COMPARISONS

BETWEEN THREE OR
MORE GROUPS (PART II)
DADANG SUDANA
(SOURCE: HATCH AND FARHADY, 1982; HATCH AND LAZARATON, 1991)
• On the other hand, if the value of between-groups variance is greater than within-group
variance, we are set to go.
• There is some difference between the five groups. Now the problem is discovering
whether they are different enough that the difference is not due to chance.
• So (guess what), we will again place the difference in a sampling distribution to discover
the probability of finding a difference as large as that we obtained.
• The sampling distribution for ANOVA is called the F distribution.
• Fortunately for us, mathematicians have already constructed this distribution (see
appendix C, table 5).
• Like the t distribution, it is made up of families with the same number of degrees of
freedom in the sample size, but, unlike the t distribution, the F distribution also considers
the degrees of freedom for number of groups.
• We don’t have to consider the degrees of freedom for number of groups in the t-test
because in the t- test there are always only two groups and the df for groups is always I.
• To discover whether we can reject the null hypothesis, we once again consult the critical values
given in the table for the sampling distribution.
• In the t-test. we obtained a t-value. In ANOVA, we will obtain an F value. The F value is the ratio
of the two sources of variance--between-groups variance over within-group variance.

• Another way of saying the same thing—

• Thus, we need an F ratio > 1 to show any difference at all among groups.
• How much larger than 1 we need depends on the number of degrees of freedom within
the groups and the degrees of freedom for the number of groups.
• Let's apply this to the example to see how it works. The data from the exam scores of the five different
groups follow.

Exam Scores for Five Methods

Mean 64 60 61 65 57
s2 110 90 95 100 110

n 15 15 15 15 15
df 14 14 14 14 14
Number of groups (K) = 5
Total observations (N) = 75 (15 per group)
• Arithmetically, the means of the groups differ. It appears that the situational-functional approach is
best, while the natural approach is second best. Worst appears to be drama.
• We cannot, however, be confident that they are truly different without taking the n size and variance
into account. Look at the following two visuals.
• In the first figure, there is a lot of variability within groups, but not much between. There
is a lot of overlap. These differences probably are not statistically significant.
• However, in the second diagram, there is little variability within groups and much more
between. There is not much overlap.
• And since statistical significance for ANOVA is determined by the ratio of variability
between to variability within, the second diagram probably shows differences which are
statistically significant.
• In any case, we cannot know for sure without calculatingthe F ratio.
• Remember that s2 is used to represent population variance. This value is estimated by
squaring the standard deviation.
• You know that we must look at the F ratio--the ratio of variance between over the
variance within groups. We won’t do the calculations just yet, but let's assume that we
found a between-groups variance of 264.
• This represents the total variability present between the scores of the five groups. The
within-group variance turned out to be 95. This represents the total variability of scores
within each of the five groups.
• We can, then, compute the F ratio:
• The subscripts B and W stand for between and within. The ratio is larger than l. So, there
is some effect for treatment.
• To discover whether the F value is large enough to give us confidence in rejecting the H0,
turn to the F distribution in table 5, appendix C.
• We know the probability level will depend on the number of df for groups.
• There are five groups (K). The df for groups is K- l = 4. There are 15 Ss in each group.
• To find the df for each group we use n- 1 or 14. There are five groups, so we multiply n- 1
by 5, 14 x 5; the df for within-group variance is 70.
• The numbers across the top of the table are the degrees of freedom for groups. The numbers down
the side of the table are the degrees of freedom for Ss or number of observations.
• For our example, look across the top of the table for the F distribution to column 4 (df for K). Then
go down the column to row 70 (df for the number of observations).
• Notice that there are two sets of numbers in place. The top set are for a critical value where p = .05.
• The lower set is for a critical value where p = .01. We should use the .05 critical values of F
because that's what we specified before the analyses.
• Fcrit for an .05 level of confidence (for df of 4, 70) is 2.50. We can reject the H0 since Fobs is
greater than Fcrit. We can, therefore, claim that scores on the exam differ across the groups.
• In this example, the data are from a relatively small number of Ss. Remember that the
smaller the number of Ss (or observations), the larger the differences required to obtain
statistically significant results.
• Now that we have gone over the basic concept of ANOVA, let's actually compute an F
ratio for a One-way ANOVA working from raw data.