You are on page 1of 24

Click on the desired subsection below to position to that page.

You may also


scroll to browse the document.

Supplemental Topic 4: Two-Way Analysis of Variance


S4.1 Assumptions and Models for Two-Way ANOVA
S4.2 Testing for Main Effects and Interactions
Exercises
S4-W3527 9/28/05 4:05 PM Page S4-1

SUPPLEMENTAL
TOPIC 4

Wide Group/Iconica/Getty Images

What simple gesture might increase her tip?


S e e E x a m p l e S 4 . 4 (p. S4-16)

S4-1
S4-W3527 9/28/05 4:05 PM Page S4-2

Two-Way Analysis
Chapter
of Title
Variance

In Chapter 16, we learned how to compare means of three or more groups using analy-
sis of variance. Now we learn how to compare means for a continuous variable when
individuals are grouped by combinations of two categorical variables. The procedure is
called two-way analysis of variance.
Chapter opeing quote (8pts b/r to box end, 36pts to ch op text)

I
n Chapter 16, we introduced two-way analysis of variance (ANOVA) and
described some of the relevant ideas. However, we did not furnish de-
tails about the models and analyses. In this chapter, we provide some
of those details.

S4.1 Assumptions and Models


for Two-Way ANOVA
Two-way analysis of variance is appropriate when the response variable is
continuous and there are two categorical explanatory variables. For instance,
we might want to compare grade point averages for students categorized by
whether they attend school full-time or part-time and whether they are work-
ing in addition to attending school. Or we might want to compare pulse rates for
adults on the basis of gender (male, female) and smoking behavior (yes, no).

Example S4.1 Sleep Hours Based on Gender and Seat Location In Exercise 16.31, we
showed the relationship between responses to the survey question hours of
sleep the previous night (continuous) and the two categorical explanatory
variables gender and seat location (using the UCDavis1 dataset on the CD).
Throughout the chapter, this icon Figure S4.1 shows the mean responses to the question How many hours did
introduces a list of resources on the you sleep last night? categorized by gender and preferred seat location. Notice
StatisticsNow website at http://
1pass.thomson.com that will: that the mean hours of sleep for students in the front of the class differ consid-
Help you evaluate your knowledge erably for males and females, with females reporting more than a full hour more
of the material sleep than males. The differences between males and females are smaller for
Allow you to take an exam- those who prefer the middle and back of the class, but there are still differences.
prep quiz
Provide a Personalized Learning Questions of interest include the following:
Plan targeting resources that Is there an overall difference in mean hours of sleep for males and
address areas you should study
females?

S4-2
S4-W3527 9/28/05 4:05 PM Page S4-3

Two-Way Analysis of Variance S4-3

Are there overall differences in mean hours of sleep for those who prefer
to sit in the front, middle, and back of the class?
Is there an interaction between gender and seat location, so that the dif-
ference in mean hours of sleep for males and for females changes on the
basis of seat location preference? This certainly appears to be the case,
but the question of interest is about the population these students repre-
sent, and the difference observed in the sample data may or may not
reect a real difference in the population.

7.2 Female

Mean hours of sleep


6.7
Male

6.2

Front Middle Back


Seat location

Figure S4.1 Main effects and interaction effects 

Main Effects and Interaction Effects


In Chapter 16, we dened the terms main effect and interaction. A main effect
describes the individual relationship between one of the categorical variables
and the response variable. In Example S4.1, one main effect is the relationship
between gender and average hours of sleep. The other main effect is the rela-
tionship between preferred seat location and average hours of sleep.
An interaction effect describes the extent to which the relationship between
one explanatory variable and the response variable depends on the other ex-
planatory variable. In Example S4.1, if the difference in mean hours of sleep for
males and females is similar for all three preferred seat location categories, then
there is little or no interaction between gender and seat location in their rela-
tionship with hours of sleep. On the other hand, if the malefemale sleep dif-
ference changes for the seat location groups, as appears to be the case, then
there is an interaction effect.

Assumptions and Conditions for Two-Way ANOVA


The assumptions and conditions that are required for two-way ANOVA are the
same as those listed for one-way analysis of variance in Section 16.1. The only
difference is in how the populations of interest are dened. In one-way ANOVA,
we are interested in comparing the means of k populations. In two-way ANOVA,
we are interested in comparing the means of ab populations, where a is the
number of categories of the rst explanatory variable and b is the number of
categories of the second explanatory variable. For instance, in Example S4.1,
the rst explanatory variable had a  2 categories (male, female), and the sec-
ond explanatory variable had b  3 (front, middle, back) categories. We are in-
S4-W3527 9/28/05 4:05 PM Page S4-4

S4-4 Supplemental Topic 4

terested in comparing the mean hours of sleep for the ab  6 populations de-
ned by the combinations of the two explanatory variables.
The comparisons in two-way ANOVA are more complicated than those
made in one-way ANOVA, but the assumptions are not. Here are the assump-
tions about the populations and the samples representing them, repeated from
Section 16.1:
The samples are independent random samples.
The distribution of the response variable is a normal curve within each
population.
The different populations may have different means.
All populations have the same standard deviation, s.
As in Chapter 16, these assumptions may be relaxed somewhat in practice. For
instance, it is rarely the case that all of the population standard deviations are
identical. As long as they are similar, the procedures will work. Similarly, the as-
sumption of normal curves can be relaxed if the sample sizes are moderate and
the samples dont have major outliers. The assumption about independent ran-
dom samples is stricter than is usual in practice. In general, separate represen-
tative samples or one combined sample that can be split into separate samples
using the two categorical variables will sufce.

Model for One-Way ANOVA


In Chapter 16, we did not explicitly state the model for observations in one-
way ANOVA. We do so now to facilitate understanding of the two-way ANOVA
model. In one-way ANOVA, we assume that there is a random sample from
each of k normal populations and that population i has mean mi and standard
deviation s.

definition Let xij be the notation for observation j from population i, where i  1, 2, . . . , k
and j  1, 2, . . . , ni. The one-way ANOVA model is
xij  mi  Pij
where the collection of Pij are independent normal random variables with mean
0 and standard deviation s.

This one-way ANOVA model is less complicated than it might appear. It says
that each observation can be thought of as being made up of the population
mean and an additional term that may be positive or negative. For instance,
suppose a womans height is 67 and she belongs to a population of women
whose mean height is 65. Then her height can be written as 67  65  2. In
the format of the one-way ANOVA, her height would be written as 67  xij 
mi  Pij  65  2. Another womans height might be 64, in which case for her,
Pij  1. As in the regression model, P is sometimes called the error term, which
is a misnomer because the amount by which an individual differs from the
population mean is certainly not an error in most cases.
S4-W3527 9/28/05 4:05 PM Page S4-5

Two-Way Analysis of Variance S4-5

Models for Two-Way ANOVA


Before dening a model for two-way ANOVA, lets dene some additional terms
and notation. The categorical explanatory variables in ANOVA are called fac-
tors. In two-way ANOVA there are two factors, and it is standard to call them
Factor A and Factor B. In Example S4.1, Factor A is gender and Factor B is seat
location. (There is nothing specic about this order; we could dene Factor A
to be seat location and Factor B to be gender.) Factor A has a categories, and
Factor B has b categories. Therefore, there are ab populations, with a random
sample from each one for which the response variable is measured.
We use the letter i to indicate the category level for Factor A and the let-
ter j to indicate the category level for Factor B. Thus, mij is the population mean
for level i of Factor A and level j of Factor B. For instance, in Example S4.1, Fac-
tor A (gender) has level 1  male and level 2  female. Factor B (seat location)
has level 1  front, level 2  middle, and level 3  back. Thus, the mean hours
of sleep for the population of male students who prefer to sit in the front of the
room is m11. The notation m23 represents the mean hours of sleep for the popu-
lation of female students (i  2) who prefer to sit in the back of the room ( j  3).
We also need notation and a denition for the combined mean for each level
of one factor averaged across the levels of the other factor. We use a dot (.)
to represent the factor over which the average has been taken. The denitions
follow.

definition The population mean for Factor A, level i is


1 b
mi .  a mij
b j1
The population mean for Factor B, level j is
1 a
m. j  a mij
a i1
The overall population mean is
1 a b
m. .  a a mij
ab i1 j1

We use the letter k to indicate which of the individual observations we are


writing. The sample observations are thus denoted as xijk, where i and j specify
the levels of Factors A and B, respectively, and k species the individual obser-
vations. For instance, in Example S4.1, observation x125 represents the hours of
sleep reported by the fth student (k  5) in the sample of male (i  1) students
who prefer to sit in the middle of the room ( j  2).
The size of the sample from the population with level i of Factor A and level j
of Factor B is nij. In many cases, all of the ab sample sizes will be identical, and
we will simply use the letter n for each one. The combined sample size is de-
noted by N. If all of the sample sizes are the same, then N  nab. If not, then N
is the sum of all of the individual nij.
S4-W3527 9/28/05 4:05 PM Page S4-6

S4-6 Supplemental Topic 4

in summary Notation for Two-Way ANOVA

There are a levels of Factor A and b levels of Factor B. Individual observations


are denoted by xijk where i, j, and k represent the level of Factor A, the level
of Factor B, and the number of the individual observation within the sample,
respectively. Notation for means and sample sizes is as follows:
Population means:
mij  Population mean for Factor A level i and Factor B level j.
mi.  Population mean for Factor A level i.
m.j  Population mean for Factor B level j.
m..  Overall population mean for all ab combinations of Factor A and
Factor B.
Sample sizes:
nij  Sample size for Factor A level i and Factor B level j.
ni.  Combined sample size for Factor A level i.
n.j  Combined sample size for Factor B level j.
N  Total sample size
Sample means:
The notation for sample means is the same as for population means, but
with m replaced by x. For instance, the sample mean for Factor A level i
and Factor B level j is xij.

definition Let xijk represent observation k from population ij, where i  1, 2, . . . , a; j  1,


2, . . . , b; and k  1, 2, . . . , nij. The general two-way ANOVA model is
xijk  mij  Pijk
where the collection of Pijk are independent normal random variables with
mean 0 and standard deviation s.

If we were simply to study the general two-way ANOVA model, it would pro-
vide no more information than the one-way model. They are identical except
for the additional subscript required for the notation in the general two-way
model. What makes the two-way ANOVA situation more useful is that we break
up the population mean into pieces. There are four pieces: the overall mean m. .
and an adjustment for each of Factor A, Factor B, and the possible interaction
between Factor A and Factor B. Before introducing notation for these pieces,
lets consider a simple example.

Example S4.2 Pulse Rates, Gender, and Smoking How are resting pulse rates different for
males and females? Does it depend on whether people smoke? On the basis of
data collected in various statistics classes, the following are reasonable guesses
for the population mean pulse rates, in beats per minute, for male and female
Watch a video example at http:// smokers and nonsmokers.
1pass.thomson.com or on your CD.
S4-W3527 9/28/05 4:05 PM Page S4-7

Two-Way Analysis of Variance S4-7

Smoker Nonsmoker Combined


Male m11  72 m12  70 m1.  71
Female m21  80 m22  74 m2.  77
Combined m.1  76 m.2  72 m..  74

Notice that the average of 71 for male pulse rates is 3 beats lower than the
overall mean of 74, and the average for females is 3 beats higher than the over-
all mean. Also, the mean of 76 for smokers is 2 beats above the overall mean
of 74, and for nonsmokers it is 2 beats below the overall mean. Thus we might
think we can dene the mean for male smokers as overall mean of 74 beats 
3 beats (for males)  2 beats (for smokers)  73 beats. But notice that in fact the
mean for male smokers isnt completely explained by adding those compo-
nents. The mean is 72 beats, one beat less than the added components pro-
duced. The additional 1 is due to the interaction effect. Even after accounting
for the fact that males have lower pulse rates and smokers have higher pulse
rates, an adjustment was needed for the combination of male and smoker. 

Trying to describe the individual means using only the overall mean and
a contribution for each factor is called the additive model. Including an inter-
action term produces the full model. If the additive model describes the data
almost as well as the full model, then there is little or no interaction between the
two factors in their relationship to the response variable.

definition The main effects for Factor A are dened as ai  mi .  m. . for i  1, 2, . . . , a.


The main effects for Factor B are dened as bj  m. j  m. . for j  1, 2, . . . , b.
The interaction effects are dened as abij  mij  (m. .  ai  bj) for all pairs i, j.
The additive model for two-way ANOVA is dened as
mij  m. .  ai  bj
The full model for two-way ANOVA is dened as
mij  m. .  ai  bj  abij
Notice that the interaction effects are equivalent to the differences between the
full model and the additive model for each mean.

It is easy to show by using algebra that the sum of the ai  0, the sum of the
bj  0, and, for each xed value of i, the sum over j of abij  0. Similarly, for each
xed value of j, the sum over i of abij  0.
The additive model and full model can be written for individual observa-
tions as well by replacing the mean mij in the general two-way ANOVA model
with the additive or full model version. The full model for each observation in
two-way ANOVA is
xijk  mij  Pijk  m..  ai  bj  abij  Pijk
S4-W3527 9/28/05 4:05 PM Page S4-8

S4-8 Supplemental Topic 4

where the Pijk are independent normal random variables with mean 0 and stan-
dard deviation s. The additive model for each observation is
xijk  mij  Pijk  m. .  ai  bj  Pijk
where the Pijk are independent normal random variables with mean 0 and stan-
dard deviation s.

E x a m p l e S 4 . 2 (cont.) Models for Pulse Rates Lets nd the main effects and the interaction terms
for the mean pulse rates in Example S4.2. The main effects for Factor A are
a1  m1 .  m. .  71  74  3 for males
a2  m2 .  m. .  77  74  3 for females
The main effects for Factor B are
b1  m. 1  m. .  76  74  2 for smokers
b2  m. 2  m. .  72  74  2 for nonsmokers
Notice that in both cases, the sum of the main effects across the levels of the fac-
tor is 0. This makes sense both intuitively and algebraically based on the de-
nition of the main effects. These values represent the amount by which each
mean differs from the overall mean, either as a positive or a negative amount.
These main effects tell us that, averaged over smoking behaviors, males
have an average pulse rate that is about 3 beats lower than the overall average,
whereas females have an average pulse rate about 3 beats higher. Similarly,
smokers have average pulse rates that are about 2 beats higher than the overall
average, and nonsmokers have an average pulse rate that is about 2 beats lower.
As was noted in Example S4.2, the additive model is not adequate to de-
scribe the individual population means. For instance, using the additive model
for male smokers would result in m11  m. .  a1  b1  74  3  2  73, which is
not correct because m11  72. The remainder is the interaction effect. The inter-
action effect for male smokers is
ab11  m11  1 m. .  ai  bj 2  72  73  1
You can verify that ab12  ab21  1 and ab22  1. In fact, the interaction ef-
fects must sum to 0 over each row and column.
Figure S4.2 displays an interaction plot for this situation, as described in
Chapter 16. If there were no interaction effect, the lines would be parallel. No-

Interaction plotMean pulse rate

80
Mean pulse rate

Female

75

Male

70

Figure S4.2 Interaction plot Smokers Nonsmokers


for Example S4.2 Smoking behavior
S4-W3527 9/28/05 4:05 PM Page S4-9

Two-Way Analysis of Variance S4-9

tice that the lines do not cross, but they are not parallel either. The plot indi-
cates an interaction between gender and smoking behavior in their relationship
with pulse rate. This can be seen by noticing that the difference between the
means for males and females is larger for smokers than it is for nonsmokers.

Interpreting Main Effects and Interactions


The main effects for each factor measure how much the means differ for the
levels of that factor, averaged over all levels of the other factor. The interaction
effects measure how much the mean for each combination of the two explana-
tory variables differs from what it would be if the variables were considered
separately.
If there are interaction effects, it may be misleading to average over one fac-
tor when comparing the means for the other factor. An important point to keep
in mind in interpreting main effects is that the interpretation is different when
interactions are present than when they are not. For instance, in Example S4.2,
the mean pulse rate of 76 for smokers is 4 beats per minute higher than the
mean of 72 for nonsmokers. However, the increase in means for smokers over
nonsmokers is only 2 beats per minute for males, whereas it is 6 beats per min-
ute for females. Thus, it would be misleading to say that smokers have a mean
pulse rate that is 4 beats higher than that of nonsmokers. The effect of smoking
(Factor B) should be reported separately for males and females, the categories
of Factor A. Similarly, the effect of Factor A should be reported separately for
the categories of Factor B. It is necessary to report the means for one factor
separately for the levels of the other factor only when interactions are present.
Sometimes the interaction is so strong that the differences in means for one
factor actually occur in the opposite direction for some levels of the other fac-
tor than for others. We saw this in Example S4.1 and Figure S4.1. In that ex-
ample, among students who prefer to sit in the front of the room, women slept
almost an hour more than men, on average. But for students who preferred the
back of the room, the men slept more on average than the women. This ex-
ample again illustrates that it is misleading to compare the means, in this case
mean hours of sleep, for the categories of one of the factors without taking into
account the other factor when interaction is present.

S4.2 Testing for Main Effects


and Interactions
There are three sets of hypotheses of interest in two-way ANOVA: one set for
each main effect and one set for the interactions. Each set of hypotheses is
tested by using an F-test. We will cover the intermediate details only for the sit-
uation in which all nij are equal because the formulas are more complicated
when the sample sizes differ. When all sample sizes are equal in ANOVA, we say
that we have a balanced ANOVA or a balanced design. When the sample sizes
are not equal, we say that we have an unbalanced design.
S4-W3527 9/28/05 4:05 PM Page S4-10

S4-10 Supplemental Topic 4

In fact, there are multiple approaches for testing the hypotheses in two-way
ANOVA with an unbalanced design, and there is some debate about which ap-
proach is best. The details are beyond the scope of this book, and we will simply
use the most common approach.

Hypotheses in Two-Way ANOVA


In one-way ANOVA, we were interested in whether the population means were
all equal. In two-way ANOVA, we are also interested in whether the ab popula-
tion means are equal. But in addition, if they do differ, we want to know whether
it is Factor A, Factor B, the interaction, or some combination that is responsible
for the difference. Therefore, we usually forego the overall test for a difference
in means (which can be done using one-way ANOVA on the ab groups) and con-
duct separate tests for Factor A, Factor B, and the interaction.
In each case, we can write the hypotheses in words or in symbols. We will
show how to write them in words and then illustrate two ways to write them in
symbols.

Hypotheses for Factor A


H0: The population means are equal for all levels of Factor A.
Ha: The population means are not all equal for all levels of Factor A.
or
H0: m1 .  m2 .   ma.
Ha: Not all mi . are equal.
or
H0: a1  a2   aa  0
Ha: Not all ai  0.

Hypotheses for Factor B


H0: The population means are equal for all levels of Factor B.
Ha: The population means are not all equal for all levels of Factor B.
or
H0: m. 1  m. 2   m. b
Ha: Not all m. j are equal.
or
H0: b1  b2   bb  0
Ha: Not all bj  0.

Hypotheses for Interaction


H0: The population means t the additive model.
Ha: The population means do not t the additive model.
or
H0: mij  m. .  ai  bj for all i, j.
Ha: mij  m. .  ai  bj for at least one i, j pair.
or
H0: All abij  0.
Ha: Not all abij  0.
S4-W3527 9/28/05 4:05 PM Page S4-11

Two-Way Analysis of Variance S4-11

Sums of Squares, Degrees of Freedom,


and Mean Squares
Recall from Chapter 16 that in one-way ANOVA, the total variability among
sample observations, SS Total, can be divided into two parts: SS Groups and SS
Error. SS Groups measures the amount by which the group means differ from
the overall mean, and SS Error measures the amount by which individual ob-
servations differ from their own group mean.
In two-way ANOVA, we can still divide SS Total into the same two pieces. But
now SS Groups, which represents the amount by which the ab group means dif-
fer from the overall mean, is further divided into pieces reecting mean differ-
ences for Factor A, Factor B, and the AB interaction.
We need some additional notation before explaining how to divide SS
Groups into pieces. The sample observations can be used to estimate the main
effect and interaction terms in the model, and those estimates are then used to
construct sums of squares for the two factors and the interaction. The estimates
are constructed by replacing the population means that dene the terms with
their corresponding sample versions.

formula Estimating Main Effect and Interaction Terms


Note: These estimates are appropriate when the sample sizes are all equal
that is, for a balanced design.
The estimate of the overall mean m is m  x. ..
The estimates of the Factor A effects ai are ai  xi .  x. . for i  1, 2, . . . , a.
The estimates of the Factor B effects bj are bj  x. j  x. . for j  1, 2, . . . , b.
The estimates of interaction effects (ab)ij are ab ij  xij  1 m  ai  bj 2 for
all i, j pairs.

These estimates are used to allocate the variability among the sample means
to sources for the two factors and interaction. Using two-way ANOVA notation,
SS Groups, which measures the variability among the sample means, is written
as follows:

SS Groups  n a a 1xij  x. . 2 2
i j

The summand, which is the amount by which the sample mean for group ij dif-
fers from the overall sample mean, can be rewritten by adding together a con-
tribution for each factor and for interaction:
1xij  x. . 2  1xi .  x. . 2  1x. j  x. . 2  3 1xij  1 m  ai  bj 2 4  ai  bj  1
ab 2 ij
If the group mean and overall sample mean are equal, then this summand
will be 0. This could happen if all three of the contributions on the right are 0 or
if they somehow cancel each other out. If the latter is the case, then when the
terms are squared to form sums of squares, it will become evident that they are
not all 0.
The SS Groups can be rewritten as follows, using the above formulation for
the summand and some algebra.
S4-W3527 9/28/05 4:05 PM Page S4-12

S4-12 Supplemental Topic 4

formula Sums of Squares for Two-Way ANOVA

SS Groups  n a a 1xij  x. . 2 2  nb a a2i  na a b2j  n a 1


ab 2 2ij
i j i j i, j

 SSA  SSB  SSAB


SS Error  a 1nij  12s 2ij  a a a 1xijk  xij 2 2
i, j i j k

SS Total  SS Groups  SS Error

Notice that SSA is dened as the sum of the squares of the individual estimated
Factor A terms, and similarly for SSB and SSAB. The sum is actually taken over
all nab individual observations, but because the term is the same for all nb in-
dividuals at level i, the sum over the subscripts j and k is summarized in the
multiple nb. A similar simplication occurs for SSB and SSAB.
The degrees of freedom can be divided in a similar way. The technical details
are beyond the level of discussion here, but the results are as follows:

formula Degrees of Freedom for Two-Way ANOVA


Note: These formulas apply whether or not the sample sizes are equal for all
groups
Total degrees of freedom  N  1  df for groups  df for error  (ab  1) 
(N  ab)
df for groups  (ab  1)  (a  1)  (b  1)  (a  1)(b  1)  (df for A) 
(df for B)  (df for AB)
When the sample sizes are equal, df for error can be written as (N  ab) 
(nab  ab)  ab(n  1), and the degrees of freedom for all of the other sources
remain the same.

As in one-way ANOVA, the mean square for each source is MS  SS>df.


Thus, MSA  SSA>(a  1), MSB  SSB>(b  1), MSAB  SSAB>(a  1)(b  1), and
MS Error  SS Error>(N  ab).

F-Tests for Two-Way ANOVA


The mean squares for Factor A, Factor B, and the AB interaction represent the
total variability among the response variable measurements that can be attrib-
uted to each of those sources. If there is really no effect in the population for A,
B, or AB, then the mean square for that term should be small in relation to the
error mean square (MSE). In that case, we would not want to reject the null hy-
pothesis for that source. For instance, if H0: all ai  0 is true, then it is reason-
able to expect SSA  nb g i a2i to be small (relative to MSE). On the other hand,
if some of the ai are large, then SSA should also be large, and in that case, the
correct decision would be to reject the null hypothesis and conclude that at
least some of the Factor A means do differ.
To determine whether the Factor A, Factor B, and AB interaction mean
squares are large enough for us to reject the corresponding null hypotheses, we
construct an F-test statistic for each one. The denominator in each case is MSE,
S4-W3527 9/28/05 4:05 PM Page S4-13

Two-Way Analysis of Variance S4-13

Table S4.1 Two-Way ANOVA Table

Degrees of Sum of
Source Freedom Squares Mean Square F
Factor A a1 SSA MSA  SSA>(a  1) MSA>MSE
Factor B b1 SSB MSB  SSB>(b  1) MSB>MSE
AB interaction (a  1)(b  1) SSAB MSAB  SSAB>(a  1)(b  1) MSAB>MSE
Error N  ab SSE MSE  SSE>(N  ab)
Total N1 SSTO

and the numerator is the MS for A, B, or AB. Table S4.1 summarizes the relevant
information and is called a two-way analysis of variance table. In practice, the
formulas in the table are replaced with their numerical values, and the terms
Factor A and Factor B are replaced by descriptive names for the variables that
they represent.
In each case, test the hypotheses about the effect (A, B, AB) using the corre-
sponding F-test statistic. Computer software will provide p-values, or Table A.4
in the Appendix can be used to nd the critical values for a  .05 or .01. (Note
that we are using the symbol a here to represent the signicance level as we have
throughout the book, not in the two-way ANOVA context of a Factor A effect. The
distinction should be clear from the context.) In each case, the denominator de-
grees of freedom  (N  ab), corresponding to MSE, and the numerator degrees
of freedom are found from the appropriate row of the ANOVA table. The details
of the hypotheses, test statistics, and p-values are shown in Table S4.2.

Table S4.2 Hypotheses and Test Statistics for Two-Way ANOVA

F-Test p-value  Area Above Value of F-Test


Effect Null Hypothesis Statistic Statistic for F-Distribution with df of:
Factor A H0: All mi. equal MSA>MSE (a  1), N  ab
Factor B H0: All m.j equal MSB>MSE (b  1), N  ab
AB interaction H0: All (ab)ij  0 MSAB>MSE (a  1)(b  1), N  ab

MINITAB t i p Two-Way ANOVA


Use StatbANOVAbTwo-way if the sample sizes are all equal. Identify the
response variable and the two factors (labeled as row factor and col-
umn factor) in the dialog box.
You can use StatbANOVAbGeneral linear model whether or not the
sample sizes are equal, and there are some features not available in the
Two-way procedure. Identify the response variable. For the model, write
A B A*B, where A and B are the names of the columns containing the lev-
els of Factors A and B, respectively. In the Options box, click Adjusted
(Type III) for Sums of Squares (the default).
If you would like to compare the means for one or both factors, or for the
ab individual means, use StatbANOVAbGeneral linear model and open
the Comparisons box. Enter the terms for which you want comparisons in
the box labeled Terms, using the names of the columns for Factor A and
Factor B and using the names with * between them for the ab means.
Click the Tukey box, and the display will show pairwise comparisons us-
ing the Tukey method, as described in Chapter 16 for one-way ANOVA.
S4-W3527 9/28/05 4:05 PM Page S4-14

S4-14 Supplemental Topic 4

Examples and Interpretation


of Two-Way ANOVA Results
It is instructive to consider three types of examples because the results are in-
terpreted differently in each case. The interpretation depends on whether there
are signicant interactions and whether there are signicant main effects for
one or both variables. We cover examples for each of these three situations:
1. Signicant main effects but no signicant interaction
2. One signicant main effect, one nonsignicant main effect, and signi-
cant interaction
3. Signicant main effects and signicant interaction
Be very careful about interpreting main effects if interaction is present. In
that case, the differences in means for the levels of Factor A should be examined
separately for each level of Factor B, and vice versa. We have already seen this
concept in Examples S4.1 and S4.2. In both cases, there appeared to be an in-
teraction between the two factors, and we examined the difference in means for
one factor separately for the levels of the other factor.

Example S4.3 Nature Versus Nurture in IQ Scores Are the IQs of children related to the
socioeconomic status (SES) of their parents? If children are adopted, are their
IQs related to the SES of their biological parents, their adoptive parents, or some
combination? These are the questions that French researchers set out to in-
vestigate in a study. They identied children from adoption registers in France
whose biological and adoptive parents were from the highest and the lowest
SES categories. SES was determined by occupation and years of education. Fac-
tor A is the SES of the adoptive parents (high, low), and Factor B is the SES of
the biological parents (high, low). The researchers attempted to create a bal-
anced design with ten children for all groups but found only eight for the com-
bination with high SES for biological parents but low SES for the adoptive par-
ents. The IQs and categories for the 38 children are shown in Table S4.3. (Source:
Data reported in Ramsey and Schafer, 1997, p. 396; adapted from Capron and
Duyme, 1991.)

Table S4.3 Data for Example S4.3

Biological Parents SES (Factor B)


Adoptive Parents
SES (Factor A) High Low
High 136, 99, 121, 133, 125, 131, 103, 115, 94, 103, 99, 125, 111, 93, 101, 94, 125, 91
116, 117
Low 98, 99, 91, 124, 100, 116, 113, 119 92, 91, 98, 83, 99, 68, 76, 115, 86, 116

The null hypotheses for this example can be written in context as follows:
For Factor A (adoptive parents SES):
H0: The mean IQs for the populations of children with high-SES and low-SES
adoptive parents are equal.
S4-W3527 9/28/05 4:05 PM Page S4-15

Two-Way Analysis of Variance S4-15

For Factor B (biological parents SES):


H0: The mean IQs for the populations of adopted children with high-SES
and low-SES biological parents are equal.
For the AB interaction:
H0: The relationship between mean IQ and adoptive parents SES status
does not depend on the biological parents SES.
Figure S4.3 displays an interaction plot for this example, showing the means
for the four categories of the two factors. Notice that the lines are parallel. This
indicates that there is little or no interaction. The difference between mean IQ
of children of high- and low-SES adoptive parents is about the same, regardless
of the SES of their biological parents.

120 Adoptive parents SES = High


Mean IQ of child

110

100
Adoptive parents SES = Low

High Low
SES of biological parents

Figure S4.3 Interaction plot for SES and IQ in Example S4.3

The plot also indicates that both main effects are present. The gap between
the lines for low and high adoptive parents SES indicates that the means for the
two levels of Factor A differ (by the amount shown by the gap, about 12 points).
The slope of the lines indicates that the means for the two levels of Factor B dif-
fer. Notice that for both lines, the mean IQ for children of high-SES biological
parents is about 16 points higher than the mean IQ for children of low-SES bio-
logical parents.
The individual means, analysis of variance table, and other results produced
by Minitab for this example are shown in Figure S4.4. The two different SS col-
umns, Seq SS and Adj SS, represent two methods for analyzing an unbal-
anced design. The results are based on the Adj SS, as is common practice.
When the design is not balanced, the three hypothesis tests of interest are not
independent. The Adj SS method tests for each effect after adjusting for the
presence of the other effects. The Seq SS method tests the effects in order,
starting with Factor A; thus, the results depend on which factor is designated as
A and which is designated as B.
The p-value for the test of the Factor A effect is .011, so the null hypothesis
can be rejected. Remember that the null hypothesis for Factor A stated that the
population mean IQs are equal for children of high-SES and low-SES adoptive
parents. This is clearly not the case. In fact, the sample means for the children
of high- and low-SES adoptive parents are 111.60 and 99.11, respectively, a dif-
S4-W3527 9/28/05 4:05 PM Page S4-16

S4-16 Supplemental Topic 4

ference of over 12 IQ points. The p-value conrms that the difference seen in
the sample results reects a real difference in the population means.
The p-value for the test of Factor B effect is .001, so the null hypothesis can
be rejected. The null hypothesis for Factor B stated that the population mean
IQs are equal for children of high-SES and low-SES biological parents. This is
also clearly not true; the sample means are 114.22 and 98.00, respectively, and
the low p-value conrms that the difference in the sample means reects a real
difference in the population means.
The p-value for the test of AB interaction is .917, so the null hypothesis can-
not be rejected. Consistent with the interaction plot, this p-value conrms that
the difference in mean IQs for the levels of one factor does not depend much on
the level of the other factor.
In summary, the mean IQs of children are related to the SES of both their
biological and adoptive parents. Children whose adoptive parents have high
SES have average IQs that are about 6.6 points higher than the overall average,
and children whose biological parents have high SES have average IQs that are
about 7.7 points higher than the overall average. The combined effect is addi-
tive, so children for whom both sets of parents have high SES have average IQs
that are about 14 points above the overall average. This can be seen in Figure
S4.4; the mean for the children with high SES for both sets of parents is 119.6,
whereas the overall mean is 105.68.

Mean IQ for Adopted Children


Rows: ADOPTIVE Columns: BIOLOGIC

High Low All


High 119.60 103.60 111.60
Low 107.50 92.40 99.11
All 114.22 98.00 105.68

Two-Way ANOVA Using General Linear Model


General Linear Model: IQ versus ADOPTIVE, BIOLOGIC

Factor Type Levels Values


ADOPTIVE fixed 2 High Low
BIOLOGIC fixed 2 High Low

Analysis of Variance for IQ, Using Adjusted SS for Tests

Source DF Seq SS Adj SS Adj MS F P


ADOPTIVE 1 1477.6 1277.4 1277.4 7.31 0.011
BIOLOGIC 1 2291.5 2275.8 2275.8 13.02 0.001
ADOPTIVE*BIOLOGIC 1 1.9 1.9 1.9 0.01 0.917
Error 34 5941.2 5941.2 174.7
Total 37 9712.2

Figure S4.4 Minitab results for Example S4.3

Example S4.4 Happy Faces and Restaurant Tips Revisited In Chapter 16, we presented
an example in which a male server and a female server in a restaurant drew a
happy face on some checks presented to customers and not on others, ran-
domly determined. The percent tip accompanying each check was recorded.
The researcher was interested in knowing whether the happy face would in-
crease or decrease the tip percentage and if it depended on whether the server
was male or female. One shortcoming of the study was that only one server
S4-W3527 9/28/05 4:05 PM Page S4-17

Two-Way Analysis of Variance S4-17

of each sex was used, so the results may apply to those two servers only. None-
theless, it is a useful example for illustrating a somewhat unusual situation.
There is a signicant interaction effect, but only one of the main effects was
signicant.
The response variable for this example is tip percent. Factor A is the message
on the check, with level 1  none and level 2  happy face. Factor B is sex, with
level 1  female and level 2  male.
Figure S4.5, reprinted from Chapter 16, shows the interaction plot. Notice
that the lines are not parallel. For the female, tips were higher with the happy
face message, whereas for the male, tips were lower with the happy face mes-
sage. This difference indicates that there is an interaction effect. The plot also
demonstrates that there is a sex effect. Overall, the mean tip for the female is
much higher than that for the male. However, it appears from the plot that there
is not a message effect. Notice that the mean tip percentage for no message is
about the same as the mean tip percentage for the happy face message when
averaged across the two sexes.

Interaction PlotMeans for tip percentage


33
Mean tip percentage

Female
28

23

Male

18

None Happy face


Message

Figure S4.5 Mean tip percentage related to sex of server and


message on check (none or happy face)

The means and analysis of variance results are shown in Figure S4.6 on the
next page. The numerical results conrm what the plot illustrated. The p-value
for testing for the AB interaction is .049, so the null hypothesis of no inter-
action can be rejected using a  .05. The p-value for the Factor B (sex) effect is
.000, indicating a strong difference in average tip percents for the male and fe-
male servers.
However, the p-value for testing for a Factor A effect is .715, so the null hy-
pothesis cannot be rejected. Remember that the null hypothesis for Factor A in
this example is that the population mean tip percentage is the same when no
message is written on the check as when the happy face is written. We have seen
that this is not true for the female, whose tips went up with the happy face, nor
is it true for the male, whose tips went down for the happy face. It is true only
when averaged across them. This illustrates the danger of interpreting a main
effect overall, in the presence of interaction. When interaction is present, it is
very important to examine the means for the different levels of one factor sepa-
rately for the levels of the other factor. In this example, it is important to exam-
ine the difference in means for the two messages separately for the male and the
female.
S4-W3527 9/28/05 4:05 PM Page S4-18

S4-18 Supplemental Topic 4

Means for Tip Percent


Message Female Male All
Happy Face 33.041 17.776 25.239
None 27.784 21.406 24.740
All 30.354 19.508 24.992

Two-Way ANOVA Using General Linear Model


General Linear Model
Factor Type Levels Values
Message fixed 2 Happy Face None
Sex fixed 2 Female Male

Analysis of Variance for TipPct, Using Adjusted SS for Tests

Source DF Seq SS Adj SS Adj MS F P


Message 1 5.5 14.7 14.7 0.13 0.715
Sex 1 2628.1 2602.0 2602.0 23.69 0.000
Message*Sex 1 438.7 438.7 438.7 3.99 0.049
Error 85 9335.5 9335.5 109.8
Total 88 12407.9

Figure S4.6 Means and analysis of variance table for Example S4.4

Example S4.5 Does Smoking Lead to More Errors? This example is adapted from a study
of the relationship between smoking and performance on three types of tasks.
The study was done by Spilich, June, and Renner (1992). Howell (1997, p. 416)
simplied the data for illustrative purposes; the data presented here are a slight
modication based on communication with Professor Howell, who thought
that this adaptation was more similar to the original results.
Participants were asked to perform one of three types of tasks. Factor A is
the type of task performed. The rst type (level 1) is pattern recognition, in
which participants had to locate a target on a computer screen. The second type
(level 2) is a cognitive task involving recall of a passage read earlier. The third
type (level 3) is a driving simulation computer game. For Factor B, participants
were categorized by their smoking behavior. Level 1 is nonsmoker, level 2 is for
smokers who did not smoke within 3 hours of the task, and level 3 is for partici-
pants who were smoking during the task. The response variable is the number
of errors made while performing the task.
The data are presented in Table S4.4. An interaction plot is shown in Fig-
ure S4.7, and analysis of variance results are shown in Figure S4.8. From the in-

Table S4.4 Data for Example S4.5

Pattern Recognition
Nonsmoker 9, 8, 12, 10, 7, 10, 9, 11, 8, 10, 8, 10, 8, 11, 10
b3 Hours 12, 7, 14, 4, 8, 11, 16, 17, 5, 6, 9, 6, 6, 7, 16
Active Smoker 8, 8, 9, 1, 9, 7, 16, 19, 1, 1, 22, 12, 18, 8, 10
Cognitive Task
Nonsmoker 27, 34, 19, 20, 56, 35, 23, 37, 4, 30, 4, 42, 34, 19, 49
b3 Hours 48, 29, 34, 6, 18, 63, 9, 54, 28, 71, 60, 54, 51, 25, 49
Active Smoker 34, 65, 55, 33, 42, 54, 21, 44, 61, 38, 75, 61, 51, 32, 47
Driving Simulation
Nonsmoker 3, 2, 0, 0, 6, 2, 0, 6, 4, 1, 0, 0, 6, 2, 3
b3 Hours 7, 0, 6, 0, 12, 17, 1, 11, 4, 4, 3, 5, 16, 5, 11
Active Smoker 15, 2, 2, 14, 5, 0, 16, 14, 9, 17, 15, 9, 3, 15, 13
S4-W3527 9/28/05 4:05 PM Page S4-19

Two-Way Analysis of Variance S4-19

Interaction plotData means for errors made


Mean Number of Errors Made
Nonsmoker >3hrs During All
40
Cognitive
Pattern Rec 9.400 9.600 9.933 9.644
Mean number of errors

Cognitive 28.867 39.933 47.533 38.778


30 Driving 2.333 6.800 9.933 6.356
All 13.533 18.778 22.467 18.259

20 Analysis of Variance for Errors Made, Using Adjusted SS for Tests


Source DF Seq SS Adj SS Adj MS F P
Pattern recognition
10 Task 2 28661.5 28661.5 14330.8 132.90 0.000
SmokeGrp 2 1813.7 1813.7 906.9 8.41 0.000
Driving Task*SmokeGrp 4 1269.5 1269.5 17.4 2.94 0.023
0 Error 126 13587.2 13587.2 107.8
Total 134 45331.9
Never >3 Hours ago During task
Smoking behavior
Figure S4.8 Means and analysis of variance results for
Figure S4.7 Interaction plot for Example S4.5 Example S4.5

teraction plot, the most obvious result is that there are many more errors in the
cognitive task than in the other two tasks. Is there an interaction effect? It ap-
pears that the mean number of errors increases as smoking behavior gets more
immediate for the cognitive task and the driving task, but does change much for
the pattern recognition task. If that pattern is statistically signicant, it is inter-
esting because it indicates that smoking is related to performance in memory
(the cognitive task) and in simulated driving, but not in pattern recognition.
The analysis of variance results in Figure S4.8 conrm that there is indeed a
signicant interaction (p-value  .023) and that there are signicant Factor A
and Factor B effects. Because of the signicant interaction effect, interpretation
of the differences in factor level means should be done within the levels of the
other factor. Notice that the means shown in Figure S4.8 conrm that for the
sample, the mean number of errors for the pattern recognition task remained
relatively constant across smoking behaviors. The mean number of errors for
the other two tasks was highest for active smokers and lowest for nonsmokers,
with those who had not smoked recently falling between them.

Cautions About Nonsignificant Results


We end this discussion with a reminder that applies to all hypothesis-testing
situations. Remember that a nonsignicant result does not mean that the null
hypothesis is true. It simply means that the sample data didnt provide enough
evidence for us to conclude that there is an effect in the population.

E x a m p l e S 4 . 1 (cont.) Seat Location, Gender, and Hours of Sleep We began our discussion of
two-way ANOVA with an example for which the interaction plot indicated that
there would be a strong interaction as well as main effects. Figure S4.1 showed
the mean hours of sleep for male and female students who prefer to sit in the
front, middle, and back of the class. The ANOVA results for this example are pre-
sented in Figure S4.9. Notice that none of the effects is statistically signicant,
as reected by the p-values, which are all much greater than .05. Why do the
ANOVA results appear to be contradictory to the interaction plot? The answer is
S4-W3527 9/28/05 4:05 PM Page S4-20

S4-20 Supplemental Topic 4

Means and Standard Deviations


Front Middle Back All

Female 7.1731 7.1085 6.9667 7.1037


S.D. 1.7085 1.6623 2.0219 1.7174

Male 6.1333 6.7250 7.1818 6.7403


S.D. 1.2743 1.9080 1.4601 1.6987

All 6.7927 6.9435 7.0946 6.9401


S.D. 1.6278 1.7724 1.6867 1.7136

Analysis of Variance for Sleep, Using Adjusted SS for Tests


Source DF Seq SS Adj SS Adj MS F P
Sex 1 5.592 5.589 5.589 1.91 0.169
Seat 2 2.959 3.410 1.705 0.58 0.560
Sex*Seat 2 7.273 7.273 3.637 1.24 0.292
Error 165 483.374 483.374 2.930
Total 170 499.198

Figure S4.9 Means, standard deviations, and analysis of variance


for sleep by gender and seat location

that there is a large amount of variability in the data. The interaction plot does
not take the standard deviations into account; it is a plot of the means.
If you look at the standard deviations in Figure S4.9, you will see that they
range from about 1.3 hours to 2.0 hours of sleep. The natural variability in sleep
times, reected in the denominator of the F-test through MSE, is so large that
any difference that might exist in the population means cannot be detected
with this sample.
The lesson is that it is not enough to look at either an interaction plot or
an analysis of variance table alone. The combination provides a picture of the
main effects and interactions in the sample, as well as a gauge for whether we
can conclude that they exist in the population from which the sample was
drawn.

Key Terms

Section S4.1 population mean for Factor A, level i, S4-5 Section S4.2
main effect, S4-3, S4-7 population mean for Factor B, level j, S4-5 balanced ANOVA, S4-9
interaction effect, S4-3, S4-7 overall population mean, S4-5 balanced design, S4-9
one-way ANOVA model, S4-4 general two-way ANOVA model, S4-6 unbalanced design, S4-9
factor, S4-5 additive model for two-way ANOVA, two-way analysis of variance table, S4-13
factor A, S4-5 S4-7, S4-8
factor B, S4-5 full model for two-way ANOVA, S4-7
S4-W3527 9/28/05 4:05 PM Page S4-21

Two-Way Analysis of Variance S4-21

Exercises
Denotes basic skills exercises S4.4 Draw an interaction plot of these means. Comment
Denotes dataset is available in StatisticsNow at http:// on whether there appear to be effects for Factor A
1pass.thomson.com or on your CD but is not required to (age) and Factor B (coffee-drinking behavior), and
solve the exercise. whether there appears to be an interaction be-
Bold-numbered exercises have answers in the back of the text and
tween them.
fully worked solutions in the Student Solutions Manual.
S4.5 Give numerical values for each of the following:
a. m12
b. m. .
Go to the StatisticsNow website at c. mi . for i  1, 2
http://1pass.thomson.com to: d. m. j for j  1, 2
Assess your understanding of this chapter S4.6 Give numerical values for each of the following:
Check your readiness for an exam by taking the Pre-Test quiz and a. ai for i  1, 2
exploring the resources in the Personalized Learning Plan b. bj for j  1, 2
c. abij for all i, j pairs
S4.1 For each of the following two-way ANOVA situations, S4.7 On the basis of relevant numerical values found in
specify the response variable, the two factors and val- Exercises S4.5 and S4.6, write each of the means mij
ues for a and b, and the number of levels for Factors A using the additive model. Is the additive model ade-
and B, respectively. quate in this situation?
a. One hundred overweight men are classied by S4.8 Write each of the means mij using the full model,
whether they smoke or not. They are randomly as- showing the numerical value of each of the terms in
signed to participate in a swimming program, a jog- the model.
ging program, or a yoga program. Weight loss after S4.9 Explain in words what the table of means indicates in
two months is measured. terms of main effects and interaction in this situation.
b. A random sample of kindergarten children in a Use the following scenario for Exercises S4.10 to S4.14.
certain state is given a reading test. The children A study was done to investigate the effect of the pes-
are categorized by whether they attended preschool ticide DDT in falcons. The response variable was the
(regularly, some, not at all) and whether they have amount of DDT measured in each bird. Factor A was
older siblings at home (yes, no). the nesting site of the bird (1  United States, 2 
S4.2 Researchers measure the salaries and years of experi- Canada, 3  Arctic region) and Factor B was the age
ence for a random sample of male and female profes- of the bird (1  Young, 2  Middle-aged, 3  Old).
sors at a university. They want to know whether there The data, means, and analysis of variance table are
is a difference in the mean salaries of males and fe- shown on the next page; the interaction plot is shown
males, adjusted for years of experience. Is this a situa- in the accompanying gure. (Source: Falcon.MTW
tion for which two-way ANOVA is appropriate? If so, dataset in the Student1 Directory of Minitab les.)
specify the response variable and the two factors. If S4.10 Based on the interaction plot, comment on whether
not, explain why not. there appears to be
S4.3 A plant biologist randomly assigns seedlings to one a. A Factor A (nesting site) main effect.
of four conditions: sunlight and drip watering, arti- b. A Factor B (age) main effect.
cial light and drip watering, sunlight and hand water- c. An interaction between the factors.
ing, articial light and hand watering. He measures the S4.11 State the null and alternative hypotheses in the con-
biomass of each plant after three weeks. Is this a sit- text of this problem for
uation for which two-way ANOVA is appropriate? If a. A Factor A effect.
so, specify the response variable and the two factors. If b. A Factor B effect.
not, explain why not. c. An interaction between the factors.
Use the following scenario for Exercises S4.4 to S4.9. S4.12 Conduct and interpret the hypothesis test for this sit-
Suppose that the population mean hours of sleep uation for
for college students, categorized by whether they are a. A Factor A effect.
coffee drinkers and whether they are over 21, are as b. A Factor B effect.
follows: c. An interaction between the factors.
S4.13 Based on examining the sample means and the anal-
ysis of variance table, explain in words
Coffee Drinker?
a. The effect of nesting site on DDT in falcons.
Yes No b. The effect of age on DDT in falcons.
c. In this example, does the interpretation of the ef-
18 21 6.0 8.0
fects of one factor depend on the levels of the other
Over 21 7.0 7.4
factor? Explain why or why not.

Basic skills Dataset available but not required Bold-numbered exercises answered in the back
S4-W3527 9/28/05 4:05 PM Page S4-22

S4-22 Supplemental Topic 4

Data
92
Rows: Site Columns: Age
82 Arctic
1 2 3 72
1 19 25 34
62

Mean DDT
17 24 37
16 29 36 52
2 16 28 36 42
13 24 39 32
16 26 37 U.S.
22
3 74 82 91 Canada
69 80 88 12
72 80 90
Young Middle Old
Means Age

Rows: Site Columns: Age

1 2 3
1 17.333 26.000 35.667
2 15.000 26.000 37.333
3 71.667 80.667 89.667

Two-Way ANOVA
Source DF SS MS F P
Site 2 17785.41 8892.70 2581.75 0.000
Age 2 1721.19 860.59 249.85 0.000
Interaction 4 17.70 4.43 1.28 0.313
Error 18 62.00 3.44
Total 26 19586.30

Output and Figure for Exercises S4.10 to S4.14

S4.14 Show numerically how the F-test statistic was created S4.16 Construct an interaction plot for this situation.
for the A, B, and AB interaction tests by plugging in S4.17 State the null and alternative hypotheses in the con-
the appropriate mean squares. text of this problem for
Use the following scenario for Exercises S4.15 to S4.20. a. A Factor A effect.
b. A Factor B effect.
Jennifer lives in the San Francisco Bay Area and c. An interaction between the factors.
drives to work. She has a choice of three routes (Fac- S4.18 Conduct and interpret the hypothesis test for this sit-
tor A) and wants to know whether any of them is uation for
faster than the others. She suspects that it may de- a. A Factor A effect.
pend on whether its Monday, Friday, or a mid-week b. A Factor B effect.
day (Factor B), due to vacationers leaving the area on c. An interaction between the factors.
Fridays and returning on Mondays. She randomly S4.19 Based on examining the sample means and the anal-
chooses which route to take each day until she has ysis of variance table, explain in words
taken each route 10 times for each of Monday, mid- a. Whether there is a signicant difference in com-
week day, and Friday, for a total of 90 measurements mute time for different routes.
(3 routes  3 types of day  10 observations). The b. Whether there is a signicant difference in com-
response variable is commute time. The means are mute times for different days.
shown in the following table. c. In this example, does the interpretation of the ef-
fects of one factor depend on the levels of the other
Monday Mid-week Friday factor? Explain why or why not.
Route 1 34.2 30.8 32.1 S4.20 What route would you recommend that Jennifer take
Route 2 22.7 24.6 26.0 to work? Does it depend on the day of the week?
Route 3 38.6 34.1 32.9 Explain.
S4.21 Explain why an interaction plot may show lines that
are not parallel, or even that cross, yet the analysis
She computes these sums of squares: SSA  1868.9, of variance results show a nonsignicant interaction
SSB  65, SSAB  229.33, SSE  552.8. effect.
S4.15 Construct the analysis of variance table for this
problem.

Basic skills Dataset available but not required Bold-numbered exercises answered in the back
S4-W3527 9/28/05 4:05 PM Page S4-23

Two-Way Analysis of Variance S4-23

S4.22 Fill in the missing values in the following two-way c. The sample size for each group, n.
analysis of variance table, where the letters occupy- d. The overall sample size, N.
ing the space indicate the parts to this exercise. For S4.25 Refer to Exercises S4.22, S4.23, and S4.24. If a one-
instance, for part (a), give the degrees of freedom for way analysis of variance had been done using the
Factor B, and for part (b), give MSA. ab groups, instead of a two-way analysis of variance,
give a numerical value for each of the following:
Analysis of Variance a. SS Groups.
b. SS Error.
Source DF SS MS F c. Degrees of freedom for groups.
Factor A 2 450 b f d. Degrees of freedom for error.
Factor B a 840 c g
Interaction 2 130 d h
Error 48 1050 e Preparing for an exam? Assess your
Total 53 2470 progress by taking the post-test at http://1pass.thomson.com.

S4.23 Refer to Exercise S4.22. Carry out the hypothesis test


for each of the following: Do you need a live tutor for homework problems? Access
a. Factor A. vMentor at http://1pass.thomson.com for one-on-one tutoring from
b. Factor B. a statistics expert.
c. AB interaction.
S4.24 Refer to Exercises S4.22 and S4.23. The results came
from a balanced design. Based on the analysis of vari-
ance table, give a numerical value for each of the
following:
a. The number of levels of Factor A.
b. The number of levels of Factor B.

Basic skills Dataset available but not required Bold-numbered exercises answered in the back

You might also like