Add method=quad or method=laplace to the proc glimmix statment.  Then in the random statement, rearrange so that you have subject=rep, something like: random intercept spd*cv/subject=rep type=<insert the type of covariance structure you wish to impose>; Now, you will have to make several runs with the different error structure types.  For a split plot, type=vc and type=un are the only ones I can think of that really apply, unless you have some sort of spatial correlation that you wish to fit. By using the integral based methods, you get quasi-likelihoods, and the information criteria (AIC, AICc) are appropriate for ranking the models. Steve Denham ... View more