feols computes infeasibly small standard errors

[NickG-e61]

exampledata.csv

As an example, running the following line on the linked dataset (where DV and EV are randomly generated) reports a p-value of 1.8e-15. There should be plenty of degrees of freedom.

feols(DV ~ EV | state + as.factor(month), data=exampledata)

I'm also getting funny looking standard errors in other cases, e,g, converting a factor to a set of dummies and then getting the same standard error on every dummy coefficient. Can give more details if you'd like.

[lrberge]

Hi, and thanks for making the data available, but could you report the output you are referring to with ready to go code?

Did you compare with other software (lm/lfe/Stata/Julia)?

[NickG-e61]

Sure, the output is attached, for fixest and for lfe.

[lrberge]

Thanks, now it is clear:

• it's because the standard-errors are clustered on a cluster of two values

By default, if the argument vcov is not provided, fixest clusters the SEs wrt the first fixed-effect, here this is state. The current lfe SEs are "iid" so the comparison doesn't hold.

You need to know what kind of standard-errors you want ("iid", "heteroskedasticity-robut", "clustered"?) and then pass it into the vcov argument. The problematic SEs come from clustering with such a few number of clusters (it shouldn't be done).

[NickG-e61]

OK, thanks for clarifying. Since there is no default argument for vcov, my assumption was that it assumes iid errors. Others might assume this too.

[lrberge]

Sure. The use case is individual panel data.

The type of SE is clearly stated though:

Closing then. Thanks for reporting the outputs.