I have ~3,000 subjects from 44 practices (clusters). The practices were randomized to two interventions. I need to test whether intervention is associated with outcomes measured at the subject level, controlling for both practice and subject characteristics and clustering of the outcome within practices. I thought this would be a straight forward application of the mixed command, but I have become seriously confused about where to put the main predictor variable (intervention). My first thought was that the intervention should go in the 2nd level with other practice characteristics:

Code:
mixed outcome subject_characteristics || practice_id: intervention practice_characteristics
But the model essentially calculates 44 estimates of the impact of the intervention (maybe?) rather than a single coefficient, so I'm not clear on how to interpret it.

Does that mean I should put the main predictor (intervention) BEFORE the double bars?

Code:
mixed outcome intervention subject_characteristics || practice_id: practice_characteristics
This yields a coefficient and SE, but it feels like I'm ignoring the variability of the effect across practices. Do I put it in BOTH places?

Code:
mixed outcome intervention subject_characteristics || practice_id: intervention practice_characteristics
I even tried putting it in as a third level (patients clustered within practices within treatments) but that is even less clear. If you have advice on how I should proceed, I will be very grateful.

Thanks

Ben Littenberg
Department of Medicine
University of Vermont