areg , robust >>> standard errors are clustered what?

[Yasutake Homma]

Hello could you please help for the following question?
I am working for company level panel data but would like to test country level too.

First, I set up the panel dataset as follows:

xtset corporate_number year

However, I would like to test the panel specification with country and year.
Then, I use

areg xxx xxx i.year, absorb (country_number) robust

But in this case the standard error would be clustered by company or country?
If you could please help me understand the above question, it would highly be appreciated.

[Andrew Musau]

areg will behave exactly as regress as it is not a panel estimator in the Stata sense (panel estimators have the -xt- prefix). Therefore, -robust- implements White standard errors. The following example uses esttab (Stata Journal) to collate results.

webuse grunfeld, clear
eststo: xtreg invest mvalue kstock, fe robust
eststo: regress invest mvalue kstock i.company, robust
eststo: areg invest mvalue kstock,  a(company) robust
esttab, drop(*.company _cons)

Res.:

esttab, drop(*.company _cons)

------------------------------------------------------------
                      (1)             (2)             (3)  
                   invest          invest          invest  
------------------------------------------------------------
mvalue              0.110***        0.110***        0.110***
                   (7.25)          (5.68)          (5.68)  

kstock              0.310***        0.310***        0.310***
                   (5.88)          (7.25)          (7.25)  
------------------------------------------------------------
N                     200             200             200  
------------------------------------------------------------
t statistics in parentheses
* p<0.05, ** p<0.01, *** p<0.001

[Jeff Wooldridge]

You can only answer the clustering question by knowing whether the data are from a true cluster sampling scheme — which sounds unlikely in this case — or knowing the level of assignment of the key explanatory variables. Are those generated at the firm level or country level? If the latter, clustering at the country level is desirable. But don’t cluster at the country level unnecessarily.

[Yasutake Homma]

I really appreciate you Andrew and Jeff. I understand your answers.