One or More Way Analysis of Variance

Wolde M.
 Look the following research questions:

 Are the birth weights of children in different geographical

regions the same?

 Are the responses of patients to different medications and

placebo different?

 Are people with different age groups have different

proportion of body fat?

 Do people from different ethnicity have the same BMI?

 One way-Analysis Of Variance
(One-way ANOVA)

 All the above research questions have one common

characteristic: That is each of them has two variables: one
categorical and one quantitative

 Main question: Are the averages of the quantitative variable

across the groups (categories) the same?

 Because there is only one categorical independent variable

which has two or more categories (groups), the name one
way ANOVA comes.
 One-way ANOVA cont…

 One way ANOVA is also called Completely

Randomized Design

 Experimental units (subjects) are assigned randomly

to treatments/groups. Here subjects are assumed to
be homogeneous

4
 Analysis of variance cont…

 One way ANOVA is a method for testing the hypothesis:

There is no difference between two or more population
means (usually at least three); or there is no difference
between a number of treatments

 More formally, we can state hypotheses as:

H0: There is no difference among the mean of treatments effects

HA: There is difference at least between two treatments effects
or
Ho: µ1 = µ2 = µ3 =…. = µa (if there are „a‟ groups)
HA: at least one group mean is different
5
 Why Not Just Use t-tests?
 Since t-test considers two groups at a time, it will be tedious
when many groups are present

 Conducting multiple t-tests can lead also to severe inflation of

the Type I error rate (false positives) and is not recommended

 However, ANOVA is used to test for differences among several

means without increasing the Type I error rate

 The ANOVA uses data from all groups at a time to estimate

standard errors, which can increase the power of the analysis
 Assumptions of One Way ANOVA)

 The data are normally distributed or the samples have come

from normally distributed populations and are independent.

 The variance is the same in each group to be compared (equal

variance) homosidasity

 Moderate departures from normality may be safely ignored,

but the effect of unequal standard deviations may be serious.

 In the latter case, transforming the data may be useful.

 Analysis of variance cont…
 We test the equality of means among groups by using the
variance

 The difference between variation within groups and

variation between groups may help us to compare the
means

 If both are equal, it is likely that the observed difference is

due to chance and not real difference

Note that:
Total Variability = Variability between + Variability within
 Analysis of variance

Basic model: Data are deviations

μ from the global mean, μ:
Xij = μ + Ɛij
 Sum of vertical deviations squared is

G-1 G-2 the total sum of squares = SSt

 One way model: Data are deviations
A2
from treatment means, Ais:
Xij = μ + Ai + Ɛij
 Sum of vertical deviations squared = SSe
A1
 Note that ΣAi = ΣƐij = 0
G-1 G-2
Decomposing the total variability
n a n a na
 Total SS = Σ Σ (xij – )2 = ΣiΣjxij2 - (ΣiΣjxij)2 /na = SST
i=1 j=1

n a n a a n
 Within SS = Σ Σ (xij – j ) 2 = Σ Σ x 2 - Σ (Σ x )2/n
i j ij j i ij = SSW
i=1 j=1

n a a n
 Between SS = Σ Σ ( i j– )2 = Σj(Σixij)2/n - (ΣiΣjxij)2 /na = SSB
i=1 j=1

This is assuming each of the „a‟ groups has equal size, „n‟.

SST = SSW + SSB

10
 Data of one way ANOVA
Groups/variable
G-1 G-2 G-3 ….. G-a
X11 X12 X13 ….. X1a
…..
Participants
X21 X22 X23 X2a
X31 X32 X33 ….. X3a
. . . . .
. . . . .
. . . . .
Xn1 Xn2 Xn3 …. Xna
Totals T.1 T.2 T.3 …. T.a

Computational formula
T= ΣiΣjxij2 Correction Factor = CF = (ΣiΣjxij)2 /na = T2../na
A = Σj(Σixij)2/n = Σj(T.j)2/n if the groups‟ (cells‟) size are equal, or
A = Σj(Σixij)2/nj = Σj(T.j)2/nj ; if unequal group size

Where, Xij = ith observation in the jth group of the table

i = 1, 2, 3,…, nj, j = 1, 2, 3,…,a, Σ jn j = N
 Sum of squares and ANOVA Table

Source of df. SS MS F
variation
Between groups a-1 SSB = A - CF SSB/(a-1) MSB/MSW
Within groups na-a SSW = T - A SSW/(na –a)
Total na-1 SST = T - CF

 If there are real differences among groups‟ means, the between

groups variation will be larger than the within variation
 Example on one-way ANOVA
The following table shows the red cell folate levels (μg/l) in three groups of
cardiac bypass patients who were given three different levels of nitrous oxide
ventilation. (Level of nitrous oxide for group I > group II‟s > group III‟s)

Group I Group II Group III

(n=8) (n=9) (n=5)
243 206 241
251 210 258
275 226 270
291 249 293
347 255 328
354 273
380 285
392 295
309
Total=2533 2308 1390
Mean =316.6 256.4 278.0
SD = 58.7 37.1 33.8
 Example Cont….
We can see the box plot just to have some
impression about it
 Example cont…
Ho: μ1 = μ2 = μ3
HA: Differences exist between at least two of the means
Source of variation df SS Mean F P
square
Between groups 2 15516 7758
3.71 0.044
Within groups 19 39716 2090

Total 21 55232

Since the P-value is less than 0.05, the null hypothesis is rejected
 Pair-wise comparisons of group means
post hoc tests or multiple comparisons

 ANOVA test tells us only whether there is statistically significant

difference among groups means, but

 It doesn‟t tell us which particular groups are significantly

different

 To identify them, we use either a priori (pre-planed) or post hoc

tests
 Pair-wise comparisons of group means
(post hoc tests) cont…
 Whether to use a priori or post hoc tests depends on whether the
researcher has previously stated the hypotheses to test.

 If you have honestly stated beforehand the comparisons between

individual pairs of means which you intend to make, then you are
entitled to use a priori test such as a t-test.
 In this case, only one pair of groups or few will be tested

 However, when you look at the data it may seem worth comparing
all possible pairs. In this case, a post hoc test such as Scheffe,
Benferroni (modified t-test), Tuckey methods, Least Significant
Difference (LSD), Duncan, Dannett, etc. will be employed.
Benferroni method or Modified t-test (Steps)

I. Find tcalc for the pairs of groups of interest (to be compared)

II. The modified t-test is based on the pooled estimate of

variance from all the groups (which is the residual variance
in the ANOVA table), not just from pair being considered.

III. If we perform k paired comparisons, then we should

multiply the P value obtained from each test by k; that is, we
calculate P' = kP with the restriction that P' cannot exceed 1.
Where, , that is the number of possible comparisons
 Benferroni method or Modified t-test
 Returning to the red cell folate data given above, the residual
standard deviation is = 45.72.

(a) Comparing groups I and II

t = (316.6 - 256.4) / (45.72 x √(1/9 +1/8)
= 2.71 on 19 degrees of freedom.

 The corresponding P-value = 0.014 and the

corrected P value is P' = 0.014x3
= 0.042
Group I and II are different
 Benferroni method or Modified t-test
(b) Comparing groups I and III

 t = (316.6 - 278.0) / (45.72 x √(1/8+1/5)

= 38.6/26.06
= 1.48 on 19 degrees of freedom.

 The corresponding P value = 0.1625 and

 The corrected P value is P' = 0.1625x3

= 0.4875
Group I and III are not different
 Benferroni method or Modified t-test
(c) Comparing Groups II and III

 t = (278 - 256.4) / (45.72 x √(1/5+1/9)

= 21.6/25.5
= 0.85 on 19 degrees of freedom.

 The corresponding P value = 0.425 and the corrected P value

is P' = 1.00

Group I and III are not different

Therefore, the main explanation for the difference between

the groups that was identified in the ANOVA is thus the
difference between groups I and II.
 One way ANOVA’s limitations

 This technique is only applicable when there is one

treatment used.

 Note that this single treatment can have 3, 4,… ,many

levels.

 Thus nutrition trial on children weight gain with 4

different feeding styles could be analyzed this way,
but a trial of BOTH nutrition and mothers health
status could not
Two-way ANOVA (axb Factorial Design)
Two-way ANOVA (axb Factorial Design)

 Suppose we have two treatments „A‟ and „B‟ in an

experimental question.

 Thus, the interest is in determining the combined effect of „A‟

and „B‟ as well as „A‟ considered separately and „B‟
considered separately.

 That means it will have the same results as the separate one-
way ANOVA on each variable, except that here the interaction
effect is also included in the model
Two-way ANOVA (axb Factorial Design)

 Again, here experimental units (subjects) are assigned at

random to each of the axb combinations, and

 The subjects to be assigned are also assumed to be

homogeneous

 „A‟ and „B‟ are the set of treatments and are called factors or
main effects and have „a‟ and „b‟ different treatments
respectively called treatment levels or groups
 Main effects, simple effects, interaction
 Main effect is a difference in population means for a factor
collapsed (or pooled) over the levels of all other factors in the
design

 Thus, a significant main effect demonstrates that an independent

variable influences the dependent variable.

 However, it does not establish which of the independent variable

levels are significantly different from another

 To identify those significantly different levels, we use a priori

(pre-planed) or post hoc tests
 Main effects, simple effects, cont…
 An interaction occurs when the effect of one independent
variable is affected by another independent variable; or

 When the effect of one factor is not the same at the levels of
another

 Or it is when the effect on the dependent variable at different

levels of an independent variable is influenced by the
corresponding levels of another variable.

 E.g. Grade score for females will be greater than males‟ when
tutor is present than absent
 Main effects, simple effects, interaction
cont…
 A simple effect is the effect of one independent variable on
the dependent variable at one particular level of another
independent variable.

 We can test the differences between effects by two levels of

an independent variable at each level of another independent
variable separately
 If the difference is significant ,we can confirm the existence
of simple effects at that particular level of second variable.

 Usually we only look at simple effects if you have found a

significant interaction.
 Two way ANOVA cont…
Calculate total Sum of Squares (for equal cell size =n)
T = ΣiΣjΣkxijk2 S = ΣiΣjTij.2/n CF = T2…/abn
A = ΣiT2i../na B = ΣjT.j.2/nb

Source df SS MS F
Factor A a-1 SSa =A - CF SSa/(a-1) MSa/MSe
Factor B b-1 SSb = B - CF SSb/(b-1) MSb/MSe
AxB (a-1)(b-1) SSab= S-A-B+CF SSab/(a-1)(b-1) MSab/MSe
Error ab(n-1) SSe = T - S SSe/ab(n-1)
Total abn-1 SSt = T - FC

Where “a” and “b”, are the number of groups in treatment A and
B respectively and „n‟ is the repilication
 The two-way Linear model

 The formal model underlying two-Way ANOVA, with 2

treatments A and B is: Xijk = μ + Ai + Bj +ABij + Ɛijk

 Xijk is the kth replicate of treatment A level i and treatment B level j

 Ai is the effect of the ith level of treatment A (= difference

between μ and mean of all data in this treatment).

 Bj is the effect of the jth level of treatment B (= difference

between μ and mean of all data in treatment B).

 Ɛijk is the unexplained error in Observation Xijk

 Two way ANOVA cont…

The null hypotheses in a two-way ANOVA are these:

– The population means for the dependent variable are

equal across levels of the first factor
– The population means for the dependent variable are
equal across levels of the second factor
– The effects of the first and second factors on the
dependent variable are independent of one another
 Two way ANOVA cont….
Handling treatment
(B)
B1 B2
Example (Steel & Torrie p. 343)
8.53 17.53
20.53 21.07
Effect of 2 treatments on blood 12.53 20.80
phospholipids in lambs. The A1 14.0 17.33
10.8 20.07
first was a handling treatment, Time
and the second one was the time treatment
(A) 39.14 32.00
treatment. 26.20 23.80
31.33 28.87
A2 45.80 25.06
40.20 29.33
 Two-Way ANOVA con…
Start by a preliminary eyeballing of the data:

They are continuous, plausibly normally distributed.

There are 2 handling treatments and 2 time treatments,
Both are combined in a factorial design so that each of the 4
combinations is replicated 5 times.

The basic findings are the following:

n = 20
T… = ΣΣΣxijk = 484.92
T = ΣΣΣxijk2 = 13676.7
T…2/N = (ΣΣΣxijk)2/N = (484.92)2 / 20 = 11757.37
SST = 13676.7 – 11757.37= 1919.33
Now get totals for treatments A and B

B1 B2 Ti.
A1 66.39 96.80 163.19
A2 182.67 139.06 321.73
T.j 249.06 235.86 484.92

Hence the sums of squares for A and B can be calculated:

SSa = 163.192/10 + 321.732 / 10 - 11757.37 = 1256.75

SSb = 249.062/10 + 235.862/10 - 11757.37 = 8.712

A alone
Source Df SS MS F
A 1 1256.75 1256.75 34.14**
error 18 662.58 36.81
total 19 1919.33

B alone
Source Df SS MS F
B 1 8.71 8.71 0.08
error 18 1910.62 106.15
total 19 1919.33
Pooled
Source Df SS MS F
A 1 1256.75 1256.75 32.67**
B 1 8.71 8.71 0.24
error 17 653.87 38.86
total 19 1919.33
 Interaction terms
 We now meet a unique, powerful feature of factorial ANOVA.
It can examine data for interactions between treatments -
synergism or antagonism.
 Some treatments intensify each others‟ effects:

 The classic examples come from pharmacology.

 Alcohol alone is lethal (toxic) at the 20-40 unit range.
Barbiturates are lethal. Together they are a vastly more lethal
combination, as the 2 drugs synergize.

 In ecology, SO2 + NO2 is more damaging than the additive

effects of each gas alone - a synergism.
 Antagonism.

 It is the opposite - two treatments nullifying each

other.

 Drought antagonizes effects of air pollution on plants,

as drought leads to closed stomata excluding the
noxious/harmful gas.
 No interaction
Response B2
I
I Treatment B categories
I
I B1
I
I
I
1 2 3 Treatment A
Synergism
Response
Antagonism I
I
I I
I
I I
I I
I I

1 2 3 Treatment A 1 2 3
 Incorporating interaction into the
ANOVA model
 First, work out a sum of squares caused by ALL treatments at
ALL levels.

 Thus for a 2*2 design, there are actually 4 treatments and for
3*3 design there are 9 treatments, etc. Call this SStrt
 Now we can partition this Sum of squares:
SStrt = SSa + SSb + SSi , Where, SSi is for Sum Square
interaction
 We know SSa, we know SSb, so we get SSi by subtraction after
calculating the SStrt using the computational formula

But SStrt = ΣΣxij.2/n - CF

 For the lamb blood data:
 We have 4 separate treatments: A1B1, A1B2, A2B1, A2B2

 The data within these 4 groups add to: 66.39, 182.67,

96.80 and 139.06.

 There are 5 replicates in each treatment

 SStrt = 66.392/5 + 182.672/5 + 96.82/5 + 139.062/5 -

11757.37 = 1539.407
 Two-Way ANOVA table with
interaction

Source df SS MS F
All treat 3 1539.07 …………………………..
A 1 1256.75 1256.75 52.93**
B 1 8.71 8.71 0.37
AxB 1 273.95 273.95 11.54**
Error 16 379.92 379.92
Total 19 1919.33 ………………………
 Example Two Way ANOVA
(4x3 factorial ANOVA)
Twenty four men each weighting 35 kg overweight, are
assigned to the 24 treatments that arise from 4 diets and
3 levels of jogging with replication of 2. Each man
consumes the same amount of food per day, but the
diets differ in their proportion of proteins, fat, and
carbohydrate. The aim of the experiment is to reduce
weight .

(Data in the following slide)

 4x3 Factorial Design

Diet

Equal High High Fat High Total

Protein Carbohyd.
0’ 8.5 15.5 8.5 15.5 97
11.5(20) 16.5(32) 7.5(16) 13.5 (29)
Data
Jogging
14 20 13 21 136
30’ 16(30) 23(43) 11(24) 18 (39)
(in minutes) 24.5 27 22 24.5 196
60’ 19.5(44) 24(51) 27(49) 27.5 (52)

Total 94 126 89 120 429

 Example-Three way ANOVA
(2x2x3 factorial ANOVA)

Maxwell and Delaney (1990) studied the effect of three different

treatments:
1. Drug (a medication having three levels: X, Y and Z)
2. Biofeed (a psychological feedback : present and absent)
3. Diet (special diet as given or not given)
 Seventy-two subjects suffering from hypertension were
recruited to the study with being randomly allocated to each of
the 12 treatment combinations. Blood pressure measurements
were made on each subject after treatments.
(Data in the following slide)
 Data…. which is a 3X2x2 Design
Treatments Special Diet

Biofeedback Drug No Yes

X 170,175,165,180,160,158 161,173,157,152,181,190
.
Present Y 186,194,201,215,219,209 164,166,159,182,187,174

Z 180,187,199,170,204,194 162,184,183,156,180,173

X 173,194,197,190,176,198 164,190,169,164,176,175

Y 189,194,217,206,199,195 171,173,196,169,199,180,203
Absent
Z 202,228,190,206,224,204 205,199,170,160,179,179