Why Variance Tests Use Chi-Square or F Critical Values

The common assumption is that a single sampling distribution covers all hypothesis tests. That assumption is wrong. Variance tests require chi-square or F critical values because the question being asked, whether you are testing one variance or comparing two variances, determines the relevant sampling distribution. When you measure a single sample variance against a hypothesized population variance, the test statistic follows a chi-square distribution. When you compare two sample variances, the ratio follows an F-distribution. These distributions are asymmetric, non-negative, and their critical values come from the right tail for hypothesis tests, not the left. Using a Z or t critical value for a variance test would be a category error: those distributions describe means, not squared deviations from the mean.

How does the question about one or two variances determine the relevant sampling distribution? The answer is built into the formula for the test statistic. For a single variance test, you compute χ² = (n−1)s² / σ₀². This statistic is a sum of squared standard normal variables, which by definition follows a chi-square distribution with df = n−1. The chi-square distribution is the exact sampling distribution for that calculation. There is no approximation, no Central Limit Theorem invocation. The shape is fixed by the degrees of freedom, for n = 20, the df is 19, and the distribution is heavily skewed to the right.

For a two-variance test, you compute F = s₁² / s₂². This ratio of two chi-square variables, each divided by its degrees of freedom, follows an F-distribution with numerator df = n₁−1 and denominator df = n₂−1. The F-distribution is the exact sampling distribution for that ratio. Again, no approximation. The question, one variance or two, maps directly to chi-square or F, not to any other distribution. The mapping is not a convention; it is a mathematical consequence of how the test statistic is constructed.

A concrete example makes the mapping clear. Suppose you want to test whether the variance of a manufacturing process exceeds a target value. You take a sample of n = 20 units and measure a sample variance s² = 12. The hypothesized population variance is σ₀² = 8. You choose α = 0.05 for a one-tailed right test because you only care if the variance is too large. The test statistic is χ² = (n−1)s² / σ₀² = (19)(12) / 8 = 28.5. The critical value is the 0.95 quantile of the chi-square distribution with 19 degrees of freedom, which is χ²_{0.95,19} = 30.14. Since 28.5 is less than 30.14, you fail to reject the null hypothesis. The process variance is not statistically significantly larger than 8 at the 0.05 level.

Now change the question to a two-variance test. You have two independent samples: n₁ = 10 with s₁² = 15, and n₂ = 12 with s₂² = 10. You want to test whether the variances are equal, using a two-tailed test at α = 0.05. The test statistic is F = s₁² / s₂² = 15 / 10 = 1.5. The degrees of freedom are df₁ = 9 and df₂ = 11. The critical values are the 0.025 and 0.975 quantiles of the F-distribution: F_{0.025,9,11} = 0.248 and F_{0.975,9,11} = 3.59. Since 1.5 falls between 0.248 and 3.59, you fail to reject the null hypothesis. The variances are not significantly different. In both examples, the sampling distribution is determined by the number of variances being tested, one or two, and the critical values come from the corresponding distribution.

To perform the calculation yourself, you need the correct inverse function. For a chi-square critical value, use CHISQ.INV.RT in Excel or Google Sheets, which takes the right-tail probability and degrees of freedom. For an F critical value, use F.INV.RT, which takes the right-tail probability and both numerator and denominator degrees of freedom. The TI-84 Plus CE does not have native chi-square or F inverse functions, so you must use the chi-square cdf and solve manually or switch to spreadsheet software. At the point where you need the next calculation, you enter the sample variances and degrees of freedom into a variance calculator that returns the test statistic, then compare it to the critical value from the appropriate distribution.

The chi-square distribution is asymmetric and non-negative. For a one-tailed right test, the critical value is the (1−α) quantile. For a two-tailed F-test, you need both the lower quantile at α/2 and the upper quantile at 1−α/2. The lower F critical value is near zero, for df₁=9 and df₂=11 at α=0.05, it is 0.248. The upper is 3.59. These are not symmetric around 1.0 because the F-distribution is skewed. Using the wrong tail is the most common error: a left-tail chi-square critical value for df=19 at α=0.05 is 10.12, while the correct right-tail value is 30.14. Substituting 10.12 for 30.14 would cause you to reject the null when you should not, inflating the Type I error rate.

The reason boils down to the nature of variance as a measured object. Variance is a squared deviation, so its sampling distribution is not normal. The chi-square distribution arises from the sum of squared standard normal variables, which is exactly what the single-variance test statistic is. The F-distribution arises from the ratio of two independent chi-square variables, which is exactly what the two-variance test statistic is. These are not arbitrary choices, they are the exact sampling distributions that describe the behavior of the test statistics under the null hypothesis.

The degrees of freedom tell you which specific distribution to use. For a single variance, df = n−1. For an F-test, df₁ = n₁−1 and df₂ = n₂−1. The critical value at a given alpha changes with degrees of freedom. For a chi-square test with df=5 at α=0.05, the critical value is 11.07. For df=19, it is 30.14. For an F-test with df₁=3 and df₂=20 at α=0.05, the critical value is 3.098. The number itself has no intrinsic goodness, it is simply the threshold. A smaller alpha makes the critical value larger, making it harder to reject the null. The choice of alpha is a trade-off between Type I and Type II error, and it is a disciplinary convention, not a mathematical requirement. The ASA 2016 statement clarifies that critical-value-based decisions are not a substitute for effect size or practical importance.

The most common failure case is using a left-tail chi-square critical value instead of a right-tail value for a one-tailed hypothesis test. For df=19 at α=0.05, the left-tail critical value is 10.12. The correct right-tail value is 30.14. If you mistakenly use 10.12, every test statistic above that threshold, which includes most of the sampling distribution, leads to a rejection. The Type I error rate becomes roughly 0.95 instead of 0.05. This error occurs because some software defaults to left-tail quantiles for chi-square, and users who do not verify the tail direction accept the output without question.

A second limit: the F-test assumes independent samples and normal distributions. If the populations are not normal, the sampling distribution of the ratio of variances is not an F-distribution. The critical values from the F-table are no longer valid, and the test's Type I error rate can deviate substantially from the nominal alpha. Non-parametric tests like Levene's test or Brown-Forsythe test are alternatives when normality is violated. The NIST/SEMATECH e-Handbook, Section 1.3.5.3, states that the chi-square distribution is used for tests on a single variance and the F-distribution for tests on two variances, but it also warns that these tests are sensitive to departures from normality.

The single most practical thing you can do is verify the tail direction every time you look up a chi-square or F critical value. For a hypothesis test, you need the right-tail quantile. For a confidence interval on a variance, you need the left-tail quantile. These are not interchangeable, and using the wrong one produces the wrong decision. Open the NIST/SEMATECH e-Handbook or use CHISQ.INV.RT and F.INV.RT in your spreadsheet. Confirm that the critical value you have matches the tail you need. That verification takes ten seconds and prevents the error that most commonly destroys variance tests.