A friend of mine approached me to help her to interpret her multinomial logistic regression model. They had measured people as 1 of 2 states at 2 time periods. So, each person can have 1 of 4 configurations: start at state 1, end at state 1; start at state 1, end at state 2; start state 2, end state 1; start state 2, end state2. They had performed an analysis where the outcome had 4 levels corresponding to each of these configurations. This struck me as a somewhat unnatural way to do this. I feel that this is a situation similar to if you had a continuous outcome, measured at 2 time points. If you were interested in the change from baseline, you can model the end value and adjust for the baseline and other covariates.
Does this translate to logistic regression? Can you model the end state, and adjust for the beginning state with other covariates? I couldn't think of a good reason why this would not work, but I also had this problem. Is this better than modeling the outcome as 4 possible categories?