Hi Colin, A couple thoughts--I think GLIMMIX defaults to the G matrix when counting subjects. Not really sure, but that is what it looks like. For repeated measures in time, we often use AR+RE models (autoregressive plus random effect), which gives the correct count of subjects. These look like: proc glimmix data=for_Stats1 ic=pq; by param ; class sex grp_no studyday anml_nbr ; model value = grp_no sex studyday grp_no*sex sex*studyday grp_no*studyday grp_no*sex*studyday cov/ ddfm=kr(firstorder); random studyday /residual type=&covtype subject= anml_nbr ;*group=grp_no; %if "&covtype" = "AR(1)" OR "&covtype" = "ARH(1)" %then %do; random intercept/subject=anml_nbr; %end; ods output fitstatistics = &outdata._a convergencestatus = &outdata.status ; run; You may need something similar to the second random statement to get the subject number where you want it. My other thought is: Why AR(1) as a structure? I would really go with the spatial covariances like sp(exp) or sp(pow), as you indicated above, given that the indexing reflects some sort of measurement. These are a lot better at handling missing values (my opinion only). The other alternative might be a heterogeneous compound symmetry (CSH), as both AR(1) and the spatial covariances really depend on the indexing. If things are unevenly spaced in reality, but indexed as 1, 2, 3, ..., you are making an assumption about the correlations that might be misleading. Good luck. Steve Denham Message was edited by: Steve Denham
... View more