MANOVA

Multivariate Statistics

manova

wilks-lambda

multivariate-test

Multivariate analysis of variance: testing group differences across multiple correlated outcomes jointly

Published

April 17, 2026

Introduction

MANOVA — multivariate analysis of variance — tests whether group means differ across several dependent variables simultaneously, accounting for the correlations among the outcomes. It is more powerful than running separate one-way ANOVAs when outcomes are correlated, because it pools information across the response set instead of testing each in isolation. It also controls Type I error at the family-wise level for the joint hypothesis, removing the need for multiple-comparison correction across a battery of univariate ANOVAs.

Prerequisites

A working understanding of one-way ANOVA, multivariate Normal distributions, and covariance matrices.

Theory

MANOVA decomposes the total cross-product matrix into hypothesis ($H$) and error ($E$) components and constructs four standard multivariate test statistics from their joint structure:

Wilks’ lambda: $\Lambda = |E| / |H + E|$, the classical choice; small values reject the null.
Pillai’s trace: $V = \mathrm{tr}(H (H + E)^{-1})$, robust to covariance heterogeneity and the recommended default.
Hotelling–Lawley trace: $T = \mathrm{tr}(H E^{-1})$, more powerful when one dimension dominates.
Roy’s greatest root: $= $ largest eigenvalue of $H E^{-1}$, most powerful when group differences are concentrated in one direction but liberal otherwise.

All four test the same omnibus null hypothesis but have slightly different operating characteristics under non-Normality and unequal covariances.

Assumptions

Multivariate Normal within groups, homogeneous covariance matrices across groups (Box’s M test), and independent observations.

R Implementation

fit <- manova(cbind(Sepal.Length, Sepal.Width, Petal.Length, Petal.Width) ~
              Species, data = iris)
summary(fit)                     # Pillai default
summary(fit, test = "Wilks")
summary(fit, test = "Hotelling-Lawley")

# Follow-up univariate ANOVAs
summary.aov(fit)

# car Anova for Type II/III
library(car)
Manova(lm(cbind(Sepal.Length, Sepal.Width) ~ Species, data = iris))

Output & Results

summary(fit) returns the Pillai trace by default, with $F$ approximation, degrees of freedom, and $p$-value. Specifying other tests via the test argument produces Wilks, Hotelling-Lawley, and Roy variants. summary.aov(fit) runs follow-up univariate ANOVAs on each outcome.

Interpretation

A reporting sentence: “MANOVA showed a significant multivariate species effect (Wilks’ $\Lambda = 0.02$, $F(8, 288) = 199$, $p < 0.001$, multivariate $\eta^2 = 0.86$); univariate follow-up ANOVAs confirmed significant species differences on each of the four flower dimensions ($p < 0.001$ for each), with petal length carrying the strongest effect ($\eta^2 = 0.94$).” Always pair MANOVA with both univariate follow-ups and a multivariate effect-size measure.

Practical Tips

Pillai’s trace is the recommended default — it is the most robust to covariance heterogeneity and unequal sample sizes.
For a significant MANOVA, follow up with univariate ANOVAs on each outcome to identify which dimensions drive the effect; corrected $p$-values (Bonferroni or Holm) are appropriate at this stage.
Test homogeneity of covariance with Box’s M (biotools::boxM) before relying on Wilks; a significant Box’s M motivates Pillai.
For unequal sample sizes and multifactorial designs, use car::Manova() with explicit Type II or Type III sums of squares; base R’s summary.manova defaults to sequential (Type I).
Discriminant analysis follows naturally as a post-MANOVA exploration; the linear discriminants visualise where the multivariate group differences live.
For few groups with many highly-correlated outcomes, MANOVA is more powerful than Bonferroni-corrected univariate tests; for many groups with few outcomes, separate ANOVAs may be more interpretable.

R Packages Used

Base R manova() and summary.manova(); car::Manova() for Type II / III sums of squares; biotools::boxM() for the homogeneity-of-covariance test; MASS::lda() for discriminant-analysis follow-up; heplots for visualising hypothesis-error decompositions.