Statistics 106

Solutions for Homework 6

Due : Nov. 30, 2018, In Class

19.31
Refer to Eye contact effect Problems 19.12 and 19.13.

a. Estimate µ21 with a 99 percent confidence interval. Interpret your interval estimate.

b. Estimate µ1. with a 99 percent confidence interval. Interpret your interval estimate.

c. Prepare a bar graph of the estimated factor B level means. What does this plot suggest about the
factor B main effects?

d. Obtain confidence intervals for µ.1 and µ.2 , each with a 99 percent confidence coefficient. Interpret
your interval estimates.

e. Prepare a bar graph of the estimated factor A level means. What does this plot suggest about the
factor A main effects?

f. Obtain confidence intervals for D1 = µ2. − µ1. and D2 = µ.2 − µ.1 ; use the Bonferroni procedure and
a 95 percent family confidence coefficient. Summarize your findings. Are your findings consistent
with those in parts (c) and (e)?

g. Is the Bonferroni procedure used in part (f) the most efficient one that could be used here? Explain.
(Hint: Scheffe’s procedure is also applicable for those comparisons in part (f), so calculate S and
compare it with B)

h. How many pairwise comparisons are there between treatment means? Construct Tukey’s confidence
intervals for these pairwise comparisons at family-wise confidence coefficient 95%. Based on these
intervals, which pairs of treatment means are significantly different from each other at level 0.05?

Solution:

√
a. Ȳ21. = 13, s(Ȳ21. ) = MS E/n = 1.1023, t(1 − α/2; (n − 1)ab) = t(.995; 16) = 2.921.
Thus, the 99% CI for µ21 is 13 ± 2.921(1.1023) = [9.780, 16.220].
√
b. Ȳ1.. = 11.4, s(Ȳ1.. ) = MS E/(bn) = .7794, t(1 − α/2; (n − 1)ab) = t(.995; 16) = 2.921.
Thus, the 99% CI for µ1. is 11.4 ± 2.921(.7794) = [9.123, 13.677].

c. The bar graph of the estimated factor B level means is shown below:

1
 The bar graph shows that Ȳ.1. and Ȳ.2. are different, thus µ.1 and µ.2 are very likely to be different,
which means the factor B main effects probably exist.
√
d. Ȳ.1. = 11.1, Ȳ.2. = 15, s(Ȳ.1. ) = s(Ȳ.2. ) = MS E/(an) = .7794, t(1 − α/2; (n − 1)ab) = t(.995; 16) =
2.921.
Thus, the 99% CI for µ.1 and µ.2 are 11.1 ± 2.921(.7794) = [8.823, 13.377] and 15 ± 2.921(.7794) =
[12.723, 17.277].

e. The bar graph of the estimated factor A level means is shown below:

2
 The bar graph shows that Ȳ1.. and Ȳ2.. are different, thus µ1. and µ2. are very likely to be different,
which means the factor A main effects probably exist.
√
D̂1 = Ȳ2.. − Ȳ1.. = 3.3, D̂2 = Ȳ.2. − Ȳ.1. = 3.9, s(D̂1 ) = 2MS E/(bn) = 1.1023, s(D̂2 ) =
f. √
2MS E/(an) = 1.1023, B = t(1 − α/(2g); (n − 1)ab) = t(.9875; 16) = 2.473.
Thus, the 95% family-wise CI for D1 and D2 using Bonferroni procedure are 3.3 ± 2.473(1.1023) =
[.574, 6.026] and 3.9 ± 2.473(1.1023) = [1.174, 6.626].
We find that at family confidence coefficient 95%, µ1. and µ2. are different, and µ.1 and µ.2 are differ-
ent, which are consistent with the findings in parts (c) and (e): that is to say, personal officer’s gender
and eye contact of applicant contribute a real effect to the forecasted success of an applicant.
√
g. Scheffe’s
√ procedure is also applicable for those comparisons in part (f), and S = (ab − 1)F(1 − α; ab − 1, (n − 1)ab) =
3F(.95; 3, 16) = 3.117.
Since B < S , the Bonferroni procedure is the most efficient one that could be used here.
h. Since there are ab = 4 treatment means, there are a total of 4×32 = 6 pairwise comparisons between
treatment means.
Let D1 = µ11 − µ12 , D2 = µ11 − µ21 , D3 = µ11 − µ22 , D4 = µ12 − µ21 , D5 = µ12 − µ22 , D6 = µ21 − µ22 .
The unbiased estimators D̂1 = −4.4, D̂2 = −3.8, D̂3 = −7.2, D̂4 = 0.6, D̂5 = −2.8, D̂6 = −3.4.

3
 √
The standard errors s(D̂1 ) = s(D̂2 ) = s(D̂3 ) = s(D̂4 ) = s(D̂5 ) = s(D̂6 ) = 2MS E/n = 1.559.
T = √1 q(1 − α; ab, (n − 1)ab) = √1 q(.95; 4, 16) = 2.861.
2 2
Thus, the 95% family-wise CI for D1 to D6 are: [-8.86, 0.06], [-8.26, 0.66], [-11.66, -2.74], [-3.86,
5.06], [-7.26, 1.66], [-7.86, 1.06].
Since the CI for D3 = µ11 − µ22 is the farthest away from 0, µ11 and µ22 are significantly different
from each other at level .05.

20.5
Brainstorming. A researcher investigated whether brainstorming is more effective for larger groups
than for smaller ones by setting up four groups of agribusiness executives, the group sizes being two,
three, four, and five, respectively. He also set up four groups of agribusiness scientists, the group sizes
being the same as for the agribusiness executives. The researcher gave each group the same problem:
”How can Canada increase the value of its agricultural exports?” Each group was allowed 30 minutes
to generate ideas. The variable of interest was the number or different ideas proposed by the group. The
results, classified by type of group (factor A) and size of group (factor B), were:

Factor B (size of group)

Factor A j=1 j=2 j=3 j=4
(type of group) Two Three Four Five
i=1 Agribusiness executives 18 22 31 32
i=2 Agribusiness scientists 15 23 29 33
Assume that no-interaction ANOVA model is appropriate.

a. Plot the data in the format of Figure 20.1 in the textbook (plot of the observations Yi j ). Does it appear
that interaction effects are present? Does it appear that factor A and factor B main effects are present?
Discuss.

b. Conduct separate tests for type of group and size of group main effects. In each test, use level of
significance α = .01 and state the alternatives, decision rule, and conclusion. What is the P-value for
each test?

c. Obtain confidence intervals for D1 = µ.2 − µ.1 , D2 = µ.3 − µ.2 , and D3 = µ.4 − µ.3 ; use the Bonferroni
procedure with a 95 percent family confidence coefficient. State your findings.

d. Is the Bonferroni procedure used in pan (c) the most efficient one here? Explain.

Solution:

a. The plot of the observations are shown below:

4
 It appears that interaction effects are very little, since the two curves are nearly parallel. Factor A
main effects are also very little, since the two curves are nearly overlapped. Factor B main effects are
present, since the two curves have large departure from horizontal.

b. We first construct the ANOVA table. Note that the error sum of squares is given by S S AB, which
can be obtained by S S AB = S S T O − S S A − S S B in R.
Source SS df MS
Type of group 1.125 1 1.125
Size of group 318.375 3 106.125
Error 6.375 3 2.125
Total 325.875 7
F test for type of group main effects:
H0 : α1 = α2 = 0, Ha : not both α1 and α2 are 0. F ∗ = MS A/MS AB = 1.125/2.125 = .53,
F(1 − α; a − 1, (a − 1)(b − 1)) = F(.99; 1, 3) = 34.1. If F ∗ < 34.1, conclude H0 , otherwise Ha . Here
conclude H0 . P-value=.52.
H0 : β1 = β2 = β3 = β4 = 0, Ha : not all βi ’s are 0. F ∗ = MS B/MS AB = 106.125/2.125 = 49.94,
F(1 − α; b − 1, (a − 1)(b − 1)) = F(.99; 3, 3) = 29.5. If F ∗ < 29.5, conclude H0 , otherwise Ha . Here
conclude Ha . P-value=.005.

5
 D̂1 = Ȳ.2 − Ȳ.1 = 6, D̂2 = Ȳ3. − Ȳ2. = 7.5, D̂3 = Ȳ.4 − Ȳ.3 = 2.5. s(D̂1 ) = s(D̂2 ) = s(D̂3 ) =
c. √
2MS AB/a = 1.4577. B = t(1 − α/(2g); (a − 1)(b − 1)) = t(.99167; 3) = 4.857.
Thus, the 95% family-wise CI for D1 , D2 and D3 using the Bonferroni procedure are 6±4.857(1.4577) =
[−1.08, 13.08], 7.5 ± 4.857(1.4577) = [.42, 14.58] and 2.5 ± 4.857(1.4577) = [−4.58, 9.58].

d. Scheffe’s
√ procedure and Tukey’s procedure are √ also applicable in part (c).
S = (b − 1)F(1 − α; b − 1, (a − 1)(b − 1)) = 3F(.95; 3, 3) = 5.275, T = √1 q(1 − α; b, (a − 1)(b −
2
1)) = √1 q(.95; 4, 3) = 4.822.
2
Since T < B < S , the Tukey’s procedure is the most efficient in part (c).

21.7
Fat in diets. A researcher studied the effects of three experimental diets with varying fat contents on
the total lipid (fat) level in plasma. Total lipid level is a widely used predictor of coronary heart disease.
Fifteen male subjects who were within 20 percent of their ideal body weight were grouped into five
blocks according to age. Within each block, the three experimental diets were randomly assigned to the
three subjects. Data on reduction in lipid level (in grams per liter) after the subjects were on the diet for
a fixed period of time follow.

Fat Content of Diet

Block j=1 j=2 j=3
i Extremely Low Fairly Low Moderately Low
1 Ages 15-24 .73 .67 .15
2 Ages 25-34 .86 .75 .21
3 Ages 35-44 .94 .81 .26
4 Ages 45-54 1.40 1.32 .75
5 Ages 55-64 1.62 1.41 .78

a. Why do you think that age of subject was used as a blocking variable?

b. Obtain the residuals for randomized block model (21.1) and plot them against the fitted values. Also
prepare a normal probability plot of the residuals. What are your findings?

c. Plot the responses Yi j by blocks in the format or Figure 21.2 in the textbook. What does this plot
suggest about the appropriateness of the no-interaction assumption here?

Solution:

a. The total lipid level may vary according to age, so age is a confounding factor.

b. The residuals are

e1 = (−0.0527, −0.0127, 0.004, −0.0227, 0.084),

e2 = (0.0053, −0.0047, −0.008, 0.0153, −0.008),
e3 = (0.0473, 0.0173, 0.004, 0.0073, −0.076),

6
where ei = (e1i , e2i , e3i , e4i , e5i ). The residuals vs fitted values plot is shown below:

The normal Q-Q plot is shown below:

7
c. The plot of the observations is shown below:

8
 Since the five curves corresponding to five blocks are nearly parallel, there is little interaction be-
tween block and treatment. Thus the no-interaction assumption is appropriate here.

21.8
Refer to Fat in diets Problem 21.7. Assume that randomized block model (21.1) is appropriate.
a. Obtain the analysis of variance table.
b. Prepare a bar-interval graph of the estimated treatment means, using 95 percent confidence intervals.
Does it appear that the treatment means differ substantially here?
c. Test whether or not the mean reductions in lipid level differ for the three diets; use α = .05. State the
alternatives, decision rule, and conclusion. What is the P-value of the test?
d. Estimate L1 = µ.1 − µ.2 and L2 = µ.2 − µ.3 using the Bonferroni procedure with a 95 percent family
confidence coefficient. State your findings.
e. Test whether or not blocking effects are present; use α = .05. State the alternatives, decision rule,
and conclusion. What is the P-value of the test?

9
f. A standard diet was not used in this experiment as a control. What justification do you think the
experimenters might give for not having a control treatment here for comparative purposes?

Solution:

a. The ANOVA table is given below:

Source SS df MS
Blocks 1.41896 4 .35474
Fat content 1.32028 2 .66014
Error .01932 8 .002415
Total 2.75856 14
√
b. Ȳ.1 = 1.11, Ȳ.2 = .992, Ȳ.3 = .43, s(Ȳ.i ) = MS BL.T R/5 = .022, i = 1, 2, 3, t(.975; 8) = 2.306.
95 percent confidence intervals for treatment means are:

Ȳ.1 ± t(.975; 8)s(Ȳ.1 ) = 1.11 ± 2.306(.022) = [1.059, 1.161],

Ȳ.2 ± t(.975; 8)s(Ȳ.2 ) = .992 ± 2.306(.022) = [.941, 1.043],
Ȳ.3 ± t(.975; 8)s(Ȳ.3 ) = .43 ± 2.306(.022) = [.379, .481].

It appears that the treatment means differ substantially here.

c. H0 : τ1 = τ2 = τ3 = 0, Ha : not all τ j ’s are 0.

F ∗ = MS T R/MS BL.T R = .66014/.002415 = 273.35, F(.95; 2, 8) = 4.46.
If F ∗ < 4.46 conclude H0 , otherwise Ha . Conclude Ha . P-value≈ 0.
√
d. L̂1 = .118, L̂2 = .562, s(L̂i ) = 2MS BL.T R/5 = .03108, i = 1, 2, B = t(.9875; 8) = 2.7515.
95 percent family wise confidence intervals are:

L̂1 ± B · s(L̂1 ) = .118 ± 2.7515(.03108) = [.032, .204],

L̂2 ± B · s(L̂2 ) = .562 ± 2.7515(.03108) = [.476, .648].

e. H0 : all ρi equal 0 (i = 1, · · · , 5), Ha : not all ρi equal 0.

F ∗ = MS BL/MS BL.T R = .35474/.002415 = 146.89, F(.95; 4, 8) = 3.84.
If F ∗ < 3.84 conclude H0 , otherwise Ha . Conclude Ha . P-value≈ 0.

10