Small sample correction for AbsorbingLS

[arielgoldszmidt]

When AbsorbingLS is estimated with debiased = True, the covariance estimator is scaled by nobs / (nobs - nvar), where (nobs, nvar) is the shape of exog_resid.

So nvar here is the number of non-absorbed variables in the regression. Shouldn't the appropriate scaling factor instead be nobs / (nobs - nvar - nabsorbed), where nabsorbed is the number of absorbed variables (or areg.approx_rank)?

[bashtage]

Thanks for noticing this. I'll need to do some thinking about this since Absorbing LS will estimate modes with redundancy columns and so it isn't always trivial to compute the DoF adjustment simply by computing the number of RHS variables. For example, if you have multiple categorical variables absorbed, it regresses on the complete sets of dummies from each. The LSDV method provides an identification method in this collinear problem since it imposes some additional parameter constraints. The parameters are not meaningful, but since these are simply purged it doesn't affect the DoF.