Polydesigns in Studies of Causal Effects

• Activity Number: 213
• Type: Contributed
• Date/Time: Tuesday, August 5, 2003 : 8:30 AM to 10:20 AM
• Sponsor: Section on Statistics in Epidemiology
• Author(s): Fan Li*+ and Ravi Varadhan and Constantine E. Frangakis
• Companies: Johns Hopkins University and Johns Hopkins University and Johns Hopkins University
• Keywords: polydesign ; causal effects ; needle exchange program

Abstract

In an increasingly common class of studies, the goal is to evaluate a causal effect using treatments that are only partially controlled by the investigator. Such studies have two conflicting features: (1) a careful model on the full design identifies the causal effect, but its extrapolation to extreme regions of the data (e.g., where exposure to a treatment is rare) can be sensitive to model mispecification, (2) the likelihood on a reduced design (e.g., conditional likelihood on matched subsets of data), which is typically more robust to misspecification, does not full the causal effect in these studies. We propose using both the full and the reduced designs of the data. We show that using such polydesigns creates a rich class of methods that can identify the causal effects and also retain robustness to model misspecification. We discuss two methods of analysis of polydesigns, and draw connections with, as well as point differences from, composite likelihoods and mixture models. We discuss applications of polydesigns in evaluating needle exchange programs, as well as in more general studies.