Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at statalist.org.
[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]
Re: Re: st: Propensity scoring using -teffects- in Stata 13; puzzling features, documentation issues
From
"Lacy,Michael" <[email protected]>
To
"[email protected]" <[email protected]>
Subject
Re: Re: st: Propensity scoring using -teffects- in Stata 13; puzzling features, documentation issues
Date
Wed, 4 Sep 2013 17:09:41 +0000
Thanks to David Drukker for the prompt and thorough response, and for (in good StataCorps
tradition) not taking a defensive posture regarding questions. I'd like to push a bit on
some of my remarks, though.
First, thank to Tom Weichle for in a followup lending some support re my questioning the
enforcement of a matching with replacemement policy. It's good to know that someone
more knowledgeable than I had similar thoughts. One other thought I would have here
is that if StataCorp does not want give countenance to without replacement analyses,
perhaps it could offer a "match cases only" option, without actually doing the analyses.
My own motivation was to do a matched analysis not available in the -teffects-package.
It might have been a bad idea, but I'd like to have the opportunity to do something bad
now and then <grin>. I do understand that user-written packages like -psmatch2- offer
some choices in similar functionality, but it's nice to have something from StataCorp itself.
David Drukker had written in response to my posting:
>From "David M. Drukker" <[email protected]>
>To "[email protected]" <[email protected]>
>Subject Re: st: Propensity scoring using -teffects- in Stata 13; puzzling features, documentation issues,
>Date Tue, 3 Sep 2013 17:12:22 -0500 (CDT)
>
>Mike's second question is essentially, why do the implemented estimators
>include all the observations whose distances are tied. [referring here to
controls matched to treated cases]
>discussion of ties is given in Abadie et al (2004, page 293). There are two
>arguments for including all the tied observations. First, following from
>the quote above, including all the ties provides a more precise estimator.
Several responses from me:
1) I had always understood that the gain in precision from multiple matches
in a matched analysis is not that great beyond 5 or so, and I think
this would be particularly true when matching with replacement is used,
so that (e.g.) my 100 matches may be the same as your 100 matches.
I don't know how the SE is calculated, though.
2) I hadn't realized (my fault) that the "excess" matches only occur
when there are ties in propensity scores of potential matches. Nevertheless,
my experience requesting 2 controls was that I thought something was wrong
with the software, as doing the calculations with (in many cases) 100s of
controls/treated case when I had requested only 2 apparently really bogged
things down and confused me. I thought I had done something wrong and that
the program had hung. Perhaps a warning of some kind is in order?
3) I presume there's nothing theoretically wrong with randomly choosing among tied
controls. I'd give up some precision (well beyond all the inherent errors in my data)
to have a procedure that runs in seconds rather than minutes. Others might feel
differently, and I might feel differently at different times myself.
Anyway, I'd still be interested in comments from others with deeper expertise
than I can claim.
Regards,
Mike Lacy
Dept. of Sociology
Colorado State University
Fort Collins CO 80523-1784
*
* For searches and help try:
* http://www.stata.com/help.cgi?search
* http://www.stata.com/support/faqs/resources/statalist-faq/
* http://www.ats.ucla.edu/stat/stata/