When would you choose a nonparametric method?

Why

Start with the scientific question and data-generating process. Are observations independent, paired, repeated or nested? Is the outcome ordinal, a count, heavily skewed or continuous with outliers?

A Mann-Whitney test, for example, can be interpreted as testing distributional differences or probabilistic ordering, and only under additional shape assumptions can it be read as a pure location-shift comparison. For paired measurements, the structure calls for a paired method. For clustered experimental units, simple rank tests may still violate independence.

Alternatives include generalized linear models, transformations, robust estimators, bootstrap confidence intervals and permutation tests aligned to the randomization design. In a small experiment with non-normal observations, a single Shapiro-Wilk p-value is not an adequate decision rule. For example, a highly skewed concentration outcome might be better represented by a log-scale model estimating a geometric mean ratio if that answers the biology.

Explain the target estimand and method assumptions, not just the name of a test.