I understand that this coding allows for a decrease in the slope from baseline to the post-baseline measurements and then all post-BL measurements are the same.
As I understand it, the random intercept term in this model refers to the (conditional) post-BL status. We are also interested in the (conditional) BL status, which I tried to obtain by recoding the slope coefficients to:
I was expecting that the the model chi-square and slope results would be the same across the models (which they are for unconditional versions of the two models), but they aren't. Can anyone tell me why the results are different across these two models and what I would need to do to obtain estimates of the regression of initial status on the ncogen1 covariate and the regression of the ncogen2 distal outcome onto initial status from the first model? Thanks!
Your expectations are correct but you need to add "int with slope;" in both models.
Tor Neilands posted on Tuesday, December 16, 2014 - 11:03 pm
Thanks! The log-likelihoods, chi-square, mean slope, etc. are now the same for the two models as expected. I have one remaining confusion: Why are the regressions of the distal outcome identical for the intercept-as-predictor, but different for the slope-as-predictor?
NCOGEN2 ON INT -0.019 0.010 SLOPE 0.001 0.007
NCOGEN2 ON INT -0.019 0.010 SLOPE 0.020 0.012
I would have thought that with the intercept latent variable representing different status points, the regressions of NCOGEN2 on INT would differ, but the regressions of NCOGEN2 on SLOPE would be the same since the slope is the same in the two models.
The fact that the slope stays the same doesn't mean that the coefficient for it stays the same. Since it is a multiple regression the coefficient depends on the covariances with the other predictors and that changes.