Non-parametric tests: what they assume, and what they cost They are widely described as assumption-free and widely believed to be much weaker than their parametric counterparts. Neither is quite true, and both misconceptions lead people to the wrong test.

A parametric test assumes the data come from a distribution of a particular family, usually normal, and works with its parameters: the mean and standard deviation. A non-parametric test makes no such assumption about shape. Most of the ones in routine use achieve that by throwing away the values and keeping only their order.

The rank transformation, and what it buys

Replace every observation by its rank in the combined sample — smallest is 1, next is 2, and so on — and then work with the ranks. Ranks are always 1, 2, 3… regardless of what the original numbers were. So the null distribution of the test statistic can be worked out exactly, without knowing anything about the distribution the data came from.

Two properties follow immediately. The test is robust to outliers: the most extreme value in the dataset becomes simply the highest rank, no matter how far out it sits. And it is invariant to monotonic transformation: taking logarithms changes no rank, so the test gives an identical answer on the log scale as on the original one. That invariance is a considerable convenience when the right scale is uncertain.

Two number lines. The upper one shows raw observations clustered together with one value far to the right, labelled outlier dominates the mean and standard deviation. The lower one shows the same observations converted to evenly spaced ranks one through eight, with the former outlier now simply at rank eight, labelled outlier becomes just the highest rank.
Ranking discards the spacing and keeps the order. That is how ranking works: the extreme value still counts as the largest, but it can no longer drag an estimate towards itself.

The main tests, and their parametric counterparts

The pairings are direct. The Wilcoxon-Mann-Whitney test replaces the independent-samples t-test. The Wilcoxon signed-ranks test replaces the paired t-test, and the sign test is the still more robust option that uses only the direction of each difference. Kruskal-Wallis replaces one-way ANOVA, and Friedman replaces the within-subjects version. Spearman’s rs and Kendall’s τ replace Pearson correlation, as covered in Pearson, Spearman or Kendall.

Follow-up comparisons exist too, and are just as necessary here as anywhere. After a significant Kruskal-Wallis, the Dwass-Steel-Critchlow-Fligner procedure compares all pairs and Steel’s procedure compares against a control, both controlling the family-wise error rate the way multiple comparison procedures require.

They cost far less power than people think

The persistent belief is that non-parametric tests are much less powerful. Under the conditions where the parametric test is exactly right (genuinely normal data), the Wilcoxon-Mann-Whitney test retains about 95% of the efficiency of the t-test. In practical terms the rank test needs roughly one extra subject in twenty to reach the same power, which is a small price.

And when the data are not normal, the comparison reverses. For heavy-tailed distributions, or data with occasional extreme values, the rank test can be more powerful than the t-test, substantially so, because the t-test’s standard deviation is inflated by the very observations the rank test neutralises.

So the real trade is not power. The real trade is interpretation.

They are not assumption-free

Rank tests are not assumption-free. Rank tests drop the assumption about shape and keep the others. Independence still matters exactly as much as before, and no rank test will detect or forgive clustered observations.

What the test compares depends on an assumption that is easy to overlook. Wilcoxon-Mann-Whitney is properly a test of whether one group tends to produce larger values than the other: stochastic dominance. The test is a test of medians only under the additional assumption that the two distributions have the same shape and differ by a shift. If one group is far more spread out than the other, the test can return a significant result while the medians are identical, because the distributions differ in a way that is not a shift.

Ties reduce the information. Heavily tied data, such as coarse scales or many repeated values, weakens rank tests, and Kendall’s τ handles ties more gracefully than Spearman does.

Very small samples cannot reach significance at all. With three observations per group, no arrangement of ranks produces a two-sided p below 0.05. The test is not failing; there is simply not enough ordering information in six numbers.

Report an estimate, not just a p-value

The most common complaint about non-parametric tests is that they give a p-value and nothing you can quote as an effect. The complaint describes a habit rather than a limitation.

The Hodges-Lehmann location shift is the natural estimate: the median of all pairwise differences between the groups, on the original measurement scale, with a confidence interval — Moses for independent samples, Tukey for paired. The Hodges-Lehmann shift is to the rank tests what the mean difference is to the t-test, and it should be reported the same way. A median difference with a Thompson-Savur interval is the alternative when the median itself is the quantity of interest.

When to reach for one

Ordinal data. The clearest case, and not really a choice: if the spacing between categories is not meaningful, the mean is not a quantity and only rank methods apply.

Small samples with visible skew or outliers. Small n is where non-normality actually costs you, since the central limit theorem is not yet helping.

Data with genuine extreme values you cannot justify removing. Robustness by design beats a judgement call about deletion.

When the median is the more meaningful summary. For strongly skewed quantities, such as times to event, costs or concentrations, the median often describes the typical case better than the mean does.

Against which: if the mean is the quantity that matters (a total yield, an average cost per unit), then a test about ranks is answering a different question, and a transformation or a robust parametric approach may serve better. And do not choose between them by running a normality test first; the reasons are in testing normality and what to do when it fails.

Downloads

Download the paired comparison example workbook (.xlsx): 28 before-and-after observations with box plots and paired lines, a difference plot with the Hodges-Lehmann shift, and the Wilcoxon signed-ranks test, ready to open in the Analyse-it trial.

Common mistakes

Believing they assume nothing. Independence still applies, and interpreting the result as a difference in medians requires similar distribution shapes.

Avoiding them because of a supposed power penalty. Around 5% under ideal normality, and often an advantage otherwise.

Choosing one after a normality test. Conditioning the analysis on a preliminary test distorts the properties of whichever you run.

Reporting only the p-value. The Hodges-Lehmann shift with its interval is an effect estimate on the original scale. Quote it.

Using them on tiny samples and reading non-significance as no effect. With three per group, significance is unreachable whatever the data show.

Using them to fix non-independence. Clustered or repeated observations need a design-level answer, not a different test.

Run a non-parametric test with Analyse-it

Analyse-it runs them with the estimate attached, not just the p-value, inside Excel:

  • Wilcoxon-Mann-Whitney and Kruskal-Wallis in Compare Groups; Wilcoxon signed-ranks, sign and Friedman in Compare Pairs
  • The Hodges-Lehmann location shift with a Moses or Tukey confidence interval alongside every p-value
  • Dwass-Steel-Critchlow-Fligner and Steel for the follow-up comparisons

Every feature from all five editions for 15 days, with no sign-up and no licence key. Hypothesis testing is part of the Standard edition, so it is in every Analyse-it edition, from US$ 155 a year. Validated against NIST Standard Reference Datasets.