When the data do not show the characteristics of a distribution from a particular parametric family, non-parametric tests are used instead of parametric ones. One important test of this type is the Wilcoxon-Mann-Whitney test, which is based on the ranking of the data. However, in some cases, classical ranking methods may return paradoxical results, since in case of paired comparisons of distributions with c.d.f. $F_1, \dots, F_a$ transitivity among paired comparisons does not hold. As a solution to this problem, a comparison with the unweighted average of the distribution functions $F_1,\dots,F_a$ is proposed, where $a$ is the number of groups. Methods of this type are called pseudo-ranking methods. In the last part, an alternative effect measure, the so-called win ratio, is presented.
|
