0
$\begingroup$

I have ordinal data (4 levels) which I have fit with:

a) 1 proportional-odds logistic regression model

and also

b) 3 separate binary logistic regression models fit to values >= levels 2, 3, and 4.

The binary model curves and coefficients are 'very close' to those of the ordinal model, which has substantive implications. My question is whether there is a formal way to describe 'very close', other than looking at a plot of the curves and looking at the coefficients.

$\endgroup$
  • $\begingroup$ Not sure how to develop this in a formal way, but I'd suggest comparing the odds ratio of the ordinal model with the respective binary ones. $\endgroup$ – Lucas Farias Nov 16 '18 at 11:04
  • $\begingroup$ Your option b looks very like the continuation ratio model but it is not clear to me exactly what you did. $\endgroup$ – mdewey Nov 16 '18 at 21:32
1
$\begingroup$

b) is not recommended method. You can fit one model called generalized logit model to replace 3 models in b. You already have proportional-odds logit model. Then you are ready to perform likelihood ratio test. No significant means no difference between two models. Therefore we will select the simple one: proportional-odds logit model. If significant, maybe you can try partial proportional-odds logit model. if you have more than one covariate.

$\endgroup$

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.