Disclaimer: not a statistician; went to school for numerical analysis and haven't actively practiced rigorous mathematics in a couple of years. My guess is that the source of the difference isn't a change in the algorithm, or even the version of SAS, but a combination of a change from a 32-bit SAS (your 9.3 results show an environment of W32_7PRO) to a 64-bit one (9.4 results show X64_S08R2) and some precision hijinks caused by the zero-variance component noted above. To do a little bit of confirmation, I ran the code above (slightly modified to output the g matrix and a detailed iteration history of GLIMMIX) on a single machine with both 32-bit 9.3 TS1M2 and 64-bit 9.3 TS1M2 running on the same Win7 workstation. Same thing: results don't match, and they don't match what was posted above, either. The scary part, to me, is that I see statisticians advocate for leaving zero-variance components in a model, as 'the model takes care of it' - see http://support.sas.com/resources/papers/proceedings12/332-2012.pdf, page 10 for an example. I don't know if I can agree with that sentiment based on what I've seen here. But again, not a statistician. I just know that if you fill a matrix with a number of zero or near-zero entries and then try to work with it from there, you're going to run into issues.
... View more