meologit and parallel odds assumption

[Francesca]

Dear Listers,

I also have an ordinal outcome measure, which assesses whether their life satisfaction has 'improved', 'stayed the same', or 'declined'. The information was collected at week 12 and week 52. I aim to assess whether the change from week 12 to 52 differs as a function of study arm.

I plan to use the meologit command in Stata 13 after reshaping the data to a long format. I know I should test whether the parallel odds assumption has been met but this does not seem possible after the -meologit- command. Any suggestions on how this can be achieved or any alternative approach?

[Adam Mayer]

You can do this in GLAMM as described here: http://www.statalist.org/forums/foru...-ordered-logit

In principle I don't see why this couldn't be extended to meologit.

[Francesca]

Hi Adam,

Thanks for the link. I found that the proportional odds assumption is violated for some variables. Would you know if it is possible to run a partial-proportional odds model using -meologit- or -gllamm- ?

Thanks

[Mike Lacy]

I have a suggestion in a different direction: My experience with ordinal logit models outside the panel context is that departures identified by an hypothesis test as "significant" are commonly trivial in size. My jocular formulation here is that if the sample size is large enough to trust the maximum likelihood estimation process to be valid, then an hypothesis test will almost always have sufficient power to cast doubt on the proportional odds assumption.

Now, if the tradition in your research area or discipline emphasizes hypothesis tests, this doesn't help you. However, if that is not the case, I think it's worth looking at other measures of model fit in making a decision. For example, one thing I would suggest is to examine proportional changes in one of the pseudo-R2 measures (which I suppose are probably not available for -meologit-, sorry) between your two model specifications. Another possibility is to compare the predicted probability vectors for your two models across all your individual respondents. Something I find intuitively appealing is to compare the mean difference in the predicted probability of Y = 1, Y = 2, etc. across the two models. If the mean differences are small, say differing by only a few percent, why would we care about the more complicated model? In most cases I would argue that the unfixable errors of measurement associated with response variables such as you suggest considerably exceed any differences associated with model specification.

[Francesca]

Hi Mike,

Thanks for your reply. Just to double-check, in your post you suggest running separate regressions for the levels of the predictor variables which do not meet the parallel odds assumption to assess whether the pseudo R2 or the predicted means are largely different - is this correct?

I am in a field that emphaizes hypothesis testing but I think it may still be worth exploring your approach if even -gllamm- can't deal with partial proportional odds.

[Mike Lacy]

I was thinking of comparing the predictions from -meologit-, which entails the proportional odds assumption, to a model that does not. I don't know what kind of alternative model you can specify with -gllamm- in the panel context. In the non-panel context, the alternative would be a generalized ordered logit or even a multinomial logit. But: Even if there is no good way to specify an alternative, I'd go back to my implicit point, namely that I think that departures from proportional odds are overemphasized.

[Francesca]

Thanks Mike!