How to check the proportional odds assumption in multilevel ordered logit

[Sixten Maximillian Thestrup]

I am currently running a multilevel ordered logit with individuals nested within seven countries, where the DV has three outcomes (0 = disagree, 1 = indifferent, 2 = agree). For simplicity, let's say that I add two IVs:
X1 = gender dummy variable (0 = man, 1 = woman)
X2 = patriotism on a 9 point scale, considered to be continuous

Lets forget about the multilevel structure for a moment, and just run an ordered logit:

ologit attitude gndr patriotism

I can then use the Brant test command (part of the 'spost'-add-on, installed using -findit spost-), to check the proportional odds assumption (that the cumulative odds ratio is constant across response categories):

brant, detail

However, I want to test the proportional odds assumption with a multilevel structure. First I run the model of interest:

meologit attitude gndr patriotism || country:, intmethod(ghermite)

Then, if I run the Brant test command as before, I get an error:

brant, detail
brant can only be used after ologit

I am not sure if it is at all possible to check the proportional odds assumption in multilevel models. If it is, do you know of any (possibly user-written) commands that I could use in this instance?

[Joseph Luchman]

Hi Sixten,

You could try fitting an equivalent model using gllamm (SSC) relaxing the parallel lines assumption and do a set of test postestimation commands to discern whether there's evidence of the assumption being violated.

something like:

eq thresh: gndr patriotism
gllamm attitude, i(country) family(binomial) link(ologit) thresh(thresh)
test [_cut11]_b[gndr] = [_cut21]_b[gndr]
test [_cut11]_b[patriotism] = [_cut21]_b[patriotism]

The test commands should give you some sense of whether the assumption is violated for those variables.

- joe

[Martijn Hogerbrugge]

Hi Joseph,

Would it be possible to run a partial-proportional odds model in gllamm?
I.e., once significance tests reveal the coefficient for threshold _cut11 and _cut12 do significantly differ, but the thresholds for _cut12 and _cut13 [and _cut14, etc.] do not,
is there some way to account for this in a new model?

Thanks.
Martijn

[Amanda Wyant]

Has there been any additional solutions to this issue since this was written? Can you use the test command after running meologit or can it only be used for gllamm?