Mixture Modeling of Treatment Effects with Multiple Compliance Classes and Missing Data

• Activity Number: 326
• Type: Invited
• Date/Time: Tuesday, August 2, 2011 : 10:30 AM to 12:20 PM
• Sponsor: Social Statistics Section
• Author(s): Michael Sobel*+ and Bengt Muthen
• Companies: Columbia University and University of California at Los Angeles
• Keywords: causal inference ; complier average causal effect

Abstract

Randomized experiments are the gold standard for making causal inferences. Researchers design treatments to affect mediators lying on one or more presumed pathways to the outcome. Investigators typically want to know the effect of offering the treatment and also the effect of the treatment itself. To address the latter question, recent attention has focused on the effect among subjects who will comply with their treatment assignment. In many cases, there is little reason to believe that the mediators targeted by the treatment will produce effects for all complier subjects. Therefore, we estimate the proportion of compliers unaffected by treatment as well as the proportion affected and the effect. Missing data further complicate estimation and we consider various missing data assumptions, including the assumption that the missing data are missing at random and the assumption of latent ignorability.