Robustness with skewed data.

Rates of type I errors in UKS test for 3 representative experimental designs (lines) and the 4 skewed distributions shown in Figure 3B (columns). In each design, the UKS test was applied before and after log-transforming the random datasets. The rates of each design are equal to the percentages of 60000 random datasets with null factor effect that were found significant at the 0.05 threshold by the UKS test. The type I error rates obtained for the same data with Kruskal-Wallis test substituted to Anova are also indicated for the third design. Overall, either log-transformation of skewed data or use of a per-individual nonparametric test guards the UKS test against excessive type I errors.