Development branch throws new error: "e(df_r) doesn't match" about reghdfe

[sergiocorreia]

Hi Nils,

This is caused by what I believe is a subtle bug within Stata. I'll give an example with the -hdfe- command (available from this repo, as described in the readme.md)

* Create your dataset
input x y i j
0 1 1 1
1 1 1 2
1 2 2 2
1 3 3 3
1 3 3 2
end

* Apply within transformation
hdfe y x, a(i j) clear

* Regress; notice that both X and cons are dropped
reg y x

* However, Stata still counts X as not dropped when computing df_m!!!
di e(df_m)

* The reason is due to numerical accuracy:
* The part of regress that decides whether to drop or not uses more less accuracy
* Than the part of regress that computes e(df_m)
su x
* If you run the hdfe command with the tol(1e-10) option, you will see that now e(df_m)==0

* Also, in the case below e(df_m) IS zero (just to make sure that they do exclude dropped variables when computing df_m)
gen z = 0
reg y z
di e(df_m)

I'm not entirely sure about how to proceed. One option would be to override the e(df_m) from regress, thus fixing their bug manually.

[sergiocorreia]

By the way, I added the above solution to the latest build. Not the most "kosher" solution, but is probably the best for now.