Estimating Population Treatment Effects from a Survey Subsample

• Activity Number: 325
• Type: Topic Contributed
• Date/Time: Tuesday, August 5, 2014 : 10:30 AM to 12:20 PM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Kara Rudolph*+ and Ivan Diaz and Michael Rosenblum and Elizabeth A. Stuart
• Companies: Johns Hopkins Bloomberg School of Public Health and Johns Hopkins Bloomberg School of Public Health and Johns Hopkins Bloomberg School of Public Health and Johns Hopkins Bloomberg School of Public Health
• Keywords: causal inference ; survey ; inverse probability weighting ; targeted maximum likelihood estimation

Abstract

We consider the problem of estimating an average treatment effect for a target population from a survey subsample. Our motivating example is generalizing the effect of living in a disadvantaged neighborhood on cortisol estimated from a subsample of a nationally representative survey to the population of U.S. adolescents. Inverse probability weighting (IPW) remains the most widely used method for accomplishing this task--possibly because it is straightforward to implement. However, IPW has well-known efficiency problems that may be exacerbated in our scenario, which consists of 1) survey weights, 2) subsample selection, 3) non-random treatment assignment, and 4) effect heterogeneity. Using simulation we compare its performance to two other estimators that are similarly straightforward to implement but have not been as widely disseminated--double-robust weighted least squares (DRWLS) and targeted maximum likelihood (TMLE) estimators. We demonstrate that DRWLS and TMLE outperform IPW in terms of bias, variance, and mean squared error under various data generating distributions and model misspecifications. We also illustrate the methods in the context of our motivating example.