Tests group differences on several related continuous outcomes simultaneously — e.g., do departments differ on the combination of satisfaction, commitment, and engagement?
Compare GroupsMultivariatealso known as: Multivariate analysis of variance
✓ When to use
Two or more conceptually related continuous outcomes and one or more categorical factors.
You want one protected omnibus test of the outcome profile before univariate follow-ups.
Outcomes are moderately correlated (~.3–.7) — that is where MANOVA has power that separate ANOVAs lack.
✗ When NOT to use
Outcomes are essentially uncorrelated — separate ANOVAs (with α correction) are simpler and clearer.
Outcomes are near-duplicates (r > .8–.9) — multicollinearity muddies the multivariate test; combine or drop.
Small samples relative to the number of DVs.
You truly care about each outcome separately from theory — pre-registered separate ANOVAs may serve better.
Data requirements
Dependent / outcome variable
Two or more continuous outcomes, ideally moderately intercorrelated.
Independent / grouping variable
One or more categorical factors.
Design
Between-subjects (extensions exist for repeated measures).
Sample size guidance
More cases than DVs in every cell — comfortably so; a common floor is n per cell > number of DVs + 20.
Assumptions
Independence of observations.
Multivariate normality of the DVs within groups (check univariate normality plus Mardia's test where available).
Homogeneity of covariance matrices across groups — Box's M (evaluated at a conservative α such as .001); with unequal ns and significant Box's M, prefer Pillai's Trace.
Linear relationships among DVs; no extreme multicollinearity; screen multivariate outliers via Mahalanobis distance.
Hypotheses
H₀ — The population mean vectors (the profile of DV means) are equal across groups.
H₁ — At least one group's mean vector differs.
The concept
MANOVA forms the linear combination of the outcomes that maximally separates the groups and tests whether that separation exceeds chance. Four statistics summarize it — Wilks' Λ, Pillai's Trace, Hotelling's T², Roy's largest root; they usually agree, and Pillai is the most robust to assumption violations.
A significant multivariate effect is followed either by univariate ANOVAs per DV (with Bonferroni-style protection) or, more informatively, by descriptive discriminant analysis showing which weighted combination of outcomes drives the separation. Multivariate effect size: partial η² based on the chosen statistic.
Worked example
Comparing three job grades (n = 50 each) on a profile of three outcomes: job satisfaction, affective commitment, and engagement (intercorrelations .45–.60).
Result: Pillai's Trace = .12, F(6, 292) = 3.05, p = .007, ηp² = .059. Follow-up ANOVAs (α = .017) show grade differences on satisfaction and engagement but not commitment.
How to run it
model <- manova(cbind(satisfaction, commitment, engagement) ~ grade, data = df)
summary(model, test = "Pillai") # multivariate test
summary.aov(model) # univariate follow-ups
library(effectsize); eta_squared(model)
from statsmodels.multivariate.manova import MANOVA
m = MANOVA.from_formula("satisfaction + commitment + engagement ~ grade", data=df)
print(m.mv_test()) # Wilks, Pillai, Hotelling, Roy
# follow-ups: pg.anova per DV with Bonferroni-adjusted alpha
Analyze → General Linear Model → Multivariate.
Move all DVs into 'Dependent Variables', the factor into 'Fixed Factor(s)'.
Options: Descriptives, Effect size, Homogeneity tests (Box's M, Levene per DV).
Read 'Multivariate Tests' (prefer Pillai's Trace), then 'Tests of Between-Subjects Effects' for univariate follow-ups at an adjusted α.
Report the multivariate statistic, its F conversion, df, p, ηp², then the univariate pattern.
Not feasible in native Excel — MANOVA requires matrix operations on covariance matrices.
Use R, Python, SPSS, or jamovi (free) instead; Excel can at most hold the raw data.
Interpreting the output
Choose and report one multivariate statistic (Pillai when in doubt) with its F approximation, df, p, ηp².
Only if the multivariate test is significant, examine per-DV ANOVAs at a corrected α — this is the 'protected' logic.
Check Box's M: if violated with unequal ns, note the reliance on Pillai.
Describe which DVs carry the group separation.
APA-style reporting
A one-way MANOVA revealed a significant multivariate effect of job grade on the combined outcomes, Pillai's Trace = .12, F(6, 292) = 3.05, p = .007, ηp² = .06. Follow-up univariate ANOVAs (Bonferroni-adjusted α = .017) showed significant grade differences for job satisfaction, F(2, 147) = 6.21, p = .003, and engagement, F(2, 147) = 4.98, p = .008, but not commitment, F(2, 147) = 1.42, p = .245.
Common mistakes
Using MANOVA as a ritual gateway when the DVs were chosen arbitrarily — DVs should form a meaningful construct profile.
Ignoring Box's M / relying on Wilks' Λ with unequal group sizes and heterogeneous covariances.
Following up with uncorrected univariate tests.
Including highly redundant DVs.
Too few cases per cell relative to the number of DVs.