Different way of clustering with reghdfe

[Filippo Palomba]

Hello everyone!

I'm currently running a multi-way fixed effect regression using the command reghdfe. As I have data which is clustered on two dimensions I went through the classical paper by Cameron, Gelbach and Miller on the topic. To the best of my understanding, the two-way clustering estimator they describe in the article can be implemented on Stata using (this is just one of the possible solutions of course) as

reghdfe depvar indepvars, a(fe1 fe2) vce(cluster c1#c2)

However, I've noticed that there exists the possibility of running the following command

reghdfe depvar indepvars, a(fe1 fe2) vce(cluster c1 c2)

Here is my doubt: what kind of clusterisation does vce(cluster c1 c2) imply and in which sense does it differ from vce(cluster c1#c2)?

Thanks for your help,

Filippo

[Jesse Wursten]

Say you have countries (c1) and years (c2). Then cluster(c1#c2) allows for correlation between observations within the same country in the same time period. Cluster(c1 c2) instead allows for correlation between observations in the same country (irrespective of time period) and in the same time period (irrespective of country). I'm 95% certain it is the latter that you want. See also page 19 of their working paper [PDF] (on their personal webpage).

[Filippo Palomba]

Yes, your guess is totally right, that's what I was looking for! Thank you for your help!