Re: st: weighted regression

[Richard Williams <Richard.A.Williams.5@nd.edu>]

At 11:28 AM 1/14/2004 +0100, Ernest Berkhout wrote:

> When doing regressions, I always use aweights. I'm a bit ignorant about 
> pweights but my guess is that they give the same results. However iweights 
> are very different! When you
Weights are discussed on pgs. 278-282 of the Stata 8 Users Guide.  -help 
weights- from within Stata gives a lot of the same info but not as much 
explanation as to rationale.  A few points:

* If you have a probability weighted sample and are using aweights rather 
than pweights, you are probably doing it wrong!  Or so says the users 
guide.  You have the consolation of knowing you are not alone.  As an SPSS 
user, I believe I have always been using the equivalent of aweights, as 
pweights are not an option.

* As I understand it, the main difference between aweights and pweights is 
that with pweights, you get robust standard errors.  aweights and pweights 
produce the same point estimates but different standard errors.  The users 
guide explains the rationale for this.

* I don't understand the practical difference between fweights and 
iweights.  The guide says that iweights are usually used by programmers; 
why they don't use fweights instead, I don't know.

* Back in December there was a discussion as to whether you should weight 
at all; see Dick Campbell's message at
http://www.stata.com/statalist/archive/2003-11/msg00653.html
and the followups.  Dick cited a very good 1994 article by Winship and 
Radbill saying that the answer was usually no.  Left unclear to me is 
whether that is still true with Stata 2004 and its pweights and svy 
commands.  My reading of the Users Guide is that you should use pweights 
rather than not weight at all.  If there is a more contemporary discussion 
of these issues I would like to see it.

* Finally, going back to the original question:  the idea behind pweights 
is that your sampling scheme deliberately oversampled some groups, e.g. 
minorities are overrepresented.  However, the original question made it 
sound like underrepresentation of some groups just sort of happened; it 
wasn't a matter of the research design.  If nonresponse was random then I 
suppose pweights may be ok; but if there are systematic biases in the 
sample then I'm not sure that pweights really solves them.  Another of my 
weak areas is post-sampling adjustments, but my understanding is that this 
is a weak area for Stata too, and something that is being worked 
on.  Again, I'd be interested in hearing more on this.


-------------------------------------------
Richard Williams, Associate Professor
OFFICE: (574)631-6668, (574)631-6463
FAX:    (574)288-4373
HOME:   (574)289-5227
EMAIL:  Richard.A.Williams.5@ND.Edu
WWW (personal):    http://www.nd.edu/~rwilliam
WWW (department):    http://www.nd.edu/~soc

*
*   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/