Sequential Predictive Mean Matching Method of Multiple Imputation

• Activity Number: 55
• Type: Topic Contributed
• Date/Time: Sunday, August 7, 2005 : 4:00 PM to 5:50 PM
• Sponsor: Section on Health Policy Statistics
• Author(s): Trivellore Raghunathan*+ and Florian Koller and Nathaniel Schenker and Susanne Raessler
• Companies: University of Michigan and GfK Custom Research, Inc. and National Center for Health Statistics and University of Erlangen-Nurenberg
• Keywords: Predictive Mean Matching ; Bayesian Bootstrap ; Sequential Regression ; Simulation ; Complex surveys

Abstract

Missing data are ubiquitous in many scientific investigations involving human populations. A popular method for handling incomplete data is through multiple imputation, where the set of missing values is "filled-in" by several sets of plausible values to create completed datasets. Each completed dataset is analyzed separately and the point estimates and standard errors are combined to construct a single inference. The multiple imputation is best justified when imputations are drawn from the posterior predictive distribution of missing values under an implicit or explicit model. In practical problems with a large number of variables with varying types and complex data structures such as logical relationships, skip patterns, etc., developing a model is difficult. In a sequential regression approach, imputations are drawn from posterior predictive distribution corresponding to a sequence of conditional regression models. The logical relationships and skip patterns are incorporated in these regression models. In this paper, we use the same framework, except the imputations are obtained using the predictive mean matching method through a sequence of conditional regression models.