Notes on ANOVA technique

1. ANOVA is short form of Analysis of Variance.

2. This technique is useful in comparing the mean value of a numerical variable across
various classifications. For example,
a. We want to compare sales across various modes of advertising. Let us say
that there are 4 different modes of advertising (newspapers, social media by
people, FM radio channels and display hoardings). Sales values are observed
for each of these modes. We want to know whether any one mode of
advertising leads to significantly higher or significantly lower sales. For this,
we can use ANOVA technique.
b. Similarly, we may want to compare marks across 3 divisions.
c. Three drug dosages are to be compared to see which drug dosage is effective
in reducing blood pressure
d. We want to find which plant locations (out of 8 plant locations) in a
manufacturing company have low productivity so that corrective managerial
action can be taken.

3. In each of the cases above, a numerical variable is compared across the

classifications of a categorical variable.

Sr No Numerical variable Categorical variable Number of classifications

1 Sales Mode of advertising 4 modes of advertising

2 Marks Division of students 3 Divisions

Reduction in blood
3 pressure Dosage of drug 3 levels of dosages of drugs

4 % rejection rate Plant location 8 plant locations

4. Note that each situation above has more than 2 classifications.

5. If there were only two classifications (ie only two divisions of students for comparing
marks) then two sample t test would be the method appropriate for evaluating
whether any one division has higher marks.
6. When 3 divisions are compared, 3 pairs of two divisions get compared. For example,
if there are 3 divisions viz. division A, division B and division C then comparison is
required across A vs B, B vs C and A vs C ie 3 pairs. Hence 3 T tests would be needed.
7. Advantage of ANOVA method is that one ANOVA test can accomplish the same task
done by 3 T tests.
8. ANOVA method basically works on the following logic:
If one classification level is significantly better or worse than the other
classification levels, the variation of the mean value for that level from the
overall sample mean would be far higher than the variation of sample values
within that classification level. To illustrate using the example of division
wise marks:
If Division A students have scored significantly more marks than other
divisions then sum of variation between the marks for each student of
Division A and overall average marks across all divisions would be far
higher than the sum of variation of marks within the Division A.

9. Post Hoc tests

a. ANOVA method simply shows whether there is any one classification level
that is significantly better or worse than the other levels. But it does not
identify that classification level.
b. Post Hoc tests are carried out to identify such classification levels.
c. There are two basic methods:
i. Compare all possible pairs
ii. Compare all other levels with one control level
d. Such comparisons , if unadjusted; result in loss of significance
i. To illustrate
1. When two samples are compared with 5% significance, there
is probability of 5% or less that there are different. Hence
there is probability of 95% or more that the two samples are
not different. Let us call this test for two samples as the test
for one pair of samples.
2. Now if marks for 3 divisions are compared; totally 3
comparisons are made (A with B, A with C and B with C; as
mentioned earlier). ie 3 pairs are compared. 3 ‘T’ tests are
carried out.
3. Probability that atleast one pair is different = 1 – probability
that no pair is different
4. Since 3 pairs are compared, probability that none of the 3
pairs is different = 0.95^3
5. Hence probability that atleast 1 pair is different = 1- 0.95^3 =
14.26%
 6. We can tolerate maximum of 5% for significance across the 3
divisions. 14.26% would not be acceptable.
7. Hence for each pair the acceptable significance level needs to
be adjusted below 5% so that at the aggregate level, 5%
significance is achieved. (If significance for each pair is reduced
from 5% to 1.69% then the probability that a pair is not
different is 1-1.69% ie 98.31%. Probability that none of the 3
pairs is different is .9831^3. This is same as 0.95. Hence
probability that atlease one pair is different is 1 – 0.95 = .05.
This is the desired significance at the aggregate level.)
8. This adjustment is automatically done by the SAS procedure by
selecting
a. Bonferoni or Tuckey adjustment/correction for
pairwise comparison
b. Dunnet adjustment/correction for comparison with
control level
10. ANOVA method assumes that
a. Sample is normally distributed with observations independent of each other
i. This is tested using Shapiro-Wilkinson test or Jacques Barra test for
normality
b. Variance within each classification level is same
i. This is tested using Levene’s test for squared residuals.