Chapter

13
case study

Introduction
Section 13.1 One-Way Analysis of Variance

starnes13_002-031hr.indd 2

4/11/11 2:16 PM

Analysis of
Variance
Do Pets or Friends Help Reduce Stress?
If you are a dog lover, perhaps having your dog along reduces the effect of stress. To examine how pets
affect people in stressful situations, researchers recruited 45 women who said they were dog lovers.
Fifteen of the subjects were randomly assigned to each of three groups: to do a stressful task alone
(control, C), with a good friend present (F), or with their dog present (pet, P). The stressful task was to
count backward by 13s or 17s. The subjects mean heart rate (in beats per minute) during the task is one
measure of the effect of stress. The table below shows the data.1
Group

Rate

Group

Rate

Group

Rate

Group

Rate

69.169

68.862

84.738

75.477

99.692

87.231

84.877

62.646

70.169

64.169

58.692

70.077

80.369

91.754

79.662

88.015

87.446

87.785

69.231

81.600

75.985

91.354

73.277

86.985

83.400

100.877

84.523

92.492

102.154

77.800

70.877

72.262

86.446

97.538

89.815

65.446

80.277

85.000

98.200

90.015

101.062

76.908

99.046

97.046

69.538

Here are some numerical summaries for the three groups:

Level

Mean

StDev

Control

15

82.524

9.242

Friend

15

91.325

8.341

Pet

15

73.483

9.970

Do the mean heart rates for the groups provide

convincing evidence that the presence of a pet or friend
affects heart rate during a stressful task? In this chapter,
youll develop the tools to help answer this question.

starnes13_002-031hr.indd 3

4/11/11 2:16 PM

4 CHAPTER

13

Analysis of Variance

Introduction

Which of three exercise routines helps overweight people lose the most weight? Who
spends more time on the Internetfreshmen, sophomores, juniors, or seniors? Which
of six brands of AAA batteries lasts longest? In each of these settings, we wish to compare the mean response for several populations or treatments. The statistical technique for comparing several means is called analysis of variance, or simply ANOVA.
This brief chapter examines the big ideas and the computational details of ANOVA.

Activity
Come fly with me
MATERIALS: Three paper airplane pattern sheets (in the Teachers Resource Binder), scissors, masking
tape, tape measures, graphing calculator

The purpose of this Activity is to see which of three paper airplane models
A, B, or Cflies farthest. Specifically, the object is to determine whether there is
a significant difference in the average distance flown for the three plane models.
The null hypothesis is that there is no difference among the mean distances flown
by Models A, B, and C:

H0:mA = mB = mC
The alternative hypothesis is that there is a difference in the average flight distances.
1. a class, design an experiment to determine which of the three paper airAs
plane models flies the farthest. Be sure to follow the principles of experimental
design that you learned in Chapter 4.
2.
Carry out your plan and collect the necessary data.
3.
Compare the flight distances for the three models graphically and numerically.
Does it appear that all the means are about the same, or is at least one mean different from the other means?
Note: Keep these data handy for further analysis later in the chapter.

13.1
In Section 13.1,
youll learn about:
Comparing means: the
analysis of variance F test
The idea of analysis of
variance
Conditions for ANOVA
F distributions and
degrees of freedom

One-Way Analysis
of Variance
The two-sample t procedures of Chapter 10 compare the means of two populations or the mean responses to two treatments in an experiment. Of course, studies
dont always compare just two groups. In Chapter 11, we saw that a chi-square test
for homogeneity could be used to compare the proportion of successes among
several populations or treatments. We need a method for comparing any number
of means.
Heres an example that introduces the basic idea of a significance test involving
more than two means. As always, we begin with data analysis before proceeding
to inference.

ANOVA calculations
Putting it all together:
one-way ANOVA

starnes13_002-031hr.indd 4

4/11/11 2:16 PM

13.1 One-Way Analysis of Variance 5

Example Hummingbirds and Tropical Flowers

Comparing several means
Ethan Temeles and W. John Kress of Amherst College studied the relationship
between varieties of the tropical flower Heliconia on the island of Dominica and
the different species of hummingbirds that pollinate the flowers.2 Over time, the
researchers believe, the lengths of the flowers and the forms of the hummingbirds
beaks have evolved to match each other. If that is true, flower varieties pollinated
by different hummingbird species should have distinct distributions of length.
The table below gives length measurements (in millimeters) for random samples
of three varieties of Heliconia, each pollinated by a different species of hummingbird. Do the three varieties display distinct distributions of length? In particular,
do the mean lengths of their flowers differ?
H. bihai
47.12

46.75

46.81

47.12

46.67

47.43

46.44

46.64

48.07

48.34

48.15

50.26

50.12

46.34

46.94

48.36

H. caribaea red
41.90

42.01

41.93

43.09

41.47

41.69

39.78

40.57

39.63

42.18

40.66

37.87

39.16

37.40

38.20

38.07

38.10

37.97

38.79

38.23

38.87

37.78

38.01

H. caribaea yellow
36.78

37.02

36.52

36.11

36.03

35.45

38.13

35.17

36.82

36.66

35.68

36.03

34.57

37.10

34.63

Lets follow the strategy we learned way back in Chapter 1: use graphs and numerical summaries to compare the three distributions of flower length. Figure 13.1 is
a side-by-side stemplot of the data. The lengths have been rounded to the nearest
tenth of a millimeter.
bihai
34
35
36
37
38
39
40
41
42
43
44
45
46 3 4 6 7 8 8 9
47 1 1 4
48 1 2 3 4
49
50 1 3

red
34
35
36
37 4 8
38 0 0
39 2 6
40 6 7
41 57
42 0 2
43 1
44
45
46
47
48
49
50

9
1 12289
8
99

yellow
34 6 6
35 2 5 7
36 0 0 1 5 7 8 8
37 0 1
38 1
39
40
41
42
43
44
45
46
47
48
49
50

FIGURE 13.1
Side-by-side stemplots comparing the lengths in
millimeters of random samples of flowers from three varieties of
Heliconia.

starnes13_002-031hr.indd 5

Here are the summary statistics we will use in further analysis:

Sample

Variety

Sample
size

Mean
length

Standard
deviation

bihai

16

47.60

1.213

Red

23

39.71

1.799

Yellow

15

36.18

0.975

What do we see? The three varieties differ so much

in flower length that there is little overlap among
them. In particular, the flowers of bihai are longer
than either red or yellow. The mean lengths are
47.6 mm for H. bihai, 39.7 mm for H. caribaea red,
and 36.2 mm for H. caribaea yellow. Are these
observed differences in sample means statistically
significant? We must develop a test for comparing
more than two population means.

4/11/11 2:16 PM

6 CHAPTER

13

Analysis of Variance

Comparing Means: The Analysis

of Variance F Test
Call the mean lengths for the three populations of flowers m1 for bihai, m2 for red,
and m3 for yellow. To compare these three population means, we might use the
two-sample t test several times:
Test H0:m1 = m2 to see if the mean length for bihai differs from the mean for red.
Test H0:m1 = m3 to see if bihai differs from yellow.
Test H0:m2 = m3 to see if red differs from yellow.
The weakness of doing three tests is that we get three P-values, one for each test
alone. That doesnt tell us how likely it is that three sample means are spread apart
as far as these are. This is the problem of multiple comparisons that we discussed
in Chapter 11.
We want to test the null hypothesis that there are no differences among the mean
lengths for the three populations of flowers:

Be careful when defining

parameters. For random
sampling, each mi refers to
a corresponding population
mean. In a completely randomized experiment, mi refers to the
actual mean response under
treatment i for subjects like the
ones in the study.
Analysis of variance
F test

H0:m1 = m2 = m3
The alternative hypothesis is that there is some difference. That is, not all three
population means are equal:
Ha: not all of m1, m2, and m3 are equal
As with the chi-square test for homogeneity, the alternative hypothesis is no longer
one-sided or two-sided. It is many-sided because it allows any relationship other
than all three equal. For example, Ha includes the case in which m2 = m3 but m1
has a different value.
When the conditions for inference are met, the appropriate significance test for
comparing means is the analysis of variance F test. Analysis of variance is usually
abbreviated as ANOVA.

The Idea of Analysis of Variance

The details of analysis of variance are a bit complicated, but the main idea of
ANOVA is more straightforward and much more important to understand. Here
it is: when we ask whether a set of sample means gives evidence for differences
among the population means, what matters is not how far apart the sample means
are but how far apart they are relative to the variability of individual observations.
Look at the two sets of boxplots in Figure 13.2. For simplicity, these distributions
are all symmetric, so that the mean and median are the same. The center line in
each boxplot is therefore the sample mean. Both sets of boxplots compare three
samples. The sample means are the same in parts (a) and (b). Could differences
this large easily arise due just to chance, or are they statistically significant?
The boxplots in Figure 13.2(a) have long boxes, which means that there is lots
of variation among the individuals in each group. With this much variation among
individuals, we would not be surprised if another set of samples gave quite
different sample means. The observed differences among the sample means
could easily happen just by chance if the three population means are equal.
The boxplots in Figure 13.2(b) have the same centers as those in Figure 13.2(a),
but the boxes are much shorter. That is, there is much less variation among the
individuals in each group. It is unlikely that any sample from the first group

starnes13_002-031hr.indd 6

4/11/11 2:16 PM

13.1 One-Way Analysis of Variance 7

would have a mean as small as the mean of
the second group. Because means as far apart
as those observed would rarely arise just by
chance in repeated sampling, they are good
evidence of real differences among the means
of the three populations we are sampling
from.
This comparison of the two parts of
Figure13.2 is too simple in one way. It ignores
FIGURE 13.2
Boxplots for two sets of three samples each. The sample
the effect of the sample sizes, an effect that
means are the same in parts (a) and (b). Analysis of variance will find a
boxplots do not show. Small differmore significant difference among the means in part (b) because there is
ences among sample means can be
Starnes/Yates/Moore: The individuals Statistics, 4E samples.
less variation among the Practice of within those
significant if the samples are large.
New Fig.: 13-02
Perm. Fig.: 13005
Large differences among sample means can fail
First Pass: 2011-02-24
to be significant if the samples are small. All we
can be sure of is that for the same sample size, Figure 13.2(b) will give a much
smaller P-value than Figure 13.2(a). Despite this qualification, the big idea
remains: if sample means are far apart relative to the variation among individuals
in the same groups, thats evidence that something other than chance is at work.
(a)

(b)

Definition:The Analysis of Variance Idea

Analysis of variance (ANOVA) tests whether several populations or treatments have
the same mean by comparing how far apart the sample means are with how much
variation there is within the samples.

One-way ANOVA
Analysis of variance
F statistic

In Chapter 10, we saw that the

two-sample t statistic usually
x x2
has the form t = 1
.
2
2
s1 s 2
+
n1 n 2
The numerator measures variation among the sample means,
and the denominator measures
variation among individuals
within the same sample. Not
surprisingly, the ANOVA F statistic for comparing the means
of several groups or populations has a similar form.

starnes13_002-031hr.indd 7

It is one of the oddities of statistical language that methods for comparing means
are named after the variance. The reason is that the test works by comparing two
kinds of variation. Analysis of variance is a general method for studying sources of
variation in responses. Comparing several means is the simplest form of ANOVA,
called one-way ANOVA.
The analysis of variance F statistic for testing the equality of several means has
this form:
F=

variation among the sample means

variation among individuals within the same sample
n

Well give more details about how to calculate this test statistic shortly. For now,
note that the F statistic can take only values that are zero or positive. It is zero only
when all the sample means are identical and gets larger as they move farther apart.
Large values of F are evidence against the null hypothesis H0 that all population
or treatment means are the same. Although the alternative hypothesis Ha is manysided, the ANOVA F test is one-sided because any violation of H0 tends to produce
a large value of F.

Conditions for ANOVA

Like all inference procedures, ANOVA is valid only in some circumstances. Here
are the conditions under which we can use ANOVA to compare population or
treatment means.

4/11/11 2:16 PM

8 CHAPTER

13

Analysis of Variance

Conditions for ANOVA Inference

Random The data are produced by random samples of size n1 from
Population 1, size n2 from Population 2, . . . , and size nk from Population k or
by groups of size n1, n2, . . . , nk in a randomized experiment.
Normal All the population distributions (or the true distributions of
r
esponses to the treatments) are Normal, or the sample sizes are large and the
distributions show no strong skewness or outliers.
Independent Both the samples/groups themselves and the individual
observations in each sample/group are independent. When sampling without
replacement, check that all the populations are at least 10 times as large as
the corresponding samples (the 10% condition).
Same SD The response variable has the same standard deviation s (with a
value that is generally unknown) for all the populations or treatments.

Robust

The first three conditions should be somewhat familiar from our study of the
two-sample t procedures for comparing two means. As usual, the design of the
data production is the most important condition for inference. Biased sampling or
confounding can make any inference meaningless. If we do not actually
draw separate random samples from each population or carry out a randomized comparative experiment, our scope of inference is very limited.
Because no real population has an exactly Normal distribution, the usefulness
of inference procedures that assume Normality depends on how sensitive they are
to departures from Normality. As long as there are no outliers or strong skewness,
the ANOVA F test is fairly robust against non-Normality. This is especially true
when the sample sizes are equal. In this setting, ANOVA becomes safer as the
sample sizes get larger.
The fourth condition is annoying: ANOVA assumes that the variability of observations, measured by the standard deviation, is the same for all populations or
treatments. The t test for comparing two means (Chapter 10) does not require
equal standard deviations. Unfortunately, the ANOVA F test for comparing more
than two means is less broadly valid. It is not easy to check the condition that the
populations (or treatments) have equal standard deviations. You must either seek
expert advice or rely on the robustness of ANOVA.
How serious are unequal standard deviations? ANOVA is not too sensitive to
violations of the condition, especially when all samples have the same or similar
sizes and no sample is very small. When designing a study, try to use samples or
groups of about the same size. The sample standard deviations estimate the population standard deviations, so check before doing ANOVA that the sample standard deviations are similar to each other. We expect some variation among them
due to chance. Here is a rule of thumb that is safe in almost all situations.

Checking the Same SD Condition in ANOVA

The results of the ANOVA F test are approximately correct when the largest sample standard deviation is no more than twice as large as the smallest
sample standard deviation.

starnes13_002-031hr.indd 8

4/11/11 2:16 PM

13.1 One-Way Analysis of Variance 9

Now were ready to check conditions for performing an ANOVA F test using the
Heliconia data.

EXAMPLE Hummingbirds and Tropical Tlowers

Checking conditions
Review the details of the Heliconia study from the previous example
(page 5).
Problem: Check conditions for carrying out an ANOVA F test in this setting.
Solution:
RandomResearchers took separate random samples of 16 bihai, 23 red, and 15 yellow Heliconia
flowers.
Normal We entered the data into a TI-84 and made side-by-side boxplots of the flower lengths
in the three samples. Although the distributions for the bihai and red varieties show some
right-skewness, we dont see any strong skewness or outliers that would prevent use of
one-way ANOVA.

Bihai
Red
Yellow

IndependentResearchers took independent samples of bihai, red, and yellow Heliconia.

Because sampling without replacement was used, there must be at least 10(16) =
160 bihai, 10(23) = 230 red, and 10(15) = 150 yellow flowers. This is pretty safe
to assume.
Starnes/Yates/Moore: The Practice of Statistics, 4E
New Fig.: 13UN01
Perm. Fig.: 13006
Same SDThe sample standard deviations for the three varieties are
First Pass: 2011-02-24
s1 = 1.213 s2 = 1.799 s3 = 0.975
These standard deviations satisfy our rule of thumb:
largest s
1.799
=
= 1.85 (less than 2)
smallest s 0.975
We can safely use ANOVA to compare the mean lengths for the three populations.

For Practice Try Exercise 1

starnes13_002-031hr.indd 9

4/11/11 2:16 PM

10 CHAPTER

13

Analysis of Variance

Check Your Understanding

Bromeliads are tropical flowering plants. Many are epiphytes that attach to trees and
o
btain moisture and nutrients from air and rain. Their leaf bases form cups that collect water and are home to the larvae of many insects. As a preliminary step to a study of changes
in the nutrient cycle, Jacqueline Ngai and Diane Srivastava examined the effects of adding
fertilizer in the form of nitrogen, phosphorus, or both to the cups. They randomly assigned
8 bromeliads growing in Costa Rica to each of four treatment groups, including an unfertilized control group. A monkey destroyed one of the plants in the control group, leaving
7 bromeliads in that group. Here are the numbers of new leaves on each plant over the
7 months following fertilization:3
Nitrogen

Phosphorus

Both

Neither

15

14

14

11

14

14

16

13

15

14

15

16

16

11

14

15

17

13

14

15

18

12

13

11

17

15

17

12

13

15

14

1.
State hypotheses for an ANOVA F test in this setting. Be sure to define your parameters.
2.
Check conditions for carrying out the test in Question 1.

F Distributions and Degrees

of Freedom
Recall that the ANOVA F statistic is
F=

variation among the sample means

variation among individuals within the same sample
n

To find the P-value for this statistic, we must know the sampling distribution of F
when the null hypothesis (all population or treatment means are equal) is true and
F distribution the conditions are satisfied. This sampling distribution is an F distribution.
The F distributions are a family of right-skewed distributions that take only values greater than 0. The density
curves in Figure 13.3 illustrate their shapes. A specific F
F(9, 10)
F(2, 51)
distribution is determined by the degrees of freedom of
density curve
density curve
thenumerator and denominator of the F statistic. When
describing an F distribution, always give the numerator
degrees of freedom first. Our brief notation will be F (df1,
df2) for the F distribution with df1 degrees of freedom
in the numerator and df2 degrees of freedom in the
0
3.02
0
3.18
denominator. Interchanging the degrees of freedom changes the distribution, so the order is
FIGURE 13.3
Density curves for two F distributions. Both
important.
are right-skewed and take only positive values. The upper
5% critical values are marked under the curves.

Starnes/Yates/Moore: The Practice of Statistics, 4E

New Fig.: 13-03
Perm. Fig.: 13008
First Pass: 2011-02-24

starnes13_002-031hr.indd 10

4/11/11 2:16 PM

13.1 One-Way Analysis of Variance 11

The degrees of freedom of the ANOVA F statistic depend on the number of
means we are comparing and the number of observations in each sample. Here
are the details.

Definition: Degrees of Freedom for the ANOVA F Test

Suppose that we want to compare the means of k populations or treatments. The total
number of observations in all samples or groups combined is
N = n1 + n2 + . . . + nk
If the null hypothesis that all the population or treatment means are equal is true and
the conditions for inference are met, the ANOVA F statistic has the F distribution with
k - 1 degrees of freedom in the numerator and N - k degrees of freedom in the
denominator.
Lets find the degrees of freedom for the ANOVA F test in the Heliconia study.

EXAMPLE

Hummingbirds and Flower Lengths

Degrees of freedom for F
In the two earlier examples, we compared the mean lengths for three varieties of
flowers, so k = 3. The three sample sizes are n1 = 16, n2 = 23, and n3 = 15. The
total number of observations is therefore N = 16 + 23 + 15 = 54. The ANOVA
F test has numerator degrees of freedom
k-1=3-1=2
and denominator degrees of freedom
N - k = 54 - 3 = 51

Check Your Understanding

Refer to the previous Check Your Understanding. Find the degrees of freedom for the
ANOVA F test in the bromeliad study. Show your work.

ANOVA Calculations
Now we will give the actual formula for the ANOVA F statistic. Suppose the
Random condition is met. Subscripts from 1 to k tell us which sample or group a
statistic refers to:
Population

Sample size

Sample mean

Sample std. dev.

x1

s1

n2

x2

starnes13_002-031hr.indd 11

n1

nk

xk

s2

sk

4/11/11 2:16 PM

12 CHAPTER

13

Analysis of Variance
You can find the F statistic from just the sample sizes ni, the sample means xi, and
the sample standard deviations si. You dont have to go back to the individual
observations.
As weve already seen, the ANOVA F statistic has the form
F=

Mean squares

variation among the sample means

variation among individuals within the same sample
n

The measures of variation in the numerator and denominator of F are called

mean squares. A mean square is a more general form of a sample variance.
2
Recall that an ordinary sample variance sx is a kind of average (or mean) of
the squared deviations of observations from their mean, so it qualifies as a
mean square.
The first step in calculating the F statistic is to determine the overall mean
response x. That is, x is the mean of all N observations together. You can
find x from the k sample means by
x=

sum of all sample observations n1x1 + n2 x2 + + nk xk

=
N
N

This expression works because multiplying a group mean xi by the number

of observations ni it represents gives the sum of the observations in that
group.
The numerator of F is a mean square that measures variation among the k
sample means x1, x2 , , xi . To measure this variation, look at the k deviations of
the means of the samples from x:
x1 - x , x2 - x , , xk - x
The mean square in the numerator of F is an average of the squares of these deviations. Each squared deviation is weighted by ni, the number of observations it
represents. We call this result the mean square for groups (MSG):
MSG =

n1( x1 x )2 + n2 ( x2 x )2 + + nk ( xi x )2
k 1

The mean square in the denominator of F measures variation among individual observations in the same sample. For any one sample, the sample variance
si2 does this job. For all k samples together, we use an average of the individual
sample variances. It is another weighted average, in which each si2 is weighted
by its degrees of freedom ni - 1. The resulting mean square is called the mean
square for error (MSE):
MSE =

2
2
( n1 1)s1 + ( n2 1)s2 + + ( nk 1)si2
N k

Error doesnt mean a mistake has been made. Its a traditional term for chance
variation.

starnes13_002-031hr.indd 12

4/11/11 2:16 PM

13.1 One-Way Analysis of Variance 13

The ANOVA FTest

To test the null hypothesis that all k populations or treatments have the same
mean against the alternative hypothesis that not all the means are equal, calculate the ANOVA F statistic
F=

MSG
MSE

The numerator of F is the mean square for groups

MSG =

n1( x1 x )2 + n2 ( x2 x )2 + + nk ( xi x )2
k 1

The denominator of F is the mean square for error

MSE =

2
2
( n1 1)s1 + ( n2 1)s2 + + ( nk 1)si2
N k

When H0 is true and the Random, Normal, Independent, and Same SD conditions are met, the F statistic has the F distribution with k - 1 and N - k
degrees of freedom.

Sums of squares
ANOVA table

The denominators in the formulas for MSG and MSE are the two degrees of
freedom k - 1 and N - k of the F test. The numerators are called sums of squares,
from their algebraic form. It is usual to present the results of ANOVA in an
ANOVA table. Output from software usually includes an ANOVA table, like the
one in Figure 13.4 for the Heliconia study. You can check that each mean square
MS is the corresponding sum of squares SS divided by its degrees of freedom df,
and that the F statistic is MSG divided by MSE.

FIGURE 13.4
Minitab output from a one-way ANOVA for the flower length data.

starnes13_002-031hr.indd 13

4/11/11 2:16 PM

14 CHAPTER

13

Analysis of Variance
We can use tables of F critical values to get the P-value for an ANOVA F test.
Doing so is awkward, however, because we need a separate table for every pair of
degrees of freedom df1 and df2. Fortunately, software gives you P-values for the
ANOVA F test without the need for a table. That leaves us free to interpret the
results of the test.

EXAMPLE

Hummingbirds and Flower Lengths

Comparing tropical flowers: Conclusion
Figure 13.4 tells us that for the flower length data in the previous examples, the
test statistic is F = 259.12 with P-value P < 0.0001.

The Minitab ANOVA output in

Figure 13.4 reports the
P-value as 0.000 to three
decimal places. This does
not mean that the P-value is
actually 0, just that it is very
close to 0.

Problem:
(a)Interpret the P-value in context.
(b)What conclusion would you draw? Justify your answer.
Solution:
(a)If the null hypothesis H0 : m1 = m2 = m3 of no difference in the population mean flower lengths
is true, the probability of getting a difference among the sample mean flower lengths as large as or
larger than the one observed in the study just by the chance involved in the random sampling is less
than 1 in 10,000.
(b)Because the P-value is so small (less than any of our usual a levels), we would reject H0.
There is very strong evidence that the three varieties of flowers do not all have the same population
mean length.
For Practice Try Exercise 13

The F test doesnt say which of the three means are significantly different. It
appears from our earlier data analysis that bihai flowers are distinctly longer than
red or yellow. Red and yellow are closer in length, but the red flowers tend to be
longer. The Minitab output in Figure 13.4 gives confidence intervals for all three
means that help us see which means differ and by how much. None of the intervals overlap, and bihai is far above the other two. Note: These are 95%
confidence intervals for each mean separately. We are not 95% confident
that all three intervals cover the three means. This is another example of
the peril of multiple comparisons.

Check Your Understanding

The figure at the top of page 15 shows Minitab ANOVA output for the bromeliad study
described in the Check Your Understanding on page 10.
1.
Interpret the P-value in context.
2.
What conclusion would you draw?

starnes13_002-031hr.indd 14

4/11/11 2:16 PM

13.1 One-Way Analysis of Variance 15

THINK
ABOUT
IT

How does Minitab get those confidence intervals? Because MSE is an

average of the individual variances, it is also called the pooled sample variance,
2
written as sp . When all k populations have the same population variance s2, as
2
ANOVA assumes they do, sp estimates s2. The square root of MSE is the pooled
standard deviation sp. It estimates the common standard deviation s of each
population or treatment. The Minitab output in Figure 13.4 (page 13) gives the
value sp = 1.446.
The pooled standard deviation sp is a better estimator of the common s than any
of the individual sample standard deviations si because it combines (pools) the
information in all k samples. We can get a confidence interval for any one of the
means mi from the usual formula:
statistic (critical value) (standard deviation of statistic)
The confidence interval for mi is
xi t

sp
ni

Use the critical value t* from the t distribution with N - k degrees of freedom.
For the bihai variety, our 95% confidence interval for the population mean
length is
47.598 2.008

1.446
= 47.598 0.726 = ( 46.872, 48.324)
4
16

using df = 54 - 3 = 51. This is the confidence interval for bihai that appears in
the Minitab ANOVA output of Figure 13.4.

starnes13_002-031hr.indd 15

4/11/11 2:16 PM

16 CHAPTER

13

Analysis of Variance
As the following example shows, it is possible to calculate the F statistic directly
from the formulas. But its pretty unpleasant to do so! We recommend using
technology.

EXAMPLE Hummingbirds and Flower Lengths

ANOVA calculations by hand
We can do the ANOVA test comparing the mean lengths of bihai, red, and yellow
flower varieties using only the sample sizes, sample means, and sample standard
deviations. There are k = 3 groups with a total of N = 54 flowers.
The overall mean of the 54 lengths is
n1x1 + n2 x2 + n3 x3
N
(16)( 47.598) + (23)(39.711) + (15)(36.180)
= 41.067
=
54

x=

The mean square for groups is

n1( x1 x )2 + n2 ( x2 x )2 + n3 ( x3 x )2
k 1
1
[(16)(47.598 41.067)2 + (23)(39.711 41.067)2
4
=
3 1
+ (15)(36.180 41.067)2 ]

MSG =

1082.996
= 541.50
2

The mean square for error is

2
2
2
( n1 1)s1 + ( n2 1)s2 + ( n3 1)s3
N k
(15)(1.2132 ) + (22)(1.7992 ) + (14)(0.9752 ) 106.580
=
= 2.09
=
51
51

MSE =

Finally, the ANOVA F statistic is

F=

MSG 541.50
=
= 259.09
MSE
2.09

Our work differs slightly from the output in Figure 13.4 because of roundoff
error.

You can carry out a one-way ANOVA test on the TI-83/84 and TI-89, as long as
you are given the raw data (TI-83/84) and are comparing no more than six means.
The following Technology Corner shows you how.

starnes13_002-031hr.indd 16

4/11/11 2:16 PM

13.1 One-Way Analysis of Variance 17

Technology CornerOne-way ANOVA on the calculator

We will repeat the one-way ANOVA procedure of the previous examples
u
sing the Heliconia data. Begin by entering the data into lists: bihaiL1/list1,
redL2/list2, and yellowL3/list3.

TI-83/84

TI-89

 ress STAT , arrow to TESTS,

P
and choose ANOVA(. Then
enter the lists that contain
the data. The completed
command is shown below.

In the Statistics/List Editor,

press 2nd F1 ( F6 ), and choose
C:ANOVA. Set Data Input
Method to Data and Number
of Groups to 3.

The results of the one-way ANOVA are shown in the following screens. The
calculator reports that the F statistic is 259.12 and the P-value is 1.92 1027.
The numerator degrees of freedom are k - 1 = 2, and by scrolling down, you
see that the denominator degrees of freedom are N - k = 54 - 3 = 51.

What the calculator calls

Factor is what we refer to
as Group.

If you know the F statistic and the numerator and denominator degrees
offreedom, you can find the P-value with the command Fcdf, under the DISTR menu on the TI-83/84, and in the CATALOG under FlashApps on the TI89. The syntax is Fcdf (leftendpoint,rightendpoint,df numerator,df
denominator). See the following screen shots.

starnes13_002-031hr.indd 17

4/11/11 2:16 PM

18 CHAPTER

13

Analysis of Variance

Putting It All Together:

One-Way ANOVA
You can use the One-Way ANOVA applet to explore the analysis of variance idea
for yourself. The applet allows you to change both the group means and the
spread within groups. You can watch the ANOVA F statistic and its P-value
change as you work.

AP

PLET

Activity One-Way ANOVA applet

The
MATERIALS: Computer with Internet connection

The One-Way ANOVA applet displays the observations in three groups, with the
group means highlighted by black dots. When you open or reset the applet, the
scale at the bottom of the display shows that for these groups the ANOVA F statistic is F = 31.74, with P < 0.001. (The P-value is marked by a red dot that moves
along the scale.)

starnes13_002-031hr.indd 18

4/11/11 2:16 PM

13.1 One-Way Analysis of Variance 19

Part I: ANOVA compares several means
1.
The middle group has a larger mean than the other two. Grab its mean point
with the mouse. How small can you make F? What did you do to the mean to
make F small? Roughly how significant is your small F?
2.
Starting with the three means aligned from your configuration at the end of
step 1, drag any one of the group means either up or down. What happens to F?
What happens to the P-value? Convince yourself that the same thing happens
if you move any one of the means, or if you move one slightly and then another
slightly in the opposite direction.
Part II: ANOVA uses within-group variation
Reset the One-Way ANOVA applet to its original state. As in Figure 13.2(b), on
page 7, the differences among the three means are highly significant (large F,
small P-value) because the observations in each group cluster tightly about the
group mean.

For another look at the big

ideas of ANOVA, try the
Simulating ANOVA Tables
applet at http://statweb.calpoly.
edu/chance/applets/anova/
anova.html.

3.
Use the mouse to slide the Pooled Standard Error at the top of the display
to the right. You see that the group means do not change, but the spread of the
observations in each group increases. What happens to F and P as the spread
among the observations in each group increases? What are the values of F and P
when the slider is all the way to the right? This is similar to Figure 13.2(a): variation within groups hides the differences among the group means.
4.
Leave the Pooled Standard Error slider at the extreme right of its scale, so
that spread within groups stays fixed. Use the mouse to move the group means
apart. What happens to F and P as you do this?

Remember what you learned earlier about significance tests: a fail to reject
H0 conclusion does not mean that H0 is true. All it says is that the data do
not provide convincing evidence against the null hypothesis. In the case
of the ANOVA F test, failing to reject H0 says that there is not enough evidence
to conclude that the population/treatment means differ. As with other kinds of tests,
we sometimes make a Type I or a Type II error as a result of an ANOVA F test.
Here is a final example that shows the entire one-way ANOVA process.
As always, we follow the familiar four-step process.

EXAMPLE Does Thinking about Money Change Behavior?

STEP

starnes13_002-031hr.indd 19

One-way ANOVA
Kathleen Vohs of the University of Minnesota and her coworkers carried out several randomized comparative experiments on the effects of thinking about money.
Heres an outline of one of the experiments. Ask student subjects to unscramble
30 sets of five words to make a meaningful phrase from four of the five. The control group unscrambled phrases like cold it desk outside is into it is cold outside. The play money group unscrambled similar sets of words, but a stack of
Monopoly money was placed nearby. The money prime group unscrambled
phrases that led to thinking about money, turning high a salary desk paying into
a high-paying salary. Then each subject worked a hard puzzle, knowing that
they could ask for help.

4/11/11 2:16 PM

20 CHAPTER

13

Analysis of Variance
The table below shows the time in seconds that each subject worked on the puzzle before asking for help.4 Psychologists think that money tends to make people
self-sufficient. If so, the two groups that were encouraged in different ways to think
about money should take longer on average to ask for help. Do the data support
this idea?
Group

Time

Group

Time

Group

Time

Prime

609

Play

455

Control

118

Prime

444

Play

100

Control

272

Prime

242

Play

238

Control

413

Prime

199

Play

243

Control

291

Prime

174

Play

500

Control

140

Prime

55

Play

570

Control

104

Prime

251

Play

231

Control

55

Prime

466

Play

380

Control

189

Prime

443

Play

222

Control

126

Prime

531

Play

71

Control

400

Prime

135

Play

232

Control

92

Prime

241

Play

219

Control

64

Prime

476

Play

320

Control

88

Prime

482

Play

261

Control

142

Prime

362

Play

290

Control

141

Prime

69

Play

495

Control

373

Prime

160

Play

600

Control

156

Play

67

Minitab output from a one-way ANOVA on these data is shown below.

starnes13_002-031hr.indd 20

4/11/11 2:16 PM

13.1 One-Way Analysis of Variance 21

STATE: We want to perform a test of H0 : m1 = m2 = m3 versus Ha : the means are not all equal,
where mi = the actual mean time until subjects like the ones in this study ask for help; i = 1: money
prime, i = 2: play money, and i = 3: control. Since no significance level was given, well use
a = 0.05.
PLAN:If conditions are met, we should perform an ANOVA F test to compare the
means.
Random Subjects were randomly assigned to the three treatment groups.
Normal Figure 13.5 shows side-by-side stemplots of the data (rounded to the nearest ten
seconds) for the three groups. We expect some irregularity in small samples, but there are no
outliers or strong skewness that would prevent use of ANOVA.
Prime
0
6
1
4
2
0
3
6
4
4
5
3
6
1

7
6 7
4 4 5
4 7 8 8

Play
0
7
1
0
2
2
3
2
4
6
5
0
6
0

7
2 3 3 4 4 6 9
8
0 7

Control
0 6 6 9 9
1 0 2 3 4 4 4 6 9
2 7 9
3 7
4 0 1
5
6

FIGURE 13.5
Side-by-side stemplots comparing the time until subjects asked for help with
a puzzle under each of the three treatments.

IndependentThe random assignment yields independent groups. If the experiment was conducted
properly, individual times to ask for help should also be independent: knowing one subjects time
should give no additional information about another subjects time.
Same SDThe Minitab ANOVA output shows that the group standard deviations easily satisfy our
rule of thumb: the largest (172.8) is no more than twice the smallest (118.1).
DO: From the computer output, we have
Test statistic F = 3.73
P-value 0.031 using the F distribution with degrees of freedom k - 1 = 2 and
N - k = 49.
CONCLUDE: Since the P-value (0.031) is less than a = 0.05, we reject H0. The experiment gives
convincing evidence that there is a difference in the true mean amount of time until people like the
ones in this study ask for help under the three treatments.

For Practice Try Exercise 17

Although the ANOVA F test in the previous example doesnt tell us which
means are different, the data suggest that reminding people of money in either of
two ways does make them less likely to ask others for help. This is consistent with
the idea that money makes people feel more self-sufficient.

starnes13_002-031hr.indd 21

4/11/11 2:16 PM

22 CHAPTER

13

Analysis of Variance

case closed
Do Pets or Friends Help
Reduce Stress?
In the chapter-opening Case Study (page 2), we described details of an
e
xperiment investigating whether the presence of a pet or a friend affects subjects heart rates during a stressful task. Comparing the mean heart rates in
the three groups calls for one-way analysis of variance. Figure 13.6 gives the
Minitab ANOVA output for these data.

FIGURE 13.6
Minitab output for the heart rate data (beats per minute) during stress. The control group
worked alone, the Friend group had a friend present, and the Pet group had a pet dog present.

STEP

starnes13_002-031hr.indd 22

State: We want to perform a test of H0:mC = mF = mP versus Ha:the means

are not all equal, where mC = the actual mean heart rate of subjects like the
ones in this study when performing a stressful task alone, and mF and mP are
the corresponding parameters for performing the task with a friend present and
a pet dog present. Since no significance level was given, well use a = 0.05.
PLAN: If conditions are met, we should perform an ANOVA F test to compare
the means.
Random Subjects were randomly assigned to the three treatment groups.
Normal Side-by-side boxplots of the data are shown below. Unfortunately,
there is one high outlier in the pet group. The ANOVA F test may not give
accurate results in this case. To be safe, well perform the analysis with and
without the outlier.

4/11/11 2:16 PM

13.1 One-Way Analysis of Variance 23

Independent The random assignment yields independent groups. If the
experiment was conducted properly, individual heart rates should also be
independent: knowing one subjects average heart rate should give no additional
information about another subjects.
Same SD The Minitab ANOVA output shows that the group standard
deviations easily satisfy our rule of thumb: the largest (9.97) is no more than
twice the smallest (8.341).
DO: From the Minitab output, we have
Test statistic F = 14.08
P-value Approximately 0, using the F distribution with degrees of freedom 2
and 42.
When we ran the ANOVA F test again without the outlier, we obtained F = 19.34
and a P-value of approximately 0.
CONCLUDE: With a P-value of approximately 0, we would reject H0 at the
a = 0.05 level. The experiment gives strong evidence that the average mean
heart rates of subjects like the ones in this study differ when performing a task
alone, with a friend present, or with a pet present. From the Minitab output, the
stress levels appear to be highest when a friend is present and lowest when a pet
is present.

starnes13_002-031hr.indd 23

4/11/11 2:16 PM

24 CHAPTER

13

Analysis of Variance

Summary

Section 13.1

One-way analysis of variance (ANOVA) compares the means of several

populations or treatments. The ANOVA F test tests the null hypothesis that
all the populations or treatments have the same mean. If the F test shows
significant differences, examine the data to see where the differences lie and
whether they are large enough to be important.
The conditions for ANOVA are
Random The data are produced by random samples of size n1 from
Population 1, size n2 from Population 2, . . . , and size nk from
Population k or by groups of size n1, n2, . . . , nk in a randomized
experiment.
Normal All the population distributions (or the true distributions of
responses to the treatments) are Normal, or the sample sizes are large
and the distributions show no strong skewness or outliers.
Independent Both the samples/groups themselves and the individual
observations in each sample/group are independent. When sampling
without replacement, check that all the populations are at least 10 times
as large as the corresponding samples (the 10% condition).
Same SD The response variable has the same standard deviation s
(with a value that is generally unknown) for all the populations or
treatments.
In practice, ANOVA inference is relatively robust when the populations are
non-Normal, especially when the samples are large. Before doing the F test,
check the observations in each sample for outliers or strong skewness. Also
verify that the largest sample standard deviation is no more than twice as
large as the smallest standard deviation.
When the null hypothesis is true, the ANOVA F statistic for comparing
k means from a total of N observations in all samples combined has the
F distribution with k - 1 and N - k degrees of freedom.
ANOVA calculations are reported in an ANOVA table that gives sums of
squares, mean squares, and degrees of freedom for variation among groups
and for variation within groups, along with the F test statistic and P-value.
In practice, we use software to do the calculations.

13.1 Technology Corner

One-way ANOVA on the calculator......................................... page 17

starnes13_002-031hr.indd 24

4/11/11 2:16 PM

13.1 One-Way Analysis of Variance 25

Exercises

1. Which color attracts beetles best? To detect the

presence of harmful insects in farm fields, we can put
up boards covered with a sticky material and examine
pg 9
the insects trapped on the boards. Which colors
attract insects best? Experimenters placed six boards
of each of four colors at random locations in a field of
oats and measured the number of cereal leaf beetles
trapped. Here are the data:5
Board Color

3. Choose your parents wisely To live long, it helps to

have long-lived parents. One study of this inheritance
of longevity looked at records of families in a small
mountain valley in France where the style of life
changed little over time.7 All married couples in which
both husband and wife were born between 1745 and
1849 were divided into four groups as follows:

Blue

16

11

20

21

14

Green

37

32

20

29

37

32

White

21

12

14

17

13

20

Yellow

45

59

48

46

38

47

a)
( State hypotheses for an ANOVA F test in this
s
etting. Be sure to define your parameters.
b)
( Check conditions for carrying out the test in
part (a).
2. More rain for California? The changing climate
will probably bring more rain to California, but we
dont know whether the additional rain will come
during the winter wet season or extend into the long
dry season in spring and summer. Kenwyn Suttle
of the University of California at Berkeley and his
coworkers randomly assigned plots of open grassland
to three treatments: added water equal to 20% of
annual rainfall either during January to March
(winter) or during April to June (spring), and no
added water (control). Here are some of the data, for
plant biomass (in grams per square meter) produced
by each plot in a single year.6

Long-Lived Wife

Short-lived husband

Insects Trapped

Short-Lived Wife
Group A

Group C

Long-lived husband

Group B

Group D

Long-lived parents were in the top 70% of the

d
istribution of age at death, short-lived parents
were in the bottom 30%. To examine extended life,
the investigators looked at the ages at death of the
children of these parents who lived to at least age 55.
The report includes a bar graph like the one below.
76

Childrens longevity (years)

Section 13.1

74

72

70

68
B

Parents longevity group

Explain why it is not wise to carry out an ANOVA F

test in this setting.
Starnes/Yates/Moore: The Practice of Statistics, 4E

Winter

Spring

Control

264.1514

318.4182

129.0538

187.7312

281.6830

144.6578

291.1431

288.8433

172.7772

176.2879

382.6673

113.2813

141.7525

326.8877

142.1562

169.9737

293.8502

117.9808

a)
( State hypotheses for an ANOVA F test in this
s
etting. Be sure to define your parameters.
b)
( Check conditions for carrying out the test in
part (a).

starnes13_002-031hr.indd 25

4. Marital status and salary Married men tend to

New Fig.: 13UN14
Perm. Fig.: 13025
earn more than single men. An investigation of the
First Pass: 2011-02-24
relationship between marital status and income collected data on all 8235 men employed as managers or
professionals by a large manufacturing firm. Here are
descriptive statistics for the salaries of these men:8
Single

Married

Divorce

Widowed

ni

337

7,730

126

42

xi

$21,384

$26,873

$25,594

$26,936

si

$5,731

$7,159

$6,347

$8,119

Explain why it is not wise to carry out an ANOVA

F test in this setting.

4/11/11 2:16 PM

26 CHAPTER

13

Analysis of Variance

5. Road rage The phenomenon of road rage has been

frequently discussed but infrequently examined.
So begins a report based on interviews with 1382
randomly selected drivers.9 The respondents answers
to interview questions produced scores on an angry/
threatening driving scale with values between 0
and 19. What driver characteristics go with road
rage? There were no significant differences among
races or levels of education. What about the effect
of the drivers age? Here are the means and standard
d
eviations for drivers in three age groups:
Age group

Less than 30 years

244

2.22

3.11

30 to 55 years

734

1.33

2.21

Over 55 years

364

0.66

1.60

a)
( State hypotheses for an ANOVA F test in this
s
etting. Be sure to define your parameters.
b)
( Show that the Random, Independent, and Same
SD conditions are met.
c)
( Data from the study confirms that the Normal
condition is met. Explain what this means.
6. Do fruit flies sleep? Mammals and birds sleep.
Insects such as fruit flies rest, but can this rest be
considered sleep? Biologists now think that insects
do sleep. One experiment gave caffeine to fruit flies
to see if it affected their rest. We know that caffeine
reduces sleep in mammals, so if it reduces rest in fruit
flies, thats another hint that the rest is really sleep.
The paper reporting the study contains a bar graph
similar to the one below.10
600

Rest (min)

400

200

0
Control

1 mg/ml

2.5 mg/ml

5 mg/ml

Caffeine

a)
( The explanatory variable is amount of caffeine,
in milligrams per milliliter (mg/ml) of blood. The
response variable is minutes of Statistics, 4E
Starnes/Yates/Moore: The Practice of rest (measured by an
infrared motion Perm. during a
New Fig.: 13UN15 sensor)Fig.: 13026 12-hour dark period.
First Pass: 2011-02-24

starnes13_002-031hr.indd 26

Outline a completely randomized design using

36 flies per group for this experiment.
b)
( State hypotheses for an ANOVA F test in this
s
etting. Be sure to define your parameters.
c)
( Show that the Random and Independent conditions are met.
d)
( Data from the study confirm that the Normal
and Same SD conditions are met. Explain what this
means in each case.
Exercises 7 to 10 describe situations in which we want to
compare the mean responses in several populations. For
each setting, identify the parameters of interest. Then give k,
the ni, and N. Finally, give the degrees of freedom of the
ANOVA F statistic.
7. Morning or evening? Are you a morning person,
an evening person, or neither? Does this personality
trait affect how well you perform? A random sample
of 100 students took a psychological test that found
16 morning people, 30 evening people, and 54 who
were neither. All the students then took a test of their
ability to memorize at 8:00 a.m. and again at 9:00 p.m.
The response variable is the score at 8:00 a.m. minus
the score at 9:00 p.m.
8. Writing essays Do strategies such as preparing a written outline help students write better essays? College
students were divided at random into four groups of
20 students each, then asked to write an essay on an
assigned topic. Group A (the control group) received
no additional instruction. Group B was required to
prepare a written outline. Group C was given 15 ideas
that might be relevant to the essay topic. Group D
was given the ideas and also required to prepare an
outline. An expert scored the quality of the essays on a
scale of 1 to 7.
9. Test accommodations Many states require schoolchildren to take regular statewide tests to assess
their progress. Children with learning disabilities
who read poorly may not do well on mathematics
tests because they cant read the problems. Most
states allow accommodations for learning-disabled
children. Randomly assign 100 learning-disabled
children in equal numbers to three types of accommodation and a control group: math problems are
read by a teacher; by a computer; by a computer
that also shows a video; and standard test conditions.
Compare the mean scores on the state mathematics
assessment.
10. A medical study The Quebec (Canada) Cardiovascular Study recruited men aged 34 to 64 at random
from towns in the Quebec City metropolitan area. Of
these, 1824 met the criteria (no diabetes, free of heart
disease, and so on) for a study of the relationship

4/11/11 2:16 PM

13.1 One-Way Analysis of Variance 27

between being overweight and medical risks. The
719 normal-weight men had a mean triglyceride
level of 1.5 millimoles per liter (mmol/l) of blood;
the 885 overweight men had mean 1.7 mmol/l; and
the 220 obese men had mean 1.9 mmol/l.11
11. Road rage The report in Exercise 5 says that
F = 34.96, with P < 0.01.
a)
( Interpret the P-value in context.
b)
( What conclusion would you draw? Justify your
answer.
12. Do fruit flies sleep? The paper described in
Exercise 6 also states, Flies given caffeine obtained
less rest during the dark period in a dose-dependent
fashion (n = 36 per group, P < 0.0001). The
P-value in the report comes from the ANOVA F test.
a)
( Interpret the P-value in context.
b)
( What conclusion would you draw? Justify your
answer.
Exercises 13 and 14 refer to the following setting. People
often match their behavior to their social environment.
One study of this idea first established that the type of
music most preferred by black college students is R&B
and that whites most preferred rock music. Will students
hosting a small group of other students choose music that
matches the makeup of the people attending? Assign 90
black business students at random to three equal-sized
groups. Do the same for 96 white students. Each student
sees a picture of the people he or she will host. Group
1 sees 6 blacks, Group 2 sees 3 whites and 3 blacks, and
Group 3 sees 6 whites. Ask how likely the host is to play
the type of music preferred by the other race. Use ANOVA
to compare the three groups to see whether the racial mix
of the gathering affects the choice of music.12
13. What music will you play? For the black subjects,
F = 2.47.
a)
( What are the degrees of freedom?
pg 14
b)
( Find the P-value using technology. Interpret this
result in context.
c)
( What conclusion would you draw? Explain.
14. What music will you play? For the white subjects,
F = 16.48.
a)
( What are the degrees of freedom?
b)
( Find the P-value using technology. Interpret this
result in context.
c)
( What conclusion would you draw? Explain.
15. Attitudes toward math Do high school students
from different racial/ethnic groups have different
attitudes toward mathematics? Measure the level
of interest in mathematics on a five-point scale for
a national random sample of students. Here are

starnes13_002-031hr.indd 27

summaries for students who were taking math at the

time of the survey:13
Racial/ethnic group

sx

African American

809

2.57

1.40

1860

2.32

1.36

Asian/Pacific Islander

White

654

2.63

1.32

HIspanic

883

2.51

1.31

Native American

207

2.51

1.28

a)
( Show that the Random, Independent, and Same
SD conditions for ANOVA are met.
b)
( Data from the study confirm that the Normal
condition is met. Explain what this means.
c)
( Calculate the overall mean response x, the mean
squares MSG and MSE, and the F statistic.
d)
( Which F distribution would you use to find
the P-value of the ANOVA F test? Software gives
P < 0.001. What do you conclude from this
study?
6. Exercise and weight loss What conditions help
1
overweight people exercise regularly? Subjects
were randomly assigned to three treatments: a
single long exercise period 5 days per week; several
10-minute exercise periods 5 days per week; and
several 10-minute periods 5 days per week on a
home treadmill that was provided to the subjects.
The study report contains the following information
about weight loss (in kilograms) after six months of
treatment:14
Treatment

sx

Long exercise periods

37

10.2

4.2

Short exercise periods

36

9.3

4.5

Short periods with equipment

42

10.2

5.2

a)
( Show that the Random, Independent, and Same
SD conditions for ANOVA are met.
b)
( Data from the study confirm that the Normal
condition is met. Explain what this means.
c)
( Calculate the overall mean response x, the mean
squares MSG and MSE, and the F statistic.
d)
( Which F distribution would you use to find
the P-value of the ANOVA F test? Software says that
P = 0.634. What do you conclude from this study?
Exercises 17 and 18 refer to the following setting. How
does logging in a tropical rain forest affect the forest in
later years? Researchers compared forest plots in Borneo
that had never been logged (Group 1) with similar plots
nearby that had been logged 1 year earlier (Group 2)
and 8 years earlier (Group 3). Although the study was

4/11/11 2:16 PM

28 CHAPTER

13

Analysis of Variance

not an experiment, the authors explain why we can

consider the plots to be randomly selected. The data
a
ppear in the table below. The variable Trees is the
count of trees in a plot; Species is the count of tree
species in a plot.15
Never
Logged

Logged
1 Year Ago

Logged
8 Years Ago

Trees

Species

Trees

Species

Trees

Species

27

22

12

11

18

17

22

18

12

11

22

18

29

22

15

14

15

14

21

20

18

18

19

15

20

18

19

15

33

21

18

15

22

15

16

13

17

15

12

10

20

13

14

12

12

12

24

19

14

13

27

13

17

15

28

19

19

19

15

17. Logging in the rain forest

a)
( Make a graph suitable for comparing the number
of trees per plot for the three groups. Write a few
pg 19
sentences describing what you see.
b)
( The figure below shows ANOVA output for these
data. Carry out the ANOVA F test.

19. Can you hear these words? To test whether a hearing aid is right for a patient, audiologists play a tape
on which words are pronounced at low volume. The
patient tries to repeat the words. There are several
different lists of words that are supposed to be equally
difficult. Are the lists equally difficult when there is
background noise? To find out, an experimenter had
subjects with normal hearing listen to four lists with a
noisy background. The response variable was the percent of the 50 words in a list that the subject repeated
correctly. The data set contains 96 responses.16 Here
are two study designs that could produce these data:
Design A: The experimenter assigns 96 subjects to
four groups at random. Each group of 24 subjects
listens to one of the lists. All individuals listen
and respond separately.
Design B: The experimenter has 24 subjects. Each
subject listens to all four lists in random order.
All individuals listen and respond separately.
Does Design A allow use of one-way ANOVA to compare the lists? Does Design B allow use of one-way
ANOVA to compare the lists? Briefly explain your
answers.

18. Logging in the rain forest

a)
( Make a graph suitable for comparing the number
of tree species per plot for the three groups. Write a
few sentences describing what you see.
b)
( The figure below shows ANOVA output for these
data. Carry out the ANOVA F test.

starnes13_002-031hr.indd 28

20. Which test? The example on page 19 describes one

of the experiments done by Kathleen Vohs and her
coworkers to demonstrate that even being reminded
of money makes people more self-sufficient and less
involved with other people. Here are three more of
these experiments. For each experiment, which inference method from Chapters 8 to 13 would you use,
and why?
a)
( Randomly assign student subjects to money
and control groups. The control group unscrambles
neutral phrases, and the money group unscrambles
money-oriented phrases, as described in the earlier
example. Then ask the subjects to volunteer to help
the experimenter by coding data sheets, which takes

4/11/11 2:16 PM

13.1 One-Way Analysis of Variance 29

about five minutes per sheet. Subjects said how
many sheets they would volunteer to code. Participants in the money condition volunteered to help
code fewer data sheets than did participants in the
control condition.
b)
( Randomly assign student subjects to high-money,
low-money, and control groups. After playing
Monopoly for a short time, the high-money group is
left with $4000 in Monopoly money, the low-money
group with $200, and the control group with no
money. Each subject is asked to imagine a future with
lots of money (high-money group), a little money
(low-money group), or just their future plans (control
group). Another student walks in and spills a box of
27 pencils. How many pencils does the subject pick
up? Participants in the high-money condition gathered fewer pencils than subjects in the other
two groups.
c)
( Randomly assign student subjects to three groups.
All do paperwork while a computer on the desk
shows a screensaver of currency floating underwater
(Group 1), a screensaver of fish swimming underwater (Group 2), or a blank screen (Group 3). Each
subject must now develop an advertisement and
can choose whether to work alone or with a partner.
Count how many in each group make each choice.
Choosing to perform the task with a coworker was
reduced among money condition participants.
21. Can you hear these words? The figure below displays
the Minitab output for one-way ANOVA applied
to the hearing data described in Exercise 19. The
response variable is Percent, and List identifies
the four lists of words. Based on this analysis, is there
good reason to think that the four lists are not all
equally difficult? Perform an appropriate test to help
answer this question. Graphs of the data (not shown)
confirm that the Normal condition is met.

22. Which blue is most blue? The color of a fabric

depends on the dye used and also on how the dye
is applied. This matters to clothing manufacturers,
who want the color of the fabric to be just right. A
manufacturer dyes fabric made of ramie with the
same procion blue die but applies the dye in three

starnes13_002-031hr.indd 29

different ways. Then he uses a colorimeter to measure

the lightness of the color on a scale in which black
is 0 and white is 100. Here are the data for 24 pieces
of fabric divided into three groups of 8 that were
randomly assigned to be dyed in each way:17
Method A:

42.05

41.44

42.27

41.12

41.49

40.98

40.88

41.30

41.28

41.66

41.50

41.39

41.27

41.68

41.65

42.30

42.04

42.25

Method C:

41.83

41.27
Method B:

41.72

41.99

41.72

41.97

The figure below displays the Minitab output for

one-way ANOVA applied to these data. Based on
this analysis, is there good reason to think that the
three methods do not all yield the same shade of
blue? Perform an appropriate test to help answer
this question.

23. Python weights A study of the effect of nest temperature on the development of water pythons separated
python eggs at random into nests at three temperatures:
cold, neutral, and hot. Exercise 33 in Chapter 11
(page 725) shows that the proportions of eggs that
hatched at each temperature did not differ significantly. Now we will examine the weights of the little
pythons. In all, 16 eggs hatched at the cold temperature, 38 at the neutral temperature, and 75 at the hot
temperature. The report of the study summarizes the
data in the common form mean standard error of
the mean as follows:18

Temperature

Weight (grams)
at Hatching

Propensity
to Strike

Cold

18

28.89 8.08

6.40 5.67

Neutral

38

32.93 5.61

5.82 4.24

Hot

75

32.27 4.10

4.30 2.70

4/11/11 2:16 PM

30 CHAPTER

13

Analysis of Variance

Graphs of the data (not shown) confirm that the

N
ormal condition is met.
a) will compare the mean weights at hatching.
( We
Recall that the standard error of the mean is sx n .
Find the standard deviations of the weights in the
three groups, and verify that they satisfy our rule of
thumb for using ANOVA.
b)
( Starting from the sample sizes ni, the means x i,
and the standard deviations si, carry out an ANOVA.
That is, find MSG, MSE, the F statistic, and the
P-value. Is there evidence that nest temperature
affects the mean weight of newly hatched pythons?
Justify your answer.
24. Python strikes The data in the previous exercise
also describe the propensity to strike of the hatched
pythons at 30 days of age. This is the number of taps
on the head with a small brush until the python
launches a strike. (Dont try this at home!) The data
are again summarized in the form sample mean
standard error of the mean. Graphs of the data (not
shown) confirm that the Normal condition is met.
Follow the outline in parts (a) and (b) of the previous
exercise for propensity to strike. Does nest temperature appear to influence propensity to strike?
25. F versus t We have two methods to compare the
means of two groups: the two-sample t test of
Section 10.2 and the ANOVA F test with k = 2.
We prefer the t test because it allows one-sided alternatives and does not assume that both populations
have the same standard deviation. Lets apply both
tests to the same data.
There are two types of life insurance companies.
Stock companies have shareholders, and mutual
companies are owned by their policyholders. Take an
SRS of each type of company from those listed in a
directory of the industry. Then ask the annual cost per
$1000 of insurance for a $50,000 policy insuring the
life of a 35-year-old man who does not smoke. Here
are the data summaries:19
Stock companies

Mutual companies

ni

13

17

xi

$2.31

$2.37

si

$0.38

$0.58

Graphs of the data (not shown) confirm that the

N
ormal condition is met.
a)
( Calculate the two-sample t statistic for testing
H0:m1 = m2 against the two-sided alternative. Use the
conservative method to find the P-value.
b)
( Calculate MSG, MSE, and the ANOVA F statistic for the same hypotheses. What is the P-value of F?

starnes13_002-031hr.indd 30

c)
( How close are the two P-values? (The square
root of the F statistic is a t statistic with N - k =
n1 + n2 -2 degrees of freedom. This is the pooled
two-sample t mentioned on page 645. So F for k = 2
is exactly equivalent to a t statistic, but it is a slightly
different t from the one we use.)
Multiple choice: Select the best answer for
Exercises 26 to 29.
26. A study of the effects of smoking classifies subjects as
nonsmokers, moderate smokers, or heavy smokers.
The investigators interview a random sample of 200
people in each group. Among the questions is How
many hours do you sleep on a typical night? The
degrees of freedom for the ANOVA F statistic comparing mean hours of sleep are
a) and 197. (d) and 597.
( 2
3
b) and 199. (e) and 599.
( 2
3
c) and 597.
( 2
27. The alternative hypothesis for the ANOVA F test in
the previous exercise is
a) mean hours of sleep in the groups are all the
( the
same.
b) mean hours of sleep in the groups are all
( the
d
ifferent.
c) mean hours of sleep in the groups are not all
( the
the same.
d) mean hours of sleep is highest for the
( the
n
onsmokers.
e)
( there is an association between sleep and smoking
status.
Exercises 28 and 29 refer to the following setting. The
air in poultry-processing plants often contains fungus
spores. Large spore concentrations can affect the health
of the workers. To measure the presence of spores, air
samples are pumped to an agar plate and colonyforming units (CFUs) are counted after an incubation
period. Here are data from the kill room of a plant
that slaughters 37,000 turkeys a day, taken at
four seasons of the year. Each bservation was made
o
on a different day. The units are CFUs per cubic meter
of air.20
Fall

Winter

Spring

Summer

1231
1254

384

2105

3175

104

701

2526

752

251

2947

1763

1088

97

842

1090

Here is Minitab output for ANOVA to compare mean

CFUs in the four seasons:

4/11/11 2:16 PM

13.1 One-Way Analysis of Variance 31

c) air in the plant is clearly unhealthy.
( the
d)
( there is more variability in CFUs during the summer than in other seasons.
e) graphs are all roughly symmetrical.
( the

28. The most striking conclusion from the numerical

summaries for the turkey-processing plant is that
a)
( there appears to be little difference among the
seasons.
b) average, CFUs are much lower in winter than
( on
in other seasons.

starnes13_002-031hr.indd 31

29. The P-value 0.014 in the output may not be accurate

because the conditions for ANOVA are not satisfied.
Which condition is clearly not satisfied based solely
on the Minitab output?
a)
( Random
b)
( Normal
c)
( Independent
d)
( Same SD
e)
( None of these

4/11/11 2:16 PM

32 CHAPTER

13

Analysis of Variance

Section 13.1

Notes

1. These are some of the data from the EESEE story

Stress among Pets and Friends. The study results
appear in K. Allen, J. Blascovich, J. Tomaka, and
R. M. Kelsey, Presence of human friends and pet
dogs as moderators of autonomic responses to stress in
women, Journal of Personality and Social Psychology,
83 (1988), pp. 582589.
2. Ethan J. Temeles and W. John Kress, Adaptation in a
plant-hummingbird association, Science, 300 (2003),
pp. 630633. I thank Ethan J. Temeles for providing
the data.
3. Jacqueline T. Ngai and Diane S. Srivastava,
Predators accelerate nutrient cycling in a bromeliad
ecosystem, Science, 314 (2006), p. 963. I thank
Jacqueline Ngai for providing the data.
4. Kathleen D. Vohs, Nicole L. Mead, and Miranda R.
Goode, The psychological consequences of money,
Science, 314 (2006), pp. 11541156. I thank Kathleen
Vohs for providing the data.
5. Modified from M. C. Wilson and R. E. Shade,
Relative attractiveness of various luminescent colors
to the cereal leaf beetle and the meadow spittlebug,
Journal of Economic Entomology, 60 (1967),
pp. 578580.
6. K. B. Suttle, Meredith A. Thomsen, and Mary
E. Power, Species interactions reverse grassland
responses to changing climate, Science, 315 (2007),
pp. 640642.
7. Amandine Cournil, Jean-Marie Legay, and Franois
Schchter, Evidence of sex-linked effects on the
inheritance of human longevity: a population-based
study in the Valserine Valley (French Jura), 1820th
centuries, Proceedings of the Royal Society of
London: Biological Sciences, 267 (2000),
pp. 10211025.
8. Sanders Korenman and David Neumark, Does
marriage really make men more productive? Journal
of Human Resources, 26 (1991), pp. 282307.
9. Elisabeth Wells-Parker et al., An exploratory study
of the relationship between road rage and crash experience in a representative sample of U.S. drivers,
Accident Analysis and Prevention, 34 (2002),
pp. 271278.

starnes13_002-031hr.indd 32

10. Based on the online supplement to Paul J. Shaw

et al., Correlates of sleep and waking in Drosophila
melanogaster, Science, 287 (2000), pp. 18341837.
11. Annie C. St. Pierre et al., Insulin resistance syndrome, body mass index and the risk of ischemic
heart disease, Canadian Medical Association Journal,
172 (2005), pp. 13011305.
12. David B. Wooten, One-of-a-kind in a full house:
some consequences of ethnic and gender distinctiveness, Journal of Consumer Psychology, 4 (1995),
pp. 205224.
13. John P. Thomas, Influences on mathematics
learning and attitudes among African American
high school students, Journal of Negro Education,
69 (2000), pp. 165183.
14. John M. Jakicic et al., Effects of intermittent exercise
and use of home exercise equipment on adherence,
weight loss, and fitness in overweight women,
Journal of the American Medical Association,
282 (1999), pp. 15541560.
15. C. H. Cannon, D. R. Peart, and M. Leighton, Tree
species diversity in commercially logged Bornean
rainforest, Science, 281 (1998), pp. 13661367.
I thank Charles Cannon for providing the data.
16. The data and the full story can be found in the Data
and Story Library at lib.stat.cmu.edu. The original
study is by Faith Loven, A study of interlist equivalency of the CID W-22 word list presented in quiet
and in noise, MS thesis, University of Iowa, 1981.
17. Yvan R. Germain, The dyeing of ramie with fiber
reactive dyes using the cold pad-batch method,
MS thesis, Purdue University, 1988.
18. R. Shine, T. R. L. Madden, M. J. Elphick, and
P. S. Harlow, The influence of nest temperatures
and maternal brooding on hatchling phenotypes in
water pythons, Ecology, 78 (1997), pp. 17131721.
19. Mark Kroll, Peter Wright, and Pochera Theerathorn,
Whose interests do hired managers pursue? An
examination of select mutual and stock life insurers,
Journal of Business Research, 26 (1993), pp. 133148.
20. Michael W. Peugh, Field investigation of ventilation
and air quality in duck and turkey slaughter plants,
MS thesis, Purdue University, 1996.

4/11/11 2:16 PM