SAS developers apparently were not interested in determining why you got different results with different versions because all the versions were giving meaningless results, based on the chosen model and data. Machine precision can make a difference when things blow up. I know from SAS developers that 32 bit and 64 bit machines could give different (incorrect) answers when there are numerical problems with a given data set. I think you have an example of this. Results should always be the same when things are "working correctly". The issue of G not being positive definite is a tricky one. It does not mean that things blew up or that one cannot use the results. The safe answer is to change the model whenever you get the warning that this happens, but I think this is a bit of an overreaction based on experience. SAS puts in the warning in the Log to cover themselves, because theory does break down when G is NPD. However, in some cases it does not matter, especially with variance component models (which is not the model that the OP is using). For instance, I have explored this a lot with variance component models when one of the components is 0. I have compared the numerical results (test statistics, expected values, SEs, etc., etc.) for data sets when the 0 variance term was kept in the model and when it was removed. In typical experimental design scenarios, the results are identical. I won't get into the always controversial issue of negative variances (this would be a different model). For more complex random effects, it is trickier to evaluate. You should always think about your results when G is NPD.  If one is getting missing values for estimated SEs of variance-covariance parameters, then you should definitely consider another model. But things can work well numerically with positive semidefinte G.  For instance, MIXED/GLIMMIX gives a nice way to restrict an (almost) unstructed G to be at least positive semidefinite. This is done with type=CHOL or type=FA0(#) option. This can be numerically more stable than type=UN. You are not dealing with this situation, but I just bring it up to give some background. I support the statements in that 2012 Global Forum paper that was cited. All zeros in G are not equally bad. It depends. ... View more