Constructing Final Weights With IPTW and Survey-Maker Weights

[Claire Cahen]

Hi everyone,

I am trying to estimate the effect of a policy using the yearly ACS (American Community Survey) data files. I have coded the policy in question as a multi-valued treatment (with 5 possible values) and constructed IPTW to estimate its effect(s) (using regular OLS-a modified DiD, specifically). To ensure (double) robustness, I included co-variates in both the construction of my generalized propensity scores and in my final regression with IPTW. However, I am unsure what to do at this point with the PERWT the ACS provides. I think I need to find a way to construct some kind of final weight that takes both the survey-maker PERWT and my IPTW into account. How can I do this? I've looked into raking algorithms but these seem to require starting from scratch rather than building on pre-existing weights. Given the general nature of my question, I'm not including my code here, but can add it if it helps. Any advice would be greatly appreciated.

Thank you,

Claire

[Steve Samuels]

I know that this question has been studied for single treatment propensity score studies, but the recommendations differ. Dugoff et al. (2014) have a decision tree that depends on the causal estimand of interest and the target population. They believe that survey weights should be covariates in the propensity model but should not weight the model. For a weighted survey analysis, they recommend a final weight that is the product of the propensity weight and the sampling weight.

In contrast, Ridgeway et al. (2018) state that the the propensity score should be weighted with the survey weights. Their recommendation for the weighted survey analysis is the same as Dugoff's.

References:

DuGoff EH, Schuler M, Stuart EA. Generalizing observational study results: applying propensity score methods to complex surveys. Health Serv Res. 2014; 49:284–303. [PubMed: 23855598]
http://europepmc.org/articles/pmc3894255

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/

[Claire Cahen]

Hi Steve,

Thank you so much for these references-- they were very helpful. I ended up using survey weights in both the construction of my IPW and my final weight (a product of the IPW and survey weights).

Claire

[Adam Alhassan]

Hi Claire and Steve,

I found this post extremely useful for my current project. I was wondering if there was any further work in this area. Is there consensus that Ridgeway et al.'s approach is the way to go?

Thank you so much for the references. I love statalist!

Adam