Robust Repeated Measures ANOVA
Also known as: robust within-subjects ANOVA, trimmed-mean repeated measures ANOVA, robust RM-ANOVA, heteroscedastic repeated measures ANOVA
Robust repeated measures ANOVA tests whether population trimmed means differ across three or more repeated conditions or time points measured on the same subjects. By replacing ordinary means with 20% trimmed means and replacing variances with Winsorized estimates, it maintains acceptable Type I error and power when data are non-normal, skewed, or contain outliers — conditions under which classical repeated measures ANOVA routinely breaks down.
Key highlights
- Controls Type I error rate reliably under non-normal distributions and outliers where classical RM-ANOVA fails.
- Maintains good statistical power with skewed or heavy-tailed data.
- Handles heterogeneous covariance structures that violate sphericity more gracefully than the standard test.
- The 20% trimming level is a well-studied default that balances robustness and efficiency.
- Supported in R (WRS2 package) and can be adapted in SPSS syntax.
Intuition
This section is available to Pro members. Upgrade to Pro
How it works
This section is available to Pro members. Upgrade to Pro
When to use it
Use robust repeated measures ANOVA when the same subjects are measured across three or more conditions or time points and classical RM-ANOVA assumptions are in doubt: non-normal distributions, outliers, or heterogeneous covariance matrices. It is particularly appropriate with small to moderate samples where normality is hard to verify and a few extreme scores could distort standard results. Do not use it when data are clearly bimodal or when the research question concerns medians rather than trimmed means; in those cases, the Friedman test or permutation-based approaches may be more appropriate. Also avoid it when the trimmed sample per condition is very small (fewer than 10 observations after trimming), as the asymptotic approximation may not hold.
Strengths & limitations
- Controls Type I error rate reliably under non-normal distributions and outliers where classical RM-ANOVA fails.
- Maintains good statistical power with skewed or heavy-tailed data.
- Handles heterogeneous covariance structures that violate sphericity more gracefully than the standard test.
- The 20% trimming level is a well-studied default that balances robustness and efficiency.
- Supported in R (WRS2 package) and can be adapted in SPSS syntax.
- Trimmed means test a different population parameter than ordinary means; results are not directly comparable to classical RM-ANOVA when distributions are symmetric and well-behaved.
- Requires adequate sample size: after 20% trimming from each tail, each condition must retain enough observations for the asymptotic approximation.
- Less familiar to reviewers and readers than classical RM-ANOVA; requires additional explanation in manuscripts.
- Post-hoc follow-up procedures are less standardized and software support is more limited than for classical ANOVA.
Common pitfalls
This section is available to Pro members. Upgrade to Pro
Applications
This section is available to Pro members. Upgrade to Pro
Frequently asked
Why 20% trimming — can I use a different level?
The 20% default (trimming the bottom and top 20% of each sample) is a broadly validated compromise: it provides substantial protection against outliers while retaining enough data for efficient estimation. You can use 10% or 15% if distributions are only mildly skewed and sample sizes are small. Wilcox (2012) discusses the trade-offs in detail.
Is robust RM-ANOVA the same as the Friedman test?
No. The Friedman test is a rank-based nonparametric alternative that tests whether rank distributions differ across conditions; it discards magnitude information. Robust RM-ANOVA retains magnitude through trimmed means and is generally more powerful when the true effect is in the location of the distribution. Use Friedman when you genuinely want a rank-based answer or when trimming removes too many observations.
Do I still need to check sphericity?
The robust approach reduces sensitivity to heterogeneous covariance structures but does not entirely eliminate it. For severely non-spherical data, combining robust estimation with a heteroscedasticity-consistent approach (as implemented in WRS2) provides the best protection. You should still examine the covariance structure and report it.
How do I run this in R?
Install the WRS2 package (Mair & Wilcox). The function rmanova() performs robust one-way repeated measures ANOVA using trimmed means, and rmmcp() runs post-hoc pairwise comparisons with multiple-testing control.
When should I prefer classical RM-ANOVA over the robust version?
When your data are approximately normally distributed within each condition, sample sizes are adequate, and sphericity holds (or a Greenhouse-Geisser correction suffices), classical RM-ANOVA is perfectly appropriate and more familiar to readers. The robust version adds value mainly when those conditions are in doubt.
Sources
- 1.Wilcox, R. R. (2012). Introduction to Robust Estimation and Hypothesis Testing (3rd ed.). Academic Press.ISBN 978-0123869838
- 2.Keselman, H. J., Wilcox, R. R., & Lix, L. M. (2003). A generally robust approach to hypothesis testing in independent and correlated groups designs. Psychophysiology, 40(4), 586–596.
You have read it. What now?
Cite this page
ScholarGate. (2026, June 3). Robust repeated measures ANOVA. ScholarGate. https://scholargate.app/statistics/robust-repeated-measures-anova