Hi,
I have data on an ordered outcome coded from 1 to 3. Suppose I suspect
that the distinction between levels 2 and 3 is meaningless--i.e., nothing
substantive would be lost if I just recoded the 3s to 2s.
One implication would be that the proportional odds assumption
wouldn't not hold for the 3-category model because the difference between
levels 2 and 3 has nothing to do with the X variables.
I could test this, but it seems like a weak test. The
proportional odds assumption could be violated for any of several reasons.
How do I demonstrate to my psychologist colleagues (the measure is a psych
measure) that collapsing 3 categories into 2 is harmless?
thanks,
Michael Foster
Professor of Health Policy and of Demography
