The authors of -psmatch2- are careful enough in explaining that the
asymptotic standard errors they provide are only valid under this and
that context, and even though they suggest the bootstrap to estimate
the standard errors in a more general context, they still note that it
is not clear whether the bootstrap is applicable in this case (and I
am a kind of a person who would only trust the bootstrap after seeing
consistency laid out in a JASA paper). Putting a complex survey
structure on top of this is more than the estimator of this kind
(non-smooth, based on pairs and groups of observations rather than
single observations working so nicely towards -ml lf- forms, etc.) is
more than it can bear. In other words, I don't think there is any
reasonable estimator currently available for this series of
unfortunate events. I mean, there are always ways to get "something"
that would look like a standard error in the output... but whether you
can trust those standard errors at all will always be questionable.
Not that is totally impossible -- in the end, we do know how to run
logistic regression with survey data to get the propensity scores
(-svy : logit -), and then get some contrasts out of the survey data
(-svy: means- with -lincom-, most likely with some subpop settings for
different groups of estimated pscore... or fiddling with weights
somehow), but putting the two steps together is a substantial
challenge.
Also, as far as I recall, Guido Imbens (who knows approximately 10^4
times more about matching estimators than I do) was a bit skeptical
about -psmatch2- routine last year at the NASUG, and was proposing
-nnmatch- instead. It does allow for -pweights-... so I would need to
check with the SJ publication to see if their treatment of the design
information satisfies me :)). In some sense, their procedure might be
considered to be simpler, as you don't need to worry about
misspecification of the first step of estimating the propensity score,
and correcting the standard errors for serveral estimation steps; but
rather essentially using a non-parametric regression procedure... of
which much more is known in survey statistics world.
On 3/30/07, Anna Gueorguieva <[email protected]> wrote:
Dear Statalist,
After searching on the statalist, I am still struggling with how to weigh
my observations when I use psmatch2.
It currently does not take weights. Is there a way I can pre-inflate my data
or use some intermediate result of the program that I can later put into
a command that uses weights?
Here is my command:
psmatch2 migranthh landend notbornkos numyearsinkos landend $ind_var, outcome(lncon) ate
--
Stas Kolenikov
http://stas.kolenikov.name
*
* 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/
