reghdfe - difference between robust and cluster error

[Gayle Scharff]

I ran a regression with reghdfe on multiple imputation datasets, The data includes 500 organizations over 23 years, with 20 imputations.
I first estimated the regression using the vce(robust) then I re-ran the model using the option vce(cluster clustvar). In many cases, the standard errors were much smaller when I used the vce(cluster clustvar) option.

Does this seem reasonable? What can explain the difference between the cluster and the robust? Can this be caused because of the reghdfe absorb specification or the MI?

Code:

mi estimate, post cmdok:reghdfe f.women var1 var2 var3, absorb(id year) vce(robust)

mi estimate, post cmdok:reghdfe f.women var1 var2 var3, absorb(id year) vce(cluster id)


Thanks in advance!
Gayle

[Carlo Lazzaro]

Gayle:
welcome to this forum.
The reason is that, likewise -regression-, the community-contributed command -reghdfe- (as you're kindly request to mention it, for sound reasons reported in the FAQ), offers the -robust- and the -cluster- option for dealing with residual distribution heteroskedasticity and autocorrelation, respectively.
In your case, the -cluster- option should be preferred (also given the smaller resulting standard errors, that warn you about serial correlation being nastier than heteroskedasticity), since you have 500 panels (i.e., clusters).

[Gayle Scharff]

Dear Carlo,
Thank you so much for your reply.
I am sorry for my confusion - The standard errors were significantly larger when using the -cluster- option compared to the -robust-.
In this case, would you still recommend continuing with the -cluster- or the -robust- ?

Thank you,
Gayle

[Carlo Lazzaro]

Gayle:
yes, because you have so many panels (and heteroskedastcity is a minor hassle when compared to serial correlation).

[Dario Maimone Ansaldo Patti]

Gayle, you may want to have a look at this paper to have an idea about when to use robust and when to use clustered s.e.

[Joro Kolev]

I should point out that clustered standard errors are robust as well.

In fact robust only standard errors are inconsistent in fixed effects models. This is why the Stata's native -xtreg- will automatically calculate clustered by panel standard errors, even if you request robust only by writing -xtreg y x, fe robust-.