oparallel perfect prediction - Brant test

[Charlie Joyez]

Dear Stata users,

I am currently working on some ordered logit models, and I'd like to run the Brant test the parallel line assumption.
I'd try to run oparallel command (as it runs five tests)
However, an error message displays

"full model cannot be estimated due to perfect prediction".

Indeed my ordered logit command did signaled

"Note : 46 observations completely determined. Standard errors questionable".

(hence my model was not properly set).

Then I manage to remove the concerned observations (by identifying factor variables that only corresponds to one category of output, and observations with predicted probabilty very close to 0 or 1), I re-run my ologit model which didn't signal me any observation completely determined.

Yet, when re-running the oparallel command, I have the same message error "full model cannot be estimated due to perfect prediction", although I get rid of such observations.

So two questions : 1) How come oparallel signals me some perfect predictions, while ologit didn't.
2) How to identify those prediction and remove them from the oparallel test.

Thanks to anyone that would advice me on this.

Charlie Joyez

[Maarten Buis]

Oparallel compares two models: the ordered logit model you estimated and a generalized ordered logit model. It is the comparison of these models that make up the test of the parallel lines assumption. So, the error message tells us that there are observations which perfectly predict the outcome in the unconstrained, generalized ordered logit model. You should be able to identify those by estimating the unconstrained model with gologit2 (available from SSC).

[Charlie Joyez]

Thank you Maarten,
I'll try to identify those observations with gologit2.
Should I remove them from my ologit model too then?
I guess yes If I want to compare the two results on a similar sample, however, it bothers me to reduce a sample (and maybe change the results of ologit).