Ranking the data reduces the impact of outliers. Furthermore, given that normality matters only for inferring significance and computing confidence intervals, we could use a bootstrap to compute the confidence interval, then we don’t need to worry about the distribution.
