Multiple Imputation of Potential Outcomes

• Activity Number: 155
• Type: Topic Contributed
• Date/Time: Monday, August 4, 2014 : 10:30 AM to 12:20 PM
• Sponsor: Section on Bayesian Statistical Science
• Author(s): Stef van Buuren*+ and Paula van Dommelen and Hedwig Hofstetter and Donald B. Rubin
• Companies: Netherlands Organisation for Applied Scientific Research and Netherlands Organisation for Applied Scientific Research and Netherlands Organisation for Applied Scientific Research and Harvard
• Keywords: causal analysis ; observational data ; missing data ; MICE ; ANCOVA ; simulation

Abstract

I will discuss the usefulness of multiple imputation of potential outcomes to obtain causal estimates from observational data. The method assumes that the treatment assignment mechanism is ignorable, that the stable unit treatment value assumption (SUTVA) holds, and that the average treatment effect can be calculated for people to which one of the treatments is applied in the future. Drawing multiple imputations is done by the general-purpose software MICE. Simulations show that the technique provides valid estimates of the treatment effect when the assumptions are met. Bias and precision are asymptotically (m = ?) equal to closed form ANCOVA, and better if there are missing data in the covariates. Multiple imputation of the potential outcomes appears as a viable and flexible technique to estimate causal effects from observational data.