Nonparametric tests
Tests that do not assume a normal distribution: Mann-Whitney, Wilcoxon, Kruskal-Wallis, Mood's median and more.
Nonparametric tests work on ranks or signs instead of the raw values, so they do not assume a normal distribution. They are useful for skewed data, ordinal ratings and data with outliers.
Common tests and their parametric partners
- Sign test / 1-sample Wilcoxon: median vs a target (1-sample t).
- Mann-Whitney (Wilcoxon rank-sum): two independent groups (2-sample t).
- Wilcoxon signed-rank: paired data (paired t).
- Kruskal-Wallis: three or more groups (one-way ANOVA).
- Mood's median test: three or more groups, robust to outliers.
- Friedman test: blocked designs (two-way ANOVA without replication).
- Spearman and Kendall correlation: monotonic relationships (Pearson correlation).
- Runs test: randomness of a sequence.
Trade-offs
Nonparametric tests are slightly less powerful than t-tests when data really are normal, but can be much more powerful when data are skewed or contain outliers. They test medians or distributions, not means.
Questions and answers
Should I always use nonparametric tests to be safe?
No. With roughly normal data, t-tests and ANOVA are more powerful and give results in the original units. Use nonparametric tests when the assumptions clearly fail and samples are small.
My data have many ties (rounded values). Does that matter?
Ties reduce the precision of rank tests slightly. Most software applies a tie correction. Heavy rounding is usually a measurement resolution problem worth fixing.
