its an interesting complication.  survey weights are designed to infer to a population different than the one that the unweighted sample aims for.   matching changes the target quantity of interest even if there are no weights to begin with.   so you have to decide what your QOI is.   if its the one defined in matching, and its all in sample, then you don't really need weights.   if its the original population, then you might not want to match at all, but you will have to deal with model dependence.  if its the population defined by the units that are matched (i.e., those matched units are a random sample from this alternative population), then you probably want to use the weights.

finally, if you're using CEM, it produces CEM weights which you should always use.  If you wind up needing both CEM-weights and survey weights, you should be able to multiply them together.

Gary
--
Gary King - Albert J. Weatherhead III University Professor - Director, IQSS - Harvard University
GKing.Harvard.edu - King@Harvard.edu - @kinggary - 617-500-7570 - Asst 495-9271 - Fax 812-8581



On Tue, May 17, 2011 at 11:06 PM, Leslie Finger <lfinger@fas.harvard.edu> wrote:


Hi Class,

Sorry for the harassment far after the end of class.  I'm using survey data with
weights and, if I use matching and readjust my quantity of interest to the
average treatment effect on the treated, should I still be using survey
weights?  My inclination is that the answer is no, especially since I might be
using the weights from matching in the regression.

Any help is much appreciated!

Best,
Leslie

p.s. if any of you have ever used "svydesign" in R, I'm trying to figure out
what exactly i'm supposed to put for "id"...

_______________________________________________
gov2001-l mailing list
gov2001-l@lists.fas.harvard.edu
http://lists.fas.harvard.edu/mailman/listinfo/gov2001-l