Pooled OLS with cluster robust option

[Taka Sakamoto]

Dear all:

I have time-series cross-national data. I would like to know whether the command "regress" with cluster robust option only corrects for serial correlation, but doesn't not correct for heteroskedasticity, or the command corrects for both serial correlation and heteroskedasticity. I've been browsing the forum on the topic, and it looks like many recommend xtreg, fe with cluster robust option, but I'd like to take into account between effects as well.

I've seen Mr. Carlo Lazzaro touches on this topic multiple times, and I tried to understand what he recommends. But when it's said the regress command with cluster robust corrects for some of heteroskedasticity, I cannot judge whether I can use pooled OLS with cluster robust option to correct for both serial correlation and heteroskedasticity.

Thank you very for your help.

Best wishes,

Taka

[Jeff Wooldridge]

The clustered standard errors are robust to heteroskedasticity and serial correlation.

[Taka Sakamoto]

Thank you, Professor Wooldridge. I'd like to confirm that the clustered standard errors are robust to heteroskedasticity and serial correlation even with -regress- command.

I heavily rely on your textbooks. What an honor to hear directly from you.

Many thanks again.

Best,

Taka

[Carlo Lazzaro]

Taka;
what I stated in many replies on this topic is something different:
1) under -xtreg-, the options -robust- and -vce(cluster panelid) do the very same job, as they both invoke cluster-robust standard error;
2) under -regress-, the two options take heteroskedasticity and serial correlation into account, respectively.

In the unfortunate case that your -regress- is plagued with both heteroskedasticity and serial correlation, -vce(cluster clusterid)- is recommended, as Jeff highlighted.

[Taka Sakamoto]

Dear Carlo:

Could you tell me why I should opt for cluster-robust standard errors, if it is the unfortunate case that -regress- is plagued with both heteroskedasticity and serial correlation? cluster-robust standard errors doesn't take into account heteroskedasticity? Or does it take into account both heteroskedasticity and serial correlation?

Thank you for your help.

Best wishes,

Taka

[Carlo Lazzaro]

Taka:
see Stata Bookstore: Environmental Econometrics Using Stata, pages 28-29, Case 4.

[Taka Sakamoto]

Thank you, Carlo. I've read the part of the book. So, cluster-robust standard errors are robust to both heteroskedasticity and serial correlation, as Jeff says. Thank you again.

[Carlo Lazzaro]

Taka:
you're correct.
The issue is a bit tricky and creeps up on this forum (as well as elsewhere) from time to time.
I'll try to wrap things up below.
Under -regress-:
1) if you have heteroskedastcity only, you can take it into account (allowing for different types of heteroskedasticity to be modeled) via -robust-;
2) if you have autocorrelation only, you can take it into account via -vce(cluster clusterid)-;
3) if you have both, -vce(cluster clusterid)-, that assumes an arbitrary type of heteroskedasticity - rules.
Under -xtreg-, both -robust- and -vce(cluster panelid)- options deal with heteroskedasticity and/or autocorrelation of the systematic epsilon, as both these options invoke cluster-robust standard errors.

[Taka Sakamoto]

Thank you, Carlo. I see. So, the limitation of 3) also applies to the case of under -xtreg-, right? I'm not sure how I should incorporate the knowledge.

Best wishes,

Taka

[Carlo Lazzaro]

Taka:
yes, cluster-robust standard errors available from -xtreg- do not allow the researcher to model heteroskedasticity of the epsilon according to a given pattern.

[Taka Sakamoto]

Thank you, Carlo, for your help.

Best wishes,

Taka