Generalized ordered logit model

[Luis Mijares Castaneda]

Hello, I have a question about model selection, I am using the gologit2 command, and I'm getting the results below. All the variables meet the PL trend lines assumption, but the overall models fail the PL assumptions. My data is survey weighted, the models are shown below. I have three outcome variables along with the controls. Should I use a mlogit model or continue with the parallel trend lines assumption? I'm using survey data so alternative tests of model fit like LR or AIC/BIC are not avaliable, not sure what other test I can use for the PL assumption

Step  17: All explanatory variables meet the pl assumption

F( 48,  1292) =    2.30
            Prob > F =    0.0000

An insignificant test statistic indicates that the final model
does not violate the proportional odds/ parallel lines assumption

svyset [pw=weights]

local outcomes  dental_visit oral_problems food_avoidence

foreach outcome of local outcomes {
    
    local contrls i.imprisson c.age i.biosex i.maritalstatus_alt i.income ///
             i.employment_alt i.education c.householdsize i.race_alt ///
             i.military_branch
    
    
gsvy: gologit2 `outcome' `contrls', autofit(0.001) difficult waldforce or

    
    
}

[Richard Williams]

Seeming conflicts between the global test and the tests of individual coefficients can happen. Suppose, for example, that all of the individual coefficients just miss being statistically significant. Even though each just misses, it is likely that at least one coefficient does differ from zero and the global test will reflect that.

I would probably go with the PL model. If you switch to mlogit you are adding zillions of additional parameters when the gologit results suggest you may only need a few more and maybe not any.

But, if you want to pursue it further: try

gologit2 `outcome' `contrls' [pw=weights], autofit(0.001) difficult waldforce or If you only have pweights, you don't necessarily need to use svy. I'd be curious to see what happens.

Or, try using a less restrictive figure for autofit, e.g. autofit(0.01) or autofit(.005). The gologit model tends to work best when only a few variables are found to violate parallel lines. Your choice of .001 is kind of arbitrary anyway and maybe you'll find that having only one or two vars violate pl is a reasonable compromise between pl and mlogit.

Whatever you wind up doing, this handout may help you interpret results.

https://academicweb.nd.edu/~rwilliam.../Margins05.pdf