Using Driscoll and Kraay Standard Errors when T is small

[Saed Same]

Dear all,

It is my first post here. I have a panel data (N= 277 firms, T= 3 years). I used the Hausman test and chose a fixed effects model for my results. However, the modified Wald Test showed the presence of heteroscedasticity. Furthermore, I used xtcsd with peseran to test for cross-sectional dependence. The test rejected that there is no "cross-sectional dependence". If I used the cluster option with xtreg fe, I would adjust for heteroskedasticity but not for the cross-sectional dependence. Therefore, I used fixed effects with Driscoll and Kraay Standard errors (xtscc) to control for both problems. I am , however, afraid that xtscc assumes large T. Is having 3 years enough to use the xtscc as there is no particular time limit? Any other method to adjust for both heteroscedasticity and cross-sectional dependence?

Regards,

[Saed Same]

Dear all,

It is my first post here. I have a panel data (N= 277 firms, T= 3 years). I used the Hausman test and chose a fixed effects model for my results. However, the modified Wald Test showed the presence of heteroscedasticity. Furthermore, I used xtcsd with peseran to test for cross-sectional dependence. The test rejected that there is no "cross-sectional dependence". If I used the cluster option with xtreg fe, I would adjust for heteroskedasticity but not for the cross-sectional dependence. Therefore, I used fixed effects with Driscoll and Kraay Standard errors (xtscc) to control for both problems. I am , however, afraid that xtscc assumes large T. Is having 3 years enough to use the xtscc as there is no particular time limit? Any other method to adjust for both heteroscedasticity and cross-sectional dependence?

Regards,

I would appreciate a lot if any one could provide his or her guidance on my post.

[Aysun Atil]

Hi,
I wonder how you had solved your problem. I've encountered similar problems in my analysis. I would appreciate if you could help me.

[Maxence Morlet]

Use cluster-robust standard errors. Without more information, I would recommend you cluster at the firm level.

You could use the user written command "summclust" for CV3 standard errors or use the wild cluster restricted bootstrap (boottest command). That's up to you.

Poeple often ignore cross-sectional depence, just try to cluster at the coarsest possible level. You can also read the paper "When should you adjust standard errors for clustering".

[Carlo Lazzaro]

Aysun:
welcome to this forum.
Maxence gave sound advice.
I can only add that with N=277 and T=3 you're clarly in the short panels realm.
Therefore, I woud invoke -robust- or -vce(cluster firmid)- standard errors (they do the very same job in -xtreg-) that take hateroskedasticity and/or autocorrelation of the epsilon into account.

[Aysun Atil]

Thanks for the answers.
In fact, in my analysis T=13 and N is above 100,000. I would like to use driscoll-kraay errors since heteroscedasticity and autocorrelation exists. I run the regression with reghdfe command, but couldn't incorporate dkraay to reghdfe command. That is why I return to xtscc, fe command. I'm not sure about what short time period xtscc becomes problematic.
I also would like to ask also the possible endogeneity problem. How can we deal with it?
Thanks a lot.

[Aysun Atil]

Thanks for the answers.
In fact, in my analysis T=13 and N is above 100,000. I study with firm-level data and use sector*year and firm fixed effects. I would like to use driscoll-kraay errors since heteroscedasticity and autocorrelation exists. I run the regression with reghdfe command, but couldn't incorporate dkraay to reghdfe command. That is why I return to xtscc, fe command. I'm not sure about what short time period xtscc becomes problematic and if it works with multi dimensional fe.
I also would like to ask the possible endogeneity problem. How can we test and deal with it?
Thanks a lot.

[Carlo Lazzaro]

Aysun:
with your updated N and T values my previous reply still holds.
That said, I do not understand if your concern is about across panel correlation of the epsilon error.
Endogeneity issues are dealt with -xtivregress-.

[Carlo Lazzaro]

Aysun:
-xtivregress- reads -xtivreg-.
Sorry for the mishap.

[Aysun Atil]

Hi everyone,
I have a model of independent variable=productivity and dep var=participation. My data is panel of firm-year combinations.
I want to use system gmm because of the endogeneity problem.
However, I don't want to lose the insight of firm heterogeneity in fixed effects models.
How can I handle the problem?
Thanks.

[Carlo Lazzaro]

Aysun:
see help -xtivreg-.
That said, I suppose that you have >1 predictor in the right-hand side of your regression equation.