Statistics Glossary → Bartlett's test of equal variances
Bartlett's test of equal variancesconcept
A chi-square test of whether two or more groups have equal variances, assuming the data are normally distributed.
Bartlett's test compares the logarithms of the group variances with the logarithm of the pooled variance. The more the variances differ, the larger the test statistic; under the null hypothesis it approximately follows a chi-square distribution with k - 1 degrees of freedom (k being the number of groups). The null hypothesis is 'all group variances are equal'.
If the data really are normal, Bartlett's test is quite powerful at detecting differences in variance. On the other hand it is very sensitive to departures from normality: with heavy-tailed or skewed distributions it often gives significant results even when the variances are equal. When normality is doubtful, Levene's or the Brown-Forsythe test is therefore preferred. Its multivariate counterpart is Box's M test for equal covariance matrices; Bartlett's name is also attached to the sphericity test used in factor analysis and PCA, but that answers a different question.
An everyday example
Three bakeries make bread to the same recipe, and 20 loaves from each are weighed. Even if the average weights are similar, one bakery's loaves may be much more variable. Bartlett's test checks whether the three bakeries' weight variances are equal; it is a good choice if the weights are normal, but it may raise a false alarm if a few very small loaves distort the distribution.
In MerQur
In the Independent t-Test, the 'Bartlett' checkbox runs the chi-square test of equal variances, which is sensitive to the normality assumption. In One-Way ANOVA, the 'Assumption checks: Levene, Bartlett, Shapiro' option gives Bartlett's K² for homogeneity of variance. In Canonical Correlation, Bartlett's name refers not to variance homogeneity but to the chi-square approximation of the sequential Wilks Λ tests.
Analyses where this concept appears
Independent-Samples t-TestOne-Way ANOVACanonical Correlation