brant test to test parallel line assumption for survey data

[Dereje Tesfaye]

Dear,I am working on 2016 Ethiopian DHS of children dataset. I need to run brant test to test parallel line assumption for outcome variable that ordered in to three categories.
I run the command on STATA 14 like this:

svy: ologit outcome variable independent variables ,or
brant
output was "brant- does not work with ologit models with weights"

It works when I excluded the command svy or

ologit outcome variable independent variables ,or
brant : works

May I have another alternative way to run brant test by considering svyset? Or I am looking for any advice concerning my problem.

Kindly regards,

Dereje T.

[Richard Williams]

The user-written gologit2 command (available from SSC) provides an alternative to the Brant test that works when the data are svyset. See

https://www3.nd.edu/~rwilliam/stats3/Ologit01.pdf

especially pp. 10 and 13-14.

[Dereje Tesfaye]

Thanks ,
I have seen that is for partial proportional odds model
But i want to run the following before partial proportional odds model if any other options instead of Brant for svy
svy: ologit outcome variable independent variables ,or
brant

[Joseph Coveney]

I'm not sure whether someone has already written a canned procedure that will work with data from complex survey designs, but you might be able to contrive something in such a context that is directly analogous to Brant's test using suest.

But why not just inspect the regression coefficients of the two binary logistic regression models and see for yourself whether you have anything worth fussing over?

[Maarten Buis]

Thanks ,
I have seen that is for partial proportional odds model
But i want to run the following before partial proportional odds model if any other options instead of Brant for svy
svy: ologit outcome variable independent variables ,or
brant

What a Brant test does is compare a generalized ordered logit with an ordered logit. It tests the hypothesis that all coefficients except the constants (cut-off values) are equal across equations. A Brant does that by approximating a generalized ordered logit and than compute a Wald test on that approximation. With gologit2 you can fit an actual generalized ordered logit and perform a Wald test on that. So not only can you correctly include your survey information into your test using gologit2 you can also improve on the Brant test by avoiding the approximation.

Having said all that, I am a great fan of just inspecting the regression coefficients and make my own informed decision rather than rely on a test. The purpose of the test is to warn us when the "pattern" in the coefficients could just as well have been the result of random noise. We humans need such a warning because we are very good at seeing patterns in random noise (e.g. Rorschach or inkblot test in psychology). So statistical tests are just an initial screening devise. After we found significant differences, we can start the real work and determine for ourselves if the pattern we see in the coefficients warrant a special model or we think it is small enough to justify simple ologit model.

[Richard Williams]

gologit2 can estimate partial proportional odds models, but that isn't what it is doing in the handout I suggested you read. As Maarten points out, it is providing an alternative to the Brant test for testing the proportional odds/parallel regressions assumption. As an added bonus, you could do the same procedure with probit if you wanted, since it also makes the parallel regressions assumption.

As Maarten and Joseph point out, what you should do if the assumptions are violated is another matter. Sometimes the violations may be so small they aren't worth worrying about. Other times you might want to try a partial proportional odds model or use a different approach completely, e.g. mlogit. Sometimes the violations have interesting substantive implications and are worth discussing. I go over things like this in my 2016 article "Understanding and interpreting generalized ordered logit models." You can get it at

https://www3.nd.edu/~rwilliam/rwpubs...ologit2016.pdf