I am running multiple proc genmod models on a very large wildlife-survey dataset with the "repeated" statement and selecting between models using QIC in SAS 9.2. I am using the negative binomial distribution and AR(1) correlation structure.
I am finding that the difference in the values of QIC and QICu between the top two models (delta QIC and delta QICu) can be very large, over 1000 for one very common species, for some it is around 20 to 50 but for other species it is between 2 and 10, similar to what one might expect using AIC or AICc. The best models (the models with the lowest QIC and QICu) make intuitive sense. Could anyone tell me if these large differences are normal for QIC compared with AIC? A colleague also found the same with a very different (smaller) dataset with a Poisson distributions.