Clustering using reghdfe- Full Rank Error

[justin winikoff]

Hello. I am running a county-level fixed effects panel regression and I include state by year level fixed effects. Ideally, I'd like to cluster at the state level to account for some spatial correlation, but understand that doing so would give me fewer clusters than coefficients, which is problematic. Alternatively I could cluster at the county level. I have about 2500 counties and then about 100 variables, so I should be ok, right?

Nonetheless, I am getting the "Warning: estimated covariance matrix of moment conditions not of full rank. overidentification statistic not reported, and standard errors an model tests should be interpreted with caution."

My first stage F-stat is drastically smaller than it was with no clustering, and I'm wondering if I can trust this result. I'd appreciate any help you can give.

Thanks

[Sergio Correia]

Can you try using the version of reghdfe from Github? On recent iterations, I tweaked reghdfe to be less aggressive when there are too few clusters.

If you still don't see the FStat though, that probably means that (after singletons are dropped), you have less clusters than variables (including the year FEs).

[justin winikoff]

Sergio, this worked. Thank you for your help