THE ROYAL STATISTICAL SOCIETY

GRADUATE DIPLOMA EXAMINATION

NEW MODULAR SCHEME
introduced from the examinations in 2009

MODULE 5
SOLUTIONS FOR SPECIMEN PAPER B
THE QUESTIONS ARE CONTAINED IN A SEPARATE FILE

The time for the examination is 3 hours. The paper contains eight questions, of which
candidates are to attempt five. Each question carries 20 marks. An indicative mark
scheme is shown within the questions, by giving an outline of the marks available for
each part-question. The pass mark for the paper as a whole is 50%.
The solutions should not be seen as "model answers". Rather, they have been written
out in considerable detail and are intended as learning aids. For this reason, they do
not carry mark schemes. Please note that in many cases there are valid alternative
methods and that, in cases where discussion is called for, there may be other valid
points that could be made.
While every care has been taken with the preparation of the questions and solutions,
the Society will not be responsible for any errors or omissions.
The Society will not enter into any correspondence in respect of the questions or
solutions.
Note. In accordance with the convention used in all the Society's examination papers, the notation log denotes
logarithm to base e. Logarithms to any other base are explicitly identified, e.g. log10.

RSS 2008

Graduate Diploma Module 5, Specimen Paper B. Question 1

(i)

The original variables, the answers to the questions, are likely to be highly
correlated. Principal component analysis (PCA) gives linear combinations of
the variables that are uncorrelated. The first PC accounts for the largest
amount of variation in the data, the second for the next largest, and so on. If
the questions form themselves into relatively distinct clusters then PCs are
useful to define subsets, and possibly to suggest ways of combining scores.
PCs are only strictly valid for numeric data, but the data here are nearer to
being categorical at best ordinal. However, PCA is often used for data such
as these.

(ii)

A cluster analysis could be useful, using correlations (or absolute values of

them); perhaps indications of the grouping of questions would be given.

(iii)

PCA only works on complete records. If a respondent's answer to one

question is missing, that whole set of responses will be omitted. Because PCA
is based on analysis of variability in data, missing values cannot easily be
imputed. The choice in this case is between analysing a large number of
responses on a small number of questions and a small number of responses on
a large number of questions. The strategy proposed seems sensible.

(iv)

The first three eigenvalues add to 5.14, i.e. 5.14/6 or 85.7% of the total
variation, and should be enough.
The first PC (54% of total variation) is an overall score of concern about cost
note that the "direction" of questions 2, 3, 4 is opposite to that of 1, 5, 6. The
second PC (23% of total variation) measures the tendency of respondents to
answer all questions in the same way, i.e. with similar scores. The third PC
(9% of total variation and so relatively much less important) is dominated by
question 4, perhaps contrasting its answers with those for question 2, perhaps
also taking question 5 into account. The first two PCs therefore give most of
the useful, easily understood, information.

(v)

The two unsatisfactory features of the data are the large amount of missing
information, leading to 9 of the 15 questions being discarded, and the
suggestion from the second PC that the respondents do not complete the form
validly. Hence these results are not reliable. A fresh start is needed, with
reworded questions and boxes to tick as in a survey.

Graduate Diploma Module 5, Specimen Paper B. Question 2

(i)

Linear discriminant analysis can be used to produce a classification rule where

the groups are known a priori, and data are described by several variables.
Linear combinations of these variables xi can show up relations not obvious
from separate, univariate, analyses. Classifications so found can be applied to
the new sites.

(ii)

Multivariate Normal variance-covariance matrices are required to be equal for

each group (but locations will be different). This is not easy to check;
although formal tests exist, they are sensitive to non-Normality. Also,
relatively small sample sizes do not help. Univariate Normality for each
measurement can be checked in the usual ways (e.g. histograms, stem-and-leaf
plots, Normal probability plots); univariate Normality is a necessary but not
sufficient condition for multivariate Normality.

(ii)

The variance-covariance matrices are apparently different, with changes in

sign as well as size of individual entries. Normality cannot be checked on the
information given.
The means of x1 and x4 (and possibly x5) appear different for the two groups.

(iv)

Method 1
After constructing and applying the discriminant function, 14/17 (1) and 12/15
(2) are found to have been correctly classified. This is good, but is likely to be
an overestimate of the future success rate (since the same data have been used
to construct the function and to "check" it).
Cross-validation may be carried out by, for example, a jack-knife method:
calculate the function omitting one observation, and use the function to predict
class membership of that item; repeat this for each item in turn and observe
the number of correct predictions. [In a large data-set, the discriminant
function would be calculated on some of the data and then used to check the
success rate of the remainder. Here we do not have enough data for that.]
This gave 12/17 (1) and 9/15 (2) correct.
Method 2
Note that x4 was identified in (iii) as a useful variate. This method correctly
classifies 12/17 (1) and 12/15 (2), and the numbers on cross-validation are the
same. This seems the better method.
With these sample sizes, using 5 variables (Method 1) may be over-fitting.
The univariate (as it has turned out) Method (2) is more successful.

Graduate Diploma Module 5, Specimen Paper B. Question 3

If f (t) is the probability density function and F(t) the cumulative distribution function
for the lifetime, then the hazard function h(t) is defined by h(t) = f (t)/(1 F(t)).
The hazard function in this case is
h(t ) = h0 (t ) exp( 1 x1 + 2 x2 + 12 x1 x2 )
where h0(t) is the baseline hazard function.
(i)

The log likelihood (log L) is not given for the null model, but the coefficient
for x1 in model A (which contains x1 only) is large relative to its standard
error; further, the hazard ratio is high. These suggest that x1 is important.
The difference in 2logL between models B and C is small and certainly not
statistically significant. Thus it seems likely that there is not an interaction
effect (the interaction term is the only difference between the two models).
The difference in 2logL between models A and B is 2(71.518 75.640) =
8.244. This is significant as an observation from 21 . So we conclude that the
best model is model B.

(ii)

Hazard ratio = exp(coefficient).

For x2 in model B, a 95% confidence interval for the coefficient is given by
1.172 (1.96 0.429) , i.e. it is (0.331, 2.013).
The corresponding 95% confidence interval for the hazard ratio is from
exp(0.331) to exp(2.013), i.e. from 1.393 to 7.484.

(iii)

The fault reduces the expected lifetime. Given two items, both with the same
level of the chemical (x1), the one with the fault is about 3 [exp(1.172)] times
as likely to fail at any time.

(iv)

(1.528 1.4) + 1.172 = 3.311

(1.528 2.9) + 0 = 4.431

So the second is more likely to fail first.

(v)

If the proportional hazards assumption is not valid then the model is deficient.
In particular the interpretation of the hazard ratios is invalid, and the answers
to parts (iii) and (iv) may be inaccurate. The consequences depend to some
extent on the type and seriousness of any departures from the assumptions.

Graduate Diploma Module 5, Specimen Paper B. Question 4

(i)

The survival time of an individual is censored when the end-point of interest

(death in this example) has not been observed, either because the trial is
terminated before the end-point took place or because the individual has been
lost to the trial for some reason (e.g. does not respond to follow-up). The
phrase right-censoring refers to the censoring occurring after (i.e. to the right
of, in natural time order) the last known survival time.

(ii)

The Kaplan-Meier survival curve is constructed as follows. The word "death"

is used in this description generically; here it does in fact refer to death, but in
other examples it might be healing, general recovery, etc.
We seek the estimated cumulative survival function S (t ) .
The Kaplan-Meier method requires the ordered death times t(1), t(2), , t(r) to
be considered. For j = 1, 2, , r, let n(j) be the number of individuals alive
just before time t(j), and let d(j) be the number of deaths at t(j).

An estimate of the probability of survival from t(j) to t(j + 1) is

n( j ) d ( j )
n( j )

Thus (assuming independence) the probability of surviving through all the

intervals up to t(k + 1) is estimated by
k n
d( j )
S ( t ) = ( j )
,
n( j )
j =1

and this is the Kaplan-Meier estimate.

If the largest survival time [t(r)] is censored, the method above is used to give
estimates up to and including the next largest, the value for which is then
assumed to apply for all times onward. If the largest survival time is not
censored, the estimate drops to zero at that point.
In the present example, t(r) is not censored.
The calculation is shown in detail on the next page.

Solution continued on next page

There are 24 patients. The calculation is shown in detail in the table below.
Some of the detail might be omitted in practice, and is often not shown in
computer output. Rows for the censored observations have been omitted, but
care must be taken to ensure that n(j) is always correct. Users of this solution
should carefully verify the values in the table by reference to the data in the
question.
Time n(j) as defined in
t(j)
text above
[i.e. number
remaining just
before time t(j)]
6
8
12
20
24
30
42

23
19
17
10
8
4
1

Solution continued on next page

d(j) as defined in
text above
[i.e. number of
events at time
t(j)]

n( j ) d ( j )
n( j )

4
2
2
1
1
1
1

19/23
17/19
15/17
9/10
7/8
3/4
0

Cumulative
survival
estimate
S (t )
at each t(j)
0.8261
0.7391
0.6522
0.5870
0.5136
0.3852
0

Greenwood's formula for the standard error for the Kaplan-Meier estimate at
12 months follow-up (corresponding to the third row in the calculation above)
is
1

3
2
dj

SE = S (12 )
j =1 n j ( n j d j )

2
2 2
4
= 0.6522
+
+

23 19 19 17 17 15

= 0.6522 ( 0.009153 + 0.006192 + 0.007843)

1/2

= 0.6522 0.023188 = 0.0993 .

(iii)

The interval is 0.6522 (1.96 0.0993) = 0.6522 0.1946 , i.e. (0.46, 0.85).

(iv)

From the graph, the median survival time is 30 months.

(v)

A log rank test could be used for this purpose.

Graduate Diploma Module 5, Specimen Paper B. Question 5

(a)

The infant mortality rate is

number of deaths between birth and one year
(excluding fetal deaths, stillbirths)
.
total number of live births in the same year
The neonatal mortality rate is as above but only including deaths up to 28
days.
The perinatal mortality rate is
number of fetal deaths and neonatal deaths
total number of live births
[sometimes divided by the total number of live births and fetal deaths, there
being no generally accepted convention for computing this rate].
The maternal mortality rate is
number of deaths from puerperal causes
total number of live births
All these rates are usually multiplied by 1000.

Solution continued on next page

(b)

Age
100

Males

Females

90

80

70

60

50

40

30

20

10

350

300

250

200

150

100

50

50

100

150

200

250

300

350

Totals (thousands) in ten-yearly age groups

The 01 frequencies are multiplied by 10; those for 14 by 10/4; the
choice of upper limit for the 85+ group is made so as not to distort the
pyramid. [NOTE. The accuracy of representation on the diagram is
constrained by the limits of electronic reproduction.]
(c)

(i)

The sex-age-specific death rate for a country is

number for that sex in age-range of
interest during 1 year

1000

average number of persons of that sex and

age living during the year
This is calculated separately for males and females, using suitable age
ranges. "Average" (mid-year) is usually the mean of beginning and
end figures for that year.
(ii) The rates for U are generally higher than for D up to 44, and
then become lower.
The rates for males are generally higher than for females, in both
countries.
Females thus have longer expectation of life than males, and
inhabitants of D have longer expectation of life than those of U.

Graduate Diploma Module 5, Specimen Paper B. Question 6

(i)

This is because a case-control study is retrospective, whereas relative risk is

measured in a prospective study.
A retrospective study takes affected persons and explores in detail the history
of events that may have led to the condition. For example, it may thereby
enquire into whether most of those affected by a cholera epidemic have
consumed water from the same source. A retrospective study relies on having
fairly complete and reliable data on a variety of topics.
A prospective study begins with unaffected persons (e.g. those without lung
cancer), notes various characteristics (e.g. smoking habits, occupation, place
of residence) and studies future development of the condition in relation to
those characteristics. Thus it may enquire into whether the condition develops
more frequently in some groups than in others.

(ii)

Odds =

P (event happens)
.
1 P(event happens)

The odds ratio is the ratio of odds of disease in the exposed group of patients
(i.e. here the smokers) to that in the unexposed group, i.e. here

odds ratio =

(iii)

P(disease, smoker)
1 P(disease, smoker)
P(disease, non-smoker)
1 P(disease, non-smoker)

The combined data are as follows.

Cases
Controls
Total

Smokers
89
13
102

Non-smokers
394
434
828

Total
483
447
930

Using the relative frequencies from this table, the odds ratio may be calculated
as
89 434
= 7.54
13 394
which is substantially greater than 1 and indicates greater prevalence of cancer
among smokers.

Solution continued on next page

(iv)

The Mantel-Haenszel method is a simple way of adjusting for another factor,

in this case sex. [Note. Other methods for doing this are used in some
computer programs.]
a
c
with a + b + c + d = n, and keeping the
b d
two sexes separate as in the question, the Mantel-Haenszel estimate of the
odds ratio is

Representing each table by

a d /n
b c / n
i

i i

where i = 1, 2 for males, females.

This gives
58 271 31163
+
41.537
580
350
=
= 7.53 .
5.514
6 245 7 149
+
580
350
This is virtually the same as for the pooled data (7.54). This often turns out to
happen when both subsets of the data are large and of the same order of size;
also, in this case, we might suppose that sex is not in fact an important factor.
To obtain a 95% confidence interval for the odds ratio, we work via
logarithms and first use the pooled data to obtain the standard error of the log
odds ratio using the formula
Var(log of odds ratio) =

1 1 1 1
+ + +
a b c d

(= 0.093001 here) so that the standard error is 0.093001 = 0.305. The log of
the Mantel-Haenszel estimate of the odds ratio is log(41.537/5.514) = 2.0193,
so the 95% confidence interval for the logarithm is given by

2.0193 (1.96 0.305) = 2.0193 0.5978 ,

i.e. it is (1.421(5), 2.617), and thus the interval for the odds ratio itself is
(4.14, 13.69).
(v)

The confidence interval does not contain 1.00, so we may reject the null
hypothesis that smoking status and occurrence of lung cancer are unrelated.
There is definite evidence of an association. As mentioned above, the odds
ratio strongly indicates greater prevalence of cancer among smokers than
among non-smokers. It does not appear that sex is an important factor in this.

Graduate Diploma Module 5, Specimen Paper B. Question 7

Part (i)

( ) ( )

( )

( )

E R E ( x ) = E ( y ) E R E ( x ) = Y X E R .
Cov R , x = E Rx

(a)

( )

1
Y
This gives E R = Cov R , x + ,
X
X

f =

(b)

( )

i.e. E R R =

Cov R , x
X

).

n
y
= y or y Rx
=0.
and R = , so that Rx
x
N

( )

Hence the estimator of Var R given in the question is

1 f 1
.

nx 2 n 1

{( y y ) R ( x x )}
i

{ ( y y )

1 f 1
.
nx 2 n 1

1 f
2
2
sY 2 R s X sY + R 2 s X
2
nx

2
2 R ( yi y )( xi x ) + R 2 ( xi x )

2
2
in which sY , s X are the estimated variances of Y and X, and is the estimated
correlation coefficient for X and Y.

Part (ii)
The ratio method works well when Y is proportional to X, with the relation passing
through the origin. It will not be better than a simple random sample when is less
than 0 or when the relation does not pass through the origin (in which case a
regression estimator is required instead).

See next page for solution to (iii)

Part (iii)
(a)
The sugar content of an individual fruit should be roughly proportional to its
weight, in fruit from the same source and batch.
(b)
Since we are not told N, the total number of oranges, a ratio estimator is used
rather than regression. Counting the whole batch would take a very long time for
what might be a very small improvement in precision.

x = 1975, y = 110.9.
y y
R = =
= 0.05615.
x x

( )

X T = 820.

= 46.045 (kg).
YT = RX
T

( )

2
We have Var YT = X T Var R .

Also, on neglecting f which will be very small (as n is only 10), we have that the value

( )

of the estimator of Var R is

1 1
2
y + R 2 x 2
. yi 2 Rx
i i
i
2
10 x 9

1
1
2
.
1268.69 2 0.05615 22194.8 + ( 0.05615 ) 392389
2
90 (197.5 )

1
13.34687
(1268.69 2492.476 + 1237.133) =
351056.25
351056.25

( )

which on multiplying by (820)2 gives that the value of the estimator of Var T is
25.5641, i.e. the standard error is 5.056.
(c)
The half-width of the interval, ts / n , is to be less than 2. Thus s / n < 1
and 25 oranges will achieve this approximately.

Graduate Diploma Module 5, Specimen Paper B. Question 8

[solution continues on next page]
(i)
As is clear from the data, the six strata split into three with fairly low density
of caribou and three with much higher density. There are also some variations in the
values of sh between the six strata. Stratified sampling ensures that all these six strata
will be represented adequately, and that an estimate of the total number of animals
will have a smaller standard deviation than for simple random sampling.
L

(ii)

The estimated total is Yst = N yst = N h yh (where there are L strata), so

h =1

Yst = ( 400 24.1) + ( 40 25.6 ) + (100 267.6 ) + ( 40 179.0 )

+ ( 70 293.7 ) + (120 33.2 ) = 69127 .

The estimated variance of Yst is given by

2

sh
74.7 2
= 400 ( 400 98 )
+ ... = 84123268.3 ,
N h ( N h nh )

98
nh
h =1
L

so the estimated standard error is 9171.9.

(iii)
Nh is the true number in stratum h. Sh is the true standard deviation in stratum
h. wh is nh/n, the proportion of the whole sample that comes from stratum h. V is the
value specified for Var( Yst ).
(iv)
Optimal allocation minimises the variance Var( Yst ) (equivalently, Var( yst ))
for fixed total sample size n.
As well as allocating more sampling to strata with larger population sizes, it allocates
more to those with larger standard deviations, so the precision is comparable with
those having lower variability. In the present survey, there are wide variations among
the stratum sizes and standard deviations; proportional allocation with the same
sample size as optimal allocation is likely to lead to a considerably larger value of
Var( Yst ).
(v)
We use the formula quoted in part (iii) of the question, taking the estimates sh
from the preliminary aerial survey as though they were the true values Sh.
Optimal allocation with constant cost of sampling any unit has wi (= ni/n) given by
NS
wi = L i i . We have NhSh = 133903, using the preliminary survey values.
N h Sh
h =1

Further, we see that

N h Sh
wh

= ( N h S h ) N h Sh , so that
h =1

N S
hw h
h

( N S )
h

Also we have NhSh2 = 47882186 (this appears in the denominator of the formula).
Finally, we need V. The criterion of d = 8000 with (one-sided) tail probability 0.025
gives V = (8000/1.96)2.
n =

(133903)

8000

+ 47882186
1.96

= 277.804 .

So we take n = 278. The allocation in each stratum is then given by

ni = 278wi = 278

N i Si
,
N h Sh

which gives n1 = 62.03, n2 = 5.29, n3 = 122.39, n4 = 12.54, n5 = 51.08, n6 = 24.66.

However, the total size of stratum 3, N3, is only 100; so we must take n3 = 100.
The remaining 178 are then allocated in the same ratios as before, by multiplying each
by 178/155.61. This gives n1 = 70.96, n2 = 6.05, n4 = 14.34, n5 = 58.43, n6 = 28.21.
Finally,
n1 = 71, n2 = 6, n3 = 100, n4 = 14, n5 = 58, n6 = 28.

With this allocation, the estimated variance of Yst is given by

L

N h ( N h nh )
h =1

sh
74.7 2
= 400 ( 400 71)
+ ... = 18610118.04
71
nh

(note there is a ZERO contribution to the sum from stratum 3, where we have a 100%
sample), so the estimated standard error is 4313.9.