Statistics Glossary → Multivariate test statistics (Wilks' Λ, Pillai's trace, Hotelling-Lawley, Roy)
Multivariate test statistics (Wilks' Λ, Pillai's trace, Hotelling-Lawley, Roy)concept
Four classical test statistics that compare the mean vectors of several dependent variables across groups at once.
In MANOVA, between-group variation and within-group (error) variation are represented not by single numbers but by the sums-of-squares-and-cross-products matrices H and E. All four statistics are computed from the eigenvalues (λᵢ) of HE⁻¹. Wilks' Λ = Π 1/(1 + λᵢ) is the proportion of variance left unexplained, so smaller values mean a larger effect. Pillai's trace is V = Σ λᵢ/(1 + λᵢ) and the Hotelling-Lawley trace is T = Σ λᵢ. Roy's largest root uses only the largest eigenvalue. Each is converted to an approximate F statistic to obtain a p value.
When there is only one non-zero eigenvalue (e.g. two groups or a single within-subject factor), all four give the same exact F. Otherwise their results can differ. Pillai's trace is the most robust to violated assumptions (unequal covariance matrices, small or unbalanced groups) and is often recommended. Roy's test is the most powerful when the differences are concentrated in a single dimension, but its F is only an upper bound and, read on its own, it can inflate the type I error rate. Wilks' Λ is the one most commonly reported.
An everyday example
Students taught with three different methods are measured on both mathematics and science scores. Instead of testing each score with a separate ANOVA, MANOVA considers them together. A Wilks' Λ of 0.72 is the ratio of the within-group generalised variance to the total generalised variance (|E|/|H + E|); the smaller Λ is, the more the methods differ. 1 - Λ = 0.28 should not be read as a percentage of variance: it does not mean that 28% of the variance in each score is explained by method. In this design (s = 2 dimensions), the partial η² reported by SPSS is 1 - Λ^(1/2) ≈ 0.15.
In MerQur
In MANOVA, the 'Reference test' option chooses which statistic the overall decision is taken from, while the table always shows all four. In canonical correlation, 'Additional output: Multivariate tests' gives the four tests of the hypothesis that all canonical correlations are 0, and repeated measures and mixed-design ANOVA have a 'Multivariate tests' checkbox. In discriminant analysis, 'Additional output: Tests of equality of group means' gives a univariate Wilks' Λ for each predictor, and stepwise variable selection also uses the Wilks' Λ criterion.
Analyses where this concept appears
Repeated Measures ANOVAMixed-Design ANOVA, Split-PlotMANOVACanonical CorrelationDiscriminant Analysis (LDA / QDA)
Related concepts
Hotelling's T²Box's M test (equality of covariance matrices)