Nonparametric Tests¶
Statistical test — instantiates Assumption-Light Inference
Compares groups or distributions with distribution-free tests chosen against a named assumption threat, not by software default.
Nonparametric Tests are the packaged significance tests — Mann–Whitney U, Wilcoxon signed-rank, Kruskal–Wallis, the sign test — that answer a comparison question (do these groups differ? is this shift real?) without assuming the data follow a specific parametric distribution. As a mechanism of this archetype, the defining move is not the arithmetic of any particular test but the deliberate act of selecting one in response to a named assumption threat: a t-test's normality assumption looks unsafe here, so a Mann–Whitney is chosen instead, on purpose, with its own residual assumptions acknowledged. The test consumes a distribution-free comparison rule and returns a p-value; what makes it this mechanism rather than a ritual is that the selection is justified by the audit and its power cost is stated, not that the software happened to have the function.
Example¶
An education researcher compares two teaching methods using end-of-term project scores rated on a 1–5 rubric. A t-test on the mean rubric scores is the default, but the scores are ordinal, clumped at 4, and clearly not normal. Rather than force interval assumptions the rubric cannot support, the researcher selects a Mann–Whitney U test, which asks only whether scores from one method tend to rank above the other — a question the ordinal data can honestly answer.
The test reports that Method B tends to produce higher-ranked outcomes, with a significance that survives the fragile-normality worry the t-test could not. The write-up names what was traded: by testing stochastic dominance rather than a mean difference, it gains credibility on ordinal data but loses the ability to say how many rubric points better B is — and if that magnitude were the decision variable, this test would be the wrong tool. That paired justification — chosen against a named threat, with its information cost stated — is what turns a stock test into an instance of the archetype.
How it works¶
- Select against a threat. Start from the assumption the parametric test would violate (normality, interval scale, equal variance) and pick the distribution-free test that sidesteps exactly that threat.
- Reframe the hypothesis. Recognize what the nonparametric test actually asks — typically a stochastic-ordering or median-shift claim, not a mean difference — and confirm that reframed question still serves the decision.
- Run on the distribution-free comparison. The test operates through a rank or sign comparison rule; it reads that rule's output and converts it to a p-value under mild exchangeability-style conditions.
- Book the power cost. Note the efficiency given up relative to the parametric test when the parametric assumptions would in fact have held, so the trade is on the record.
Tuning parameters¶
- Test choice — which nonparametric test matches the design (two-sample, paired, multi-group). The wrong pairing answers a different question; matching the test to the design is the primary dial.
- One- vs two-sided — whether a direction is pre-committed. One-sided gains power when the direction is genuinely pre-specified but invites fishing if chosen after seeing the data.
- Tie and continuity handling — exact versus asymptotic p-values, and how ties are corrected. Exact methods are honest on small samples; asymptotic ones are convenient but fray when n is tiny.
- Effect-size companion — whether a rank-based effect size accompanies the p-value. Reporting one guards against significance-without-magnitude, at a little extra work.
When it helps, and when it misleads¶
Its strength is that it delivers a defensible comparison when the parametric assumptions are shaky, and it does so with surprisingly little power loss under many real distributions — the Wilcoxon–Mann–Whitney test is only marginally less efficient than the t-test even when normality holds, and far more trustworthy when it does not.[n1] For small, skewed, or ordinal evidence it is often the honest default.
Its failure mode is mechanism ritualization: reaching for a nonparametric test because it is familiar or because a reviewer likes it, without an audit establishing that a parametric assumption was actually threatened. Used that way it needlessly discards power and can obscure a magnitude the decision needed. It also carries assumptions people forget it has — many rank tests still assume similar distribution shapes across groups or exchangeable observations, so calling them "assumption-free" is the archetype's cardinal error. The classic misuse is switching to a nonparametric test only after a t-test failed to reach significance, hunting for a p-value across methods. The guarding discipline is to tie the test choice to a named threat before seeing the result and to state the residual assumptions the test still makes.
How it implements the components¶
assumption_light_method_choice— selecting the specific distribution-free test in deliberate response to a named parametric-assumption threat is the mechanism's core act.power_or_information_loss_note— booking the efficiency and magnitude information surrendered relative to the parametric test keeps the trade-off explicit and guards against ritual use.
It selects and runs the test but does not itself define the rank-substitution rule the test computes on, nor build the null by resampling. It does NOT implement rank_or_distribution_free_rule — that belongs to Rank-Based Methods, the transform this test consumes — nor resampling_or_permutation_plan, which is Permutation Tests.
Related¶
- Instantiates: Assumption-Light Inference — supplies a group comparison that survives fragile distributional assumptions.
- Consumes: Rank-Based Methods — supplies the rank transform many nonparametric tests compute on.
- Sibling mechanisms: Rank-Based Methods · Permutation Tests · Bootstrap-Like Checks · Robust Statistics · Median-Based Summaries · Model Comparison Table · Assumption Audit Checklist · Diagnostic Plot Review
Editorial Notes¶
Form Classification¶
Form family: Analysis, Modeling & Optimization
Rationale: Nonparametric Tests operates as an analytical, modeling, inference, comparison, or optimization procedure that derives insight or a solution because it compares groups or distributions with distribution-free tests chosen against a named assumption threat, not by software default.
Independent corroboration: The frozen evidence defines Nonparametric Tests as 'Compares groups or distributions with distribution-free tests chosen against a named assumption threat, not by software default', so its operative form is Analysis, Modeling & Optimization.
Review outcome: Independent reviewer agreement; high confidence.
Origin Attribution¶
Primary origin: Statistics & Experimental Design
Origin pattern: Single lineage
Present-day reach: Multi-domain
Rationale: Mathematical statistics developed rank, permutation, sign, and distribution-free tests for inference under named threats to parametric assumptions.
Review resolution: Both independent reviews agree on primary origin statistics_experimental_design; reconciliation resolves domain_reach_disagreement. Formative alternate lineages retained: none. The broader reach of later applications is kept separate as domain_reach=multi_domain; origin_mode=single_lineage describes the historical relationship among lineages. Confidence is conservatively reconciled to high, and encyclopedia_synthesis=false preserves the reviewers' boundary judgment.
Review outcome: Reconciled after independent review; high confidence.
Notes¶
[n1] The Wilcoxon–Mann–Whitney test compares two groups by ranking all observations together and asking whether one group's ranks tend to exceed the other's. Its asymptotic relative efficiency versus the t-test is about 0.95 under normality — a small price — and it can be far more efficient under heavy-tailed data, which is why it is a defensible default when normality is doubtful. ↩