that the dierence is sig-
nicant (but not spectacularly so).
iv) Our estimate of the probability that Sabatini beats Sanchez, in one match,
is
e
.7308
/(1 +e
.7308
) = 1/(1 +e
.7308
) = .675
and note that using .7308 1.96 .6764
we could attach a condence interval to this probability
ie 1/(1 +exp(.7308 1.96 .6764)).
v) The dataset for the mens tennis doesnt t the Bradley-Terry model so well.
Note that our underlying model contains the assumption
w
ij
distributed independently as Bi(t
ij
, p
ij
).
This may not be reasonable. For example, if Graf beats Sanchez in the rst
match, then this result may aect the probability that Graf beats Sanchez in
their next match, and so on. Further, we should perhaps take account of the
surface on which the match is played (grass, clay, etc). But we do not have the
data to be able to take account of such factors in our modelling.
11. Take
0
as the true value of . The usual Taylor series expansion of the rst
derivative of l
p
() at

about
0
shows that, to rst order,
0 = U(
0
) (

0
)I(
0
)
which gives the required approximation for (

0
).
Now, as usual, E(U(
0
)) = 0, giving us the required expression for E(
). Fur-
ther, by noting that
var(
) = var(x
i
Y
i
)/(I(
0
))
2
and substituting var(Y
i
) =
i
, we obtain the required expression for var(
).
..........................
P.M.E.Altham 40
For number 7 (lager data), Figure 3.2 gives the corresponding pairs-plot. Here
is the corresponding code and output.
>lager <- read.table("lager", header=T)
>summary(lager)
>pairs(lager)
>attach(lager)
> first.lm <- lm(Price ~ og + percent + cal)
>summary(first.lm,cor=F)
Call: lm(formula = Price ~ og + percent + cal)
Residuals:
Min 1Q Median 3Q Max
-6.953 -3.217 -0.2989 1.827 16.53
Coefficients:
Value Std. Error t value Pr(>|t|)
(Intercept) -1430.4215 1487.7819 -0.9614 0.3427
og 1.4393 1.4849 0.9693 0.3389
percent -2.4402 5.0846 -0.4799 0.6342
cal -0.3201 0.5135 -0.6233 0.5370
Residual standard error: 4.856 on 36 degrees of freedom
Multiple R-Squared: 0.4136
F-statistic: 8.465 on 3 and 36 degrees of freedom,
the p-value is 0.0002186
>summary(first.lm,cor=T)
Call: lm(formula = Price ~ og + percent + cal)
Residuals:
-6.953 -3.217 -0.2989 1.827 16.53
Coefficients:
(Intercept) -1430.4215 1487.7819 -0.9614 0.3427
og 1.4393 1.4849 0.9693 0.3389
percent -2.4402 5.0846 -0.4799 0.6342
cal -0.3201 0.5135 -0.6233 0.5370
Correlation of Coefficients:
P.M.E.Altham 41
(Intercept) og percent
og -1.0000
percent 0.4782 -0.4781
cal 0.9217 -0.9218 0.1035
>next.lm <- lm(Price ~ og); summary(next.lm)
Call: lm(formula = Price ~ og)
Residuals:
-7.617 -3.135 -0.3271 2.045 16.89
Coefficients:
(Intercept) -336.5331 71.2697 -4.7220 0.0000
og 0.3475 0.0684 5.0802 0.0000
the p-value is 1.033e-05
>rat <- (rating>0)*1 ; Rat <- factor(rat)
>third.lm <- lm(Price ~ og + Rat); summary(third.lm)
Call: lm(formula = Price ~ og + Rat)
Residuals:
-7.609 -3.343 -0.22 1.934 17.05
Coefficients:
(Intercept) -330.0376 78.3167 -4.2141 0.0002
og 0.3415 0.0748 4.5667 0.0001
Rat -0.3635 1.7006 -0.2138 0.8319
the p-value is 6.696e-05
>last.lm <- lm(Price ~ Rat); summary(last.lm)
Call: lm(formula = Price ~ Rat)
Residuals:
-9.067 -4.405 -0.78 3.273 19.22
Coefficients:
P.M.E.Altham 42
Price
1030 1040 1050 1060 1070 1080 80 100 140 180
2
0
3
0
4
0
1
0
3
0
1
0
5
0
1
0
7
0
og
percent
3
4
5
6
7
8
9
8
0
1
2
0
1
6
0
2
0
0
cal
20 25 30 35 40 3 4 5 6 7 8 9 0 1 2 3 4 5
0
1
2
3
4
5
rating
Figure 3.2: The pairs plot for the lager data
(Intercept) 27.5667 1.5370 17.9359 0.0000
Rat -3.2867 1.9441 -1.6906 0.0991
>tapply(Price,Rat,mean)
0 1
27.56667 24.28

All Ex Sol

Uploaded by

Copyright:

Available Formats

All Ex Sol

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

All Ex Sol

Uploaded by

Copyright:

Available Formats

Introduction to Generalized Linear Modelling,

Example Sheets 1, 2 and 3, with solutions

are NID(0, 1).

) for i = , (i, = 1). (Here y

= 0 has to be solved iteratively. The asymptotic covariance

(). Further the matrix of 2nd derivatives of the loglikelihood

l = 0 must be THE MAX-

)) for short; our (1 ) condence region is

(). (Note added

) (expected values under the null hypothesis).

You might also like