Toothaker (1991) reported the results of a larger study (Martin et al., 1989) in
which nine multiple comparison procedures were examined under equal sample
sizes but unequal variances. Only the Scheff? procedure maintained the empiri
cal Type I error rate below the nominal .05 level, although the Dunn and Tukey
procedures maintained the error rate across all conditions within Bradley's
(1978) liberal definition of robustness (i.e., empirical Type I error rate less than
.075 for nominal alpha of .05). Even though the other procedures examined
(Shaffer, Ryan, and Peritz) did not maintain the actual Type I error rate within
these robustness limits, the extent to which alpha was inflated was not uncom
fortably large (the maximum value of alpha under these tests was only .087, with
a nominal alpha of .05).