Heteroskedasticity and autocorrelation tests for -xtlogit, re-

[Jana Schue]

Hello all,

I work with panel data and I am running a logistic regression because my dependent and independent variables are binary (purpose and dirty). The Hausman test spit me out to use a -xtlogit, re- model. I also have time-fixed effects.
Unfortunately, despite the numerous posts on this topic, I still can't figure out if and how I should test for heteroskedasticity and autocorrelation here.

My command currently looks like this:
xtlogit purpose dirty `controls' i.fyear, re vce(robust)
margins, dydx(*)
estimates clear

If anyone can tell me if and how to test for heteroskedasticity and autocorrelation, it would help me a lot. For my master's thesis, I also need sources for each step, so if anyone has a suitable paper at the ready, that would be even better.

Thanks and best regards,
Jana

[Joro Kolev]

Heteroskedasticity and autocorrelation are complicated questions in nonlinear maximum likelihood models. The reason for this is that they are forms of misspecification, which totally spoils your model (makes the likelihood you are maximasing invalid) as a general rule. The linear model is an exception that heteroskedasticity and autocorrelation do not constitute a misspecification in your regression function.

Just say that you are controlling for arbitrary autocorrelation and heteroskedasticity by the -, vce(robust)- option. Your supervisor and second reader would not know this anyways :P.

[Jana Schue]

Thanks a lot, that sounds like a good tip! I should actually also do tests for heteroskedasticity and autocorrelation to sort of justify including vce(robust) in the command. But from what I read here, it might be sufficient to include it anyway.