Use of baseline as a covariate in GLIMMIX

[ANKH1]

Hello, 

Our study is a repeated measures design where APP (dependent variable) is measured at different time points (0, 30, 60 and 120) in response to different treatments (TRT).  We are using time 0 as baseline (BL) and treating it as a covariate. 

PROC GLIMMIX DATA=datlong;

class ID TRT TIME ;
model APP = TRT|TIME BL;
Random TIME /residual subject = ID type=CS;
LSMEANS TRT|TIME/adjust=tukey;
RUN;

When run like this, the tukey comparisons make no sense when looking at the lsmeans. However, when we removed BL from the model, the tukey's makes sense. Why is that? Is it possible that I have to include time 0 as part of the time and not as a covariate?

Any help is deeply appreciated. 

Thank you 

[sld]

Help us out here: in what way do the Tukey comparisons make no sense? It could be useful to post an example dataset that illustrates the problem.

[ANKH1]

Thank you for looking into this. This is the output and means with BL in the model:

And this is the Tukey's when I removed BL from the model. Comparisons make sense in this case

[sld · Accepted solution]

I would focus more on the lsmeans estimates, rather than the results of the Tukey comparisons, because...

In the model that includes BL, the lsmeans are estimated at the overall mean of BL (averaged over observations) by default. (You can control the value using the AT option on the LSMEANS statement.) If the overall BL mean is similar in value to the BL mean for a treatment group, then the treatment lsmean from the model with BL will not be much different than the treatment lsmean from the model without BL; i.e., the adjusted mean will be similar to the unadjusted mean. So, I'm guessing that mean BL varies among treatment groups, in which case you may want to ponder group-mean centering to obtain sensible adjusted means and comparisons among them.

Let me know if that seems plausible.

[ANKH1]

Thank you very much for your response.  To see if I had differences at BL, I ran an one-way anova but they were not statistically different. I should mention that this is a crossover design (so all subjects are exposed to all treatments). I just checked the raw data and it looks like one subject's BL when exposed at one particular treatment is higher compared to its other baselines (this only by looking at box plots). How should I approach this? 

Thanks again!

[sld]

The fact that this is a crossover design is a really important detail, and the information you've provided so far about the design does not tell me enough about your experimental design that I can suggest an appropriate statistical model. You could be (and probably are) looking at a relatively complicated statistical model, and the complexity may be beyond what can be dealt with in this forum. If you have access to statistical support at your institution, I would definitely recommend that you pursue that.