Before-after measurement and ordinal longitudinal analysis Two types of treatment were considered to relieve pain after surgery. 
Pain is measured as follows:
 first 30 min after first interpleural (IP) injection (hour zero) in recovery room,  then every 4 hours in resting position before and 30 min after IP injection (i.e., hours 4, 8, 12, 16, 20, and 24 postoperatively) in ICU using faces pain scale. 
However, 


*

*The repeated measurements are both before-after and longitudinal (0, 4, 8, 12, 16, 20 and 24 hours).

*The dependent variable (pain) is recorded on an ordinal scale, and we can't subtract  difference before-after use this difference in ordinal logistic models.
Which analysis is better to highlight the difference between the two treatment groups?
 A: This question seems rather closely related regarding the nature of the experiment (i.e. analysis depends on whether this was a randomized study) and the ordinal nature of the assessments (ordinal regression as an option, likely monotonic effect of baseline measurement on subsequent measurements).
The one extra-feature is the repeated post-intervention assessments. That has two aspects to it:

*

*What extra model terms might be needed to describe what is likely to be different between these different assessments? For this, the conventional wisdom (applicable whether the outcome variable is ordinal or not) would be to add a timepoint main effect, a treatment by timepoint interaction, and a baseline by timepoint interaction to a regression model. That would be in addition to the treatment and baseline main effects you would already have.

*How to reflect that measurements on the same subject are correlated (and likely more correlated the closer together in time)? For this, one would typically want to have a random(-subject) effect for each timepoint that would ideally be arbitrarily correlated (aka "unstructured correlation structure"). I.e. the random effect for each patient would be a 6-dimensional vector with a component for each assessment timepoint. As an approximation a random(-subject) effect on the intercept might do (i.e. a single scalar random effects value for each patient that's used for all timepoints).

