2 way ANOVA(Analysis Of VAriance

www.stat.ufl.edu/~ssaha/3024.html Ch 13 – Two Way ANOVA
CHAPTER 13 – TWO-WAY ANALYSIS OF VARIANCE
Two-way ANOVA has many of the same ideas as one-way ANOVA, with
the main difference being the inclusion of another factor (or explanatory
variable) in our model.
In the two-way ANOVA model, there are two factors, each with its own
number of levels. When we are interested in the effects of two factors, it is
much more advantageous to perform a two-way analysis of variance, as
opposed to two separate one-way ANOVAs.
There are three main advantages of two-way ANOVA:
- It is more efficient to study two factors simultaneously rather than
separately.
- We can reduce the residual variation in a model by including a
second factor thought to influence the response.
- We can investigate interactions between factors.
The interaction between two variables is usually the most interesting
feature of a two-way analysis of variance. When two factors interact, the
effect on the response variable of one explanatory variable depends on the
specific value or level of the other explanatory variable.
For example, the statement “being overweight caused greater increases in
blood pressure for men than for women” is a statement describing
interaction. The effect of weight (factor #1, categorical – overweight or not
overweight) on blood pressure (response) depends on gender (factor #2,
categorical – male or female).
The term main effect is used to describe the overall effect of a single
explanatory variable. For our previous example, there would be two main
effects: the effect of weight on blood pressure and the effect of gender on
blood pressure.
1

The presence of a main effect might not necessarily be useful when an
interaction effect exists. For example, it might be sensible to report the
effect of being overweight on blood pressure without reporting that there is a
difference in the effect of being overweight on blood pressure for men and
women.
There are many types of interactions which can often be seen from patterns
in graphs. We will study some types in the following examples.
Example #1(a)
The Bureau of Labor Statistics collects data on earnings of workers in the
US classified according to various characteristics. Here are the mean weekly
earnings in dollars of men and women in two age groups who were working
full-time in the first quarter of 1997:
Age Women Men Mean
16-19
20-24
239
302
264
333
251.5
317.5
Mean 270.5 298.5 284.5
Figure 13.1 is a plot of the group
means. It clearly shows an existence of
main effects.
From the plot we can see that on
average men earn more than women
indicating an effect of gender. We can
also see an effect of age since the older
groups makes on average more than the
younger age group.
What about the interaction between
gender and age?
An interaction is present when the
main effects cannot provide a complete
description of the data.
Figure 13.1
2

In our graph, the only two effects that are present are an effect due to age and
an effect due to gender. We know that men make, on average, more than
women, but there is no dependence on age to determine the amount that men
make more than women. Therefore, since there is no dependence between
gender and age, there is no interaction effect.
Parallel lines in our plot usually imply little or no interaction.
Example #1 (b)
The survey described in 1(a) also gave weekly earnings for groups of older
workers. Here is the complete table:
Figure 13.2
Age Women Men Mean
16-19
20-24
25-34
35-44
45-54
55-64
239
302
422
475
496
426
264
333
514
639
692
660
251.5
317.5
468.0
557.0
594.0
543.0
Mean 393 517 455.0
Figure 13.2 is a plot of the group means. It demonstrates both main effects
and an interaction between gender and age.
In this case, the main effects do not provide a complete description of the data
since now the gender difference in earnings depends on which age group we
examine. As the age of the workers increases, the earnings increase (except in
the last age group). We also see a main effect of gender on earnings.
However, our interaction shows that the amount by which the men earn over the
women depends on the age group.
Therefore, there is a very strong interaction present and our main effects don’t
provide us with enough information to get an accurate idea of the relationship
between age and earnings as well as gender and earnings.
3

Example #2
A study of the energy expenditure of farmers in Burkina Faso collected data
during the wet and dry seasons. The farmers grow millet during the wet
season. In the dry season there is very little activity because the ground is
too hard to grow crops. The mean energy expended (in calories) by men
and women in Burkina Faso during the wet and dry seasons is given in the
following table:
Season Men Women Mean
Dry
Wet
2310
3460
2320
2890
2315
3175
Mean 2885 2605 2745
During the dry season both men and women
use about the same amount of energy.
However, during the wet season, both
genders burn more calories. Therefore, there
is an effect of season on the number of
calories burned.
During the wet season, since it is the custom
for men to do most of the field work, they
use more energy on average than women.
Therefore, there is an effect of gender on the
number of calories burned, but only in the
wet season.
So there is an interaction between gender
and season, as well as main effects of both
factors.
Figure 13.3
4

THE TWO-WAY ANOVA MODEL
When discussing two-way ANOVA models, we will use the labels A and B
to represent our two factors. In examples, we will use the factor names so
that we can easily see the meaning of each variable.
We use I to represent the number of levels of factor A and J to represent the
number of levels of factor B. Therefore, we call the general two-way
problem an I x J ANOVA, since every level of A is combined with every
level of B, so that I x J groups are compared.
The sample size for level i of factor A and level j of factor B is nij. The total
number of observations is N = Σ nij.
Example #3
A company wants to compare three different training programs for its new
employees. Each of these programs takes 8 hours to complete. The training
can be given for 8 hours on one day or for 4 hours on two consecutive days.
The next 120 employees that the company hires will be the subjects for this
study. After the training is completed, the employees are asked to evaluate
the effectiveness of the program on a 7-point scale. Identify the response and
explanatory variables, state the number of levels for each factor (I and J) and
the total number of observations (N).
5

Solution:
Explanatory variables:
1) Training program (I = 3 levels)
2) The manner in which the employee does the training (J = 2 levels: 8
hrs f or one day, 4 hrs on two consecutive days).
Response variable: the 7-point scaled effectiveness score
Total number of observations N = 120
_____________________________________________________________
For our two-way ANOVA, we assume to have independent SRSs of size nij
from each of I x J normal populations. The population means µij may differ,
but all populations must have the same standard deviation σ.
Let xijk be the kth
observation from the population having factor A at level i
and factor B at level j. Since our observations differ from their mean by a
value of εijk (xijk – µij = εijk), we can write our two-way ANOVA model in
the form: xijk = µij + εijk.
In this case, i = 1,2, …, I, j = 1,2,…, J, k = 1,2,…,nij and the deviations εijk
are normally distributed with mean 0 and standard deviation σ.
As in the previous chapters, we can describe our model by DATA = FIT +
RESIDUAL. In this case, the fit part of our model is the means µij, (since
we use each mean to describe each population) and the residual part is the
deviations εijk of the individual observations from their group means.
Our population means µij and the population standard deviations are all
unknown. Therefore, we pick simple random samples (SRSs) to learn about
the relationship between our factors and the response in our samples and
then use that to estimate the relationship in our population.
6

PARAMETER ESTIMATES
The population mean for the group with level i of factor A and level j of
factor B is represented by µij. It can be estimated by
1
1 ijn
ij ijk
ij k
x x
n =
= ∑ .
Since every σ (population standard deviation) is assumed to be the same in
two-way ANOVA, we can pool all of our estimates (sample standard
deviations) to give one estimate of the population standard deviation.
2
2 ( 1)
( 1)
ij ij
p
ij
n s
s
n
−
=
−
∑
∑ and
2
p ps s=
Just like in one-way ANOVA, sp
2
= MSE, and so the numerator of sp
2
is
equal to SSE and the denominator is equal to DFE (since we have ( 1ijn )−∑ =
N - IJ, so we have N observations being compared to IJ sample means).
7

Example #4
Students in a statistics class gave information on their height and weight and
also reported their perception of their weight (about right, underweight,
overweight). The body mass index, BMI, was calculated for each
combination of student gender and perception of weight category (BMI =
Weight/Height). Calculate each sample mean based on the data.
Perception of Weight
Underweight About Right Overweight
Female 20.2 20.6 18.1 21.4 22.5 23.1 20.9 24.5 26.0 22.7 24.6
Male 23.8 20.5 23.4 26.5 22.2 24.0 28.1 26.4 27.7
Solution:
Let factor A be the perception of weight and let factor B be gender. Within
factor A, let level 1 be underweight, level 2 be about right, and level 3 be
overweight. Within factor B (gender), let level 1 be female and let level 2 be
male. So µ11 corresponds to the population mean BMI of underweight
females, µ12 corresponds to the population mean BMI of underweight males,
and so on. So to estimate these population values, we use xij(bar) to estimate
µij for i=1,2,3 and j=1,2.
Now,
x11(bar) = (20.2 + 20.6 + 18.1 + 21.4) / 4 = 20.075
x12(bar) = (23.8 + 20.5) / 2 = 22.15
x21(bar) = (22.5 + 23.1 + 20.9) / 3 = 21.167
x22(bar) = (23.4 + 26.5 + 22.2 + 24.0) / 4 = 24.025
x31(bar) = (24.5 + 26 + 22.7 + 24.6) / 4 = 24.45
x32(bar) = (28.1 + 26.4 + 27.7) / 3 = 27.4
Note: (bar) just means that the x has a bar over it (indicating a mean).
8

TWO-WAY ANOVA TABLE
In two-way ANOVA, there are three F statistics that are calculated and used
in significance tests: two that test for the main effects and one that tests for
an interaction.
Therefore, the sum of squares for our FIT (SSM) is made up of three parts:
SSA - sum of squares for factor A
SSB - sum of squares for factor B
SSAB - sum of squares for the interaction between factor A
and factor B
Our total sum of squares and total degrees of freedom are still the sum of the
sources of variation and degrees of freedom in our model.
SST = SSA + SSB + SSAB + SSE
DFT = DFA + DFB + DFAB + DFE
When our sample sizes are different, do not be alarmed if our sums of
squares do not add up to our given SST. When our sample sizes are
different, this can cause sums of squares which don’t add.
Since we have I levels of factor A, DFA = I – 1.
Since we have J levels of factor B, DFB = J – 1.
Since SSM = SSA + SSB + SSAB and DFM = IJ – 1,
DFAB = (IJ – 1) – (I – 1) – (J – 1) = (I – 1)( J – 1).
DFE = N – IJ (we have N observations and IJ sample means)
DFT = N – 1
9

TWO-WAY ANOVA TABLE
Source
Degrees of
Freedom Sums of Squares Mean Square F
A I – 1 SSA SSA/DFA MSA/MSE
B J – 1 SSB SSB/DFB MSB/MSE
AB (I – 1)(J – 1) SSAB SSAB/DFAB MSAB/MSE
Error N – IJ SSE SSE/DFE
Total N - 1 SST
Remember that when we do a two-way ANOVA, we need to assume that
our data is normally distributed and that the population standard
deviations are equal.
Therefore, we must make sure that twice our smallest sample standard
deviation is larger than our largest standard deviation.
10

HYPOTHESES FOR TWO-WAY ANOVA
To test the main effect of A:
H0: No main effect of A Ha: There exists a main effect of factor A
F = MSA Compare to F(I-1, N-IJ)
MSE
To test the main effect of B:
H0: No main effect of B Ha: There exists a main effect of factor B
F = MSB Compare to F(J-1, N-IJ)
MSE
To test the interaction of A and B:
H0: No interaction Ha: There exists an interaction between factor A and
factor B
F = MSAB Compare to F((I-1)(J-1), N-IJ)
MSE
11

Example #5
When a restaurant server writes a friendly note or draws a “happy face” on
your restaurant check, is this just a friendly act or is there a financial
incentive? Psychologists conducted a randomized experiment to investigate
whether drawing a happy face on the back of a restaurant bill increased the
average tip given to the server. One female server and one male server in a
Philadelphia restaurant either did or did not draw a happy face on checks
during the experiment. In all they drew happy faces on 45 checks and did not
draw happy faces on 44 checks. The sequence of drawing the happy faces or
not was random.
a) Identify the response and explanatory variables, state the number of levels
for each factor (I and J) and the total number of observations (N).
Solution:
Response variable: tip
Explanatory variables: gender of server (2 levels), message (2 levels, yes or
no)
Total number of observations: N = 89
b) Complete the following two-way ANOVA table and then perform the
appropriate F tests for main effects and interaction and state your
conclusions.
Try to fill the table in on your own (answer is below):
Source DF SS MS F
Message 14.7
Gender 2602.0
Interaction 438.7
Error 109.8
Total 12407.9
12

Solution:
Source DF SS MS F
Message 1 14.7 14.7 0.134
Gender 1 2602.0 2602.0 23.7
Interaction 1 438.7 438.7 4.0
Error 85 9333.0 109.8
Total 88 12407.9 ---
H0: No main effect of message
Ha: A main effect of message exists
Test Statistic: F0 = 14.7/109.8 = 0.134 with numerator degrees of freedom 1
and denominator degrees of freedom 85.
Decision Rule: This test statistic corresponds to a p-value of 0.7152. We do
not have any evidence to reject the null hypothesis that there is main effect
of message on the average amount a server gets tipped.
_____________________________________________________________
H0: No main effect of gender
Ha: A main effect of gender exists
Test Statistic: F0 = 2602/109.8 = 23.7 with numerator df = 1 and
denominator df = 85.
Decision Rule: This test statistic corresponds to a p-value of less than .0001.
We have very strong evidence that a main effect of gender does exist.
_____________________________________________________________
H0: No interaction effect between gender and message
Ha: An interaction effect between gender and message exists.
Test Statistic: F0 = 438.7/109.8 = 4.0 with numerator df = 1 and
denominator df = 85.
Decision Rule: This test statistic corresponds to a p-value of .0487. We
have evidence to reject the null hypothesis of no interaction at the α = 0.05
level. We have reason to believe that there is an interaction effect.
_____________________________________________________________
13

2 way ANOVA(Analysis Of VAriance

More Related Content

What's hot (20)

Similar to 2 way ANOVA(Analysis Of VAriance (20)

More from musadoto (20)

Recently uploaded (20)

2 way ANOVA(Analysis Of VAriance