Robustness to non-normality of common tests for the many-sample location problem (Q1769996)

The authors focused on a comparison of two statistical tests of zero-hypotheses on median equality under violation of common presumptions on the probability distribution. The two studied tests are the ANOVA (analysis of variance) and Kruskal-Wallis test, where a utilization of the first one assumes that the processed data are independent, identically distributed normal random variables, while the other test needs not the assumption of normal distribution. This study was motivated by the fact that many practitioners use the ANOVA test even if its assumptions are not fully satisfied. The authors intended to demonstrate the behaviour of both tests under various measures of assumption violations To be able to control the size of violations as concerns the skewness and kurtosis, they used the g-and -k distributions by MacGillivray and Cannon. The core of the study and the comparison consists in enumeration of power the functions for both tests of the zero-hypotheses for various combinations of population means and population distribution deviations from normality. The paper concludes by summarizing recommendations fot the circumstances under which it is better to use the Kruskal-Wallis test and when to employ the classical ANOVA approach.