Thanks for the numerous answers to my question on non-parametric
mean-sensitive tests. I have one further remark concerning Martin
Bland's answer:
> Apologies for applying to the list, but I think the comments about
> medians in rank tests may be misleading. As far as I can see,
> rank-based tests are only about medians under the assumption that the
> difference between means is the same as the difference between medians
(two-sample,
> distributions same shape, differ only in location).
For differently-shaped distributions the Wilcoxon-Test can also be
median-sensitive if the difference in means is *different* to the
difference
in medians (and it's exactly this case where median- and
mean-sensitivity
become distinguishable). Example:
sample1: 1 2 3 4 5 6, sample2: 1 2 3 4 100 1000
mean1 = 3.5, mean2 = 185
median1 = 3.5, median2 = 3.5
significance from Wilcoxon-Test: 0.9805
Volker
|