Introduction
Non-parametric methods provide statistical tests that do not rely on assuming a specific underlying distribution (such as the normal distribution assumed by many standard, “parametric” tests) — essential when data is ranked/ordinal, clearly non-normal, or when sample sizes are too small to verify distributional assumptions confidently.
1. Rank-based methods
Rank-based methods replace actual data values with their ranks (their position when data is ordered from smallest to largest) before analysis, making the test valid regardless of the data’s original distribution shape, and also more robust to outliers (since an extreme value’s rank is bounded, unlike its actual magnitude). This approach underlies procedures such as the Mann–Whitney U test (comparing two independent samples) and the Kruskal–Wallis test (comparing more than two independent samples) — the non-parametric analogues of the parametric t-test and ANOVA respectively.
2. Sign-type procedures
Sign tests examine only the direction (positive or negative sign) of differences between paired observations, ignoring the magnitude of each difference entirely — this makes them applicable even when the actual size of a difference is unreliable or unavailable, but also means they use less of the available information than a method that does account for magnitude (such as the more powerful Wilcoxon signed-rank test, which ranks the absolute differences before applying signs).
3. Chi-square-type categorical methods
The chi-square test compares observed category frequencies against those expected under some hypothesis, testing goodness-of-fit (whether one categorical variable follows a specified distribution) or independence (whether two categorical variables are related, using a contingency table of observed counts) — appropriate specifically for categorical (count) data, in contrast to the ranked/continuous data handled by the other methods in this chapter.
4. Situations for choosing non-parametric tests
Non-parametric tests are preferred over their parametric counterparts when: the data is ordinal (naturally ranked rather than measured on an interval scale), the sample size is small and normality cannot be verified or reasonably assumed, or the data contains extreme outliers that would unduly distort a parametric test’s results. The trade-off is that non-parametric tests are generally somewhat less statistically powerful than an appropriate parametric test would be if the parametric test’s assumptions genuinely held — non-parametric methods are the safer, more robust choice, at some cost in efficiency, when those assumptions are in genuine doubt.
5. Interpretation of test results
As with any hypothesis test (developed formally in Chapter 7), interpreting a non-parametric test’s result means relating its p-value to a pre-chosen significance level, and — importantly — restating the conclusion in terms of the ranks or signs actually analysed (e.g., “the median of population A is significantly different from that of population B”), rather than incorrectly claiming a difference in means, which these particular methods do not directly test.
Key Takeaways
- Non-parametric methods avoid assuming a specific underlying distribution, at some cost in statistical power compared to a valid parametric alternative.
- Rank-based tests (Mann–Whitney, Kruskal–Wallis) are robust non-parametric analogues of the t-test and ANOVA; sign-type tests use only the direction, not magnitude, of paired differences.
- The chi-square test is specifically suited to categorical/count data, testing goodness-of-fit or independence between categorical variables.
- Non-parametric tests are preferred for ordinal data, small samples of uncertain distribution, or data with extreme outliers.
- Results should be interpreted in terms of the actual quantity tested (often medians or ranks), not incorrectly generalised to means.