EViews.com

Dear all,

I'm trying to estimate an ordered logit model. Now I have some problems in testing the specifications of my data. The following questions will touch both topics: "How to do it in EVIEWS?" and "How to do it [full stop]?". As my first try in the "estimation"- category was no success, I' ll try here.

First problem: How can I find out if the "proportional odds/parallel regression assumption" holds? I think one way to do this is to apply a Brant-test (even if it seems to me that this test is seen as to strict in detecting a violation of the assumption). Is this test implemented in EVIEWS? If I would arrive at the conclusion that the assumption is violated, I would have to use a general ordered logit regression instead of the "normal" one? Is there an easy way to do this in eviews?

Second problem: To "hedge" my estimations I also computed an OLS estimation with the data which I use. In this case I detected heteroscedasticity using a White-test, so I corrected the estimation by using White-coefficients. So I got the ideas of testing the ordered logit model for heteroscedasticity, too. As when I changed the type of estimation procedure from OLS (White corrected) to ordered logit (omitting the intercept) in the "Equation Estimation"-window, the ordered logit model will be estimated with "Robust covariances: Huber/White", I suppose that "White" and "Huber/White" are somehow related or do correspond, right?
- How can I separately test the ordered logit model for heteroscedasticity? (or is my procedure viable to test heteroscedasticity in OLS and then concluding that heteroscedasticty is also a problem in ordered logit?)
- Searching in econometric journals I found the suggestion to test ordered logit models for misspecification by using LaGrange-Multiplier tests. Would this be a totally different approach from "Huber/White" and "GLM" (= Generalized Linear Models")?

Yeah, that are quite a few questions..
Thank you for your help (and your patience)!!

Best wishes

Lama

EViews doesn't perform the Brant test automatically, but you could certainly perform the estimation for each of the binary submodels with little difficulty using the dependent variable expression

y>k

for each value of k. Getting the asymptotic variance matrix of the coefficients is a bit trickier, but can be done with a bit of work. Depending on how many categories there are in the original ordered model, this can be a bit of a pain. To be honest, from my reading of the literature, the evidence on the Brant test seems to be that it not worth the effort, but I'll leave that for you to decide.

As the the heteroskedasticity issue, you should note that LS on a binary or ordered model *should* exhibit heteroskedasticity, so your test using the linear model may not be telling you what you want.

Having said that, the computed robust standard errors are somewhat different than the LM tests for misspecification since the former allows for a limited set of departures from the ML specification. The various forms of the LM tests appear to be straightforward auxiliary regression tests that should be easy to carry out in EViews (though we don't yet have them built in).

Hello Glen,

thanks a lot for your advices!You helped me a great deal!
(And you are probably right about the Brant-test, but I just wanted to make sure, if there was an easy way to perform it which I simply overlooked)

Best regards

Lama