--- Richard Williams wrote:
> Why not always specify -robust- when using
> OLS regression? My initial reaction is to say that you shouldn't
> relax restrictions unnecessarily; and there are various
> post-estimation commands where Stata will at least whine at you if
> you've used robust standard errors (e.g. -lrtest-). But in practice,
> your model is probably at least a little mis-specified and/or there
> may be some degree of heteroskedasticity, so maybe robust is a good
> idea. Any thoughts on the matter?
Part of the answer is discussed in [U] 20.14. I interpret it as
follows: If your model is correct in every respect, the parameter
estimates represent causal effects. Combine this with what we know
about (simple random) sampling, and you get the normal standard
errors. The robust standard errors start from a different point:
the regression coefficient measures the difference in expected
value for individuals that are one unit of the explanatory
variable apart. It doesn't care whether this difference is
(partially) causal or not. I tend to be rather sceptical about
arguments that take estimates of causal effects too seriously; in
the end we only observe differences in means, odds ratios, etc. As
a consequence I cannot get very excited about this distinction,
but others can.
-----------------------------------------
Maarten L. Buis
Department of Social Research Methodology
Vrije Universiteit Amsterdam
Boelelaan 1081
1081 HV Amsterdam
The Netherlands
visiting address:
Buitenveldertselaan 3 (Metropolitan), room Z434
+31 20 5986715
http://home.fsw.vu.nl/m.buis/
-----------------------------------------
*
* For searches and help try:
* http://www.stata.com/support/faqs/res/findit.html
* http://www.stata.com/support/statalist/faq
* http://www.ats.ucla.edu/stat/stata/