> I am working on some data that is giving me fits. I have problems with > heterogeneity of variance. I have already omitted the outliers and have > tried every transformation in the book as well as weighted least squares. > The largest variance is associated with the smallest groups. My thought is > to do a nonparametric test - Kruskall-Wallis ANOVA. This of course does not > tell me where the difference is between by 7 means or mean ranks here now > with the K-W. My thought is to do the Games-Howell post hoc which assumes > heterogeneity of variance with my DV in ranks since that is the K-W does > ranks. > > Does this sound reasonable or are there better ways to go about this? What's your design? By your considering a Kruskall-Wallis, I'm assuming a single-factor between-subjects. How have you established heterogeneity of variance? Did you try a power transformation? Remember that the Kruskall-Wallis assumes that the distributions for all groups are similar--it doesn't matter what shape they are, as long as they're similar. Assuming that you make the decision that non-parametrics are the way to go, I would skip the Kruskall-Wallis entirely, and do a series of Mann-Whitney's using Holm's Sequentially Rejective Bonferroni procedure, although with seven groups--gee, that's a lot of pairwise comparisons (21 to be exact), which would make the critical significance level of the first pairwise comparison .0024. Still, the biggest difference is probably going to be <.001, so you'll get something. If you would like me to walk you through the Holm's procedure off-list, let me know. John +----------------------------------------------------------------+ |John Reece, PhD | |Dept. of Psychology & Intellectual Disability Studies | |RMIT University Phone: +61.3.9925.7512 | |PO Box 71 Fax: +61.3.9925.7303 | |Bundoora VIC 3083 mailto:[EMAIL PROTECTED] | |AUSTRALIA http://www.rmit.edu.au/departments/ps/reece.htm | +----------------------------------------------------------------+
