Regarding normality: As big as this dataset is, even trivial differences will turn up significant for most of the tests of normality. I think you would be better off looking at QQ plots. What you probably have are some extreme values (I hesitate to call them outliers) that might be influencing these tests. I would say that an assumption of normality is justifiable. So, on to fitting the model. If the residuals are normal, then for a model this complex, you may wish to look at HPMIXED. Perhaps something like: proc hpmixed data=kurt; CLASS BLOCK SITE YEAR TRT DIST rep week; MODEL MOISTURE = TRT DIST week trt*week trt*dist trt*dist*week; random block site year block*site block*year block*site*year; repeated week/ residual TYPE=AR(1) SUBJECT=REP(TRT*DIST); test trt dist trt*dist; run; While HPMIXED does not accept the STORE statement, you can pass the covariance parameters to GLIMMIX, and use the results. See Example 46.3 in the HPMIXED documentation for STAT12.3 as a start. Steve Denham
... View more