When the Assumptions Won't Hold
Real-world data is messy. Heavy tails, skew, contamination by outliers, sample sizes too small to reliably check normality — all of this is the norm, not the exception.
Nonparametric tests don't require assumptions about the underlying distribution. Their power is generally lower than parametric tests when parametric assumptions hold — but they're robust when those assumptions don't. Often, a nonparametric test is the responsible choice.
Nonparametric Tests Work on Ranks
Most nonparametric tests work by converting data to ranks — the 1st, 2nd, 3rd largest value — rather than operating on the raw values. This makes them robust to outliers and non-normality, since a single extreme value only contributes rank 1 (or rank n), not its actual magnitude.