Using Heteroskedacticity Consistent Standard Errors HC1 vs HC3 in Stata vs R

[Kasey Zapatka]

Hey everyone,

Does anyone know why Stata uses HC1 as default for robust standard errors instead of HC3, which is the default for the sandwich package in R? The package maintainers cite the Long and Ervin paper below but I can't find much documentation from Stata as to why they use HC1. I don’t think there is usually a very big difference but I would like to know the rationale as to why R and Stata have different defaults here.

Thanks!

Long and Ervin paper (pg 4)
https://www.tandfonline.com/doi/abs/...dvYsbDA9MAAAAA :plWvLmIxKX4IoQqPxosxJBjNNDgTk4StsXlWrJWG0PR_68_hI go9LG-pyx8dMDvyL2NCamvQRgeSDg

R package
http://sandwich.r-forge.r-project.or...html#arguments

[Leonardo Guizzetti]

There was a similar thread about this some time ago, but the rationale was not resolved as I recall. If the default choice isn't justified in the documentation, then this is a question on StataCorp can answer. You are of course free to use whichever method you prefer.

[Kasey Zapatka]

Great. I just emailed them and will report back. Thanks Leonardo!

[Enrique Pinzon (StataCorp)]

Dear Kasey,

To clarify, there is no such thing as a default robust standard error computation in Stata. All commands have a default standard error computation and then there are -vce()- options. If you want anything other than the default -vce(ols)-, for -regress-, you need to type a -vce()- option. Both HC1 and HC3 are readily available and optional. We have made no specific stance as to what the "correct" robust standard error computation should be. We provide options and let the researchers decide what is best for their particular set up.