xtgee - heteroskedasticity, autocorrelation, cross-sectional dependence

[Katharina Koe]

Hi,

I am estimating a model using:

xtgee DV IV, family(bin) link(logit) corr(independent) i(panelvar) t(timevar)

I suspect to have heteroskedastic s.e., I think I could fix this by estimating robust s.e. by using -vce(robust) and thus solve the problem of heteroskedasticity and autocorrelation.
However, for now I only suspect it, ho can I verify this? (preferably with a test, not only a plot)

My question is, how do I test for heteroskedasticity, and also for autocorrelation, and cross-sectional dependence in -xtgee?

My DV is fractional, but just for comparison I estimated a linear model, too.
If I estimate the model by OLS, and use the -xttest3 command after, I reject the nullhypothesis of homoskedasticity. Then I have a valid reason to use -vce(robust).

I am hoping for something similar in -xtgee

Many thanks in advance!
Katharina

[Katharina Koe]

Does someone have an idea?
I would be very grateful for some advice, I assume other people must have already come across this, too.
Many thanks!

[Andrew Musau]

I suspect to have heteroskedastic s.e., I think I could fix this by estimating robust s.e. by using -vce(robust) and thus solve the problem of heteroskedasticity and autocorrelation.

You can test for heteroscedasticity and autocorrelation and there are panel data commands to do this. However, with a large N and small T panel, including robust standard errors is sufficient. This is because these are valid in the absence of heteroscedasticity and autocorrelation and are also valid in the presence of heteroscedasticity and arbitrary forms of autocorrelation. So, no one bothers to do the tests. If you are worried that you will specify an incorrect correlation structure in xtgee, this does not matter given that you have specified the robust option as the manual explains

vce(robust) specifies that the Huber/White/sandwich estimator of variance is to be used in place of the default conventional variance estimator (see Methods and formulas in [XT] xtgee). Use of this option causes xtgee to produce valid standard errors even if the correlations within group are not as hypothesized by the specified correlation structure. Under a noncanonical link, it does, however, require that the model correctly specifies the mean.

So what is important is the link function. From your previous post, you want to estimate a fractional logit and I referred you to the do-files in the 2008 paper by Papke and Wooldridge. Here, you will see that you need to compute the average partial effects (APEs) for comparison with the linear model's coefficients and that you need to bootstrap the standard errors of the APEs.

[Katharina Koe]

Thanks a lot, Andrew!!