Every semester I watch grad students run a t-test or ANOVA the moment their data is entered — then get flagged in peer review for skipping assumption checks. Here's the checklist I actually use before touching any parametric test, condensed from what committees ask about:
Sample size vs. group count. Below roughly 15-20 per cell, parametric tests get shaky — consider the nonparametric equivalent (Mann-Whitney instead of independent t, Kruskal-Wallis instead of one-way ANOVA) before you finish collecting the rest of your sample.
Normality. Run Shapiro-Wilk on residuals, not raw scores, for regression/ANOVA. For n > 50, eyeball a Q-Q plot instead — Shapiro-Wilk gets oversensitive to trivial deviations at large n.
Homogeneity of variance. Levene's test for t-tests/ANOVA. If it fails, don't switch tests — use Welch's correction instead.
Independence. The one assumption no software checks for you. Repeated measurements on the same subject, or nested data (students within classrooms), violate this even when everything else passes — you need repeated-measures or mixed models instead.
Effect size, decided before you see the p-value. Committees ask for Cohen's d / eta-squared / r regardless of significance. Compute it as part of your plan, not as an afterthought when p = .06.
None of this needs paid software — R's shapiro.test() and leveneTest() (car package) cover 1-4 for free. If you'd rather have the checks run automatically alongside the test and the write-up formatted in APA style, StatMate's calculators (statmate.org) do that step for you — but the checklist above holds regardless of what tool you use.
Top comments (0)