A Practical Method for Estimating Population Average Treatment Effects in Observational Studies

Sharon Christ , Purdue University
Luhao Wei, Purdue University

Theory and method development for strengthening causal validity in observational studies has flourished in recent decades. In the population and life-course sciences, most exposures (“treatments”) of interest involve non-random selection into exposure groups. Quasi-experimental methods have been developed that generally mimic randomized experimental design through creating synthetic balance across exposure groups. Among others, two of these methods include propensity score matching (PSM) and inverse-probability weighting (IPW). Yet, research has shown that these methods do not necessarily perform better than traditional statistical control on variables that induce selection. In addition, there has been little work on establishing external validity of the sample average treatment effects (SATEs) to population average treatment effects (PATEs) using these methods. Population science is mostly focused on external validity and utilizes random selection into samples along with corrections for complex sample design features such as sample weights and cluster and strata corrected standard error estimators. In this study, we utilize a practical method of traditional model control on covariate confounds of the exposure-outcome relationship and centering to estimate the PATE. The method allows for estimation corrections for unequal sample selection, clustering, and stratification thereby retaining external validity to the population. We compare the traditional control method to PSM and IPW methods empirically using a large, longitudinal representative sample of U.S. residents and using a simulation study. The flexibility of the proposed method for estimating PATEs for varying subpopulations and both categorical and continuous treatment is also demonstrated.

See extended abstract

 Presented in Session P3. Migration, Economics, Policies, History