The assumptions of standard tests, such as the t-test, ANOVA, and ordinary least squares regression, are frequently violated, which may be consequential for the desired error rates. Using robust tests instead does not rely on assuming normality and equal variances. Employing them from the outset bypasses the pitfalls of deciding on the usability of a standard test with data. In this opinion piece, I summarize the epistemic benefits of robust alternatives, like the Mann–Whitney U test or robust linear regression. Restricting on a robust test instead of conducting it in addition to the standard test avoids multiple results, thus counteracts fishing for the desired result, which can occur subtly. It counteracts nonreplicated findings that are just due data anomalies like extreme values and outliers that occur differently across studies. From a practical standpoint, running a single test simplifies analysis, and there are many robust methods readily available in R. However, it is important to understand what a robust method does and what is actually robust against. I also address common defenses of standard tests, discuss why they remain widespread, and suggest how these arguments should be countered.
Michael Höfler (Thu,) studied this question.