You could be running into a quasi-separation situation, such that one time-point is all in one response category.  Then CS has a hard time fitting that point in, lots of tries are made, and eventually you do blow things up.  Only explanation I can come up with. Other approaches--try it in GLIMMIX with method=laplace, fitting the repeated nature as a G-side matrix rather than as an R-side.  With only two time points, you may wish to consider an unstructured covariance matrix.  If so, try type=CHOL to get the Cholesky root, which is at least positive semidefinite, and will help avoid convergence issues. Steve Denham ... View more