I NTRODUCTION TO S TATISTICAL M ODELLING T RINITY 2002

The Analysis of Variance

Introduction
Researchers often perform experiments to compare two treatments, for example, two different
fertilizers, machines, methods or materials. The objectives are to determine whether there is
any real difference between these treatments, to estimate the difference, and to measure the
precision of the estimate. In the Descriptive Statistics course we have discussed comparisons of
two means. It is often important to compare more than two means. For example, we may be
interested in determining if there is any evidence for real differences among the mean values
associated with various different treatments that have been randomly allocated to the
experimental units. This corresponds to a hypothesis of the form

H 0 : µ1 = µ 2 = K = µ I
H A : at least two of the µ i ' s are different

where µ i is the mean of the ith treatment or population. Analysis of variance (ANOVA)
provides the framework to test hypotheses like the one above, on the supposition that the data
can be treated as random samples from I normal populations having the same variance σ 2 and
possibly only differing in their means. The sample sizes for the treatment groups are possibly
different, say J i . The analysis resulting from these assumptions may be approximately
justified by randomisation, to guarantee inferential validity. It is normally the case in
performing an ANOVA that the data come from an experiment rather than an observational
study, since the experimental conditions imply that balance has been achieved through
randomisation. The calculations to explore these hypotheses are set out in an analysis of
variance table. Essentially, this calculation determines whether the discrepancies between the
treatment averages are greater than could reasonably be expected from the variation that
occurs within the treatment classifications.

One-way analysis of variance

Balanced versus unbalanced layouts

Before we proceed to define the one-way layout we just need to distinguish between balanced
and unbalanced layouts. A balanced one-way ANOVA refer to the special case of one-way
ANOVA in which there are equal numbers of observations in each group, say
J 1 = J 2 = K = J I . An experimental layout involving different numbers of observations in
each group is referred to as unbalanced. Below we will specify the one-way layout in its most
general form, allowing for different numbers of observations in each group.

IAUL & D EPARTMENT OF S TATISTICS P AGE

1
I NTRODUCTION TO S TATISTICAL M ODELLING T RINITY 2002

The one-way ANOVA model and assumptions

The one-way layout refers to the simplest case in which analysis of variance is applied and
involves the comparison of the means of several (univariate) populations. One-way analysis of
variance gets its name from the fact that the data are classified in only one way, namely, by
treatment. We shall assume that the I populations are normal with equal variance σ 2 , and that
we have independent random samples of sizes J 1 , J 2 , K , J I from the respective populations
or treatment groups with n = ∑iI=1 J i . Furthermore, let µ i denote the mean of the ith
population. If Yij is the random variable denoting the jth measurement from the ith
population, we can specify the one-way analysis of variance model as

Yij = µ i + ε ij ( )
i = 1, K, I ; j = 1, K , J i , ε ij i. i. d. N 0, σ 2 , (1)

Note that this model has the same main assumption as the standard linear model in that the
unobservable random errors ε ij are independent and follow a normal distribution with mean
zero and unknown constant variance. If this assumption is not satisfied, then the validity of
the results of an ANOVA is in question.

The hypothesis that is associated with this model is

H 0 : µ1 = µ 2 = K = µ I
H A : at least two of the µ i ' s are different

Alternatively, model (1) is often written as

Yij = µ + α i + ε ij ( )
i = 1,K , I ; j = 1, K , J i , ε ij i. i. d. N 0, σ 2 , (2)

where µ i = µ + α i . The parameter µ is viewed as a grand mean, while α i is an effect for the
ith treatment group. The hypothesis associated with this model is then specified as

H 0 : α1 = α 2 = K = α I = 0
(3)
H A : at least one α i ≠ 0

It is important to note that the parameters µ and {α i } are not uniquely defined in model (2).
We say that the parameters µ and {α i } are not completely specified by the model. However,
I
we can assume without loss of generality that α . = ∑ α i = 0 , since we can write
i =1

η ij = E ( yij ) = µ + α i = (µ + α . ) + (α i − α . ) ,

and take as new µ and {α i } the quantities µ

~ = µ + α and α~ = α − α , then ∑ α~ = 0 .
. i i . i i

IAUL & D EPARTMENT OF S TATISTICS P AGE

2
I NTRODUCTION TO S TATISTICAL M ODELLING T RINITY 2002

It follows from the general theory that there is a unique solution satisfying ∑i αˆ i = 0 , and that
~ and {α~ } is estimable.
every parametric function of the new parameters µ i

One-way ANOVA table

The results from fitting model (2) are typically summarised in an ANOVA table, which can be
obtained from most statistical software packages such as SPSS. Below we give a typical
template of an ANOVA table for the one-way ANOVA classification.

Source of Variation Df Sum of Squares Mean Square F

Intercept 1 SSA MSA

Treatments I-1 SST MST MST/MSE

Error N-I SSE MSE

Total N TSS

In the table above SSA, SST, SSE, TSS, MST and MSE are calculated from the observed
data. Furthermore, I is the number treatments, and N is the total number of
observations. The last column contains the value of the test statistic for the hypothesis
in (3). We will discuss these quantities below without going into too much
mathematical detail.

In order to give a short explanation of the above ANOVA table, we start off by noting
that a measure of the overall variation could have been obtained by ignoring the
separation into treatments and calculating the sample variance for the aggregate of N
observations. This would be done by calculating the total sum of squares of deviations
about the overall mean y

2
( )
I Ji
SSD = ∑ ∑ y ij − y
i =1 j =1

and by dividing by the appropriate degrees of freedom ν D = N − 1 . Based on the

algebraic identity

( )2 = ∑ J i ( yi − y )2 + ∑ ∑ (yij − y j )2
I Ji I I Ji
SSD = ∑ ∑ y ij − y
i =1 j =1 i =1 i =1 j =1

IAUL & D EPARTMENT OF S TATISTICS P AGE

3
I NTRODUCTION TO S TATISTICAL M ODELLING T RINITY 2002

or

SSD = SST + SSE

the total sum of squares of deviations from the overall mean can be divided into sum of
the between treatment sum of squares (SST) and the within treatment or residual sum
of squares (SSE). The SSD can also be written as the total sum of squares (TSS) minus
the sum of squares due to the average or correction factor (SSA)

I Ji
SSD = ∑ ∑ yij2 − Ny 2 = TSS − SSA
i =1 j =1

Therefore

TSS = SSA + SSD

so that, we can split up the sum of squares of the original N observations into three
additive parts:

TSS = SSA + SST + SSE .

In words this means that the total sum of squares (TSS) can be written as the sum of
squares due to the average, the between treatment sum of squares and the residual
sum of squares.

The associated degrees of freedom are

N = 1 + k −1 + N −I.

Inference and interpretation

Once we have checked the assumptions underlying the model, as we will discuss in the
next subsection, we can use the analysis of variance table to draw inference about the
treatment means. The table provides the basis for a formal test of the hypothesis that
the treatment effects are all equal to zero, that is, α 1 = α 2 = K = α I = 0 . If the null
hypothesis of zero treatment effects is true and the errors are identically independently
( )
distributed N 0, σ 2 , then the test statistic is given by the ratio

SST SSE MST

FC = = .
I − 1 N − 1 MSE

and follows an F-distribution on I-1 and N-I degrees of freedom. Intuitively, we would
expect the test statistic, FC , to be approximately 1 if there is no difference between the
treatments, and considerably greater than 1 if there is a difference. If we wish to test
H 0 at the 100(1 − α )% level, then the criterion for the test is to reject H 0 if

IAUL & D EPARTMENT OF S TATISTICS P AGE

4
I NTRODUCTION TO S TATISTICAL M ODELLING T RINITY 2002

FC > F1−α , I −1, N − I ,

where Fα , I −1, N − I is the100(1 − α ) percentile of a F-distribution with I-1 and N-I degrees
of freedom. Its value can be obtained from standard tables or from a software package.
Software packages usually provide an exact p-value for the test.

Diagnostic checking of the model

It is important that we check the assumptions underlying our model, namely, errors
that are independent and identically normally distributed with constant variance. In
order to investigate the validity of these assumptions there are a few standard plots of
the residuals that can be used to evaluate their distribution and to check for systematic
patterns in the residuals. When the assumptions concerning the adequacy of the model
are true, we expect the residuals to vary randomly. Recall from Lecture 2 that the
residuals are simply the differences between the observed and fitted values, y ij − yˆ ij .
Below we consider a few useful plots to investigate the above properties of the
residuals. These particular discrepancies should be looked for as a matter of routine,
but the experimenter should also be on the alert for other abnormalities. Refer to
Lecture Notes 3 for a more detailed discussion.

Normal probability plot of the residuals

The first plot that we consider for use in checking whether we have satisfied the
assumptions is something called a normal probability plot or a quantile-normal plot of
the residuals. This plot was introduced in Lecture 3. This plot is used to check the
normality assumption since it allows us to investigate whether the residuals are
normally distributed: since the residuals can be thought of as estimates of the errors,
evidence that the residuals are non-normal might lead us to suspect that the errors are
not normal. The interpretation of this plot was discussed in Lecture 3.

Scatterplot of residuals versus predicted values

The variability of the residuals should be unrelated to the levels of the response, as we
assumed constant variance for the residuals. This can be investigated by plotting the
residuals, y ij − yˆ ij , against the fitted values, ŷ ij . Sometimes the variance increases as
the value of the response increases, which is indicative of non-constant variance.

Plot of the residuals in time sequence

Sometimes an experimental factor may drift or the skill of the experimenter may
improve as the experiment proceeds. Tendencies of this kind may be uncovered by
plotting the residuals against time order where it is appropriate.

Example

IAUL & D EPARTMENT OF S TATISTICS P AGE

5
I NTRODUCTION TO S TATISTICAL M ODELLING T RINITY 2002

Consider an experiment that is set-up to compare melon varieties. Each variety is

grown 6 times under identical conditions with identical feeding patterns1. The yields
in kg are displayed in Table 1.

Table1: Yields of four different varieties of melons.

Variety A Variety B Variety C Variety D

25.1 40.2 18.3 28.0

17.2 35.3 22.6 28.6
26.4 32.0 25.9 33.2
16.1 36.5 15.1 31.7
22.2 43.3 11.4 30.3
15.9 37.1 23.7 27.6

We analysed this data set and obtained the following ANOVA table as output:

ANOVA

YIELD
Sum of
Squares df Mean Square F Sig.
Between Groups 1290.951 3 430.317 23.458 .000
Within Groups 366.888 20 18.344
Total 1657.840 23

Test of Homogeneity of Variances

YIELD
Levene
Statistic df1 df2 Sig.
2.544 3 20 .085

We now verify the assumptions underlying the above model by considering residual
plots discussed above and a homogeneity of variance test. From the quantile-quantile
plot of the standardised residuals in figure 1 it seems as though the assumption of
normality for the residuals is a reasonable one. Figure 2 shows that variability of the
residuals may be a little larger for the first variety or perhaps a little smaller for the
third variety, but in general the variability of the residuals seems to be reasonably
similar across the different varieties. For a more formal analysis we can consider the
Levine test for equal variances given above. From the above SPSS output we see that
the test does not reject the null hypothesis of equal variances at a 5% significance level,

1
Data taken from Borja, M. C. Introduction to Statistical Modelling for Research: Course notes at the Department of Statistics
at Oxford University.

IAUL & D EPARTMENT OF S TATISTICS P AGE

6
I NTRODUCTION TO S TATISTICAL M ODELLING T RINITY 2002

but it does reject it at a 10% level. Thus, it seems that the variability of the residuals is
somewhat different for the different varieties, but that the difference is not significant
at a 5% level.

In this example, the result of the hypothesis test is highly significant, as indicated by
the very small p-value. Thus, the null hypothesis that the yields of all varieties are zero
will be rejected in favour of the alternative that at least one variety has a non-zero
effect.
6
4
2
Standardised residuals

0
-2
-4
-6
-8

-2 -1 0 1 2
Quantiles of Standard Normal

Figure 1: QQ-plot for the standardised residuals of the melon yield example.

IAUL & D EPARTMENT OF S TATISTICS P AGE

7
I NTRODUCTION TO S TATISTICAL M ODELLING T RINITY 2002

15

6
3

4
2
Residuals
0
-2
-4
-6
-8

17

20 25 30 35

Fitted : Variety

Figure 2: Residuals plotted against the fitted values of the expected response.

Treatment contrasts and multiple comparisons

The analysis of variance test above involves only one hypothesis, namely, that of equal
treatment means µ i (or treatment effects α i ). If the hypothesis is rejected in an actual
application of the F-test for the equality of means in the one-way layout, the resulting
conclusion that the means µ1 , µ 2 , K , µ I are not all equal would by itself usually be
insufficient to satisfy the experimenter. Methods of making further inferences about the
mean are then desirable. In more complicated situations than the one-way ANOVA,
the analysis of variance table becomes a very useful tool for identifying aspects of a
complicated problem that deserve more attention. It also introduces the SST as a
measure of treatment differences. The SST can be broken into components
corresponding to the sums of squares for individual orthogonal contrasts. These
components of the SST can then be used to explain the differences in the means.

A contrast among the parameters µ1 , µ 2 , K , µ I is a linear function of the µ i , ∑iI=1 λi µ i ,

with known contrast coefficients, λi , subject to the condition ∑iI=1 λi = 0 . Furthermore,
two contrasts are defined as orthogonal if

I λi1λi 2
∑ = 0.
i =1 N i

Contrasts are only of interest when they define interesting functions of the µ i s. There
are many ways to choose contrasts and these depend on the question the researcher is
interested in answering. Treatment contrasts are commonly used contrast to compare
all of the treatments to a control. Other commonly used contrasts are polynomial

IAUL & D EPARTMENT OF S TATISTICS P AGE

8
I NTRODUCTION TO S TATISTICAL M ODELLING T RINITY 2002

contrasts and Helmert contrasts. The choice of contrasts can be a rather technical
subject and so we will not go into too much detail.

As an example, consider a one-way ANOVA set-up where there are five different diets
A, B, C, D and E. Suppose that A is an existing standard diet that serves as the control
and that B, C, D and E are new diets. An example of an interesting contrast may be to
compare the control diet, A, with the four new diets. This means that we are interested
in comparing the control to the average of the other four diets. This contrast would
then be

µ B − µC − µ D − µ E
µA −
4

and by multiplying the contrast by four gives the equivalent contrast

4µ A − µ B − µ C − µ D − µ E .

Software packages such as SPSS accommodate commonly used contrasts such as

polynomial contrasts and also allow the user to specify any a priori contrast that may be
of interest. Ultimately the choice of which comparison to make is up to the researcher.

One very important final remark we will make involves the issue of multiple
comparisons. The specification several different contrasts leads to multiple hypothesis
tests that are performed using the same data set. This means that we have to remember
to adjust the significance level of any multiple hypothesis tests that we conduct to
ensure that the overall level of significance for carrying out all of the tests is equal to
the desired level of significance. Recall that the reason for this is that the probability of
making at least one type I error is greater than the original significance level when we
conduct multiple tests using the same data. There are a number of different
significance correction methods described in the literature. Five of the most common
ones include adjustments suggested by Bonferroni, Scheffe, Dunnett, Tukey and Sidak.
The advantages and disadvantages of using different methods are quite complex and it
is common practice to use all of the available methods and then report the most
conservative one. Bonferroni’s method is perhaps the simplest and most widely used
method. We have discussed this method in the Descriptive Statistics Course.

Non-parametric one-way ANOVA

As was the case in the discussion for linear models, transformations of the response
variable can be useful tools for dealing with the problems of heteroscedasticity and
non-normality. However, sometimes a transformation cannot solve the problem, in
cases like these we can proceed by using a non-parametric test. In the case of two
samples, the Mann-Whitney U statistic is the non-parametric equivalent of the two-
sample t-test, and in the case of more than two samples, the Kruskal-Wallis test is the
non-parametric equivalent of the one-way ANOVA. This approach can be applied
when the assumption of normality or equal variance does not hold.

IAUL & D EPARTMENT OF S TATISTICS P AGE

9
I NTRODUCTION TO S TATISTICAL M ODELLING T RINITY 2002

The intuitive idea for the Kruskal-Wallis test is that if you rank all of the data and sum
the ranks in each group, then if the groups have no real differences, the sum of the
ranks should be the same in each group. If there were differences between the groups,
then you would expect the sum of the ranks within each group to be different.

As an example, consider the following experiment investigating LDL cholesterol in

quails2. Thirty-nine quails were randomly assigned to four diets, each diet containing a
different drug compound, which would hopefully reduce LDL cholesterol. The drug
compounds are labelled I, II, III and IV. At the end of the experimental time the LDL
cholesterol of each quail was measured. Thus, there are two values relating to each
quail, one recording the diet and one indicating the LDL cholesterol level. The data are
displayed in Table 2.

Table 2: LDL cholesterol levels in quails exposed to four different diets.

Drug I Drug II Drug III Drug IV

52 36 52 62
67 34 55 71
54 47 66 41
69 125 50 118
116 30 58 48
79 31 176 82
68 30 91 65
47 59 66 72
120 33 61 49
73 98 63

The main interest is to see if whether or not diet has any effect on the mean value of
LDL cholesterol level. If we consider the boxplots in figure 3 of the LDL level for each
diet, there appear to be significant differences between the diets. We start of by fitting
the usual two-way ANOVA model to the data. It is clear from the normal Q-Q plot of
the standardised residuals in figure 4 that the residuals do not satisfy the assumption of
normality. Any inferences drawn from the results of this ANOVA model will not be
valid. Consequently, we perform the Kruskal-Wallis test. The results from this
procedure are also presented below.

2Data taken from Hettmansperger, T. P. and McKean, J. W. (1998). Robust Nonparametric Statistical Methods: Kendall’s
Library of Statistics 5. London: Arnold.

IAUL & D EPARTMENT OF S TATISTICS P AGE

10
I NTRODUCTION TO S TATISTICAL M ODELLING T RINITY 2002

180
160
140
120
100
80
60
40

Diet I Diet II Diet III Diet IV

Figure 3: Boxplots for the quail data.

100
80
Standardised residuals

60
40
20
0
-20

-2 -1 0 1 2
Quantiles of Standard Normal

Figure 4: QQ-plot of the standardised residuals of the one-way ANOVA.

Test Statistics a,b

LDL
Chi-Square 7.188
df 3
Asymp. Sig. .066
a. Kruskal Wallis Test
b. Grouping Variable: DIET

IAUL & D EPARTMENT OF S TATISTICS P AGE

11
I NTRODUCTION TO S TATISTICAL M ODELLING T RINITY 2002

The p-value of 0.066 for Kruskal-Wallis test, although not significant at a 5% level,
indicates some difference between the diets as suggested by the diets. A standard one-
way ANOVA gives a p-value of 0.35 for the F-test of equal treatment effects. This does
not agree with the boxplots in figure 3. The reason for this is that the long right tail of
the errors shown in figure 4 adversely affects the test statistic. Unfortunately, there is
no easy extension to the problem of multiple comparisons that we will introduce in
more details in a later section.

Complete randomised blocks

Introduction
In this section we extend the ideas of the previous section by comparing more than two
treatments, using randomised designs with larger block sizes. In blocked designs there
are two kinds of effects of interest. The first is the treatment effects, which are of
primary interest to the experimenter. The second is the blocks, whose contribution
must be accounted for. In practice, blocks might be, for example, different litters of
animals, blends of chemical material, strips of land, or contiguous periods of time. In
the next section we will consider a replicated factorial design in which the main effects
of two factors and the interaction are all of equal interest.

Randomised complete block designs: (1) identify blocks of homogeneous experimental

material (units) and (2) randomly assign each treatment to an experimental unit within
each block. The blocks are complete in the sense that each block contains all of the
treatments. Random assignment of treatments to experimental units allows us to infer
causation from a designed experiment. If treatments are randomly assigned to
experimental units, then the only systematic differences between the units are the
treatments. In observational studies where the treatments are not randomly assigned it is
much more difficult to infer causation. So the advantage of this procedure is that
treatment comparisons are subject only to the variability within blocks. Block to block
variation is eliminated in the analysis. In a completely randomised design applied to
the same experimental material, the treatment comparisons would be subject to both
within block and between block variability.

Let us consider an example. We consider a randomised block experiment in which a

process for the manufacture of penicillin was being investigated, and yield was the
response of primary interest3. There were 4 treatments to be studied, denoted by A, B,
C and D. It was known that an important raw material, corn steep liquor, was quite
variable. Fortunately sufficient blends could be made to allow researchers to run all 4
treatments within each of 5 blocks (blends of corn steep liquor). Furthermore, the
experiment was protected against extraneous unknown sources of bias by running the

3Example taken from Box, Hunter and Hunter. 1978. Statistics for Experimenters: An Introduction to Design, Data
Analysis and Model Building. John Wiley, New York.

IAUL & D EPARTMENT OF S TATISTICS P AGE

12
I NTRODUCTION TO S TATISTICAL M ODELLING T RINITY 2002

treatments in random order within each block. The randomised block design is given
in Table 3.

Table 3: Randomised block design on penicillin manufacture4

Treatment
Block
A B C D

Blend 1 89(1) 88(3) 97(2) 94(4)

Blend 2 84(4) 77(2) 92(3) 79(1)

Blend 3 81(2) 87(1) 87(4) 85(3)

Blend 4 87(1) 92(3) 89(2) 84(4)

Blend 5 79(3) 81(4) 80(1) 88(2)

The model and assumptions

The model for a randomised complete block design is given by

( )
y ij = µ + α i + β j + ε ij , i = 1, K, I ; j = 1, K , J , ε ij i.i.d. N 0, σ 2 .

There are J blocks with I treatments observed within each block. As before the
parameter µ is viewed as a grand mean, α i is an unknown fixed effect for the ith
treatment, and β j is an unknown fixed effect for the jth block. The theoretical basis for
the analysis of this model is precisely as in the balanced one-way ANOVA. As before
the computations can be summarised in an ANOVA table, as we will show in the
following section.

ANOVA table
As was the case for the one-way layout, the results of fitting model the model for a complete
randomised block design are typically represented in an ANOVA table, which is a summary of
the modelling procedure and can be calculated using most statistical software packages. As we
are interested in the interpretation rather than theory, we only consider an example of what an
ANOVA table looks like for a complete block design, thereby avoiding unnecessary
mathematics.

Table 4: ANOVA table for a complete randomised block design

4 The superscripts in parentheses associated with the observations indicate the random order in which the
experiments were run within each blend.

IAUL & D EPARTMENT OF S TATISTICS P AGE

13
I NTRODUCTION TO S TATISTICAL M ODELLING T RINITY 2002

Source of Variation Df Sum of Squares Mean Square F

Intercept 1 SSA MSA

Treatments I-1 SST MST MST/MSE

Blocks J-1 SSB MSB MSB/MSE

Residuals (I-1)(J-1) SSE MSE

Total IJ TSS

In the table above SSA, SST, SSE, TSS, MST and MSE are simply summaries calculated
from the observed data, similar to what we saw for the one-way ANOVA table.
Furthermore, I is the number of treatments and J is the number of blocks. The two test
statistics are produced in a similar way to the one-way case. We will consider the two
hypotheses involved in a little more detail in the following section.

Inference and interpretation

Table 4 provides the basis for a formal test of the hypothesis that the treatment effects
are all equal to zero, as well as a formal test for the hypothesis that the block effects are
all equal to zero. The F-statistic, FT = MST MSE , is used to test whether there are
significant treatment effects, i.e., it is used to test

H 0 : α1 = α 2 = K = α I = 0
.
H A : at least one α i ≠ 0

The statistic follows an F-distribution on I-1 and (I-1)(J-1) degrees of freedom. If we

wish to test H 0 at the 100(1 − α )% level, then the criterion for the test is to reject H 0 if

FT > F1−α , I −1,( I −1)( J −1) .

Fα , I −1,( I −1)( J −1) is the 100(1 − α ) percentile of a F-distribution with I-1 and (I-1)(J-1) degrees
of freedom and its value can be obtained from standard tables or from a software
package.

Similarly, the F-statistic, FB = MSB MSE , is used to test whether there are significant
block effects, i.e., it is used to test

IAUL & D EPARTMENT OF S TATISTICS P AGE

14
I NTRODUCTION TO S TATISTICAL M ODELLING T RINITY 2002

H 0 : β1 = β 2 = K = β J = 0
.
H A : at least one β j ≠ 0

The F-statistic provides a test of whether we can isolate comparative differences in the
block effects. Thus, a significant test indicates that blocking was a worthwhile exercise.
If we wish to test H 0 at the 100(1 − α )% level, then the criterion for the test is to reject
H 0 if

FT > F1−α , J −1,( I −1)( J −1) .

Fα , J −1,( I −1)( J −1) is the100(1 − α ) percentile of a F-distribution with J-1 and (I-1)(J-1) degrees
of freedom.

Diagnostic checking of the model

The assumptions underlying the randomised complete block design model are similar
to those of the one-way ANOVA model. These assumptions need to be verified in
order for the model to be valid. We can use the same methods as for the one-way
ANOVA model to achieve this.

Example
Let us return to the penicillin example that we introduced earlier in this section. We
will fit the randomised block design model on this data set and validate the underlying
assumptions. Below we include the resulting ANOVA table and some diagnostic plots.

Tests of Between-Subjects Effects

Dependent Variable: YIELD

Sum of
Source Squares df Mean Square F Sig.
Intercept 147920.000 1 147920.000 7854.159 .000
BLOCKS 264.000 4 66.000 3.504 .041
TREATMEN 70.000 3 23.333 1.239 .339
Error 226.000 12 18.833
Total 148480.000 20

The residual plots for the penicillin example in figures 4 and 5 do not reveal anything
of special interest. The assumptions of normality and constant variance for the
residuals seem to be reasonable. Sometimes the plot of the residuals versus the
predicted values shows a curvilinear relationship. For example the residuals may tend
to be positive for low values of ŷ , become negative for intermediate values, and be
positive again for high values. This often suggests nonadditivity between the block and
treatment effects and might be eliminated by a suitable transformation of the response.
However, it is not the case in this example.

IAUL & D EPARTMENT OF S TATISTICS P AGE

15
I NTRODUCTION TO S TATISTICAL M ODELLING T RINITY 2002

20 12

6
4
2
Residuals
0
-2
-4

15

80 85 90 95

Fitted : 1 + TREATMEN + BLOCKS

Figure 5: Residuals versus fitted values for penicillin example.

6
4
2
0
-2
-4

-2 -1 0 1 2

Quantiles of Standard Normal

Figure 6: Q-Q plot of the standardised residuals in the penicillin example.

The p-value of 0.339 for the hypothesis of zero treatment effects, suggests that the four
different treatments have not resulted in different yields. The variability among the
treatment averages can be reasonably attributed to experimental error. We reject the
null hypothesis of no blend-to-blend variation, as suggested by the small p-value
(0.041), thus there are significant block effects.

IAUL & D EPARTMENT OF S TATISTICS P AGE

16
I NTRODUCTION TO S TATISTICAL M ODELLING T RINITY 2002

The two-way factorial design

Introduction
In this section we are interested in applying two different treatments (each
treatment having a number of levels). We are trying to discover if there are any
differences within each treatment and whether the treatments interact. The main
effects of the two treatments and their interaction are all of equal interest. An easy way
to understand this topic is by means of an example. Consider an agricultural
experiment where the investigator is interested in the corn yield when three different
fertilisers are available, and corn is planted in four different soil types. The researcher
would like to establish if the fertiliser has an effect on crop yield, if the soil type has an
effect on crop yield and whether the two treatments interact. The presence of an
interaction in this example means that there may be no difference between fertiliser 1
and fertiliser 2 in soil type 1, but that fertiliser 1 may produce a greater crop yield than
fertiliser 2 in soil type 2. We will only consider balanced designs here, although the
theory extends to non-balanced designs.

The model and assumptions

Assume that we have two experimental factors, named A (with a levels) and B (with b
levels). The model for the balanced two-way factorial design with interaction is given
by

y ijk = µ + α i + β j + (αβ )ij + ε ijk ,

(
i = 1,2, K , a; j = 1,2, K, b; k = 1,2, K , n, ε i.i.d. N 0, σ 2 )
In the above model y ijk is the kth response value subject to the ith level of factor A, and
the jth level of factor B. It is assumed that these y ijk are independent and
( )
y ijk ~ N µ ij , σ 2 . It represents a balanced design because we have the same number, n ,
observations for each treatment combination. The global mean or intercept term is
again represented by µ . In the above model α i is the treatment effect of the ith level of
factor A, while β j is the treatment effect of the jth level of factor B. The α i are often
referred to as the row effects and the β j as the column effects. This stems from the fact
that the ith row of the data table often represents the observations made for the ith level
of factor A, and the jth column of the data table often represents the observations made
for the jth level of factor B. They are called the treatment main effects. The interaction
effects are represented by the (αβ )ij , where (αβ )ij represents the interaction effect of the
ith level of factor A and the jth level of factor B. We will discuss interaction in some
more detail in a later subsection.

IAUL & D EPARTMENT OF S TATISTICS P AGE

17
I NTRODUCTION TO S TATISTICAL M ODELLING T RINITY 2002

ANOVA table
Again we consider the ANOVA table for the above model by avoiding unnecessary
mathematical detail. The general form of the ANOVA table is presented in table 5.

Table 5: ANOVA table for a balanced two-way factorial design with interaction

Source of Variation Df Sum of Squares Mean Square F

Intercept 1 SSInt MSInt

Factor A a-1 SSA MSA MSA/MSE

Factor B b-1 SSB MSB MSB/MSE

Interaction (a-1)(b-1) SS(AB) MS(AB) MS(AB)/MSE

Error ab(n-1) SSE MSE

Total TSS

As before the quantities in the table above are simply summaries calculated from the
observed data, similar to what we saw for the one-way ANOVA table and the block
design. The three test statistics are produced in a similar way than before, and is based
on the same intuitive approach. That is, we are considering estimating how much of
the overall variation each factor and the interaction explain, compared to the residual
(error) variation. The next subsection will look at the relevant hypotheses in some
more detail.

Inference and interpretation

Again the ANOVA table provides the basis for formal tests of all the relevant
hypotheses that we may be interested in. There are three main hypothesis tests of
interest here, namely, the test for significant treatment effects for factor A, the test for
significant treatment effects for factor B and the test for significant interaction effects.
As with the previous sections, to test each of the hypotheses, we compare the test
statistic given in the last column with the F-distribution with appropriate degrees of
freedom. The appropriate degrees of freedom are the degrees of freedom associated
with the source of variation and the degrees of freedom associated with the error.

IAUL & D EPARTMENT OF S TATISTICS P AGE

18
I NTRODUCTION TO S TATISTICAL M ODELLING T RINITY 2002

The F-statistic, FA = MSA MSE , is used to test whether there are significant treatment
effects for factor A, i.e.

H 0 : α1 = α 2 = K = α a = 0
.
H A : at least one α i ≠ 0

Similar to the one-way ANOVA the statistic follows an F-distribution on a-1 and (a-
1)(b-1) degrees of freedom. If we wish to test H 0 at the 100(1 − α )% level, then the
criterion for the test is to reject H 0 if

FA > F1−α ,a −1,(a −1)(b−1)

where F1−α ,a −1,(a −1)(b −1) is the 100(1 − α ) percentile of a F-distribution with a-1 and (a-1)(b-
1) degrees of freedom.

Similarly, the F-statistic, FB = MSB MSE , is used to test whether there are significant
treatment effects for factor B, i.e.

H 0 : β1 = β 2 = K = β b = 0
.
H A : at least one β j ≠ 0

If we wish to test H 0 at the 100(1 − α )% level, then the criterion for the test is to reject
H 0 if

FB > F1−α ,b−1,(a −1)(b−1)

where F1−α ,b −1,(a −1)(b −1) is the 100(1 − α ) percentile of a F-distribution with b-1 and (a-1)(b-
1) degrees of freedom.

The third hypothesis that we want to test is that of no significant interaction effects.
The F-statistic, FAB = MS ( AB ) MSE , provides us with the basis of doing that. The
corresponding hypothesis can be formulated as

H 0 : αβ ij = 0, i = 1,2,K , a; j = 1,2, K, b,
H A : at least one αβ ij ≠ 0

When we test the null hypothesis of no interaction at the 100(1 − α )% level, the criterion
for the test is to reject H 0 if

FAB > F1−α ,ab (n −1),(a −1)(b −1)

IAUL & D EPARTMENT OF S TATISTICS P AGE

19
I NTRODUCTION TO S TATISTICAL M ODELLING T RINITY 2002

where F1−α ,ab (n −1),(a −1)(b −1) is the 100(1 − α ) percentile of a F-distribution with ab(n-1) and
(a-1)(b-1) degrees of freedom. Normally all these values are provided by the software
package and thus there is no need to calculate them yourself.

Example
Our example involves an experiment in which haemoglobin levels in the blood of
brown trout were measured after treatment with four rates of sulfamerazine. Two
methods of administering the sulfamerazine were used. Ten fish were measured for
each rate and each method. The data are given in Table 65.

Table 6: Data for the haemoglobin example.

Rate
Method
1 2 3 4

6.7 9.9 10.4 9.3

7.8 8.4 8.1 9.3
5.5 10.4 10.6 7.2
8.4 9.3 8.7 7.8
7.0 10.7 10.7 9.3
A
7.8 11.9 9.1 10.2
8.6 7.1 8.8 8.7
7.4 6.4 8.1 8.6
5.8 8.6 7.8 9.3
7.0 10.6 8.0 7.2

7.0 9.9 9.9 11.0

7.8 9.6 9.6 9.3
6.8 10.2 10.2 11.0
7.0 10.4 10.4 9.0
7.5 11.3 11.3 8.4
B
6.5 9.1 10.9 8.4
5.8 9.0 8.0 6.8
7.1 10.6 10.2 7.2
6.5 11.7 6.1 8.1
5.5 9.6 10.7 11.0

If we now fit a two-way ANOVA to the data we obtain the following results:

5
This data is taken from Rencher, A. C. (2000). Linear Models in Statistics. Wiley: New York.

IAUL & D EPARTMENT OF S TATISTICS P AGE

20
I NTRODUCTION TO S TATISTICAL M ODELLING T RINITY 2002

ANOVA output

Dependent Variable: HEMOGLOB

Sum of
Source Squares df Mean Square F Sig.
Intercept 6102.271 1 6102.271 3956.626 .000
RATE 90.304 3 30.101 19.517 .000
METHOD 2.346 1 2.346 1.521 .221
RATE * METHOD 4.803 3 1.601 1.038 .381
Error 111.045 72 1.542
Total 6310.770 80

The approach is generally to first test the hypothesis that there is an interaction, since
the significance of the main effects cannot be tested in the presence of an interaction.
If there is a significant interaction we can use something called an interaction plot to
allow us to examine the interaction and explain what is happening. If there is no
interaction we then consider testing for the effects of the two treatments.

In the above example the interaction term is not significant. If we just consider the
effect of method on its own, it appears to be strongly insignificant, whereas the effect of
rate on its own is very significant. The underlying normality and constant variance
assumption have to be verified in order for the ANOVA results to be valid. This has
been done for this example, but the results are not reported here.

Interactions
Although the interaction term is not significant in the above example, we present two
interaction plots from the example to illustrate how interaction affects can be
graphically represented. An interaction plot basically plots the mean of each level of
one treatment variable, at each level of the other treatment variable. If all the means
follow the same general pattern, then there will be no interaction. If some levels follow
different patterns, there will be an interaction. In the haemoglobin example we get the
two plots in Figure 7.

IAUL & D EPARTMENT OF S TATISTICS P AGE

21
I NTRODUCTION TO S TATISTICAL M ODELLING T RINITY 2002

11 11

10 10

Marginal mean of haemoglobin

Marginal mean of haemoglobin

9 9

RATE
8 8
1.00

2.00 METHOD
7 7

3.00 A

6 4.00 6 B
A B 1.00 2.00 3.00 4.00

METHOD RATE

Figure 7: Interaction plots for the haemoglobin example.

The means in the above plots all follow the same general pattern, as we would expect
from the non-significant interaction term. The use of interaction plots can be very
useful to explore the practical importance of interactions between two variables. As an
example, let us consider the following hypothetical graphs in Figure 8.

Non-significant A, Significant B Significant A, Non-significant B

35 25
30
Average Response

Average Response

20
25
Level 1 of B Level 1 of B
20 15
Level 2 of B Level 2 of B
15 10 Level 3 of B
Level 3 of B
10
5
5
0 0
1 2 3 1 2 3
Factor A Factor A

Both A,B Significant, No interaction Both A,B Significant and an interaction

30 30

25 25
Average Response
Average Response

20 20 Level 1 of B
Level 1 of B
15 Level 2 of B 15 Level 2 of B

10 Level 3 of B 10 Level 3 of B

5 5

0 0
1 2 3 1 2 3
Factor A Factor A

Figure 8: Some hypothetical interaction plots.

It is quite clear from these graphs, the effect that a significant main effect of a variable
and an interaction would have on the interaction plot.

You should always pay particular attention to interaction terms and their
interpretation. In particular, you should never consider a model that has an interaction

IAUL & D EPARTMENT OF S TATISTICS P AGE

22
I NTRODUCTION TO S TATISTICAL M ODELLING T RINITY 2002

between two variables without having both of the variables included; not including the
main effects of an interaction does not make intuitive sense and is dangerous, although
a lot of software packages will allow you to do this. You should also be careful when
interpreting the results of any particular modelling procedure, if you have a significant
interaction between two variables, you can not say anything about the effect of one
variable on the response, however, you can say that your variable of interest and its
interaction with another variable has an effect on the response. In many observational
studies, interactions can often be one of the most interesting findings, since they imply
that the two variables do not act in isolation on the response, they in fact act together
and this joint interaction can give a researcher great insight into the effect of the
variables on the response.

Treatment contrasts and multiple comparisons

As an example consider a two-factor study where the response is the score on a 20 item
multiple choice test over a taped lecture, and the two factors are cognitive style
(FI=Field Independent, FD=Field Dependent) and study technique (NN=No notes,
SN=Student Notes, PO=Partial Outline Supplied, CO=Complete Outline)6.

We initially consider an interaction plot for this data to see whether there is any
evidence that there may be an interaction.
18

studytec

SN
17

CO
PO
NN
16
mean of score

15
14
13

FD FI

cogstyle

It appears from the above plot that there is an interaction between cognitive style and
study technique, so we consider the ANOVA with the interaction term included. The
ANOVA table is as follows:

6 The data were collected by Frank (1984) “Effects of field independence-dependence and study technique on
learning from a lecture” Amer.Educ.Res.J., 21, 669-678

IAUL & D EPARTMENT OF S TATISTICS P AGE

23
I NTRODUCTION TO S TATISTICAL M ODELLING T RINITY 2002

ANOVA output

Dependent Variable: SCORE

Sum of
Source Squares df Mean Square F Sig.
COGSTYLE 25.010 1 25.010 7.784 .006
STUDYTEC 320.183 3 106.728 33.216 .000
COGSTYLE * STUDYTEC 27.260 3 9.087 2.828 .043
Error 308.462 96 3.213

Clearly from this table, the interaction term is significant. It is now of interest to
perform an analysis of effects. In this example we are interested in testing for
differences in each of the four study techniques. However, as these techniques occur in
each of the two cognitive style groups we will have to carry out the multiple
comparisons in each group. The results from this analysis are given below. The
interesting result from this analysis is that it appears that the differences in study
technique are the same in each of the cognitive method groups; thus suggesting that
most of the differences in the response variables come from differences in study
technique rather than cognitive style.
95 % simultaneous confidence intervals for specified
linear combinations, by the Sidak method

critical point: 2.9238

response variable: cogstyle

intervals excluding 0 are flagged by '****'

Estimate Std.Error Lower Bound Upper Bound

CO.adj1-NN.adj1 4.4600 0.703 2.4100 6.520 ****
CO.adj1-PO.adj1 0.7690 0.703 -1.2900 2.820
CO.adj1-SN.adj1 2.1500 0.703 0.0982 4.210 ****
NN.adj1-PO.adj1 -3.6900 0.703 -5.7500 -1.640 ****
NN.adj1-SN.adj1 -2.3100 0.703 -4.3600 -0.252 ****
PO.adj1-SN.adj1 1.3800 0.703 -0.6710 3.440
CO.adj2-NN.adj2 4.3800 0.703 2.3300 6.440 ****
CO.adj2-PO.adj2 0.0769 0.703 -1.9800 2.130
CO.adj2-SN.adj2 -0.3850 0.703 -2.4400 1.670
NN.adj2-PO.adj2 -4.3100 0.703 -6.3600 -2.250 ****
NN.adj2-SN.adj2 -4.7700 0.703 -6.8200 -2.710 ****
PO.adj2-SN.adj2 -0.4620 0.703 -2.5200 1.590

The confidence intervals above are represented graphically in figure 8 below.

IAUL & D EPARTMENT OF S TATISTICS P AGE

24
I NTRODUCTION TO S TATISTICAL M ODELLING T RINITY 2002

CO.adj1-NN.adj1 ( )
CO.adj1-PO.adj1 ( )
CO.adj1-SN.adj1 ( )
NN.adj1-PO.adj1 ( )
NN.adj1-SN.adj1 ( )
PO.adj1-SN.adj1 ( )
CO.adj2-NN.adj2 ( )
CO.adj2-PO.adj2 ( )
CO.adj2-SN.adj2 ( )
NN.adj2-PO.adj2 ( )
NN.adj2-SN.adj2 ( )
PO.adj2-SN.adj2 ( )

-7 -6 -5 -4 -3 -2 -1 0 1 2 3 4 5 6 7
simultaneous 95 % confidence limits, Sidak method
response variable: cogstyle

Figure 8: Graphical representation of confidence intervals above.

Non-parametric two-way ANOVA

Just as in the one-way case, we can equivalently consider a non-parametric two-way
ANOVA. However, it now becomes considerably more complex and the common test
statistic in this setting is restricted to the following special case. Consider the case
where you have two factors, one representing the treatment of interest and the other
representing a “blocking” variable (see the section on complete block designs). The
interest is then in assessing whether there are any differences between the treatments
and to do this we can use the Friedman rank sum test. The generalisation to more
complex settings is possible but requires a lot of work and is not generally
implemented in the computing packages. The interpretation is identical to that of the
Kruskal-Wallis test for the one-way ANOVA, you will get a test statistic and also a p-
value, and as usual you can conclude that there is a significant difference between your
treatments if the p-value is less than 0.05.

More than two factors and covariates

Although all of the techniques described above only cover the case of either one or two-
factor models the same ideas can be extended to three- or more factor models or to
models that include continuous covariates. The main issue to consider is which model
you are going to fit – i.e. do you fit the full factorial model which includes all the main
effects and all of the interactions (two-way and higher) or do you only fit a subset of
this model, perhaps one which only contains the main effects and two-factor
interactions. These decisions will be guided by the aim of your analysis and what
hypothesis you wish to test and is called model selection. You have the same choices if
you include covariates in the model. In the next section we will consider the case
where model includes continuous covariates. This type of analysis is called analysis of
covariance.

IAUL & D EPARTMENT OF S TATISTICS P AGE

25
I NTRODUCTION TO S TATISTICAL M ODELLING T RINITY 2002

Analysis of covariance

Introduction
Analysis of covariance incorporates one or more regression variables into an analysis of
variance. The regression variables are typically continuous and are referred to as
covariates, hence the name analysis of covariance. In this course we will only examine
the use of a single covariate. We discuss the analysis of covariance at the hand of an
example that involves one-way analysis of variance and a covariate.

Example
We consider data on the body weights (in kilograms) and heart weights (in grams) for
domestic cats of both sexes that were given digitalis. Part of the original data is
presented in Table 7 7, 8. The primary interest is to determine whether females’ heart
weights differ from males’ when both have received digitalis. A first step in the
analysis might be to fit a one-way ANOVA model by ignoring the body weight
variable. Such a model would be given by

Yij = µ + α i + ε ij i = 1,2; j = 1,K,24,

where the y ij are the heart weights. The ANOVA output for this model is given below.

Table 7: Body weights (kg) and heart weights (g) of domestic cats with digitalis.

7The data was originally published by Fisher, R. A. (1947). The analysis of covariance method for the relation
between a part and the whole. Biometrics, 3, 65-68.
8 The data is taken from Christensen, R. (1996). Analysis of Variance, Design and Regression: Applied statistical methods.

New York: Chapman & Hall. Here a subset of the original data is used to illustrate the application of the analysis of
covariance.

IAUL & D EPARTMENT OF S TATISTICS P AGE

26
I NTRODUCTION TO S TATISTICAL M ODELLING T RINITY 2002

Females Males

Body Heart Body Heart Body Heart Body Heart

2.3 9.6 2.0 7.4 2.8 10.0 2.9 9.4

3.0 10.6 2.3 7.3 3.1 12.1 2.4 9.3
2.9 9.9 2.2 7.1 3.0 13.8 2.2 7.2
2.4 8.7 2.3 9.0 2.7 12.0 2.9 11.3
2.3 10.1 2.1 7.6 2.8 12.0 2.5 8.8
2.0 7.0 2.0 9.5 2.1 10.1 3.1 9.9
2.2 11.0 2.9 10.1 3.3 11.5 3.0 13.3
2.1 8.2 2.7 10.2 3.4 12.2 2.5 12.7
2.3 9.0 2.6 10.1 2.8 13.5 3.4 14.4
2.1 7.3 2.3 9.5 2.7 10.4 3.0 10.0
2.1 8.5 2.6 8.7 3.2 11.6 2.6 10.5
2.2 9.7 2.1 7.2 3.0 10.6 2.5 8.6

one-way ANOVA output

BRAIN
Sum of
Squares df Mean Square F Sig.
Between Groups 56.117 1 56.117 23.444 .000
Within Groups 110.106 46 2.394
Total 166.223 47

We see that the effect due to sexes is highly significant in the one-way ANOVA. Until
now we have not made use of the extra value of body weight that we have in addition
to the brain weight variable. It is natural to consider whether effective use can be made
of this value.

In order to make use of the body weight observations, we add a regression term to the
one-way ANOVA model above and fit the traditional analysis of covariance model,

(
y ij = µ + α i + γz ij + ε ij , i = 1,2; j = 1,K ,24, ε ij i.i.d. N 0, σ 2 )
In the above model the z ij ’s are the body weights and γ is a slope parameter
associated with body weights. Note that this model is an extension of the simple linear
regression model between the y s and z s in which we allow a different intercept µ i for
each sex. The ANOVA table for this model is given below.

IAUL & D EPARTMENT OF S TATISTICS P AGE

27
I NTRODUCTION TO S TATISTICAL M ODELLING T RINITY 2002

ANOVA output for the covariance model

Dependent Variable: BRAIN

Sum of
Source Squares df Mean Square F Sig.
Intercept 5.433 1 5.433 3.383 .072
SEX 4.499 1 4.499 2.801 .101
BODY 37.828 1 37.828 23.551 .000
Error 72.279 45 1.606
Total (adjusted) 166.223 47

The interpretation of the above ANOVA table is different from the earlier ANOVA
tables. We note that the sum of squares for body weights, sex and error do not add up
to the total sum of squares, for example. The sums of squares in the above ANOVA
table are referred to as adjusted sums of squares because the body weight sum of squares
is adjusted for sexes and the sex sum of squares is adjusted for body weights. We do
not discuss the computation here. Interested readers can refer to Christiansen (1996)
for the computational details. The error line in the table above is simply the error from
fitting the covariance model. The only difference between the one-way model and the
covariance model is that the one-way model does not involve the regression of on body
weights, so by testing the models we are testing whether there is a significant effect due
to the regression on body weights. The standard way of comparing a full and a
reduced model is by comparing their error terms. We see from the ANOVA table
above that there is a major effect due to the regression on body weights. Figures 9, 10
and 11 contain residual plots for the covariance model. The plot of the residuals versus
the predicted values (figure 9) looks good, while figure 10 shows slightly less
variability for females than males. The difference is not very large though and we need
perhaps not worry about it too much. The normal plot of the residuals (figure 11) also
seems reasonably acceptable.
13
12
11
residuals
10
9

-2 -1 0 1 2
fitted values

IAUL & D EPARTMENT OF S TATISTICS P AGE

28
I NTRODUCTION TO S TATISTICAL M ODELLING T RINITY 2002

Figure 9: Residuals versus predicted values of the covariance model

2
1
residuals
0
-1
-2

FEMALE MALE

sex

Figure 10: Box plots of residuals by sex for the residuals of the covariance model.
2
1
0
-1
-2

-2 -1 0 1 2

Quantiles of Standard Normal

Figure 11: Normal Q-Q plot of the standardised residuals of the covariance model

Analysis of covariance in designed experiments

IAUL & D EPARTMENT OF S TATISTICS P AGE

29
I NTRODUCTION TO S TATISTICAL M ODELLING T RINITY 2002

In the previous example we are dealing with an observational study as opposed to a

designed experiment. In a designed experiment the role of covariates is solely to
reduce the error of treatment comparisons. For a covariate to be of help in an analysis
it must be related to the dependent variable. Unfortunately, improper use of covariates
can invalidate, or alter, comparisons among treatments. In the observational study
above the very nature of what we were comparing changed when we adjusted for body
weights. Originally we investigated whether heart weights were different for females
and males. The analysis of covariance examined whether there were differences
between female and male heart weights beyond what could be accounted for by the
regression on body weights. These are quite different interpretations.

In a designed experiment, we want to investigate the effects of the treatments and not
the treatments adjusted for some covariates. Cox (1958) refers to a supplementary
observation that may be used to increase precision as a concomitant observation. It is
stated that an important condition for the use of a concomitant observation is that after
its use, estimated effects for the desired main observation shall still be obtained. This
condition means that the concomitant observations should be unaffected by the
treatments. In practice this means that either the concomitant observations are taken
before the assignment of the treatments, or the concomitant observations are made
after the assignment of the treatments, but before the effect of treatments has had time
to develop. These requirements on the nature of covariates in a designed experiment
are imposed so that the treatment effects do not depend on the presence or absence of
the covariates in the analysis. The treatment effects are logically independent
regardless of whether covariates are measured or incorporated in the analysis.

Multivariate analysis of variance (MANOVA)

The multivariate approach to analysing data that contain repeated measurements on
each subject involves using the repeated measures as separate dependent variables in a
collection of standard analyses of variance. The method of analysis, known as
multivariate analysis of variance (MANOVA), then combines results from the several
ANOVAs. A detailed discussion of MANOVA is beyond the scope of this course.
Readers who are interested to learn more about the subject are referred to Johnson and
Wichern (1992).

A. Roddam(2000), K. Javaras and W. Vos (2002)

References
Box, G., J. S., Hunter, W. G., Hunter, J. S. (1978). Statistics for Experimenters: an
Introduction to design, Data Analysis, and Model Building. New York: John Wiley.

Christensen, R. (1996). Analysis of Variance, Design and Regression: Applied statistical

methods. New York: Chapman & Hall.

IAUL & D EPARTMENT OF S TATISTICS P AGE

30
I NTRODUCTION TO S TATISTICAL M ODELLING T RINITY 2002

Cox, D. R. (1958). Planning of Experiments. New York: John Wiley.

Fisher, R. A. (1947). The analysis of covariance method for the relation between a part
and the whole. Biometrics, 3, 65-68.

Hettmansperger, T. P. and McKean, J. W. (1998). Robust Nonparametric Statistical

Methods: Kendall’s Library of Statistics 5. London: Arnold.

Johnson, R. A. and Wichern, D. W. (1992). Applied Multivariate Statistical Analysis. New

York: Prentice Hall.

Rencher, A. C. (2000). Linear Models in Statistics. Wiley: New York.

Scheffé, H. (1959). The Analysis of Variance. New York: John Wiley.

IAUL & D EPARTMENT OF S TATISTICS P AGE

31