Statistics › Section M · The F distribution and ANOVA › Topic M.3
Test of two variances
Which machine fills more consistently? Does this group's spread match that one's? The F distribution's second job compares two variances directly - a simple ratio test, and the tool that audits ANOVA's own equal-spreads assumption.
Comparing spreads
Some questions are about spread, compared BETWEEN two groups: which production line fills bottles more consistently, does a new process reduce measurement variability, do two teaching methods produce equally variable scores. The null takes the familiar form \(H_0: \sigma_1^2 = \sigma_2^2\), and the test statistic is the F family's purest expression - a ratio of the two sample variances.
The statistic
The test statistic divides one sample variance by the other:
with degrees of freedom \((n_1 - 1, n_2 - 1)\) - numerator sample first. Under the null, the two variances estimate the same quantity and the ratio wanders near 1; a ratio far above 1 says group 1 is more variable, far below 1 the reverse. For the two-sided question - "do the spreads differ?" - the convention puts the LARGER sample variance on top, so the statistic reads from the single right tail and the reported p-value is doubled.
Worked example: two filling machines
Machine A fills \(n_1 = 25\) bottles with \(s_1 = 2.4\) ml; machine B fills \(n_2 = 21\) with \(s_2 = 1.4\) ml. Do the machines differ in consistency? 1. \(H_0: \sigma_1^2 = \sigma_2^2\); \(H_a: \sigma_1^2 \ne \sigma_2^2\) - two-sided. 2. Larger variance on top: \(F = \frac{2.4^2}{1.4^2} = \frac{5.76}{1.96} \approx 2.94\) with \((24, 20)\) df. 3. The F(24, 20) curve's mean is \(20/18 \approx 1.11\); 2.94 sits well out the tail - software gives the upper-tail area as about 0.009, doubled to \(p \approx 0.018\). 4. \(p \le 0.05\): reject - convincing evidence the machines' fill spreads differ, with A the noisier by roughly 70% (2.4 against 1.4).
Worked example: auditing ANOVA's assumption
The two-variance test's in-course job is checking an earlier topic's condition. ANOVA assumed roughly equal group spreads, judged there by a rule of thumb (largest SD not much past twice the smallest). The F test prices that judgement. The teaching methods' sample SDs were all 5, so comparing any two gives \(F = 25/25 = 1\) at the centre of its curve - no evidence against the assumption. Contrast a different study with two groups of 16 each, \(s_1 = 8\) against \(s_2 = 4\): \(F = 64/16 = 4\) at \((15, 15)\), two-sided \(p \approx 0.011\) - the imbalance is formal, not a judgement call, and the analysis needs a transformation or a variance-friendly alternative. The auditor's test, applied to the auditor's own condition.
One-sided variance claims
Some questions name a direction: "is the new process LESS variable than the old?" - a one-sided claim, and one where the larger-on-top convention must be set aside. Put the hypothesised-larger variance on top deliberately: for \(H_a: \sigma_1^2 > \sigma_2^2\), the statistic is \(s_1^2/s_2^2\) however it comes out, and the p-value is the single right-tail area beyond it, undoubled. If the data point the wrong way - the ratio lands below 1 - the p-value is large and the verdict is "no evidence for the claim", which is exactly what a one-sided promise owes. The direction is chosen with the hypothesis, before the data, as every one-sided test in this course has demanded.
Conditions
The F test of two variances demands what the single-variance chi-square test demanded, twice over: BOTH populations roughly normal - and, like its one-sample cousin, this test is highly sensitive to departures from normality, because squared deviations amplify every long tail, now in two samples at once. Independent random samples, and a plot of each group before trusting any p-value: visibly skewed or heavy-tailed data void the arithmetic at any sample size, and alternatives (Levene-style tests, transformations, or rank methods) exist precisely because this test is so easily misled.
Reading the ratio in context
The verdict lands best back in original units: rejecting equality with \(s_1 = 2.4\) and \(s_2 = 1.4\) ml says machine A's typical fill deviation is about 1.7 times machine B's - a ratio a process engineer can act on, since B's consistency is what to copy. Variance ratios themselves translate awkwardly (\(F \approx 2.94\) in squared units); the SD ratio, \(\sqrt{F} \approx 1.7\), is the interpretable version. As everywhere, the test says the difference is real; the SD ratio says what it is worth.
Common mistakes to avoid
First, putting the SDs in the ratio instead of the variances - square first, always. Second, forgetting to order the samples: the df pair must match whichever variance went on top. Third, reading a one-sided tail as a two-sided p-value - double it (or say which side was intended, in advance). Fourth, trusting the test on non-normal data - one of the most fragile assumptions in this course. Fifth, quoting F without its df pair - the family's standing rule. Sixth, concluding the more variable group has the larger mean: spread says nothing about centre, and the t-test chapter owns that question.
Key takeaways
Two-variance questions run through \(F = s_1^2/s_2^2\) on \((n_1 - 1, n_2 - 1)\) degrees of freedom, with the larger sample variance on top for two-sided reading and the tail area doubled. The machines: \(F = 5.76/1.96 \approx 2.94\) at (24, 20), \(p \approx 0.018\) - the spreads differ, A noisier, SD ratio \(\sqrt{F} \approx 1.7\). The test's in-course job is auditing ANOVA's equal-spreads condition; its standing weakness is fragility to non-normality, shared with the single-variance test and now present in both samples. Variances in the formula, SDs in the interpretation - and spread never speaks about centre.
Flashcards
Flashcard
The two-variance test statistic
Test your understanding
Question 1 of 5
With s1 = 2.4 (n1 = 25) and s2 = 1.4 (n2 = 21), the two-variance F statistic is:
Question 2 of 5
For that ratio, the degrees of freedom are:
Question 3 of 5
For a TWO-SIDED test, the convention is:
Question 4 of 5
The test gives upper-tail 0.009. The two-sided p-value is about:
Question 5 of 5
Which in-course assumption does the two-variance test audit directly?
