I agree with Steve. Also, you need to separate modeling the mean part (what terms of time to include... linear, quadratic, etc) and modeling the covariance part. The book by Fitzmaurice et al has a good discussion of steps to follow when modeling the mean and covariance components of GLMM. I usually like to identify the 'best' mean model under a more unstructured covariance model and then work towards a more parsimonious covariance model using my 'best' mean model. Hope this helps. Mulugeta ... View more

Dears Steve and Sandhu, While I do not have access to your data to make more concrete suggestions, I am not clear why you need to structure the R-side covariance matrix to the extent of what you have "type=sp(pow)(xtime)". Is the outcome variable, pain assessment using 100 mm VAS scale, measured  with a lot of error that could not be accounted for using a diagonal covariance matrix as in "random _residual_/ sub=id type=VC;"? I would go with a well structured G-side covariance matrix with the default R-side (you still have a non-diagonal V matrix that accounts for both with-in and between group covariation, BTW) for both model parsimony and computational simplicity purposes. Thanks. Mulugeta ... View more

I just saw this question and thought the following might help. If you would like to fit the model without an R side matrix (without the random _residual_ statement in your code) which will lead to a random coefficient model with a G side matrix, I think the non-convergence problem might be resolved using NLOPTIONS. With the given data example above, the following code leads to convergence with four iterations. However, the problem of nonconvergence with very large datasets still remains to be a problem. You have to of course check if that is what you want. proc glimmix data=subtype; class center; model y = height / dist=bin link=logit solution; random intercept height / type =un subject= center; nloptions tech=NRRIDG; run; ... View more