On 10/9/08, Austin Nichols <[email protected]> wrote:
> Matching (which addresses only selection on
> observables), esp. propensity-score matching, can be thought of as
> reweighting,
As Archimedes had said before a Roman axe treated him, "Do not disturb
my sampling weights!". They have a certain meaning, and one needs to
be sure whatever (s)he does to them makes sense within that meaning;
or alternatively that those weights are used according to the new
meaning. If you can still get a Horvitz-Thompson estimator (and the
accompanying variance formula) in the end, then I am fine with all
that weight tumbling. If I can think of a superpopulation from which
the population ALONG WITH the treatment assignments was sampled, then
this reweighting procedure does indeed produce the weights that would
allow one to generalize to that superpopulation (although there are
still issues of the weight adjustment groups cutting through the
existing strata; that's usually frowned upon by survey statisticians,
as it messes up the pairwise probabilities necessary for the proper
variance estimation).
But as I said I would love to see (should I say, I am yet to see?) the
technical reference that would marry those concepts seamlessly.
--
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/
