Since you have created your own dummies, not sure you want those on the CLASS statement.  I think you are pretty close with your ESTIMATE statement. A common mistake is to not include the intercept term in the estimate. I ran the following simulation data test; call streaminit(9871534); do id = 1 to 50; x1=rand("normal"); trt=ceil(rand("uniform")*2); do time=1 to 3; x2=rand("normal"); if time=1 and trt=1 then trt1=1; else trt1=0; if time=2 and trt=1 then trt2=1; else trt2=0; if time=3 and trt=1 then trt3=1; else trt3=0; y=x1 + x2 + trt*time + rand("normal"); output; end; end; run; proc mixed data=test; class time; model y=time trt1 trt2 trt3 x1 x2; estimate "trt1 and time1" int 1 time 1 trt1 1 / e; lsmeans time / at trt1=1 e; run; The ESTIMATE statement gives you estimate you suggest in your write-up. Notice, though, that the covariates for X1 and X2 are zero-ed out. If you want to evaluate the estimate at values other than 0 for your covariates, you would need to add those to your ESTIMATE statement. The LSMEANS statement is doing something similar, but evaluating all the covariates at their means. Including the covariates TRT2 and TRT3.  ... View more

If CLASSROOM is a random effect, then you do not list that effect on the MODEL statement. Only fixed effects appear on the MODEL statement. If you have multiple observations for each student in a classroom, then you could try ID(CLASSROOM) as an additional random effect. If you only have one observation per student, then you do not have enough data to include ID(CLASSROOM) in the model.    Yes, you can include both student-level and classroom-level covariates in the model. GLIMMIX will assign the correct df to those effects based on how they change or do not change within levels of the random effect. ... View more

The test formed by the LSMESTIMATE statement is in the form KLB, where K is the matrix formed by the rows of your LSMESTIMATE coefficients. B is the parameter estimate vector that you can see by adding the SOLUTION option to the MODEL statement. Looking at the results of the ELSM option and the SOLUTION option shows you how your LSMESTIMATE coefficients line up with the elements of the parameter estimates vector and will tell you if you have the correct set of coefficients for your test.   It would be unusual to have 1's on the levels of the coefficients that you are not interested in. We usually see 0's there. Also unusual to see 11 and -11 on the levels you want to compare. We usually see 1 and -1 there. But that might have to do with the weighting you have for this data.  ... View more

What @SteveDenham said. You will need a lot of data to segment out the slopes into 10 6 month intervals.  ... View more

You will need to be careful with the merging. If the merging is done correctly then SSCOEF will have the slope for each subject. You might want to take out the intercept terms in the two SOLUTION tables before doing the merge, since that is not of interest to you.  ... View more

With this mixed model:   proc mixed data=mock_dataset; class ID; model eGFR=time/solution; random int time/subject=ID type=un solution; TIME is a covariate and EGFR is the dependent variable. You can get the slope on TIME for each level of ID by combining the results from the two SOLUTIONs tables. SOLUTIONF, from the MODEL statement, gives you the overall slope on TIME while SOLUTIONR, from the RANDOM statement, gives you the adjustment to that overall slope for each level of ID.    If REGISTRY is a nominal or ordinal variable, then you can use PROC GLIMMIX to model that using a multinomial logistic regression. Put COVA on the MODEL statement as a predictor, and GLIMMIX will use the correct DF to test that effect. You do not need an option to specify COVA as a time-varying covariate. GLIMMIX will detect that.  ... View more

Conover's Technometrics article from 1979 or so introduced the idea of Latin Hypercube Sampling.  ... View more

That code should do what you want. I am not sure how you will apply the DOSC effect, though. TYPE is your repeated effect in the design you describe. If you have missing data, meaning that you might only collect one sample for some subjects, then include TYPE on the REPEATED statement before the / to make sure you identify the TYPE observations correctly for each subject.  ... View more

This paper  provides some guidance when dealing with the warning message you receive with your model and data. You can modify the INITITER= option on the PROC GLIMMIX statement to use more than the default 4 iterations to find starting values for this model. You might try specifying your own starting values as well using the PARMS statement.    If you have not done so already, try PROC SGPLOT to plot the trajectories for each subject. Use week as your X variable and your response as the Y. You could group the id's by the treatment they received, too. This plot would give you some idea as to how the response varies over the weeks and if your model is appropriate for this data.  ... View more

Do you want to treat block*id and block*tmt*id as random effects? That is what your RANDOM statement is doing. Do id's see different levels of block and block*tmt? If not, then take SUBJECT=ID off the RANDOM statement to fit block and block*tmt as random effects. RANDOM int tmt / subject=block will do that. ARH(1) fits a covariance parameter for each week. It appears that you have 120 animals in the study, so the full data set has 15 weeks (120*15=1800)? If so, that number of covariance parameters should not pose a problem, unless demand on the server is high. You could try AR(1) to see if you can get that to fit. Knowing the number of levels to each class effect will help get you a better answer from the community, too. ... View more

Finding the cause of the non-estimable lsmeans will be difficult without the data. You can add the /E option to the LSMEANS statement to see the linear combination used to estimate the problematic lsmean. You can let GLIMMIX do the dirty work of forming the spline basis using the EFFECT statement as in this example Fitting random coefficients on top of the spline fit might be more than you can do with this data. If you do want those random coefficients for each subject, do you want those on the spline effect rather than a linear effect of days? That might be more than the data can handle, too. You could try fitting just the random intercept and see if that does better for you.  ... View more

I'm in favor of your second model, with random clinic effect and repeated measures on the subjects. We see that model quite a lot in interactions with other SAS users.  Your coding of the subject effect is giving MIXED fits with memory when you include SUBJID(CLINIC) as a random effect. That is trying to fit a 160x40 level interaction. But that is not what you have. Recoding the subject effect so that you have subjects 1,2,3 in clinic A and subjects 1,2,3,4 in clinic B will save you a lot of memory ... and fit the model correctly. By nesting subject within clinic, MIXED will know that subject 2 in clinic A is different from subject 2 in clinic B. Now, in that second model you can use SUBJ(CLINIC) as the SUBJECT= effect on the REPEATED statement and possibly get better behavior out of your DF calculations.  ... View more