[Carlo Lazzaro]

Ferhart:
yes, they are (but not to cross-sectional dependence).
That said, with 2000 clusters, I would be more worried about within panel autocorrelation.

[Ferhat Demirci]

Ferhart:
yes, they are (but not to cross-sectional dependence).
That said, with 2000 clusters, I would be more worried about within panel autocorrelation.

Mr. Lazzaro,
As far as I know (I don't have in-depth econometrics knowledge), isn't Driscoll-Kraay a robust method for all three cases (cross-sectional dependence, autocorrelation, and heteroskedasticity)?

[Jeff Wooldridge]

Ferhat: Driscroll-Kraay relies on large-T asymptotics. That's how it can be robust to cross-sectional dependence, because, after estimation, it collapses the data to a time series and applies Newey-West. Because it uses a heteroskedasticity and autocorrelation consistent (HAC) estimator, it doesn't allow any kind of time series correlation: it can only be weakly dependent. In my view, that's usually fine, but not as general as clustering as Carlo suggested. In any case, T = 16 is not enough to justify Driscoll-Kraay.

What is your cross-sectional unit? Is it geography? If so, then you should look for spatial HAC solutions. I've found the user-written command -acreg- to be very useful, but it relies on having latitude and longitude, which typically means having access to a "shape" matrix. If your unit is something else -- such as firms -- you should rethink the whole need for allowing cross-sectional dependence. You should always include full time dummies and hope that accounts for common aggregate shocks.

If you're interested in learning more about these issues, I sometimes offer a panel data course where I discuss applications and tradeoffs of clustering by unit, Driscoll-Kraay, and spatial HAC.

[Ferhat Demirci]

Ferhat: Driscroll-Kraay relies on large-T asymptotics. That's how it can be robust to cross-sectional dependence, because, after estimation, it collapses the data to a time series and applies Newey-West. Because it uses a heteroskedasticity and autocorrelation consistent (HAC) estimator, it doesn't allow any kind of time series correlation: it can only be weakly dependent. In my view, that's usually fine, but not as general as clustering as Carlo suggested. In any case, T = 16 is not enough to justify Driscoll-Kraay.

What is your cross-sectional unit? Is it geography? If so, then you should look for spatial HAC solutions. I've found the user-written command -acreg- to be very useful, but it relies on having latitude and longitude, which typically means having access to a "shape" matrix. If your unit is something else -- such as firms -- you should rethink the whole need for allowing cross-sectional dependence. You should always include full time dummies and hope that accounts for common aggregate shocks.

If you're interested in learning more about these issues, I sometimes offer a panel data course where I discuss applications and tradeoffs of clustering by unit, Driscoll-Kraay, and spatial HAC.

It made me happy to see you here. I will consider your suggestions. I'll try two-way fixed effects estimator, including full time dummies, but then what should I do if autocorrelation, heteroskedasticity, and cross-section dependence continue to be a problem?

[Jeff Wooldridge]

Heteroskedasticity and autocorrelation are taken care of by vce(cluster id). You have to be careful about testing for cross-sectional dependence, as those tests can reject when there is no issue. What is your unit of observation? If it's not geography, I doubt you have a case where you need to deal with cross-sectional dependence.

[Ferhat Demirci]

Heteroskedasticity and autocorrelation are taken care of by vce(cluster id). You have to be careful about testing for cross-sectional dependence, as those tests can reject when there is no issue. What is your unit of observation? If it's not geography, I doubt you have a case where you need to deal with cross-sectional dependence.

Mr. Wooldridge,
I am really grateful for your interest. I am working on a financial data. My units are public firms listed on stock exchanges in different countries.

[Jasmine Xu]

I also have a problem with autocorrelation, heteroskedasticity, and cross-section dependence.
I have experimental data: 8 independent sessions, 14 participants per session where they played multi-player game for 80 periods.
I am currently using vce(cluster session), and I am advised to use some more conservative method to take care of correlation between participants within a session.
Therefore, I am thinking to run regressions for each session and use standard error which will be robust to autocorrelation, heteroskedasticity, and cross-panel correlation.
I am wondering if Driscroll-Kraay works in my case.

Thanks.