09-04-2014 02:15 PM
Disclaimer: I am in the process of teaching myself mixed effects modeling, so apologies if I've not conceptualized part of the process properly.
I am attempting to model change in blood antigen levels over time vs. infection status for patients undergoing a certain therapy, using proc mixed. I have blood levels at baseline (t = 0) and at several different time points after initiation of therapy. I also have infection status at each of those time points (infection/no infection), and that status can change over time. It is my understanding that, in general, it may be potentially usefully to include the intercept as a random effect (i.e. different intercept for each subject) when searching for the 'best' model. I'm my case however, since I have the baseline measurements, I already know the intercepts for each patient. Is there any way for me to make sure these values are included in the model? Does using the noint option accomplish this? Are there other procs that might allow me to manually specify the intercepts?
On a related note, how does one interpret the effect of infection on antigen levels, when infection status can change over time? Is it simply a matter of never infection vs. any infection?
09-04-2014 03:24 PM
I assume that you are fitting the following:
This is functionally equivalent to:
Both estimate a single variance component due to patient. There is not a different intercept for each patient that can be fixed. The variance component estimates variability among subjectid's in excess of the residual variance. Now, the BLUP for each patient can be estimated, but the baseline value is not the intercept in this sense.
A good read here is SAS for Mixed Models, 2nd ed.
The infection status situation is probably best addressed as per your last statement, but through what is called a means model, it might be addressed as a time dependent factor.
What covariance structure are you fitting for the repeated nature of this design?