Test
Design
Effect to detect

What this calculator does

A one-way MANOVA compares several groups on several outcome variables at once, rather than running a separate ANOVA on each one. This page answers the question you have to settle before collecting any data: how many subjects do I need?

It searches for the smallest total sample size whose power reaches the target you set, for the statistic you pick, and then reports the balanced design that total implies. The power curve underneath shows how power grows with sample size, so you can see how much a few more subjects would actually buy you.

The three statistics

MANOVA summarises the differences between groups in a matrix, and there is more than one sensible way to reduce a matrix to a single number. The three on offer here are the three that R's summary(manova(...)) reports:

StatisticWhat it doesWhen to prefer it
Wilks' LambdaThe ratio of the error variance to the total variance, as determinants. Mapped onto an F by Rao's approximation.The historical default, and what most textbooks report.
Pillai's TraceAdds up the share of variance each dimension explains.The most robust choice when the groups have unequal covariance matrices or the sample is small. This page's default.
Hotelling-Lawley TraceAdds up the eigenvalues themselves, so it weights the single strongest dimension most heavily.When you expect the effect to be concentrated in one direction.

None of the three is exactly F-distributed, so each is mapped onto an F through its own formula for the degrees of freedom. That is why the three give slightly different sample sizes for the same design — the difference is the approximation, not a disagreement about the data. Roy's Maximum Root is deliberately not offered: it is an upper bound rather than an exact test, and its F approximation is only a bound too.

How to use the calculator

  1. Significance level (α) — a p-value below this counts as statistically significant. Most researchers use 0.05; use something smaller, such as 0.01, when a false positive would be costly.
  2. Target power — the chance you want the test to have of detecting the effect when it is really there. 0.8 is the usual convention, 0.9 or 0.95 when missing the effect would be costly.
  3. Groups (g) — how many groups the one-way MANOVA compares.
  4. Dependent variables (p) — how many outcome variables are analysed together. See the warning about this below — more is not always better.
  5. Effect — if you do not know the effect size you need, pick Small, Medium or Large and the calculator fills in Cohen's conventional value and switches the effect type to f. This dropdown is only a shortcut: the calculation uses the effect type and effect size fields.
  6. Effect type — whether you are entering f, f² or η².
  7. Effect size — the smallest effect you want the test to be able to find. A smaller effect needs a larger sample, roughly with the square of the effect: halving the effect you want to detect multiplies the sample size by about four.
  8. Rounding (under More options) — how many digits the results are displayed to. It does not change the calculation.

Effect sizes, and how they convert

The three scales describe the same effect, so you may enter whichever one you have and the results table shows all three:

ScaleFrom f²To f²
Cohen's ff = √f²f² = f²
η²η² = f² / (1 + f²)f² = η² / (1 − η²)

Cohen's conventions for f, borrowed from ANOVA, are 0.1 small, 0.25 medium, 0.4 large. They are rules of thumb across the social sciences, not facts about your field — if there is a published effect size for the measure you are using, use that instead.

Why more dependent variables can mean a bigger sample

This is the result that surprises people most, and it is worth understanding before you add another outcome measure to a study.

With g groups, the group means can differ in at most g − 1 directions — two groups can only differ along the line between them, three groups can only span a plane, and so on. So the number of dimensions the effect can actually occupy is

s = min(p, g − 1)

Once p passes g − 1, adding another dependent variable adds no new direction for the effect to live in — but it still adds numerator degrees of freedom, which raises the critical value. The net result is usually that the required sample size goes up. Measure the outcomes your hypothesis is about, not every outcome you can record.

Wilks' Lambda is the exception to the formula, not to the idea: its Rao approximation uses an exponent b in the same place, and b is generally not a whole number. The results table names whichever of the two your chosen statistic uses.

How the power is calculated

Under H0 the F statistic follows a central F distribution with df1 and df2 degrees of freedom. Under H1 it follows a noncentral F with the same degrees of freedom and a noncentrality parameter

λ = f² · n · s

The power is the chance that a noncentral F exceeds the critical value taken from the central one. Power rises with n, so the calculator finds the smallest whole n whose power reaches the target — by doubling until it passes the target and then bisecting — and reports the power at n − 1 as well, which is what shows that n really is the smallest sample that works.

Balanced groups, and why the two totals differ

The search is over the total sample, and that total need not divide evenly by the number of groups. Equal groups therefore need ⌈n / g⌉ subjects each, which is a slightly larger total and a slightly higher power. Both figures are reported.

A balanced design is worth the extra subject or two for a reason beyond arithmetic: MANOVA assumes the groups share one covariance matrix, and it is at its most tolerant of that assumption being wrong when the groups are the same size.

How to read the power curve

The curve shows power against total sample size for the design you entered. The dashed line is your target power and the marker is the answer. Hover anywhere on the curve to read the sample size and the power at that point; the buttons in the corner download the chart as PNG or SVG.

The shape is the useful part. The curve is steep below the answer — where a few more subjects buy a lot of power — and flattens above it, where they buy very little. If your answer sits on the flat part, a larger study is mostly wasted effort.

Example

For a one-way MANOVA using Pillai's Trace with 4 groups and 2 dependent variables, at α = 0.05 and a target power of 0.8, against a medium effect of f = 0.25: a total sample of 113 reaches a power of 0.8034. For a balanced design that is 29 per group, i.e. 116 in total, with a power of 0.8158.

If the sample size is too large to be practical

There are only four levers, and they are not equally cheap:

  • Target a larger effect. By far the cheapest — the sample falls roughly with the square of the effect. Only honest if a larger effect is genuinely the smallest one worth detecting.
  • Measure fewer dependent variables. Especially if p is above g − 1, where the extra variables are costing degrees of freedom without adding a dimension.
  • Accept a lower power. Going from 0.9 to 0.8 saves a substantial fraction of the sample, at the price of missing more real effects.
  • Raise α. The cheapest arithmetically and the most costly scientifically — more false positives.

Related calculators