Do "Driscoll-Kraay standard errors" correct heteroskedasticity and autocorrelation in Panel Data?

[Diego Castano]

Hi!

I was using xtreg, fe command on my Panel Data with N = 33, T = 25 and it had heteroskedasticity, autocorrelation and cross sectional depedence. So I used xtscc but I´m not sure if Driscoll-Kray errors correct heteroskedascity and autocorrelation

[Carlo Lazzaro]

Diego:
due to such a rilevant T dimensione, which is only slightly lower than the N one, you should probably consider -xtgls- or -xtregar- for your analysis.

[Jeff Wooldridge]

Actually, this is a good scenario for xtscc, as it requires “large T.” Its whole purpose is to compute stand errors that are robust to spatial correlation and serial correlation. You always get robustness to heteroskedasticity for free.

Using GLS with a specific AR structure on the errors is less robust, and large N makes the statistical properties suspect.

[Carlo Lazzaro]

Interesting, thanks Jeff!

[Diego Castano]

Ok guys. So do I have to keep my regression and analysis with xtssc?

By the way, thanks for the advices!

[Jeff Wooldridge]

That’s what I recommend in my post.

[Diego Castano]

Ok. Thank you, Jeff.

[Meike Stephan]

Dear all,

I am running a fixed effects orderd logit model with the feologit command in order to test whether the employment type of life partner has effects on the subjective well-being of individuals. Unfortuantely, the hettest tells me that I have a problem of heteroskedasticity. Do you have any idea how I can solve this problem, maybe by using another model command?
I appreciate help on this problem!

[Carlo Lazzaro]

Meike:
welcome to this forum.
Please start a new thread, as your query is not consistent with the original post's title.

[Meike Stephan]

Thanks Carlo!

no problem!
I will open a new thread!

[Sabrin Rahman]

I am using xtreg, fe command on my Panel Data with N = 78 countries, T = 23 a years and it has heteroskedasticity and autocorrelation but when I am using Xttest2 for cross sectional depedence it does not work. My Dependent variable is Female Labour force participation and independent variable is FDI. For robustness check when I am using vce(robust) all my control variables become insignificant but my main variable (FDI) is significant. As a PhD student shall I show these results, or will it create problems. Moreover I have read about the Driscoll-Kray errors correct heteroskedascity and autocorrelation. When I use this my results are significant, can I use this as a robustness check?

I am new to statalist.

[Joro Kolev]

If we start to worry about countries' residuals being cross sectionally correlated, I do not know what we can assume that can be possibly cross sectionally uncorrelated -- observations coming from alternative universes?

I think that with your data you should stick to -xtreg- because your data is more of the large N variety. I would not worry about cross sectional correlation in your case, and just do -xtreg, robust- which will give you standard errors robust to heteroskedasticity and arbitrary within country correlation.

I am using xtreg, fe command on my Panel Data with N = 78 countries, T = 23 a years and it has heteroskedasticity and autocorrelation but when I am using Xttest2 for cross sectional depedence it does not work. My Dependent variable is Female Labour force participation and independent variable is FDI. For robustness check when I am using vce(robust) all my control variables become insignificant but my main variable (FDI) is significant. As a PhD student shall I show these results, or will it create problems. Moreover I have read about the Driscoll-Kray errors correct heteroskedascity and autocorrelation. When I use this my results are significant, can I use this as a robustness check?

I am new to statalist.

[Ferhat Demirci]

Diego:
due to such a rilevant T dimensione, which is only slightly lower than the N one, you should probably consider -xtgls- or -xtregar- for your analysis.

If N>T (eg N=2000, T=16) is Driscoll-Kraay suitable?

[Carlo Lazzaro]

Ferhat:
I would go -xtreg- with cluster-robust standard errors here.

[Ferhat Demirci]

Ferhat:
I would go -xtreg- with cluster-robust standard errors here.

Are clustered standard errors robust to autocorrelation and heteroskedasticity? Wouldn't it be better to use Driscoll-Kraay for cross-sectional dependence?