Heteroskedasticity and Serial Correlation in logistic analyses of panel data

[Simon Schmid]

Hallo,

my panel data is suffering heteroskedasticity and serial correlation.
By using standard error robust regressions, I get insignificant models (F-test) and insignificant independent variables.
xtreg y x , vce(robust) re
xtreg y x , vce(robust) fe

I've read that logistic regression resolves the violation of OLS assumptions (serial correlation and heteroskedasticity) and thus I'm wondering, if that holds for panel data and thus I am allowed to use xtlogit without any robust standard errors!?
xtlogit y x, fe
xtlogit y x, re

Thanks,
Simon

PS:
By using random effects model with robust standard errors, I get again insignificant results. (xtlogit y x, vce(robust) re)
I haven't heard anything about an FE Logit model with robust standard errors.

[Stephen Jenkins]

I haven't heard anything about an FE Logit model with robust standard errors.

Have you looked at clogit? Among the options section of the help file, it is stated:

SE/Robust
vce(vcetype) vcetype may be oim, robust, cluster clustvar, opg, bootstrap, or jackknife

[Simon Schmid]

thanks for that advice!

But I'm still looking for an answer to my main question:
"I've read that logistic regression resolves the violation of OLS assumptions (serial correlation and heteroskedasticity) and thus I'm wondering, if that holds for panel data and thus I am allowed to use xtlogit without any robust standard errors!?"

[Nick Cox]

Using logit rather than classical linear regression does not in itself solve any problems of serial correlation.

Please note, from the Advice Guide, that we do ask for use of full real names.

[Simon Schmid]

@ Stephen Jenkins: I do not have a matched sample. (300 vs. 1700 observation) So clogit seems to me inappropriate.

@Nick Cox: ok, I will creat a new account soon

[Nick Cox]

Don't create a new account. Contact the administrators to change your identifier. See "Contact us" on the home page.

[Stephen Jenkins]

Simon, isn't clogit is the same as xtlogit, fe ? See help and Manual entries

[Alfonso Sánchez-Peñalver]

Hi all,

Stephen is right in that xtlogit, fe estimates conditional fixed effects (see help xtlogit). However, this doesn't take care of the serial correlation problem, which is what I think Nick was pointing out. Serial correlation cannot be modeled by either fixed effects or random effects estimations in xtlogit, or by clogit for that matter. I'm very confused that the whole point of using xtlogit is to model heteroskedasticity and serial correlation, because xtlogit requires a binary variable and xtreg does not, and estimating a linear probability model if you do have a binary response variable (i.e. using xtreg with a binary response variable) produces heteroskedasticity of its own. The reason for considering xtlogit, thus, is that your response variable is binary, which means that you should have never considered xtreg to start with. I will continue with the assumption that you do have a binary response variable, but honestly this is confusing. The only way you can capture serial correlation is through population averaged estimation, specifying the corr option. You can do population averaged with either xtlogit, pa or xtgee, family(binomial) link(logit). They are identical. They allow for robust, bootstrap or jacknife estimation of the variance-covariance matrix, thus also adressing the heteroskedasticity issue in the data.