Ordered Logistic Regression in multiply-imputed (mi) dataset: Verifying the proportional odds assumption

[Musarrat Nahid]

I'm trying to do ordered logistic regression with “mi” data and wanted to verify the underlying assumption: “the proportional odds assumption or the parallel regression assumption" but I’m unable to use the relevant commands for the tests alongside “mi estimate” command: “omodel” and “brant” (I already installed both of these in my Stata). How can I verify that assumption in the multiply-imputed (mi) data? or how to make these commands (“omodel” and “brant”) work in mi data?
I would greatly appreciate it if someone could provide some suggestions!

[Brendan Halpin]

You can likely do it by hand, using mi estimate: logit, if you create the contrasts by hand (i.e. create variables indicating 1 vs 2+, 1&2 vs 3+, etc, and fitting binary logistic regressions on each one).

[Musarrat Nahid]

Thank you very much, Mr. Halpin, for your prompt reply! Is there a more straight-forward/ efficient way to check that assumption in Stata? I mean, right now I'm trying to find out what works better- linear regression, binary logistic, multinomial, and ordered logistic regression (our study team wants to explore various different approaches); so I'm conducting all of those analyses on our 3 different datasets (each one has 5 imputed datasets); moreover, we have 3 outcome variables (DVs) and we want to investigate which outcome might yield better results/ be more useful. So it would really be helpful if there was a less tedious way to do it in Stata. Thanks again.
P.S. I tried to use “brant” in different ways alongside “mi estimate” but it did not work; I could make “omodel” work alongside “mi estimate” using “cmdok” but it did not give me the necessary output (“Approximate likelihood-ratio test of proportionality of odds across response categories”) as would usually be expected if my dataset was not mi set.

Thanks!

[Richard Williams]

Personally I don't think diagnostic tests have to be accurate to 12 decimal places. You could do something like this and see if there appears to be a problem.

webuse nhanes2f, clear
mi set mlong
mi register imputed health
mi impute ologit health weight, add(5)
mi xeq: omodel logit health weight

[Musarrat Nahid]

Dear Mr. Williams,
Thanks very much for your suggestions! This sounds great. I'll try the way you showed. Thanks again!
Sincerely
Musarrat

[Kevin Kovach]

Dear Dr. Williams,

I am using stata's mi xeq: omodel logit commands to test the proportional odds assumption while using multiple imputation. However, I'm not sure how to interpret this because it test for proportional odds in each imputation separately. How would I interpret a case where the proportional odds assumption was met for some imputations and not for others?

I really appreciate your help.

Sincerely,

-Kevin