Multiple Imputation for causal Inference

• Activity Number: 143
• Type: Contributed
• Date/Time: Monday, August 2, 2010 : 8:30 AM to 10:20 AM
• Sponsor: Section on Survey Research Methods
• Author(s): Irina Bondarenko*+ and Trivellore Raghunathan
• Companies: University of Michigan and University of Michigan
• Keywords: Causal inference ; Randomized Studies ; Potential Outcomes ; Direct effect ; Mediation ; Observational Studies

Abstract

The potential outcome framework for causal inference is fundamentally a missing data problem with a special "file-matching" pattern of missing data. This article uses the sequential regression or chained equation methodology to impute the potential outcomes based on the observed data. The causal inference parameters are formulated based on the models for the completed data and standard multiple imputation (MI) combining rules are applied to infer about the direct and mediated effects. MI framework is modified to incorporate constraints or prior information in terms of augmented complete-data. Given the ability of the multiple imputation framework to handle several types of variables, missing values in covariates and the availability of software for performing multiple imputations, this approach makes easier to perform causal inference from both observational and randomized studies.