I almost never look at the unstructured/Cholesky structure unless there are fewer than 5 timepoints, since most of the data I work with have a limited number of subjects. Remember, 5 timepoints require estimation of 15 parameters. If your data is "intense", say with 20 timepoints, you are talking 210 parameters. Using a rule of thumb I came across in some R vignette, that would require roughly 2100 subjects to be able to get stable estimates (R side) or have a positive definite G matrix (G side).
So, I tend to think about the process that generates the data. If you can exchange any two time points, then compound symmetry type covariance structures are good. If there is some sort of correlation between errors, then the autoregressive or spatial structures are usually a better choice. One thing is that if the data are unchanged and the errors in PROC MIXED are normal, you can use information criteria (AIC, AICc, BIC) to rank the models in order of retained information. For generalized mixed models, it becomes more difficult, but as long as you deal with single parameter distributions, this method is well defined.
For more information, I recommend Walt Stroup's Generalized Linear Mixed Models: Modern Concepts, Methods and Applications, as well as SAS for Mixed Models, 3rd edition.
SteveDenham
... View more
