Re: st: Combining a survey weight and a frequency weight

[Steve Samuels <sjsamuels@gmail.com>]

After rereading David's post and the -help- for -psmatch2- (which I've
never used), I realize that I don't know what the _weight variable is.
My modified advice is below.


S.

James:

I suggest that, subsequent to running -psmatch2-, you create the
following ATT weights

For treated subject i:             
weight = survey weight

For the controls of treated subject i: 
weight = 
(survey weight for  treated subject i)/(no. controls matched to subject i)


This forces the weights of the controls matched to a treated subject to
sum to the survey weight of that subject. Then -svyset- data with these
weights and the rest of the design information and use Stata's survey
programs. If you restricted treated subjects to those whose propensity
scores overlapped those of the controls, you will need the subpop()
option in your survey analysis commands. That means that data from the
dropped treated subjects must be in your analysis data set.


References:

Orwin, R, Hornik, R, Judkins, D, Zador, P, Sridharan, S, and Baskin, R.
2003. Innovative design and analysis strategies in the evaluation of the
national youth anti-drug media campaign: Propensity scores and
counterfactual projection weights in a national probability survey FCSM
Statistical Policy Working Paper 37. Federal Committee On Statistical
Methodology Research Conference 6-11. http://fcsm.gov/03papers/Orwin.pdf

DuGoff, Eva H., Megan Schuler, and Elizabeth A. Stuart. 2013.
Generalizing Observational Study Results: Applying Propensity Score
Methods to Complex Surveys. Health Services Research

This was a simulation study of a non-stratified, single-stage design.