Robust Multiple Imputation Based on Bayesian Bootstrap Predictive Mean Matching

• Activity Number: 413
• Type: Contributed
• Date/Time: Tuesday, August 3, 2010 : 2:00 PM to 3:50 PM
• Sponsor: Section for Statistical Programmers and Analysts
• Author(s): Florian Koller-Meinfelder*+
• Companies: Universität Bamberg
• Keywords: Missing Data ; Multiple Imputation ; Predictive Mean Matching ; Fully Conditional Specification

Abstract

Multiple Imputation has become a generally accepted way to handle statistical analysis of incomplete data. This paper describes an MI procedure that combines an FCS (fully conditional specification) approach with Bayesian Bootstrapping and Predictive Mean Matching (PMM). The algorithm bears similarities to IVEware and MICE, but in contrast to these software packages, multiple imputations are consequently based on variants of PMM for all variable types. An additional benefit is that the algorithm ensures that imputed values are plausible and more robust to model misspecification than purely parametric MI algorithms, as PMM is a nearest neighbor matching technique. The algorithm is therefore particularely suited to impute missing survey data which typically contain mixed-scale, non-normal and bounded discrete variables.