whereas the approximate type one error rate is 0.0221 which is too conservative for a
ratio of σ1 :…:σnf = 3:…:1).
Using permutation tests with small sample sizes and a high number of permutations (e.g.
10000) is problematic. Because of the rather low number of possible rearrangements of
data the probability to create all possible datasets for several times is high. So it would be
better to use exact probability values, but this is not available in the coin package for
more than 2 factor levels.
In Table 2 an equivalent situation as in Table 1 is presented. In difference to Table 1 the
ratio of standard deviations was fixed to 3:1:…:1.
Based on the results of Table 2 none of the presented methods except Hotelling’s T 2 test
can be recommended.
As already Box (1954) pointed out, the influence of heteroscedasticity is much higher in
unbalanced designs than in balanced ones. In the following some results for unbalanced
designs are presented. Table 3 illustrates the situation if all standard deviations are 1
except for the first one which is 3 and all sample sizes are 3 except for the second one
which is 5.