Well, sorry to be late to answer, but perhaps this will help. In SAS for Mixed Models, 2nd ed., Chapter 7, there is a great discussion about the use of covariates.  Sometimes the best model has an interaction, sometimes not.  Suppose the F test for the interaction is significant - that implies that the slope coefficients differ depending on the level of the other factor in the interaction.  This has profound implications on comparisons between least squares means in that differences between levels of the other factor are not constant, but instead are dependent on the value of the covariate. Section 7.3 walks through the steps to consider - testing for unequal slopes model, analysis if the unequal slopes model should be used, fitting the equal slopes model if it is deemed adequate, and follow-up analyses for additional information.  I can't suggest anything beyond what is covered there - it is the most succinct and complete chapter in the whole book, in my opinion.   SteveDenham ... View more

@PaigeMiller , the sound you just heard was me slapping my own face.  This is an EXCELLENT observation. The RMSE from PROC GLM is the standard deviation of the residuals.   SteveDenham ... View more

This will be a two-step process.   First generate the dataset containing the residuals by adding two options to the MODEL statement (outp= and residual)   ods output lsmeans=estimate diffs=diff; proc mixed data = mydata (where=(treatn in (1,2,3))); by treat avisit; class treatn; model aval = treatn avisitn base/outp=predicts residual; lsmeans treatn / diff=control('3'); run; ods output close; Then, use PROC MEANS (or SUMMARY or UNIVARIATE or SQL) to calculate the SD for the residual variable.  Off the top of my head, I don't recall the variable name for the residuals, but it probably defaults to something like 'resid'.   proc means data=predicts noprint; var resid; output out=resid_sd std(resid)=std_resid; run; If you are looking for standard deviations by treatn, you can add a CLASS statement.   SteveDenham ... View more

That really does look like an ANTE(1) structure.  Consider a matrix where the "non-significant" correlations are set to zero (or are at least diminishing rapidly).  That would result in means that are separated in time by more than about 3 periods to start looking "independent", which is what you see here.  Look at the parameterization of an ANTE(1) structure in the documentation and you'll see that it builds in heterogeneity of both variance and covariance.  It is like an unstructured matrix parameterized in a way that saves on the number of parameters estimated, but also shows a "ridge" that gets broader for later time periods.   Does that even make sense?   SteveDenham ... View more

I am a big believer in the use of corrected AIC in selection of covariance structures.  In this case it is really a clear-cut case for using ANTE(1).  The next best model (AR(1) + RE) only retains 0.006% of the information in the ANTE(1).   So, the most parsimonious structure isn't always the best.  Finding a best fitting structure is, at least at this point, a random search through candidate structures.  You need to look at all of them, considering the processes that generate the data.  One other thing to consider-the autoregressive and ante-dependence structures are contingent on the distance between time points being constant.  If this is not the case, then those sort of structures should probably not be included in the options to choose from.   SteveDenham ... View more

Hi @Malthy  Those records will not contribute to the fit of the model, just as in STATA.  You could impute them using PROC MI and then model average, but that is not for anyone but an experienced user, as feeding in the correct values and variance-covariance matrices is not simple.   Note that the mixed model procedures in SAS will not delete records with missing values for the random variables or the dependent variable.   SteveDEnham ... View more