MANOVA & Multivariate Tests
Real studies rarely measure one outcome. A therapy trial tracks anxiety and depression and sleep; a memory experiment logs accuracy and reaction time. MANOVA (Multivariate ANOVA) tests whether groups differ on a whole set of outcomes at once, and it can see differences that testing each outcome separately will flat-out miss.
Why not just run several ANOVAs?
Two reasons. First, the familiar multiple-testing problem: five outcomes at α = .05 gives you roughly a 23% chance of at least one false alarm. The second is less obvious: separate ANOVAs are blind to the correlation structure. Outcomes usually travel together, and a pattern that's unremarkable variable-by-variable can be wildly improbable as a combination. MANOVA looks at the cloud of outcomes as a shape, not as a list.
🎮 Two Outcomes, One Verdict
Every dot is a person measured on two outcomes. Slide the group differences and the correlation between outcomes, then compare the three verdicts below. The arrow is the discriminant direction, the single best combination of the two outcomes for telling the groups apart. (The same people stay on screen while you drag; "New sample" redraws.)
How it works: the smartest composite
MANOVA effectively asks: is there any weighted combination of the outcomes on which the groups clearly differ? It finds the combination that maximizes the between-group difference relative to the within-group spread (the discriminant direction in the plot) and tests the groups on that. Distance is measured the multivariate way (Mahalanobis distance): not "how far apart are the means?" but "how far apart are they given the shape of the cloud?"
That's the secret behind the ✨ button. When two outcomes are strongly correlated, they normally move together, so a treatment that pushes them in opposite directions is extremely surprising, even if each individual shift is small. Each univariate ANOVA sees a small, non-significant nudge; MANOVA sees a pattern the correlation structure says should almost never happen.
Don't skip the ✨ demo: small opposite differences (+0.35 and −0.35) with a strong correlation (r ≈ 0.8). Both univariate p-values hover far above .05 — separately, nothing to see. The MANOVA p-value collapses below .001. Now drag the correlation to 0 and watch the advantage evaporate: with independent outcomes, MANOVA is roughly just the two tests glued together.
The statistics you'll meet in the output
Software reports four multivariate test statistics, all summarizing how much of the outcome-space variation the group factor explains:
- Wilks' Λ (lambda) — the share of variance not explained by groups; smaller = stronger effect. The most commonly reported.
- Pillai's trace — the most robust to assumption violations; many texts recommend it as your default.
- Hotelling–Lawley trace and Roy's largest root — variations on the same theme; Roy's focuses only on the single best discriminant direction.
With exactly two groups they all collapse into one number: Hotelling's T², the multivariate cousin of the t-test, which is what the playground above computes. With three-plus groups and several outcomes they can disagree slightly; that's when Pillai's robustness earns its keep.
How big is the effect?
Those four statistics say whether the groups differ, not by how much, and a multivariate table of p-values with no effect size beside it has the same weakness as any other. The standard multivariate effect size falls straight out of Wilks' Λ:
η² = 1 − Λ1/s (s = the smaller of: number of outcomes, number of groups − 1)
With two groups s = 1 and the formula collapses to η² = 1 − Λ, so a Wilks' Λ of 0.62 says the grouping accounts for 38% of the variance in the outcome set taken as a whole. Statistical software will print a partial η² beside each multivariate test if you ask it to. Report one for the multivariate result, then an effect size per outcome in the follow-ups, since that is where a reader finds out which measures actually moved. The effect-size converter will turn those per-outcome numbers into d or r if that is the currency of your field.
The wider multivariate family
- MANCOVA = MANOVA + covariates: compare groups on several outcomes at once while adjusting for baselines, exactly like ANCOVA but multivariate.
- Multivariate multiple regression is the regression counterpart: several outcome variables predicted by several predictors simultaneously; MANOVA is the special case where the predictors are group dummies.
- Discriminant analysis is MANOVA run in reverse: instead of testing whether groups differ, it uses the discriminant directions to classify new cases into groups.
- Canonical correlation is the fully symmetric version: find the linked combinations of two whole sets of variables.
A significant MANOVA is a license to look closer, not a final answer. Standard follow-ups: univariate ANOVAs on each outcome (with a correction) to see which outcomes carry the difference, or a discriminant analysis to see what combination separates the groups.
Assumptions, briefly: independent observations, multivariate normality (roughly: each outcome normal, jointly well-behaved), and similar covariance matrices across groups (Box's M test, though it's oversensitive; with equal group sizes MANOVA is fairly forgiving). And choose outcomes that belong together conceptually. Stuffing ten unrelated variables into a MANOVA dilutes real effects instead of revealing them.
Why it matters: multivariate outcomes are the rule, not the exception — batteries of symptoms, panels of biomarkers, bundles of test scores. MANOVA and its family are how you test the profile honestly, without either alpha inflation or throwing away the structure that makes your measures informative.
Common questions
Does a significant MANOVA protect my follow-up ANOVAs from multiple testing?
It doesn't, and the belief that it does is one of the most durable habits in the multivariate literature. The idea borrows from the omnibus F in ANOVA: pass the overall test, and you have earned the right to look closer. The borrowing fails here because the multivariate test and the univariate ones ask different questions of different quantities, so clearing the first says nothing about the false-positive rate of the several tests that follow. Correct those on their own terms, with Bonferroni or an FDR procedure across the outcomes. The logic breaks in the other direction too: MANOVA can be significant when no single outcome is, which is what the ✨ demo above is built to show.
Should I report Wilks' lambda or Pillai's trace?
Wilks' Λ is the traditional default and what most textbooks tabulate. Pillai's trace is the most robust when assumptions wobble (unequal covariance matrices, unequal group sizes), so many methodologists recommend it outright. With two groups they (and Hotelling's T²) agree exactly; when they disagree materially with 3+ groups, that itself hints at assumption trouble, and Pillai is the safer citation.
What should I do after a significant MANOVA?
Localize the effect. The standard route: univariate ANOVAs on each outcome with a multiplicity correction, to see which variables carry the difference. The more multivariate route: descriptive discriminant analysis, which reveals what combination of outcomes best separates the groups — often the more faithful summary, since a combination is what MANOVA actually tested.