This is a mixed question, between theory and practice. I have a dataset with a continous variable grouped by a 33 levels factor. After having log-tranformed my original data I can assume the normality of my data but I have two strong departures from the basic assumptions for anova and t tests: *unbalanced data* (some groups contain ten samples, others hundreds) and *non homogenity of variances* (tested with a kruscal test just for a qualitative assessment). Is it possible, and how, to make multiple comparisons when these conditions are met? In past anaylises I've simply used the pairwise.t.test function to do multiple t-test, but know I have to consider the above situation... Thanks a lot, giovanni
I've forgot to cite the Games and Howell procedure, which some literature (eg in "Pairwise Multiple Comparison Procedures: A Review" [1]) referes as a good one for uneual samples and non homogeneus variance. Yet I haven't found an implementation for R... giovanni [1] psycnet.apa.org/journals/bul/96/3/589.pdf 2009/11/30 G. Allegri <giohappy at gmail.com>:> This is a mixed question, between theory and practice. > I have a dataset with a continous variable grouped by a 33 levels > factor. After having log-tranformed my original data I can assume the > normality of my data but I have two strong departures from the basic > assumptions for anova and t tests: *unbalanced data* (some groups > contain ten samples, others hundreds) and *non homogenity of > variances* (tested with a kruscal test just for a qualitative > assessment). > Is it possible, and how, to make multiple comparisons when these > conditions are met? In past anaylises I've simply used the > pairwise.t.test function to do multiple t-test, but know I have to > consider the above situation... > > Thanks a lot, > giovanni >