I have a general question about regression in Stata.
What is the main difference between commands:
rreg (Huber regression) and reg, robust?
What is the explanation for one and the other?
Asked
Active
Viewed 6,413 times
1
-
this website goes through it in more detail http://www.ats.ucla.edu/stat/stata/webbooks/reg/chapter4/statareg4.htm – charles Nov 28 '13 at 14:52
-
1actually it goes through it in less detail, but with code examples – charles Nov 28 '13 at 16:09
1 Answers
5
You are confusing quite different things, but the main reason for your confusion is that terminology in statistical science is indeed inconsistent here.
The Stata command rreg implements one flavour of robust regression that is (in a very limited sense) robust to outliers in the data. What it does is well documented in the Stata manuals and also discussed elsewhere in this forum at
Quantile regression vs. Li's regression: which should I use, and when?. It remains visible as a matter of continuity but even among Stata programs it has (in my opinion) been superseded long since by alternatives. The thread just cited offers an independent and authoritative opinion in agreement.
The Stata option (not command) robust implements robust (Huber-White-sandwich) standard errors that are offered as more honest standard errors in the face, principally, of heteroscedasticity. Asking for robust standard errors does not amount to robust regression in the sense just discussed, and outliers and long tails in any of the variables will have exactly the effect they have otherwise on coefficient estimates; the difference is that your standard errors may differ, affecting any quantities that depend on them.
-
Thank you for this helpful explanation. Please can you confirm that "reg, vce(robust)" is the same as "reg, robust"? – bobmcpop Mar 06 '18 at 13:06
-
1@bobmcpop I think that's so in any version of Stata that supports both. At the time of writing (Stata 15.1) `vce(robust)` is documented and `robust` undocumented. – Nick Cox Mar 06 '18 at 13:14
-
Not directly related, I had been thinking of asking a question regarding a critique of rreg. It seems that initially discarding values if they have a Cook’s distance greater than one, compounded by not making this Cook’s distance threshold (one) depend on the number of observations are perhaps weaknesses. Stata 3.0 introduced rreg, are there commands you would recommend instead of it? – Single Malt Dec 05 '21 at 19:21
-
1See my postings on that command on Statalist. I suggested several years ago that it should go undocumented. But there never has emerged, in the 30+ years since the original paper, a single widely accepted flavour of robust regression— nor even would there be universal acceptance of the goal. So, no, I have no recommendation independently of anyone’s dataset and model. – Nick Cox Dec 05 '21 at 20:22