I am having some serious problems with my analyses and was hoping someone might be able to help me out!

I have a sample of 95 participants, who have been classified into three groups using a cluster analysis. I used hierarchical cluster analysis for this, and had 26 variables on which I clustered the participants. I am now wanting to run analyses to find out which of these variables are the best predictors for category membership.

First I looked at discriminant analysis. However, due to my sample size I am violating the assumptions of discriminant analysis, namely that my data is not normally distributed and I do not have 5 times the number of IVs for number of cases in each group (IVs=26; DVs: group1= 32; group2=36; group3=27). The output from the discriminant analysis tells me the fit is good. However, how much of a problem is the violation on number of IVs and cases???

As I had violated the assumptions for discriminant analysis, I looked to Multinomial Logistic Regression. Again here I am violating the assumptions of regression: I have incomplete information from the predictors (e.g. 64% of cells with zero frequencies), I have complete separation of the data and it looks like I also have underdispersion. I tried to correct for some of these by selecting the deviance dispersion parameter. I also ran a principle components analysis on the IVs to reduce the number of variables, and used these instead. However I still have complete separation of the data.

Can someone tell me if there is anything else I can do??? Can i use the discriminant analysis as it is? Is there another analysis I can run to assess what I am looking for??

Any suggestions grately appreciated!

Mlb