Nonparametric Statistics

Nonparametric statistics are a class of statistical methods that do not assume the data follow a specific probability distribution, such as the normal distribution. They are based on ranks, signs, or other order based properties of the data and are suitable for analyzing ordinal data, small samples, or data that violate the assumptions required by parametric methods.

Parametric methods such as the t test and ANOVA assume that the data are drawn from a known distribution (usually normal) with specific properties. When these assumptions are not met, whether due to small sample sizes, skewed distributions, ordinal measurement scales, or the presence of outliers, parametric tests can produce misleading results. Nonparametric methods provide valid alternatives that are less sensitive to distributional violations. They are essential tools for clinical researchers analyzing ordinal outcome scales (such as Likert items or pain ratings), for ecologists working with count data that violate normality, and for any analyst whose data do not meet the stringent requirements of parametric inference.

Nonparametric methods typically operate on the ranks of the data rather than the raw values. By replacing observed values with their positions in the ordered sequence, these methods limit the influence of extreme values and do not assume a particular distribution, such as the normal. The price of this flexibility is a small loss of statistical power compared to parametric tests when the data are actually normally distributed (for the Mann Whitney U and Wilcoxon signed rank tests, about 5% less efficient in large samples).

Common nonparametric tests include the Wilcoxon signed rank test (paired comparison of two related groups, analogous to the paired t test), the Mann Whitney U test (comparison of two independent groups, analogous to the independent t test), the Kruskal Wallis test (comparison of three or more independent groups, analogous to one way ANOVA), and the Friedman test (comparison of three or more related groups, analogous to repeated measures ANOVA). For correlation, Spearman's rank correlation replaces Pearson's correlation when the relationship is monotonic but not necessarily linear, or when one or both variables are ordinal.

Nonparametric confidence intervals and effect sizes can be calculated using bootstrap methods, which resample the observed data with replacement to estimate the sampling distribution of a statistic without distributional assumptions. Permutation tests construct the null distribution by rearranging the data across groups, providing exact p values (when every rearrangement is counted) that assume only that observations are exchangeable under the null hypothesis, not a particular distribution.

Use nonparametric methods when the data are measured on an ordinal scale, when sample sizes are too small to verify distributional assumptions, when the data are clearly non normal (heavily skewed or contain influential outliers), or when the parametric assumptions have been tested and found to be violated. They are a common choice for analyzing single Likert items, pain scores, satisfaction ratings, and other bounded ordinal outcomes. When sample sizes are moderate to large and the data are approximately normal, parametric tests are preferred because they are slightly more powerful. For continuous data with large samples, the central limit theorem means that parametric tests of means are often approximately valid even with non normal data.

A common error is assuming that nonparametric tests make no assumptions at all. While they do not assume a specific distribution, they do assume independence of observations (of pairs, for paired tests), and some tests assume that the distributions in compared groups have the same shape (differing only in location). Another mistake is defaulting to nonparametric methods when parametric methods are appropriate, sacrificing power unnecessarily. Some researchers apply the Kruskal Wallis test and then use parametric post hoc tests, which is inconsistent. Failing to report effect sizes alongside nonparametric tests (because the researcher is unsure how to calculate them for rank based methods) leaves the magnitude of the effect unreported.

A physical therapy study compares pain levels (on a 0 to 10 numeric rating scale) before and after treatment in 15 patients. The differences between pre and post scores are not normally distributed (Shapiro Wilk p = 0.02). None of the 15 differences is zero. The Wilcoxon signed rank test is used, yielding a test statistic of W = 12, p = 0.004 (exact), indicating a statistically significant reduction in pain after treatment. The median pain score decreases from 7 (IQR: 5 to 8) before treatment to 4 (IQR: 3 to 6) after treatment. The matched pairs rank biserial correlation is 0.80, indicating a large effect.