Can it compare more than two groups directly?

This specific two-sample ratio test does not; other methods address multiple groups.

Does p > 0.05 prove equality?

No. The data may simply be insufficiently informative.

AI Glossary / Statistics and probability / Intermediate

F-test of equality of variances

Definition

The F-test of equality of variances is a classical hypothesis test comparing the population variances underlying two independent normal samples. It uses the ratio of their sample variances and an F reference distribution under the null hypothesis of equal population variances. The test concerns variability, not equality of population means, and its exact calibration is sensitive to departures from normality.

What it means

Two processes can have the same average output but different consistency. A variance comparison asks whether the spread of outcomes differs. Conversely, processes can have different means but equal variances.
The F-test turns a sample-variance ratio into a probability statement under a specified null model. A ratio away from one may arise through sampling variation, especially with small samples. The p-value helps assess that variation but does not measure the business importance of a difference.

How it works

Define the two populations and collect samples whose observations are independent within and between groups. Use the same measurement units and comparable conditions. Paired observations, repeated measurements, and time-dependent records generally require a different analysis or additional modeling.
Compute each sample variance with denominator n − 1. For groups one and two, form F = s_1²/s_2². Under independent normal sampling and H_0: σ_1² = σ_2², the statistic follows F(n_1 − 1,n_2 − 1).
Choose the alternative before inspecting the data. An upper-tail test addresses σ_1² > σ_2²; a lower-tail test addresses the reverse; a two-sided test addresses either difference. Keep group ordering explicit so that the statistic, degrees of freedom, and direction match.
For a conventional equal-tail two-sided test, reject outside the lower α/2 and upper 1 − α/2 cumulative quantiles. Equivalently, the p-value is twice the smaller tail probability, capped at one. Some tables label quantiles by upper-tail area, so check their convention.
Report the estimated variance ratio and a confidence interval. Under the same normal model, a two-sided interval for σ_1²/σ_2² divides the observed ratio by the upper and lower F quantiles, respectively. Taking square roots gives an interval for the standard-deviation ratio.
Inspect whether normality is credible. Heavy tails and outliers can strongly affect sample variances and the F test. Methods such as median-based Levene or Brown - Forsythe procedures may be more suitable for some nonnormal data, but they also have assumptions and finite-sample limitations.
Do not use failure to reject equality as proof that variances are practically equivalent. If practical similarity is the goal, define a meaningful tolerance and use a method designed for that question. Avoid using a preliminary variance test as an automatic switch between later analyses without considering the combined procedure.

Three practical examples

A manufacturer compares the consistency of two stable measurement processes when independent normal samples are plausible.
A laboratory compares repeatability variances from two instruments using compatible units and sampling conditions.
An analyst demonstrates how a large sample-variance ratio can still be uncertain when each sample is very small.

Think of it this way

Think of comparing the spread of two dart patterns rather than their centers. The test evaluates how much evidence there is for different spread under a particular model, not whether either pattern hits the desired target.

Worked example

Take invented samples A = [0, 2, 4] and B = [1, 2, 3]. Both means are 2. For A, the sum of squared deviations is 8, so s_A² = 8/2 = 4. For B, it is 2, so s_B² = 2/2 = 1.
The variance ratio is F = 4, with numerator and denominator degrees of freedom both equal to two. For F(2,2), the cumulative probability at x is x/(1 + x), giving a lower-tail probability of 0.8 and an upper-tail probability of 0.2 at four.
The equal-tail two-sided p-value is 2 × 0.2 = 0.4. A prespecified upper-tail test would instead have p = 0.2. Neither rejects at 5%.
These tiny samples do not demonstrate equal population variance. They illustrate the calculation conditional on independent normal sampling, an assumption that cannot be established from three observations per group.

Illustrative case study

Fictional case: Crest Instruments obtains a nonsignificant variance test comparing two devices and initially plans to call them equally precise. A reviewer asks for a confidence interval and a practical tolerance for acceptable precision differences. The team recognizes that its small study is inconclusive and plans a more informative repeatability assessment. This invented case distinguishes absence of evidence from evidence of equivalence.

Common misconceptions and mistakes

Do not confuse the variance F-test with an ANOVA test of group means. The same reference-distribution family can support different hypotheses.
Do not automatically put the larger variance in the numerator and then use an unadjusted one-sided p-value. Data-driven ordering changes the procedure.
The classical test's normality requirement matters; it is not generally a robust default for arbitrary skewed measurements.

Business use

Variance comparisons can support process consistency and measurement-quality decisions. Practical reporting should identify the variance or standard-deviation ratio, its uncertainty, and the consequences of a plausible difference. A statistical result should be considered alongside outliers, measurement resolution, operating conditions, and whether the sampling reflects routine use.

Check your understanding

What is the ratio in the example?

Answer: Four for variances, corresponding to two for sample standard deviations.

Why are both degrees of freedom two?

Answer: Each sample has three observations.

What makes the p-value exact under the model?

Answer: Independent normal-sample variance theory supplies the F distribution.

Related terms

Learn next

Study the chi-square distribution of normal-sample variance. Then compare variance tests with robust alternatives and learn confidence intervals for variance ratios.

Sample variance

Sources

NIST/SEMATECH, “F-Test for Equality of Two Variances”: https://www.itl.nist.gov/div898/handbook/eda/section3/eda359.htm R stats, var.test: https://stat.ethz.ch/R-manual/R-devel/library/stats/html/var.test.html NIST/SEMATECH, “Levene Test”: https://www.itl.nist.gov/div898/handbook/eda/section3/eda35a.htm

Previous: F-measure Next: Fabric computing