Unusual p-values with -reghdfe-

[Bader Alhashel]

Hi everyone,

I am new to using reghdfe. I am running a regression in which I am using absorbing with fixed effects two variables. I am also clustering the errors on those two variables. I get large t-stats such as 4.73 for which the p-value is only 0.018. Why is it that the p-values are not as we would get with a regular regression? I tried canceling the error clustering and using just robust and that resulted in the p-values and t-stats being aligned to what you'd normally see.

Thanks for your help in understanding this issue.

[Jesse Wursten]

It's rather difficult to say anything about this without some data and or code illustrating the issue. Sergio Correia might know what's going on...

[Nick Cox]

Seems to imply 2 df???

[Bader Alhashel]

Nick,

It seems so because when I use robust SEs instead of clustered ones, I get p-values that correspond to the t-stats shown in the table.

Jesse,

This is the regression I am running.

reghdfe i_4 rep_match dem_match ln_pu3 q cf sg gdp_g election_year fg_mean indexofconsumersentimen, cluster(gvkey quarter) absorb(gvkey quarter fqtr)

[Nick Cox]

The code doesn't help much, I am afraid, given that rather obviously we don't have your data. It's the output that will give the context.

[Sergio Correia]

Given that you are doing multi-way clustering and have several FEs, I wouldn't be surprised if reghdfe thinks you only have 2 DFs, as Nick says.

It seems you are absorbing for quarter and fqtr (financial quarter?), so these variables might overlap a lot, and as per Cameron et al, I set the number of "real observations" when running the Ftests as the minimum number of clusters - 1.

[Bader Alhashel]

Nick,

Attached is a screen capture of the output from one of the regressions.

Sergio,

I am absorbing for calendar quarter and fiscal quarter, and yes for many firms those would overlap perfectly.

[Sergio Correia]

Well you only have four quarters; I'm not sure how correct is it to cluster by quarter. In fact, the thumb rule is that if you have less than 50 quarters you should reconsider what you are doing, or perhaps switch to other methods (see Cameron et al's paper).

More importantly: is your quarter variable correct? You should be having {2001q1, 2001q2 ... 2016q4}; not {1,2,3,4} as quarters (!!)

[Jesse Wursten]

I think this is the first time I've seen a regression with standard errors clustered on a time variable?

[Bader Alhashel]

Thanks your replies everyone.

Sergio,

Yes this is the way to do it as there is another variable that is time variant but cross-sectionally invariant so using your suggested fixed effects would result in one being canceled. The idea behind using quarter fixed effects is to capture seasonality across the quarters. I could try not to cluster the SEs by quarter and see if I still have this problem.

Thanks again !