Statistical test for change in prognostic score pre-treatment vs. post-treatment I am working with medical data and specifically determining whether Treatment_X is safe in certain prognostic groups of patients with chronic liver disease. 
To summarize: a common prognostic score for liver disease patients is the Child Pugh Score, which takes on integer score 5-15, and is classified into categories A, B, and C depending on the numeric score (A, 5-6; B, 7-9; C 10-15). Please note that lower is better
We are interested in the change in numeric CP Score from pre-treatment to post-treatment after treatment with Treatment_X. 
If we define a clinically meaningful increase in CP score as 2, we are interested in testing whether the rate of clinically meaningful increase in CP score is greater among patients who are "Child Pugh B" category before treatment compared to those who are in the more favorable "Child Pugh A" category. 
What would be the best statistical method to analyze this data? Thank you so much for your help.
 A: This raises a large number of worrisome issues.  Foremost, a pre-post design is ill-suited for this type of question.  It is too sensitive to time trends, Hawthorne effects, and regression to the mean.  Second, an increase of 2 units has absolutely nothing to do with the way you analyze the data.  It is used only at the final interpretation, or if you compute the Bayesian posterior probability of a difference of at least 2.  Third, the classification into categories A, B, C is statistically invalid and clinically nonsensical.  
Redesign the study as a parallel-group randomized trial.  If you are unable to do that, then just ask the most general question you can.  For a patient who begins at level X=x (y=5-15) what is the probability that she will end up at level Y=y.  To do that I'd suggest a proportional odds model to estimate the P(Y >= y at follow-up given a baseline value of x) and model x as a quadratic expression in 5-15.  You are likely to see a nonlinear effect (which also demonstrates that computing the change from baseline is a mistake).  You can also use the prop. odds model to plot the mean of Y as a function of x.
Doing anything that loses information in the 5-15 levels is a crime against data.  And using a pre-post design will not land the research in the best medical journals.
