I'm looking to analyze data coming from a study for reduction in dependence on an addictive substance. We're looking at 4 subjective measures of clinical dependence called clinical global impression (CGI) scores. They have a domain of at most 7 possible outcomes.
We're interested in looking at continuous differences before and after treatment. We're also possibly interested in a more robust categorical approach, since the distance between categorical scores may be egregiously different from a constant difference in latent addiction levels.
We would like to use the proportional odds model (GLM with cumulative logit link) to do this. In order to measure changes from baseline, should we use the difference in categorical scores as the outcome or adjust for the baseline scores? I'm assuming we include time as a factor. Then to determine treatment efficacy, do we adjust for the interaction of time and treatment and measure the ratio of odds ratios for cumulative decrease in CGI scores?
Lastly, how does one report the results from a proportional odds model interaction term? As I stated it is my best known approach, however, I would be curious to know if there is a more accurate interpretation.
Information on these types of models is found in Categorical Data Analysis, by Alan Agresti ch. 5.