Statistics › Section M · The F distribution and ANOVA › Topic M.1
One-way ANOVA and the F ratio
Two groups were compared with t tests; three or more demand a bigger instrument. Analysis of variance asks whether several group means differ by weighing variation BETWEEN the groups against variation WITHIN them - one F ratio, one verdict.
Three means, one question
Do three teaching methods produce different average scores? Does output vary across four machines? The t-test chapter compared two groups, but comparing three or more means pairwise means multiple tests - and multiple tests multiply false alarms (three independent comparisons at \(\alpha = 0.05\) give a collective false-alarm chance of \(1 - 0.95^3 \approx 14\%\)). Analysis of variance (ANOVA) solves the problem with ONE test: are ALL these means drawn from one population, or does at least one group stand apart?
The idea: between versus within
ANOVA's logic weighs two sources of variation. BETWEEN-group variation: how far the group means sit from each other - the signal, if real differences exist. WITHIN-group variation: how far individual scores scatter around their own group means - the noise, ordinary sampling wobble. If the groups share one true mean, the between-group variation should be no larger than chance explains; if a real difference exists, the means spread wider than the noise can account for. The F ratio measures exactly that comparison.
The F ratio
Both variations are first averaged over their degrees of freedom, giving MEAN SQUARES:
where \(MS_{\text{between}}\) divides the between-group sum of squares by \(k - 1\) (\(k\) groups), and \(MS_{\text{within}}\) pools each group's squared deviations, divided by \(n - k\). \(F \approx 1\): between-group spread looks like ordinary noise - no evidence of differences. Large \(F\): the means sit apart by more than the noise explains - evidence that not all means are equal. The ratio is never negative, and the verdict reads from the right tail of the F distribution (next topic's subject) with \(k-1\) and \(n-k\) degrees of freedom.
Computing the pieces
1. Grand mean: the average of all \(n\) scores (the mean of the \(k\) group means when the groups are equal-sized). 2. Between-group sum of squares: each group mean's squared distance from the grand mean, weighted by the group size: \(SS_{\text{between}} = \sum n_j(\bar{x}_j - \bar{x})^2\). 3. Within-group sum of squares: the pooled squared deviations of scores about their own group means - with equal SDs across groups, roughly \((n-k)\) times the common variance. 4. Divide each by its df, take the ratio, and the F is done.
Worked example: three teaching methods
Test scores, \(n = 10\) students per method: Method A \(\bar{x} = 75\), B \(\bar{x} = 79\), C \(\bar{x} = 82\), each with \(s = 5\). 1. Grand mean: \((75 + 79 + 82)/3 = 78.67\). 2. \(SS_{\text{between}} = 10[(75-78.67)^2 + (79-78.67)^2 + (82-78.67)^2] = 10(13.44 + 0.11 + 11.11) \approx 246.7\); \(MS_{\text{between}} = 246.7/2 \approx 123.3\). 3. \(MS_{\text{within}}\): pooled variance \(\approx 25\) (each group's \(s^2\)). 4. \(F = 123.3/25 \approx 4.93\) with \((2, 27)\) df. 5. The F distribution's right tail gives \(p \approx 0.015 \le 0.05\): reject - convincing evidence that the three methods' true means are not all equal.
Worked example: reading the methods' verdict
The methods' \(F \approx 4.93\) at \((2, 27)\) df gives \(p \approx 0.015\), and the verdict is deliberately vague: NOT ALL three means are equal - ANOVA weighed the whole collection at once. Which methods differ is follow-up work: pairwise comparisons run with the multiple-testing discipline of the effect-size topic, since each extra comparison inflates the false-alarm rate again. Had the scores scattered about twice as widely (\(s \approx 10\)), \(F\) would fall to about 1.2 with \(p \approx 0.32\), and the verdict would read: the spread of means (still 75, 79, 82) is within noise's reach at these sample sizes; no evidence of difference, not proof of equality. The verdict, as always, is about population means, not these samples.
The ANOVA table
Software reports the whole computation in one table: source (between, within, total), sum of squares, degrees of freedom, mean squares, the F ratio, and the p-value. Reading it is reading the topic: between-group df = \(k - 1\), within-group df = \(n - k\), the total SS splits exactly into the two parts, and F sits in the mean-squares' ratio. The table is worth learning because research reports, software output, and the coming F-distribution topic all speak it.
Conditions
Three conditions license the F test. The samples are random and independent - no student in two methods, no shared units across groups. Each group's population is roughly normal - moderately robust at decent sizes, the t-test's usual grading. And the groups share roughly equal population spreads (the strongest condition): the pooled \(MS_{\text{within}}\) assumes one common variance, and badly unequal spreads make the F verdict unreliable - check the group SDs, and if the largest is much past twice the smallest, consider alternatives or a transformation.
Common mistakes to avoid
First, running all pairwise t-tests instead of one ANOVA - the false-alarm inflation ANOVA was invented to prevent. Second, reading a significant F as "every mean differs from every other": at least one difference exists, and follow-ups locate it. Third, wrong degrees of freedom: \(k - 1\) and \(n - k\), numerator and denominator respectively. Fourth, ignoring unequal group spreads - the condition ANOVA is most sensitive to. Fifth, expecting F near 1 to mean "all equal": 1 means between-matches-within in this sample; the p-value prices the doubt. Sixth, quoting F without its degrees of freedom - the same number means different things on different F curves.
Key takeaways
One-way ANOVA tests whether \(k\) group means are all equal, with one F ratio instead of many pairwise tests: \(F = MS_{\text{between}}/MS_{\text{within}}\) on \((k-1, n-k)\) degrees of freedom, verdict from the F distribution's right tail. Between-group variation measures the signal (means apart), within-group the noise (scores scattering about their own means); \(F \approx 1\) means signal no louder than noise. The methods example: grand mean 78.67, \(SS_{\text{between}} \approx 246.7\), \(MS_{\text{between}} \approx 123.3\), \(MS_{\text{within}} = 25\), \(F \approx 4.93\), \(p \approx 0.015\) - the methods differ, with follow-up comparisons next. Conditions: random independent samples, roughly normal groups, roughly equal spreads.
Flashcards
Flashcard
One test for equality of k group means
Test your understanding
Question 1 of 5
Why use one ANOVA instead of three pairwise t tests for three group means?
Question 2 of 5
The F ratio compares:
Question 3 of 5
With k = 3 groups and n = 30 total, the F ratio's degrees of freedom are:
Question 4 of 5
For the methods data (MS_between ≈ 123.3, MS_within = 25), the F ratio is about:
Question 5 of 5
A significant ANOVA F means:
