Well, the infinity problem comes from what mentioned regarding model complexity--too many parameters being estimated leads to zeroes being put in for standard errors of the lsmeans. You may have to collapse some of your time categories to get around this, given that obtaining more data is probably out of the question. As far as the AR vs. AR+RI--my preference there stems from fitting Gaussian data (see, for example, Littell et al., Statist. Med. 2000, 19:1793-1819). I honestly don't know if it will lead to better or worse fits for other distributions that are fit as G side models. However, it is an additional parameter to estimate, and it may be causing the infinity problem. Steve Denham
... View more