Evaluating a pretreatment predictor with repeated measures ordinal data

Question

I am currently attempting to study whether particular baseline factors predict pre/post changes in relation to a substance use intervention. However, a few things have me stumped.

1. Our substance use measure uses ordinal responses for past 30 day substance use (i.e. 0 = 0 days of use; 1 = 1-3 days; 2=4-8 days; 3=9-15 days; 4=16-30 days).

2. Our predictor variables are all scaled.

First, I'd like to ensure I'm utilizing the best test for a repeated measures design for our ordinal substance use variable. We have one condition (treatment) without a control. I've seen some comments that perhaps a GEE or Wilcoxon analysis would work for the initial aim (does the treatment result in significant reduction of substance use).

Second, and more central to our aims, what might be the most suitable method to determine, "does baseline characteristic x (i.e. age, weight, etc.) have an effect of reducing the effectiveness of the intervention?" given that my pre/post assessments are ordinal?

Answer 1

I'm not sure that I agree that your responses are ordinal. They actually seem more like an interval-censored logarithmic response. Consider: your zero means something real and the intervals are roughly exponential.

If you elect to treat them as ordinal, I would suggest that you apply a forward-continuation-ratio logistic regression. See Agresti or O'Connell.

An excellent alternative to ordinal would be Brunner, Domhof, and Langer nonparametric methods which would bring mixed models to bear.