Selection of Propensity Functions for Causal Inference

• Activity Number: 392
• Type: Topic Contributed
• Date/Time: Tuesday, August 3, 2010 : 2:00 PM to 3:50 PM
• Sponsor: Biometrics Section
• Author(s): Wei (Peter) Yang*+ and Marshall M. Joffe and Sean Hennessy
• Companies: University of Pennsylvania and University of Pennsylvania and University of Pennsylvania
• Keywords: Propensity score ; Continuous treatment ; Confouding

Abstract

Propensity scores (PS) method is widely used to control for confounding of the effect of a binary treatment in observational studies. They have been generalized to ordinal and continuous treatments (Imai 2004, Imbens 2000, Joffe 1999) in recent literature. Following the definition of propensity functions in Imai 2004, we explored the sufficient conditions of selecting propensity functions to estimate the causal effect. We showed that the propensity functions that are required to get unbiased estimates of the causal treatment effect are only those that correspond to the parameters characterizing the expectation of the variables in the causal model. In addition, we showed that using propensity functions alone is not sufficient when there are interactions between the treatment and baseline covariates in the causal model.