Of course we would like to see that the dependent variable is correlated with the independent covariate. Otherwise, it really doesn't add anything to the analysis. The problem I see in this case is that some of the correlation is artificially induced mathematically.  I believe the problem gets worse when there is an interaction between the covariate and any categorical variables in the model - there is a greater chance of calling regression to the mean a covariate by class level interaction.   Shoot - it's just that I was taught someplace along the line to not fit both a change score (equivalent to a regression coefficient of -1) and the pretreatment baseline (with regression coefficient = beta_hat).   SteveDenham ... View more

The covariance parameters.  Be sure they are in the correct order.  Take a look at the examples for the PARMS statement in the GLIMMIX documentation.   Based on what you started this with, I would try:   PARMS (0.2813) (0.6921);   SteveDenham ... View more

From Frank Harrell's online version of Regression Modeling Strategies, I pulled this from the annotated bibliography:   Kung-Yee Liang and Scott L. Zeger.“Longitudinal Data Analysis of Continuous and Discrete Responses for Pre-Post Designs”. In: Sankhy¯a 62 (2000). makes an error in assuming the baseline variable will have the same univariate distribution as the response except for a shift; Baseline may have for example a truncated distribution based on a trial’s inclusion criteria. If correlation between baseline and response is zero, ANCOVA will be twice as efficient as simple analysis of change scores;if correlation is one they may be equally efficient, pp. 134–148 (cit. on p. 7-5).   That last sentence lets you know that fitting a change score, with the baseline as a covariate, is never as good as fitting the actual data.  Once you get your marginal means for treatment groups, you can calculate change from baseline.   Another way to think about this is to just do some simple algebraic rearrangement.  You have this model:   Change = Bo + B1*Y1 + B2*Grp2 +B2*Grp3.  Plugging in the definition of Change, you get: Y2 - Y1 = Bo + B1*Y1 + B2*Grp2 +B2*Grp3.  Now add Y1 to both sides, and get: Y2 = Bo + B1*Y1 + B2*Grp2 +B2*Grp3 + Y1.  Rearranging terms, this gives: Y2 = Bo + (B1 + 1)*Y1 + B2*Grp2 +B2*Grp3.  Redefine B1 + 1 as B1' and you get: Y2 = Bo + B1'*Y1 + B2*Grp2 +B2*Grp3.  But this holds if and only if there is no correlation between Y1 and Y2.  If there is correlation (and there usually is in most pre/post designs), then the estimates for B2 and B3 will be biased as a function of the amount of correlation in the response variables after removing the true effects of B2 and B3..   SteveDenham ... View more

Hi @edhuang  - This error is, for me, the most frustrating of the "can't get started" errors.  It is saying that the default starting values for all of the covariance  parameters don't allow for a MIVQUE estimate to get started.  This means using the PARMS statement to feed in better values.  The question then arises, "Where do I get better values, and how close do they have to be to get things started?"  I think you are lucky in that your integral method converged and gave you values (which might be biased like most MLE variance estimates).  You could plug those in as starting values.   And if that doesn't help, you can grid search for starting values.  @jiltao or @STAT_Kathleen  might have better ideas.   SteveDenham ... View more

@edhuang , that is what @jiltao was suggesting.  It is in line with a paper by Stroup and Claassen that talks about the linearized method (RSPL) with an R side variance for repeated measures as being usually less biased than the integral methods (Laplace and adaptive quadrature), where the RESIDUAL option is not supported.  What you fit in the original code would be the equivalent of the glme or glmmTMB package in R.   SteveDenham ... View more

In the documentation for the type= option for the RANDOM statement, you'll find the parameterization.  Two factors are included: sigma and rho.  Sigma squared is the variance at each time point (labeled Variance in the table) and rho is the correlation between the variance at any two adjacent time points, and is labeled AR(1) in the table.   SteveDenh ... View more

Echoing @PaigeMiller here, could you provide your definition of "adjusted" mean.  I can think of several ways to adjust - subtract baseline, use baseline as a covariate, calculate marginal means to accommodate unequal sample sizes, weight factors in the model by (number of observations, 1/standard deviation, 1/variance, custom weights, any of these in conjunction with the first two).  Since I am not really familiar with standards in this field, I don't know how to advise.you.   SteveDenham ... View more

Thanks, @Rick_SAS .  I missed that there was no closing parenthesis, so that the exp(yhat) is the only term that transforms back to the original scale in this representation.   So, this code calculates the log-likelihood as -yhat+y*log(yhat)-LFACT(y), and thus is off by a factor of y/2 after removing the identical parts of the calculation (check me here, Rick).   SteveDenham ... View more

I don't know if it is a mistake, but it seems unusual that these two lines are in there:   lk=exp(-yhat)*(yhat**y)/fact(y); ll=-log(lk); That would imply that the log likelihood is just   lk=(yhat)*(yhat**y)/fact(y); and that doesn't seem right somehow.   SteveDenham     ... View more

And to follow up on what @PaigeMiller was saying about an interval, it doesn't depend on the unrealistic assumption that the underlying population values are identical, which is what the hypothesis test p value assumes.  You know a priori that they are not identical, and so whatever p value comes out, it isn't "correct", and you don't know how far away from correct the p value is (or else you could apply some non-central correction).  For some things (say a binomial underlying variable), it isn't even a smooth curve as you move away from identical.   (Thanks and a tip of the hat to my Math Stat prof Dr. Norm Matloff)   SteveDenham ... View more