Ordinal logit model with panel data

[Thibaud Marcesse]

Hi,

I am trying to run an ordinal logit model with panel data, but can’t perform a Brant test with xtologit. I ran one using one year of my panel, and unsurprisingly the proportional odds assumption has been violated, which makes me think that this is going to be a problem for the whole panel. Is gologit2 the solution here? And does it work with panel data? Or is it safer to run a multinomial model (with a different interpretation).

Thanks for letting me know!
Thibaud.

[Richard Williams]

See https://www3.nd.edu/~rwilliam/gologit2/tsfaq.html Specifically, the section labeled Can I do a random effects or multilevel model with gologit2?

[Thibaud Marcesse]

Thanks a lot Richard!

[Bruce Weaver]

What is the sample size? If it is quite large, then it is not surprising that the Brant test is statistically significant.

Richard Williams, if Frank Harrell is right when he says that Violation of Proportional Odds is Not Fatal, wouldn't -meologit- be an option here too?

[Richard Williams]

My FAQ does have several suggestions, including

Note: If you have a multilevel model and aren't too worried about violations of the parallel lines assumption, then consider using the xtologit, meologit, xtoprobit, or meoprobit commands. You might also consider using xtmlogit, available in newer versions of Stata. Or,you can use gsem to estimate a multilevel mlogit model.

[Stephen Jenkins]

A minimal amount of searching (-search-ing) also yields "feologit: A new command for fitting fixed-effects ordered logit models", by Baetschmann and others, Stata Journal, 20 (2).
" https://journals.sagepub.com/doi/ful...36867X20930984. The article draws on the authors' article published in JRSS(A).

[Thibaud Marcesse]

Bruce, the dataset is very large (close to 50K observations). I have been trying to decide whether to treat the data as a panel, or alternatively, run the ordered logit model for individual years of my dataset. For context, I am studying the impact of affirmative action at the village level in India on a large poverty alleviation scheme (implemented at the village level). I have an ordinal/categorical variable that measures outcomes (low level of spending, middling level of spending and high level of spending). On the independent variable side, I use a categorical variable that codes villages according to their reservation level (for certain population categories). I decided against using a multinomial model as I worried about not being able to interpret the coefficients. This analysis is for a book chapter, and I make no causal claims (the reservation of villages is not randomized). I had not run feologit but will give it a try as a random effects model does not make much sense with the data I have.

[Jeff Wooldridge]

It’s a misconception that just pooling the data and using ologit creates some sort of bias. The usual inference is incorrect, but that’s easily fixed by clustering at the village level. A full set of time fixed effects should be included, for example, i.year. This actually requires fewer assumptions for consistency than random effects, which requires independence across time conditional on the random effects.

Fixed effects ologit requires the same assumption, but allows correlation between the village effect and the x(i,t). But it’s hard to recover marginal effects. A correlated random effects model is a nice compromise. Add the time averages of the time-varying explanatory variables and use ologit or probit.

[Maria Kaneva]

I am running oparallel command and got the folloing results:
Tests of the parallel regression assumption

| Chi2 df P>Chi2
-----------------+----------------------
Wolfe Gould | 2646 57 0.000
Brant | 2739 57 0.000
score | 2799 57 0.000
likelihood ratio | 2632 57 0.000
Wald | 2748 57 0.000

What is the null hypothesis for these tests? Does p-value 0.000 indicates the assumption of the parallel trends is violated?