Default robust regression standard errors

[Sebastian Kripfganz]

I think the following comment might be of interest to the Stata community:
http://datacolada.org/99

"Hyping Fisher: The Most Cited 2019 QJE Paper Relied on an Outdated STATA Default to Conclude Regression p-values Are Inadequate"

Aside from the observation that they misspelled Stata, it is a good reminder to not blindly trust default settings in Stata (or any other software, for that matter). Whether the default should be changed is a different story. Even if vce(hc3) is superior to vce(robust), there is always the issue of ensuring consistency/replicability over different Stata versions.

[Jackson Monroe]

Seeing as robust is not a default in the first place I am not really sure what the gripe is. "There are multiple ways to correct for heteroskedasticity and they picked a bad one," is not the software's fault. The "default" heteroskedasticity correction is the first-derived (I think) and most direct one, whoopdie doo. Also, having higher than nominal coverage in some cases cannot really remedy other cases having lower than nominal coverage, which seems to be the result the author is pushing with their graph. Why not use the MAE in coverage?

Maybe the take home should really be "if you think asymptotic results are guaranteed to hold for N=20 or N=200 I have a bridge to sell you."

[George Ford]

Might be interesting to have a -set robust HC3- option (or whatever option you prefer).