Actually, the two procedures are giving you the "same" results, with slight variations because of different estimation algorithms. The apparent difference is due to the way the two procedures parameterize the class variable. In GLIMMIX, the parameter for the last level of the class variable (1 here) is forced to be 0. In PANEL, the parameter for the first level of the class variable (0 here) is forced to be 0. This reverses sign for the class parameter, and changes the intercept.   Try running GLIMMIX by removing dep_var1 from the class statement. CLASS CUSTOMER_ID  time_period; MODEL dep_var2 = ind_VAR1 ind_VAR2 dep_var1 /LINK = IDENTITY DIST = NORMAL SOLUTION; RANDOM INTERCEPT/ SUBJECT = CUSTOMER_ID TYPE= VC;   You now get the same results as with PANEL. This can only be done when there are two levels to the class variable, coded as 0 and 1.   Note that the random effect variance is 0. That also is contributing to the equivalence of the two methods. ... View more

There are several ways to do this. In general, you will need two runs, one for Poisson and one for geometric distribution, and then compare the -2 log-likelihoods. I demonstrate with PROC GLIMMIX, with my example code. data a; input x count; datalines; 0 10 1 16 2 14 3 21 4 17 5 16 6 13 7 13 8 10 9 6 10 5 11 1 12 2 ; run; proc glimmix data=a; title 'Poisson'; model x = / dist=poisson s; freq count; run; proc glimmix data=a; title 'Geometric'; model x = / dist=geometric s; freq count; run; proc glimmix data=a; title 'Negative Binomial'; model x = / dist=nb s; freq count; run; *another way to get geometric (neg bin with scale=1); proc glimmix data=a; title 'Neg.Bin. with Scale=1 (identical to Geometric)'; model x = / dist=nb s; freq count; parms (1) / hold=1; *hold Scale to 1 to get geometric; run; Note that Poisson and geometric are 1-parameter distributions, so one can directly compare the -2LL values in the output. I also show the fit of the negative binomial, which is a two-parameter distribution which reduces to the Poisson as the Scale goes to 0 (the way GLIMMIX parameterizes it). I bring it up because another special case of the negative binomial is the geometric (when the neg bin Scale = 1). So, I show how to fit this in GLIMMIX also. Note that the results are identical for dist=geometric and dist=nb with the parms (1) statement (last one).   Because negative binomial has two parameters, use AIC to compare fits of different distributions. one can also use other aspects of the fit (e.g., Pearson chi-square) for comparisons, but I won't get into that.   There are other programs that could also be used. ... View more

You definitely want the "residual" in the model statement: this gives you the correlated R-side pattern that would be appropriate for your dataset. That warning message can be caused by many things, and can be data-dependent. Your model is not apparently over-parameterized, so some of the common causes would not apply. For a similar problem (reported elsewhere on the web), the issue was the detection of singularity in solving the mixed-model equations. One solution from SAS was to add the following option to the GLIMMIX statement: singular=1e-9 As in proc glimmix data=.... singular=1e-9;   Another fix is to switch to PROC MIXED. There you use REPEATED statement instead of the random statement.   It seems that you have a different question related to the multiple comparisons of means. To get only the contrasts you want, you could use ESTIMATE, CONTRAST, or newer LSMESTIMATE statements. There is a lot on the web about learning these. In glimmix, you could also use SLICE command: slice time*trt / sliceby=time sliceby=trt lines; This gives comparisons of treatments at each level of time, and comparisons of times at each level of treatment.     ... View more