Testing proportional odds assumption for longitudinal data

[Job04]

I have longitudinal data with 3 ordered classes and I'm running proc genmod (interested in marginal trend).  I need to test the assumption of odds proportionality but proc genmod. Do you know another method that compares models in terms in terms of this assumption?

Thanks 

[SteveDenham]

The type 3 test for the interaction of time and your class variable (assuming you are treating this as a GEE) will tell you if the log odds are the same across time for each class.  If this test is not significant, then the test of the main effect of time will tell you if the log odds (averaged over all classes) is homogeneous over time.

So to test proportionality, you could set up ESTIMATE or LSMESTIMATE statements that look for a significant trend over time within each class.  If the trend is not significant, then either the log odds are constant over time, or are not proportional.  Only examination of the marginal means for the log odds will distinguish between these two cases.  If there is a significant trend over time, then there is a significant constant of proportionality in the log odds.

I hope this helps. There may be a specific option available to test this, but I am not aware of it.

SteveDenham

[Job04]

Thank you so much I'll try to incorporate that. 

[StatDave]

The proportional odds assumption implies that the slopes on all of the cumulative logits are the same. More particularly, that the parameters of any particular predictor on the logits are the same. To avoid testing and the need to deal with the correlations among the repeated measures in longitudinal data, you can use the graphical method described in this note. The macro there can produce a plot of the empirical logits for each of the predictors. 

As far as doing tests for the assumption, I can only suggest fitting the ordinal model in PROC NLMIXED and including an RANDOM statement to deal with the repeated measures. You could then use CONTRAST statements to test the equality of the parameters of each predictor among the logits. I illustrate fitting the ordinal model in NLMIXED in the Addendum at the end of this note, but not for the repeated measures case, so it does not include a RANDOM statement.