Extended definition
MANOVA (Multivariate Analysis of Variance) is the multivariate extension of ANOVA: it tests whether means of multiple dependent variables differ across groups defined by one or more categorical independent variables, accounting for the correlation structure across outcomes. Instead of running separate ANOVAs (each with its own type I error rate, inflating the overall false-positive rate), MANOVA tests a joint hypothesis: is there a difference in any multivariate aspect of group means? Standard test statistics: Wilks’ Lambda (most used), Pillai’s trace (most robust to assumption violations), Hotelling-Lawley trace, Roy’s largest root (most powerful when a dominant dimension exists). Tabachnick and Fidell (2019, Using Multivariate Statistics, 7th ed., Pearson) offer the consolidated didactic reference; Bray and Maxwell (1985, Multivariate Analysis of Variance, Sage) is the classical technical treatment. Central assumptions: multivariate normality, homogeneity of covariance matrices (tested via Box’s M), observation independence, linearity across outcomes. When MANOVA is significant, post-hoc analysis follows — usually subsequent univariate ANOVAs, or discriminant analysis to identify where the difference lies.
When it applies
MANOVA applies when there are multiple correlated dependent variables measuring the same theoretical construct or related aspects of a phenomenon. It is standard in experimental research in psychology (treatment effect on multiple cognitive dimensions), in education (intervention effect on multiple skills), in health (treatment effect on multiple correlated biomarkers), in organizational management (program effect on multiple performance dimensions). It also applies in repeated-measures multivariate designs and in factorial experiments with multiple outcomes. Reporting MANOVA instead of multiple separate ANOVAs is good editorial practice in top-tier journals in experimental psychology and behavioral sciences.
When it does not apply
It does not apply with a single outcome — ANOVA is appropriate. It does not apply when multiple outcomes are theoretically independent: in those cases, separate ANOVAs with multiple-testing correction (Bonferroni, FDR) can be more transparent. It does not apply when assumptions are severely violated: Box’s M rejecting covariance homogeneity → consider transformations or robust MANOVA. It does not apply with small relative to number of outcomes: dimension close to or exceeding destabilizes estimation. It does not replace SEM when the interest is the structure of relations among latent constructs. In a continuous outcome combined with categorical, standard MANOVA does not accommodate — generalized models or specific approaches are needed.
Applications by field
— Experimental psychology: treatment effect on multiple cognitive/behavioral measures; standard in top-tier journals (JEP, Psychological Science). — Education: program evaluation with multiple skills as outcomes. — Health: intervention effects on biomarker panels; trials with multiple clinical endpoints. — Organizational management: organizational intervention effects on correlated performance dimensions.
Common pitfalls
The first pitfall is using MANOVA as a shortcut to avoid pre-registration of specific hypotheses: if the hypothesis is about a specific outcome, pre-registered ANOVA is more transparent. The second is failing to check assumptions: Box’s M rejection can invalidate Wilks-based inference; Pillai is a more robust alternative. The third is interpreting significant MANOVA as a universal claim about differences in all outcomes — post-hoc analysis is needed to localize where the difference lies. The fourth is treating MANOVA as more “powerful” than separate ANOVAs with multiple-testing correction: power gains occur only when outcomes are informatively correlated; with independent outcomes, no gain. The fifth is failing to report multivariate effect size (multivariate partial eta-squared): p-value alone is editorially weak reporting in modern research.