Hi there,
I am running some multilevel ordinal regression models and am not sure how to test the proportional odds assumption. In single level modelling the brant command performs this test but I have been unsuccessful using this command in conjunction with runmlwin.
Any help would be gratefully received.
Many thanks,
Beth
Testing the proportional odds assumption in runmlwin

 Site Admin
 Posts: 422
 Joined: Fri Apr 01, 2011 2:14 pm
Re: Testing the proportional odds assumption in runmlwin
Hi Beth,
The brant test is not implemented in MLwiN and therefore is not available in MLwiN.
Presumably the brant test simply tests the equality of the parameters across the various logodds contrasts which you could make.
So you could fit a model where you relax (at least partially) the proportional odds assumption and then use Wald tests to test the equality of the parameters across the contrasts.
Alternatively, if you fit the models by MCMC you could fit two models and compare their DIC statistics. The first model would be the ordinal model where you impose the proportional odds assumption, the second model would be the same model, but where you relax the proportional odds assumption
Take a look at the LEMMA online course which has a module (Module 9) on multilevel ordinal models.
http://www.cmm.bris.ac.uk/lemma/login/index.php
Best wishes
George
The brant test is not implemented in MLwiN and therefore is not available in MLwiN.
Presumably the brant test simply tests the equality of the parameters across the various logodds contrasts which you could make.
So you could fit a model where you relax (at least partially) the proportional odds assumption and then use Wald tests to test the equality of the parameters across the contrasts.
Alternatively, if you fit the models by MCMC you could fit two models and compare their DIC statistics. The first model would be the ordinal model where you impose the proportional odds assumption, the second model would be the same model, but where you relax the proportional odds assumption
Take a look at the LEMMA online course which has a module (Module 9) on multilevel ordinal models.
http://www.cmm.bris.ac.uk/lemma/login/index.php
Best wishes
George
Re: Testing the proportional odds assumption in runmlwin
Hi George,
Thanks for your suggestions, they are very helpful and I'll give them a try. Just to clarify, a model using the proportional odds assumption would look like this (to use the example given):
quietly runmlwin use4 cons (age lc1 lc2 lc3plus urban, contrast(1/3)), ///
level2(district: (cons, contrast(1/3))) ///
level1(woman:) ///
discrete(dist(multinomial) link(ologit) denom(cons) basecategory(4)) ///
nopause
While a model not making the assumption would be specified as follows:
quietly runmlwin use4 age lc1 lc2 lc3plus urban cons, ///
level2(district: cons) ///
level1(woman) ///
discrete(dist(multinomial) link(ologit) denom(cons) basecategory(4)) ///
nopause
The difference being that the command 'contrast (1/3)' in the first model states that a single coefficient will be estimated for all contrasts, implying proportional odds? I just want to check I am understanding you correctly.
Many thanks,
Beth
Thanks for your suggestions, they are very helpful and I'll give them a try. Just to clarify, a model using the proportional odds assumption would look like this (to use the example given):
quietly runmlwin use4 cons (age lc1 lc2 lc3plus urban, contrast(1/3)), ///
level2(district: (cons, contrast(1/3))) ///
level1(woman:) ///
discrete(dist(multinomial) link(ologit) denom(cons) basecategory(4)) ///
nopause
While a model not making the assumption would be specified as follows:
quietly runmlwin use4 age lc1 lc2 lc3plus urban cons, ///
level2(district: cons) ///
level1(woman) ///
discrete(dist(multinomial) link(ologit) denom(cons) basecategory(4)) ///
nopause
The difference being that the command 'contrast (1/3)' in the first model states that a single coefficient will be estimated for all contrasts, implying proportional odds? I just want to check I am understanding you correctly.
Many thanks,
Beth

 Site Admin
 Posts: 422
 Joined: Fri Apr 01, 2011 2:14 pm
Re: Testing the proportional odds assumption in runmlwin
Hi Beth,
Yes that's correct.
In the first model only the intercept varies across the logodds contrasts
In the second model all parameters vary across the logodds contrasts
It would be worth using runmlwin to replicate this test which is illustrated in the lemma online course I referred to in an earlier post in this thread.
Best wishes
George
Yes that's correct.
In the first model only the intercept varies across the logodds contrasts
In the second model all parameters vary across the logodds contrasts
It would be worth using runmlwin to replicate this test which is illustrated in the lemma online course I referred to in an earlier post in this thread.
Best wishes
George
Re: Testing the proportional odds assumption in runmlwin
Thank you George, that's much appreciated.
Many thanks,
Beth
Many thanks,
Beth