Robust Pearson Correlation
Also known as: winsorized correlation, percentage bend correlation, robust r, outlier-resistant correlation
The robust Pearson correlation is an outlier-resistant measure of linear association between two continuous variables. By applying Winsorizing, trimming, or percentage-bend transformations before computing the classic Pearson r, it retains the interpretability of a correlation coefficient while dramatically reducing the distortion caused by extreme values.
Key highlights
- Substantially reduces the influence of outliers and leverage points without deleting observations.
- Retains the familiar −1 to 1 scale and linear-association interpretation of classical Pearson r.
- Multiple robust variants (Winsorized, percentage-bend, biweight midcorrelation) allow tuning the degree of resistance.
- Hypothesis testing and confidence intervals are straightforward, especially when combined with bootstrapping.
- Works well even when distributional assumptions of classical Pearson r are mildly violated.
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 Pearson correlation when you suspect or observe outliers in one or both continuous variables and want a measure of linear association that is not driven by those extremes. It is appropriate when your sample is moderately sized (n ≥ 20 recommended) and your data are continuous but not guaranteed to be free of contamination. Do not use it as a default replacement for Pearson r when data are well-behaved; and note that it still assumes a primarily linear relationship — for monotone but non-linear patterns, Spearman or Kendall tau may be more natural choices.
Strengths & limitations
- Substantially reduces the influence of outliers and leverage points without deleting observations.
- Retains the familiar −1 to 1 scale and linear-association interpretation of classical Pearson r.
- Multiple robust variants (Winsorized, percentage-bend, biweight midcorrelation) allow tuning the degree of resistance.
- Hypothesis testing and confidence intervals are straightforward, especially when combined with bootstrapping.
- Works well even when distributional assumptions of classical Pearson r are mildly violated.
- Still assumes the primary relationship is linear; robust variants do not capture non-linear association.
- Choice of trim proportion (e.g., 10 % vs 20 %) is somewhat arbitrary and can affect results.
- Less widely implemented in standard software than classical Pearson r; often requires R packages such as WRS2.
- With very heavy contamination or extreme departures from linearity, Spearman or Kendall measures may outperform it.
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 does the Winsorized correlation differ from simply removing outliers?
Instead of deleting extreme values — which changes the sample size and can introduce bias — Winsorizing replaces them with the nearest non-extreme value. This retains all observations in the computation while preventing the extremes from exerting disproportionate influence.
Which robust variant should I choose?
The Winsorized correlation is the most common and easiest to test. The percentage-bend correlation (Wilcox, 1994) offers a slightly different contamination model. The biweight midcorrelation is popular in genomics. All three typically agree closely on clean-to-moderately-contaminated data; the choice matters most when contamination is severe.
Does robust Pearson correlation replace Spearman rho?
Not necessarily. Spearman rho is rank-based and handles any monotone relationship, not just linear ones, making it more broadly applicable. Robust Pearson r is preferable when you specifically want a measure of linear association that resists outliers but need the linear interpretation of r.
What software can I use?
In R, the WRS2 package (Mair & Wilcox) provides wincor(), pbcor(), and related functions. The robustbase and robust packages offer additional options. Native implementations in SPSS and Excel are absent, making R or Python (via pingouin or statsmodels) the practical choices.
How do I choose the trim proportion?
A trim proportion of 0.10 to 0.20 (10–20 % per tail) is conventional in behavioral research. Lower values offer minimal resistance; higher values discard too much information in clean samples. If you have no prior knowledge, 0.20 is a common default used by Wilcox.
Sources
- 1.Wilcox, R. R. (2012). Introduction to Robust Estimation and Hypothesis Testing (3rd ed.). Academic Press.ISBN 978-0123869838
- 2.Shevlyakov, G. L., & Oja, H. (2011). Robust Correlation: Theory and Applications. Wiley.ISBN 978-1118493458
You have read it. What now?
Cite this page
ScholarGate. (2026, June 3). Robust Pearson correlation. ScholarGate. https://scholargate.app/statistics/robust-pearson-correlation