Propensity score matching considering sampling weights

[Zhijing Cai]

Hi everyone,
I am using the NLSY97 data to do a matching analysis, and I think it important to consider the sampling weights of NLSY97. I found some useful information here: https://www.statalist.org/forums/for...e-psmatch2-etc, especially Melissa's answer at #4. However, when it comes to my own model, I am confused by the command psmatch2, teffects, svy:logit and glm and don't know which to use.
My question is: if I multiply propensity score weight by survey weight like the post mentioned above said, and want to look at the ATT with a model of good extrapolation, which command should I use for matching?
(I read the Appendix B of DuGoff et al. but it makes me even more confused. What does "subpop" mean in the svy command? I probably will not involve a subpopulation in the analysis.)
Thank you!
Sincerely,
Zhijing

[Steve Samuels]

Welcome to Statalist, Zhiling!

Be sure to read and follow the FAQ, especially FAQ 12, which asks that you paste code, results, and data listings between CODE delimiters. .

I haven't done much matching in recent years,, so cannot recommend specific commands. However propensity score matching has serious problems in my opinion, based on this article, which says that Mahalanobis and Coarsened Exact Matching (CEM) are preferable. You can get the contributed command cem by typing in Stata

ssc install cem

See also this article by Ridgeway and others about how to combine propensity and sampling weights.


Ridgeway, Greg et al. “Propensity Score Analysis with Survey Weighted Data.” Journal of causal inference 3.2 (2015): 237–249. PMC. Web. 11 June 2018.
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5802372/

[Zhijing Cai]

Welcome to Statalist, Zhiling!

Be sure to read and follow the FAQ, especially FAQ 12, which asks that you paste code, results, and data listings between CODE delimiters. .

I haven't done much matching in recent years,, so cannot recommend specific commands. However propensity score matching has serious problems in my opinion, based on this article, which says that Mahalanobis and Coarsened Exact Matching (CEM) are preferable. You can get the contributed command cem by typing in Stata

ssc install cem

See also this article by Ridgeway and others about how to combine propensity and sampling weights.


Ridgeway, Greg et al. “Propensity Score Analysis with Survey Weighted Data.” Journal of causal inference 3.2 (2015): 237–249. PMC. Web. 11 June 2018.
https://www.ncbi.nlm.nih.gov/pmc/articles/PMC5802372/

Thank you so much, Steve! I read Blackwell et al.'s paper about CEM and found it very useful. Now I am using CEM instead of propensity score matching.
I also found a useful post here discussing how to incorporate sampling weight into CEM. After I multiplied the sampling weight and cem_weight and regress using pweight, the number of observations dropped a lot compared to not using [pw= newweight ]. My code is like this:

cem covariates , treatment(treatment)
gen newweight= cem_weights* samplingweight
regress logwage treatment [pw= newweight ]

But compared to the propensity score matching, which drop a large part of observations in my treatment group, CEM is more satisfactory since it helps match 90% of the observations in the treatment group.
Thank you again, Steve!

[Steve Samuels]

Great! I'm glad to hear it worked out for you.