For a small number of factor levels no method keeps nominal α value. Welch-Test and
weighted ANOVA perform best in this situation, but with a higher number of factor
levels, the results get worse. Kruskal-Wallis test cannot be recommended, because it gets
more conservative with increasing number of factor levels. Permutation variant of this
test seems to become better with increasing factor levels, but with a higher number of
observations (e.g. sample sizes: 5,25,10 observed α=0.118 at the 5%-level) the problem
get worse. Overall no method is appropriate for the situation of inhomogeneous variances
as soon as the biggest sample size does not correspond to the highest variance.
Situation for Table 4 is contrary to that of Table 3 in that sense, as the biggest sample
size (5) corresponds with the highest standard deviations (3).