MANOVA, the multivariate analysis of variance, tests whether groups differ on several dependent variables at once, treating those outcomes as a single combined response rather than testing each one separately. It asks whether the groups differ on the overall profile of outcomes, and only if that combined test is significant do you examine the individual variables. The point is to capture differences that live in the pattern across outcomes, which separate analyses can miss.
Why not just run separate ANOVAs?
When a study has two or more related outcome variables that are conceptually part of the same construct, for example testing whether teaching methods affect both exam scores and coursework grades together, the tempting route is a separate univariate ANOVA on each outcome. That route inflates the false-positive rate across the family of tests and ignores the correlations between the outcomes. MANOVA controls that error rate with a single omnibus test and can detect a group difference that is spread thinly across several correlated measures.
ANOVA versus MANOVA
The difference is in the number of dependent variables. An ANOVA has one outcome; a MANOVA has two or more, analysed jointly. MANOVA produces multivariate test statistics, most commonly Wilks' lambda, along with Pillai's trace, Hotelling's trace, and Roy's largest root, each summarising the group difference across the whole set of outcomes. Because of this multivariate structure, MANOVA is sometimes described simply as the multivariate extension of ANOVA, which is also the answer to the common question of what another name for it is.
| Test | Dependent variables | Headline statistic |
|---|---|---|
| ANOVA | One | F ratio |
| MANOVA | Two or more, analysed jointly | Wilks' lambda |
Interpreting Wilks' lambda
Wilks' lambda is the most commonly reported multivariate statistic, and reading it trips up many students because it runs the opposite way to an F ratio. Lambda is the proportion of variance in the combined outcomes not explained by the grouping variable, so it ranges from 0 to 1 and smaller values mean a stronger group effect. A lambda near 1 says the groups barely differ on the outcome profile; a lambda near 0 says they differ sharply. Software converts lambda into an approximate F statistic with two sets of degrees of freedom and a p-value, and that converted F is what you test for significance. You can also derive a multivariate effect size, multivariate eta-squared, as one minus lambda, which puts the result on the familiar 0 to 1 variance-explained scale described in the guide to effect size.
When your groups have only two levels, Wilks' lambda, Pillai's trace, and the other three statistics all agree exactly, so the choice between them only matters with three or more groups. Pillai's trace is the most robust when the equal-covariance assumption is shaky, which is why many methodologists report it alongside lambda. Whatever you lead with, state the statistic, its value, the approximate F, the two degrees of freedom, and the p-value together, in the order set out in the APA reporting guide.
When to choose MANOVA over separate ANOVAs
Having multiple dependent variables does not by itself justify a MANOVA. The test earns its place only when the outcomes are genuinely correlated and belong to one underlying construct, so that the group difference lives in their joint pattern. If your outcomes are unrelated, a MANOVA can actually lose power relative to well-chosen separate analyses, and it complicates the follow-up without buying you anything. A useful check is whether the outcomes would load on a single factor, the kind of question explored in factor analysis and scale reliability. When the outcomes are only loosely linked, it is often cleaner to run one analysis per outcome with a corrected alpha, deciding each design with the help of the test selection tool.
Is MANOVA a two-way ANOVA?
No, though the names invite confusion. A two-way ANOVA has one outcome and two independent factors. A one-way MANOVA has one factor but several outcomes. The "multi" in MANOVA refers to multiple dependent variables, not multiple factors. You can of course combine the ideas into a two-way MANOVA with two factors and several outcomes, but the feature that makes it a MANOVA is always the set of dependent variables analysed together.
Assumptions and the follow-up analysis
MANOVA assumes multivariate normality, homogeneity of the covariance matrices across groups (checked with Box's M), independence of observations, and that the outcomes are correlated but not so highly that they are redundant. A significant multivariate test is only the first step: you then probe which outcomes drive the effect, traditionally with follow-up ANOVAs using a corrected alpha, or with discriminant analysis. Report the multivariate statistic, its approximate F and degrees of freedom, the p-value, and a multivariate effect size, following APA conventions. If your outcomes are not genuinely related, separate analyses guided by the test-selection guide are often the cleaner choice, and you can sanity-check each single-outcome comparison with the one-way ANOVA calculator.