Analysis of Variance

School of Public Health

 Objectives
At this end of this session students are able to:

• Understand the basic concepts, purposes & assumptions of ANOVA

• Differentiate between One way ANOVA and Two way ANOVA

• Familiarize the different post hoc tests

• Test hypothesis about the comparison of means of three or more

groups

School of Public Health 2

 Analysis of Variance
• A analysis of variance is a technique that partitions the total sum of
squares of deviations of the observations about their mean into
portions associated with independent variables in the experiment
and a portion associated with error.
• A factor refers to a categorical quantity under examination in an
experiment as a possible cause of variation in the response
variable.
• Levels refer to the categories, measurements, or strata of a factor
of interest in the experiment.
School of Public Health 3
 Types of Experimental Designs

Experimental
Designs

Completely Randomized Factorial

Randomized Block

One-Way Two-Way
Anova Anova

School of Public Health 4

 Completely Randomized Design

1. Experimental Units (Subjects) Are Assigned

Randomly to Treatments
– Subjects are Assumed Homogeneous
2. One Factor or Independent Variable
– 2 or More Treatment Levels or groups
3. Analyzed by One-Way ANOVA

School of Public Health 5

 Completely randomized …

School of Public Health

 One-Way ANOVA F-Test

1. Tests the Equality of 2 or More (p)

Population Means

2. Variables
– One categorical Independent Variable
– One Continuous Dependent Variable

School of Public Health 7

 Assumptions
1. Randomness & Independence of Errors
2. Normality
– Populations (for each condition) are Normally
Distributed
3.Homogeneity of Variance
– Populations (for each condition) have Equal
Variances

School of Public Health 8

 Hypotheses
H0: 1 = 2 = 3 = ... = p
– All Population Means are Equal
– No Treatment Effect
Ha: Not All j Are Equal
– At Least 1 Pop. Mean is Different
– Treatment Effect
– NOT 1  2  ...  p

9
 Hypotheses

H0: 1 = 2 = 3 = ... = p
– All Population Means
f(X)
are Equal
– No Treatment Effect
X
Ha: Not All j Are Equal 1 = 2 = 3
– At Least 1 Pop. Mean
is Different f(X)
– Treatment Effect
– NOT 1 = 2 = ... = p
X
– Or i ≠ j for some i, j. 1 =  2  3
10
 One-Way ANOVA Basic Idea
1. Compares 2 Types of Variation to Test
Equality of Means
2. If Treatment Variation Is Significantly Greater
Than Random Variation then Means Are Not
Equal
3.Variation Measures Are Obtained by
‘Partitioning’ Total Variation

School of Public Health 11

 One-Way ANOVA Partitions Total Variation

Total variation

Variation due to Variation due to

treatment random sampling

Sum of Squares Among Sum of Squares Within

Sum of Squares Between Sum of Squares Error
Sum of Squares Treatment (SST) (SSE)
Among Groups Variation Within Groups Variation
12
13
 The Model

• μ represents the mean of all the k population means

and is called the grand mean.
• Tj represents the difference between the mean of
the jth population and the grand mean and is called
the treatment effect.
• eij represents the amount by which an individual
measurement differs from the mean of the
population to which it belongs and is called the
error term.
14

School of Public Health

 Total Variation

SS Total  x11  x   x21  x     xij  x 

2 2 2

Response, x

x

Group 1 Group 2 Group 3

School of Public Health 15
 Treatment Variation

SST n1 x1  x   n2 x2  x     n p x p  x 

2 2 2

Response, x
x3
x
x2
x1

Group 1 Group 2 Group 3

School of Public Health 16
 Random (Error) Variation

SSE x11  x1   x21  x1     x pj  x p 

2 2 2

Response, x

x3
x2
x1

Group 1 Group 2 Group 3

School of Public Health 17
 F-Test Test Statistic
• 1. Test Statistic SST /  p  1

– F = MST / MSE=V.R. SSE / n  p 
• MST Is Mean Square for Treatment
• MSE Is Mean Square for Error
• 2. Degrees of Freedom
– df1 = p -1
– df2 = n - p
• p = # Groups, or Levels
• n = Total Sample Size

School of Public Health 18

 One-Way ANOVA Summary Table

Source of Degrees Sum of Mean F

Variation of Squares Square
Freedom (Variance)
Treatment p-1 SST MST = MST
SST/(p - 1) MSE
Error n-p SSE MSE =
SSE/(n - p)
Total n-1 SS(Total) =
SST+SSE
19
 One-Way ANOVA F-Test Critical Value

If means are equal,

F = MST / MSE  1.
Only reject large F! Reject H0

Do Not 
Reject H0

0 F
Fa ( p  1, n  p)

Always One-Tail!
© 1984-1994 T/Maker Co.
School of Public Health 20
 HOW TO CALCULATE ANOVA’S BY HAND…

Treatment 1 Treatment 2 Treatment 3 Treatment 4

y11 y21 y31 y41 n=10 obs./group
y12 y22 y32 y42
y13 y23 y33 y43 k=4 groups
y14 y24 y34 y44
y15 y25 y35 y45
y16 y26 y36 y46
y17 y27 y37 y47
y18 y28 y38 y48
y19 y29 y39 y49
y110 y210 y310 y410
10

y
10 10 10

1j  y2 j  y3 j y 4j The group
j 1 j 1
y1  y 2 
j 1
y 3 
j 1 y 4  means
10 10 10 10
10 10 10
 (y
10
 y 2 ) 2   (y
j 1
( y1 j  y1 ) 2

j 1
2j
j 1
( y 3 j  y 3 ) 2
j 1
4j  y 4 ) 2
The (within) 21
10  1 10  1 10  1 10  1 group
variances
School of Public Health
 SSW/SSE

10 10 10
 (y
10 2
 ( y1 j  y1 ) 2

j 1
2j  y 2 ) 
j 1
( y 3 j  y 3 ) 2
 (y
j 1
4j  y 4 ) 2
The (within)
j 1
group
10  1 10  1 10  1 10  1 variances

10 10
10 10

 (y 1j  y1 ) +2
 (y 2j  y 2 ) + 2
 ( y 3 j  y 3 ) 2 + j 1
( y 4 j  y 4 ) 2
j 1 j 3
j 1

4 10

  (y 2 Sum of Squares Within (SSW)

 ij  y i ) (or SSE, for chance error)

i 1 j 1

22
 SSB/SSR

4 10
Overall
mean of all  y
i 1 j 1
ij
40
observation y 
s (“grand 40
mean”)

4 Sum of Squares

 (y  y  ) 2 Between (SSB).
10x i
Variability of the
group means
i 1 compared to the
grand mean (the
variability due to the
treatment).

23
 Total Sum of Squares (SST)

Total sum of
4 10 squares(TSS).


Squared difference
( y ij  y   ) 2 of every observation
from the overall
i 1 j 1 mean. (numerator of
variance of Y!)

24
 Partitioning of Variance

4 10 4 4 10
 ( y
i 1 j 1
ij  y i ) 2

+10x ( y i  y  ) 2 =  ( y ij  y  ) 2
i 1 i 1 j 1

SSW + SSB =
TSS

25
 ANOVA Table

Mean Sum
Source of Sum of
of Squares
variation d.f. squares
F-statistic p-value

Between k-1 SSB SSB/k-1 Go to

(k groups) (sum of squared SSB Fk-1,nk-k
k1
deviations of group SSW
means from grand nk  k chart
mean)

Within nk-k SSW s2=SSW/nk-k

(sum of squared
(n individuals per
deviations of
group)
observations from
their group mean)

TSS
Total nk-1 (sum of squared deviations of
variation observations from grand mean)

TSS=SSB + SSW
26
 Example
Treatment 1 Treatment 2 Treatment 3 Treatment 4
60 inches 50 48 47
67 52 49 67
42 43 50 54
67 67 55 67
56 67 56 68
62 59 61 65
64 67 61 65
59 64 60 56
72 63 59 60
71 65 64 65

27
 Example

Step 1) calculate the Treatment 1 Treatment 2 Treatment 3 Treatment 4

sum of squares between 60 inches 50 48 47
groups: 67 52 49 67
42 43 50 54
Mean for group 1 =
67 67 55 67
62.0 56 67 56 68
Mean for group 2 = 59.7 62 59 61 65
64 67 61 65
Mean for group 3 = 56.3
59 64 60 56
Mean for group 4 = 61.4 72 63 59 60

Grand mean= 59.85 71 65 64 65

SSB = [(62-59.85)2 + (59.7-59.85)2 + (56.3-59.85)2 + (61.4-59.85)2 ] xn per

group= 19.65x10 = 196.5

28
 Example

Treatment 1 Treatment 2 Treatment 3 Treatment 4

Step 2) calculate the
60 inches 50 48 47
sum of squares within
groups: 67 52 49 67

42 43 50 54

(60-62) 2+(67-62) 2+ (42- 67 67 55 67

62) 2+ (67-62) 2+ (56-62) 56 67 56 68
2
+ (62-62) 2+ (64-62) 2+
62 59 61 65
(59-62) 2+ (72-62) 2+ (71-
62) 2+ (50-59.7) 2+ (52- 64 67 61 65
59.7) 2+ (43-59.7) 2+67-
59 64 60 56
59.7) 2+ (67-59.7) 2+ (69-
59.7) 2…+….(sum of 40 72 63 59 60

squared deviations) = 71 65 64 65
2060.6

29
 Fill in the ANOVA table

Source of d.f. Sum of squares Mean Sum of F-statistic p-value

variation Squares

Between 3 196.5 65.5 1.14 .344

Within 36 2060.6 57.2

Total 39 2257.1

30
 Fill in the ANOVA table

Source of d.f. Sum of squares Mean Sum of F-statistic p-value

variation Squares

Between 3 196.5 65.5

2060.6 57.2 1.14 .344

Within 36

Total 39 2257.1

INTERPRETATION of ANOVA:
How much of the variance in height is explained by treatment
group?
R2=“Coefficient of Determination” = SSB/TSS =
196.5/2275.1=9%
31
 Coefficient of Determination

2 SSB SSB
R  
SSB  SSE SST
The amount of variation in the outcome variable
(dependent variable) that is explained by the predictor
(independent variable).

32

School of Public Health

 Multiple Comparisons/Post hoc tests

• ANOVA only tells us if there is an effect

– That is, are the means of the groups not all equal?
• ANOVA does not tell us which means are different
from which other means
• Multiple comparisons are used to determine which
means are probably different from which other means
33

School of Public Health

 Multiple Comparisons…
• Multiple comparisons are not fundamentally
different from performing a t-test
• That is, both multiple comparisons and t-tests
answer the same question: are two means different
from each other
• The difference is that the multiple comparisons
protect us from making a Type-I error even though
we perform many such comparisons

34

School of Public Health

 Multiple Comparisons
• There are many types of multiple comparison
– E.g. Tukey, Newman-Keuls, Dunnett, Scheffé,
Bonferroni
• There is little agreement about which particular
one is best
• The different tests trade off statistical power for
protection from Type-I errors

35

School of Public Health

 Tukey Tests

• We will consider only the Tukey test; other people

may feel that other tests are more appropriate
• The Tukey test offers reasonable protection from
Type-I errors while maintaining reasonable statistical
power

36

School of Public Health

 Tukey Tests
• You should perform Tukey tests only when two
criteria have been met:
– There are more than two levels to the IV
– The IV is statistically significant
• Write the hypotheses
– H0: m1 = m2
– H1: m1 ¹ m2
• Specify the a level
– a = .05 37

School of Public Health

 Steps in Performing Tukey Tests
Calculate the honestly significant difference (HSD)
MS within groups
HSD q
 , k , df with in g ro u p s n
q,k,dfwithin-groups = tabled q value
 =  level
k = number of levels of the IV
dfwithin-groups = degrees of freedom for MSwithin-groups
MSwithin-groups = within-groups variance estimate
n = number of participants

38
 Steps in Performing Tukey Tests
• Take the difference of the means of the
conditions you are comparing
• If the difference of the means is at least as large
as the HSD, you can reject H0
• Repeat for whatever other comparisons need to
be made

39

School of Public Health

 Tukey’s Test for Unequal Sample Sizes

• Any absolute value of the difference between two

sample means that exceeds HSD* is declared
significant.

40

School of Public Health

41
42
 Example
• Let us illustrate the use of the HSD test with the data
from Example 8.2.1.
• Solution: The first step is to prepare a table of all
possible (ordered) differences between means. The
results of this step for the present example are
displayed in Table 8.2.5.

43
44
 Bonferroni and Sidak test
 Bonferroni procedures
 Bonferroni adjustment simply reduces the
comparisons of interest based on the number of
comparisons being made
Use ` = where c is the number of comparisons
Technically we could adjust in such a fashion that
some comparisons are more liberal than others, but
this is the default approach in most statistical
packages

45

School of Public Health

 Bonferroni
• While the traditional Bonferroni adjustment is widely used
(and makes for an easy approach to eyeball comparisons
yourself), it generally is too conservative in its standard from

• It is not recommended that you use it if there are a great

many comparisons, as your pairwise comparisons would be
using very low alpha levels
– E.g. 7 groups: each comparison would be tested at alpha = .002

46

School of Public Health

 TWO-WAY ANALYSIS OF VARIANCE
• Two-way ANOVA is a statistical with one numerical
outcome(dependent) variable and two categorical
explanatory (independent)variables.
• Example – In a completely randomised design we may
wish to compare outcome by age, gender or disease
severity. Subjects are grouped by one such factor and
then randomly assigned one treatment.
• Technical term for such a group is block and the study
design is also called randomised block design
47

School of Public Health

 RANDOMISED BLOCK DESIGN

• Blocks are formed on the basis of expected homogeneity

of response in each block (or group).
• The purpose is to reduce variation in response within
each block (or group) due to biological differences
between individual subjects on account of age, sex or
severity of disease.

48

School of Public Health

 RANDOMISED BLOCK DESIGN
• Randomised block design is a more robust
design than the simple randomised design.
• The investigator can take into account
simultaneously the effects of two factors on an
outcome of interest.
• Additionally, the investigator can test for
interaction, if any, between the two factors.

49

School of Public Health

 STEPS IN PLANNING A RANDOMISED
BLOCK DESIGN
1) Subjects are randomly selected to constitute a random
sample.
2) Subjects likely to have similar response
(homogeneity) are put together to form a block.
3) To each member in a block intervention is assigned
such that each subject receives one treatment.
4) Comparisons of treatment outcomes are made within
each block

50

School of Public Health

 Randomized complete block design

School of Public Health

 where

52

School of Public Health

 Two- Way ANOVA
• The technique for analyzing the data from a randomized complete
block design is called two-way analysis of variance
– since an observation categorized on the basis of two criteria the
block to which it belongs as well the treatment group to which
it belongs.
– The treatments are assigned at random to the experimental
units within each block.
• In the two-way ANOVA model, there are two factors, each with
its own number of levels.
– When we are interested in the effects of two factors, it is much
more advantageous to perform a two-way analysis of variance,
as opposed to two separate one-way ANOVAs.
53

School of Public Health

 Cont…
• There are three main advantages of two-way
ANOVA:
– It is more efficient to study two factors
simultaneously rather than separately.
– We can reduce the residual variation in a model by
including a second factor thought to influence the
response.
– We can investigate interactions between factors.

54

School of Public Health

 Cont…
• The steps for hypothesis testing when the randomized
complete block design is used are as follows:
– Data. After identifying the treatments, the blocks, and the
experimental units, the data, for convenience, may be
displayed as in Table below

55

School of Public Health

Assumptions. The model for the randomized complete
block design and its underlying assumptions are as follows:

56

School of Public Health

 Cont…
• Hypotheses. We may test
– H0: Tj = 0. j =1,2,..., k
against the alternative
– HA : not all Tj = 0
• Test statistic. The test static V. R.
• Distribution of test statistic. When H0 is true and the
assumptions are met, V.R. follows an F distribution.
• Decision rule. Reject the null hypothesis if the computed value
of the test statistic V.R. is equal to or greater than the critical
value of F.

57

School of Public Health

The SST for the randomized complete block design can be
partitioned into three components, one each attributable to
treatments (SSTr), blocks (SSB1), and error (SSE). That is,

SST SSBl  SSTr  SSE

k n
Degrees of freedom
SST   ( xij  x.. ) 2
• Total = kn - 1
j 1 i 1
k n
SSBl   ( xi.  x.. ) 2 • Blocks = n - 1
j 1 i 1
k n
SSTr  ( x . j  x ..) 2 • Treatments = k – 1
j 1 i 1

SSE SST  SSBl  SSTr • Residual(error) = (n-1)(k-1)

58
 TWO-WAY ANOVA
TOTAL VARIATION PARTITIONING

Total
Total Variation
Variation
SS(Total)

Variation
VariationDue
Dueto
to Variation
VariationDue
Dueto
to
Treatment
TreatmentAA Treatment
Treatment BB
SSA SSB

Variation
VariationDue
Dueto
to
Random
Random Sampling
Sampling
SSE
59
Two-Way Summary Table

Source of Degrees of Sum of Mean F

Variation Freedom Squares Square
Bl n-1 SSBl MSBl MSBl
(blocks) MSE
Tr SSTr MSTr MSTr
(treatments) k-1 MSE

Error (k-1)(n-1) SSE MSE

Total kn - 1 SS(Total)
60
Statistical decision. When the null hypothesis is true, therefore,
the quantity MSTr/MSE is distributed as F with k — 1 numerator
degrees of freedom and (n — 1) X (k — 1) denominator degrees of
freedom. The computed variance ratio, therefore, is compared with
the critical-value of F.
Conclusion. If we reject H0, we conclude that the alternative
hypothesis is true. If we fail to reject H0, we conclude that H0 may
be true.
p value.

61
 Example
• A physical therapist wished to compare three methods
for teaching patients to use a certain prosthetic device.
He felt that the rate of learning would be different for
patients of different ages and wished to design an
experiment in which the influence of age could be taken
into account.

62

School of Public Health

 .

63

School of Public Health

 .

64

School of Public Health

 .

65

School of Public Health

 .

66

School of Public Health

 .

67

School of Public Health

 The Repeated measures design
• One of the most frequently used experimental
designs in the health sciences field is the repeated
measures design.
• Definition: A repeated measures design is one in
which measurements of the same variable are made
on each subject on two or more different occasions.
• The different occasions during which measurement
are taken may be either points in time or different
conditions such as different treatments.
68

School of Public Health

 When to Use Repeated Measurement
• The usual motivation for using a repeated measures
design is a desire to control for variability among
subjects.
• Advantages
– ability to control for extraneous variation among subjects.
– fewer subjects are needed than for a design in which
different subjects are used for each occasion on which
measurements are made.
• Disadvantages
– carry-over effect.
– position effect. 69

School of Public Health

 Single-Factor Repeated Measures Design

• The simplest repeated measures design is the one

in which, in addition to the treatment variable,
one additional variable is considered.
• The reason for introducing this additional
variable is to measure and isolate its
contribution to the total variability among the
observations.

70

School of Public Health

 Assumptions
• The subjects under study constitute a simple random sample
from a population of similar subjects.
• Each observation is an independent simple random sample of
size I from each of kn populations, where n is the number of
subjects and k is the number of treatments to which each
subject is exposed.
• The kn populations have potentially different means, but
they all have the same variance.
• Tire k treatments are fixed: that is, they are the only
treatments about which we have an interest in the current
situation.
• There is no interaction between treatments and subjects 71

School of Public Health

 Example

Licciardone et al. (A-11) examined subjects with chronic,

nonspecific low back pain. In this study, 18 of the subjects
completed a survey questionnaire assessing physical functioning at
baseline, and after 1, 3, and 6 months. Table 8.4.1 shows the data
for these subjects who received a sham treatment that appeared to
be genuine osteopathic manipulation. Higher values indicate better
physical functioning. The goal of the experiment was to determine
if subjects would report improvement over time even though the
treatment they received would provide minimal improvement. We
wish to know if there is a difference in the mean survey values
among the four points in time
72

School of Public Health

 .

73

School of Public Health

 Solution

74

School of Public Health

 .

75

School of Public Health

 The factorial analysis of variance
• In the experimental designs that we have considered up to
this point, we have been interested in the effects of only
one variable—the treatments.
• Frequently, however, we may be interested in studying,
simultaneously, the effects of two or more variables.
• We refer to the variables in which we are interested as
factors.
• The experiment in which two or more factors are
investigated simultaneously is called a factorial experiment.
• The different designated categories of the factors are called
levels. 76

School of Public Health

• Suppose, for example. that we are studying the effect on
reaction time of three dosages of some drug.
– The drug factor, then, is said to occur at three levels.
– Suppose the second factor of interest in the study is age, and it is
thought that two age groups, under 65 years and 65 years and
older, should be included.
– We then have two levels of the age factor.
• In general, we say that factor A occurs at a levels and factor
B occurs at b levels.
• In a factorial experiment we may study not only the effects
of individual factors but also, if the experiment is properly
conducted. the interaction between factors. 77

School of Public Health

 Advantages
• The advantages of the factorial experiment include the
following.
1. The interaction of the factors may be
studied.
2. There is a saving of time and effort.
3. Because the various factors are
combined in one experiment, the
results have a wider range of
application.
78

School of Public Health

 79

School of Public Health

 80

School of Public Health

 The Two-Factor Completely Randomized Design
• A factorial arrangement may be studied with either of
the designs that have been discussed.
• Factorial experiment could be illustrated by means of a
two-factor completely randomized design.
• Data. The results from a two-factor completely
randomized design may be presented in tabular form as
shown in the next table

81

School of Public Health

 The Two-Factor Completely Randomized Design

Total
Total Variation
Variation
SS(Total)

Variation
VariationDue
Dueto
to Variation
VariationDue
Dueto
to
Treatment
TreatmentAA Treatment
Treatment BB
• SSA SSB

Variation
VariationDue
Duetoto Variation
VariationDue
Dueto
to
Interaction
Interaction Random
Random Sampling
Sampling
SS(AB) • SSE
82
 83

School of Public Health

 Cont…

• The Model The fixed-effects model for the two-factor

completely randomized design may be written as
xijk = μ + ai + βj+ (αβ)ij +lijk
i= 1, 2,…..a; j = 1,2,…b; k =1.2,…. n
• Assumptions of the Model
a. The observations in each of the ab cells constitute a random
independent sample of size n drawn from the population
defined by the particular combination of the levels of the two
factors.
b. Each of the ab populations is normally distributed.
c. The populations all have the same variance.
84

School of Public Health

 Two-Way ANOVA Null Hypotheses
1.No Difference in Means Due to Factor A
– H0: 1. = 2. =... = a.
2.No Difference in Means Due to Factor B
– H0: .1 = .2 =... = .b
3.No Interaction of Factors A & B
– H0: ABij = 0
• Before collecting data, the researchers may
decide to test only one of the possible
hypotheses.
School of Public Health 85
 Calculation of test statistic

86

School of Public Health

 87

School of Public Health

 Cont…
• Statistical decision. If the assumptions stated earlier hold
true, and if each hypothesis is true, it can be shown that
each of the variance ratios shown in Table 8.5.4 follows
an F distribution with the indicated degrees of freedom.
• We reject HO if the computed VR. values are equal to
greater than the corresponding critical values as
determined by the degrees of freedom and the chosen
significance levels.
• Conclusion. If we reject H0, we conclude that HA is true.
If we fail to reject H0, we conclude that H0 may be true.
• p value.
88

School of Public Health

 Example
• In a study of length of time spent on individual home
visits by public health nurses, data were reported on
length of home visit, in minutes. by a sample of 80
nurses. A record was made also of each nurse‘s age and
the type of illness of each patient visited, The
researchers wished to obtain from their investigation
answers to the following questions:
– Does the mean length of home visit differ among different age
groups of nurses?
– Does the type of patient affect the mean length of home visit?
– Is there interaction between nurse‘s age id type of patient?
89

School of Public Health

 90

School of Public Health

 Cont…
• Statistical decision. The variance ratios are
V.R. (A) =997.5/ 14.7 = 67.86, V.R. (B) =
400.4/14.7 = 27.21. and V.R. (AB) = 67.6/
14.7= 4.60. Since the three computed values
of V.R. are all greater than the
corresponding critical values, we reject all
null hypotheses.

91

School of Public Health

 Cont…
• Conclusion. When H0: a1 =a2 =a3= a4 Is rejected, we
conclude that there are differences among the levels of
A, that is, differences in the average amount of time
spent in home visits with different types of patients.
• Similarly, when H0. β1 = β 2 = β 3 = β 4 is rejected, we
conclude that there are differences among the levels of
B, or differences in the average amount of time spent on
home visits among the different nurses when grouped
by age.
• When H0: (a β) ij= 0 is rejected, we conclude that
factors A and B interact; that is, different combinations
of levels of the two factors produce different effects.
92

School of Public Health