Nonparametric statistics

Nonparametric statistics replaces strong distributional assumptions with weaker conditions suited to the data and research question. A t test, for example, is commonly motivated by a normal-error model, whereas the Mann–Whitney U test compares the relative ordering of observations from two groups. The term “nonparametric” does not mean that parameters are absent; some methods estimate medians, quantiles, or other population features, while others test distributional differences directly.

The approach is especially appropriate when measurements are ordinal, when a histogram shows pronounced skewness, or when a sample is too limited to assess a model’s assumptions confidently. It can also complement parametric analysis: transforming data, fitting a robust model, and applying a rank-based test answer related but not identical questions.

The main procedures differ according to the number of groups, the pairing of observations, and the target of inference. The Wilcoxon signed-rank test evaluates paired differences, while the sign test uses only whether each difference is positive or negative. For several independent groups, the Kruskal–Wallis test extends rank comparison beyond two groups; for repeated or matched groups, the Friedman test provides a corresponding rank-based design.²

Permutation tests generate a reference distribution by rearranging labels under a null hypothesis, and bootstrap methods repeatedly resample observed data to quantify uncertainty. Rank correlation measures such as Spearman’s rho and Kendall’s tau assess monotonic association without requiring a linear relationship or normally distributed variables. Software implementations often report an exact or asymptotic p-value, depending on sample size and ties.³

Rank tests usually assess ordering or distributional separation, not automatically a difference in means or medians. The Mann–Whitney procedure has a simple location interpretation only under additional conditions, such as similarly shaped group distributions; with unequal shapes, a significant result may reflect dispersion or broader distributional differences.¹ Likewise, a small p-value indicates incompatibility with a specified null model, not the magnitude or practical importance of an effect.

Ties, zero differences, dependence, and unequal sampling probabilities require explicit treatment. Exact calculations can be preferable for small samples, but ties may require specialized algorithms or approximations. Effect sizes, confidence intervals, graphical displays, and the sampling design should accompany the test result. Multiple nonparametric tests can still inflate false-positive risk, so multiplicity adjustment remains relevant.

Nonparametric analysis includes more than familiar rank tests. Quantile regression models conditional medians or other quantiles, making it useful when effects differ across the outcome distribution; semiparametric survival models, including the Cox proportional-hazards model, also avoid specifying a complete baseline hazard distribution. Empirical-distribution methods can compare entire samples, while robust estimators such as the median and trimmed mean reduce sensitivity to extreme observations.

Rank methods are often more efficient than parametric alternatives when their assumptions fit, but they are not universally safer or more powerful. Under a well-specified normal model, a t test can outperform a rank test, particularly when the target is a mean. The practical choice therefore depends on the estimand, scale, design, distribution, missingness, and whether the analysis seeks a global distributional comparison or a specific location effect.

Rank test

A procedure that replaces numerical observations with their order positions before calculating a test statistic.

Permutation test

A test that constructs a null reference distribution by rearranging observed labels or assignments.

Ordinal data

Measurements whose values have a meaningful order but not necessarily equal intervals between categories.

Semiparametric model

A model combining a finite-dimensional parameter component with an unspecified or flexibly modeled distributional component.

Robust statistic

A statistic designed to be relatively insensitive to outliers or modest violations of model assumptions.