How to Perform MANOVA in Clinical Research

Q: What is the difference between MANOVA and ANOVA?

ANOVA tests whether groups differ on a single continuous outcome. MANOVA tests whether groups differ on two or more continuous outcomes considered together as a combined, correlated set, producing one omnibus test rather than several separate ones. This matters because MANOVA can detect a multivariate pattern of group separation that no single outcome shows clearly on its own, and it controls the overall false-positive rate in a way that running several independent ANOVAs does not.

Q: How many dependent variables do I need for MANOVA?

A minimum of two continuous dependent variables is required, since MANOVA with a single outcome mathematically reduces to ordinary ANOVA. There is no strict upper limit, but each additional outcome adds complexity to interpretation and increases the sample size needed per group, so most applied clinical MANOVAs use somewhere between two and about five conceptually related outcomes rather than an unbounded set.

Q: Which MANOVA test statistic should I report?

Wilks' Lambda is the most commonly reported statistic in published research and works well when the assumptions of multivariate normality and homogeneous covariance matrices are reasonably met. Pillai's Trace is generally recommended instead when those assumptions are questionable or when group sizes are unequal, since it is the most robust of the four statistics to such violations, even at some cost in statistical power when assumptions genuinely hold. With exactly two groups being compared, all four statistics are mathematically equivalent and yield the same result.

Q: What if my dependent variables aren't correlated?

MANOVA is built around dependent variables that are conceptually and statistically related to each other; if the outcomes are essentially uncorrelated, much of MANOVA's practical rationale and its statistical power advantage over separate ANOVAs disappears, since there is no shared multivariate structure to model. In that situation, running separate, appropriately corrected univariate tests for each outcome is usually a more transparent and defensible choice than forcing unrelated variables into a single multivariate model.

Q: What is Box's M test and how strict should I be about it?

Box's M test formally checks whether the covariance matrices of the dependent variables are equal across groups, which is the multivariate equivalent of the homogeneity of variance assumption in ordinary ANOVA. Box's M is known to be extremely sensitive to sample size and to departures from multivariate normality, frequently flagging a significant violation even in datasets that are otherwise perfectly usable, so a significant result is usually treated as a signal to prefer the more robust Pillai's Trace statistic rather than as a reason to abandon MANOVA altogether.

Q: Do I need post hoc tests after MANOVA?

A significant multivariate result does not specify which individual outcome or outcomes drove the difference, or which specific groups differ, so a two-stage follow-up is standard practice: first, univariate ANOVAs (with a correction for the number of outcomes tested) to identify which dependent variables show significant group differences, and second, standard post hoc pairwise comparisons (such as Tukey's HSD) on each of those significant outcomes to identify which specific groups differ.

Q: What effect size should I report for MANOVA?

Partial eta squared (partial η²), derived from the chosen multivariate test statistic, is the standard effect size for the overall multivariate result, representing the proportion of multivariate variance in the combined outcomes explained by group membership. For the univariate follow-up tests on each individual dependent variable, a separate partial eta squared is reported for each outcome, since the multivariate and univariate effect sizes answer different questions and are not interchangeable.

Q: Can MANOVA be used for repeated measures?

The standard MANOVA covered in this guide assumes independent, between-subjects groups, with each subject contributing to exactly one group. When the same continuous outcomes are measured repeatedly on the same subjects, a related but more specialized approach, sometimes called doubly multivariate repeated measures analysis, or a multivariate mixed model, is required instead, since these explicitly model both the correlation between the multiple outcomes and the correlation between repeated time points.

Q: What is the difference between MANOVA and ANCOVA?

MANOVA compares groups on two or more correlated continuous dependent variables simultaneously. ANCOVA (analysis of covariance) compares groups on a single continuous outcome while statistically adjusting for one or more continuous covariates that are not the primary variables of interest. A design combining both — multiple correlated outcomes plus adjustment for covariates — is called MANCOVA, and follows the same core logic as MANOVA with the covariate adjustment layered in.

Q: What sample size do I need for MANOVA?

A commonly cited minimum guideline is more subjects than dependent variables in every group, but this is a bare mathematical floor rather than an adequate sample for real statistical power. A more practical rule of thumb suggests at least 20 subjects per group beyond the number of dependent variables being analyzed, and a formal multivariate power calculation based on the expected effect size, number of groups, and number of outcomes provides a far more reliable estimate than any general rule of thumb.

Clinical outcomes rarely travel alone. A drug trial might measure disease activity, inflammation, and patient-reported wellbeing together; a rehabilitation study might track strength, mobility, and pain as one connected picture of recovery. When you have two or more correlated continuous outcomes and want to know whether groups differ on the whole pattern — not just each measure in isolation — MANOVA (Multivariate Analysis of Variance) is the test built for exactly that. This guide is a complete, standalone walkthrough: what MANOVA actually tests, its assumptions, the four test statistics you'll see in output, a full worked clinical example, and exactly how to run, interpret, and report it using StatClinic.

Key Takeaways

MANOVA tests whether groups differ on two or more correlated continuous outcomes considered together, in a single omnibus test — it is the multivariate extension of ANOVA.
It is the right choice when your dependent variables are conceptually related and correlated — running separate ANOVAs on unrelated outcomes doesn't need MANOVA.
MANOVA output reports up to four multivariate test statistics — Pillai's Trace, Wilks' Lambda, Hotelling's Trace, and Roy's Largest Root — and they can disagree when assumptions are violated.
A significant multivariate result does not tell you which outcome differs — follow-up univariate ANOVAs and post hoc tests, both corrected for multiple comparisons, are required.
Pillai's Trace is the most robust statistic to assumption violations and is the safer default when covariance homogeneity or normality is in doubt.

What Is MANOVA?

MANOVA (Multivariate Analysis of Variance) is a statistical test that compares two or more groups on two or more continuous outcome variables simultaneously, testing whether the groups differ on the combined, correlated set of outcomes considered as a single multivariate pattern. It extends ordinary ANOVA — which handles exactly one outcome — to situations where several related outcomes are measured on the same subjects and analyzed together rather than one at a time.

Conceptually, instead of comparing group means on a single number line (as ANOVA does), MANOVA compares group centroids — the multivariate average position of each group across all outcomes at once — within the multidimensional space formed by the combined outcomes. Two groups can look similar on every individual outcome yet still be genuinely separable once the outcomes are considered jointly, and MANOVA is specifically built to detect that kind of pattern.

In Plain Terms

Instead of asking "does Group A differ from Group B on outcome X?" three separate times for three related outcomes, MANOVA asks one question: "considering all three outcomes together as one combined profile, do the groups occupy meaningfully different positions?" — a single test that also accounts for how the outcomes correlate with each other.

When Should You Use MANOVA?

MANOVA is the appropriate choice when your study meets these conditions:

You have two or more continuous outcome variables that are conceptually related and correlated with each other.
You have one or more categorical independent variables (factors) defining independent, between-subjects groups.
You want a single overall test of group differences across the combined outcomes, rather than several separate, uncorrected tests.
You're specifically interested in whether groups differ on the multivariate pattern, not just on any one outcome individually.

If you only have one outcome, use ordinary ANOVA. If your outcomes are essentially unrelated to each other, separate, appropriately corrected univariate tests are usually more transparent than forcing them into a single multivariate model.

Multiple Related Continuous Dependent Variables

MANOVA's statistical logic depends on the outcomes being genuinely related — not identical, but drawn from a shared underlying construct or physiological system. Disease-activity scores, inflammatory markers, and patient-reported wellbeing in the same condition are a natural fit; unrelated outcomes like blood pressure and a mood questionnaire, bundled together only for convenience, are not.

Before running MANOVA, it's worth inspecting the correlation matrix between your candidate outcomes. Moderate correlations (roughly 0.2 to 0.7) are typical and healthy for a MANOVA; correlations near zero suggest the outcomes don't share enough multivariate structure to benefit from being analyzed together, while correlations above about 0.9 raise a different problem — multicollinearity between the outcomes themselves, covered further in the assumptions section below.

MANOVA vs ANOVA and Multiple Separate ANOVAs

Feature	Multiple Separate ANOVAs	MANOVA
Outcomes tested	One at a time, each with its own test	All outcomes simultaneously, one omnibus test
Overall false-positive control	Inflated with each additional test, unless corrected	Controlled by the single multivariate test
Accounts for correlation between outcomes	No	Yes
Can detect a multivariate-only pattern	No	Yes
Typical use case	A single outcome of primary interest	Several correlated outcomes reflecting one construct

Take-Home Points Running three separate ANOVAs on three correlated outcomes doesn't just risk inflating the false-positive rate — it also throws away the information contained in how those outcomes move together, which is exactly the information MANOVA is built to use.

Suitable Study Designs

MANOVA fits an independent-groups (between-subjects) design: two or more groups made up of different, unrelated subjects, each measured on the same set of two or more continuous outcomes. A one-way MANOVA has a single grouping factor (such as treatment arm); a factorial MANOVA extends this to two or more grouping factors, testing main effects and interactions across the combined outcome set, analogous to how two-way ANOVA extends one-way ANOVA for a single outcome.

Standard MANOVA is not built for repeated measurements of the same outcomes on the same subjects — that situation calls for a more specialized doubly multivariate repeated measures model or a multivariate mixed model, which is beyond the scope of this guide.

Assumptions, Including Multivariate Normality, Covariance Homogeneity, Independence, and Multicollinearity

Two or more continuous, correlated outcomes

Conceptually related dependent variables measured on the same interval or ratio scale.

Independent groups

Different, unrelated subjects contribute to each group; each subject belongs to exactly one group.

Independence of observations

One subject's set of outcomes shouldn't influence another subject's.

Multivariate normality

Each outcome, and their linear combinations, should be approximately normally distributed within each group — approximated in practice by checking univariate normality per outcome per group and screening for multivariate outliers.

Homogeneity of covariance matrices

The pattern of variances and covariances among the outcomes should be similar across groups — checked formally with Box's M test.

No severe multicollinearity or singularity among outcomes

Outcomes correlated above roughly 0.9 with each other make the analysis unstable and largely redundant.

Linearity between all pairs of outcomes

MANOVA assumes the relationships among the dependent variables are approximately linear.

Adequate sample size per group

More subjects than outcomes in every group is a bare minimum; roughly 20 or more subjects per group beyond the number of outcomes is a more realistic planning guideline.

Box's M test deserves a specific caution: it is notoriously sensitive to sample size and to non-normality, frequently returning a significant result even in datasets that are otherwise perfectly usable. A significant Box's M is typically treated as a signal to favor Pillai's Trace over Wilks' Lambda, not as an automatic reason to abandon MANOVA.

Pillai's Trace, Wilks' Lambda, Hotelling's Trace, and Roy's Largest Root

MANOVA output reports up to four multivariate test statistics, each converted into an approximate F-statistic with its own p-value. They usually agree closely when assumptions are well met, but can diverge when they aren't — understanding what each one measures helps you choose which to lead with.

Statistic	What It Measures	When to Prefer It
Wilks' Lambda (Λ)	Ratio of within-group to total variance-covariance; ranges 0–1, smaller values indicate larger group differences	Most commonly reported; reliable when assumptions hold
Pillai's Trace	Sum of explained variance across all dimensions of group separation	Most robust to violated assumptions and unequal group sizes — the safer default
Hotelling's Trace (Hotelling-Lawley Trace)	Similar to Pillai's Trace but more sensitive to assumption violations	Commonly used with exactly two groups, where it is mathematically related to Wilks' Lambda
Roy's Largest Root	Based only on the single largest eigenvalue (the strongest dimension of separation)	Most statistically powerful when one dominant dimension truly separates groups; least robust otherwise

With exactly two groups being compared, all four statistics are mathematically equivalent and produce an identical F-value and p-value — the choice between them only matters once three or more groups are being compared.

Caution Roy's Largest Root is the most powerful statistic when the group separation genuinely concentrates on one dimension, but it can also be the most misleading when it doesn't — it ignores information the other three statistics use. Don't select it simply because it produces the smallest p-value.

How to Prepare the Dataset

MANOVA needs long format: one row per subject, one column identifying the group, and one column for each continuous outcome.

Patient ID	Treatment Group	DAS28	CRP (mg/L)	Global Health VAS
001	Standard DMARD	5.1	19.4	65
002	+ Biologic	3.4	8.7	39
003	+ New Biologic	2.7	5.9	28
...	...	...	...	...

Confirm the group variable is a consistent categorical label, that every subject appears exactly once (independent groups), and that every outcome column has a value for every subject — MANOVA, like most multivariate methods, requires complete cases on all outcomes.

A Realistic Medical Example With Multiple Outcomes

A rheumatology trial compares three treatment regimens for rheumatoid arthritis — standard DMARD therapy alone, standard therapy plus a biologic, and standard therapy plus a newer biologic — in 90 patients (30 per group). Three correlated outcomes are measured at 12 weeks, all reflecting disease activity from different angles: the DAS28 composite disease-activity score, CRP (an inflammatory marker), and patient-reported global health on a 0–100 VAS. Because these three outcomes are clinically and statistically related, MANOVA is the appropriate choice over three separate ANOVAs.

Group	n	Mean DAS28	Mean CRP	Mean Global Health VAS
Standard DMARD	30	4.8	18.2	62
+ Biologic	30	3.6	9.4	41
+ New Biologic	30	2.9	6.1	30

What This Pattern Suggests

All three outcomes move in the same direction across the groups — lower disease activity, lower inflammation, and better patient-reported health as treatment intensity increases — exactly the kind of correlated, multivariate pattern MANOVA is designed to test as a single combined effect.

Step-by-Step Analysis Using StatClinic

Enter your dataset in long format

One row per patient, with a Treatment Group column and one column for each of the three outcomes.

Confirm your study design

Tell StatClinic you have independent groups and multiple correlated continuous outcomes — this routes the analysis to MANOVA.

Let StatClinic check assumptions automatically

StatClinic screens normality per outcome per group, runs Box's M for covariance homogeneity, and checks the correlation matrix among outcomes for multicollinearity.

Run the MANOVA

Review all four multivariate statistics; StatClinic highlights Pillai's Trace as the recommended default if Box's M flags a violation.

Review the multivariate effect size

Partial η² for the overall multivariate effect is calculated and reported alongside the test statistic and p-value.

Run follow-up analyses automatically

If the multivariate result is significant, StatClinic runs corrected univariate ANOVAs and post hoc comparisons, then generates ready-to-use Methods and Results text.

You can also run the calculation directly with the MANOVA Calculator once your grouped, multi-outcome data are ready.

Interpretation of the Multivariate Result

For the rheumatoid arthritis example, StatClinic's output might read: Wilks' Λ = 0.62, F(6, 170) = 8.14, p < .001, partial η² = 0.22.

Output	What It Tells You
Wilks' Lambda (or chosen statistic)	A summary of how distinct the group centroids are across the combined outcomes — for Wilks' Λ, closer to 0 means a stronger effect
F, df	The approximate F-test converting the multivariate statistic into a familiar significance test, with degrees of freedom reflecting the number of outcomes and groups
P-value	The probability of seeing this much multivariate separation between groups if there were truly no difference
Partial η² (multivariate effect size)	Proportion of multivariate variance explained by group; roughly 0.01 small, 0.06 medium, 0.14+ large

Put together: the significant p-value confirms the three treatment groups occupy meaningfully different positions in the combined DAS28–CRP–global-health space, and partial η² = 0.22 indicates a large multivariate effect. What this result does not tell you is which of the three outcomes — or which specific pair of groups — is driving that difference; that requires the follow-up step below.

Follow-Up Univariate Analyses and Post Hoc Tests

A significant MANOVA is followed by two further steps. First, run a univariate ANOVA for each outcome separately, using a correction (commonly Bonferroni: 0.05 divided by the number of outcomes) to control the false-positive rate across the follow-up tests. Second, for every outcome that shows a significant univariate effect, run standard post hoc pairwise comparisons (such as Tukey's HSD) to identify exactly which groups differ on that specific outcome.

Follow-Up Results (Example)

With a Bonferroni-adjusted alpha of .017 (.05 ÷ 3 outcomes): DAS28, F(2,87) = 24.1, p < .001; CRP, F(2,87) = 19.8, p < .001; Global Health VAS, F(2,87) = 21.3, p < .001 — all three outcomes show significant group differences. Tukey post hoc comparisons then show both biologic-containing regimens outperform standard DMARD alone on every outcome, with the new biologic showing the largest improvement.

Effect Sizes and Multiple-Comparison Control

MANOVA analyses report effect sizes at two levels, and both should be included in a complete report: the overall multivariate partial η² from the omnibus test, and a separate univariate partial η² for each outcome from the follow-up ANOVAs — these answer different questions and are not interchangeable.

Multiple-comparison control is needed at two separate stages: once when correcting the univariate follow-up tests across the number of outcomes (commonly Bonferroni), and again when correcting the pairwise post hoc comparisons within each significant outcome across the number of group pairs. Skipping either correction re-introduces the inflated false-positive risk that running the single multivariate test was meant to control in the first place.

How to Report the Methods and Results

Sample Methods Text

"A one-way MANOVA was conducted to compare the effects of three rheumatoid arthritis treatment regimens (standard DMARD, standard plus biologic, standard plus new biologic) on three correlated disease-activity outcomes: DAS28 score, CRP, and patient-reported global health (VAS)."

Sample Results Text

"There was a statistically significant effect of treatment group on the combined disease-activity outcomes, Wilks' Λ = 0.62, F(6, 170) = 8.14, p < .001, partial η² = 0.22. Follow-up univariate ANOVAs, using a Bonferroni-adjusted alpha of .017, showed significant group differences for DAS28 (F(2,87) = 24.1, p < .001, partial η² = 0.36), CRP (F(2,87) = 19.8, p < .001, partial η² = 0.31), and global health VAS (F(2,87) = 21.3, p < .001, partial η² = 0.33). Tukey post hoc comparisons indicated both biologic-containing regimens produced significantly better outcomes than standard DMARD alone across all three measures, with the new biologic showing the greatest improvement."

Common Mistakes

Mistake 1: Running Separate ANOVAs on Correlated Outcomes

Testing several related outcomes with independent, uncorrected ANOVAs instead of one MANOVA, inflating the overall false-positive rate.

✓ Fix: Use MANOVA when outcomes are genuinely correlated and conceptually related.

Mistake 2: Bundling Unrelated Outcomes Together

Including outcomes with little conceptual or statistical relationship purely to run "one big test," losing the interpretability of the result.

✓ Fix: Check the correlation matrix first; only combine outcomes that genuinely share multivariate structure.

Mistake 3: Treating a Significant Box's M as Disqualifying

Abandoning MANOVA entirely after a significant Box's M test, without recognizing its well-known sensitivity to sample size and non-normality.

✓ Fix: Switch to the more robust Pillai's Trace rather than discarding the analysis.

Mistake 4: Ignoring Multicollinearity Between Outcomes

Including two outcomes that are almost perfectly correlated with each other, making the multivariate model unstable and largely redundant.

✓ Fix: Screen the outcome correlation matrix; remove or combine outcomes correlated above roughly 0.9.

Mistake 5: Assuming a Significant Multivariate Result Means Every Outcome Differs

Reporting "groups differed significantly" from the omnibus MANOVA alone, without running the univariate follow-up tests that identify which outcome actually drove the effect.

✓ Fix: Always follow a significant MANOVA with corrected univariate ANOVAs for each outcome.

Mistake 6: Skipping Correction on Follow-Up Tests

Running uncorrected univariate ANOVAs and post hoc comparisons after a significant MANOVA, undoing the Type I error control the multivariate test was meant to provide.

✓ Fix: Apply a correction (commonly Bonferroni) at both the univariate follow-up stage and the pairwise post hoc stage.

Mistake 7: Defaulting to Roy's Largest Root for the Smallest P-Value

Selecting Roy's Largest Root simply because it produces the most significant result, without confirming that group separation truly concentrates on one dimension.

✓ Fix: Choose the statistic based on your data's structure and assumption checks, not on which p-value looks best.

❌ Incorrect

"MANOVA was significant (p < .001), so all three treatment groups differ on DAS28, CRP, and global health."

✅ Correct

"Wilks' Λ = 0.62, p < .001; corrected univariate follow-ups confirmed significant group differences on all three individual outcomes, with Tukey post hoc tests identifying which specific groups differed on each."

Frequently Asked Questions

What is the difference between MANOVA and ANOVA? +

ANOVA tests group differences on a single continuous outcome. MANOVA tests group differences on two or more correlated continuous outcomes together, as one omnibus test that also accounts for how the outcomes correlate — something separate ANOVAs cannot do.

How many dependent variables do I need for MANOVA? +

At least two — with one outcome, MANOVA reduces to ordinary ANOVA. There's no strict upper limit, but most applied clinical MANOVAs use two to about five conceptually related outcomes, since more outcomes add complexity and require larger samples.

Which MANOVA test statistic should I report? +

Wilks' Lambda is most commonly reported and works well when assumptions hold. Pillai's Trace is recommended when normality or covariance homogeneity is questionable, since it's the most robust of the four. With exactly two groups, all four statistics are equivalent.

What if my dependent variables aren't correlated? +

MANOVA's rationale and power advantage depend on correlated outcomes. If outcomes are essentially uncorrelated, separate, appropriately corrected univariate tests are usually more transparent than forcing them into one multivariate model.

What is Box's M test and how strict should I be about it? +

Box's M tests whether covariance matrices are equal across groups. It's highly sensitive to sample size and non-normality, often flagging violations in otherwise usable datasets. A significant Box's M usually just means preferring Pillai's Trace, not abandoning MANOVA.

Do I need post hoc tests after MANOVA? +

Yes. A significant MANOVA doesn't identify which outcome or which groups differ. Follow with corrected univariate ANOVAs per outcome, then standard post hoc pairwise comparisons on any outcome that shows a significant univariate effect.

What effect size should I report for MANOVA? +

Partial eta squared for the overall multivariate effect, plus a separate partial eta squared for each outcome from the univariate follow-ups. The multivariate and univariate effect sizes answer different questions and should both be reported.

Can MANOVA be used for repeated measures? +

Standard MANOVA assumes independent, between-subjects groups. For the same outcomes measured repeatedly on the same subjects, a doubly multivariate repeated measures model or multivariate mixed model is needed instead.

What is the difference between MANOVA and ANCOVA? +

MANOVA compares groups on multiple correlated continuous outcomes. ANCOVA compares groups on a single outcome while adjusting for continuous covariates. Combining both — multiple outcomes plus covariate adjustment — is called MANCOVA.

What sample size do I need for MANOVA? +

More subjects than outcomes per group is a bare minimum. A more realistic guideline is at least 20 subjects per group beyond the number of outcomes, though a formal multivariate power calculation is far more reliable than any rule of thumb.

Once your MANOVA is run, these guides take you through the surrounding decisions:

Ready to Run Your MANOVA?

Let StatClinic's AI Statistical Assistant confirm the right test, run your analysis, and write your results sentence. Free, no registration required.

Try StatClinic Free →

How to Perform MANOVA in Clinical Research

What Is MANOVA?

When Should You Use MANOVA?

Multiple Related Continuous Dependent Variables

MANOVA vs ANOVA and Multiple Separate ANOVAs

Suitable Study Designs

Assumptions, Including Multivariate Normality, Covariance Homogeneity, Independence, and Multicollinearity

Two or more continuous, correlated outcomes

Independent groups

Independence of observations

Multivariate normality

Homogeneity of covariance matrices

No severe multicollinearity or singularity among outcomes

Linearity between all pairs of outcomes

Adequate sample size per group

Pillai's Trace, Wilks' Lambda, Hotelling's Trace, and Roy's Largest Root

How to Prepare the Dataset

A Realistic Medical Example With Multiple Outcomes

Step-by-Step Analysis Using StatClinic

Enter your dataset in long format

Confirm your study design

Let StatClinic check assumptions automatically

Run the MANOVA

Review the multivariate effect size

Run follow-up analyses automatically

Interpretation of the Multivariate Result

Follow-Up Univariate Analyses and Post Hoc Tests

Effect Sizes and Multiple-Comparison Control

How to Report the Methods and Results

Common Mistakes

Mistake 1: Running Separate ANOVAs on Correlated Outcomes

Mistake 2: Bundling Unrelated Outcomes Together

Mistake 3: Treating a Significant Box's M as Disqualifying

Mistake 4: Ignoring Multicollinearity Between Outcomes

Mistake 5: Assuming a Significant Multivariate Result Means Every Outcome Differs

Mistake 6: Skipping Correction on Follow-Up Tests

Mistake 7: Defaulting to Roy's Largest Root for the Smallest P-Value

Frequently Asked Questions

Related Articles

Ready to Run Your MANOVA?