the idea of CEM, and matching in general, is to prune observations from the
data set. That often changes the quantity of interest in return for
reducing model dependence. If you'd like to keep the same quantity of
interest (pate or patt), then you can divide the sample into 2 parts -- the
part CEM identifies as matched and which you can estimate with little model
dependence, and the part which is much more model dependent. you can
produce estimates within each and then take the weighted average of the two
to produce your ultimate estimate. you'll find some examples of this in
our papers at my web site.
best,
Gary
On Thursday, June 14, 2012, Fernando Terrés wrote:
Hello, CEM list:
I would like to use CEM with survey data. The sampling design is
stratified with survey weights (1500 observations). I would like, if
possible, to estimate population estimates (PATE and PATT).
My questions:
1. Is it possible to get this estimations with CEM? (if yes, how?)
2. Can imbalance be used with survey data? (if yes, how?)
3. I see that, with CEM, the researcher must specify interactions. Exist
any kind of advice on which interactions to include with CEM?
After the matching, I would like to test mediator effects (SEM with
Mplus), using the CEM matched observations and weights (perhaps the weights
to be used are a product of CEM weights and survey weights?), and the
correct sampling design (stratified, on the subpopultation of matched
cases?).
Thanking you in advance,
Fernando
--
Gary
--
*Gary King* - Albert J. Weatherhead III University Professor - Director, IQSS
<http://iq.harvard.edu>- Harvard University
GKing.Harvard.edu <http://gking.harvard.edu/> - King(a)Harvard.edu -
@kinggary<http://twitter.com/kinggary>- 617-500-7570 - Asst 495-9271 -
Fax 812-8581