MerQuR+Statistics Glossary

Statistics Glossary → Homogeneity of variance assumption

Homogeneity of variance assumptionconcept

The assumption that the groups being compared have equal population variances (spreads); the classical t-test and ANOVA rely on it.

The Student t-test and classical ANOVA pool the group variances into a single error variance. This is valid only if the groups are about equally spread. Its counterpart in regression is the constant-variance (homoscedasticity) assumption. It is assessed with the Levene, Brown-Forsythe or Bartlett tests, but more usefully by looking at the group standard deviations and at residual plots.

The impact of a violation depends on the group sizes. With equal group sizes the classical tests are fairly robust to moderate differences. With unequal sizes, the type I error rate is inflated when the larger variance is in the smaller group, and the test becomes overly conservative when it is in the larger group. The usual remedies are the Welch correction, the Games-Howell post-hoc test or robust standard errors. Choosing the method based on a preliminary variance test can itself distort error rates, which is why many authors recommend using Welch directly.

An everyday example

The weights of loaves from three bakeries are compared. The first two bakeries produce loaves of almost exactly the same weight, while the third bakery's loaves can be 20 g lighter or heavier. If this difference in spread is ignored when comparing the means, the ANOVA p value may be misleading.

In MerQur

The independent-samples t-test has the 'Variance (Auto (Levene), Equal (classic t), Welch (unequal))' option and the Levene, Bartlett and folded F assumption tests. In one-way ANOVA, the 'Assumption checks' test homogeneity of variance with Levene's and Bartlett's tests, while the Welch variant and 'Extra output: Brown-Forsythe F*' do not assume equal variances. Two-way ANOVA gives Levene and Brown-Forsythe tests across cells, mixed-design ANOVA gives Levene's and Box's M tests, and the nested mixed model gives Levene and Brown-Forsythe tests under 'Residual normality and homogeneity of variance'. In the Mann-Whitney U test, 'Extra outputs: Brunner-Munzel' adds a test that does not require equal variances.

Analyses where this concept appears

Independent-Samples t-TestOne-Way ANOVATwo-Way ANOVAMixed-Design ANOVA, Split-PlotMann-Whitney UNested Mixed Model (Nested LMM, Multi-Level Hierarchical Mixed Model)

Related concepts

Levene's testWelch correctionBartlett's test of equal variances