Analyses That Inform Policy Decisions Are, De Facto, Causal

• Activity Number: 2
• Type: Invited
• Date/Time: Sunday, August 4, 2013 : 2:00 PM to 3:50 PM
• Sponsor: ENAR
• Author(s): Roee Gutman*+ and Donald B. Rubin
• Companies: Brown University and Harvard University
• Keywords: Rubin Causal Model ; Observational Studies ; Bayesian adjustment for confounding ; Multiple imputation

Abstract

To inform major policy-related decisions, it is imperative that the goal of a statistical procedure be to estimate the effects of a policy on outcomes of interest; that is, the goal must be to estimate the causal effects of interventions on outcomes. A common statistical approach for causal estimation is based on Rubin Causal Model (RCM). The RCM is a framework that was partially introduced by Neyman (1923) in the context of randomized experiments, and was extended by Rubin (1978) to include observational studies and Bayesian inference. Wang, Parmigiani, and Dominici (2012) proposed a Bayesian method to estimate the association between contaminants and hospitalization rates, while controlling for "confounders" (called "BAC"). BAC is claimed to be aimed at informing major policy-related decisions. Using the RCM, we describe the implicit and explicit assumptions made by BAC when attempting to draw causal conclusions, and we discuss limitations of BAC-like methods for addressing causal questions. Also, we briefly describe a method that directly confronts the casual question by multiply imputing the unobserved potential outcomes, thereby addressing the possible pitfalls of BAC.