Robust Mann-Whitney U Test
Also known as: robust Wilcoxon rank-sum test, robust two-sample rank test, outlier-resistant Mann-Whitney test, robust nonparametric two-group comparison
The Robust Mann-Whitney U test is a nonparametric two-group comparison that combines the rank-based logic of the classic Mann-Whitney U test with modern robust techniques — such as outlier screening, trimmed means, or robust variance estimation — to produce reliable inferences when data contain extreme values, heavy-tailed distributions, or other violations that compromise the standard test.
Key highlights
- Maintains correct Type I error rates under outlier contamination and heavy-tailed distributions.
- Requires no normality assumption, making it safe for ordinal and non-normal continuous data.
- Combines the broad applicability of nonparametric rank tests with the stability benefits of robust statistical methods.
- Can be paired with bootstrap or permutation reference distributions for exact small-sample inference.
- Effect sizes such as the rank-biserial correlation provide meaningful, outlier-resistant summaries of group differences.
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 the Robust Mann-Whitney U test when you need to compare two independent groups on a continuous or ordinal outcome and you suspect or detect outliers or heavy-tailed distributions that would distort the standard test. It is especially valuable in small-to-moderate samples where a single outlier can shift ranks substantially, in exploratory clinical or behavioral data where extreme values are common, and whenever replication of results across plausible outlier-handling strategies is a priority. Do not use it when the standard Mann-Whitney U test already demonstrates robustness to the distributional features present, or when a parametric robust test (e.g. the Welch-corrected t-test with trimmed means) is more appropriate for your data type.
Strengths & limitations
- Maintains correct Type I error rates under outlier contamination and heavy-tailed distributions.
- Requires no normality assumption, making it safe for ordinal and non-normal continuous data.
- Combines the broad applicability of nonparametric rank tests with the stability benefits of robust statistical methods.
- Can be paired with bootstrap or permutation reference distributions for exact small-sample inference.
- Effect sizes such as the rank-biserial correlation provide meaningful, outlier-resistant summaries of group differences.
- More computationally demanding than the standard Mann-Whitney U test, especially when using bootstrap or permutation reference distributions.
- The choice of trimming proportion or outlier-detection rule is not standardized, introducing analyst degrees of freedom.
- Trimming reduces effective sample size, lowering power when the trimmed observations are not actually problematic.
- Results can be harder to explain to applied audiences unfamiliar with robust methods.
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
How is this different from the standard Mann-Whitney U test?
The standard Mann-Whitney U test assigns ranks to all observations and computes U without any pre-screening for outliers. The robust version adds a step — trimming, Winsorization, or a reweighting scheme — that reduces the influence of extreme observations on the rank order and on the variance estimate used for the p-value, producing more stable results when outliers are present.
What trimming proportion should I choose?
Wilcox recommends 20% trimming (removing the top and bottom 20% from each group) as a general default that balances robustness and power across a wide range of contamination scenarios. However, the choice should ideally be guided by diagnostic plots and sensitivity analyses, and should always be reported explicitly.
Should I use a bootstrap p-value or the normal approximation?
The normal approximation is adequate for large samples (n > 30 per group) with no ties. For small samples, or when a trimming step has been applied and effective sample sizes are reduced, a bootstrap or permutation p-value is more accurate and is the approach recommended by Wilcox.
Can I use this test with tied data?
Yes. Ties are handled by assigning average ranks, which is the same convention as the standard Mann-Whitney U test. If ties are very frequent, consider whether an ordinal model would be more appropriate.
What effect size should I report?
Report the rank-biserial correlation r (which ranges from -1 to 1 and equals 1 - 2U / (n1 * n2)) or the probability of superiority. Both are interpretable, assumption-light, and not inflated by outliers in the way that mean-based effect sizes can be.
Sources
- 1.Wilcox, R. R. (2005). Introduction to Robust Estimation and Hypothesis Testing (2nd ed.). Academic Press.ISBN 978-0127515427
- 2.Wilcox, R. R., & Keselman, H. J. (2003). Modern robust data analysis methods: Measures of central tendency. Psychological Methods, 8(3), 254–274.
You have read it. What now?
Cite this page
ScholarGate. (2026, June 3). Robust Mann-Whitney U test. ScholarGate. https://scholargate.app/statistics/robust-mann-whitney-u-test