Statistics 502 Lecture Notes

Peter D. Hoff
c
December
9, 2009

Contents
1 Principles of experimental design
1.1 Induction . . . . . . . . . . . . . . . .
1.2 Model of a process or system . . . . . .
1.3 Experiments and observational studies
1.4 Steps in designing an experiment . . .

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.

1
1
2
2
6

2 Test statistics and randomization distributions

2.1 Summaries of sample populations . . . . . . . .
2.2 Hypothesis testing via randomization . . . . . .
2.3 Essential nature of a hypothesis test . . . . . .
2.4 Sensitivity to the alternative hypothesis . . . . .
2.5 Basic decision theory . . . . . . . . . . . . . . .

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

9
10
13
17
18
23

.
.
.
.
.
.
.

25
25
29
30
36
42
43
43

.
.
.
.

47
47
49
52
54

3 Tests based on population models

3.1 Relating samples to populations .
3.2 The normal distribution . . . . .
3.3 Introduction to the t-test . . . . .
3.4 Two sample tests . . . . . . . . .
3.5 Checking assumptions . . . . . .
3.5.1 Checking normality . . . .
3.5.2 Unequal variances . . . . .

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.

4 Confidence intervals and power

4.1 Confidence intervals via hypothesis tests
4.2 Power and Sample Size Determination .
4.2.1 The non-central t-distribution . .
4.2.2 Computing the Power of a test .

.
.
.
.

.
.
.
.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.
.
.
.

.
.
.
.

.
.
.
.

.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.

.
.
.
.

.
.
.
.
.
.
.

.
.
.
.

CONTENTS

ii

5 Introduction to ANOVA
5.1 A model for treatment variation . . . . . . . . .
5.1.1 Model Fitting . . . . . . . . . . . . . . .
5.1.2 Testing hypothesis with MSE and MST .
5.2 Partitioning sums of squares . . . . . . . . . . .
5.2.1 The ANOVA table . . . . . . . . . . . .
5.2.2 Understanding Degrees of Freedom: . . .
5.2.3 More sums of squares geometry . . . . .
5.3 Unbalanced Designs . . . . . . . . . . . . . . . .
5.3.1 Sums of squares and degrees of freedom .
5.3.2 ANOVA table for unbalanced data: . . .
5.4 Normal sampling theory for ANOVA . . . . . .
5.4.1 Sampling distribution of the F -statistic .
5.4.2 Comparing group means . . . . . . . . .
5.4.3 Power calculations for the F-test . . . .
5.5 Model diagnostics . . . . . . . . . . . . . . . . .
5.5.1 Detecting violations with residuals . . .
5.5.2 Checking normality assumptions: . . . .
5.5.3 Checking variance assumptions . . . . .
5.5.4 Variance stabilizing transformations . . .
5.6 Treatment Comparisons . . . . . . . . . . . . .
5.6.1 Contrasts . . . . . . . . . . . . . . . . .
5.6.2 Orthogonal Contrasts . . . . . . . . . . .
5.6.3 Multiple Comparisons . . . . . . . . . .
5.6.4 False Discovery Rate procedures . . . . .
5.6.5 Nonparametric tests . . . . . . . . . . .
6 Factorial Designs
6.1 Data analysis: . . . . . . . . . . . . . .
6.2 Additive effects model . . . . . . . . .
6.3 Evaluating additivity: . . . . . . . . . .
6.4 Inference for additive treatment effects
6.5 Randomized complete block designs . .
6.6 Unbalanced designs . . . . . . . . . . .
6.7 Non-orthogonal sums of squares: . . . .
6.8 Analysis of covariance . . . . . . . . .
6.9 Types of sums of squares . . . . . . . .

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

60
62
63
66
70
72
73
76
78
79
81
83
85
88
90
92
93
94
96
100
106
107
110
112
115
115

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.

.
.
.
.
.
.
.
.
.

116
. 117
. 123
. 126
. 130
. 140
. 146
. 153
. 155
. 159

CONTENTS

iii

7 Nested Designs
163
7.1 Mixed-effects approach . . . . . . . . . . . . . . . . . . . . . . 171
7.2 Repeated measures analysis . . . . . . . . . . . . . . . . . . . 174

List of Figures
1.1

Model of a variable process . . . . . . . . . . . . . . . . . . . .

2.1
2.2
2.3

Wheat yield distributions . . . . . . . . . . . . . . . . . . . . .

Approximate randomization distribution for the wheat example
Histograms and empirical CDFs of the first two hypothetical
samples. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Randomization distributions for the t and KS statistics for
the first example. . . . . . . . . . . . . . . . . . . . . . . . .
Histograms and empirical CDFs of the second two hypothetical samples. . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Randomization distributions for the t and KS statistics for
the second example. . . . . . . . . . . . . . . . . . . . . . . .

12
16

3.1
3.2
3.3
3.4
3.5
3.6

The population model . . . . . . . . . . . . . . . .

2 distributions . . . . . . . . . . . . . . . . . . . .
t-distributions . . . . . . . . . . . . . . . . . . . . .
The t-distribution under H0 for the wheat example
Randomization and t-distributions for the t-statistic
Normal scores plots. . . . . . . . . . . . . . . . . .

27
33
34
39
40
44

4.1
4.2
4.3

A t10 distribution and two non-central t10 -distributions. . . .

Critical regions and the non-central t-distribution . . . . . .
and power versus sample size, and the normal approximation
to the power. . . . . . . . . . . . . . . . . . . . . . . . . . .
Null and alternative distributions for another wheat example,
and power versus sample size. . . . . . . . . . . . . . . . . .

2.4
2.5
2.6

4.4
5.1
5.2

. . . . . .
. . . . . .
. . . . . .
. . . . . .
under H0
. . . . . .

20
21
22
22

. 52
. 55
. 57
. 59

Response time data . . . . . . . . . . . . . . . . . . . . . . . . 61

Randomization distribution of the F -statistic . . . . . . . . . . 70
iv

LIST OF FIGURES

5.3
5.4
5.5
5.6
5.7
5.8
5.9
5.10
5.11
5.12
5.13
5.14
5.15

Coagulation data . . . . . . . . . . . . . . . . . . . . . . . . . 83
F-distributions . . . . . . . . . . . . . . . . . . . . . . . . . . 87
Normal-theory and randomization distributions of the F -statistic 88
Power as a function of n for m = 4, = 0.05 and 2 / 2 = 1 . 92
Power as a function of n for m = 4, = 0.05 and 2 / 2 = 2 . 92
Normal scores plots of normal samples, with n {20, 50, 100} 95
Crab data . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 97
Crab residuals . . . . . . . . . . . . . . . . . . . . . . . . . . . 98
Fitted values versus residuals . . . . . . . . . . . . . . . . . . 99
Data and log data . . . . . . . . . . . . . . . . . . . . . . . . . 101
Diagnostics after the log transformation . . . . . . . . . . . . 102
Mean-variance relationship of the transformed data . . . . . . 107
Yield-density data . . . . . . . . . . . . . . . . . . . . . . . . 110

6.1
6.2
6.3
6.4
6.5
6.6
6.7
6.8
6.9
6.10
6.11
6.12
6.13
6.14
6.15
6.16
6.17
6.18
6.19

Marginal Plots. . . . . . . . . . . . . . . . . . . . . . . . . .
Conditional Plots. . . . . . . . . . . . . . . . . . . . . . . .
Cell plots. . . . . . . . . . . . . . . . . . . . . . . . . . . .
Mean-variance relationship. . . . . . . . . . . . . . . . . . .
Mean-variance relationship for transformed data. . . . . . .
Plots of transformed poison data . . . . . . . . . . . . . . .
Comparison between types I and II, without respect to delivery. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
Comparison between types I and II, with delivery in color. .
Marginal plots of the data. . . . . . . . . . . . . . . . . . . .
Three datasets exhibiting non-additive effects. . . . . . . . .
Experimental material in need of blocking. . . . . . . . . . .
Results of the experiment . . . . . . . . . . . . . . . . . . .
Marginal plots, and residuals without controlling for row. .
Marginal plots for pain data . . . . . . . . . . . . . . . . . .
Interaction plots for pain data . . . . . . . . . . . . . . . . .
Oxygen uptake data . . . . . . . . . . . . . . . . . . . . . .
ANOVA and ANCOVA fits to the oxygen uptake data . . .
Unbalanced design: Controlling eliminates effect. . . . . . .
Unbalanced design: Controlling highlights effect. . . . . . . .

7.1
7.2
7.3

Potato data. . . . . . . . . . . . . . . . . . . . . . . . . . . . . 165

Diagnostic plots for potato ANOVA. . . . . . . . . . . . . . . 166
Potato data . . . . . . . . . . . . . . . . . . . . . . . . . . . . 167

.
.
.
.
.
.

118
119
120
120
121
122

.
.
.
.
.
.
.
.
.
.
.
.
.

131
132
136
139
141
142
143
150
151
156
157
159
161

LIST OF FIGURES
7.4
7.5
7.6
7.7
7.8

Potato data . . . . . . . . . . . . . . . . . .
Sitka spruce data. . . . . . . . . . . . . . . .
ANCOVA fit and residuals . . . . . . . . . .
Within-tree dependence . . . . . . . . . . .
Reduction to tree-specific summary statistics

vi
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

.
.
.
.
.

172
174
176
177
178

Chapter 1
Principles of experimental
design
1.1

Induction

Much of our scientific knowledge about processes and systems is based on

induction: reasoning from the specific to the general.
Example (survey): Do you favor increasing the gas tax for public transportation?
Specific cases: 200 people called for a telephone survey
Inferential goal: get information on the opinion of the entire city.
Example (Womens Health Initiative): Does hormone replacement improve health status in post-menopausal women?
Specific cases: Health status monitored in 16,608 women over a 5-year
period. Some took hormones, others did not.
Inferential goal : Determine if hormones improve the health of women
not in the study.

CHAPTER 1. PRINCIPLES OF EXPERIMENTAL DESIGN

x1
x2

Process


Figure 1.1: Model of a variable process

1.2

Model of a process or system

We are interested in how the inputs of a process affect an output. Input

variables consist of
controllable factors x1 : measured and determined by scientist.
uncontrollable factors x2 : measured but not determined by scientist.
noise factors : unmeasured, uncontrolled factors, often called experimental variability or error.
For any interesting process, there are inputs such that:
variability in input variability in output
If variability in an input factor x leads to variability in output y, we say x is
a source of variation. In this class we will discuss methods of designing and
analyzing experiments to determine important sources of variation.

1.3

Experiments and observational studies

Information on how inputs affect output can be gained from:

Observational studies: Input and output variables are observed from a
pre-existing population. It may be hard to say what is input and what
is output.
Controlled experiments: One or more input variables are controlled
and manipulated by the experimenter to determine their effect on the
output.

CHAPTER 1. PRINCIPLES OF EXPERIMENTAL DESIGN

Example (Womens Health Initiative, WHI):

Population: Healthy, post-menopausal women in the U.S.
Input variables:
1. estrogen treatment, yes/no
2. demographic variables (age, race, diet, etc.)
3. unmeasured variables (?)
Output variables:
1. coronary heart disease (eg. MI)
2. invasive breast cancer
3. other health related outcomes
Scientific question: How does estrogen treatment affect health outcomes?
Observational Study:
1. Observational population: 93,676 women enlisted starting in 1991,
tracked over eight years on average. Data consists of x= input variables,
y=health outcomes, gathered concurrently on existing populations.
2. Results: good health/low rates of CHD generally associated with estrogen treatment.
3. Conclusion: Estrogen treatment is positively associated with health outcomes, such as prevalence of CHD.
Experimental Study (WHI randomized controlled trial):
1. Experimental population:
373,092 women determined to be eligible
, 18,845 provided consent to be in experiment
, 16,608 included in the experiment

CHAPTER 1. PRINCIPLES OF EXPERIMENTAL DESIGN


16,608 women randomized to either

x = 1 (estrogen treatment)
x = 0 (no estrogen treatment)

Women were of different ages and were treated at different clinics.

Women were blocked together by age and clinic, and then treatments
were randomly assigned within each agetreatment block. This type
of random allocation is called a randomized block design.
age group
1 (50-59) 2 (60-69) 3 (70-79)
clinic 1
n11
n12
n13
2
n21
n22
n23
..
..
..
..
.
.
.
.

ni,j = # of women in study, in clinic i and in age group j

= # of women in block i, j
Randomization scheme: For each block, 50% of the women in that
block were randomly assigned to treatment (x = 1) and the remaining
assigned to control (x = 0).
Question: Why did they randomize within a block?
2. Results: JAMA, July 17 2002. Also see the NLHBI press release . Women
on treatment had a higher incidence rate of
CHD
breast cancer
stroke
pulmonary embolism
and a lower incidence rate of
colorectal cancer
hip fracture

CHAPTER 1. PRINCIPLES OF EXPERIMENTAL DESIGN

3. Conclusion: Estrogen treatment is not a viable preventative measure

for CHD in the general population. That is, our inductive inference is
(specific) higher CHD rate in treatment population than control
suggests
(general) if everyone in the population were treated, the incidence
rate of CHD would increase.
Question: Why the different conclusions between the two studies? Consider the following possible explanation: Let
x = estrogen treatment
 = health consciousness (not directly measured)
y = health outcomes

correlation

Association between x and y may be due to an unmeasured variable .


randomization

correlation=
causation

CHAPTER 1. PRINCIPLES OF EXPERIMENTAL DESIGN

Randomization breaks the association between  and x.

Observational studies can suggest good experiments to run, but cant
definitively show causation.
Randomization can eliminate correlation between x and y due to a
different cause , aka a confounder.
No causation without randomization

1.4

Steps in designing an experiment

1. Identify research hypotheses to be tested.

2. Choose a set of experimental units, which are the units to which treatments will be randomized.
3. Choose a response/output variable.
4. Determine potential sources of variation in response:
(a) factors of interest
(b) nuisance factors
5. Decide which variables to measure and control:
(a) treatment variables
(b) potential large sources of variation in the units (blocking variables)
6. Decide on the experimental procedure and how treatments are to be
randomly assigned.
The order of these steps may vary due to constraints such as budgets, ethics,
time, etc..

CHAPTER 1. PRINCIPLES OF EXPERIMENTAL DESIGN

Three principles of Experimental Design

1. Replication: Repetition of an experiment.
Replicates are runs of an experiment or sets of experimental units that
have the same values of the control variables.
More replication more precise inference
Let
yA,i = response of the ith unit assigned to treatment A
yB,i = response of the ith unit assigned to treatment B
i = 1, . . . , n.
Then yA 6= yB provides evidence that treatment affects response, i.e.
treatment is a source of variation, with the amount of evidence increasing with n.
2. Randomization: Random assignment of treatments to experimental units.
This removes potential for systematic bias on the part of the researcher,
and removes any preexperimental source of bias. Makes confounding
the effect of treatment with an unobserved variable unlikely (but not
impossible).
3. Blocking: Randomization within blocks of homogeneous experimental
units. The goal is to evenly distribute treatments across large potential
sources of variation.
Example (Crop study):
Hypothesis: Tomato type (A versus B) affects tomato yield.
Experimental units: three plots of land, each to be divided into a 2 2 grid.
Outcome: Tomato yield.
Factor of interest: Tomato type, A or B.
Nuisance factor : Soil quality.

bad soil

good soil

CHAPTER 1. PRINCIPLES OF EXPERIMENTAL DESIGN

Questions for discussion:

What are the benefits of this design?
What other designs might work?
What other designs wouldnt work?
Should the plots be divided up further? If so, how should treatments
then be assigned?

Chapter 2
Test statistics and
randomization distributions
Example: Wheat yield
Question: Is one fertilizer better than another, in terms of yield?
Outcome variable: Wheat yield.
Factor of interest: Fertilizer type, A or B. One factor having two levels.
Experimental material: One plot of land to be divided into 2 rows of 6
subplots each.

1. Design question: How should we assign treatments/factor levels to

the plots? We want to avoid confounding a treatment effect with another
potential source of variation.
2. Potential sources of variation: Fertilizer, soil, sun, water, etc.

CHAPTER 2. TEST STATISTICS AND RANDOMIZATION DISTRIBUTIONS10

3. Implementation: If we assign treatments randomly, we can avoid
any pre-experimental bias in results: 12 playing cards, 6 red, 6 black were
shuffled and dealt:
1st card black 1st plot gets B
2nd card red 2nd plot gets A
3rd card black 3rd plot gets B
..
.
This is the first design we will study, a completely randomized design.
4. Results:
B
A
B
A
B
B
26.9 11.4 26.6 23.7 25.3 28.5
B
A
A
A
B
A
14.2 17.9 16.5 21.1 24.3 19.6
How much evidence is there that fertilizer type is a source of yield variation?
Evidence about differences between two populations is generally measured by
comparing summary statistics across the two sample populations. (Recall, a
statistic is any computable function of known, observed data).

2.1

Summaries of sample populations

Distribution:

Empirical distribution: Pr(a,

b] = #(a < yi b)/n
Empirical CDF (cumulative distribution function)

F (y) = #(yi y)/n = Pr(,

y]
Histograms
Kernel density estimates
Note that these summaries more or less retain all the information in
the data except the unit labels.
Location:

CHAPTER 2. TEST STATISTICS AND RANDOMIZATION DISTRIBUTIONS11

sample mean or average : y =

1
n

Pn

i=1

yi

sample median : A/the value y.5 such that

#(yi y.5 )/n 1/2

#(yi y.5 )/n 1/2

To find the median, sort the data in increasing order, and call
these values y(1) , . . . , y(n) . If there are no ties, then
if n is odd, then y( n+1 ) is the median;
2

if n is even, then all numbers between y( n2 ) and y( n2 +1) are

medians.
Scale:
sample variance and standard deviation:
n

1 X
s =
(yi y)2 ,
n 1 i=1
2

s=

interquantile range:
[y.25 , y.75 ] (interquartile range)
[y.025 , y.975 ] (95% interval)
Example: Wheat yield
All of these sample summaries are easily obtained in R:
> yA<c ( 1 1 . 4 , 2 3 . 7 , 1 7 . 9 , 1 6 . 5 , 2 1 . 1 , 1 9 . 6 )
> yB<c ( 2 6 . 9 , 2 6 . 6 , 2 5 . 3 , 2 8 . 5 , 1 4 . 2 , 2 4 . 3 )

> mean (yA)

[ 1 ] 18.36667
> mean (yB)
[ 1 ] 24.3
> median (yA)
[ 1 ] 18.75
> median (yB)
[ 1 ] 25.95

s2

20
y

25

0.04
Density
0.02
0.00
10

15

20
y

25

30

F(y)
0.4 0.6

0.8

10

15

20

25

20
y

25

10

20
y

30

40

0.00

0.0

0.2

Density
0.00 0.10

yA

15

0.12

Density
0.00 0.08

1.0

15

Density
0.04
0.08

0.0

0.00

0.2

0.02

F(y)
0.4 0.6

Density
0.04

0.8

0.06

1.0

CHAPTER 2. TEST STATISTICS AND RANDOMIZATION DISTRIBUTIONS12

10

15

20
yB

25

30

10

Figure 2.1: Wheat yield distributions

15

20

25

30

35

CHAPTER 2. TEST STATISTICS AND RANDOMIZATION DISTRIBUTIONS13

> sd (yA)
[ 1 ] 4.234934
> sd (yB)
[ 1 ] 5.151699
> q u a n t i l e (yA , prob=c ( . 2 5 , . 7 5 ) )
25%
75%
16.850 20.725
> q u a n t i l e (yB , prob=c ( . 2 5 , . 7 5 ) )
25%
75%
24.550 26.825

So there is a difference in yield for these wheat fields.

Would you recommend B over A for future plantings?
Do you think these results generalize to a larger population?

2.2

Hypothesis testing via randomization

Questions:
Could the observed differences be due to fertilizer type?
Could the observed differences be due to plot-to-plot variation?
Hypothesis tests:
H0 (null hypothesis): Fertilizer type does not affect yield.
H1 (alternative hypothesis): Fertilizer type does affect yield.
A statistical hypothesis test evaluates the plausibility of H0 in light of the
data.
Suppose we are interested in mean wheat yields. We can evaluate H0 by
answering the following questions:
Is a mean difference of 5.93 plausible/probable if H0 is true?
Is a mean difference of 5.93 large compared to experimental noise?
To answer the above, we need to compare
{|
yB yA | = 5.93}, the observed difference in the experiment

CHAPTER 2. TEST STATISTICS AND RANDOMIZATION DISTRIBUTIONS14

to
values of |
yB yA | that could have been observed if H0 were true.
Hypothetical values of |
yB yA | that could have been observed under H0 are
referred to as samples from the null distribution.
Finding a null distribution: Let
g(YA , YB ) = g({Y1,A , . . . , Y6,A }, {Y1,B , . . . , Y6,B }) = |YB YA |.
This is a function of the outcome of the experiment. It is a statistic. Since
we will use it to perform a hypothesis test, we will call it a test statistic.
Observed test statistic: g(11.4, 23.7, . . . , 14.2, 24.3) = 5.93 = gobs
Hypothesis testing procedure: Compare gobs to g(YA , YB ) for values of YA
and YB that could have been observed, if H0 were true.
Recall the design of the experiment:
1. Cards were shuffled and dealt B, R, B, R, . . . and fertilizer types planted
in subplots:
B

2. Crops were grown and wheat yields obtained:

B
A
B
A
B
B
26.9 11.4 26.6 23.7 25.3 28.5
B
A
A
A
B
A
14.2 17.9 16.5 21.1 24.3 19.6
Now imagine re-doing the experiment in a universe where H0 : no treatment
effect is true:
1. Cards are shuffled and dealt B, R, B, B, . . . and wheat types planted in
subplots:

CHAPTER 2. TEST STATISTICS AND RANDOMIZATION DISTRIBUTIONS15

B

2. Crops are grown and wheat yields obtained:

B
A
B
B
A
A
26.9 11.4 26.6 23.7 25.3 28.5
A
B
B
A
A
B
14.2 17.9 16.5 21.1 24.3 19.6
Under this hypothetical treatment assignment,
(YA , YB ) = {11.4, 25.3, . . . , 21.1, 19.6}
|YB YA | = 1.07
This represents an outcome of the experiment in a universe where
The treatment assignment is B, A, B, B, A, A, A, B, B, A, A, B;
H0 is true.

IDEA: To consider what types of outcomes we would see in universes where

H0 is true, compute g(YA , YB ) under every possible treatment assignment
and assuming H0 is true.
Under our randomization scheme, there were
 
12
12!
=
= 924
6!6!
6
equally likely ways the treatments could have been assigned. For each one
of these, we can calculate the value of the test statistic that wouldve been
observed under H0 :
{g1 , g2 , . . . , g924 }
This enumerates all potential pre-randomization outcomes of our test statistic, assuming no treatment effect. Along with the fact that each treatment

0.00

0.00

Density
0.04
0.08

Density
0.10
0.20

0.12

CHAPTER 2. TEST STATISTICS AND RANDOMIZATION DISTRIBUTIONS16

10

0
YB YA

10

4
6
|YB YA|

Figure 2.2: Approximate randomization distribution for the wheat example

assignment is equally likely, these value give a null distribution, a probability
distribution of possible experimental results, if H0 is true.
F (x|H0 ) = Pr(g(YA , YB ) x|H0 ) =

#{gk x}
924

This distribution is sometimes called the randomization distribution, because it is obtained by the randomization scheme of the experiment.
Comparing data to the null distribution:
Is there any contradiction between H0 and our data?
Pr(g(YA , YB ) 5.93|H0 ) = 0.056
According to this calculation, the probability of observing a mean difference of 5.93 or more is unlikely under the null hypothesis. This probability
calculation is called a p-value. Generically, a p-value is
The probability, under the null hypothesis, of obtaining a result as or more
extreme than the observed result.

CHAPTER 2. TEST STATISTICS AND RANDOMIZATION DISTRIBUTIONS17

The basic idea:
small p-value evidence against H0
large p-value no evidence against H0
Approximating a randomization distribution:


B
We dont want to have to enumerate all nAn+n
possible treatment assignA
ments. Instead, repeat the following S times for some large number S:
(a) randomly simulate a treatment assignment from the population of possible treatment assignments, under the randomization scheme.
(b) compute the value of the test statistic, given the simulated treatment
assignment and under H0 .
The empirical distribution of {g1 , . . . , gS } approximates the null distribution:
#(gs gobs )
Pr(g(YA , YB ) gobs |H0 )
S
The approximation improves if S is increased.
Here is some R-code:
y<c ( 2 6 . 9 , 1 1 . 4 , 2 6 . 6 , 2 3 . 7 , 2 5 . 3 , 2 8 . 5 , 1 4 . 2 , 1 7 . 9 , 1 6 . 5 , 2 1 . 1 , 2 4 . 3 , 1 9 . 6 )
x<c ( B , A , B , A , B , B , B , A , A , A , B , A )
g . n u l l <r e a l ( )
for ( s in 1:10000)
{
xsim<sample ( x )
g . n u l l [ s]< abs ( mean ( y [ xsim==B ] ) mean ( y [ xsim==A ] ) )
}

2.3

Essential nature of a hypothesis test

Given H0 , H1 and data y = {y1 , . . . , yn }:

CHAPTER 2. TEST STATISTICS AND RANDOMIZATION DISTRIBUTIONS18

1. From the data, compute a relevant test statistic g(y): The test statistic
g(y) should be chosen so that it can differentiate between H0 and H1 in
ways that are scientifically relevant. Typically, g(y) is chosen so that

small under H0
g(y) is probably
large under H1
2. Obtain a null distribution : A probability distribution over the possible
outcomes of g(Y) under H0 . Here, Y = {Y1 , . . . , Yn } are potential
experimental results that could have happened under H0 .
3. Compute the p-value: The probability under H0 of observing a test
statistic g(Y) as or more extreme than the observed statistic g(y).
p-value = Pr(g(Y) g(y)|H0 )
If the p-value is small evidence against H0
If the p-value is large not evidence against H0
Even if we follow these guidelines, we must be careful in our specification of
H0 , H1 and g(Y) for the hypothesis testing procedure to be useful.
Questions:
Is a small p-value evidence in favor of H1 ?
Is a large p-value evidence in favor of H0 ?
What does the p-value say about the probability that the null hypothesis is true? Try using Bayes rule to figure this out.

2.4

Sensitivity to the alternative hypothesis

In the previous section we said that the test statistic g(y) should be able
to differentiate between H0 and H1 in ways that are scientifically relevant.
What does this mean?
Suppose our data consist of samples yA and yB from two populations A and
A |. Lets consider two different
B. Previously we used g(yA , yB ) = |
yB y
test statistics:

CHAPTER 2. TEST STATISTICS AND RANDOMIZATION DISTRIBUTIONS19

t-statistic:
|
y yA |
p B
, where
sp 1/nA + 1/nB
nB 1
nA 1
s2A +
s2
=
(nA 1) + (nB 1)
(nA 1) + (nB 1) B

gt (yA , yB ) =
s2p

This is a scaled version of our previous test statistic, in which we compare the difference in sample means to a pooled version of the sample
standard deviation and the sample size. Note that this statistic is
A |;
increasing in |
yB y
increasing in nA and nB ;
decreasing in sp .
A more complete motivation for using this statistic will be given in the
next chapter.
Kolmogorov-Smirnov statistic:
gKS (yA , yB ) = max |FB (y) FA (y)|
yR

This is just the size of the largest gap between the two sample CDFs.
Comparing the test statistics:
Suppose we perform a CRD and obtain samples yA and yB like those in
Figure 2.3. For these data,
nA = nB = 40
yA = 10.05, yB = 9.70.
sA = 0.87, sB = 2.07
The main difference between the two samples seems to be in their variances
and not in their means. Now lets consider evaluating
H0 : treatment does not affect response
using our two new test statistics. We can approximate the null distributions
of gt (YA , YB ) and gKS (YA , YB ) by randomly reassigning the treatments but
leaving the responses fixed:

10
yA

12

14

10
yB

12

14

F(y)

0.00

0.0

0.2

Density
0.10

0.4

0.6

0.0

0.8

1.0

Density
0.3
0.6

CHAPTER 2. TEST STATISTICS AND RANDOMIZATION DISTRIBUTIONS20

10
y

12

14

Figure 2.3: Histograms and empirical CDFs of the first two hypothetical
samples.
Gsim<NULL
for ( s in 1:5000)
{
xsim<sample ( x )
yAsim<y [ xsim==A ] ; yBsim<y [ xsim==B ]
g1< g . t s t a t ( yAsim , yBsim )
g2< g . ks ( yAsim , yBsim )
Gsim<r b i n d ( Gsim , c ( g1 , g2 ) )
}

These calculations give:

t-statistic :

gt (yA , yB ) = 1.00 , Pr(gt (YA , YB ) 1.00) = 0.321

KS-statistic: gKS (yA , yB ) = 0.30 , Pr(gKS (YA , YB ) 0.30) = 0.043

The hypothesis test based on the t-statistic does not indicate strong evidence
against H0 , whereas the test based on the KS-statistic does. The reason is
that the t-statistic is only sensitive to differences in means. In particular, if yA = yB then the t-statistic is zero, its minimum value. In contrast, the
KS-statistic is sensitive to any differences in the sample distributions.
Now lets consider a second dataset, shown in Figure 2.5, for which
nA = nB = 40

Density
3 4
0

0.0

0.2

Density
0.4

0.6

CHAPTER 2. TEST STATISTICS AND RANDOMIZATION DISTRIBUTIONS21

2
3
t statistic

0.1

0.2
0.3
KS statistic

0.4

Figure 2.4: Randomization distributions for the t and KS statistics for the
first example.
yA = 10.11, yB = 10.73.
sA = 1.75, sB = 1.85
The difference in sample means is about twice as large as in the previous example, and the sample standard deviations are pretty similar. The B-samples
are slightly larger than the A-samples on average. Is there evidence that this
is caused by treatment? Again, we evaluate H0 using the randomization
distributions of our two test statistics.
t-statistic :

gt (yA , yB ) = 1.54 , Pr(gt (YA , YB ) 1.54) = 0.122

KS-statistic: gKS (yA , yB ) = 0.25 , Pr(gKS (YA , YB ) 0.25) = 0.106

This time the two test statistics indicate similar evidence against H0 . This
is because the difference in the two sample distributions could primarily be
summarized as the difference between the sample means, which the t-statistic
can identify.

10

12

14

16

12

14

16

0.00

0.0

0.2

Density
0.15

0.4

yA

F(y)

0.6

0.00

0.8

1.0

Density
0.15 0.30

CHAPTER 2. TEST STATISTICS AND RANDOMIZATION DISTRIBUTIONS22

10

10

12

14

16

yB

Density
3 4
0

0.0

Density
0.2
0.4

0.6

Figure 2.5: Histograms and empirical CDFs of the second two hypothetical
samples.

2
3
t statistic

0.1

0.2 0.3 0.4

KS statistic

0.5

Figure 2.6: Randomization distributions for the t and KS statistics for the
second example.

CHAPTER 2. TEST STATISTICS AND RANDOMIZATION DISTRIBUTIONS23

Discussion: These last two examples suggest we should abandon gt in favor
of gKS if we are interested in comparing the following hypothesis:
H0 :
H1 :

treatment does not affect response

treatment does affect response

This is because, as we found, gt is not sensitive all violations of H0 , it is only

sensitive to violations of H0 where there is a difference in means. However,
in many situations we are actually interested in comparing the following
hypotheses:
H0 :
H1 :

treatment does not affect response

treatment increases responses or decreases responses

In this case H0 and H1 are not complementary, and we are only interested
in evidence against H0 of a certain type, i.e. evidence that is consistent with
H1 . In this situation we may want to use a statistic like gt .

2.5

Basic decision theory

Task: Accept or reject H0 based on data.

truth
action
accept H0
reject H0

H0 true
correct decision
type I error

H0 false
type II error
correct decision

As we discussed,
the p-value can measure of evidence against H0 ;
the smaller the p-value, the larger the evidence against H0 .
Decision procedure:
1. Compute the p-value by comparing observed test statistic to the null
distribution.
2. Reject H0 if the p-value , otherwise accept H0 .

CHAPTER 2. TEST STATISTICS AND RANDOMIZATION DISTRIBUTIONS24

This procedure is called a level- test. It controls the pre-experimental probability of a type I error, or for a series of experiments, controls the type I
error rate.
Pr(type I error|H0 ) = Pr(reject H0 |H0 )
= Pr(p-value |H0 )
=
Single Experiment Interpretation: If you use a level- test for your experiment where H0 is true, then before you run the experiment
there is probability that you will erroneously reject H0 .
Many Experiments Interpretation: If level- tests are used in a large
population of experiments, then H0 will be declared false in (100 )%
of the experiments in which H0 is true.

Pr(H0 rejected|H0 true)

Pr(H0 accepted|H0 true)
Pr(H0 rejected|H0 false)
Pr(H0 accepted|H0 false)

=
=
=
=

1
?
?

Pr(H0 rejected|H0 false) is the power. Typically we need to be more specific

than H0 false in order to calculate the power. We need to specify how it
is false.

Chapter 3
Tests based on population
models
3.1

Relating samples to populations

If the experiment is
complicated,
non- or partially randomized, or
includes nuisance factors
then a null distribution based on randomization may be difficult to obtain.
An alternative approach to hypothesis testing is based on formulating a sampling model.
Consider the following model for our wheat yield experiment:
There is a large/infinite population of plots of similar size/shape/composition as the plots in our experiment.
When A is applied to these plots, the distribution of plot yields can be
represented by a probability distribution pA with
R
expectation = E[YA ] = ypA (y)dy = A ,
variance =Var[YA ] = E[(YA A )2 ] = A2 .

25

CHAPTER 3. TESTS BASED ON POPULATION MODELS

26

When B is applied to these plots, the distribution of plot yields can be

represented by a probability distribution pB with
R
expectation = E[YB ] = ypB (y)dy = B ,
variance =Var[YB ] = E[(YB B )2 ] = B2 .
The plots that received A in our experiment can be viewed as independent samples from pA , and likewise for plots receiving B.
Y1,A , . . . , YnA ,A i.i.d. pA
Y1,B , . . . , YnB ,B i.i.d. pB

Recall from intro stats:

nA
1 X

E[YA ] = E[
Yi,A ]
nA i=1

nA
1 X
E[Yi,A ]
nA i=1

nA
1 X
=
A = A
nA i=1

We say that YA is an unbiased estimator of A . Furthermore, if Y1,A , . . . , YnA ,A

are independent samples from population A, then
1 X
Var[YA ] = Var[
Yi,A ]
nA
1 X
=
Var[Yi,A ]
n2A
1 X 2
=
A = A2 /nA .
n2A
This means that as nA ,
Var[YA ] = E[(YA A )2 ] 0
which, with unbiasedness, implies
YA A

CHAPTER 3. TESTS BASED ON POPULATION MODELS

0.10
0.06

0.06

random sampling

yA=18.37

0.04

0.04

sA=4.23

0.02

0.02

0.00

0.00

Experimental samples

0.08

0.08

All possible' A wheat yields

All possible' B wheat yields

10

15

20
yA

25

30

35

Experimental samples

0.08

random sampling

yB=24.30
sB=5.15

0.00

0.00

0.02

0.04

0.04

0.06

0.12

0.08

27

10

Figure 3.1: The population model

15

20
yB

25

30

35

CHAPTER 3. TESTS BASED ON POPULATION MODELS

28

and we say that YA is a consistent estimator for A . Several of our other

sample characteristics are also consistent for the corresponding population
characteristics. As n ,
YA A
s2A A2
#{Yi,A x}
= FA (x) FA (x) =
nA

pA (y)dy

Back to hypothesis testing:

We can formulate null and alternative hypotheses in terms of population
quantities. For example, if B > A , we would recommend B over A, and
vice versa.
H 0 : A = B
H1 : A 6= B
The experiment is performed and it is observed that
18.37 = yA < yB = 24.3
This is some evidence that A > B . How much evidence is it? Should we
reject the null hypothesis? Consider evaluating evidence against H0 with our
t-statistic:
|
yB yA |
gt (yA , yB ) = p
sp 1/nA + 1/nB
To decide whether to reject H0 or not, we need to know the distribution of
g(YA , YB ) under H0 . Consider the following setup:
Assume:
Y1,A , . . . , YnA ,A i.i.d. pA
Y1,B , . . . , YnB ,B i.i.d. pB
Evaluate: H0 : A = B versus H1 : A 6= B , i.e., whether or not
Z
Z
ypA (y)dy = ypB (y)dy.
To make this evaluation and obtain a p-value, we need the distribution of
g(YA , YB ) under A = B . This will involve assumptions about/approximations to pA and pB .

CHAPTER 3. TESTS BASED ON POPULATION MODELS

3.2

29

The normal distribution

The normal distribution is useful because in many cases,

our data are approximately normally distributed, and/or
our sample means are approximately normally distributed.
These are both due to the central limit theorem. Letting P (, ) denote a
population with mean and variance 2 , then

X1
P1 (1 , 12 )

m
X
X
X 
X2
P2 (2 , 22 )

normal

,
j2 .
..
i
j

j=1

2
Xm Pm (m , m
)
Sums of varying quantities are approximately normally distributed.
Normally distributed data
Consider crop yields from plots of land:
Yi = a1 seedi + a2 soili + a3 wateri + a4 suni +
The empirical distribution of crop yields from a population of fields with
varying quantities of seed, soil, water, sun, etc. will be approximately normal
(, ), where and depend on the effects a1 , a2 , a3 , a4 , . . . and the variability
of seed, soil, water, sun, etc..
Additive effects normally distributed data
Normally distributed means
Consider the following scenario:
(1)

(1)

(2)

(2)

Experiment 1: sample y1 , . . . , yn i.i.d. p and compute y(1) ;

Experiment 2: sample y1 , . . . , yn i.i.d. p and compute y(2) ;
..
.

CHAPTER 3. TESTS BASED ON POPULATION MODELS

(m)

(m)

Experiment m: sample y1 , . . . , yn

30

i.i.d. p and compute y(m) .

A histogram of {
y (1) , . . . , y(m) } will look approximately normally distributed
with
sample mean {y (1) , . . . , y (m) }
sample variance {y (1) , . . . , y (m) } 2 /n
i.e. the sampling distribution of the mean is approximately normal(, 2 /n),
even if the sampling distribution of the data are not normal.
Basic properties of the normal distribution:
Y normal(, 2 ) aY + b normal(a + b, a2 2 ).
Y1 normal(1 , 12 ), Y2 normal(2 , 22 ), Y1 , Y2 independent
Y1 + Y2 normal(1 + 2 , 12 + 22 )
if Y1 , . . . , Yn i.i.d. normal(, 2 ), then Y is statistically independent
of s2 .
How does this help with hypothesis testing?
Consider testing H0 : A = B (treatment doesnt affect mean). Then regardless of distribution of data, under H0 :

YA

YB

YB YA

normal(, A2 /nA )
normal(, B2 /nB )
2
normal(0, AB
)

2
where AB
= A2 /nA + B2 /nB . So if we knew the variances, wed have a null
distribution.

3.3

Introduction to the t-test

Consider a simple one-sample hypothesis test:

Y1 , . . . , Yn i.i.d. P , with mean and variance 2 .

CHAPTER 3. TESTS BASED ON POPULATION MODELS

31

H 0 : = 0
H1 : 6= 0
Examples:
Physical therapy
Yi = muscle strength after treatment - muscle score before.
H0 : E[Yi ] = 0
Physics
Yi = boiling point of a sample of an unknown liquid.
H0 : E[Yi ] = 100 C.
To test H0 , we need a test statistic and its distribution under H0 .
|
y 0 | might make a good test statistic:
it is sensitive to deviations from H0 .
its sampling distribution is approximately known:
E[Y ] =
Var[Y ] = 2 /n
Y is approximately normal.
Under H0

(Y 0 ) normal(0, 2 /n),

but we cant use this as a null distribution because 2 is unknown. What if

we scale (Y 0 )? Then
Y 0

f (Y) =
/ n
is approximately standard normal and we write f (Y) normal(0, 1). Since
this distribution contains no unknown parameters we could potentially use
it as a null distribution. However, having observed the data y, is f (y) a
statistic?
y is computable from the data and n is known;

CHAPTER 3. TESTS BASED ON POPULATION MODELS

32

0 is our hypothesized value, a fixed number that we have chosen.

is not determined by us and is unknown.
The solution to this problem is to approximate the population variance 2
with the sample variance s2 .
One-sample t-statistic:
t(Y) =

Y 0

s/ n

For a given value of 0 this is a statistic. What is the null distribution of

t(Y)?
Y 0
Y 0

s so
s/ n
/ n

0 is normal(0, 1), and so it would

If Y1 , . . . , Yn i.i.d. normal(0 , 2 ) then Y/
n
seem that t(Y) is approximately distributed as a standard normal distribution under H0 : = 0 . However, if the approximation s is poor, like
when n is small, we need to take account of our uncertainty in the estimate
of .

The 2 distribution
Z1 , . . . , Zn i.i.d. normal(0, 1)

Zi2 2n , chi-squared dist with n degrees of freedom

X
2 2n1
(Zi Z)
Y1 , . . . , Yn i.i.d. normal(, ) (Y1 )/, . . . , (Yn )/ i.i.d. normal(0, 1)
1 X
(Yi )2 2n

1 X
(Yi Y )2 2n1
2

Some intuition: Which vector do you expect to be bigger: (Z1 , . . . , Zn ) or

. . . , Zn Z)?
(Z1 Z,
. . . , Zn Z)
is a vector of length n but lies
(Z1 Z,
in an n 1 dimensional space. In fact, it is has a singular multivariate
normal distribution, with a covariance matrix of rank (n 1).

0.08

CHAPTER 3. TESTS BASED ON POPULATION MODELS

33

0.00

p (X )
0.04

n=9
n=10
n=11

10

15
X

20

25

30

Figure 3.2: 2 distributions

Getting back to the problem at hand,
n1 2 n1 1 X
s =
(Yi Y )2 2n1 ,
2
2 n 1
which is a known distribution we can look up on a table or with a computer.
The t-distribution
If
Z normal (0,1) ;
X 2m ;
Z, X statistically independent,
then
Z
p

X/m

tm , the t-distribution with m degrees of freedom

How does this help us? Recall that if Y1 , . . . , Yn i.i.d. normal(, 2 ),

34

0.3

0.4

CHAPTER 3. TESTS BASED ON POPULATION MODELS

0.0

0.1

p (t )
0.2

n=3
n=6
n=12
n=

0
t

Figure 3.3: t-distributions

n(Y )/ normal(0,1)

n1 2
s
2

2n1

Y , s2 are independent.

Let Z = n(Y )/, X =

n1 2
s.
2

Then

Z
n(Y )/
p
= q
n1 2
X/(n 1)
s /(n 1)
2
=

Y
tn1
s/ n

This is still not a statistic because is unknown. However, under a specific

hypothesis like H0 : = 0 , it is a statistic:
t(Y) =

Y 0
tn1
s/ n

It is called the t-statistic.

Some questions for discussion:

if E[Y ] = 0

CHAPTER 3. TESTS BASED ON POPULATION MODELS

35

What does X/m converge to as m ?

What happens to the distribution of t(Y) when n ? Why?
Consider the situation where the data are not normally distributed.

What is the distribution of n(Y )/ for large and small n?

What is the distribution of n1
s2 for large and small n?
2

Are n(Y ) and s2 independent for small n? What about for

large n?
Two-sided, one-sample t-test:
1. Sampling model: Y1 , . . . , Yn i.i.d. normal(, 2 )
2. Null hypothesis: H0 : = 0
3. Alternative hypothesis: H1 : 6= 0

4. Test statistic: t(Y) = n(Y 0 )/s

Pre-experiment we think of this as a random variable, an unknown
quantity that is to be randomly sampled from a population.
Post-experiment this is a fixed number.
5. Null distribution: Under the normal sampling model and H0 , the sampling distribution of t(Y) is the t-distribution with n 1 degrees of
freedom:
t(Y) tn1
If H0 is not true, then t(Y) does not have a t-distribution. If the data
are normal but the mean is not 0 , then t(Y) has a non-central tdistribution, which we will use later to calculate power, or the type II
error rate. If the data are not normal then the distribution of t(Y) is
not a t-distribution.

CHAPTER 3. TESTS BASED ON POPULATION MODELS

36

6. p-value: Let y be the observed data.

p-value =
=
=
=
=

Pr(|t(Y)| |t(y)||H0 )
Pr(|Tn1 | |t(y)|)
2 Pr(Tn1 |t(y)|)
2 (1 pt(tobs, n 1))
t.test(y, mu = mu0)

7. Level- decision procedure: Reject H0 if

p-value or equivalently
|t(y)| t(n1),1/2 (for = .05, t(n1),1/2 2 ).
The value t(n1),1/2 2 is called the critical value value for this test.
In general, the critical value is the value of the test statistic above
which we would reject H0 .
Question: Suppose our procedure is to reject H0 only when t(y) t(n1),1 .
Is this a level- test?

3.4

Two sample tests

Recall the wheat example:

B
A
B
A
B
B
26.9 11.4 26.6 23.7 25.3 28.5
B
A
A
A
B
A
14.2 17.9 16.5 21.1 24.3 19.6
Sampling model:
Y1A , . . . , YnA A
Y1B , . . . , YnB B

i.i.d. normal(A , 2 )
i.i.d. normal(B , 2 ).

In addition to normality we assume for now that both variances are equal.

CHAPTER 3. TESTS BASED ON POPULATION MODELS

HA : A 6= B

Hypotheses: H0 : A = B ;
Recall that
YB YA N

37




1
1
2
B A ,
+
.
nA nB

Hence if H0 is true then

YB YA N


0,

1
1
+
nA nB


.

How should we estimate 2 ?

P B
PnA
(yi,B yB )2
A )2 + ni=1
2
i=1 (yi,A y
sp =
(nA 1) + (nB 1)
nB 1
nA 1
s2A +
s2B
=
(nA 1) + (nB 1)
(nA 1) + (nB 1)
This gives us the following two-sample t-statistic:
t(YA , YB ) =

Y YA
qB
tnA +nB 2
sp n1A + n1B

Self-check exercises:
1. Show that (nA + nB 2)s2p / 2 2nA +nB 2 (recall how the 2 distribution was defined).
2. Show that the two-sample t-statistic has a t-distribution with nA +nB
2 d.f.

Numerical Example (wheat again)

Suppose we want to have a type-I error rate of = 0.05.
Decision procedure:
Level test of H0 : A = B , with = 0.05
Reject H0 if p-value < 0.05
Reject H0 if |t(y A , y B )| > t10,.975 = 2.23

CHAPTER 3. TESTS BASED ON POPULATION MODELS

38

Data:
yA = 18.36, s2A = 17.93, nA = 6
yB = 24.30, s2B = 26.54, nB = 6
t-statistic:
s2p = 22.24, sp = 4.72
p
t(yA , yB ) = 5.93/(4.72 1/6 + 1/6) = 2.18
Inference:
Hence the p-value= Pr(|T10 | 2.18) = 0.054
Hence H0 : A = B is not rejected at level = 0.05
> t . t e s t ( y [ x==A ] , y [ x==B ] , var . e q u a l=TRUE)
Two Sample tt e s t
data : y [ x == A ] and y [ x == B ]
t = 2.1793 , d f = 1 0 , pv a l u e = 0 . 0 5 4 3 1
a l t e r n a t i v e h y p o t h e s i s : t r u e d i f f e r e n c e i n means i s not e q u a l t o 0
95 p e r c e n t c o n f i d e n c e i n t e r v a l :
11.999621
0.132954
sample e s t i m a t e s :
mean o f x mean o f y
18.36667 24.30000

Always keep in mind where the p-value comes from: See Figure 3.4.
Comparison to the randomization test:
Recall that we have already compared the two-sample t-statistic to its randomization distribution. A sample from the randomization distribution were
obtained as follows:
1. Sample a treatment assignment according to the randomization scheme.
2. Compute the value of t(YA , YB ) under this treatment assignment and
assuming the null hypothesis.

39

0.0

0.1

p (T )
0.2

0.3

0.4

CHAPTER 3. TESTS BASED ON POPULATION MODELS

0
T

Figure 3.4: The t-distribution under H0 for the wheat example

The randomization distribution of the t-statistic is then approximated by
the empirical distribution of
t(1) , . . . , t(S)
To obtain the p-value, we compare tobs = t(y A , y B ) to the empirical distribution of t(1) , . . . , t(S) :
p-value =

#(|t(s) | |tobs |)
S

t . s t a t . obs<t . t e s t ( y [ x==A ] , y [ x==B ] , var . e q u a l=T) $ s t a t

t . s t a t . sim<r e a l ( )
for ( s in 1:10000)
{
xsim<sample ( x )
tmp<t . t e s t ( y [ xsim==B ] , y [ xsim==A ] , var . e q u a l=T)
t . s t a t . sim [ s]<tmp$stat
}
mean ( abs ( t . s t a t . sim ) >= abs ( t . s t a t . obs ) )

40

0.0

0.1

Density
0.2

0.3

0.4

CHAPTER 3. TESTS BASED ON POPULATION MODELS

0
t( Y A ,Y B )

Figure 3.5: Randomization and t-distributions for the t-statistic under H0

When I ran this, I got
#(|t(s) | 2.18)
= 0.058 0.054 = Pr(|TnA +nB 2 | 2.18)
S
Is this surprising? These two p-values were obtained via two completely
different ways of looking at the problem!
Assumptions: Under H0 ,
Randomization Test:
1. Treatments are randomly assigned
t-test:
1. Data are independent samples
2. Each population is normally distributed
3. The two populations have the same variance
Imagined Universes:
Randomization Test: Numerical responses remain fixed, we
imagine only alternative treatment assignments.

CHAPTER 3. TESTS BASED ON POPULATION MODELS

41

t-test: Treatment assignments remain fixed, we imagine an

alternative sample of experimental units and/or conditions, giving
different numerical responses.
Inferential Context / Type of generalization
Randomization Test: inference is specific to our particular experimental units and conditions.
t-test: under our assumptions, inference claims to be generalizable to other units / conditions, i.e. to a larger population.
Yet the numerical results are often nearly identical.
Keep the following concepts clear:
t-statistic : a scaled difference in sample means, computed from the data.
t-distribution : the probability distribution of a normal random variable
divided by the square-root of a 2 random variable.
t-test : a comparison of a t-statistic to a t-distribution
randomization distribution : the probability distribution of a test statistic under random treatment reassignments and H0
randomization test : a comparison of a test statistic to its randomization
distribution
randomization test with the t-statistic : a comparison of the t-statistic
to its randomization distribution
Some history:
de Moivre (1733): Approximating binomial distributions T =
{0, 1}.

Pn

i=1

Yi , Y i

Laplace (1800s): Used normal distribution to model measurement error in

experiments.
Gauss (1800s) : Justified least squares estimation by assuming normally
distributed errors.

CHAPTER 3. TESTS BASED ON POPULATION MODELS

42

Gosset/Student (1908): Derived the t-distribution.

Gosset/Student (1925): testing the significance
Fisher (1925): level of significance
Fisher (1920s?): Fishers exact test.
Fisher (1935): It seems to have escaped recognition that the physical act of
randomisation, which, as has been shown, is necessary for the validity of
any test of significance, affords the means, in respect of any particular
body of data, of examining the wider hypothesis in which no normality
of distribution is implied.
Box and Andersen (1955): Approximation of randomization null distributions (and a long discussion).
Rodgers (1999): If Fishers ANOVA had been invented 30 years later or
computers had been available 30 years sooner, our statistical procedures
would probably be less tied to theoretical distributions as what they
are today

3.5

Checking assumptions

For our t-statistic

t(YA , YB ) =

Y YA
qB
sp n1A + n1B

we showed that if Y1,A , . . . , YnA ,A and Y1,B , . . . , YnB ,B are independent samples
from pA and pB respectively, and
(a) A = B
(b) A2 = B2
(c) pA and pB are normal distributions
then
t(YA , YB ) tnA +nB 2
So our null distribution really assumes conditions (a), (b) and (c). Thus if
we perform a level- test and reject H0 , we are really just rejecting that (a),
(b), (c) are all true.

CHAPTER 3. TESTS BASED ON POPULATION MODELS

43

one of (a), (b) or (c) is not true

reject H0

For this reason, we will often want to check if conditions (b) and (c) are
plausibly met. If
(b) is met
(c) is met
H0 is rejected, then
this is evidence that H0 is rejected because A 6= B .

3.5.1

Checking normality

Normality can be checked by a normal probability plot

Idea: Order observations within each group:
y(1),A y(2),A y(nA ),A
compare these sample quantiles to the quantiles of a standard normal
distribution:
z 1 1 z 2 1 z nA 1
nA

here Pr(Z z

k
12
nA

)=

k
nA

nA

1
.
2

The

12

nA

is a continuity correction.

If data are normal, the relationship should be approximately linear. Thus

we plot the pairs
(z

1
12
nA

, y(1),A ), . . . , (z nA 1 , y(nA ),A ).

nA

Normality may be checked roughly by fitting straight lines to the probability

plots and examining their slopes

3.5.2

Unequal variances

For now, we will use the following rule of thumb:

If 1/4 < s2A /s2B < 4, we wont worry too much about unequal variances.

1.0
0.0
1.0
Theoretical Quantiles
Sample Quantiles
0.5 0.0 0.5

Sample Quantiles
0.0 0.5 1.0

1.0
0.0
1.0
Theoretical Quantiles

1.0
0.0
1.0
Theoretical Quantiles

1.0
0.0
1.0
Theoretical Quantiles

1.0
0.0
1.0
Theoretical Quantiles

1.0
0.0
1.0
Theoretical Quantiles

1.0
0.0
1.0
Theoretical Quantiles

44

Sample Quantiles
0.0 0.5 1.0 1.5

1.0
0.0
1.0
Theoretical Quantiles

Sample Quantiles
1 0
1
2

1.0
0.0
1.0
Theoretical Quantiles

Sample Quantiles
0.5 0.5 1.5 2.5

Sample Quantiles
1.0 0.0
1.0

1.0
0.0
1.0
Theoretical Quantiles
Sample Quantiles
1.0 0.0 1.0 2.0

1.0
0.0
1.0
Theoretical Quantiles

1.0
0.0
1.0
Theoretical Quantiles

1.0
0.0
1.0
Theoretical Quantiles

Figure 3.6: Normal scores plots.

Sample Quantiles
1.5 0.5 0.5

1.0
0.0
1.0
Theoretical Quantiles
Sample Quantiles
2.0 1.0 0.0
1.0

Sample Quantiles
2.0
0.5 0.5

1.0
0.0
1.0
Theoretical Quantiles

Sample Quantiles
1.5
0.5
0.5

Sample Quantiles
1.0
0.0

Sample Quantiles
1.0
0.0 0.5

Sample Quantiles
0.5
0.5
1.5

Sample Quantiles
18 22 26

14

Sample Quantiles
12
16
20
24

CHAPTER 3. TESTS BASED ON POPULATION MODELS

1.0
0.0
1.0
Theoretical Quantiles

CHAPTER 3. TESTS BASED ON POPULATION MODELS

45

This may not sound very convincing. In later sections, we will show how
to perform formal hypothesis tests for equal variances. However, this wont
completely solve the problem. If variances do seem unequal we have a variety
of options available:
use the randomization null distribution;
transform the data to stabilize the variances (to be covered later);
use a modified t-test that allows unequal variance.
The modified t-statistic is
yB yA
tw (y A , y B ) = q 2
sA
s2
+ nBB
nA
This statistic looks pretty reasonable, and for large nA and nB its null distribution will indeed be a normal(0, 1) distribution. However, the exact null
distribution is only approximately a t-distribution, even if the data are
actually normally distributed. The t-distribution we compare tw to is a tw distribution, where the degrees of freedom w are given by
2
 2
s2B
sA
+
nA
nB
w =
 2 2
 2 2 .
sA
s
1
1
+ nB 1 nBB
nA 1 nA
This is known as Welchs approximation; it may not give an integer as the
degrees of freedom.
This t-distribution is not, in fact the exact sampling distribution of tdiff (yA , yB )
under the null hypothesis that A = B , and A2 6= B2 . This is because the
null distribution depends on the ratio of the unknown variances, A2 and B2 .
This difficulty is known as the Behrens-Fisher problem.
Which two-sample t-test to use?
If the sample sizes are the same (nA = nB ) then the test statistics
tw (y A , y B ) and t(y A , y B ) are the same; however the degrees of freedom
used in the null distribution will be different unless the sample standard
deviations are the same.

CHAPTER 3. TESTS BASED ON POPULATION MODELS

46

If nA > nB , but A2 < B2 , and A = B then the two sample test based
on comparing t(y A , y B ) to a t-distribution on nA + nB 2 d.f. will
reject more than 5% of the time.
If the null hypothesis that both the means and variances are equal,
i.e.
H0 : A = B and A2 = B2
is scientifically relevant, then we are computing a valid p-value,
and this higher rejection rate is a good thing! Since when the
variances are unequal the null hypothesis is false.
If however, the hypothesis that is most scientifically relevant is
H0 : A = B
without placing any restrictions on the variances, then the higher
rejection rate in the test that assumes the variances are the same
could be very misleading, since p-values may be smaller than they
are under the correct null distribution (in which A2 6= B2 ).
Likewise we will underestimate the probability of type I error.
If nA > nB and A2 > B2 , then the p-values obtained from the test using
t(y A , y B ) will tend to be conservative (= larger) than those obtained
with tw (y A , y B ).
In short: one should be careful about applying the test based on t(yA , yB ) if
the sample standard deviations appear very different, and it is not reasonable
to assume equal means and variances under the null hypothesis.

Chapter 4
Confidence intervals and power
4.1

Confidence intervals via hypothesis tests

Recall that
H0 : E[Y ] = 0 is rejected if

n|(
y 0 )/s| t1/2
H0 : E[Y ] = 0 is not rejected if

n|(
y 0 )/s| t1/2
1
|
y 0 | s t1/2
n
s
s
y t1/2 0 y + t1/2
n
n
If 0 satisfies this last line, then it is in the acceptance region. Otherwise it
is in the rejection region. In other words, plausible values of are in the
interval
s
y t1/2
n
We say this interval is a 100 (1 )% confidence interval for . This
interval contains only those values of that are not rejected by this level-
test.

47

CHAPTER 4. CONFIDENCE INTERVALS AND POWER

48

Main property of a confidence interval

Suppose you are going to
1. gather data;
2. compute a 100 (1 )% confidence interval.
Further suppose H0 : E[Y ] = 0 is true. What is the probability that 0
will be in your to-be-sampled (random) interval? In other words, what is the
probability that the random interval will contain the true value?

Pr(0 in interval|E[Y ] = 0 ) =
=
=
=

1 Pr(0 not in interval|E[Y ] = 0 )

1 Pr(reject H0 |E[Y ] = 0 )
1 Pr(reject H0 |H0 is true)
1

The quantity 1 is called the coverage probability. It is

the pre-experimental probability that your confidence interval will cover
the true value;
the large sample fraction of experiments in which the confidence interval
covers the true mean.
Confidence interval for a difference between treatments
Recall that we may construct a 95% confidence interval by finding those null
hypotheses that would not be rejected at the 0.05 level.
Sampling model:
Y1,A , . . . , YnA ,A i.i.d. normal(A , 2 )
Y1,B , . . . , YnB ,B i.i.d. normal(B , 2 ).
Consider evaluating whether is a reasonable value for the difference in
means:
H0 : B A =
H1 : B A 6=

CHAPTER 4. CONFIDENCE INTERVALS AND POWER

49

Under H0 , you should be able to show that

(YB YA )
p
tnA +nB 2
sp 1/nA + 1/nB
Thus a given difference is accepted at level if
|
y yA |
pB
tc
sp 1/nA + 1/nB
r
r
1
1
1
1
(
yB yA ) sp
+
tc (
yB yA ) + sp
+
tc
nA nB
nA nB
where tc = t1/2,nA +nB 2 is the critical value.
Wheat example:
yB yA = 5.93
p
sp = 4.72, sp 1/nA + 1/nB = 2.72
t.975,10 = 2.23
A 95% C.I. for B A is
5.93 2.72 2.23
5.93 6.07 = (0.13, 11.99)
Questions:
What does the fact that 0 is in the interval say about H0 : A = B ?
What is the interpretation of this interval?
Could we have constructed an interval via a randomization test?

4.2

Power and Sample Size Determination

Suppose that we are designing a study where we intend to gather data on

two groups, and will decide if there is a difference between the groups based
on the data. Further suppose our decision procedure will be determined by
a level- two-sample t-test.

CHAPTER 4. CONFIDENCE INTERVALS AND POWER

50

Two sample experiment and t-test:

H 0 : A = B

H1 : A 6= B

Randomize treatments to the two groups via a CRD.

Gather data.
Perform a level hypothesis test: reject H0 if
|tobs | t1/2,nA +nB 2 .
Recall, if = 0.05 and nA , nB are large then t1/2,nA +nB 2 2.
We know that the type I error rate is = 0.05, or more precisely:
Pr(type I error|H0 true) = Pr(reject H0 |H0 true) = 0.05
What about
Pr(type II error|H0 false) = Pr(accept H0 |H0 false)
= 1 Pr(reject H0 |H0 false)
This is not yet a well-defined problem: there are many different ways in which
the null hypothesis may be false, e.g. B A = 0.0001 and B A = 10, 000
are both instances of the alternative hypothesis. However, clearly we have
Pr(reject H0 |B A = .0001) < Pr(reject H0 |B A = 10, 000)
if the treatment doesnt affect the variance.
To make the question concerning Type II error-rate better defined we need to
be able to refer to a specific alternative hypothesis. For example, in the case
of the two-sample test, for a specific difference we may want to calculate:
1 Pr(type II error|B A = ) = Pr(reject H0 |B A = ).
We define the power of a two-sample t-test test under a specific alternative to be:
Power() = Pr(reject H0 | B A = )

= Pr( |t(Y A , Y B )| t1/2,nA +nB 2  B A = ).

CHAPTER 4. CONFIDENCE INTERVALS AND POWER

51

Remember, the critical value t1/2,nA +nB 2 above which we reject the null
hypothesis was computed from the null distribution.
However, now we want to work out the probability of getting a value of
the t-statistic greater than this critical value, when a specific alternative
hypothesis is true. Thus we need to compute the distribution of our tstatistic under the specific alternative hypothesis.
If we suppose Y1,A , . . . , YnA ,A i.i.d. normal(A , 2 ) and Y1,B , . . . , YnB ,B
i.i.d. normal(B , 2 ), where B A = then to calculate the power we need
to know the distribution of
YB YA
.
t(Y A , Y B ) = q
sp n1A + n1B
We know that if B A = then
YB YA
q
tnA +nB 2
sp n1A + n1B
but unfortunately
t(Y A , Y B ) =

YB YA

q
+ q
sp n1A + n1B
sp n1A +

(4.1)

1
nB

The first part in the above equation has a t-distribution, which is centered
around zero. The second part moves the t-statistic away from zero by an
amount that depends on the pooled sample variance. For this reason, we
call the distribution of the t-statistic under B A = the non-central
t-distribution. In this case, we write

t(Y A , Y B ) tnA +nB 2 q

n1A + n1B
{z
}
|
non-centrality
parameter.

Note that this distribution is more complicated than just a t-distribution

plus a constant shift away from zero. For the t-statistic, the amount of the
shift depends on the (random) pooled sample variance.

52

0.4

CHAPTER 4. CONFIDENCE INTERVALS AND POWER

0.0

0.1

0.2

0.3

=0
=1
=2

2
t

Figure 4.1: A t10 distribution and two non-central t10 -distributions.

4.2.1

The non-central t-distribution

A noncentral t-distributed random variable can be represented as

Z +
T =p
X/
where
is a constant;
Z is standard normal;
X is 2 with degrees of freedom, independent of Z.
The quantity is called the noncentrality parameter.
Exercise: Using the above representation, show that the distribution of
the t-statistic is a non-central t distribution, assuming the data are normal
and the variance is the same in both groups.
For a non-central t-distribution,
the mean is not zero;

CHAPTER 4. CONFIDENCE INTERVALS AND POWER

53

the distribution is not symmetric.

It can be shown that
( 1
)
2

( 2 )

E[t(Y A , Y B )|B A = ] = q
n1A +

1
nB

where = nA + nB 2, the degrees of freedom, and (x) is the gamma

function, a generalization of the factorial:
(n + 1) = n! if n is an integer
(r + 1) = r(r)
(1) = 1, ( 21 ) =

Anyway, you can show that for large ,

r
1

(
)/( ) 1 so
2
2
2

E[t(Y A , Y B )|B A = ]

1
nA

1
nB

This isnt really such a big surprise, because we know that:

YB YA normal(, 2 [1/nA + 1/nB ]).
Hence

Y YA

qB
normal q
n1A + n1B
n1A +

1
nB

, 1 .

We also know that for large values of nA , nB , we have s , so the noncentral t-distribution will (for large enough nA , nB ) look approximately normal with
p
mean /( (1/nA ) + (1/nB ));
standard deviation 1.

CHAPTER 4. CONFIDENCE INTERVALS AND POWER

54

Another way to get the same result is to refer back to the expression for the
t-statistic given in 4.1:
t(Y A , Y B ) =

YB YA

q
+ q
sp n1A + n1B
sp n1A +

anA ,nb

1
nB

bnA ,nB

The first term anA ,nB has a t-distribution, and becomes standard normal as
nA , nB . As for bnA ,nB , since s2p 2 as nA or nB , we have
1
bnA ,nB

4.2.2

p
1 as nA , nB .
sp 1/nA + 1/nB

Computing the Power of a test

Recall our level- testing procedure using the t-test:

1. Sample data, compute tobs = t(Y A , Y B ) .
2. Compute the p-value, Pr(|TnA +nB 2 | > |tobs |).
3. Reject H0 if the p-value |tobs | t1/2,nA +nB 2 .
For this procedure, we have shown that
Pr(reject H0 |B A = 0) =
=
=
=

Pr( p-value |B A = 0)
Pr(|t(Y A , Y B )| t1/2,nA +nB 2 |H0 )
Pr(|TnA +nB 2 | t1/2,nA +nB 2 )

But what is the probability of rejection under H : B A = ? Hopefully

this is bigger than ! Let tc = t1/2,nA +nB 2 , the 1 /2 quantile of a
t-distribution with nA + nB 2 degrees of freedom.
Pr(reject H0 | B A = ) =
=
=
=

Pr ( |t(Y A , Y B )| > tc | B A = )
Pr (|T | > tc )
Pr (T > tc ) + Pr (T < tc )
[1 Pr (T < tc )] + Pr (T < tc )

55

0.4

CHAPTER 4. CONFIDENCE INTERVALS AND POWER

0.0

0.1

0.2

0.3

=0
=1

0
t

Figure 4.2: Critical regions and the non-central t-distribution

where T has the non-central t-distribution with = nA + nB 2 degrees of
freedom and non-centrality parameter

= q
n1A +

.
1
nB

We will want to make this calculation in order to see if our sample size is
sufficient to have a reasonable chance of rejecting the null hypothesis. If we
have a rough idea of and 2 we can evaluate the power using this formula.
t . c r i t < qt ( 1a l p h a /2 , nA + nB 2 )
t . gamma< d e l t a / ( sigma s q r t ( 1 /nA + 1/nB ) )
t . power < 1 pt ( t . c r i t , nA+nB2 , ncp=t . gamma )
pt( t . c r i t , nA+nB2 , ncp=t . gamma )

When you do these calculations you should think of Figure 4.2. Letting T
and T be non-central and central t-distributed random variables respectively,
make sure you can relate the following probabilities to the figure:
Pr(T > tc )

CHAPTER 4. CONFIDENCE INTERVALS AND POWER

56

Pr(T < tc )
Pr(T > tc )
Pr(T < tc )
Note that if the power Pr(|T | > tc ) is large, then one of Pr(T > tc ) or
Pr(T < tc ) will be very close to zero.
Approximating the power
Recall that for large nA , nB ,
.

t(Y A , Y B ) normal(, 1)
The normal approximation to the power is thus given by
Pr(|X| > tc ) = [1 Pr(X < tc )] + Pr(X < tc )
where X normal(, 1). This can be computed in R as
t . norm . power <

1 pnorm ( t . c r i t , mean=t . gamma ) +

pnorm( t . c r i t , mean=t . gamma )

This will be a reasonable approximation for large nA , nB . It may be an overestimate or under-estimate of the power obtained from the t-distribution.
Finally, keep in mind that in our calculations we have assumed that the
variances of the two populations are equal.
Example (selecting a sample size): Suppose the wheat researchers wish
to redo the experiment using a larger sample size. How big should their sample size be if they want to have a good chance of rejecting the null hypothesis
B A = 0 at level = 0.05, if the true difference in means is B A = 5
or more?
B A = 5
2 is unknown: Well assume the pooled sample variance from the first
experiment is a good approximation: 2 = 22.24.

15

20

25

power
normal approx

10

57

1.0

0.4

2.0

power
0.6
0.8

2.5

3.0

3.5

4.0

CHAPTER 4. CONFIDENCE INTERVALS AND POWER

30

10

15

20

25

30

Figure 4.3: and power versus sample size, and the normal approximation
to the power.
Under these conditions, if nA = nB = n, then
=

B A

1/nA + 1/nB

5
p
=
= .75 n
4.72 2/n

What is the probability well reject H0 at level = 0.05 for a given sample
size?
d e l t a <5 ; s2< (

(nA1) var (yA) + (nB1) var (yB) ) / ( nA1+nB1)

alpha <0.05 ; n<s e q ( 6 , 3 0 )

t . c r i t < qt (1 a l p h a / 2 , 2 n2)
t . gamma<d e l t a / s q r t ( s 2 ( 1 / n+1/n ) )
t . power<1pt ( t . c r i t , 2 n2, ncp=t . gamma)+
pt( t . c r i t , 2 n2, ncp=t . gamma)
t . normal . power< 1 pnorm ( t . c r i t , mean=t . gamma ) +
pnorm( t . c r i t , mean=t . gamma )

CHAPTER 4. CONFIDENCE INTERVALS AND POWER

58

So we see that if the true mean difference were B A = 5, then the original
study only had about a 40% chance of rejecting H0 . To have an 80% chance
or greater, the researchers would need a sample size of 15 for each group.
Note that the true power depends on the unknown true mean difference and
true variance (assuming these are equal in the two groups). Even though our
power calculations were done under potentially inaccurate values of B A
and 2 , they still give us a sense of the power under various parameter values:
How is the power affected if the mean difference is bigger? smaller?
How is the power affected if the variance is bigger? smaller?
Example (power as a function of the effect): Suppose a chemical
company wants to know if a new procedure B will yield more product than
the current procedure A. Running experiments comparing A to B are expensive and they are only budgeted to run an experiment with at most 10
observations in each group.
Is running the experiment worthwhile? To assess this we can calculate the
power under nA = nB = 10 for a variety of values of B A and . The
first panel plots power as a function of the mean difference for three different
values of . From this plot, we can see that if the mean difference is 1 and
the variance is 1, then we have almost a 60% chance of rejecting the null
hypothesis, although we only have about a 23% chance of doing so if the
variance is 9 ( = 3).
Because the power varies as the ratio of effect size to the standard deviation,
it is often useful to plot power in terms of this ratio. The scaled effect size
, where
= (B A )/,
represents the size of the treatment effect scaled by the experimental variability (the standard deviation). The noncentrality parameter is then
p
= / 1/nA + 1/nB .
With nA = nB = 10, we have = 2.24 . A plot of power versus for a
level-0.05 test appears in the first panel of Figure 4.4. From this we see that
H0 will be rejected with probability 80% or more only if || is bigger than
about 1.33. In other words, for a sample size of 10 in each group, the effect
must be at least 1.33 times as big as the standard deviation in order to have
an 80% chance of rejecting H0 .

1.0

59

power
0.4 0.6

power
0.4 0.6

0.8

0.8

1.0

CHAPTER 4. CONFIDENCE INTERVALS AND POWER

0.2

0.0

0.2

=1
=2
=3
4

0
B A

1
0
1
(
B A)/

Figure 4.4: Null and alternative distributions for another wheat example,
and power versus sample size.
Increasing power
As weve seen by the normal approximation to the power, for a fixed type I
error rate the power is a function of the noncentrality parameter
B A
,
= p
1/nA + 1/nB
so clearly power is
increasing in |B A |;
increasing in nA and nB ;
decreasing in 2 .
The first of these we do not generally control with our experiment (indeed, it
is the unknown quantity we are trying to learn about). The second of these,
sample size, we clearly do control. The last of these, the variance, seems
like something that might be beyond our control. However, the experimental
variance can often be reduced by dividing up the experimental material into
more homogeneous subgroups of experimental units. This design technique,
known as blocking, will be discussed in an upcoming chapter.

Chapter 5
Introduction to ANOVA
Example (Response times):
Background: Psychologists are interested in how learning methods affect
short-term memory.
Hypothesis: Different learning methods may result in different recall time.
Treatments: 5 different learning methods (A, B, C, D, E).
Experimental design: (CRD) 20 male undergraduate students were randomly assigned to one of the 5 treatments, so that there are 4 students
assigned to each treatment. After a learning period, the students were
given cues and asked to recall a set of words. Mean recall time for each
student was recorded in seconds.
Results:
Treatment
A
B
C
D
E

sample mean sample sd

6.88
0.76
5.41
0.84
6.59
1.37
5.46
0.41
5.64
0.83

Question: Is treatment a source of variation?

60

CHAPTER 5. INTRODUCTION TO ANOVA

61

response time (seconds)

5
6
7
8

C
treatment

Figure 5.1: Response time data

Possible data analysis method: Perform t-tests of
H0i1 i2 : i1 = i2 versus H1i1 i2 : i1 6= i2


for each of the 52 = 10 possible pairwise comparisons. Reject a hypothesis
if the associated p-value .
Problem: If there is no treatment effect at all, then
Pr(reject H0i1 i2 |H0i1 i2 true) =
Pr(reject any H0i1 i2 | all H0i1 i+2 true ) = 1 Pr( accept all H0i1 i2 | all H0i1 i2 true )
Y
1 (1 )
i<j

If = 0.05, then
Pr(reject one or more H0i1 i2 | all H0i1 i2 true ) 1 .9510 = 0.40
So, even though the pairwise error rate is 0.05 the experiment-wise error rate
is about 0.40. This issue is called the problem of multiple comparisons and
will be discussed further in Chapter 6. For now, we will discuss a method of
testing the global hypothesis of no variation due to treatment:
H0 : i1 = i2 for all i1 , i2

versus

H1 : i1 6= i2 for some i1 , i2

CHAPTER 5. INTRODUCTION TO ANOVA

62

To do this, we will compare

treatment variability: variability across treatments, to
experimental variability: variability among experimental units.
First we need to have a way of quantifying these things.

5.1

A model for treatment variation

Data:
yij
i
j

= measurement from the jth replicate under the ith treatment.

= 1, . . . , m indexes treatments
= 1, . . . , n indexes observations or replicates.

Treatment means model:

yij = i + ij
E[ij ] = 0
Var[ij ] = 2
i is the ith treatment mean,
ij s represent within treatment variation, also known as error or noise.
Treatment effects model:
yij = + i + ij
E[ij ] = 0
Var[ij ] = 2
is the grand mean;
1 , . . . , m are the treatment effects, representing between treatment variation
ij s still represent within treatment variation.

CHAPTER 5. INTRODUCTION TO ANOVA

In this model, we typically restrict
rameterized.

63

i = 0, otherwise the model is overpa-

The treatment means and treatment effects models represent two parameterizations of the same model:
i = + i i = i
Null (or reduced) model:
yij = + ij
E[ij ] = 0
Var[ij ] = 2
This is a special case of the above two models with
= 1 = m , or equivalently
i = 0 for all i.
In this model, there is no variation due to treatment.

5.1.1

Model Fitting

What are good estimates of the parameters? One criteria used to evaluate
different values of = {1 , . . . , m } is the least squares criterion:
SSE() =

m X
n
X

(yij i )2

i=1 j=1

The value of that minimizes SSE() is called the least-squares estimate,

and will be denoted .

Estimating treatment means: SSE is the sum of a bunch of quadratic
satisfies
terms, and so is a convex function of . The global minimizer

CHAPTER 5. INTRODUCTION TO ANOVA

64

= 0. Taking derivatives, we see that

SSE()
n

X
SSE() =
(yij i )2
i
i j=1
X
= 2
(yij i )
= 2n(
yi i ), so
SSE() = 2n(
y )
=y
= {
Therefore, the global minimum occurs at
y1 , . . . , yt }.
provides a measure of experimental variability:
Interestingly, SSE()
XX
XX
=
SSE()
(yij
i )2 =
(yij yi )2
P
Recall s2i = (yij yi )2 /(n 1) estimates 2 using data from group i. If we
have more than one group, we want to pool our estimates to be more precise:

=
=
=
=

(n 1)s21 + + (n 1)s2m
(n 1) + + (n 1)
P
P
(y1j y1 )2 + + (y1j y1 )2
m(n 1)
PP
(yij
i )2
m(n 1)

SSE()
MSE
m(n 1)

and s2 have various interpretations depending on the assumpThe values

tions we are willing to make. Consider the following assumptions:
A0: Data are independently sampled from their respective populations
A1: Populations have the same variance
A2: Populations are normally distributed
Then
is an unbiased estimator of
A0

CHAPTER 5. INTRODUCTION TO ANOVA

65

A0+A1 s2 is an unbiased estimator of 2

A0+A1+A2
s2 ) are the minimum variance unbiased estimators of (, 2 )
(,
n1
(,
s2 ) are the maximum likelihood estimators of (, 2 )
n
Within-treatment variability:
SSE is a measure of within treatment variability:
SSE()
SSE =

m X
n
X

(yij yi )2

i=1 j=1

It is the sum of squared deviation of observation values from their group

mean.
1
MSE = SSE/m(n 1) = (s21 + + s2m )
m
Clearly MSE is a measure of average (or mean) within-treatment variability.
Between-treatment variability:
The analogous measure of between treatment variability is
m X
n
m
X
X
2
SST =
(
yi y ) = n
(
yi y )2
i=1 j=1

i=1

where
1 XX
yij
mn
1
=
(
y1 + + ym )
m
is the grand mean of the sample. We call SST the treatment sum of squares.
We also define MST = SST/(m 1) as the treatment mean squares or mean
squares (due to) treatment. Notice that MST is simply n times the sample
variance of the sample means:
"
#
m
1 X
MST = n
(
yi y )2
m 1 i=1
y =

CHAPTER 5. INTRODUCTION TO ANOVA

5.1.2

66

Testing hypothesis with MSE and MST

Consider evaluating the null hypothesis of no treatment effect:

H0 : {1 , . . . , m } all equal versus H1 : {1 , . . . , m } not all equal
Note that
{1 , . . . , m } not all equal

m
X
(i
)2 > 0
i=1

Probabilistically,
m
m
X
X
2
(i
) > 0 a large
(
yi y )2 will probably be observed.
i=1

i=1

Inductively,
a large

(
yi y )2 observed

i=1m

(i
)2 > 0 is plausible

i=1m

So a large value of SST or MST gives evidence that there are differences
between the true treatment means. But how large is large? We need to
know what values of MST to expect under H0 .
MST under the null:
Suppose H0 : 1 = = m = is true . Then

E[(Yi ] = E[ nYi ] = n

Var[Yi ] = 2 /n Var[ nYi ] = 2

So under the null nY1 , . . . , nYm are m independent random variables

having the same mean and variance. Recall that if X1 , . . . , Xn i.i.d. P ,
then an unbiasedPestimate of the variance of population P is given by the
2 /(n 1). Therefore,
sample variance (Xi X)
P 2

( nYi nY )
is an unbiased estimate of Var[ nYi ] = 2 .
m1

CHAPTER 5. INTRODUCTION TO ANOVA

67

Notice that
P 2
P
( nYi nY )
n (Yi Y )2
=
m1
m1
SST
=
m1
= MST,
so E[MST|H0 ] = 2 .
MST under an alternative:
We can show that under a given value of ,
P
)2
n m
2
i=1 (i
E[MST|] = +
Pmm 2 1

= 2 + n i=1 i
m1
2 + nv2
So E[MST|] 2 , with equality only if there is no variability in treatment means, i.e. v2 = 0.
ExpectedP
value of MSE:
1
2
MSE = m m
i=1 si , so
1 X
E[s2i ]
m
1 X 2
= 2
=
m

E[MSE] =

Lets summarize our potential estimators of 2 :

If H0 is true:
E[MSE|H0 ] = 2
E[MST|H0 ] = 2
If H0 is false:
E[MSE|H1 ] = 2

CHAPTER 5. INTRODUCTION TO ANOVA

68

E[(MST|H1 ] = 2 + nv2
This should give us an idea for a test statistic:
If H0 is true:
MSE 2
MST 2
If H0 is false
MSE 2
MST 2 + nv2 > 2
So
under H0 , MST/MSE should be around 1,
under Hc0 , MST/MSE should be bigger than 1.
Thus the test statistic F (Y ) = MST/MSE is sensitive to deviations from the
null, and can be used to measure evidence against H0 . Now all we need is a
null distribution.
Example (response times):
ybar . t<t a p p l y ( y , x , mean )
s 2 . t<t a p p l y ( y , x , var )
SSE< sum ( ( n1) s 2 . t )
SST< nsum ( ( ybar . tmean ( y ) ) 2 )
MSE<SSE / (m ( n1))
MST<SST/ (m1)

> SSE
[ 1 ] 12.0379
> SST
[ 1 ] 7.55032

CHAPTER 5. INTRODUCTION TO ANOVA

69

> MSE
[ 1 ] 0.8025267
> MST
[ 1 ] 1.88758

It is customary to summarize these calculations in a table:

Source of variation Sums of Squares Mean Squares F -ratio
Treatment
7.55
1.89
2.35
Noise
12.04
0.80
The F -ratio deviates a bit from 1. Possible explanations:
H1 : The F -value is a result of actual differences between treatments.
H0 : The F -value is a result of a chance assignment of treatments, eg.
the slow students were randomized to A, the fast students to D.
To evaluate H0 we look at how likely such a treatment assignment is:
Randomization test:
F . obs<anova ( lm ( y a s . f a c t o r ( x ) ) ) $F [ 1 ]
> F . obs
[ 1 ] 2.352046
$
set . seed (1)
F . n u l l <NULL
f o r ( nsim i n 1 : 1 0 0 0 )
{
x . sim<sample ( x )
F . n u l l <c (F . n u l l , anova ( lm ( y a s . f a c t o r ( x . sim ) ) ) $F [ 1 ] )
}
$
> mean (F . n u l l >=F . obs )
[ 1 ] 0.102

The observed between-group variation is larger than the observed withingroup variation, but not larger than the types of F -statistics wed expect
to get if the null hypothesis were true.

70

0.0

0.2

0.4

0.6

CHAPTER 5. INTRODUCTION TO ANOVA

Figure 5.2: Randomization distribution of the F -statistic

5.2

Partitioning sums of squares

Informally, we can think about the total variation in a datasets as follows:

Total variability

= variability due to treatment

= between treatment variability

+ variability due to other things

+ within treatment variability

This can in fact be made formal:

m X
n
X
SSTotal
(yij y )2 = SST + SSE
i=1 j=1

Proof:
m X
n
X

(yij y )2 =

i=1 j=1

XX
i

[(yij yi ) + (
yi y )]2

XX

XX

(yij yi )2 + 2(yij yi )(
yi y ) + (
yi y )2

(yij yi )2 +

= (1) + (2) + (3)

XX
i

2(yij yi )(
yi y ) +

XX
i

(
yi y )2

CHAPTER 5. INTRODUCTION TO ANOVA

71

P P
(1) =
(y yi )2 = SSE
Pi Pj ij
P
(3) =
(
yi y )2 = n i (
yi y )2 = SST
i
j
P P
P
P
(2) = 2 i j (yij yi )(
yi y ) = 2 i (
yi y ) j (yij yi )
but note that
!
X

(yij yi ) =

yij

n
yi,

= n
yi, n
yi, = 0
for all j. Therefore (2) = 0 and we have
total sum of squared deviations from grand mean

between trt sums of squares +

within trt sums of squares

or more succinctly,
SSTotal = SST + SSE.
Putting it all together:
H0 : i1 = i2 for all i1 , i2
Reduced model : yij = + ij
H1 : i1 6= i2 for some i1 6= i2
Full model : yij = i + ij
A fitted value or predicted value of an observation yij is denoted yij and
represents the modeled value of yij , without the noise.
A residual ij is the observed value minus the fitted value, ij = yij yij .
If we believe H1 ,
our estimate of i is
i = yi
the fitted value of yij is yij =
i = yi
the residual for (ij) is ij = (yij yij ) = (yij yi ).
the model lack-of-fit is measured by the sum of squared errors:
XX
(yij yi )2 = SSEF
i

CHAPTER 5. INTRODUCTION TO ANOVA

72

If we believe H0 ,
our estimate of is
= y
the fitted value of yij is yij =
= y
the residual for (ij) is ij = (yij yij ) = (yij y ).
the model lack-of-fit in this case is
XX
(yij y )2 = SSER = SSTotal
i

The improvement in model fit by including treatment parameters is

error in reduced model
SSTotal

- error in full model

SSE

= improvement in model fit

=
SST

The main idea: Variance can be partitioned into parts representing different sources. The variance explained by different sources can be compared
and analyzed. This gives rise to the ANOVA table.

5.2.1

The ANOVA table

ANOVA = Analysis of Variance

Source
Treatment
Noise
Total

Degrees of Freedom Sums of Squares Mean Squares

m1
SST
MST=SST/(m 1)
m(n 1)
SSE
MSE=SSE/m(n 1)
mn 1
SSTotal

F-ratio
MST/MSE

Reaction time example:

Source of variation Degrees of Freedom Sums of Squares Mean Squares F-ratio
Treatment
4
7.55
1.89
2.352
Noise
15
12.04
0.80
Total
19
19.59

CHAPTER 5. INTRODUCTION TO ANOVA

73

Uses of the ANOVA table:

No model assumptions table gives a descriptive decomposition of
variation.
Assuming Var[Yij ] = 2 in all groups
E[MSE] = 2
E[MST] = 2 + nv2 , where v2 = variance of true group means.
Assuming treatments were randomly assigned hypothesis tests, pvalues.
Assuming data are random samples from normal population hypothesis tests, p-values, confidence intervals for parameters, power calculations.

5.2.2

Understanding Degrees of Freedom:

Consider an experiment with m treatments/groups :

Data
Group means

y11 . . . , y1n
y1

n
y1 + + n
ym
y1 + + ym
y21 . . . , y2n
y2
y =
=

mn
m
..
..

.
.

ym1 . . . , ymn
ym
We can decompose each observation as follows:
yij = y + (
yi y) + (yij yi )
This leads to
(yij y)
=
(
yi y)
+
(yij yi )
total variation = between group variation + within group variation
All data can be decomposed this way, leading to the decomposition of the
data vector of length m n into two parts, as shown in Table 5.1. How do we
interpret the degrees of freedom? Weve heard of degrees of freedom before,
in the definition of a 2 random variable:

CHAPTER 5. INTRODUCTION TO ANOVA

Treatment
(
y1. y.. )
(
y1. y.. )
.
.
.
(
y1. y.. )
(
y2. y.. )
.
.
.
(
y2. y.. )
..
.

74

Total
y11 y..
y12 y..
.
.
.
y1n y..
y21 y..
.
.
.
y2n y..
..
.

=
=
=
=
=
=
=
=
=
=
=

ym1 y..
.
.
.
ymn y..

= (
ym. y.. )
=
.
=
.
=
.
= (
ym. y.. )

+ (ym1 ym. )
+
.
+
.
+
.
+ (ymn ym. )

SSTotal
mn 1

=
=

+
+

SSTrt
m1

+
+
+
+
+
+
+
+
+
+
+

Error
(y11 y1. )
(y12 y1. )
.
.
.
(y1n y1. )
(y21 y2. )
.
.
.
(y2n y2. )
..
.

SSE
m(n 1)

Table 5.1: ANOVA decomposition

CHAPTER 5. INTRODUCTION TO ANOVA

75

dof = number of statistically independent elements in a vector

In the ANOVA table, the dof have a geometric interpretation:
dof = number of components of a vector that can vary independently
To understand this latest definition, consider x1 , x2 , x3 and x = (x1 + x2 +
x3 )/3:

x1 x
c1
x2 x = c2
x3 x
c3
How many degrees of freedom does the vector (c1 , c2 , c3 )T have? How many
components can vary independently, if we know the elements are equal to
some numbers minus the average of the numbers?
c1 + c2 + c3 =
=
=
=

x1 x + x2 x x3 x
(x1 + x2 + x3 ) 3
x
3
x 3
x
0

Thus we must have c1 + c2 = c3 , and so c1 , c2 , c3 cant all be independently

varied, only two at a time can be arbitrarily changed. This vector thus lies in
a two-dimensional subspace of R3 , and has 2 degrees of freedom. In general,

x1 x

..

.
xm x

is an m-dimensional vector in an m 1 dimensional subspace, having m 1

degrees of freedom.
Exercise: Return to the vector decomposition of the data and obtain the
degrees of freedom of each component. Note that
dof = dimension of the space the vector lies in
SS = squared length of the vector
We will soon see the relationship between the geometric interpretation of
degrees of freedom and the interpretation involving 2 random variables.

CHAPTER 5. INTRODUCTION TO ANOVA

5.2.3

76

More sums of squares geometry

Consider the following vectors:

y11
y1
.
.
..
..

y
y1
1n

.
..
.
trt =
y=
y
.
.

ym1
ym

..
..
.
.
ymn
ym

y
.

..

..
= 1
y =
y

..

.
y

We can express our decomposition of the data as follows:

) = (
) + (y y
trt )
(y y
ytrt y
a
=
b
+
c
This is just vector addition/subtraction on vectors of length mn. Recall that
two vectors u and v are orthogonal/perpendicular/at right angles if
uv

m
X

ui vi = 0

i=1

We have already shown that b and c are orthogonal:

bc =

mn
X

b l cl

k=1

=
=
=

m X
n
X

(
yi y )(yij yi )

i=1 j=1
m
X

n
X
(
yi y )
(yij yi )

i=1
m
X

j=1

(
yi y ) 0

i=1

= 0
So the vector a is the vector sum of two orthogonal vectors. We can draw
this as follows:

CHAPTER 5. INTRODUCTION TO ANOVA

)
a = (y y
trt )
c = (y y

)
b = (
ytrt y
Now recall
||2 = SSTotal
||a||2 = ||y y
||b||2 = ||
ytrt y ||2 = SST
trt ||2 = SSE
||c||2 = ||y y
What do we know about right triangles?
||a||2
= ||b||2 + ||c||2
SSTotal = SST + SSE
So the ANOVA decomposition is an application of Pythagoras Theorem.
One final observation: recall that
) = m 1
dof (
ytrt y
trt ) = m(n 1)
dof (y y
) and (y y
trt ) are orthogonal.
(
ytrt y
The last lines means the degrees of freedom must add, so
) = (m 1) + m(n 1) = mn 1
dof(y y

77

CHAPTER 5. INTRODUCTION TO ANOVA

5.3

78

Unbalanced Designs

If the number of replications is constant for all levels of the treatment/factor

then the design is called balanced. Otherwise it is said to be unbalanced.
Unbalanced data:
yij , i = 1, . . . , m, j = 1, . . . , ni
P
let N = m
i=1 ni be the total sample size.
How does the ANOVA decomposition work in this case? What are the parameter estimates for the full and reduced model?
Null model:
yij = + ij

Var[ij ] = 2

You should be able to show that the least-squares estimators are

P

= y = N1
yij
P
s2 = N11 (yij y )2

Full model:
yij = i + ij
= + i + ij
Var[ij ] = 2
How should we estimate these parameters? When ni = n for all i, we had
Treatment means parameterization:
i = yi
Treatment effects parameterization:
= y , i = (
yi y )
which meant that y =

1
y .
m i

Similarly, we had
Pm Pn
i )2
i=1
j=1 (yij y
2
P
s =
m
i=1 (n 1)

CHAPTER 5. INTRODUCTION TO ANOVA

which implied that s2 =

1
m

79

s2i . However, if ni1 6= ni2 , then in general

Pm Pni
m
m ni
m
(yij yi )2
1 X
1 XX
1 X 2
i=1
Pmj=1
yi =
6
si 6=
yij , and
m i=1
N i=1 j=1
m i=1
i=1 (ni 1)
P

What should the parameter estimates be? With a bit of calculus you can
show that the least squares estimates of i or (, i ) are

i = yi
i = yi y ,
= y
We no longer have
=

1
m

i , or

i = 0, but we do have

Pm
ni
m X
X
i
i=1 ni y
P
=
yij /N
ni
i=1 j=1
= y , so
X

ni
i /
ni =
, and
X
ni i = 0.

So
is a weighted average of the
i s, and a weighted average of the i s is
zero. Similarly,
Pm Pni
Pm
i )2
(ni 1)s2i
i=1
j=1 (yij y
2
Pm
= Pi=1
s =
m
i=1 (ni 1)
i=1 (ni 1)
so s2 is a weighted average of the s2i s.

5.3.1

Sums of squares and degrees of freedom

The vector decomposition is shown in table 5.2. Let a, b and c be the three
vectors in the table. We define the sums of squares as the squared lengths of
these vectors:
P Pni
SSTotal = ||a||2 = m
)2
i=1
j=1 (yij y
P P i
P
SSTrt = ||b||2 = ti=1 nj=1
(
yi y )2 = m
)2
i=1 ni (yi y
P Pni
SSE = ||c||2 = m
i )2
i=1
j=1 (yij y

CHAPTER 5. INTRODUCTION TO ANOVA

80

Total
y11 y..
y12 y..
.
.
.
y1n1 y..
y21 y..
.
.
.
y2n2 y..
..
.

=
=
=
=
=
=
=
=
=
=
=

Treatment
(
y1. y.. )
(
y1. y.. )
.
.
.
(
y1. y.. )
(
y2. y.. )
.
.
.
(
y2. y.. )
..
.

+
+
+
+
+
+
+
+
+
+
+

Error
(y11 y1. )
(y12 y1. )
.
.
.
(y1n1 y1. )
(y21 y2. )
.
.
.
(y2n2 y2. )
..
.

ym1 y..
.
.
.
ymnm y..

= (
ym. y.. )
=
.
=
.
=
.
= (
ym. y.. )

+
+
+
+
+

(ym1 ym. )
.
.
.
(ymnm yt. )

=
=

+ P SSE
m
+
i=1 (ni 1)

SSTotal
N 1

SSTrt
m1

Table 5.2: ANOVA decomposition, unbalanced case

CHAPTER 5. INTRODUCTION TO ANOVA

81

Lets see if things add in a nice way. First, lets check orthogonality:
bc =

ni
m X
X

(
yi y )(yij yi )

i=1 j=1

=
=

m
X

ni
X
(
yi y )
(yij yi )

i=1

j=1

m
X

(
yi y ) 0 = 0

i=1

Note: c must be orthogonal to any vector that is constant within a group.

So we have

a=b+c
||a||2 = ||b||2 + ||c||2
bc=0
and so SSTotal = SST + SSE as before. What about degrees of freedom?
P
dof(c) = m
i=1 (ni 1) = N m should be clear
dof(a) = N 1 should be clear
dof(b) = ?
Note that in general,
m

1 X
yi 6= y ,
m i=1
But in the vector b we have ni copies of (
y y ), and
1 X
P
ni yi = y
ni
and so the vector b does sum to zero. Another way of looking at it is that
the vector b is made up of m numbers, which dont sum to zero, but their
weighted average sums to zero, and so the degrees of freedom are m 1.

5.3.2

ANOVA table for unbalanced data:

CHAPTER 5. INTRODUCTION TO ANOVA

82

Source

Deg. of Freedom

Sum of Squares

Mean Square

F-Ratio

Treatment

m1

SST

SST
MST= m1

MST/MSE

Noise

N m

SSE

Total

N 1

SSTotal

MSE =

SSE
N m

Now suppose the following model is correct:

yij = i + ij

Var[ij ] = 2

Does MSE still estimate 2 ?

MSE = SSE/(N m)
P m
Pn1
(ymj ym )2
1 )2 + nj=1
j=1 (y1j y
=
(n1 1) + + (nm 1)
(n1 1)s21 + + (nm 1)s2m
=
(n1 1) + + (nm 1)
So MSE is a weighted average of a bunch of unbiased estimates of 2 , so it
is still unbiased.
Is the F -statistic still sensitive to deviations from H0 ? Note that a group
with more observations contributes more to the grand mean, but it also
contributes more terms to the SST. One can show
E[MSE] = 2
E[MST] = 2 +

v2

Pm

N
v2
m1

, where

ni i2
N

i=1

i = i

Pm
n
i=1
P i i.
ni

So yes, MST/MSE will still be sensitive to deviations from the null, but the
groups with larger sample sizes have a bigger impact on the power.

CHAPTER 5. INTRODUCTION TO ANOVA

5.4

83

Normal sampling theory for ANOVA

Example (Blood coagulation): From a large population of farm animals,

24 animals were selected and randomly assigned to one of four different diets
A, B, C, D. However, due to resource constraints, rA = 4, rB = 6, rC =
6, rD = 8.

coagulation time
60
65

70

A
A

C
C
C
C

B
B
B
B
B

D
D
D
D
D
D

A
A

D
diet

Figure 5.3: Coagulation data

Questions:
Does diet have an effect on coagulation time?
If a given diet were assigned to all the animals in the population, what
would the distribution of coagulation times be?
If there is a diet effect, how do the mean coagulation times differ?
The first question we can address with a randomization test. For the second
and third we need a sampling model:
yij = i + ij
11 . . . mnm i.i.d. normal(0, 2 )
This model implies

CHAPTER 5. INTRODUCTION TO ANOVA

84

independence of errors
constant variance
normally distributed data
Another way to write it is as follows:
yA1 , . . . , yA4 i.i.d. normal(A , 2 )
yB1 , . . . , yB6 i.i.d. normal(B , 2 )
yC1 , . . . , yC6 i.i.d. normal(C , 2 )
yD1 , . . . , yD8 i.i.d. normal(D , 2 )
So we are viewing the 4 samples under A as a random sample from the
population of coagulation times that would be present if all animals got A
(and similarly for samples under B, C and D).
> anova ( lm ( c t i m e d i e t ) )
A n a l y s i s o f V a r i a n c e Table
Response : c t i m e
Df Sum Sq Mean Sq F v a l u e
diet
3 228.0
76.0 13.571
R e s i d u a l s 20 1 1 2 . 0
5.6

The F -statistic is large, but how unlikely is it under H0 ?

Two viewpoints on null distributions:
The 24 animals we selected are fixed. Other possible outcomes of the
experiment correspond to different assignments of the treatments.
The 24 animals we selected are random samples from a larger population of animals. Other possible outcomes of the experiment correspond
to different experimental units.
Randomization tests correspond to the first viewpoint, sampling theory to
the second.

CHAPTER 5. INTRODUCTION TO ANOVA

5.4.1

85

Sampling distribution of the F -statistic

Recall the 2 distribution:

Y1 , . . . , Yn i.i.d. normal(, 2 )

1 X
(Yi Y )2 2n1
2

Also,

X1 2k1

X2 2k2
X1 + X2 2k1 +k2

X1 , X2 independent

Distribution of SSE:
PP

(Yij Yi )2
2

1
2

(Y1j Y1 )2 +
+
2n1 1
2
N m

+
+

1
2

P
(Ymj Ym )2
2nm 1

So SSE/ 2 2N m .
Distribution of SST under the null: Under H0 ,
Yi normal(, 2 /ni )
ni Yi normal(ni , 2 )
1
1 X
SST
=
ni (Yi Y )2
2
2
m
1 X 2
=
( ni Yi ni Y ) 2m1
2 i=1

Results so far:
SSE/ 2 2N m
SST/ 2 2m1
SSE, SST independent (why?)

CHAPTER 5. INTRODUCTION TO ANOVA

86

Introducing the F -distribution: If

X1 2k1
X1 /k1
X2 2k2
Fk1 ,k2

X2 /k2
X 1 X2
Fk1 ,k2 is the F -distribution with k1 and k2 degrees of freedom.
Application: Under H0


SST
/(m 1)
M ST
2


=
Fm1,N m
SSE
M SE
/(N m)
2
A large value of F is evidence against H0 , so reject H0 if F > Fcrit . How to
determine Fcrit ?
Level- testing procedure:
1. gather data
2. construct ANOVA table
3. reject H0 : i = for all i if F > Fcrit
where Fcrit is the 1 quantile of an Ft1,N t distribution, available in
R via qf(1alpha, dof.trt , dof. err) . Under this procedure (and a host of
assumptions),
Pr(reject H0 |H0 true) =
Plots of several different F -distributions appear in Figure 5.4. Study these
plots until you understand the relationship between the shape of the curves
and the degrees of freedom. Now lets get back to the data analysis:
> anova ( lm ( c t i m e d i e t ) )
A n a l y s i s o f V a r i a n c e Table
Response : c t i m e
Df Sum Sq Mean Sq F v a l u e
Pr(>F)
diet
3 228.0
7 6 . 0 1 3 . 5 7 1 4 . 6 5 8 e 05
R e s i d u a l s 20 1 1 2 . 0
5.6

87

0.6

CHAPTER 5. INTRODUCTION TO ANOVA

0.0

0.2

density
0.4

F(3,20)
F(3,10)
F(3,5)
F(3,2)

10
F

15

20

15

20

CDF
0.4 0.6

0.8

1.0

0.2

F(3,20)
F(3,10)
F(3,5)
F(3,2)

10
F

Figure 5.4: F-distributions

88

p (F 3,, 20)
0.4

0.6

CHAPTER 5. INTRODUCTION TO ANOVA

0.0

0.2

randomization 95th percentile

normal theory 95th percentile

4
F

Figure 5.5: Normal-theory and randomization distributions of the F -statistic

Fobs<anova ( lm ( c t i m e d i e t ) ) $F [ 1 ]
Fsim<NULL
f o r ( nsim i n 1 : 1 0 0 0 ) {
d i e t . sim<sample ( d i e t )
Fsim<c ( Fsim , anova ( lm ( c t i m e d i e t . sim ) ) $F [ 1 ] )
}
> mean ( Fsim>=Fobs )
[1] 0
> 1p f ( Fobs , 3 , 2 0 )
[ 1 ] 4 . 6 5 8 4 7 1 e 05

5.4.2

Comparing group means

If H0 is rejected, there is evidence that some population means are different

from others. We can explore this further by making treatment comparisons.
If H0 is rejected we
estimate i with yi
estimate i2 with
s2i : if variances are very unequal, this might be a better estimate.

CHAPTER 5. INTRODUCTION TO ANOVA

89

MSE : if variances are close and ni is small, this is generally a

better estimate.
Standard practice: Unless there is strong evidence to the contrary, we typically assume Var[Yij ] = Var[Ykl ] = 2 , and use s2 MSE to estimate 2 . In
this case,
Var[
i ] = Var[Yi ] = 2 /ni
s2 /ni
p
The standard
error
of
the
mean
=
SE[

]
=
s2 /ni , is an estimate of
i
p

SD[
i ] = Var[Yi ] = / ni . The standard error is a very useful quantity.
Standard error: The usual definition of the standard error of an estimator
of a parameter is an estimate of its sampling standard deviation:
)
= (Y
= 2
Var[]
\
= 2
Var[
]
=
SE[]
where 2 is an estimate of 2 . For example,

i = Yi
Var[
i ] = 2 /ni
\
Var[
i ] =
2 /ni = s2 /ni

SE[
i ] = s/ ni
Confidence intervals for treatment means: Obtaining confidence intervals is very similar to the one-sample case. The only difference is that we
use data from all of the groups to estimate the variance. As a result, the
degrees of freedom changes.
Yi i
Yi i
Yi, i
p
=
=

s/ ni
SE[Yi ]
MSE/ni
tN m

CHAPTER 5. INTRODUCTION TO ANOVA

90

Note that degrees of freedom are those associated with MSE, NOT ni 1.
As a result,
Yi SE[Yi ] t1/2,N m
is a 100 (1 )% confidence interval for i .
A handy rule of thumb: If is an estimator of , then in many situations

2 SE[]
is an approximate 95% CI for .
Coagulation Example:
Yi SE[Yi ] t1/2,N m
For 95% confidence intervals,
t1,N m = t.975,20 = qt (.975,20) 2.1
p
p

SE[Yi ] = s2 /ni = 5.6/ni = 2.37/ ni

Diet diet
C
68
B
66
A
61
D
61

5.4.3

ni
6
6
4
8

SE[
diet ]
95% CI
0.97
(65.9,70.0)
0.97
(63.9,68.0)
1.18
(58.5,63.5)
0.84
(59.2,62.8)

Power calculations for the F-test

Recall, the power of a test is the probability of rejecting H0 when it is false.

Of course, this depends on how the null hypothesis is false.
Power(, 2 , n) = Pr(reject H0 |, 2 , n)
= Pr(Fobs > F1,m1,N m |, 2 , n)

CHAPTER 5. INTRODUCTION TO ANOVA

91

The noncentral F-distribution:

Y11 . . . , Y1n i.i.d. normal(1 , 2 )
.. .. ..
. . .
Ym1 . . . , Ymn i.i.d. normal(t , 2 )

Under this sampling model, F = F (Y ) has a non-central F distribution with

degrees of freedom n 1, N m
noncentrality parameter
=n

i2 / 2

where i = i
is the ith treatment effect.
In many texts, power is expressed as a function of the quantity :
sP
r P
2
i2 /m p
n i
=
= /m
=
2m
2 /n
Lets try to understand what represents:
P 2
i /m
2 =
2 /n
treatment variation
=
experimental uncertainty
= treatment variation experimental precision
Note that treatment variation means average squared treatment effect
size. We can gain some more intuition by rewriting as follows:
X
= n
i2 / 2
P 2 
i
1
= nm
m
2
between-treatment variation
= N
within-treatment variation

6
n

10

power
0.8
0.6
0.4

3.0

F.crit
3.5

4.0

10 15 20 25 30 35 40

m = 4 = 0.05 var.b/var.w=1

92
1.0

CHAPTER 5. INTRODUCTION TO ANOVA

6
n

10

6
n

10

Figure 5.6: Power as a function of n for m = 4, = 0.05 and 2 / 2 = 1

6
n

10

0.95
power
0.85
0.75

3.0

F.crit
3.5

4.0

20 30 40 50 60 70 80

m = 4 = 0.05 var.b/var.w=2

6
n

10

6
n

10

Figure 5.7: Power as a function of n for m = 4, = 0.05 and 2 / 2 = 2

So presumably power is increasing in , and .
Power(, 2 , n) = Pr(reject H0 |, 2 , n)
= Pr(Fobs > F1,m1,N m |, 2 , n)
=

5.5

1pf( qf(1alpha,m1,Nm) , m1 , Nm , ncp=lambda )

Model diagnostics

Our model is
yij = j + ij .
We have shown that, if

CHAPTER 5. INTRODUCTION TO ANOVA

93

A1: {ij }s are independent;

A2: Var[ij ] = 2 for all j;
A3: {ij }s are normally distributed.
then
F = MST/MSE Fm1,N m,
where is the noncentrality parameter. If in addition
H0 : i = for all i = 1, . . . , m.
then the noncentrality parameter is zero and F = MST/MSE Fm1,N m .
We make these assumptions to
do power calculations when designing a study,
test hypotheses after having gathered the data, and
make confidence intervals comparing the different treatments.
What should we do if the model assumptions are not correct?
We could have written
A0 : {ij } i.i.d. normal(0, 2 )
to describe all of A1 A3. We dont do this because some violations of
assumptions are more serious than others. Statistical folklore says the order
of importance is A1, A2 then A3. We will discuss A1 when we talk about
blocking. For now we will talk about A2 and A3.

5.5.1

Detecting violations with residuals

Violations of assumptions can be checked via residual analysis.

CHAPTER 5. INTRODUCTION TO ANOVA

94

Parameter estimates:
yij = y + (
yi y ) + (yij yi )
=
+ i + ij
Our fitted value for any observation in group i is yij =
+ i = yi
Our estimate of the error is ij = yij yi .
ij is called the residual for observation i, j.
Assumptions about ij can be checked by examining the values of ij s:

5.5.2

Checking normality assumptions:

Two standard graphical ways of assessing normality are with the following:
Histogram:
Make a histogram of ij s. This should look approximately bell-shaped
if the (super)population is really normal and there are enough observations. If there are enough observations, graphically compare the
histogram to a N (0, s2 ) distribution.
In small samples, the histograms need not look particularly bell-shaped.
Normal probability, or qq-plot:
If ij N (0, 2 ) then the ordered residuals ((1) , . . . , (mn) ) should correspond linearly with quantiles of a standard normal distribution.
How non-normal can a sample from a normal population look? You can
always check yourself by simulating data in R. See Figure ??
Example (Hermit Crab Data): Is there variability in hermit crab population across six different coastline sites? A researchers sampled the population in 25 randomly sampled transects in each of the six sites.
Data: yij = population total in transect j of site i.
Model: Yij = + i + ij
Note that the data are counts so they cannot be exactly normally distributed.

CHAPTER 5. INTRODUCTION TO ANOVA

95

Sample Quantiles
1
0
1

Density
0.0 0.1 0.2 0.3 0.4 0.5

0
y

0
1
Theoretical Quantiles

Sample Quantiles
1
0
1

Density
0.0 0.1 0.2 0.3 0.4 0.5
2

0
y

1
0
1
Theoretical Quantiles

Density
0.0 0.1 0.2 0.3 0.4 0.5

Sample Quantiles
1
0
1
2
2

0
y

1
0
1
Theoretical Quantiles

Figure 5.8: Normal scores plots of normal samples, with n {20, 50, 100}

CHAPTER 5. INTRODUCTION TO ANOVA

96

Data description: See Figure 5.9.

site sample mean
1
33.80
2
68.72
50.64
3
4
9.24
5
10.00
6
12.64

sample median
17
10
5
2
2
4

sample std dev

50.39
125.35
107.44
17.39
19.84
23.01

ANOVA:
> anova ( lm ( c r a b [ , 2 ] a s . f a c t o r ( c r a b [ , 1 ] ) ) )
A n a l y s i s o f V a r i a n c e Table
Response : c r a b [ , 2 ]
Df Sum Sq Mean Sq F v a l u e Pr(>F)
as . f a c t o r ( crab [ , 1 ] )
5 76695
15339 2 . 9 6 6 9 0 . 0 1 4 0 1
Residuals
144 744493
5170

Residuals:
ij = yij
i = yij yi
Residual diagnostic plots are in Figure 5.10. The data are clearly not
normally distributed.

5.5.3

Checking variance assumptions

The null distribution in the F-test is based on Var[ij ] = 2 for all groups i.
(a) tabulate residual variance in each treatment:
Trt
1
..
.

Sample var. of (i1 , . . . , in )

s21
..
.

s2m

CHAPTER 5. INTRODUCTION TO ANOVA

97

site2

0.000

0.000

Density
0.004 0.008

Density
0.010
0.020

0.012

site1

100

200
300
population

400

100

400

site4

0.00

0.000

0.04

Density
0.08

Density
0.005
0.010

0.12

0.015

site3

200
300
population

100

200
300
population

400

100

400

site6

0.00

0.00

0.02

0.02

Density
0.04

Density
0.04

0.06

0.06

site5

200
300
population

100

200
300
population

400

Figure 5.9: Crab data

100

200
300
population

400

0.000

0.004

0.008

Sample Quantiles
100 200 300 400

0.012

CHAPTER 5. INTRODUCTION TO ANOVA

100

100 200 300 400

residuals

98

2 1 0
1
2
Theoretical Quantiles

Figure 5.10: Crab residuals

Rule of thumb:
If s2largest /s2smallest < 4: dont worry;
If s2largest /s2smallest > 7: worry: need to account for the non-constant
variance, especially if sample sizes are different. Consider a randomization test, or converting the data to ranks for the overall
test of treatment effects.
(b) Residual vs. fitted value plots
For many types of data there is a mean-variance relationship: typically, groups with large means tend to have large variances. This is
especially true when the underlying distributions are skewed.
To check this, plot ij (residual) vs. yij = yi (fitted value).
Example: Crab data
Site
4
5
6
1
3
2

Sample mean
9.24
10.00
12.64
33.80
50.64
68.72

Sample standard deviation

17.39
19.84
23.01
50.39
107.44
125.35

residual
100 200 300 400

CHAPTER 5. INTRODUCTION TO ANOVA

99

10

20

30

40
fitted value

50

60

70

Figure 5.11: Fitted values versus residuals

Here we have s2largest /s2smallest 50. This is a problem.
(c) Statistical tests of equality of variance:
Levenes Test: Let dij = |yij yi |, where yi is the sample median
from group i.
These differences will be large in group i, if group i has a large
variance; and small in group i, if group i has a small variance.
We compute:
Pt
2 /(t 1)
(di d)
F0 = Pt Pni=1
2
j=1 (dij di ) /(t(n 1))
i=1
n

which is the ratio of the between group variability of the dij to the
within group variability of the dij .
Reject H0 : Var[ij ] = 2 for all i, j if F0 > Ft1,t(n1),1
Crab data:
F0 =

14, 229
= 2.93 > F5,144,0.95 = 2.28
4, 860

hence we reject the null hypothesis of equal variances at the 0.05 level.
See also

CHAPTER 5. INTRODUCTION TO ANOVA

100

the F Max test

Bartletts test;
the normal-theory test of equality for two variances (although this
depends on normality).
Crab data: So the assumptions that validate the use of the F -test are
violated. Now what?

5.5.4

Variance stabilizing transformations

Recall that one justification of the normal model,

Yij = i + ij ,
was that if the noise ij = Xij1 + Xij2 + was the result of the addition of
unobserved additive, independent effects then by the central limit theorem
ij will be approximately normal.
However, suppose the effects are multiplicative, so that in fact:
Yij = i ij = i (Xij1 Xij2 )
In this case, the Yij will not be normal, and the variances will not be constant:
Var[Yij ] = 2i Var[Xij1 Xij2 ]

Log transformation:
log Yij = log i + (log Xij1 + log Xij2 + )
Var[log Yij ] = Var[log i + log Xij1 + log Xij2 + ]
= Var[log Xij1 + log Xij2 + ]
2
= log
y
So that the variance of the log-data does not depend on the mean i . Also
note that by the central limit theorem the errors should be approximately
normally distributed.

crab population
0
200 400

CHAPTER 5. INTRODUCTION TO ANOVA

101

log crab population

2 0 2 4 6

site

3
site

Figure 5.12: Data and log data

Crab data: Let Yij = log(Yijraw + 1/6)
Site Sample mean
6
0.82
4
0.91
5
1.01
3
1.75
1
2.16
2
2.30

Sample standard deviation

2.21
1.87
1.74
2.41
2.27
2.44

> anova ( lm ( l o g ( c r a b [ , 2 ] + 1 / 6 ) a s . f a c t o r ( c r a b [ , 1 ] ) ) )
A n a l y s i s o f V a r i a n c e Table
Response : l o g ( c r a b [ , 2 ] + 1 / 6 )
Df Sum Sq Mean Sq F v a l u e Pr(>F)
as . f a c t o r ( crab [ , 1 ] )
5 54.73
10.95 2.3226 0.04604
Residuals
144 6 7 8 . 6 0
4.71
> anova ( lm ( l o g ( c r a b [ , 2 ] + 1 / 6 ) a s . f a c t o r ( c r a b [ , 1 ] ) ) )
A n a l y s i s o f V a r i a n c e Table

4
2

residual
0

0.10
0.05

102

0
2
residual

0.00
4

Sample Quantiles
0
2

0.15

0.20

CHAPTER 5. INTRODUCTION TO ANOVA

1
0
1
2
Theoretical Quantiles

1.0

1.5
fitted value

2.0

Figure 5.13: Diagnostics after the log transformation

Response : l o g ( c r a b [ , 2 ] + 1 / 6 )
Df Sum Sq Mean Sq F v a l u e Pr(>F)
as . f a c t o r ( crab [ , 1 ] )
5 54.73
10.95 2.3226 0.04604
Residuals
144 6 7 8 . 6 0
4.71

Other transformations: For data having multiplicative effects , we showed

that
i i ,
and taking the log stabilized the variances. In general, we may observe:
i i i.e. the standard deviation of a group depends on the group mean.
The goal of a variance stabilizing transformation is to find a transformation
of yij to yij such that yij (i )0 = 1, i.e. the standard deviation doesnt
depend on the mean.
Consider the class of power transformations, transformations of the form
Yij = Yij . Based on a Taylor series expansion of g (Y ) = Y around i , we
have
Yij = g (Yij )
i + (Yij i )1
i

E[Yij ] i
Var[Yij ] E[(Yij i )2 ](1
)2
i
SD[Yij ] i 1
= +1
i
i

CHAPTER 5. INTRODUCTION TO ANOVA

103

So if we observe i i , then i +1
. So if we take = 1 then
i
we will have stabilized the variances to some extent. Of course, we typically
dont know , but we could try to estimate it from data.
Estimation of :
i i i = ci
log i = log c + log i ,
so log si log c + log yi
Thus we may use the following procedure:
(1) Plot log si vs. log yi
(2) Fit a least squares line: lm( log si log yi )
(3) The slope
of the line is an estimate of .
(4) Analyze yij = yij1 .
Here are some common transformations:
mean-var relation

=1

transform

y const.

no transform!

y
y

1/2
i
3/4
i

y i
3/2

y i

y 2i

1/2

1/2

square root

yij

yij
1/2
yij

3/4

1/4

quarter power

1/4
yij

log

log yij

3/2

-1/2

reciproc. sqr. root

-1

reciprocal

yij

1/2

yij

1/yij

Note that
all the mean-variance relationships here are examples of power-laws.
Not all mean-variance relations are of this form.
= 1 is the multiplicative model discussed previously.

CHAPTER 5. INTRODUCTION TO ANOVA

104

More about the log transform

How did = 1 give us yij = log yij , shouldnt it be yij1 = yij = yij0 = 1 in
there?
Everything will make sense if we define for any 6= 0:
y () =

y 1
y + c.

For = 0, its natural to define the transformation as:

y 1
0


y ln y 
=
= ln y
1 =0

y (0) = lim y () = lim

0

Note that for a given 6= 0 it will not change the results of the ANOVA on
the transformed data if we transform using:
y = y

or

y () =

y 1
= ay + b.

To summarize the procedure: If the data present strong evidence of

nonconstant variance,
(1) Plot log si vs. log yi . If the relationship looks linear, then
(2) Fit a least squares line: lm( log si log yi )
(3) The slope
of the line is an estimate of .
(4) Analyze yij = yij1 .
This procedure is called the Box-Cox transformation (George Box and
David Cox first proposed the method in a paper in 1964). If the relationship
between log si and log yi does not look linear, then a Box-Cox transformation
will probably not help.

CHAPTER 5. INTRODUCTION TO ANOVA

105

A few words of caution:

For variance stabilization via the Box-Cox procedure to do much good,
the linear relationship between means and variances should be quite
tight.
Remember the rule of thumb which says not to worry if the ratio of
the largest to smallest variance is less than 4, i.e. dont use a transform
unless there are drastic differences in variances.
Dont get too carried away with the Box-Cox transformations. If
=
0.53 dont analyze yij = yij0.47 , just use yij0.5 . Remember,
is just an
estimate anyway (See Box and Cox Rebuttal, JASA 1982.)
Remember to make sure that you describe the units of the transformed
data, and make sure that readers of your analysis will be able to understand that the model is additive in the transformed data, but not
in the original data. Also always include a descriptive analysis of the
untransformed data, along with the p-value for the transformed data.
Try to think about whether the associated non-linear model for yij
makes sense.
Dont assume that the transformation is a magical fix: remember to
look at residuals and diagnostics after you do the transform. If things
havent improved much, dont transform.
Remember that the mean-variance relationship might not be cured by
a transform in the Box-Cox class. For example, if the response is a
binomial proportion
(= proportion of successes out of n), we have mean
p
= p, s.d. = p(1 p); the variance stabilizing transformation in this

case is y = arcsin y.
Keep in mind that statisticians disagree on the usefulness of transformations: some regard them as a hack more than a cure:
It can be argued that if the scientist who collected the data had a good
reason for using certain units, then one should not just transform the
data in order to bang it into an ANOVA-shaped hole. (Given enough
time and thought we could instead build a non-linear model for the
original data.)

CHAPTER 5. INTRODUCTION TO ANOVA

106

The sad truth: as always you will need to exercise judgment while
performing your analysis.
These warnings apply whenever you might reach for a transform, whether in
an ANOVA context, or a linear regression context.
Example (Crab data): Looking at the plot of means vs. sd.s suggests
1, implying a log-transformation. However, the zeros in our data lead
to problems, since log(0) = .
Instead we can use yij = log(yij +1/6). For the transformed data this gives us
a ratio of the largest to smallest standard deviation of approximately 2 which
is acceptable based on the rule of 4. Additionally, the residual diagnostic
plots (Figure 5.13) are much improved
This table needs to be fixed: Third column needs to be sd(log(y)).
site sample sd
4
17.39
19.84
5
6
23.01
50.39
1
3
107.44
2
125.35

sample mean
9.24
10.00
12.64
33.80
50.64
68.72

log(sample sd)
2.86
2.99
3.14
3.92
4.68
4.83

log(sample mean)
2.22
2.30
2.54
3.52
3.92
4.23

lm ( f o r m u l a = l o g s d log mean )
Coefficients :
( Intercept )
log mean
0.6652
0.9839

5.6

Treatment Comparisons

Recall the coagulation time data from the beginning of the chapter: Four
different diets were assigned to a population of 24 animals, with n1 = 4,
n2 = 6, n3 = 6 and n4 = 8.
> anova ( lm ( c t i m e d i e t ) )
A n a l y s i s o f V a r i a n c e Table

CHAPTER 5. INTRODUCTION TO ANOVA

107

4.0

3.5

log(sd)

4.5

3.0

2.5

3.0

3.5

4.0

log(mean)

Figure 5.14: Mean-variance relationship of the transformed data

Response : c t i m e
Df Sum Sq Mean Sq F v a l u e
Pr(>F)
diet
3 228.0
7 6 . 0 1 3 . 5 7 1 4 . 6 5 8 e 05
R e s i d u a l s 20 1 1 2 . 0
5.6

We conclude from the F -test that there are substantial differences between
the population treatment means. How do we decide what those differences
are?

5.6.1

Contrasts

Differences between sets of means can be evaluated by estimating contrasts.

A contrast is a linear function of the means such that the coefficients sum to
zero:
m
m
X
X
C = C(, k) =
ki i , where
ki = 0
i=1

Examples:
diet 1 vs diet 2 : C = 1 2

i=1

CHAPTER 5. INTRODUCTION TO ANOVA

108

diet 1 vs diets 2,3 and 4 : C = 1 (2 + 3 + 4 )/3

diets 1 and 2 vs diets 3 and 4 : C = (1 + 2 )/2 (3 + 4 )/2
Contrasts are functions of the unknown parameters. We can estimate
them, and obtain standard errors for them. This leads to confidence intervals
and hypothesis tests.
Parameter estimates: Let
C =

m
X

ki i =

i=1

m
X

ki yi

i=1

= C, so C is a unbiased estimator of C.
Then E[C]
Standard errors:
m
X

=
Var[C]

i=1
m
X

Var[ki yi ]
ki2 2 /ni

i=1

m
X

ki2 /ni

i=1

is
So an estimate of Var[C]
s2C

=s

m
X

ki2 /ni

i=1

The standard error of a contrast is an estimate of its standard deviation

s
X k2
i
=s
SE[C]
ni

CHAPTER 5. INTRODUCTION TO ANOVA

109

t-distributions for contrasts: Consider

Pm
i
C
i=1 ki y
= pP

ki2 /ni
s
SE[C]
If the data are normally distributed , then under H0 : C =
C

SE[C]

ki i = 0,

tN m

Exercise: Prove this result

Hypothesis test:
H0 : C = 0 versus H1 : C 6= 0.

> t1/2,N m .
Level- test : Reject H0 if |C/SE[
C]|

p-value: Pr(|tN m | > |C(y)/SE[

C(y)]|)
= 2*(1-pt( abs(c_hat/se_c_hat),N-m))
Example: Recall in the coagulation example
1 = 61,
2 = 66, and their
95% confidence intervals were (58.5,63.5) and (63.9,68.0). Let C = A B .
Hypothesis test: H0 : C = 0.
C

y y2
p 1

s 1/6 + 1/4
SE[C]
5
=
1.53
= 3.27
=

p-value = 2*(1-pt(3.27,20)) = 0.004.

Confidence intervals: Based on the normality assumptions, a 100 (1
)% confidence interval for C is given by

C t1/2,N m SE[C]

CHAPTER 5. INTRODUCTION TO ANOVA

110

20

12

grain yield
14
16
18

10

20

30
plant density

40

50

Figure 5.15: Yield-density data

5.6.2

Orthogonal Contrasts

What use are contrasts beyond just comparing two means? Consider the
data in Figure 5.15, which show the results of a CRD for an experiment on
the effects of planting density on crop yield in which there were three fields
randomly assigned to each of 5 planting densities.
> anova(lm(y~as.factor(x))
Df Sum Sq Mean Sq F value
Pr(>F)
as.factor(x) 4 87.600 21.900 29.278 1.690e-05 ***
Residuals
10 7.480
0.748

There is strong evidence of an effect of planting density. How should we summarize the effect? In this experiment, the treatment levels have an ordering
to them (this is not always the case). Consider the following m 1 = 4
contrasts:
Contrast k1 k2 k3 k4 k5
C1
-2 -1 0 1 2
C2
2 -1 -2 -1 2
C3
-1 2 0 -2 1
C4
1 -4 6 -4 1
Note:

CHAPTER 5. INTRODUCTION TO ANOVA

111

These are all actually contrasts (the coefficients sum to zero).

They are orthogonal : Ci Cj = 0.
What are these contrasts representing? Would would make them large?
If all i s are the same, then they will all be close to zero. This is the
sum to zero part, i.e. Ci 1 = 0 for each contrast.
C1 will be big if 1 , . . . , 5 are monotonically increasing or monotonically decreasing. This contrast is measuring the linear component of
the relationship between density and yield.
C2 is big if there is a bump, up or down, in the middle of the treatments, i.e. C2 is measuring the quadratic component of the relationship.
Similarly, C3 and C4 are measuring the cubic and quartic parts of the
relationship between density and yield.
The orthogonality bit is important: Suppose the true relationship between
density and yield is linear, eg, i = + densityi . Then C2 () = C3 () =
C4 () = 0.
You can produce these contrast coefficients in R:
> t ( contr . poly ( 5 ) )
[ ,1]
[ ,2]
[ ,3]
[ ,4]
[ ,5]
. L 0.6324555 0.3162278 0 . 0 0 0 0 0 0 e+00 0 . 3 1 6 2 2 7 8 0 . 6 3 2 4 5 5 5
.Q 0 . 5 3 4 5 2 2 5 0.2672612 5.345225 e 01 0.2672612 0 . 5 3 4 5 2 2 5
. C 0.3162278 0 . 6 3 2 4 5 5 5 4.095972 e 16 0.6324555 0 . 3 1 6 2 2 7 8
4 0 . 1 1 9 5 2 2 9 0.4780914 7 . 1 7 1 3 7 2 e 01 0.4780914 0 . 1 1 9 5 2 2 9

Here each contrast has been normalized so that

change the orthogonality.
Estimating the contrasts is easy:
> t r t . means<t a p p l y ( y , x , mean )
> t r t . means
10 20 30 40 50
12 16 19 18 17
> c . hat<t r t . means%%c o n t r . p o l y ( 5 )
> c . hat
.L
.Q
.C
4
[ 1 , ] 3 . 7 9 4 7 3 3 3.741657 0 . 3 1 6 2 2 7 8 0 . 8 3 6 6 6

ki2 = 1. This doesnt

CHAPTER 5. INTRODUCTION TO ANOVA

112

> 3 c . hat 2
. L .Q . C 4
[ 1 , ] 4 3 . 2 42 0 . 3 2 . 1
> sum ( 3 c . hat 2 )
[ 1 ] 87.6

Coincidence? I think not. Recall, we represented treatment variation as a

vector that lived in m 1 dimensional space. This variation can be further
decomposed into m 1 orthogonal parts:
Source
Linear trt
Quad trt
Cubic trt
Quart trt
Total trt
Error
Total

df
1
1
1
1
4
10
14

SS
MS
F
43.20 43.20 57.75
42.00 42.00 56.15
0.30 0.30 0.40
2.10 2.10 2.81
87.60 21.90 29.28
7.48 0.75
95.08

The useful idea behind orthogonal contrasts is that the treatment variation can be decomposed into orthogonal parts. As you might expect, under
H0r : Cr = 0, the F -statistic corresponding to the rth contrast has an F distribution with 1 and N m degrees of freedom (assuming normality, constant variance, etc.). For the planting density data, we find strong evidence
of linear and quadratic components to the relationship between density and
yield.

5.6.3

Multiple Comparisons

An experiment with m treatment levels has m2 pairwise comparisons, i.e.
contrasts of the form C = i j . Should we perform hypothesis tests for
all comparisons?
The more hypotheses we test, the higher the probability that at least one of
them will be rejected, regardless of their validity.
Two levels of error: Define the hypotheses
H0 : i = j for all i, j

CHAPTER 5. INTRODUCTION TO ANOVA

113

H0ij : i = j
We can associate error rates to both of these types of hypotheses
Experiment-wise type I error rate: Pr(reject H0 |H0 is true ).
Comparison-wise type I error rate : Pr(reject H0ij |H0ij is true ).
Consider the following procedure:
1. Gather data
2. Compute all pairwise contrasts and their t-statistics
3. Reject each H0ij for which |tij | > t1C /2,N m
Letting tcrit = t1C /2,N m , the comparison-wise type I error rate is of
course
P (|tij | > tcrit |H0ij ) = C .
The experiment-wise type I error rate is the probability that we say
differences between treatments exist when no differences exist:
P (|tij | > tcrit for some i, j |H0 ) C
with equality only if there are two treatments total. The fact that the
experiment-wise error rate is larger than the comparison-wise rate is called
the issue of multiple comparisons. What is the experiment-wise rate in this
analysis procedure?
Pr(one or more H0ij rejected |H0 ) = 1 Pr(none of H0ij rejected |H0 )
Y
<
1
Pr(H0ij not rejected |H0 )
i,j
m

= 1 (1 C )( 2 )
We can approximate this with a Taylor series expansion: Let f (x) = 1 (1
C )x . Then f 0 (0) = log(1 C ) and
f (x) f (0) + xf 0 (0)
1 (1 C )x (1 1) x log(1 C )
xC for small C .

CHAPTER 5. INTRODUCTION TO ANOVA

So
Pr(one or more H0ij

114

 
m
rejected |H0 )
C
2
<

If we are worried about possible dependence among the tests, perhaps a

better way to derive this bound is to recall that
Pr(A1 A2 ) = Pr(A1 ) + Pr(A2 ) Pr(A1 A2 )
Pr(A1 ) + Pr(A2 ) and that
Pr(A1 Ak ) Pr(A1 ) + + Pr(Ak )
(subadditivity). Therefore
P (one or more H0ij rejected |H0 )

P (H0ij rejected |H0 )

i,j

 
m
=
C
2
Bonferroni error control:
The Bonferroni procedure for controlling experiment-wise type I error rate
is as follows:


1. Compute pairwise t-statistics on all m2 pairs.
2. Reject H0ij if |tij | > t1C /2,N m


where C = E / m2 .
the experiment-wise error rate is less than E
the comparison-wise error rate is C .
So for example if E = 0.05 and m = 5, then c = 0.005.
Fishers Protected LSD (Fishers least-significant-difference method):
Another approach to controlling type I error makes use of the F -test.
1. Perform ANOVA and compute the F -statistic.
2. If F (y) < F1E ,m1,N m then dont reject H0 and stop.

CHAPTER 5. INTRODUCTION TO ANOVA

115

3. If F (y) > F1E ,m1,N m then reject H0 and reject all H0ij for which
|Cij /SE[Cij ]| > t1 /2,N m .
C

Under this procedure,

Experiment-wise type I error rate is E
Comparison-wise type I error rate is
Pr(reject H0ij |H0ij ) = Pr(F > Fcrit and |tij | > tcrit |H0ij )
= Pr(F > Fcrit |H0ij ) Pr(|tij | > tcrit |F > Fcrit , H0ij )
If H0 is also true, this is less than E , and is generally between C
E and E , depending on how many treatments there are. For
example, if there only a few treatments, then F > Fcrit suggests
|tij | > tcrit .
If H0 is not true, we dont know what this is, but some simulation
work suggests this is approximately C power, again depending
on how many treatments there are and the true treatment variation.

5.6.4

False Discovery Rate procedures

5.6.5

Nonparametric tests

Chapter 6
Factorial Designs
Example (Insecticide): In developing methods of pest control, researchers
are interested in the efficacy of different types of poison and different delivery methods.
Treatments:
Type {I, II, III}
Delivery {A, B, C, D}
Response: time to death, in minutes.
Possible experimental design: Perform two separate CRD experiments,
one testing for the effects of Type and the other for Delivery. But,
which Delivery to use for the experiment testing for Type effects?
which Type to use for the experiment testing for Delivery effects?
To compare different TypeDelivery combinations, we need to do experiments under all 12 treatment combinations.
Experimental design: 48 insects randomly assigned to treatments: 4 to
each treatment combination, i.e.
4 assigned to (I, A),
4 assigned to (I, B),
116

CHAPTER 6. FACTORIAL DESIGNS

117

..
.
It might be helpful to visualize the design as follows:
Type
1
2
3

A
y I,A
y II,A
y III,A

Delivery
B
C
y I,B
y I,C
y II,B y II,C
y III,B y III,C

D
y I,D
y II,D
y III,D

This type of design is called a factorial design. Specifically, this design is a

3 4 two-factor design with 4 replications per treatment combination.
Factors: Categories of treatments
Levels of a factor: the different treatments in a category
So in this case, Type and Delivery are both factors. There are 3 levels of
Type and 4 levels of Delivery.

6.1

Data analysis:

Lets first look at a series of plots:

Marginal Plots: Based on these marginal plots, it looks like (III, A) would
be the most effective combination. But are the effects of Type consistent across levels of Delivery?
Conditional Plots: Type III looks best across delivery types. But the difference between types I and II seems to depend on delivery. For example, for delivery methods B or D there doesnt seem to be much of a
difference between I and II.
Cell Plots: Another way of looking at the data is to just view it as a CRD
with 3 4 = 12 different groups. Sometimes each group is called a cell.
Notice that there seems to be a bit of a mean-variance relationship. Lets
take care of this before we go any further: Plotting means versus standard
deviations on both the raw and log scale gives the relationship in Figure 6.4.
Computing the least squares line gives

CHAPTER 6. FACTORIAL DESIGNS

12
10
8

10
2

12

118

II

III

Figure 6.1: Marginal Plots.

> lm ( l o g ( s d s ) l o g ( means ) )
Coefficients :
( Intercept )
l o g ( means )
3.203
1.977

So i c2i . This suggests a reciprocal transformation, i.e. analyzing yi,j =

1/ time to death. Does this reduce the relationship between means and variances? Figure 6.5 shows the mean-variance relationship for the transformed
data. The transformation seems to have improved things. We select this as
our data scale and start from scratch, beginning with plots (Figure 6.6).
Possible analysis methods: Lets first try to analyze these data using
our existing tools:
Two one-factor ANOVAS: Just looking at Type, for example, the experiment is a one-factor ANOVA with 3 treatment levels and 16 reps
per treatment. Conversely, looking at Delivery, the experiment is a one
factor ANOVA with 4 treatment levels and 12 reps per treatment.
> dat$y <1/dat$y
> anova ( lm ( dat$y d a t $ t y p e ) )

dat$type

Df Sum Sq Mean Sq F v a l u e
Pr(>F)
2 0 . 3 4 8 7 7 0 . 1 7 4 3 9 2 5 . 6 2 1 3 . 7 2 8 e 08

CHAPTER 6. FACTORIAL DESIGNS

119

2.0

2.5

3.0

3.5

10

4.0

12

4.5

II

III

III

10

II

II

III

Figure 6.2: Conditional Plots.

II

III

120

10

12

CHAPTER 6. FACTORIAL DESIGNS

I.A II.A

I.B II.B

I.C II.C

I.D II.D

1.0

Figure 6.3: Cell plots.

2.0

0.0

log(sds)
1.0
0.0

means
5 6

3
2

1.0

2.0
sds

3.0

0.8

1.2
1.6
log(means)

Figure 6.4: Mean-variance relationship.

2.0

CHAPTER 6. FACTORIAL DESIGNS

3.8

0.2

0.02

0.04

0.06

log(sds)
3.4
3.0

means
0.3

0.4

2.6

121

0.08

2.2

sds

1.8
1.4
1.0
log(means)

Figure 6.5: Mean-variance relationship for transformed data.

R e s i d u a l s 45 0 . 3 0 6 2 8 0 . 0 0 6 8 1

> anova ( lm ( dat$y d a t $ d e l i v e r y ) )

Df Sum Sq Mean Sq F v a l u e
Pr(>F)
d a t $ d e l i v e r y 3 0 . 2 0 4 1 4 0 . 0 6 8 0 5 6 . 6 4 0 1 0 . 0 0 0 8 4 9 6
Residuals
44 0 . 4 5 0 9 1 0 . 0 1 0 2 5

A one-factor ANOVA: There are 3 4 = 12 different treatments:

> anova ( lm ( dat$y d a t $ t y p e : d a t $ d e l i v e r y ) )
Df Sum Sq Mean Sq F v a l u e
Pr(>F)
d a t $ t y p e : d a t $ d e l i v e r y 11 0 . 5 6 8 6 2 0 . 0 5 1 6 9 2 1 . 5 3 1 1 . 2 8 9 e 12
Residuals
36 0 . 0 8 6 4 3 0 . 0 0 2 4 0

Why might these methods be insufficient?

What are the SSE, MSE representing in the first two ANOVAS? Why
are they bigger than the value in the in the third ANOVA?

CHAPTER 6. FACTORIAL DESIGNS

122

10
8

12

10
8

12

6
4
2

II

III

0.10

0.25

0.35

0.20

0.45

0.30

0.55

II

III

II

0.10

0.15

0.25

0.20

0.35

0.30

0.45

III

II

III

II

III

0.1 0.2 0.3 0.4 0.5

I.A

II.A

III.A

I.B

II.B

III.B

I.C

II.C

III.C

I.D

Figure 6.6: Plots of transformed poison data

II.D

III.D

CHAPTER 6. FACTORIAL DESIGNS

123

In the third ANOVA, can we assess the effects of Type and Delivery
separately?
Can you think of a situation where the F -stats in the first and second
ANOVAs would be small, but the F -stat in the third ANOVA big?
Basically, the first and second ANOVAs may mischaracterize the data and
sources of variation. The third ANOVA is valid, but wed like a more
specific result: wed like to know which factors are sources of variation, and
the relative magnitude of their effects. Also, if the effects of one factor are
consistent across levels of the other, maybe we dont need to have a separate
parameter for each of the 12 treatment combinations, i.e. a simpler model
may suffice.

6.2

Additive effects model

Yi,j,k = + ai + bj + i,j,k ,

i = 1, . . . , m1 , j = 1, . . . , m2 , k = 1, . . . , n

= overall mean;
a1 , . . . , am1 = additive effects of factor 1;
b1 , . . . , bm2 = additive effects of factor 2.
Notes:
1. Side conditions: As with with the treatment effects model in the onefactor case, we only need m1 1 parameters to differentiate between
m1 means, so we usually
restrict a1 = 0, b1 = 0 (set-to-zero side conditions ), OR
P
P
restrict
ai = 0,
bj = 0 (sum-to-zero side conditions).
2. The additive model is a reduced model : There are m1 m2 groups
or treatment combinations, and a full model fits a different population
mean separately to each treatment combination, requiring m1 m2
parameters. In contrast, the additive model only has
1
parameter for
m1 1
parameters for ai s
m2 1
parameters for bj s
m1 + m2 1 parameters total.

CHAPTER 6. FACTORIAL DESIGNS

124

Parameter estimation and ANOVA decomposition:

yijk = y + (
yi y ) + (
yj y ) + (yijk yi yj + y )
bj
=
+
a
i
+
+
ijk
These are the least-squares parameter estimates, under the sum-to-zero side
conditions:
X
X
a
i =
(
yi y ) = n
y n
y = 0
To obtain the set-to-zero side conditions, add a
1 and b1 to
, subtract a
1

from the a
i s, and subtract b1 from the bj s. Note that this does not change
the fitted value in each group:
bj
fitted(yijk ) =

+
a
i
+
= (
+a
1 + b1 ) + (
ai a
1 ) + (bj b1 )

b
=

+
a
i
+
j
As you might have guessed, we can write this decomposition out as vectors
of length m1 m2 n:
+ 
+ b
y y = a
vT
= v1 + v2 + ve
The columns represent
vT variation of the data around the grand mean;
v1 variation of factor 1 means around the grand mean;
v2 variation of factor 2 means around the grand mean;
ve variation of the data around fitted the values.
You should be able to show that these vectors are orthogonal, and so
P P P
P P P 2
P P P 2
P P P 2
2
(y

)
=
a

+
b
+
i
ijk

i
i
i
j
k
i
j
k
i
j
k
i
j
k
SSTotal
=
SSA
+
SSB
+
SSE
Degrees of Freedom:
contains m1 different numbers but sums to zero m1 1 dof
a
contains m2 different numbers but sums to zero m2 1 dof
b

CHAPTER 6. FACTORIAL DESIGNS

125

ANOVA table
Source
SS
A
SSA
SSB
B
SSE
Error
Total SSTotal

df
m1 1
m2 1
(m1 1)(m2 1) + m1 m2 (n 1)
m1 m2 n 1

MS
SSA/dfA
SSB/dfB
SSE/dfE

F
MSA/MSE
MSB/MSE

> anova ( lm ( dat$y d a t $ t y p e+d a t $ d e l i v e r y ) )

A n a l y s i s o f V a r i a n c e Table
Response : dat$y
Df Sum Sq
dat$type
2 0.34877
dat$delivery 3 0.20414
Residuals
42 0 . 1 0 2 1 4

Mean Sq F v a l u e
0.17439 71.708
0.06805 27.982
0.00243

This ANOVA has decomposed the variance in the data into the variance of
additive Type effects, additive Delivery effects, and residuals. Does this
adequately represent what is going on in the data? What do we mean by
additive? Assuming the model is correct, we have:
E[Y |type=I, delivery=A] = + a1 + b1
E[Y |type=II, delivery=A] = + a2 + b1
This says that the difference between Type I and Type II is a1 a2 regardless
of Delivery. Does this look right based on the plots? Consider the following
table:
Effect of Type I vs II, given Delivery
Delivery full model
additive model
A
IA IIA ( + a1 + b1 ) ( + a2 + b1 ) = a1 a2
B
IB IIB ( + a1 + b2 ) ( + a2 + b2 ) = a1 a2
C
IC IIC ( + a1 + b3 ) ( + a2 + b3 ) = a1 a2
D
ID IID ( + a1 + b4 ) ( + a2 + b4 ) = a1 a2
The full model allows differences between Types to vary across levels
of Delivery
The reduced/additive model says differences are constant across levels
of Delivery.

CHAPTER 6. FACTORIAL DESIGNS

126

Therefore, the reduced model is appropriate if

(IA IIA ) = (IB IIB ) = (IC IIC ) = (ID IID )
How can we test for this? Consider the following parameterization of the
full model:
Interaction model:
Yijk = + ai + bj + (ab)ij + ijk
= overall mean;
a1 , . . . , am1 = additive effects of factor 1;
b1 , . . . , bm2 = additive effects of factor 2.
(ab)ij = interaction terms = deviations from additivity.
The interaction term is a correction for non-additivity of the factor effects.
This is a full model: It fits a separate mean for each treatment combination:
E(Yijk ) = ij = + ai + bj + (ab)ij
Parameter estimation and ANOVA decomposition:
yijk = y + (
yi y ) + (
yj y ) + (
yij yi yj + y ) + (yijk yij )

=
+
a
i
+
bj
+
(ab)
+
ijk
ij
Note that the interaction term is equal to the fitted value under the full
model (yij ) minus the fitted value under the additive model (
yi + yj y ).
Deciding between the additive/reduced model and the interaction/full model
s is large or
is tantamount to deciding if the variance explained by the (ab)
ij
not, i.e. whether or not the full model is close to the additive model.

6.3

Evaluating additivity:

Recall the additive model

yijk = + ai + bj + ijk

CHAPTER 6. FACTORIAL DESIGNS

127

i = 1, . . . , m1 indexes the levels of factor 1

j = 1, . . . , m2 indexes the levels of factor 1
k = 1, . . . , n indexes the replications for each of the m1 m2 treatment
combinations.
Parameter estimates are obtained via the following decomposition:
yijk = y + (
yi y ) + (
yj y ) + (yijk yi yj + y )
bj
=
+
a
i
+
+
ijk
Fitted value:
yijk =
+a
i + bj
= y + (
yi y ) + (
yj y )
= yi + yj y
Note that in this model the fitted value in one cell depends on data
from the others.
Residual:
ijk = yijk yijk
= (yijk yi yj + y )
As we discussed, this model is adequate if the differences across levels of
one factor dont depend on the level of the other factor, i.e.
y1j y2j a
1 a
2 for all j = 1, . . . , m2 , and
yi1 yi2 b1 b2 for all i = 1, . . . , m1 .
This adequacy can be assessed by looking differences between the sample
means from each cell and the fitted values of these averages under the additive
model.
yij (
+a
i + bj ) = yij yi yj + y

(ab)
ij

The term (ab)

ij measures the deviation of the cell means to the estimated
additive model. It is called an interaction. It does not measure how factor 1
interacts with factor 2. It measures how much the data deviate from the
additive effects model.

CHAPTER 6. FACTORIAL DESIGNS

128

Interactions and the full model: The interaction terms also can be
derived by taking the additive decomposition above one step further: The
residual in the additive model can be written:
i yj + y
A
ijk = yijk y
= (yijk yij ) + (
yij yi yj + y )
I

=  + (ab)
ijk

ij

This suggests the following full model, or interaction model

yijk = + ai + bj + (ab)ij + ijk
with parameter estimates obtained from the following decomposition:
yijk = y + (
yi y ) + (
yj y ) + (
yij yi yj + y ) + (yijk yij )
ij

=
+
a
i
+
bj
+
(ab)
+
ijk
Fitted value:
ij
yijk =
+a
i + bj + (ab)
= yij =
ij
This is a full model for the treatment means: The estimate of the mean
in each cell depends only on data from that cell. Contrast this to
additive model.
Residual:
ijk = yijk yijk
= yijk yij
Thus the full model ANOVA decomposition partitions the variability among
the cell means y11 , y12 , . . . , ym1 m2 into
the overall mean

additive treatment effects a

i , bj
= yij (
what is leftover : (ab)
+a
i + bj )
ij

CHAPTER 6. FACTORIAL DESIGNS

129

As you might expect, these different parts are orthogonal, resulting in the
following orthogonal decomposition of the variance.

Total SS

var explained by add model + error in add model

var explained by full model
+ error in full model
= SSA +
SSB
+ SSAB +
SSE

Example: (Poison) 3 4 two-factor CRD with 4 reps per treatment combination.

Df Sum Sq Mean Sq F v a l u e
p o i s $ d e l i v 3 0.20414 0.06805 27.982
pois$type
2 0.34877 0.17439 71.708
R e s i d u a l s 42 0 . 1 0 2 1 4 0 . 0 0 2 4 3

Df
pois$deliv
3
pois$type
2
pois$deliv : pois$type 6
Residuals
36

Sum Sq
0.20414
0.34877
0.01571
0.08643

Mean Sq F v a l u e
0.06805 28.3431
0.17439 72.6347
0.00262 1.0904
0.00240

So notice
0.10214 = 0.01571 + 0.08643, that is SSEadd = SSABint + SSEint
42 = 6 + 36, that is dof(SSEadd ) = dof(SSABint ) + dof(SSEint )
SSA, SSB, dof(A), dof(B) are unchanged in the two models
MSEadd MSEint , but degrees of freedom are larger in the additive
model. Which do you think is a better estimate of the within-group
variance?
Expected sums of squares:
If H0 : (ab)ij = 0 is true, then
E[MSE] = 2
E[MSAB] = 2
If H0 : (ab)ij = 0 is not true, then
E[MSE] = 2

CHAPTER 6. FACTORIAL DESIGNS

130

2
E[MSAB] = 2 + rAB
> 2.

This suggests
An evaluation of the adequacy of the additive model can be assessed
by comparing MSAB to MSE. Under H0 : (ab)ij = 0 ,
FAB = MSAB/MSE F(m1 1)(m2 1),m1 m2 (n1)
Evidence against H0 can be evaluated by computing the p-value.
If the additive model is adequate then MSEint and MSAB are two
independent estimates of roughly the same thing (why independent?).
We may then want to combine them to improve our estimate of 2 .
Df
pois$deliv
3
pois$type
2
pois$deliv : pois$type 6
Residuals
36

Sum Sq
0.20414
0.34877
0.01571
0.08643

Mean Sq F v a l u e
Pr(>F)
0 . 0 6 8 0 5 2 8 . 3 4 3 1 1 . 3 7 6 e 09
0 . 1 7 4 3 9 7 2 . 6 3 4 7 2 . 3 1 0 e 13
0.00262 1.0904
0.3867
0.00240

For these data, there is strong evidence of both treatment effects, and little
evidence of non-additivity. We may want to use the additive model.

6.4

Inference for additive treatment effects

Consider a two-factor experiment in which it is determined that the effects

of factor F1 and F2 are large. Now we want to compare means across levels
of one of the factors.
Recall in the pesticide example we had 4 reps for each of 3 levels of Poison
type and 4 levels of Delivery. So we have 4 4 = 16 observations for each
level of Type.
The wrong approach: The two-sample t-test is
y1 y2
p
s12 2/(4 4)
For the above example,
y1 y2 = 0.047

CHAPTER 6. FACTORIAL DESIGNS

2 22

2 2

1 1 1 111 111

0.1

2 2 2 22

11 1

0.2

131

2 2

0.3
rate

0.4

0.5

Figure 6.7: Comparison between types I and II, without respect to delivery.
s12 = 0.081, n m2 = 4 4 = 16.
t-statistic = -1.638, df=30, p-value=0.112.
Questions:
What is s212 estimating?
What should we be comparing the factor level differences to?
If Delivery is a known source of variation, we should compare differences
between levels of Poison type to variability within a treatment
combination,
2
i.e. . For the above example, s12 = .081, whereas sMSE = 0.00240 0.05,
a ratio of about 1.65.
y1 y2
p
= 2.7,
sMSE 2/(4 4)

p-value 0.01

Testing additive effects Let ij be the population mean in cell ij. The
relationship between the cell means model and the parameters in the interaction model are as follows:
ij = + (i ) + (j ) + (ij i j + )
= + ai + bj + (ab)ij
and so

CHAPTER 6. FACTORIAL DESIGNS

2 22

2 22

2 2 2 2 22

1 11 111 11
1

0.1

11 1

132

2 2

0.2

0.3
rate

0.4

0.5

Figure 6.8: Comparison between types I and II, with delivery in color.

j bj

i (ab)ij

ai = 0
=0
= 0 for each j,

j (ab)ij

= 0 for each i

Suppose we are interested in the difference between F1 = 1 and F1 = 2. With

the above representation we can show that
1 X
1 X
1 X
1j
2j =
(1j 2j )
m2 j
m2 j
m2 j
1 X
=
([ + a1 + bj + (ab)1j ] [ + a2 + bj + (ab)2j ])
m2 j
!
X
X
1 X
=
(a1 a2 ) +
(ab)1j
(ab)2j
m2
j
j
j
= a1 a2
and so the effect of F1 = 1 versus F1 = 2 = a1 a2 can be viewed
as a contrast of cell means. Note that a1 a2 is a contrast of treatment

CHAPTER 6. FACTORIAL DESIGNS

133

(population) means:
F2 = 1 F2 = 2 F2 = 3 F2 = 4
F1 = 1
11
12
13
14
4
1
F1 = 2
21
22
23
24
4
2
F1 = 3
31
32
33
34
4
3
3
1
3
2
3
3
3
4
12

So
a1 a2 =
1
2
= (11 + 12 + 13 + 14 )/4 (21 + 22 + 23 + 24 )/4
Like any contrast, we can estimate/make inference for it using contrasts of
sample means:
a1 a2 = a
1 a
2 = y1 y2 is an unbiased estimate of a1 a2
Note that this estimate is the corresponding contrast among the m1 m2
sample means:
F2 = 1 F2 = 2 F2 = 3 F2 = 4
F1 = 1
y11
y12
y13
y14
4
y1
F1 = 2
y21
y22
y23
y24
4
y2
F1 = 3
y31
y32
y33
y34
4
y3
3
y1
3
y2
3
y3
3
y4
12
y
So
a
1 a
2 = y1 y2
= (
y11 + y12 + y13 + y14 )/4 (
y21 + y22 + y23 + y24 )/4
Hypothesis tests and confidence intervals can be made using the standard
assumptions:
E[
a1 a
2 ] = a1 a2
Under the assumption of constant variance:
Var[
a1 a
2 ] =
=
=
=

Var[
y1 y2 ]
Var[
y1 ] + Var[
y2 ]
2
/(n m2 ) + 2 /(n m2 )
2 2 /(n m2 )

CHAPTER 6. FACTORIAL DESIGNS

134

Under the assumption that the data are normally distributed

p
a
1 a
2 normal(a1 a2 , 2/(n m2 ))
and is independent of MSE (why?)
So

(
a1 a
2 ) (a1 a2 )
(
a1 a
) (a1 a2 )
q2
=
t
2
SE[
a

]
1
2
MSE nm2

where are the degrees of freedom associated with our estimate of 2 ,

i.e. the residual degrees of freedom in our model.
= m1 m2 (n 1) under the full/interaction model
= m1 m2 (n 1) + (m1 1)(m2 1) under the reduced/additive
model
Review: Explain the degrees of freedom for the two models.
t-test: Reject H0 : a1 = a2 if
a
a
2
q 1
2
MSE nm
2

> t1C /2,

r
MSE
|
a1 a
2 | >

2
t1C /2,
n m2

So the quantity
LSD1 = t1C /2, SE(
1
2)
r
2
= t1C /2, MSE
n m2
is a yardstick for comparing levels of factor 1. It is sometimes called the
least significant difference for comparing levels of Factor 1. It is analogous
to the LSD we used in the 1-factor ANOVA.
Important note: The LSD depends on which factor you are looking at:
The comparison of levels of Factor 2 depends on
Var[
y1 y2 ] = Var[
y1 ] + Var[
y2 ]
2
= /(n m1 ) + 2 /(n m1 )
= 2 2 /(n m1 )

CHAPTER 6. FACTORIAL DESIGNS

135

So the LSD for factor 2 differences is

LSD2 = t1C /2,
Df
pois$deliv
3
pois$type
2
pois$deliv : pois$type 6
Residuals
36

Sum Sq
0.20414
0.34877
0.01571
0.08643

r
MSE

2
n m1

Mean Sq F v a l u e
Pr(>F)
0 . 0 6 8 0 5 2 8 . 3 4 3 1 1 . 3 7 6 e 09
0 . 1 7 4 3 9 7 2 . 6 3 4 7 2 . 3 1 0 e 13
0.00262 1.0904
0.3867
0.00240

There is not very much evidence that the effects are not additive. Lets
assume there is no interaction term. If we are correct then we will have
increased the precision of our variance estimate.
Df Sum Sq
p o i s $ d e l i v 3 0.20414
pois$type
2 0.34877
R e s i d u a l s 42 0 . 1 0 2 1 4

Mean Sq F v a l u e
Pr(>F)
0 . 0 6 8 0 5 2 7 . 9 8 2 4 . 1 9 2 e 10
0 . 1 7 4 3 9 7 1 . 7 0 8 2 . 8 6 5 e 14
0.00243

So there is strong evidence against the hypothesis that the additive effects
are zero for either factor. Which treatments within a factor are different from
each other?

effects of poison type: At C = 0.05,

LSDtype = t.975,42 SE[
a1 a
2 ]
r
2
= 2.018 .0024
44
= 2.018 0.0173 = 0.035
Type Mean
I
0.18
II
0.23
III
0.38

LSD Grouping
1
2
3

effects of poison delivery: At C = 0.05,

LSDtype = t.975,42 SE[b1 b2 ]
r
2
= 2.018 .0024
43
= 2.018 0.02 = 0.040

0.2

0.3

0.4

0.5

136

0.1

0.1

0.2

0.3

0.4

0.5

CHAPTER 6. FACTORIAL DESIGNS

II

III

Figure 6.9: Marginal plots of the data.

Type Mean
B
0.19
D
0.21
0.29
C
A
0.35

LSD Grouping
1
1
2
3

Note the differences between these comparisons and those from two-sample
t-tests.
Interpretation of estimated additive effects: If the additive model is
clearly wrong, can we still interpret additive effects? The full model is
yijk = ij + ijk
A reparameterization of this model is the interaction model:
yijk = + ai + bj + (ab)ij + ijk
where
=

1
m1 m2

P P
i

ij =

CHAPTER 6. FACTORIAL DESIGNS

ai =

1
m2

j (ij

) =
i

bj =

1
m1

i (ij

) =
j

137

(ab)ij = ij
i
j +
= ij ( + ai + bj )
The terms {a1 , . . . , am1 } , {b1 , . . . , bm2 } are sometimes called main effects.
The additive model is
yijk = + ai + bj + ijk
Sometimes this is called the main effects model. If this model is correct,
it implies that
(ab)ij = 0 i, j
i1 j i2 j = ai1 ai2 i1 , i2 , j
ij1 ij2 = bj1 bj2 i, j1 , j2
What are the estimated additive effects estimating, in either case?
a
1 a
2 = (
y1 y ) (
y2 y )
= y1 y2
m2
m2
1 X
1 X
y1j
y2j
=
m2 j=1
m2 j=1
m2
1 X
=
(
y1j y2j )
m2 j=1

Now

(2)

m2
m2
1 X
1 X
E[
(
y1j y2j )] =
(1j 2j ) = a1 a2 ,
m2 j=1
m2 j=1

so a
1 a
2 is estimating a1 a2 regardless if additivity is correct or not. Now,
how do we interpret this effect?
Regardless of additivity, a
1 a
2 can be interpreted as the estimated
difference in response between having factor 1 =1 and factor 1=2, averaged over the experimental levels of factor 2.

CHAPTER 6. FACTORIAL DESIGNS

138

If additivity is correct, a
1 a
2 can further be interpreted as the estimated difference in response between having factor 1 =1 and factor
1=2, for every level of factor 2 in the experiment.
Statistical folklore suggests that if there is significant non-additivity, then
you cant interpret main/additive effects. As we can see, this is not true:
the additive effects have a very definite interpretation under the full model.
In some cases (block designs, coming up), we may be interested in additive
effects even if there is significant interaction.
Dissecting the interaction: Sometimes if there is an interaction, we
might want to go in and compare individual cell means. Consider the following table of means from a 2 3 two-factor ANOVA.
y11 y12 y13 y14
y21 y22 y23 y24
A large interaction SS in the ANOVA table gives us evidence, for example,
that 1j 2j varies across levels j of factor 2. It may be useful to dissect
this variability further, and understand how the non-additivity is manifested
in the data: For example, consider the three plots in Figure 4.10. These all
would give a large interaction SS, but imply very different things about the
effects of the factors.
Contrasts for examining interactions: Suppose we want to compare
the effect of (factor 1=1) to (factor 1=2) across levels of factor 2. This
involves contrasts of the form:
C = (1j 2j ) (1k 2k )
This contrast can be estimated with the sample contrast:
C = (
y1j y2j ) (
y1k y2k )
As usual, the standard error of this contrast is the estimate of its standard
deviation:
= 2 /r + 2 /r + 2 /r + 2 /r = 4 2 /r
Var[C]
p
= 2 MSE/n
SE[C]
Confidence intervals and t-tests for C can be made in the usual way.

CHAPTER 6. FACTORIAL DESIGNS

139

2.1

1.2

2.2

1.3

2.3

1.4

2.4

1.1

2.1

1.2

2.2

1.3

2.3

1.4

2.4

2.2

1.3

2.3

1.4

2.4

1.1

1.1

2.1

1.2

Figure 6.10: Three datasets exhibiting non-additive effects.

CHAPTER 6. FACTORIAL DESIGNS

6.5

140

Randomized complete block designs

Recall our first design:

CRD: to assess effects of a single factor, say F1 , on response, we randomly
allocate levels of F1 to experimental units. Typically, one hopes the e.u.s
are homogeneous or nearly so:
scientifically: If the units are nearly homogeneous, then any observed
variability in response can be attributed to variability in factor levels.
statistically: If the units are nearly homogeneous, then MSE will be
small, confidence intervals will be precise and hypothesis tests powerful.
But what if the e.u.s are not homogeneous?
Example: Let F2 be a factor that is a large source of variation/heterogeneity, but is not recorded (age of animals in experiment, gender, field plot
conditions, soil conditions).
ANOVA
SS
MS
Source
F1
SS1
MS1 =SS1/(m1 1)
(F2+Error) SS2+SSE (SS2+SSE)/N m1
F-stat for F1 may be small.

F-ratio
If SS2 is large,
MS1/MSE

If a factor
1. affects response
2. varies across experimental units
then it will increase the variance in response and also the experimental error
variance/MSE if unaccounted for. If F2 is a known, potentially large source
of variation, we can control for it pre-experimentally with a block design.
Blocking: The stratification of experimental units into groups that are more
homogeneous than the whole.
Objective: To have less variation among units within blocks than between
blocks.

CHAPTER 6. FACTORIAL DESIGNS

141

dry

wet
irrigation
Figure 6.11: Experimental material in need of blocking.
Typical blocking criteria:
location
physical characteristics
time
Example(Nitrogen fertilizer timing): How does the timing of nitrogen
additive affect nitrogen uptake?
Treatment: Six different timing schedules 1, . . . , 6: Level 4 is standard
Response: Nitrogen uptake (ppm102 )
Experimental material: One irrigated field
Soil moisture is thought to be a source of variation in response.

CHAPTER 6. FACTORIAL DESIGNS

142

5
37.99

4
37.18

1
34.98

6
34.89

3
42.07

1
41.22

3
49.42

4
45.85

6
50.15

5
41.99

2
46.69

6
44.57

3
52.68

5
37.61

1
36.94

2
46.65

4
40.23

2
41.9

4
39.2

6
43.29

5
40.45

3
42.91

1
39.97

row

2
40.89

treatment and data versus location

column
Figure 6.12: Results of the experiment
Design:
1. Field is divided into a 4 6 grid.
2. Within each row or block, each of the 6 treatments are randomly allocated.
1. The experimental units are blocked into presumably more homogeneous
groups.
2. The blocks are complete, in that each treatment appears in each block.
3. The blocks are balanced, in that there are
m1 = 6 observations for each level of block
m2 = 4 observations for each level of trt
n = 1 observation for each trtblock combination.
This design is a randomized complete block design.

40

Ni

45

50

143

35

35

40

Ni

45

50

CHAPTER 6. FACTORIAL DESIGNS

3
4
treatment

row

5
1.52

4
3.44

1
3.3

6
8.33

3
4.7

1
2.94

3
2.65

4
5.24

6
6.92

5
2.48

2
2.66

6
1.34

3
5.91

5
1.9

1
1.34

2
2.62

4
0.39

2
2.13

4
1.41

6
0.06

5
0.94

3
3.86

1
1.69

row

2
3.14

treatment and residual versus location

column

Figure 6.13: Marginal plots, and residuals without controlling for row.

CHAPTER 6. FACTORIAL DESIGNS

144

Analysis of the RCB design with one rep: Analysis proceeds just as
in the two-factor ANOVA:
yij y = (
yi y ) + (
yj y ) + (yij yi yj + y )
SSTotal =
SSTrt
+
SSB
+
SSE
ANOVA table
Source
SS
dof
Trt
SST
m1 1
Block
SSB
m2 1
Error
SSE
(m1 1)(m2 1)
Total SSTotal
m1 m2 1

MS
F-ratio
SST/(m1 1)
MST/MSE
SSB/(m2 1)
(MSB/MSE)
SSE/(m1 1)(m2 1)

ANOVA for nitrogen example:

Source
SS
dof MS F-ratio p-value
Trt
201.32 5 40.26
5.59
0.004
Block 197.00 3 65.67
9.12
Error 108.01 15 7.20
Total 506.33 23
Discussion: Consider the following three ANOVA decompositions:
#######
> anova ( lm ( c ( y ) a s . f a c t o r ( c ( t r t ) )
))
Df Sum Sq Mean Sq F v a l u e Pr(>F)
as . f a c t o r ( c ( t r t ) ) 5 201.316 40.263 2.3761 0.08024 .
Residuals
18 3 0 5 . 0 1 2 1 6 . 9 4 5
#######
#######
> anova ( lm ( c ( y ) a s . f a c t o r ( c ( t r t ) ) + a s . f a c t o r ( c ( rw ) ) ) )
Df Sum Sq Mean Sq F v a l u e
Pr(>F)
a s . f a c t o r ( c ( t r t ) ) 5 2 0 1 . 3 1 6 4 0 . 2 6 3 5 . 5 9 1 7 0 . 0 0 4 1 9 1
a s . f a c t o r ( c ( rw ) )
3 1 9 7 . 0 0 4 6 5 . 6 6 8 9 . 1 1 9 8 0 . 0 0 1 1 1 6
Residuals
15 1 0 8 . 0 0 8
7.201
#######
#######
> anova ( lm ( c ( y ) a s . f a c t o r ( c ( t r t ) ) : a s . f a c t o r ( c ( rw ) ) ) )
Df Sum Sq Mean Sq F v a l u e Pr(>F)

CHAPTER 6. FACTORIAL DESIGNS

145

a s . f a c t o r ( c ( t r t ) ) : a s . f a c t o r ( c ( rw ) ) 23 5 0 6 . 3 3
Residuals
0
0.00
#######

22.01

Can we test for interaction? Do we care about interaction in this case, or just
main effects? Suppose it were true that in row 2, timing 6 is significantly
better than timing 4, but in row 3, treatment 3 is better. Is this relevant in
for recommending a timing treatment for other fields?
Did blocking help? Consider CRD as an alternative:
block
block
block
block

1
2
3
4

2
5
6
1

4
5
3
2

3
3
4
6

2
4
2
2

1
1
6
5

4
4
5
6

Advantages:
more possible treatment assignments, so power is increased in a
randomization test.
If we dont estimate block effects, well have more dof for error.
Disadvantages:
It is possible, (but unlikely) that some treatment level will get assigned many times to a good row, leading to post-experimental
bias.
If row is a big source of variation, then ignoring it may lead to
an overly large MSE.
Consider comparing the F -statistic from a CRD with that from an RCB:
According to Cochran and Cox (1957)
SSB + n(m 1)MSErcbd
1
 nm 


n1
n(m 1)
= MSB
+ MSErcbd
nm 1
nm 1

MSEcrd =

In general, the effectiveness of blocking is a function of MSEcrd /MSErcb . If

this is large, it is worthwhile to block. For the nitrogen example, this ratio
is about 2.

CHAPTER 6. FACTORIAL DESIGNS

6.6

146

Unbalanced designs

Example: Observational study of 20 fatal accidents.

Response: y = speed in excess of speed limit
Recorded sources of variation:
1. R =rainy (rainy/not rainy)
2. I =interstate (interstate/two-lane highway),

rainy
not rainy
sum
marginal mean

cell means
interstate two-lane
15
5
n11 = 8
n12 = 2
20
10
n21 = 2
n22 = 8
160
90
n1 = 10 n2 = 10
16
9

sum

marginal means

130
n1 = 10
120
n1 = 10
250
n = 20

13
12

y = 12.5

Lets naively compute sums of squares based on the decomposition:

yijk = y + (
yij y ) + (yijk yij )
yijk = y + (
yi y ) + (
yj y ) + (
yij yi yj + y ) + (yijk yij )

SSF =

XXX
i

(
yij y )2 =

k
2

XX
i
2

nij (
yij y )2

= 8(15 12.5) + 2(5 12.5) + 2(20 12.5)2 + 8(10 12.5)2 = 325

SS1 = 10 (13 12.5)2 + 10 (12 12.5)2 = 5
SS2 = 10 (16 12.5)2 + 10 (10 12.5)2 = 245
Previously we had SS1+SS2+SS12 =SSF, so it seems we should have
SS12 = SSF SS1 SS2 = 325 5 245 = 75
This suggests some interaction. But look at the cell means:
no rain - rain =5 regardless of interstate

CHAPTER 6. FACTORIAL DESIGNS

147

interstate - two=lane =10 regardless of rain

There is absolutely no interaction! The problem is that SS1, SS2, SS12 are
not orthogonal (as computed this way) and so SSF 6= SS1 + SS2 + SSR12.
There are other things to be careful about too:
Unbalanced marginal means: Supposed someone asked, are the accident speeds higher on rainy days or on clear days?
yrain yclear = 1312 = 1: Marginal means suggest speeds were slightly
higher on average during rainy days than on clear days.
Cell means show, for both road types, speeds were higher on clear
days by 5 mph on average.
Explanation: Cell frequencies are varying.
Marginal means for rain are dominated by interstate accidents
Marginal means for no rain are dominated by two-lane accidents
We say that the marginal effects are unbalanced. How can we make sense
of marginal effects in such a situation? How can we test for non-additivity?
Least-squares means: The solution to this paradox is to use least squares
means, which are based on on the cell means model:
yijk = ij + ijk

ij =

1 X
yijk ,
nij k

2 = s2 =

XXX
1
(yijk yij )2
N m1 m2 i j k

The least-squares marginal means are:

The idea:

m2
1 X

ij 6 yi
=
m2 j=1

m1
1 X
=

ij 6 yj
m1 i=1

CHAPTER 6. FACTORIAL DESIGNS

148

1. get estimates for each cell

2. average cell estimates to get marginal estimates.
1
2

1
y11
y21
..
.

2
y12
y22
..
.

m1 ym1 1 ym1 2

m2
y1m2
y2m2

ym1 m2
m1

m2

Accident example:

rainy
not rainy
marginal mean
LS mean

interstate two-lane
15
5
20
10
16
9
17.5
7.5

marginal mean
13
12

LS mean
10
15

Comparisons between means can be made in the standard way:

Standard errors:
i =

1
m2

yij , so
1 X 2
Var[i ] =
/nij
m22 j
s
1 X MSE
SE[
i ] =
m2
nij
j
s
1 X MSE
SE[
j ] =
m1
nij
i
j

and similarly

Now use the SE to obtain confidence intervals, t-tests, etc.

Testing for interaction: Consider
H0 : Yijk = + ai + bj + ijk
H1 : Yijk = + ai + bj + (ab)ij + ijk
How do we evaluate H0 when the data are unbalanced? Well first outline
the procedure, then discuss why it works.

CHAPTER 6. FACTORIAL DESIGNS

149

P P P
1. Compute SSEF = min,a,b,(ab) i j k (yijk [ + ai + bj + (ab)ij ])2
P P P
2. Compute SSEA = min,a,b i j k (yijk [ + ai + bj ])2
3. Compute SS12 = SSEA SSEF . Note that this is always positive.
Allowing for interaction improves the fit, and reduces error variance. SSI
measures the improvement in fit. If SSI is large, i.e. SSEA is much bigger than
SSEF , this suggests the additive model does not fit well and the interaction
term should be included in the model.
Testing:
F =

SSI/(m1 1)(m2 1)
MSI
=
MSE
SSEF /(N m1 m2 )

Under H0 , F F(m1 1)(m2 1),N m1 m2 , so a level- test of H0 is

reject H0 if F > F1,(m1 1)(m2 1),N m1 m2 .
Note:
SSI is the change in fit in going from the additive to the full model;
(m1 1)(m2 1) is the change in number of parameters in going from
the additive to the full model.
A painful example: A small scale clinical trial was done to evaluate the
effect of painkiller dosage on pain reduction for cancer patients in a variety
of age groups.
Factors of interest:
Dose { Low, Medium, High }
Age { 41-50, 51-60, 61-70, 71-80 }
Response = Change in pain index {5, 5}
Design: CRD, each treatment level was randomly assigned to ten patients, not blocked by age.

150

CHAPTER 6. FACTORIAL DESIGNS

50

60

70

80

Figure 6.14: Marginal plots for pain data

> t a b l e ( t r t , ageg )
ageg
t r t 50 60 70 80
1 1 2 3 4
2 1 3 3 3
3 2 1 4 3
> t a p p l y ( y , t r t , mean )
1
2
3
0 . 3 8 1 0.950 2.131
>
> t a p p l y ( y , ageg , mean )
50
60
70
80
2.200 1.265 0.922 0.139

Do these marginal plots and means misrepresent the data? To evaluate this
possibility,
compare the marginal plots in Figure 6.14 to the interaction plots in
Figure 6.15;
compute the LS means and compare to marginal means.

151

CHAPTER 6. FACTORIAL DESIGNS

1.50

3.50

2.60

1.70

3.70

2.80

Figure 6.15: Interaction plots for pain data

> c e l l m e a n s < t a p p l y ( y , l i s t ( t r t , ageg ) , mean )
> cellmeans
50
60
70
80
1 4.90 1 . 4 0 0 0 0 0 1 . 0 6 6 6 6 7 0 . 6 7 7 5 0 0
2 2 . 4 0 2.996667 1.270000 0 . 3 0 0 0 0 0
3 3.15 1.400000 2.152500 1.666667
> t r t l s m < apply ( c e l l m e a n s , 1 , mean )
> age lsm < apply ( c e l l m e a n s , 2 , mean )
> trt lsm
1
2
3
0.4389583 0.3916667 2.0922917
> age lsm
50
60
70
80
1.8833333 0.9988889 0.7852778 0.2297222

What are the differences between LS means and marginal means? Not as
extreme as in the accident example, but the differences can be explained by
looking at the interaction plot, and the slight imbalance in the design:
The youngest patients (ageg=50) were imbalanced towards the higher

CHAPTER 6. FACTORIAL DESIGNS

152

dose, so we might expect their marginal mean to be too low. Observe

the change from the marginal mean = -2.2 to the LS mean = -1.883.
The oldest patients (ageg=80) were imbalanced towards the lower dose,
so we might expect their marginal mean to be too high. Observe the
change from the marginal mean = -.14 to the LS mean = -.23 .
Lets look at the main effects in our model:
> t r t c o e f < t r t l s m mean ( c e l l m e a n s )
> a g e c o e f < age lsm mean ( c e l l m e a n s )
> trt coef
1
0.5353472

2
3
0 . 5 8 2 6 3 8 9 1.1179861

> age coef

50
60
0.90902778 0.02458333

70
0.18902778

80
0.74458333

What linear modeling commands in R will get you the same thing?
> o p t i o n s ( c o n t r a s t s=c ( c o n t r . sum , c o n t r . p o l y ) )
> f i t f u l l <lm ( y a s . f a c t o r ( ageg ) a s . f a c t o r ( t r t ) )
> fit full$coef [2:4]
a s . f a c t o r ( ageg ) 1 a s . f a c t o r ( ageg ) 2 a s . f a c t o r ( ageg ) 3
0.90902778
0.02458333
0.18902778

> fit full$coef [5:6]

as . f a c t o r ( t r t )1 as . f a c t o r ( t r t )2
0.5353472
0.5826389

Note that the coefficients in the reduced/additive model are not the same:
> f i t a d d <lm ( y a s . f a c t o r ( ageg )+ a s . f a c t o r ( t r t ) )
> fit add$coef [2:4]
a s . f a c t o r ( ageg ) 1 a s . f a c t o r ( ageg ) 2 a s . f a c t o r ( ageg ) 3
0.7921041
0.3593577
0.3049595
> fit add$coef [5:6]
as . f a c t o r ( t r t )1 as . f a c t o r ( t r t )2
1.208328354
0.002085645

CHAPTER 6. FACTORIAL DESIGNS

6.7

153

Non-orthogonal sums of squares:

Consider the following ANOVA table obtained from R:

> anova (

lm ( y a s . f a c t o r ( ageg )+ a s . f a c t o r ( t r t ) ) )
Df Sum Sq Mean Sq F v a l u e Pr(>F)
a s . f a c t o r ( ageg ) 3 1 3 . 3 5 5
4.452 0.9606 0.42737
as . f a c t o r ( t r t )
2 28.254 14.127 3.0482 0.06613 .
Residuals
24 1 1 1 . 2 3 0
4.635

It might be somewhat unsettling that R also produces the following table:

> anova (

lm ( y a s . f a c t o r ( t r t )+ a s . f a c t o r ( ageg ) ) )
Df Sum Sq Mean Sq F v a l u e Pr(>F)
as . f a c t o r ( t r t )
2 31.588 15.794 3.4079 0.0498
a s . f a c t o r ( ageg ) 3 1 0 . 0 2 1
3.340 0.7207 0.5494
Residuals
24 1 1 1 . 2 3 0
4.635

Where do these sums of squares come from? What do the F -tests represent?
By typing ?anova.lm in R we see that anova() computes
a sequential analysis of variance table for that fit. That is, the
reductions in the residual sum of squares as each term of the formula
is added in turn are given in as the rows of a table, plus the residual
sum of squares.

Sequential sums of squares: Suppose in a linear model we have three

sets of parameters, say A, B and C. For example, let
A be the main effects of factor 1
B be the main effects of factor 2
C be their interaction.
Consider the following calculation:
0. Calculate SS0 = residual sum of squares from the model
(0) yijk = + ijk
1. Calculate SS1 = residual sum of squares from the model
(A) yijk = + ai + ijk

CHAPTER 6. FACTORIAL DESIGNS

154

2. Calculate SS2 = residual sum of squares from the model

(AB) yijk = + ai + bj + ijk
3. Calculate SS3 = residual sum of squares from the model
(ABC) yijk = + ai + bj + (ab)ij + ijk
We can assess the importance of A, B and C sequentially as follows:
SSA = improvement in fit in going from (0) to (A) = SS0-SS1
SSBA = improvement in fit in going from (A) to (AB) = SS1-SS2
SSCAB = improvement in fit in going from (AB) to (ABC) = SS2SS3
This is actually what R presents in an ANOVA table:
>
>
>
>

s s 0 <sum (
s s 1 <sum (
s s 2 <sum (
ss3<

lm ( y 1 ) $ r e s 2 )
lm ( y a s . f a c t o r ( ageg ) ) $ r e s 2 )
lm ( y a s . f a c t o r ( ageg )+ a s . f a c t o r ( t r t ) ) $ r e s 2 )

> s0s s 1
[ 1 ] 13.3554
>
> s s 1 s s 2
[ 1 ] 28.25390
>
> s s 2 s s 3
[ 1 ] 53.75015
> ss3
[ 1 ] 57.47955
lm ( y a s . f a c t o r ( ageg ) a s . f a c t o r ( t r t ) ) )
Df Sum Sq Mean Sq F v a l u e Pr(>F)
a s . f a c t o r ( ageg )
3 13.355
4.452 1.3941 0.27688
as . f a c t o r ( t r t )
2 28.254 14.127 4.4239 0.02737
a s . f a c t o r ( ageg ) : a s . f a c t o r ( t r t ) 6 5 3 . 7 5 0
8.958 2.8054 0.04167
Residuals
18 5 7 . 4 8 0
3.193

> anova (

Why does order of the variables matter?

CHAPTER 6. FACTORIAL DESIGNS

155

In a balanced design, the parameters are orthogonal, and SSA = SSAB

, SSB = SSBA and so on, so the order doesnt matter.
In an unbalanced design, the estimates of one set of parameters depends
on whether or not you are estimating the others, i.e. they are not
orthogonal, and in general SSA 6= SSAB , SSB 6= SSBA.
I will try to draw a picture of this on the board.
The bottom line: For unbalanced designs, there is no variance due to
factor 1 or variance due to factor 2. There is only extra variance due
to factor 1, beyond that explained by factor 2, and vice versa. This is
essentially because of the non-orthogonality, and so the part of the variance
that can be explained by factor 1 overlaps with the part that can be explained
by factor 2. This will become more clear when you get to regression next
quarter.

6.8

Analysis of covariance

Example(Oxygen ventilation): Researchers are interested in measuring

the effects of exercise type on maximal oxygen uptake.
Experimental units: 12 healthy males between 20 and 35 who didnt exercise regularly.
Design: Six subjects randomly selected to a 12-week step aerobics program,
the remaining to 12-weeks of outdoor running on flat terrain.
Response: Before and after the 12 week program, each subjects O2 uptake
was tested while on an inclined treadmill.
y = change in O2 uptake
Initial analysis: CRD with one two-level factor. The first thing to do is
plot the data. The first panel of Figure 6.16 indicates a moderately large
difference in the two sample populations. The second thing to do is a twosample t-test:


 y y   7.705 (2.767) 
 A
B
 = 2.907, Pr(|t10 | 2.907) = 0.0156
tobs =  p
=
 sp 2/6   6.239 .577 

CHAPTER 6. FACTORIAL DESIGNS

156

residuals
0
5

o2_change
0
5 10

15

10

AA

10 5

A
B

A
A
B

20

22

grp

24

26
age

28

30

Figure 6.16: Oxygen uptake data

So we have reasonably strong evidence of a difference in treatment effects.
However, a plot of residuals versus age of the subject indicates a couple of
causes for concern with our inference:
1. The residual plot indicates that response increases with age (why?),
regardless of treatment group.
2. Just due to chance, the subjects assigned to group A were older than
the ones assigned to B.
Question: Is the observed difference in yA , yB due to
treatment?
age?
both?
A linear model for ANCOVA: Let yi,j be the response of the jth subject
in treatment i:
yi,j = + ai + b xi,j + i,j

CHAPTER 6. FACTORIAL DESIGNS

A
A B
B

B
B
20

22

24

26
age

28

AA
B

o2_change
0
5 10

10 5

10 5

o2_change
0
5 10

AA
B

15

15

157

30

A
A
A B
B

B
B
20

22

24

26
age

28

30

Figure 6.17: ANOVA and ANCOVA fits to the oxygen uptake data
This model gives a linear relationship between age and response for each
group:
intercept
slope
if i = A, Yi = ( + aA ) + b xi,j
i = B, Yi = ( + aB ) + b xi,j

error
+ i,j
+ i,j

Unbiased parameter estimates can be obtained by minimizing the residual

sum of squares:
X
(
, a
, b) = arg min
(yi,j [ + ai + b xi,j ])2
,a,b

i,j

The fitted model is shown in the second panel of Figure 6.17.

Variance decomposition Consider the following two ANOVAs:
> anova ( lm ( o 2 c h a n g e grp ) )
Df Sum Sq Mean Sq F v a l u e Pr(>F)
grp
1 328.97 328.97 8.4503 0.01565
R e s i d u a l s 10 3 8 9 . 3 0
38.93

CHAPTER 6. FACTORIAL DESIGNS

158

> anova ( lm ( o 2 c h a n g e grp+age ) )

Df Sum Sq Mean Sq F v a l u e
Pr(>F)
grp
1 3 2 8 . 9 7 3 2 8 . 9 7 4 2 . 0 6 2 0 . 0 0 0 1 1 3 3
age
1 3 1 8 . 9 1 3 1 8 . 9 1 4 0 . 7 7 6 0 . 0 0 0 1 2 7 4
Residuals 9 70.39
7.82

The second one decomposes the variation in the data that is orthogonal to
treatment (SSE from the first ANOVA) into a parts that can be ascribed to
age (SS age in the second ANOVA), and everything else (SSE from second
ANOVA). I will try to draw some triangles that describe this situation.
Now consider two other ANOVAs:
> anova ( lm ( o 2 c h a n g e age ) )
Df Sum Sq Mean Sq F v a l u e
Pr(>F)
age
1 5 7 6 . 0 9 5 7 6 . 0 9 4 0 . 5 1 9 8 . 1 8 7 e 05
R e s i d u a l s 10 1 4 2 . 1 8
14.22

> anova ( lm ( o 2 c h a n g e age+grp ) )

Df Sum Sq Mean Sq F v a l u e
Pr(>F)
age
1 5 7 6 . 0 9 5 7 6 . 0 9 7 3 . 6 5 9 4 1 . 2 5 7 e 05
grp
1 71.79
71.79 9.1788
0.01425
Residuals 9 70.39
7.82

Again, I will try to draw some triangles describing this situation.

Blocking, ANCOVA and unbalanced designs: Suppose we have a factor of interest (say F 1) that we will randomly assign to experimental material,
but it is known that there is some nuisance factor (say F 2 ) that is suspected
to be a large source of variation. Our options are:
Block according to F 2: This allows an even allocation of treatments
across levels of F 2. As a result, we can separately estimate the effects
of F 1 and F 2.
Do a CRD with respect to F 1, but account for F 2 somehow in the
analysis:
For a standard two-factor ANOVA, F 2 must be treated as a categorical variable (old, young).
ANCOVA allows for more precise control of variation due to F 2.

CHAPTER 6. FACTORIAL DESIGNS

2.5
2.0

A
A
B
A

B
B
B
A

B
1

f2

0.5

1.0

A
B
A

B
B
B
A

0.0

0.5

1.0

0.0

B
B
B
A

0.0

0.5

1.0

1.5

2.0

A
B
A

1.5

2.0

1.5

2.5

2.5

159

B
1

f2

2
f2

Figure 6.18: Unbalanced design: Controlling eliminates effect.

However, in either of these approaches the design is unlikely to be completely
balanced , and so the effects of F 1 cant be estimated separately from those
of F 2. This is primarily a problem for experiments with small sample sizes:
The probability of design imbalance decreases as a function of sample size
(as does the correlation between the parameter estimates) as long as F 1 is
randomly assigned.

6.9

Types of sums of squares

Figure 6.18 gives a simple example where an unbalanced design can lead to
results that are difficult to interpret.
Two factors:
F1 = treatment (A vs B)
F2 = block (location 1 versus 2)
10 experimental units
Balanced CRD in F1, but not in F1F2.
Looking at things marginally in F1, we have yA > yB , and there seems to be
an effect of F1=A versus B. This is highlighted in the second panel of the
figure, which shows the difference between the sample marginal means and
the grand mean seems large compared to error variance.

CHAPTER 6. FACTORIAL DESIGNS

160

> anova ( lm ( y f 1+f 2 ) )

Df Sum Sq Mean Sq F v a l u e
Pr(>F)
f1
1 3 . 5 8 2 4 3 . 5 8 2 4 2 1 . 5 7 7 0 . 0 0 2 3 5 5
f2
1 4 . 2 9 7 1 4 . 2 9 7 1 2 5 . 8 8 2 0 . 0 0 1 4 1 9
Residuals 7 1.1622 0.1660

However, notice the imbalance:

4 A observations under F2=1, 1 under F2=2
1 B observations under F2=1, 4 under F2=2
What happens if we explain the variability in y using F2 first, then F1?
Look at the black lines in the third panel of the plot: These are the marginal
means for F2. Within levels of F2, differences between levels of F1 are small.
> anova ( lm ( y f 2+f 1 ) )
Df Sum Sq Mean Sq F v a l u e
Pr(>F)
f2
1 7 . 8 0 6 4 7 . 8 0 6 4 4 7 . 0 1 8 5 0 . 0 0 0 2 4 0 5
f1
1 0.0731 0.0731 0.4405 0.5281460
Residuals 7 1.1622 0.1660

The ANOVA quantifies this: There is variability in the data that can be
explained by either F1 or F2. In this case,
SSA > SSA|B
SSB > SSB|A
Do these inequalities always hold? Consider the data in Figure 6.19. In this
case F1=A is higher than F1=B for both values of F2. But there are more
A observations in the low-mean value of F2 than the high-mean value. The
second an third plots suggest
Marginally, the difference between levels of F1 are small.
Within each group, the difference between levels of F1 are larger.
Thus controlling for F2 highlights the differences between levels of F1.
This is confirmed in the corresponding ANOVA tables:
> anova ( lm ( y f 1+f 2 ) )
Df Sum Sq Mean Sq F v a l u e
Pr(>F)
f1
1 0.1030 0.1030 0.5317 0.4895636
f2
1 5 . 7 8 5 1 5 . 7 8 5 1 2 9 . 8 7 7 2 0 . 0 0 0 9 3 9 9
Residuals 7 1.3554 0.1936

2.5

A
B
B
B

2.0

B
B
B
A

2
f2

1.5

0.5

B
1

1.0

A
A
A

0.5

1.0

A
A

0.0

1.5

0.0

0.5

B
B
B
A

0.0

1.0

A
A

1.5

161

2.0

2.5

2.0

2.5

CHAPTER 6. FACTORIAL DESIGNS

B
1

f2

2
f2

Figure 6.19: Unbalanced design: Controlling highlights effect.

> anova ( lm ( y f 2+f 1 ) )

Df Sum Sq Mean Sq F v a l u e
Pr(>F)
f2
1 4 . 4 8 0 4 4 . 4 8 0 4 2 3 . 1 3 9 1 0 . 0 0 1 9 4 3
f1
1 1.4077 1.4077 7.2698 0.030806
Residuals 7 1.3554 0.1936

Which ANOVA table to use? Some software packages combine these anova
tables to form ANOVAs based on alternative types of sums of squares.
Consider a two-factor ANOVA in which we plan on decomposing variance
into additive effects of F1, additive effects of F2, and their interaction.
Type I SS: Sequential, orthogonal decomposition of the variance.
Type II SS: Sum of squares for a factor is the improvement in fit from
adding that factor, given inclusion of all other terms at that level or
below.
Type III SS: Sum of squares for a factor is the improvement in fit from
adding that factor, given inclusion of all other terms.
So for example:
SSF11 = RSS(0) RSS(F1) if F1 first in sequence
SSF11 = RSS(F2) RSS(F1 + F2) if F1 second in sequence

CHAPTER 6. FACTORIAL DESIGNS

162

SSF12 = RSS(F2) RSS(F1 + F2)

SSF13 = RSS(F2, F1 : F2) RSS(F1, F2, F1 : F2)
Type II sums of squares is very popular, and is to some extent the default
for linear regression analysis. It seems natural to talk about the variability
due to a treatment, after controlling for other sources of variation. However, there are situations in which you might not want to control for other
variables. Consider the following (real-life) scenario:
Clinical trial:
Factor 1: Assigned dose of a drug (high or low)
Factor 2: Dose reduction (yes or no)
Weaker patients experienced higher rates of drug side-effects than stronger
patients, and side effects are worse under the high dose. As a result,
Only the strongest A patients remain in the non-reduced group;
Only the weakest B patients go into the reduced group.
Controlling or adjusting for dose reduction artificially inflates the perceived
effect of treatment.

Chapter 7
Nested Designs
Example(Potato): Sulfur added to soil kills bacteria, but too much sulfur
can damage crops. Researchers are interested in comparing two levels of
sulfur additive (low, high) on the damage to two types of potatoes.
Factors of interest:
1. Potato type {A, B}
2. Sulfur additive {low,high}
Experimental material: Four plots of land.
Design constraints:
It is easy to plant different potato types within the same plot
It is difficult to have different sulfur treatments in the same plot,
due to leeching.
Experimental Design: A Split-plot design
1. Each sulfur additive was randomly assigned to two of the four
plots.
2. Each plot was split into four subplots. Each potato type was
randomly assigned to two subplots per plot.

163

CHAPTER 7. NESTED DESIGNS

164

A B
A B

B A
A B

B A
A B

B A
A B

Randomization:
Sulfur type was randomized to whole plots;
Potato type was randomized to subplots.
Initial data analysis: Sixteen responses, 4 treatment combinations.
8 responses for each potato type
8 responses for each sulfur type
4 responses for each potatotype combination
> f i t . f u l l <lm ( y type s u l f u r ) ;
> anova ( f i t . f u l l )
Df Sum Sq Mean Sq
type
1 1.48840 1.48840
sulfur
1 0.54022 0.54022
type : s u l f u r 1 0 . 0 0 3 6 0 0 . 0 0 3 6 0
Residuals
12 1 . 3 2 8 3 5 0 . 1 1 0 7 0

f i t . add<lm ( y type+s u l f u r )

F value
Pr(>F)
1 3 . 4 4 5 9 0 . 0 0 3 2 2 5
4.8803 0.047354
0.0325 0.859897

> anova ( f i t . add )

Df Sum Sq Mean Sq F v a l u e
Pr(>F)
type
1 1 . 4 8 8 4 0 1 . 4 8 8 4 0 1 4 . 5 2 7 0 0 . 0 0 2 1 6
sulfur
1 0.54022 0.54022 5.2727 0.03893
R e s i d u a l s 13 1 . 3 3 1 9 5 0 . 1 0 2 4 6

Randomization test: Consider comparing the observed outcome to the

population of other outcomes that couldve occurred under
different treatment assignments;

5.0

5.5

165

4.5

4.5

5.0

5.5

CHAPTER 7. NESTED DESIGNS

high

low

B
type

4.5

5.0

5.5

sulfur

high.A

low.A

high.B

Figure 7.1: Potato data.

low.B

CHAPTER 7. NESTED DESIGNS

166

0.4

Sample Quantiles
0.0 0.2 0.4

fit.add$res
0.0 0.2

0.4

0.4

1
0
1
Theoretical Quantiles

4.6

4.8 5.0 5.2

fit.add$fit

5.4

Figure 7.2: Diagnostic plots for potato ANOVA.

no treatment effects.
Treatment assignment
1

..
.

..
.

Field 1
low
A B
low
B A
low
B A
low
B A
low
B A
..
.
.. ..
. .

B A
B A
B A
B A
B A
..
.

..
.

Number of possible treatment assignments:

42 = 6 ways to assign sulfur

Field 2
low
A B
low
A B
low
B A
low
A B
high
A B
..
.
.. ..
. .

B A
B A
B A
B A
B A
..
.

..
.

Field 3
high
A B
high
A B
high
A B
high
B A
low
A B
..
.
.. ..
. .

B A
B A
B A
B A
B A
..
.

..
.

Field 4
high
A B
high
A B
high
A B
high
A B
high
A B
..
.
.. ..
. .

B
B
B
B
B
..
.

167

0.0

0.00

0.1

Density
0.05
0.10

Density
0.2 0.3

0.4

0.15

CHAPTER 7. NESTED DESIGNS

10
15
F.t.null

20

10
F.s.null

15

Figure 7.3: Potato data

For each sulfur assignment there are

4 4
2

= 1296 type assignments

So we have 7776 possible treatment assignments. It probably wouldnt be to

hard to write code to go through all possible treatment assignments, but its
very easy to obtain a Monte Carlo approximation to the null distribution:
F . t . n u l l <F . s . n u l l <NULL
f o r ( ns i n 1 : 1 0 0 0 ) {
s . sim<r e p ( sample ( c ( low , low , h i g h , h i g h ) ) , r e p ( 4 , 4 ) )
t . sim<c ( sample ( c ( A , A , B , B ) ) , sample ( c ( A , A , B , B ) ) ,
sample ( c ( A , A , B , B ) ) , sample ( c ( A , A , B , B ) )
)
f i t . sim<anova ( lm ( y a s . f a c t o r ( t . sim)+ a s . f a c t o r ( s . sim ) ) )
F . t . n u l l <c (F . t . n u l l , f i t . sim [ 1 , 4 ] )
F . s . n u l l <c (F . s . n u l l , f i t . sim [ 2 , 4 ] )
}
> mean (F . t . n u l l >=F . t . obs )
[ 1 ] 0.001
> mean (F . s . n u l l >=F . s . obs )
[ 1 ] 0.352

CHAPTER 7. NESTED DESIGNS

168

What happened?
rand
anova1
Ftype
F1,13 prand
type ptype
anova1
rand
6 F1,13 prand
Fsulfur
sulfur 6 psulfur

The F -distribution approximates a null randomization distribution if treatments are randomly assigned to units. But here, the sulfur treatment is
being assigned to groups of four units.
The precision of an effect is related to the number of independent
treatment assignments made
We have 16 assignments of type, but only 4 assignments of sulfur. It
is difficult to tell the difference between sulfur effects and field effects.
Our estimates of sulfur effects are less precise than those of type.
Note:
From the point of view of type alone, the design is a RCB.
Each whole-plot(field) is a block. We have 2 observations of each
type per block.
We compare MSType to the MSE from residuals left from a model
with type effects, block effects and possibly interaction terms.
Degrees of freedom breakdown for the sub-plot analysis:
Source
dof
block=whole plot 3
type
1
subplot error
11
subplot total
15
From the point of view of sulfur alone, the design is a CRD.
Each whole plot is an experimental unit.
We want to compare MSSulfur to the variation in whole plots,
which are fields.
Degrees of freedom breakdown for the whole-plot analysis:
Source
dof
sulfur
1
whole plot error 2
whole plot total 3

CHAPTER 7. NESTED DESIGNS

169

Recall, degrees of freedom quantify the precision of our estimates and

the shape of the null distribution.
The basic idea is:
We have different levels of experimental units (fields/whole-plots for
sulfur, subfields/sub-plots for type).
We compare the MS of a factor to the variation among the experimental
units of that factors level. In general, this variation is represented
by the interaction between the experimental units label and all factors
assigned at the given level.
The level of an interaction is the level of the smallest experimental unit
involved.
> anova ( lm ( y type+type : s u l f u r+a s . f a c t o r ( f i e l d ) ) )
Df Sum Sq Mean Sq F v a l u e
Pr(>F)
type
1 1 . 4 8 8 4 0 1 . 4 8 8 4 0 5 4 . 7 0 0 5 2 . 3 2 6 e 05
a s . f a c t o r ( f i e l d ) 3 1 . 5 9 6 4 8 0 . 5 3 2 1 6 1 9 . 5 5 7 5 0 . 0 0 0 1 6 6 1
type : s u l f u r
1 0.00360 0.00360 0.1323 0.7236270
Residuals
10 0 . 2 7 2 1 0 0 . 0 2 7 2 1

Thus there is strong evidence for type effects, and little evidence that the
effects of type vary among levels of sulfur.

> anova ( lm ( y s u l f u r+a s . f a c t o r ( f i e l d ) ) )

Df Sum Sq Mean Sq F v a l u e Pr(>F)
sulfur
1 0.54022 0.54022 3.6748 0.07936 .
as . f a c t o r ( f i e l d ) 2 1.05625 0.52813 3.5925 0.05989 .
Residuals
12 1 . 7 6 4 1 0 0 . 1 4 7 0 1

The F-test here is not appropriate: We need to compare MSSulfur to the

variation among whole-plots, after accounting for the effects of sulfur, i.e.
MSField (note that including or excluding type here has no effect on this
comparison).
MSS<anova ( lm ( y s u l f u r+a s . f a c t o r ( f i e l d ) ) ) [ 1 , 3 ]
MSWPE<anova ( lm ( y s u l f u r+a s . f a c t o r ( f i e l d ) ) ) [ 2 , 3 ]
F . s u l f u r <MSS/MSWPE

CHAPTER 7. NESTED DESIGNS

170

> F. sulfur
[ 1 ] 1.022911
> 1p f (F . s u l f u r , 1 , 2 )
[ 1 ] 0.4182903

This is more in line with the analysis using the randomization test.
The above calculations are somewhat tedious. In R there are several automagic ways of obtaining the correct F -test for this type of design. One
way is with the aov command:
> f i t 1 <aov ( y type s u l f u r + E r r o r ( f a c t o r ( f i e l d ) ) )
> summary ( f i t 1 )
Error : f a c t o r ( f i e l d )
Df Sum Sq Mean Sq F v a l u e Pr(>F)
sulfur
1 0.54022 0.54022 1.0229 0.4183
Residuals 2 1.05625 0.52813
E r r o r : Within
Df Sum Sq
type
1 1.48840
type : s u l f u r 1 0 . 0 0 3 6 0
Residuals
10 0 . 2 7 2 1 0

Mean Sq F v a l u e
Pr(>F)
1 . 4 8 8 4 0 5 4 . 7 0 0 5 2 . 3 2 6 e 05
0.00360 0.1323
0.7236
0.02721

###
> f i t 2 <aov ( y type + s u l f u r + E r r o r ( f a c t o r ( f i e l d ) ) )
> summary ( f i t 2 )
Error : f a c t o r ( f i e l d )
Df Sum Sq Mean Sq F v a l u e Pr(>F)
sulfur
1 0.54022 0.54022 1.0229 0.4183
Residuals 2 1.05625 0.52813
E r r o r : Within
Df Sum Sq Mean Sq F v a l u e
Pr(>F)
type
1 1 . 4 8 8 4 0 1 . 4 8 8 4 0 5 9 . 3 8 5 9 . 3 0 7 e 06
R e s i d u a l s 11 0 . 2 7 5 7 0 0 . 0 2 5 0 6

The Error(field) option tells R that it should

treat fields as sampled experimental units, i.e. subject to sampling
heterogeneity;
identify factors assigned to fields;

CHAPTER 7. NESTED DESIGNS

171

test those factors via comparisons of MSFactor to MSField.

7.1

Mixed-effects approach

What went wrong with the normal sampling model approach? What is wrong
with the following model?
yijk = + ai + bj + (ab)ij + ijk
where
i indexes sulfur level, i {1, 2} ;
j indexes type level, j {1, 2} ;
k indexes reps, k {1, . . . , 4}
ijk are i.i.d. normal
We checked the normality and constant variance assumptions for this model
previously, and they seemed ok. What about independence? Figure 5.4 plots
the residuals as a function of field. The figure indicates that residuals are
more alike within a field than across fields, and so observations within a field
are positively correlated. Statistical dependence of this sort is common
to split-plot and other nested designs.
dependence within whole-plots
affects the amount of information we have about factors applied at
the whole-plot level: within a given plot, we cant tell the difference
between plot effects and whole-plot factor effects.
This doesnt affect the amount of information we have about factors
applied at the sub-plot level: We can tell the difference between plot
effects and sub-plot factor effects.
If residuals within a whole-plot are positively correlated, the most intuitively
straightforward way to analyze such data (in my opinion) is with a hierarchical mixed-effects model:
yijkl = + ai + bj + (ab)ij + ik + ijkl
where things are as before except

172

l
h
hh
h

l
l
l

l
l

h
h
h
h

0.4

residual
0.0 0.2

0.4

CHAPTER 7. NESTED DESIGNS

1.0

1.5

2.0

2.5
field

3.0

l
3.5

4.0

Figure 7.4: Potato data

ik represents error variance at the whole plot level, i.e. variance in
whole plot experimental units (fields or blocks in the above example).
The index k represents whole-plot reps k = 1, . . . , r1 ,
{ik } normal(0, w2 )
ijkl represents error variance at the sub plot level, i.e. variance in
sub plot experimental units The index j represents sub-plot reps l =
1, . . . , r2 ,
{ijkl } normal(0, s2 )
Now every subplot within the same wholeplot has something in common, i.e.
ik . This models the positive correlation within whole plots:
Cov(yi,j1 ,k,l1 , yi,j2 ,k,l2 ) = E[(yi,j1 ,k,l1 E[yi,j1 ,k,l1 ]) (yi,j2 ,k,l2 E[yi,j2 ,k,l2 ])]
= E[(i,k + i,j1 ,k,l1 ) (i,k + i,j2 ,k,l2 )]
2
= E[i,k
+ i,k (i,j1 ,k,l1 + i,j2 ,k,l2 ) + i,j1 ,k,l1 i,j2 ,k,l2 ]
2
= E[i,k
] + 0 + 0 = w2

This and more complicated random-effects models can be fit using the lme
command in R. To use this command, you need the nlme package:

CHAPTER 7. NESTED DESIGNS

173

l i b r a r y ( nlme )
f i t . me<lme ( f i x e d=y type+s u l f u r , random =1| a s . f a c t o r ( f i e l d ) )
>summary ( f i t . me)
Fixed e f f e c t s : y type + s u l f u r
Value Std . E r r o r DF
tv a l u e pv a l u e
( I n t e r c e p t ) 4 . 8 8 5 0 0 . 2 5 9 9 6 5 0 11 1 8 . 7 9 0 9 9 1 0 . 0 0 0 0
typeB
0 . 6 1 0 0 0 . 0 7 9 1 5 7 5 11 7 . 7 0 6 1 5 3 0 . 0 0 0 0
sulfurlow
0.3675 0 . 3 6 3 3 6 0 2 2 1.011393 0 . 4 1 8 3
> anova ( f i t . me)
numDF denDF Fv a l u e pv a l u e
( Intercept )
1
11 7 5 9 . 2 9 4 6 <.0001
type
1
11 5 9 . 3 8 4 8 <.0001
sulfur
1
2
1.0229 0.4183

Notice that this gives the same results as

the randomization test (approximately);
the by-hand comparison of mean squares for factors to the appropriate
MSE;
the results from the aov command.
But the lme command allows for much more complex models to be fit.

CHAPTER 7. NESTED DESIGNS

7.2

174

Repeated measures analysis

Sitka spruce data: Longitudinal data on 79 spruce trees

54 grown in ozone-enriched chambers
25 grown in regular atmosphere (controls)

height
4
5

The size of each tree (roughly the volume) was measured at five time points:
152, 174, 201, 227 and 258 days after the beginning of experiment.

152

174

201
time

227

258

Figure 7.5: Sitka spruce data.

How should we analyze these data?
How should we evaluate evidence of a treatment (ozone) effect?
Naive approach I: A naive approach would be to ignore the study design
and the temporal nature of the data.
5 54 = 270 ozone observations
5 25 = 125 control observations

CHAPTER 7. NESTED DESIGNS

175

> f i t <lm ( S i t k a $ s i z e S i t k a $ t r e a t )
> anova ( f i t )
Df Sum Sq Mean Sq F v a l u e Pr(>F)
Sitka$treat
1
3.810
3.810 6.0561 0.01429
Residuals
393 2 4 7 . 2 2 2
0.629

Normal QQ Plot

ozone

1
0
fit$res

1
0
1
Theoretical Quantiles

fit$res
1
0

control

20

Frequency
40 60 80

Sample Quantiles
2
1
0
1

100

Histogram of fit$res

227

258

174

201
as.factor(Sitka$Time)

fit$res
1
0

152

9 11

14

17

20

23

26

29

32

35 38 41 44 47
as.factor(Sitka$tree)

50

53

56

59

62

65

68

71

74

77

CHAPTER 7. NESTED DESIGNS

176

Naive approach II: Clearly there is some effect of time. Lets now account for growth over time, using a simple ANCOVA:
yi,j,t = 0 + ai + b t + ci t + i,j,t
for ozone (i = 1) , E[y1,j,t ] = (0 + a1 ) + (b + c1 ) t
for control (i = 2) , E[y2,j,t ] = (0 + a2 ) + (b + c2 ) t
> f i t <lm ( s i z e Time+t r e a t+Time t r e a t , data=S i t k a )
> anova ( f i t )
Df Sum Sq Mean Sq F v a l u e
Pr(>F)
Time
1 8 9 . 5 6 4 8 9 . 5 6 4 2 2 2 . 9 0 2 0 < 2 . 2 e 16
treat
1
3.810
3.810
9 . 4 8 1 3 0 . 0 0 2 2 2 2
Time : t r e a t
1
0.551
0.551
1.3703 0.242480
R e s i d u a l s 391 1 5 7 . 1 0 7
0.402

height
4
5

160

180

200

220

240

260

time

Normal QQ Plot

0 20

Frequency
60
100

Sample Quantiles
1
0

140

Histogram of fit$res

0
fit$res

1
0
1
Theoretical Quantiles

Figure 7.6: ANCOVA fit and residuals

CHAPTER 7. NESTED DESIGNS

177

fit$res

11

15

19

23

27

31 35 39 43 47
as.factor(Sitka$tree)

51

55

59

63

67

71

75

Figure 7.7: Within-tree dependence

Conservative approach: Any of our standard inference tools are suspect
because we clearly do not have independent observations. The within-tree
dependence indicates that we have
less information than in 270 independent ozone and 125 independent
control observations,
but more information than in 54 independent ozone and 24 independent
control observations.
So our worst-case scenario is analogous to having n1 = 54 and n2 = 24.
One approach to analysis in such situations is to reduce the information
to one number per unit, then compare numbers across treatments. For
example, we could compute an average and fit a regression line for each tree
j, giving
avg
yi,j
, the average of the 5 observations for tree i, j
int
yi,j
, the estimated intercept of the regr. line for tree i, j
int
yi,j
, the estimated slope of the regr. line for tree i, j

We can then compare averages, intercepts and slopes across the two treatment groups. Note that, for each type of y, there is only one observation for
each tree: we have eliminated the problem of dependent measurements.

79

178

height

CHAPTER 7. NESTED DESIGNS

160

180

200

220

240

260

0.020

0.015

intercept
2

slope
0.010

0.005

3.5

4.0

average
4.5 5.0

5.5

6.0

time

3.0

control

ozone
treatment

control

ozone

control

treatment

Figure 7.8: Reduction to tree-specific summary statistics

> anova ( lm ( y . avg t r e a t ) )
Df Sum Sq Mean Sq F v a l u e Pr(>F)
treat
1 0.7619 0.7619 2.0188 0.1594
R e s i d u a l s 77 2 9 . 0 6 1 7 0 . 3 7 7 4
> anova ( lm ( y . i n t t r e a t ) )
Df Sum Sq Mean Sq F v a l u e Pr(>F)
treat
1 0.840
0.840 1.0431 0.3103
R e s i d u a l s 77 6 1 . 9 8 9
0.805
> anova ( lm ( y . s l o p e t r e a t ) )
Df
Sum Sq
Mean Sq F v a l u e
Pr(>F)
treat
1 0 . 0 0 0 0 7 8 1 5 0 . 0 0 0 0 7 8 1 5 7 . 6 6 2 8 0 . 0 0 7 0 5 8
R e s i d u a l s 77 0 . 0 0 0 7 8 5 2 9 0 . 0 0 0 0 1 0 2 0

ozone
treatment

CHAPTER 7. NESTED DESIGNS

179

Linear random effects models: The last approach is extremely conservative, as it basically reduces all the information we have from a tree to one
number. Of course, the observations from a single tree are not completely
dependent, and so compressing the data in this way throws away potentially
valuable information. To make use of all the information from a tree, we can
use a random effects model which accounts for correlation of observations
common to a given tree.
yi,j,t = (a1 + b1,j + c1,i ) + (a2 + b2,j + c2,j ) t + i,j,t
where
(b1,j , b2,j ), j = 1, 2 are fixed-effects, measuring the heterogeneity of
the average slope and intercept across the two levels of treatment;
(c1,1 , c2,1 ), . . . , (c1,n , c2,n ) i.i.d. multivariate normal(0, ) are random
effects , inducing a within-tree correlation of observations.

> f i t <lme ( f i x e d=s i z e t r e a t+Time+t r e a t Time ,

random=Time | t r e e , data=S i t k a )
> anova ( f i t )
numDF denDF Fv a l u e pv a l u e
( Intercept )
1
314 4 9 6 7 . 0 5 6 <.0001
treat
1
77
2.118 0.1497
Time
1
314 1 2 4 6 . 5 1 8 <.0001
t r e a t : Time
1
314
7.663 0.0060