A Data-Adaptive Approach for Modeling Propensity Scores

• Activity Number: 524
• Type: Contributed
• Date/Time: Wednesday, August 3, 2011 : 10:30 AM to 12:20 PM
• Sponsor: Biometrics Section
• Author(s): Yeying Zhu*+ and Debashis Ghosh
• Companies: Penn State University and Penn State University
• Keywords: Causal inference ; Logistic regression ; Random forests ; Two-stage modeling ; Bias-variance ; Observational data

Abstract

In non-randomized observational studies, estimated differences between treatment groups may arise not only due to the treatment but also because of the masking effect of confounders. To adjust for confounding due to measured covariates, the average treatment effect is often estimated conditioning on propensity scores. In the literature, propensity scores are usually estimated by logistic regression. Alternatively, one can employ non-parametric classication algorithms, such as various tree-based methods or support vector machines. In this talk, we explore the effect of classification algorithms used to model propensity scores using ideas of bias and variance. In addition, we explore ways to combine logistic regression with nonparametric approaches to estimate propensity scores. Simulation studies are used to assess the performance of the newly proposed method and a data analysis example is presented in the end.