Should i use a paired t-test or wilcoxon signed-rank test for small n=15?

[Tyler10]

I’m trying to decide whether to use a paired t-test or a Wilcoxon signed-rank test for my pre/post intervention analysis on a small sample (n=15). My data is continuous but the post-test scores don’t look normally distributed in the Q-Q plot, which makes me question the paired t-test assumption.

[EvelynAG]

I've been there. With n=15, the paired t-test assumes the differences are roughly normal. If the Q–Q plot for post scores looks off, I start by checking the differences themselves. Sometimes the post scores look skewed but the differences are not too bad, and the test still behaves reasonably; other times the nonparametric alternative (Wilcoxon) seems safer. In small samples, the t-test can give optimistic p-values when the distribution is skewed.

[BrianL]

I actually computed the differences and then ran both tests anyway, to see how sensitive things were. In one project, the Wilcoxon test gave a p-value around .08 while the t-test gave .03; confidence intervals were not directly comparable though; I ended up reporting both results with a caveat.

[Kenneth_A]

I wonder if a simple transformation like a rank-based approach or a robust statistic would help; I tried a log transform on the post scores when they were strictly positive; it helped the Q-Q a bit but interpretation got murky.

[Timothy_M]

Do you care more about the exact p-value or about the direction and magnitude of change in the scores? If your focus is practical impact, the data's noise and what the instrument actually measures might matter more than a formal test.