08-14-2012 11:00 PM
I am attempting to model reaction time data using the Local =exp() option in Proc Mixed. Doing so, should model the residual variances as exponential functions of the covariates I enter. I.e. I am modeling the natural log residuals. My data are positively skewed, which many people might say need a transformation before analysis (like the natural log). So my question is...since I am already modeling the residual variance on the natural log scale, is it safe to use untransformed, positively skewed data as the DV? Has anyone else used local=exp for this type of data.
08-15-2012 09:17 AM
If you believe your data are skewed and that a log transform would improve things, you may want to look at PROC GLIMMIX rather than PROC MIXED. Using a log link (and not specifying a distribution) will do everything that I think the local=exp() in PROC MIXED does. Admittedly, GLIMMIX does not have anything like a local= option--I think you would have to specify a factor analytic covariance type to get the ridging done by the local option, and then it still depends on the class variable. Heterogeneity can be modeled for any of the class variables. If the heteroskedasticity is obvious when plotted against a continuous variable, then appropriate link/transform is the way to go, in my opinion.
So, can you share your PROC MIXED code? From that, it should be possible to tell if GLIMMIX offers anything that might help.