Statistics GIDP

Ph.D. Qualifying Exam

Methodology
January 10, 9:00am-1:00pm
Instructions: Put your ID (not name) on each sheet. Complete exactly 5 of 6 problems; turn in
only those sheets you wish to have graded. Each question, but not necessarily each part, is
equally weighted. Provide answers on the supplied pads of paper and/or use a Microsoft word
document or equivalent to report your software code and outputs. Number each problem. You
may turn in only one electronic document. Embed relevant code and output/graphics into your
word document. Write on only one side of each sheet if you use paper. You may use the
computer and/or a calculator. Stay calm and do your best. Good luck!

1. A manufacturer suspects that the batches of raw material furnished by her supplier differ
significantly in antibiotic content. There are a large number of batches currently in the
warehouse. Five of these are randomly selected for study. A biologist makes five
determinations on each batch and obtains the following data:

Batch 1 Batch 2 Batch 3 Batch 4 Batch 5

23.46 23.59 23.51 23.28 23.29
23.48 23.46 23.64 23.40 23.46
23.56 23.42 23.46 23.37 23.37
23.39 23.49 23.52 23.46 23.32
23.40 23.50 23.49 23.39 23.38

A portion of pertinent computer output follows:

Analysis of Variance for Content

Source DF SS MS F P
Batch 4 0.096976 0.024244 5.54 0.004
Error 20 0.087600 0.004380
Total 24 0.184576

(a) What type of design/experiment is this? Why?

(b) State the statistical model and corresponding assumptions.
(c) Write the hypothesis in mathematical notation for testing the equality of the batches.
(d) Are the batches equal in output? Use  = 0.05.
(e) Estimate the variability between batches.
(f) Estimate the experimental error variance.
2. A consumer products company relies on direct mail marketing pieces as a major component
of its advertising campaigns. The company has three different designs for a new brochure
and wants to evaluate their effectiveness, as there are substantial differences in costs between
the three designs. The company decides to test the three designs by mailing 5,000 samples
of each to potential customers in four different regions of the country. Note that there may
be regional differences in the customer base. The number of responses to each mailing is
shown below (find the dataset in the file “brochure.csv”).

Region
Design NE NW SE SW
1 225 350 220 375
2 400 525 390 550
3 275 340 200 320

(a) What design is this? Why?

(b) Write the statistical model and the corresponding assumptions.

(c) Analyze the dataset and draw a conclusion at α = 0.05.

(d) Can you check whether there exists an interaction between the Design factor and the Region
factor? If no, explain why not. If yes, write your statistical model, state the hypotheses,
analyze the data, and report your finding

(e) Attach all SAS/R code.

3. An engineer is interested in the effects of cutting speed (A), tool geometry (B), and cutting
angle on the life (in hours) of a machine tool. Two levels of each factor are chosen, and three
replicates are run for each combination. The results are as follows (find the dataset in the
file “life.csv”)

Treatment Replicate
A B C Combination I II III
- - - (1) 22 31 25
+ - - a 32 43 29
- + - b 35 34 50
+ + - ab 55 47 46
- - + c 44 45 38
+ - + ac 40 37 36
- + + bc 60 50 54
+ + + abc 39 41 47

(a) What design is this?

(b) Estimate the factor effects. Which effects appear to be large?

(c) Use the analysis of variance to confirm your informal conclusions from part (b).

(d) Write down a regression model for predicting tool life (in hours) based on the results of this
experiment.

(e) What levels of A, B, and C would you recommend using? Justify your answer.

(f) Attach your SAS/R code.

4. A major state university took a sample of n = 705 students and collected the outcome variable
Y = {Freshman Grade Point Average}, along with predictor variables X1 = {High school
class rank (as a percentile; higher percentiles indicate higher rank)} and X2 = {ACT score}.
It was felt that the interaction cross-product X3 = X1X2 was likely to also be an important
predictor. The data are found in the file “college.csv.”

(a) Examine the two predictor variables X1 and X2 to determine if multicollinearity is a concern
with these data.

(b) To assuage any concerns over multicollinearity, apply a ridge regression analysis to these
data. Begin with a ridge trace plot and use it to identify a plausible value for the ridge tuning
parameter c.
[Hint: if your plot covers sufficient values of c, the stock traceplot() function in the R
genridge package will draw vertical lines at plausible values, known as the modified Hoerl-
Kennard-Baldwin (HKB) and Lawless-Wang (LW) estimators. Decide if either could be a
useful choice here and if so, select it.]

(c) Using the tuning parameter, c, selected in part (b), calculate the ridge-regression predicted
values. From these find the raw residuals and construct a residual plot. Comment on any
patterns that may appear.
5. Consider the simple linear regression model: Yi ~ indep. N(0+1Xi, 2), i = 1, ..., n, where
X is a non-stochastic predictor variable. Suppose interest exists in estimating the X value(s)
at which E[Y] = 0. Let this target parameter be .

(a) What is  in terms of the original regression parameters?

(b) State the maximum likelihood estimator (MLE) of . Call this ̂.

(c) Recall from Section 5.5.4 in Casella & Berger that the Delta Method can be used to determine
the asymptotic features of a function of random variables. In particular, for a bivariate
random vector [U1 U2]T and a differentiable function g(u1,u2), where E[Uj] = j (j = 1,2),
2 ∂
E[g(U1,U2)] ≈ g(,2) + ∑j=1 ∂ g(,2) E(Ui – i)
j

and
2 ∂ 2 ∂ ∂
Var[g(U1,U2)] ≈ ∑j=1{∂ g(,2)} Var(Ui) + 2{∂ g(,2)}{∂ g(,2)}Cov(U1,U2)
j 1 2

From these results:

(i) show that E[̂] ≈ , and
(ii) find an approximation for Var[̂] under this regression setting.
6. Assume a simple linear regression-through-the-origin model: Yi ~ indep. N(1Xi, 2), i = 1,
..., n. Equation (4.14) in Kutner et al. shows that the maximum likelihood estimator (MLE)
of 1 is
n
∑i=1XiYi
b1 = n 2
∑j=1Xj
with standard error
MSE
s{b1} = n ,
∑k=1Xk2
n
where MSE = ∑i=1(Yi – b1Xi)2/(n–2) is an unbiased estimator of 2. Starting from the basic
principles of the Working-Hotelling-Scheffé (WHS) construction, derive the 1–WHS
simultaneous confidence band for the mean response E[Yi] with this model. What is the
“shape” of this band?
 Statistics GIDP
Ph.D. Qualifying Exam
Methodology
January 10, 9:00am‐1:00pm
Instructions: Put your ID (not name) on each sheet. Complete exactly 5 of 6 problems; turn in
only those sheets you wish to have graded. Each question, but not necessarily each part, is
equally weighted. Provide answers on the supplied pads of paper and/or use a Microsoft word
document or equivalent to report your software code and outputs. Number each problem. You
may turn in only one electronic document. Embed relevant code and output/graphics into your
word document. Write on only one side of each sheet if you use paper. You may use the
computer and/or a calculator. Stay calm and do your best. Good luck!

1. A manufacturer suspects that the batches of raw material furnished by her supplier differ
significantly in antibiotic content. There are a large number of batches currently in the
warehouse. Five of these are randomly selected for study. A biologist makes five
determinations on each batch and obtains the following data:

Batch 1 Batch 2 Batch 3 Batch 4 Batch 5

23.46 23.59 23.51 23.28 23.29
23.48 23.46 23.64 23.40 23.46
23.56 23.42 23.46 23.37 23.37
23.39 23.49 23.52 23.46 23.32
23.40 23.50 23.49 23.39 23.38

A portion of pertinent computer output follows:

Analysis of Variance for Content

Source DF SS MS F P
Batch 4 0.096976 0.024244 5.54 0.004
Error 20 0.087600 0.004380
Total 24 0.184576

(a) What type of design/experiment is this? Why?

This is a one-factor random (effect) design. We are interested in the entire population of
batches and batches are randomly chosen.

(b) State the statistical model and corresponding assumptions.

We assume yij = μ + τi + εij, i=1, …, 5, j=1, …, 5, with τij ~ N(0,σ2 ) independent of
εij ~ N(0,σ2 ).
(c) Write the hypothesis in mathematical notation for testing the equality of the batches.
H0: στ2 = 0
H1: στ2 > 0

(d) Are the batches equal in output? Use  = 0.05.

Since the P-value is given as P = 0.004, and this is less than 0.05, we reject the null
hypothesis and conclude that the batches differ significantly.

(e) Estimate the variability between batches.

0.024244 0.004380
0.003972
5

(f) Estimate the experimental error variance.

̂2 = 0.004380
2. A consumer products company relies on direct mail marketing pieces as a major
component of its advertising campaigns. The company has three different designs for a
new brochure and wants to evaluate their effectiveness, as there are substantial differences
in costs between the three designs. The company decides to test the three designs by
mailing 5,000 samples of each to potential customers in four different regions of the
country. Note that there may be regional differences in the customer base. The number of
responses to each mailing is shown below (find the dataset in the file “brochure.csv”).

Region
Design NE NW SE SW
1 225 350 220 375
2 400 525 390 550
3 275 340 200 320

(a) What design is this? Why?

This is a RCBD design (randomized complete block design) with Region as a blocking
factor.

(b) Write the statistical model and the corresponding assumptions.

2
yij = μ + τi + αj + εij, i=1,2,3, j=1, …, 4, and where ∑αj = 0, ∑τi = 0, and εij ~ i.i.d.N(0,σ ).

(c) Analyze the dataset and draw a conclusion at α = 0.05.

From the SAS output (below) the “design” factor and “region” block factor are both significant
at  = 0.05.
Source DF Sum of Squares Mean Square F Value Pr > F
Model 5 133137.5000 26627.5000 42.18 0.0001
Error 6 3787.5000 631.2500
Corrected Total 11 136925.0000

Source DF Type III SS Mean Square F Value Pr > F

design 2 84762.50000 42381.25000 67.14 <.0001
region 3 48375.00000 16125.00000 25.54 0.0008

The raw residual plot and QQ-plot (below) do not suggest any unusual patterns. Thus the
normality assumption is reasonable.
(d) Can you check whether there exists an interaction between the Design factor and the
Region factor? If no, explain why not. If yes, write your statistical model, state the
hypotheses, analyze the data, and report your finding
Yes. Use Tukey’s 1-degree of freedom test, with model yij = μ + τi + αj + γτiαj + εij. The
hypotheses are
H0: 0
H1: 0
The following SAS output shows that the interaction is not significant (P-value = 0.4741).
Therefore the additive model from above is appropriate.
Source DF Type III SS Mean Square F Value Pr > F
design 2 821.240880 410.620440 0.61 0.5808
region 3 1099.093874 366.364625 0.54 0.6745
q 1 404.852146 404.852146 0.60 0.4741

(e) Attach all SAS/R code.

data brochure ;
input design region $ response;
datalines;

1 NE 225
1 NW 350
1 SE 220
1 SW 375
2 NE 400
2 NW 525
2 SE 390
2 SW 550
3 NE 275
3 NW 340
3 SE 200
3 SW 320
;

proc glm data=brochure;

class design region;
model response=design region;
output out=myout p=pred r=res;
run;

proc sgplot data=myout;

scatter x=pred y=res;
run;

proc univariate data=myout normal;

var res;
qqplot res/normal(mu=est sigma=est color=red L=1);
run;

data two;
set myout;
q=pred*pred;

proc glm data=two;

class design region;
model response=design region q/ss3;
run;
3. An engineer is interested in the effects of cutting speed (A), tool geometry (B), and cutting
angle on the life (in hours) of a machine tool. Two levels of each factor are chosen, and
three replicates are run for each combination. The results are as follows (find the dataset in
the file “life.csv”)

Treatment Replicate
A B C Combination I II III
- - - (1) 22 31 25
+ - - a 32 43 29
- + - b 35 34 50
+ + - ab 55 47 46
- - + c 44 45 38
+ - + ac 40 37 36
- + + bc 60 50 54
+ + + abc 39 41 47

(a) What design is this?

3
A 2 factorial design.

(b) Estimate the factor effects. Which effects appear to be large?

It seems that the factors B and C and interaction AC are large, they are 11.33, 6.83 and -
8.83 respectively (just double the coefficients below obtained through a regression model):
Intercept A B C AB AC BC ABC
40.8333 0.16667 5.66667 3.41667 -0.83333 -4.41667 -1.41667 -1.08333

(c) Use the analysis of variance to confirm your informal conclusions from part (b).
The ANOVA table below confirms the significance of factors B, C and the interaction AC:
their P-values are 0.001, 0.0077, and 0.0012, respectively
Source DF Type III SS Mean Square F Value Pr > F
A 1 0.6666667 0.6666667 0.02 0.8837
B 1 770.6666667 770.6666667 25.55 0.0001
A*B 1 16.6666667 16.6666667 0.55 0.4681
C 1 280.1666667 280.1666667 9.29 0.0077
A*C 1 468.1666667 468.1666667 15.52 0.0012
B*C 1 48.1666667 48.1666667 1.60 0.2245
A*B*C 1 28.1666667 28.1666667 0.93 0.3483

(d) Write down a regression model for predicting tool life (in hours) based on the results of this
experiment.
Since the interaction AC is significant we need to include Factor A into the model, though
itself, it is not significant. Then, the prediction equation becomes
Life = 40.8333 + 0.16667XA + 5.66667XB + 3.41667XC – 4.41667XAC
(e) What levels of A, B, and C would you recommend using? Justify your answer.
For a maximum of the tool life, we would recommend high levels for factors B and C, and
low level for factor A. This is because
(i) all three main factors have a positive effect,
(ii) the interaction AC has a negative effect, and
(iii) the coefficient of factor C is larger than that for factor A.

(f) Attach your SAS/R code.

data Q3 ;
input A B C rep life;
datalines;
-1 -1 -1 1 22
-1 -1 -1 2 31
-1 -1 -1 3 25
1 -1 -1 1 32
1 -1 -1 2 43
1 -1 -1 3 29
-1 1 -1 1 35
-1 1 -1 2 34
-1 1 -1 3 50
1 1 -1 1 55
1 1 -1 2 47
1 1 -1 3 46
-1 -1 1 1 44
-1 -1 1 2 45
-1 -1 1 3 38
1 -1 1 1 40
1 -1 1 2 37
1 -1 1 3 36
-1 1 1 1 60
-1 1 1 2 50
-1 1 1 3 54
1 1 1 1 39
1 1 1 2 41
1 1 1 3 47
;

data inter;
set Q3;
AB=A*B; AC=A*C; BC=B*C; ABC=A*B*C;
run;

proc print data=inter;

run;

proc reg outest=effect data=inter;

model life=A B C AB AC BC ABC;
run;
/* or from the estimates in the glm proc */
proc glm data=inter;
class A B C AB AC BC ABC;
model life=A B C AB AC BC ABC;
estimate 'A' A -1 1;
estimate 'B' B -1 1;
estimate 'C' C -1 1;
estimate 'AB' AB -1 1;
estimate 'AC' AC -1 1;
estimate 'BC' BC -1 1;
estimate 'ABC' ABC -1 1;
run;

/* or through a more complicated procedure (skip) */

proc print data=effect;
run;

proc glm data=Q3;

class A B C;
model life=A|B|C;
output out=myout p=pred r=res;
run;
4. A major state university took a sample of n = 705 students and collected the outcome
variable Y = {Freshman Grade Point Average}, along with predictor variables X1 = {High
school class rank (as a percentile; higher percentiles indicate higher rank)} and X2 = {ACT
score}. It was felt that the interaction cross-product X3 = X1X2 was likely to also be an
important predictor. The data are found in the file “college.csv.”

(a) Examine the two predictor variables X1 and X2 to determine if multicollinearity is a

concern with these data.

Always plot the data! Sample R code:

college.df = read.csv( file.choose() )
attach( college.df )
X1 = class.rank; X2 = ACT; X3=X1*X2
Y = GPA
pairs( Y~X1+X2+X3, pch=19 )

We see some multicollinearity may exist between X1 & X3 and between X2 & X3 (no
surprise, in either case). A further check is available via the sample correlations among the
predictor variables:
cor( cbind(X1,X2,X3) )
 X1 X2 X3
X1 1.0000000 0.4425075 0.8883073
X2 0.4425075 1.0000000 0.7890032
X3 0.8883073 0.7890032 1.0000000
where we see large correlations with the interaction term X3.

While informative, however, all the above analysis is subjective. A final, objective,
definitive check employs the VIFs:
require( car )
vif( lm(Y ~ X1+X2+X3) )
X1 X2 X3
29.35675 16.40282 62.54291
where clearly max{VIFk} > 10. (Also, the mean VIF is 36.1008, which is above the target
bound of 6.) Multicollinearity is a clear problem with these data.

(b) To assuage any concerns over multicollinearity, apply a ridge regression analysis to these
data. Begin with a ridge trace plot and use it to identify a plausible value for the ridge
tuning parameter c.
[Hint: if your plot covers sufficient values of c, the stock traceplot() function in the R
genridge package will draw vertical lines at plausible values, known as the modified Hoerl-
Kennard-Baldwin (HKB) and Lawless-Wang (LW) estimators. Decide if either could be a
useful choice here and if so, select it.]

–
For a ridge regression fit, start by centering the response variable to Ui = Yi − Y and
standardizing the predictors to Z1, Z2, and Z3, each with zero mean and unit variance:
U = scale( Y, scale=F )
Z1 = scale( X1 )
Z2 = scale( X2 )
Z3 = scale( X3 )
Now load the genridge package, choose a range for c (here 0 < c < 10) and construct the
traceplot using the ridge() function:
require( genridge )
c = seq( from=.01,to=10,by=.01 )
traceplot( ridge(U~Z1+Z2+Z3, lambda=c) )
 The stock traceplot suggests that the Lawless-Wang (LW) estimate for c is far enough
along to where the traces start to flatten and stabilize. So choose cLW. The exact value can
be found as the $kLW attribute in the ridge() object:
print( ridge(U~Z1+Z2+Z3,lambda=c)$kLW )
[1] 3.531847

(c) Using the tuning parameter, c, selected in part (b), calculate the ridge-regression predicted
values. From these find the raw residuals and construct a residual plot. Comment on any
patterns that may appear.

The LW estimate is c = 3.531847. The final ridge regression object is then constructed via:
cLW = ridge(U~Z1+Z2+Z3,lambda=c)$kLW
college.ridge = ridge( U~Z1+Z2+Z3, lambda=cLW )
Fitted values and residuals can be found via direct calculation. First identify the ridge regr.
coefficients (use the $coef attribute in the ridge() object, but select its components
carefully!) and then proceed from there:
Zmtx = as.matrix( cbind(Z1,Z2,Z3) )
Uhat = Zmtx %*% college.ridge$coef[,1:3]
RawResid = U - Uhat
plot( RawResid ~ Uhat, pch=19 ); abline( h=0 )
The plot shows possible decreasing variability with increasing response (i.e. variance
heterogeneity), or perhaps a decreasing trend suggestive of a possible hidden/unidentified
predictor variable. Either way, deeper diagnostic study of these data is warranted.
5. Consider the simple linear regression model: Yi ~ indep. N(0+1Xi, 2), i = 1, ..., n, where
X is a non-stochastic predictor variable. Suppose interest exists in estimating the X
value(s) at which E[Y] = 0. Let this target parameter be .

(a) What is  in terms of the original regression parameters?

This is essentially an inverse regression problem. We have E[Y] = 0+1X.

Clearly, at E[Y] = 0 we have 0 = 0+1X, so solving for X produces = –0/1.

(b) State the maximum likelihood estimator (MLE) of . Call this ̂.

Under ML invariance, ̂ = –̂0/̂1, where ̂j is the ML estimator of j (j=0,1).

(c) Recall from Section 5.5.4 in Casella & Berger that the Delta Method can be used to
determine the asymptotic features of a function of random variables. In particular, for a
bivariate random vector [U1 U2]T and a differentiable function g(u1,u2), where E[Uj] = j
(j = 1,2),
2 ∂
E[g(U1,U2)] ≈ g(,2) + ∑ j=1 ∂ g(,2) E(Ui – i)
j

and
2 ∂ 2 ∂ ∂
Var[g(U1,U2)] ≈ ∑ j=1{∂ g(,2)} Var(Ui) + 2{∂ g(,2)}{∂ g(,2)}Cov(U1,U2)
j 1 2

From these results:

(i) show that E[̂] ≈ , and
(ii) find an approximation for Var[̂] under this regression setting.

Let g(1,2) = = –0/1. We know that the MLEs for j are unbiased such that E[̂j] = j
(j=0,1). Then from the Delta Method we see
1 ∂ 1 ∂
(i) E[̂] = E[–̂0/̂1] ≈ g(,2) + ∑ j=0 ∂ g(,1) E(̂i – i) = –0/1 + ∑ j=0 ∂ g(,1)(0)
i i

= –0/1 = .
(This is actually true in general: under the conditions of the Delta Method, E[̂0/̂1] ≈ 0/1.
See Example 5.5.27 in Casella & Berger.)
(ii) From the Delta Method,
2
2  Var[̂0] Var[̂] Cov[̂0;̂]
Var[̂] = Var[–̂0/̂1] = (–1) Var[̂0/̂1] ≈    2 + –2 .
1  0 1 2 0 1 
– 2 –
Now let SSx = ∑(Xi – X) . We know Var[̂0] = 2(n–1SSx + X2)/SSx, Var[̂1] = 2/SSx, and
–
Cov[̂0,̂1] = –2X/SSx. This allows for some simplification:
 – 
2 SSx – 2  X0 2 SSx – 2 – 
Var[̂] ≈ 2 + X + 2+2 = 2  n + X + 2 – 2X.
1 SSx  n 1 1  1 SSx  
6. Assume a simple linear regression-through-the-origin model: Yi ~ indep. N(1Xi, 2), i = 1,
..., n. Equation (4.14) in Kutner et al. shows that the maximum likelihood estimator (MLE)
of 1 is
n
∑i=1XiYi
b1 = n 2
∑j=1Xj
with standard error
MSE
s{b1} = n ,
∑k=1Xk2
n
where MSE = ∑i=1(Yi – b1Xi)2/(n–2) is an unbiased estimator of 2. Starting from the basic
principles of the Working-Hotelling-Scheffé (WHS) construction, derive the 1–WHS
simultaneous confidence band for the mean response E[Yi] with this model. What is the
“shape” of this band?

The WHS band will take the form Ŷh ± W s{Ŷh}, where the WHS critical point is based on
W2 = pF(1–; p, n–p). Here p=1, while Ŷh = b1Xh with standard error

s{Ŷh} = Var{Ŷh} = Xh2Var{b1} = |Xh| s{b1}

for all Xh. Thus the band will be b1Xh ± W |Xh| s{b1}, where W2 = F(1–;1,n–1).
But recall that F(1–;1,n–1) = t2(1–{½};n–1), so this simplifies to

b1Xh ± t(1–{½};n–1) s{b1} |Xh|,

for all Xh. Note the form of this “band”: a pair of straight lines that intersect at the origin.