MerQuR+Statistics Glossary

Statistics Glossary → Box's M test (equality of covariance matrices)

Box's M test (equality of covariance matrices)concept

A multivariate test of whether the covariance matrices of several groups are equal.

MANOVA and linear discriminant analysis assume that the groups share not only their variances but also the covariances between variables. Box's M compares each group's covariance matrix with the pooled covariance matrix through their determinants; the statistic is evaluated approximately with a chi-square or F distribution. It is the multivariate counterpart of the univariate tests of equal variances.

The test is very sensitive to non-normality and in large samples flags even small differences, so it is often read at a strict level such as α = 0.001. With equal group sizes MANOVA is fairly robust to violating this assumption, and Pillai's trace is the most robust statistic. In discriminant analysis, a clear violation suggests considering quadratic discriminant analysis (QDA).

An everyday example

Weight loss and change in waist size are measured together in three diet groups. If the two measures are tightly related in one group but only weakly in the others, the covariance matrices differ and Box's M picks this up. Because the groups are balanced at 40 people each, the researcher still reports the MANOVA, using Pillai's trace.

In MerQur

The 'Box M' checkbox in MANOVA and Mixed-Design ANOVA (with a χ² approximation; in the mixed design it compares the covariance matrices of the within-subject levels across groups); and in the assumption output of Discriminant Analysis.

Analyses where this concept appears

Mixed-Design ANOVA, Split-PlotMANOVADiscriminant Analysis (LDA / QDA)

Related concepts

Multivariate test statistics (Wilks' Λ, Pillai's trace, Hotelling-Lawley, Roy)Covariance and correlation matrix