I am trying to fit a mixed model which includes fixed effects, random effects and a command to account for repeated measurements on the same subject, using the SAS macro - glimmix. I am testing whether I need to transform a continuous fixed effect variable but I am getting strange results. When I add the squared term to the model the -2 Log likelihood increases, whereas I would have expected it to decrease since I have added an extra parameter to the model. Has anyone else come across this?