Statalist


[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: st: robust regression with robust errors


From   "Stas Kolenikov" <skolenik@gmail.com>
To   statalist@hsphsun2.harvard.edu
Subject   Re: st: robust regression with robust errors
Date   Mon, 17 Nov 2008 09:33:11 -0600

On 11/17/08, Diemo Urbig <urbig@econ.mpg.de> wrote:
> Dear stata experts:
>
>  knowing that "rreg" can be used to deal with violations of
>  distributional assumptions and "reg ,r" for dealing with
>  heteroscedasticity, I ask myself whether and to what extent there is a
>  way to deal with both at the same time. Given that "reg", "reg , r", and
>  "rreg" all yield different results regarding a model and distributional
>  assumptions and heteroscedasticity seem to play a role, is there a way
>  to combine rreg with robust errors?

-rreg- is a reasonably complicated routine (with some switching
between objective functions as far as I can recall), and the procedure
for computing standard errors is even less transparent. Technically
since -rreg- is an M-estimator, one should be able to construct the
analogue of -_robust- sandwich var-cov matrix (although you might have
issues with the -rreg-'s objective function not being smooth enough,
and the bread part of the sandwich requires a derivative).
Realistically most robust regression methods work with some sort of
cut-off points beyond which the influence of the residuals is
downweighted, and allowing for heteroskedasticity means that one
should have a way to have a new influence cut-off for any new
observation -- which is a bit too much to ask: you either believe a
large residual is an outlier that needs to be downplayed by the robust
procedure, or that this is a realization of large variance due to
heteroskedasticity. You cannot do both at the same time, those are
kinda contradictory to one another.

You might want to run a small simulation to see how -rreg- standard
errors actually perform.

-- 
Stas Kolenikov, also found at http://stas.kolenikov.name
Small print: I use this email account for mailing lists only.
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/



© Copyright 1996–2014 StataCorp LP   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index