Alternatives for Brant Test in Survey Data

[Zhianni Yang]

Hello everyone, I am currently working on a complex dataset where I apply an ordinal scale on my dependent variable, so I use ologit, like this (for example):

svy: ologit risk_behavior_numbers i.sex, base

Once I am going to put more variables in the model, I have to check the assumption of parallel odds or proportional odds. However, brant test does not work after logit.
I have heard that many people suggested using gologit2 and I tried that, but unfortunately gologit2 is exhaustive to my computer that I am unable to do it anymore.

I heard that someone suggested likelihood ratio test between ologit and mlogit, but I am not sure how to do it.


Many thanks.

[Richard Williams]

Unless you've got a very weak computer or a massive data set, I'd be really surprised if gologit2 can't handle this. See pp. 13-14 of https://www3.nd.edu/~rwilliam/xsoc73994/Ologit01.pdf

Example:

. webuse nhanes2f, clear
. quietly svy: gologit2 health weight i.female age i.race, npl
. test [#1 = #2 = #3 = #4]

Adjusted Wald test

( 1) [poor]weight - [fair]weight = 0
( 2) [poor]0b.female - [fair]0b.female = 0
[Output deleted]

F( 15, 17) = 21.04
Prob > F = 0.0000

If this really is too demanding for your computer, then I might just run tests without the svy prefix. The results would at least be suggestive of any problems.

Also make sure you have the most recent version of gologit2, available from ssc.

. which gologit2
c:\ado\plus\g\gologit2.ado
*! version 3.2.6 15feb2025 Richard Williams, [email protected]

[Zhianni Yang]

Dear Mr. Richard Williams:

What I did was:

local varlist sex age working_status wealth_quintile

foreach i of local varlist{
svy: gologit2 risky_behavior_number i.`i'
}

And the code is running for 15 minutes without further result

[Richard Williams]

That isn't what you want. You are running separate gologit2 models for each independent variable. You want one gologit2 model that includes all the independent variables. Something like

svy: gologit2 risky_behavior_number i.(`varlist')
Incidentally, I don't necessarily think 15 minutes is all that long! I'm had many jobs run for hours or days.

[Zhianni Yang]

Dear Mr Richard Williams,

Many thanks to your responses.

Yes, but initially we wanted to look for unadjusted association first -> the result of unadjusted association (p-value below 0.250) will be adjusted in the final model. Or am I doing things wrong?

Many thanks.

[Richard Williams]

Sorry, I don’t understand what you want to do. It seems very different from the original question, which I think I have now answered.

I don’t understand what your code in #3 is supposed to do, but if you give it more time it will probably run, assuming that all your independent variables are indeed categorical with a small number of categories.

[Richard Williams]

You may want to read either or both of these articles on gologit models if you haven't already:

https://www.stata-journal.com/articl...article=st0097

https://www.tandfonline.com/doi/full...X.2015.1112384