Robust Sensitivity Analysis Using Bayesian Nonparametric Modeling

• Activity Number: 510
• Type: Invited
• Date/Time: Wednesday, August 7, 2013 : 10:30 AM to 12:20 PM
• Sponsor: Section on Nonparametric Statistics
• Author(s): Nicole Carnegie and Jennifer Hill*+
• Companies: Harvard University and New York University
• Keywords: sensitivity analysis ; BART ; Bayesian nonparametrics ; causal inference ; omitted variables

Abstract

Non-randomized studies are always suspect when used to estimate causal effects because the researcher can never ascertain whether the key assumption that all confounders have been measured (ignorability) has been met. One way to alleviate this concern is to evaluate the sensitivity of causal inferences to the presence of an unknown confounder. An increasingly popular set of approaches define this omitted variable through two sensitivity parameters that represent some form of (conditional) association between 1) the omitted variable and the treatment, and 2) the omitted variable and the outcome. However these methods almost invariably rely on traditional parametric models both to reflect these associations as well as to estimate the adjusted treatment effect estimates. This paper extends this framework to allow for more complex, potentially non-linear relationships between the outcome and treatment and the other confounding covariates when specifying conditional associations as well as when estimating the adjusted treatment effects.