THE UNIVERSITY OF AUCKLAND

SECOND SEMESTER, 2004

Campus: City

STATISTICS

Advanced Statistical Modeling

Topics in Statistics 3

(Time allowed: THREE hours)

INSTRUCTIONS

SECTION A: Multiple Choice (40 marks)

Answer ALL 20 questions on the answer sheet provided.

All questions have a single correct answer and carry the same mark value.

If you give more than one answer to any question you will receive zero marks for that
question.

Incorrect answers are not penalized.

SECTION B (60 marks)

Answer ALL 3 questions. Each is worth 20 marks.

CONTINUED
 SECTION A

1. A data set consist of measurements on three variables X, Y and Z. The variables X

and Y are continuous and Z is categorical. Which of the following plots would you
expect to give the best picture of the relationship between the variables?

(1) A barchart, with bars corresponding to the frequencies of X and Y .

(2) A coplot corresponding to the formula Y Z | X.
(3) A coplot corresponding to the formula Y X | Z.
(4) A scatterplot of X versus Z, with the value of Y shown by a colour coding.
(5) A trellis plot consisting of panels corresponding to values of X, and each panel
containing a dot plot.

2. The coplot in Figure 1 displays the relationship between three continuous variables Y ,
X1 and X2. The plot was produced by the R code coplot(YX1|X2). Which of the
following statements is TRUE?

Given : X2
0.0 0.2 0.4 0.6 0.8

0.0 0.4 0.8 0.0 0.4 0.8

2.5
1.5
0.5
Y
2.5
1.5
0.5

0.0 0.4 0.8

X1

Figure 1: Coplot plot for Question 2.

CONTINUED
 (1) The plot shows that the data are not planar.
(2) The panels show linear relationships between the variables.
(3) A pairs plot would have been better for checking if the data are planar.
(4) A linear model tting non-parallel lines would be appropriate for these data.
(5) The coplot is inappropriate because the explanatory variables are continuous.

3. An industrial experiment was run to explore the eect of two factors on the life of a
machine tool. The factors were

(a) speed: Cutting speed (125 rpm, 150 rpm, 175 rpm),
(b) angle: Cutting angle (15 degrees, 20 degrees, 25 degrees).

15degrees 20degrees 25degrees

36

34
lifetime

32

30

28

125rpm 150rpm 175rpm 125rpm 150rpm 175rpm 125rpm 150rpm 175rpm

Figure 2: Trellis plot for Question 3.

The trellis plot in Figure 2 displays data from the experiment. Each treatment com-
bination was observed twice. Which of the following statements is TRUE?

(1) Slow speeds result in longer lifetimes.

(2) The shortest lifetime is at 20 degrees.
(3) There is no evidence of a angle eect.
(4) The lifetimes for 15 degree angles are longer than for 20 degree angles.
(5) There is evidence of interaction in these data.

CONTINUED
4. Figure 3 shows some diagnostic plots obtained after tting a regression. What, if
anything, is wrong with the regression?

Residuals vs Fitted Normal QQ plot

Standardized residuals
10 10

0 1 2
1
Residuals

0
1

2
19 19
3

4
91
91

0.2 0.4 0.6 0.8 1.0 1.2 2 1 0 1 2

Fitted values Theoretical Quantiles

ScaleLocation plot Cooks distance plot

0.0 0.5 1.0 1.5 2.0
Standardized residuals

91
91
19 Cooks distance

0.20
10
19
0.10

10
0.00

0.2 0.4 0.6 0.8 1.0 1.2 0 20 40 60 80 100

Fitted values Obs. number

Figure 3: Diagnostic plots for Question 4.

(1) There are outliers in the data.

(2) The errors are not normally distributed.
(3) The plots do not indicate problems with the regression.
(4) The points are not scattered about a plane.
(5) The error variances are not constant.

CONTINUED
5. For the data in Question 4, what remedial action should we take?

(1) No action is necessary as there are no problems with the regression.

(2) We should use logistic regression since the data are not normal.
(3) We should transform the response.
(4) We should transform the independent variables.
(5) We should delete the outliers.

6. Which of the following is FALSE?

(1) The hat matrix diagonals always lie between 0 and 1.

(2) The hat matrix diagonal measures how much leverage a point has.
(3) The hat matrix diagonal measures how outlying a point is as far as the explanatory
variables are concerned.
(4) The hat matrix diagonal measures the extent to which a data point is an outlier.
(5) The hat matrix diagonal measures the potential inuence of the point on its tted
value.

7. The car data set was discussed in a tutorial. In this data set, various characteristics
of several dierent makes of car were measured, along with the price of the cars. We
want to choose a model for these data to use as an equation for predicting the price of
other types of car. The following output was obtained:

> cars.lm<-lm(log(PRICE) ~ WEIGHT + CITY + DISP + COMP + HP +

TORQ+ TRANS + CYL, data=cars.df)
> summary(cars.lm)
Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 7.732e+00 4.204e-01 18.391 < 2e-16 ***
WEIGHT 4.193e-04 7.225e-05 5.804 4.75e-08 ***
CITY -2.126e-03 7.017e-03 -0.303 0.76238
DISP -4.435e-04 5.724e-05 -7.748 2.42e-12 ***
COMP 1.425e-02 3.317e-02 0.430 0.66815
HP 6.314e-03 9.317e-04 6.777 3.95e-10 ***
TORQ 3.444e-03 1.383e-03 2.490 0.01403 *
TRANS -9.478e-02 4.859e-02 -1.950 0.05329 .
CYL 8.123e-02 2.533e-02 3.207 0.00169 **
---

CONTINUED
 > all.poss.regs(cars.lm)
rssp sigma2 adjRsq Cp AIC BIC WEIGHT CITY DISP COMP HP TORQ TRANS CYL
1 10.615 0.078 0.843 153.481 289.481 292.408 0 0 0 0 1 0 0 0
1 14.956 0.110 0.778 271.052 407.052 409.979 0 0 0 0 0 1 0 0
1 21.903 0.161 0.675 459.190 595.190 598.117 0 0 0 0 0 0 0 1
2 8.618 0.064 0.871 101.391 237.391 243.246 1 0 0 0 1 0 0 0
2 9.771 0.072 0.854 132.612 268.612 274.467 0 0 1 0 1 0 0 0
2 9.948 0.074 0.851 137.417 273.417 279.272 0 0 1 0 0 1 0 0
3 5.496 0.041 0.917 18.844 154.844 163.625 1 0 1 0 1 0 0 0
3 7.618 0.057 0.885 76.314 212.314 221.096 0 0 1 0 0 1 0 1
3 7.697 0.057 0.884 78.462 214.462 223.244 1 0 0 0 1 1 0 0
4 5.193 0.039 0.921 12.647 148.647 160.356 1 0 1 0 1 0 0 1
4 5.291 0.040 0.920 15.301 151.301 163.010 1 0 1 0 1 0 1 0
4 5.348 0.040 0.919 16.845 152.845 164.554 1 0 1 0 1 1 0 0
5 4.914 0.037 0.925 7.094 143.094 157.730 1 0 1 0 1 1 0 1
5 5.000 0.038 0.924 9.414 145.414 160.051 1 0 1 0 1 0 1 1
5 5.179 0.039 0.921 14.270 150.270 164.906 1 0 1 1 1 0 0 1
6 4.773 0.036 0.926 5.266 141.266 158.829 1 0 1 0 1 1 1 1
6 4.908 0.037 0.924 8.919 144.919 162.483 1 1 1 0 1 1 0 1
6 4.911 0.037 0.924 8.992 144.992 162.556 1 0 1 1 1 1 0 1
7 4.767 0.037 0.926 7.092 143.092 163.583 1 0 1 1 1 1 1 1
7 4.770 0.037 0.926 7.185 143.185 163.675 1 1 1 0 1 1 1 1
7 4.904 0.038 0.924 10.804 146.804 167.295 1 1 1 1 1 1 0 1
8 4.763 0.037 0.926 9.000 145.000 168.418 1 1 1 1 1 1 1 1

Which model is most strongly indicated by this output?

(1) WEIGHT + DISP + HP

(2) CITY + COMP
(3) WEIGHT + DISP + HP + TORQ + TRANS + CYL
(4) WEIGHT + DISP + HP + TORQ + CYL
(5) WEIGHT + CITY + DISP + COMP + HP +TORQ + TRANS + CYL

8. In the education data studied in class, data were given on each of the 50 states of the
United States. We tted a model relating the reciprocal of spending on education per
capita (educ) to the explanatory variables per capita income (percap) and number of
residents under 18 per 1000 residents (under18). Use the output below to decide if
some states are having an undue inuence on the analysis.

> educ.lm<-lm(I(1/educ)~percap+under18, data=educ.df)

> influence.measures(educ.lm)
Influence measures of
lm(formula = I(1/educ) ~ percap + under18, data = educ.df) :

dfb.1_ dfb.prcp dfb.un18 dffit cov.r cook.d hat

6 -0.016411 0.13616 -0.034595 0.17673 1.158 1.06e-02 0.0960
10 -0.076648 0.20169 0.026691 0.43607 0.707 5.60e-02 0.0257
29 -0.069398 0.02278 0.071737 -0.07930 1.195 2.14e-03 0.1103
44 0.218843 0.02846 -0.281892 -0.32528 1.241 3.56e-02 0.1690
50 0.407739 -0.25851 -0.384531 -0.42573 1.587 6.13e-02 0.3429

CONTINUED
 Which of the following is TRUE?

(1) Point 50 is potentially inuential since it has a large hat matrix diagonal.
(2) Point 10 is having no eect on the estimation of the standard errors.
(3) Point 6 is inuential.
(4) Point 29 is having a big eect on the estimation of the coecient for under18.
(5) Point 44 is inuential because it has a large Cooks distance.

HINT: Recall that points are inuential if:

(a) Cooks D is more than F3,47 (0.1) = 0.1940,

(b) |DF BET AS| > 3/ n,

(c) |DF F IT | > 3/ p/(n p)
(d) |COV RAT IO 1| > 3p/n,
(e) h > 3p/n.

9. After tting a regression, which has response Y and explanatory variables X1 and X2,
we want to predict the value of Y for the values X1 = 16, X2 = 7.5. We get the the
following R output:

> q11.lm<-lm(y~X1 + X2)

> predict(q11.lm, data.frame(X1=16,X2=7.5), se=T)
$fit
[1] 34.15078
$se.fit
[1] 0.3367116
$df
[1] 97
$residual.scale
[1] 2.128866
> qt(0.975,97)
[1] 1.984723

Which of the following is TRUE?

(1) A 95% condence interval for the mean of Y is (29.926, 38.376).

(2) A 95% prediction interval for Y is (29.873, 38.429).
(3) A 95% condence interval for the mean of Y is (29.873, 38.429).
(4) The estimate of is 1.459.
(5) A 95% prediction interval for Y is (33.483, 34.819).

CONTINUED
10. In a chemical process, the yield of the process is thought to depend on both the
temperature and pressure at which the reaction takes place. An experiment was set
up to study the relationship between these variables, here called yield, temp and
pressure. There were 3 levels of temperature: Low, Medium and High, and three
levels of pressure 200 psi, 215 psi and 230 psi. (Note in the output below, Low is
the baseline level for temperature.) Each of the possible 9 treatment combinations was
used twice in the experiment, generating 18 observations.
The model yield temp*pressure was tted to the data, and a fragment of the
resulting output is shown below.

Coefficients:
Estimate
(Intercept) 90.300
tempMedium -0.100
tempHigh 0.300
pressure215.psi 0.350
pressure230psi 0.000
tempMedium:pressure215.psi 0.000
tempHigh:pressure215.psi -0.100
tempMedium:pressure230psi -0.200
tempHigh:pressure230psi -0.350

Which of the following is TRUE?

(1) The mean yield at pressure 230 psi is estimated as 0.000.

(2) The mean yield at medium temperature is estimated as 90.200.
(3) The mean yield at high temperature and pressure 200 psi is estimated as 0.300.
(4) The mean yield at medium temperature and pressure 215 psi is estimated as
90.550.
(5) The mean yield at high temperature and pressure 230 psi is estimated as 90.100.

11. In the experiment described in Question 10, the following ANOVA table was obtained:

Response: yield
Df Sum Sq Mean Sq F value Pr(>F)
temp 2 0.30111 0.15056 8.4687 0.0085392 **
pressure 2 0.76778 0.38389 21.5937 0.0003673 ***
temp:pressure 4 0.06889 0.01722 0.9687 0.4700058
Residuals 9 0.16000 0.01778

CONTINUED
 Which of the following is FALSE?

(1) There is strong evidence of interaction in these data.

(2) There is no eect of changing pressure from 200 psi to 230 psi.
(3) The eect of changing pressure from 215 psi to 230 psi is to decrease the yield by
0.350.
(4) The eect of changing temperature from medium to low is to increase the yield
by 0.100.
(5) The eect of changing temperature from medium to high is to increase the yield
by 0.400.

12. Suppose we want to t a linear model to data consisting of a continuous response

variable Y and covariates X, Z, A and B , where X and Z are continuous and A and
B are factors.
Which of the following statements is TRUE?

(1) The model Y A*X + A*Z ts separate planes though the data for each level of
A.
(2) The model Y A*B*X*Z ts separate planes though the data with the slopes and
intercepts depending on the levels of A and B.
(3) The model Y A*X + Z ts separate planes though the data for each level of A,
with the coecient of X being the same for each plane.
(4) The model Y A + X + Z ts a single plane though the data.
(5) The model Y A*B + X + Z ts separate planes though the data with the slopes
depending on the levels of A and B.

13. In a Poisson regression of a count response Y on a continuous explanatory variable X,

using the R code glm(YX, family=poisson), the regression coecient is 0.3. Which
of the following is TRUE?

(1) The eect of a unit increase in X is to increase our estimate of the mean count
by about 30%.
(2) The eect of a unit increase in X is to increase our estimate of the log-odds by
approximately 1.35.
(3) The eect of a unit increase in X is to increase our estimate of the log-odds by
about 35%.
(4) The eect of a unit increase in X is to increase our estimate of the the mean
count by an amount 0.3.
(5) The eect of a unit increase in X is to increase our estimate of the mean count
by about 35%.

CONTINUED
14. An experiment has been done to test the eect of an insecticide on a certain species of
beetle. For each of eight pre-specied dosages, a number (n) of beetles were exposed
to the insecticide and the number dying (r) was recorded. The following data were
obtained:

logdose r n
1.6907 5 59
1.7242 11 60
1.7552 31 62
1.7842 37 56
1.8113 48 63
1.8369 55 59
1.8610 61 62
1.8839 59 60

A logistic model of the form

log = + logdose
1
(where is the probability a beetle will die) was tted to the data with the following
results:

Call: glm(formula = cbind(r, n - r) ~ logdose, family = binomial)

Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -57.375 5.018 -11.43 <2e-16 ***
logdose 32.588 2.834 11.50 <2e-16 ***
---
Null deviance: 253.1354 on 7 degrees of freedom
Residual deviance: 4.3618 on 6 degrees of freedom
AIC: 36.985
> 1-pchisq(4.3618,6)
[1] 0.6278396

Which of the following is FALSE?

(1) The output indicates that death is more likely with increasing dose.
(2) The model predicts that the odds of death for a beetle subjected to a logdose of
1.6907 is 0.1024.
(3) The null deviance indicates that the t of the logistic model is poor.
(4) The model predicts that the probability of death for a beetle subjected to a logdose
of 1.6907 is 0.0929.
(5) The data are grouped data.

CONTINUED
15. In the beetle example in Question 14, the following additional output was obtained:

> predict(q18.glm, data.frame(logdose=1.8113), se=T)

$fit
[1] 1.65089
$se.fit
[1] 0.1747074
> qnorm(0.975)
[1] 1.959964

Which of the following is FALSE?

(1) The point estimate for the probability of death at log dose 1.8113 is 0.839.
(2) From the data given, it is not possible to calculate the condence interval for the
probability of death when the log dose is 1.8113.
(3) A 95% condence interval for the odds of death at log dose 1.8113 is (3.701,
7.340).
(4) The point estimate for the log-odds of death at log dose 1.8113 is 1.651.
(5) A 95% condence intervlal for the log-odds of death at log dose 1.8113 is (1.308,
1.993).

16. The beetle experiment in Questions 14 and 15 was repeated, using a modied form of
the insecticide (treatment 1) as well as the previous form (treatment2), but keeping
the same dosages. We now have data

logdose r n treat
1.6907 7 59 1
1.7242 17 60 1
1.7552 21 62 1
1.7842 41 56 1
1.8113 52 63 1
1.8369 57 59 1
1.8610 61 62 1
1.8839 59 60 1
1.6907 9 64 2
1.7242 21 59 2
1.7552 41 63 2
1.7842 48 64 2
1.8113 55 59 2
1.8369 54 58 2
1.8610 57 57 2
1.8839 64 64 2

We obtained the following anova.

CONTINUED
 Analysis of Deviance Table
Model: binomial, link: logit
Response: cbind(r, n - r)
Terms added sequentially (first to last)
Df Deviance Resid. Df Resid. Dev P(>|Chi|)
NULL 15 511.23
logdose 1 487.62 14 23.61 4.694e-108
treat 1 9.39 13 14.22 2.184e-03
logdose:treat 1 3.073e-03 12 14.21 0.96
> 1-pchisq(14.21,12)
[1] 0.2874996

Which of the following is TRUE?

(1) All of the main eects and interactions should be retained in the model.
(2) If we include treat as a variable, we dont need logdose.
(3) The test statistic 3.073e-03 is comparing the full model to the null model.
(4) On the log-odds scale, the eect of changing from treatment 1 to treatment 2 is
dierent at dierent dosages.
(5) The model cbind(r, n - r) logdose + treat seems to t well.

17. In a logistic regression for ungrouped data, with 3 explanatory variables, the residual
deviance was 134.986 on 96 degrees of freedom, while the null deviance was 137.989
on 99 degrees of freedom. The following additional output was obtained:

> 1-pchisq(137.989,99)
[1] 0.005884545
> 1-pchisq(134.986,96)
[1] 0.005384848
> 1-pchisq(3.003,3)
[1] 0.3911629

Which of the following is most strongly indicated?

(1) The residual deviance indicates that the model doesnt t well.
(2) The small residual deviance p-value indicates that the model ts well.
(3) Some of the explanatory variables should be retained in the model.
(4) None of the explanatory variables should be retained in the model.
(5) All of the explanatory variables should be retained in the model.

CONTINUED
18. In the Florida murder data discussed in class, defendants convicted of murder were
classied according to three factors; namely race of the defendant (defendant), race
of the victim (victim), and whether or not the death penalty was imposed (dp). The
following anova was obtained:

> murder.glm<-glm(count~defendant*dp*victim,family=poisson,
data=murder.df)
> anova(murder.glm, test="Chisq")
Analysis of Deviance Table.
Model: poisson, link: log, Response: count
Terms added sequentially (first to last)
Df Deviance Resid. Df Resid. Dev P(>|Chi|)
NULL 7 774.73
defendant 1 0.22 6 774.51 0.64
dp 1 443.51 5 331.00 1.861e-98
victim 1 64.10 4 266.90 1.183e-15
defendant:dp 1 0.43 3 266.47 0.51
defendant:victim 1 254.15 2 12.32 3.230e-57
dp:victim 1 12.30 1 0.02 4.535e-04
defendant:dp:victim 1 0.02 0 3.997e-15 0.89
> summary(murder.glm)
Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) 2.5649 0.2774 9.248 < 2e-16 ***
defendantw -2.5649 1.0377 -2.472 0.01345 *
dpno 2.7081 0.2864 9.454 < 2e-16 ***
victimw 0.5705 0.3470 1.644 0.10012
defendantw:dpno 0.2364 1.0652 0.222 0.82438
defendantw:victimw 3.0930 1.0705 2.889 0.00386 **
dpno:victimw -1.1896 0.3675 -3.237 0.00121 **
defendantw:dpno:victimw 0.1613 1.1032 0.146 0.88375
---
Null deviance: 7.7473e+02 on 7 degrees of freedom
Residual deviance: 3.9968e-15 on 0 degrees of freedom

Which of the following independence models is indicated by this output?

(1) Victims race and death penalty are conditionally independent, given defendants
race.
(2) Defendants race and death penalty are conditionally independent, given victims
race.
(3) The association between defendants race and death penalty is the same for black
and white victims.
(4) All three factors are independent.
(5) Defendants race and death penalty are independent of victims race.

CONTINUED
19. In a three-dimensional contingency table with factors A, B and C, we want to test the
hypothesis that factor A is independent of factors B and C, using an R statement of
the form, anova(model1, model2). What should the formulas dening model 1 and
model 2 be? The R vector count contains the cell counts.

(1) Model 1: count A + B + C, Model 2: count A*B*C.

(2) Model 1: count A*B + A*C, Model 2: count A + B*C.
(3) Model 1: count 1, Model 2: count A*B*C.
(4) Model 1: count A + B*C, Model 2: count A*B*C.
(5) Model 1: count A + B + C, Model 2: count A + B*C.
20. The data below are from a famous data set, which records the number of deaths per
year from horse kicks in corps of the Prussian army. The data have been grouped for
your convenience:
Number of deaths 0 1 2 3 4
Counts 109 65 22 3 1
We want to check if these data follow a Poisson distribution. We get the following R
output:
> ndeaths<-c(0,1,2,3,4)
> counts<-c(109,65,22,3,1)
> mean.deaths<-sum(ndeaths*counts)/sum(counts)
> poiss.probs<-dpois(0:4,mean.deaths)
> logL1<-sum(counts *log(pois.probs))
> logL1
[1] -206.1067
> freqs<-counts/sum(counts)
> logL2<-sum(counts *log(freqs))
> logL2
[1] -205.6726
> logL3<-sum(counts *log(1/5))
> logL3
[1] -321.8876
> 1-pchisq(0.8682,3)
[1] 0.8330943
Which of the following is FALSE?
(1) The residual deviance for the Poisson model is 0.8682.
(2) The test statistic for testing the hypothesis that the true distribution is Poisson
is 232.43 - 0.87 = 231.56.
(3) The Poisson model ts the data well.
(4) The test statistic for testing the hypothesis that the true distribution is uniform
(i.e. the probabilities of 0,1,2,3,4 deaths are all the same) is 232.43.
(5) The null deviance for the Poisson model is 232.43.

CONTINUED
 SECTION B

1. The data for this question are taken from an experiment which investigated the ascorbic
acid content of cabbages. Three variables are thought to inuence the ascorbic acid
content:

The size of the cabbage head (measured by the head weight, variable HeadWt)
the genetic line or cultivar (there were two lines, 39 and 52, recorded as the factor
Line)
the planting date (three dates were considered, labeled as 16, 20 and 21, recorded
as the factor Date).

Each row in the data set refers to an individual cabbage, and there are 10 observations
for each of the six factor level combinations, for a total of 60 observations. The response
variable is Ascorbic, the acsorbic acid content of the cabbage head.
Given : Date

21

20

16

1.0 2.0 3.0 4.0 1.0 2.0 3.0 4.0

80
70

52
60

Given : Line
50
Ascorbic

40
80
70

39
60
50
40

1.0 2.0 3.0 4.0

HeadWt

Figure 4: Coplot for Question B1.

(a) In Figure 4, we show a coplot of the data. Based on this coplot only, I would try
a model of the form Ascorbic HeadWt + Date*Line. Do you agree with this
choice? What aspects of the plot are you basing your decision on?

CONTINUED
(b) A model was tted, producing the output below.
> cabbage.lm<-lm(Ascorbic~HeadWt*Date*Line, data=cabbage.df)
> anova(cabbage.lm)
Analysis of Variance Table Response: Ascorbic
Df Sum Sq Mean Sq F value Pr(>F)
HeadWt 1 2630.53 2630.53 68.3538 8.704e-11 ***
Date 2 101.17 50.58 1.3144 0.2781
Line 1 1303.36 1303.36 33.8675 4.724e-07 ***
HeadWt:Date 2 116.50 58.25 1.5136 0.2304
HeadWt:Line 1 15.95 15.95 0.4145 0.5227
Date:Line 2 1.32 0.66 0.0171 0.9830
HeadWt:Date:Line 2 24.78 12.39 0.3220 0.7263
Residuals 48 1847.24 38.48
Does this model conrm my choice in part (a)? Give reasons, referring to specic
parts of the output.
(c) The diagnostic plots in Figure 5 on the next page refer to the model tted in part
(b). Point 30 seems to have a large Cooks distance. The following diagnostics
on this point were also obtained.
dfb.1_ dfb.HdWt dfb.Dt20 dfb.Dt21 dfb.Ln52
-1.937071e-15 1.151518e-15 3.205382e+00 7.941197e-16 1.478540e-15
dfb.HdW:D20 dfb.HdW:D21 dfb.HW:L52 dfb.D20:L52 dfb.D21:L52
-3.388566e+00 -9.064311e-16 -1.576766e-15 -2.729959e+00 -6.087945e-16
dfb.HW:D20:L52 dfb.HW:D21:L52 dffit cov.r cook.d
2.788916e+00 4 .186369e-16 -3.773335e+00 2.680588e+00 1.126440e+00
hat
8.000000e-01

Can you explain why the point is so inuential? Can you identify the point on
the coplot (Figure 4)?
NOTE: A copy of the coplot in Figure 4 is attached to the back of the examination
paper. Detach the plot and mark the point you think is point 30 on the plot. Hand
in the plot along with your answer.

CONTINUED
 Residuals vs Fitted Normal QQ plot

3
Standardized residuals
21 21

54 54

2
29

10
29

Residuals

1
5
0

0
1
10

2
45 50 55 60 65 70 2 1 0 1 2

ScaleLocation plot Cooks distance plot

1.2
21
Standardized residuals

30
1.5

54
29

Cooks distance

0.8
1.0

0.4
0.5

16 58
0.0

0.0
45 50 55 60 65 70 0 10 20 30 40 50 60

Figure 5: Diagnostic plots for Question B1(c).

(d) The model Ascorbic HeadWt+Date*Line was tted to the same data(i.e. no
points were deleted), The diagnostic plots are now satisfactory. Can you suggest
why?
(e) The following output from the t in part (d) was obtained:
> cabbage2.lm<-lm(Ascorbic~HeadWt+Date*Line, data=cabbage.df)
> summary(cabbage2.lm)
Call:
lm(formula = Ascorbic ~ HeadWt + Date * Line, data = cabbage.df)
Coefficients:
Estimate Std. Error t value Pr(>|t|)
(Intercept) 64.618 4.305 15.011 < 2e-16 ***
HeadWt -4.503 1.210 -3.721 0.000481 ***
Date20 -2.611 2.768 -0.943 0.349899
Date21 2.519 2.781 0.906 0.369254
Line52 8.058 2.948 2.733 0.008510 **
Date20:Line52 2.838 4.138 0.686 0.495742
Date21:Line52 3.224 3.884 0.830 0.410209
---
Residual standard error: 6.105 on 53 degrees of freedom
Multiple R-Squared: 0.6731, Adjusted R-squared: 0.636
F-statistic: 18.18 on 6 and 53 DF, p-value: 2.502e-11
Are the two factors having an eect on the ascorbic acid content? If so, how?
Does the eect depend on the head size?

CONTINUED
2. Framingham is an industrial town located approximately 30 km from Boston. In 1948
a study was begun with the aim of identifying factors that are related to the occurrence
of coronary heart disease (CHD). At the start of the study, a large proportion of the
towns inhabitants were examined for the presence of CHD. Measurements were made
on a number of potential risk factors. The individuals who were found to be free of
CHD at that time were followed up for twelve years and those who developed CHD
during that period were identied. The following dataset was extracted from that data
and relates the proportions developing CHD to the initial serum cholesterol level (mg
per 100 ml) of these individuals cross classied by age and sex.

> frame.df
sex age serum num tot
1 M 30-49 A 13 340
2 M 30-49 B 18 408
3 M 30-49 C 40 421
4 M 30-49 D 57 362
5 M 50-62 A 13 123
6 M 50-62 B 33 176
7 M 50-62 C 35 174
8 M 50-62 D 49 183
9 F 30-49 A 6 542
10 F 30-49 B 5 552
11 F 30-49 C 10 412
12 F 30-49 D 18 357
13 F 50-62 A 9 58
14 F 50-62 B 12 135
15 F 50-62 C 21 218
16 F 50-62 D 48 395

The variables num and tot give the number having heart disease and the total number
in the study for each of the 16 risk factor combinations.

(a) Describe the model or models you would initially try to t to these data, with a
view to identifying the main risk factors for CHD. Describe briey how you would
check if your model ts well, and how you might nd a simpler model.

CONTINUED
(b) The the plots in Figure 6 refer to tting a particular model. Can you identify
any risk factor combinations for which the model ts poorly? Can you oer any
explanation for this?
Residuals vs Fitted Normal QQ plot
13 13

10
1
8

10
0

16
1

30
5
5

4.5 3.5 2.5 1.5 2 1 0 1 2

ScaleLocation plot Cooks distance plot

4
5
Std. deviance resid.

5
5

16

3
13
4
3

2
2

1
1

13 16
0

4.5 3.5 2.5 1.5 5 10 15

Figure 6: Diagnostic plots for Question B2(b).

CONTINUED
(c) After some remedial action, The following output was obtained:
> frame.glm<-glm(cbind(num,tot-num)~ sex * age + age * serum,
family=binomial, data=frame.df)
> summary(frame.glm)
Coefficients:
Estimate Std. Error z value Pr(>|z|)
(Intercept) -4.55797 0.26742 -17.044 < 2e-16 ***
sexM 1.36270 0.18801 7.248 4.23e-13 ***
age50-62 2.23894 0.35980 6.223 4.89e-10 ***
serumB 0.05638 0.31516 0.179 0.858023
serumC 0.92448 0.27611 3.348 0.000813 ***
serumD 1.54452 0.26427 5.844 5.08e-09 ***
sexM:age50-62 -0.49769 0.24189 -2.057 0.039641 *
age50-62:serumB -0.06752 0.42424 -0.159 0.873549
age50-62:serumC -0.84739 0.38824 -2.183 0.029065 *
age50-62:serumD -1.15523 0.36796 -3.140 0.001692 **
---
Null deviance: 301.2147 on 15 degrees of freedom
Residual deviance: 1.5435 on 6 degrees of freedom
AIC: 96.225
> round(100*predict(frame.glm, type="response"),1)
[1] 3.9 4.2 9.4 16.1 18.9 18.8 20.2 25.6 1.0 1.1
[11] 2.6 4.7 9.0 8.9 9.6 12.7
Discuss the eect of the various risk factors on the probability of heart disease.
Which is the most important risk factor?

CONTINUED
3. (a) In a 22 table with factors A and B, each at two levels, we can measure association
between the factors using the odds ratio. Give a denition of the odds ratio and
discuss its connection with the concept of independence.
(b) In a 2 2 2 table, with factors A, B and C, each at two levels, we can consider
the odds ratios in the separate A B tables corresponding to the two levels of
C. Under what circumstances will these be the same? What is the name given
to the situation when they are dierent? What implications does this have for an
analysis?
(c) The table below classies admission applications to the Berkeley graduate school,
classied according to (a) the success or failure of the application, (b) the gender
of the applicant, and (c) the department.
Gender
Dept Admit Male Female
A Yes 353 17
No 207 8
B Yes 120 202
No 205 391
C Yes 138 131
No 279 244
D Yes 53 94
No 138 299
E Yes 22 24
No 351 317

Below is a data frame obtained by collapsing the above table over departments.
Ignoring departments, and using the output below, are success and gender inde-
pendent? Give a reason for your answer.
> berk.df
Count Gender Admit
1 1278 F No
2 1493 M No
3 557 F Yes
4 1198 M Yes
> berk.glm<-glm(Count~Gender*Admit, family=poisson, data=berk.df)
> anova(berk.glm, test="Chisq")
Analysis of Deviance Table
Model: poisson, link: log
Response: Count
Terms added sequentially (first to last)
Df Deviance Resid. Df Resid. Dev P(>|Chi|)
NULL 3 486.35
Gender 1 162.87 2 323.48 2.665e-37
Admit 1 230.03 1 93.45 5.879e-52
Gender:Admit 1 93.45 0 3.393e-13 4.167e-22

CONTINUED
(d) Some further output is shown below. Does this modify the conclusion you reached
in part(c)?
NOTE: The data frame berkeley.df contains the data in the table shown on the
previous page, and contains variables Count, Admit, Gender and Dept. The rst
few lines of this data frame are shown in the output.
> berkeley.df
Count Admit Gender Dept
1 353 Yes M A
2 207 No M A
3 17 Yes F A
4 8 No F A
5 120 Yes M B
...............
20 lines in all

> berkeley.glm<-glm(Count~Dept*Gender*Admit, family=poisson,

data=berkeley.df)
> anova(berkeley.glm, test="Chisq")
Analysis of Deviance Table
Model: poisson, link: log
Response: Count
Terms added sequentially (first to last)
Df Deviance Resid. Df Resid. Dev P(>|Chi|)
NULL 19 1879.78
Dept 4 111.60 15 1768.18 3.320e-23
Gender 1 5.38 14 1762.80 0.02
Admit 1 469.90 13 1292.90 3.365e-104
Dept:Gender 4 753.44 9 539.46 9.333e-162
Dept:Admit 4 536.78 5 2.68 7.427e-115
Gender:Admit 1 0.13 4 2.56 0.72
Dept:Gender:Admit 4 2.56 0 1.936e-13 0.63

ATTACHMENT FOLLOWS
 STATS 330
ATTACHMENT 23 STATS 794

COPY OF COPLOT FOR QUESTION B1(c)

Given : Date

21

20

16

1.0 2.0 3.0 4.0 1.0 2.0 3.0 4.0

80
70

52
60

Given : Line
50
Ascorbic

40
80
70

39
60
50
40

1.0 2.0 3.0 4.0

HeadWt