Panel with FE and standard errors clustered on the same variable?

[Lorenzo Caverni]

Hi everyone. I am dealing with a balanced panel data of 200 banks from 14 countries which are observed over 9 years. The panelid is bank_id.
I implemented a model with fixed effects at bank_id level (using option fe in the xtreg command) and I clustered the standard errors at same level too, i.e. at bank_id level (using the option cluster() in the xtreg command). The model as this specified produce strongly significant results. Does this make sense? I couldn't find similar questions in this forum or elsewhere in the web.
Thanks everybody, this question is of vital importance as i am right next to a fundamental deadline.

[Jeff Wooldridge]

Lorenzo: Do your variables of interest vary at the bank or country level? If the former, cluster at the bank level, as you have done. If you’re studying differences in policies across country then one can make a case for clustering at the country level. Hopefully not with only 14 countries.

[Lorenzo Caverni]

Jeff: Thanks for your answer. The variables of interest certainly vary at bank level, although some differences have been found across countries too. Anyway, you're saying that adding FE and clustering the standard errors at the same level of the FE variable is something that makes sense? I couldn't find similar evidences across the web. I see that usually one add the fixed effect for a variable (say the country) and cluster the s.e. for a different one. Thanks again for your answer.

[Alfonso Sánchez-Peñalver]

Lorenzo, in panel data with few years you should consider the unit of observation the whole panel, with all the years. So when you say that you're clustering at the panel variable level, you are in fact estimating the appropriate White robust estimator of the variance covariance matrix. This is why with -xtreg- if you specify robust it will give you the cluster robust estimates of the variance covariance matrix, with clustering at the panel level. It may also be why you don't see people talking about clustering at the panel level with panel data because, once again, the unit of observation is the actual panel. There is nothing wrong with what you've done, just think of it as the robust estimator of the variance covariance matrix for panel data.

[Jeff Wooldridge]

Doing what you have done is very standard with panel data. Using only country fixed effects in this context is kind of a cop out. It would only be done if using FE at the bank level gives unappealing results -- often imprecise results. Since you have variability in the covariates of interest at the bank level, clustering at that level seems fine. If you were studying a "treatment" assigned at the country level then you would want to cluster at the county level.