A repeated measures ANOVA tests whether the means of three or more conditions differ when the same participants are measured in every condition. Because each person acts as their own comparison, the design removes the variation between individuals from the error term, making it far more sensitive than comparing separate groups. It is the go-to test for within-subjects designs such as measuring an outcome before, during, and after an intervention on the same people.
When a repeated measures design pays off
The design earns its keep whenever the same units are measured under several conditions or at several time points: a pre-test, mid-test, and post-test on one cohort, or the same panel rating three product designs. It extends the paired-samples t-test from two measurements to many, just as the one-way ANOVA extends the independent t-test. The shared logic is that pairing the observations within a person strips out stable individual differences, which is exactly the noise that makes between-groups designs less powerful.
How it differs from a two-way ANOVA
This is a common point of confusion. A two-way ANOVA has two independent factors, each made up of different groups of participants, and tests their main effects and interaction. A repeated measures ANOVA has one factor whose levels are all measured on the same participants. The two are not rivals: you can have a two-way design where one factor is between-subjects and the other is repeated, called a mixed ANOVA. The defining feature of the repeated measures version is that the levels share participants, which changes the error structure and the assumptions.
| Design | Participants across conditions | Key assumption |
|---|---|---|
| One-way ANOVA | Different in each group | Equal variances |
| Repeated measures ANOVA | Same in every condition | Sphericity |
The sphericity assumption
The distinctive assumption of a repeated measures ANOVA is sphericity: the variances of the differences between every pair of conditions should be equal. Mauchly's test checks it, and when it is violated the F test becomes too liberal. The fix is a correction to the degrees of freedom, the Greenhouse-Geisser or Huynh-Feldt adjustment, which most software applies automatically. Reporting which correction you used, and why, is part of a defensible analysis rather than an optional footnote.
Repeated measures ANOVA assumptions
Sphericity is the headline condition, but it is not the only one. The outcome should be roughly normally distributed at each time point or condition, the observations across conditions should be related rather than independent by design, and there should be no extreme outliers within any condition, because a within-subjects design magnifies the influence of an unusual participant. You also need complete data: a participant missing even one condition is dropped from the default analysis entirely, which can quietly shrink your sample and your statistical power. For that reason it pays to plan how you will handle gaps before collection, a topic covered in dealing with missing data. Check normality at each level with a histogram and a formal test rather than assuming it carries over from one condition to the next.
Repeated measures vs mixed ANOVA
A pure repeated measures ANOVA has every factor measured on the same people, while a mixed ANOVA combines at least one within-subjects factor with at least one between-subjects factor. The classic mixed design is a before-and-after measurement (the repeated factor) compared across a treatment group and a control group (the between-subjects factor), and its most interesting output is the interaction, which tells you whether the change over time differs between the groups. Choosing between the two comes down to whether your second factor splits participants into separate groups or simply measures them again. When the added factor is a continuous covariate instead of a grouping variable, the design shifts toward covariate-adjusted analysis of variance, and untangling which structure fits your data is exactly what the test selection tool is built to help with.
Interpreting and reporting the result
A significant within-subjects F tells you the condition means are not all equal, but not which ones differ, so follow up with pairwise comparisons using a correction such as Bonferroni. Report the F statistic, the (possibly corrected) degrees of freedom, the p-value, and an effect size such as partial eta-squared, following APA reporting conventions. If your design adds a continuous covariate or a second between-subjects factor, the analysis moves toward analysis of covariance or a mixed model, and choosing between them is exactly the kind of decision the test-selection guide is built for.