Must observed sample variances match exactly?

No; sampling variability makes exact equality unusual even when population variances are equal.

AI Glossary / Statistics and probability / Intermediate

Homogeneity of variance

Definition

Homogeneity of variance is the condition that the populations or model-error distributions being compared have the same variance. It is an assumption in familiar procedures such as the pooled two-sample t-test and classical one-way analysis of variance. The condition concerns underlying variability, not a requirement that observed sample variances be numerically identical.

What it means

Groups can have different average outcomes while being equally variable around their own means. Homogeneity of variance says their spreads, measured by variance on the chosen scale, are the same.
This can justify pooling information to estimate a common variance. However, pooling is a modeling decision. Similar-looking sample spreads do not prove the assumption, and a formal test that fails to reject equality does not establish it as a fact.

How it works

Specify which variance is relevant. In a simple comparison of independent groups, the condition is sigma1² = sigma2² = ... = sigmak². In a model that adjusts for covariates, attention usually belongs to residual or conditional error variances rather than raw outcome variances.
Examine group-specific observations, sample sizes, standard deviations, and plots. Variances use squared outcome units; standard deviations use the original units. A group with twice the standard deviation has four times the variance, so the two ratios should not be confused.
Formal checks include Levene-type tests and Bartlett's test. Levene's method compares absolute deviations from group centers, with median-based variants often used for robustness. Bartlett's test is sensitive to nonnormal distributions, which can make its result reflect more than the intended variance comparison.
Interpret any test alongside the design and effect magnitude. Small samples may have little power to detect consequential heterogeneity; large samples may identify small differences that barely affect the intended analysis. A p-value alone is not a complete assessment of practical suitability.
For a pooled two-group mean comparison, estimate the common variance by weighting each sample variance by its residual degrees of freedom: s_p² = [(n1 - 1)s1² + (n2 - 1)s2²]/(n1 + n2 - 2). This differs from simply averaging the two variances when sample sizes differ.
The standard error of the mean difference is s_p*sqrt(1/n1 + 1/n2). Exact classical t inference additionally relies on appropriate independent normal sampling assumptions. Equal variances alone do not provide every condition needed for the test.
If a common variance is not justified, methods such as Welch's comparison of independent means or suitable heteroskedasticity-robust model inference may be preferable. These methods also have assumptions; replacing one procedure is not a substitute for understanding the data.
A transformation can change variance patterns and interpretation. For example, relative variability may be more stable than absolute variability. Choose a scale that answers the substantive question, and report it rather than transforming data only to obtain a desired diagnostic result.

Three practical examples

A training analyst considers whether score variability is sufficiently comparable across groups for a pooled analysis.
An AI evaluator checks residual spread across task categories before using one common error-variance estimate.
A quality team compares consistency across production lines separately from differences in their average output.

Think of it this way

Think of two groups of darts with equally wide scatter around different centers. Their average positions differ, but the amount of spread can be the same. Homogeneity of variance concerns the scatter, not whether the centers coincide.

Worked example

First consider illustrative values A = (1, 3, 5) and B = (11, 13, 15). Their means are 3 and 13. Each group has deviations -2, 0, and 2 from its own mean, so each sample variance is (4 + 0 + 4)/(3 - 1) = 4.
Adding ten to every observation changes the mean but leaves variance unchanged. This demonstrates the distinction between location and spread; it does not prove that the populations generating two tiny samples have equal variances.
For a separate pooled-variance calculation, suppose independent samples have n1 = 4, n2 = 6, and both sample variances equal 9. The pooled estimate is [(4 - 1) x 9 + (6 - 1) x 9]/(4 + 6 - 2) = 72/8 = 9.
The estimated common standard deviation is 3. The standard error of the mean difference is 3*sqrt(1/4 + 1/6) = sqrt(3.75), approximately 1.93649 outcome units. Under the usual equal-variance normal model, the comparison uses eight residual degrees of freedom.
This calculation shows what the assumption enables. It is not a test that the assumption holds, and no mean difference or significance conclusion has been supplied.

Illustrative case study

Fictional case: Willow Training compares assessment outcomes from two course formats. Their sample standard deviations are similar, but the analyst checks plots and discovers that one format includes a broader range of baseline skill.
After fitting a justified baseline-adjusted model, the team examines residual spread rather than relying solely on the raw score summaries. It compares conclusions from a common-variance analysis with an appropriate alternative and reports any material sensitivity.
Management receives estimates of the outcome difference together with a clear description of variability. A nonsignificant variance test is described as limited evidence, not certification that the groups are statistically identical.

Common misconceptions and mistakes

Common mistakes include confusing equal variance with equal mean, equating equal sample variances with equal population variances, and assuming equal group sizes guarantee equal variability.
Do not interpret failure to reject a variance test as proof of equality. Also avoid deleting valid observations solely to obtain a common spread or treating a variance diagnostic as evidence for causal comparability.

Business use

Homogeneity of variance matters when a business analysis combines groups under one uncertainty model. It can simplify estimation when justified, but operational consistency should also be examined directly. Reports should distinguish mean performance from spread and explain how the chosen analysis handles meaningful differences in variability.

Check your understanding

Why do A and B have equal sample variance?

Answer: Their deviations from their own means are identical.

What is the pooled variance in the second example?

Answer: Nine.

Does that estimate prove the population variances are equal?

Answer: No; equality is an assumption being used in the calculation.

Related terms

Learn next

Compare group means and spreads separately, then work through pooled and unequal-variance standard errors. Study how residual plots differ from raw outcome plots after adding covariates.

Sources

David Howell, glossary: https://www.uvm.edu/~statdhtx/methods8/Glossary/Glossary.html NIST, Levene test: https://www.itl.nist.gov/div898/handbook/eda/section3/eda35a.htm NIST, two-sample t-test: https://www.itl.nist.gov/div898/handbook/eda/section3/eda353.htm NIST, Bartlett test: https://www.itl.nist.gov/div898/handbook/eda/section3/eda357.htm

Previous: Homogeneity of regression Next: Homogeneity of variance in arrays