Multiple Imputation of High-Dimensional Mixed Incomplete Data

• Activity Number: 297
• Type: Contributed
• Date/Time: Tuesday, August 2, 2011 : 8:30 AM to 10:20 AM
• Sponsor: Biometrics Section
• Author(s): Ren He*+ and Thomas R. Belin
• Companies: University of California at Los Angeles and University of California at Los Angeles
• Keywords: Multiple Imputation ; Hierarchical prior ; MCMC ; Parameter-Extended algorithm

Abstract

It is common in applied research to have large numbers of variables with mixed data types (continuous, binary, ordinal or nomial) measures on a modest number of cases. Also, even a simple imputation model can be overparameterized when the number of variables is moderately large. Finding a joint model to accommodate multivariate data with mixed data types is challenging. Here we develop a joint multiple imputation model with multivariate normal components for continuous variables and latent-normal components for categorical variables. Following the strategy of Boscardin and Weiss (2003) and using Parameter-expanded Metropolis-Hastings estimation (Boscardin,Zhang and Belin 2008), we use a hierarchical prior for the covariance matrix centered around a parametric family. This not only substantially reduces the dimension of the parameter space but also allows the data to depart from a tightly defined structured covariance matrix. The method is compared in several simulation settings to available-case analysis and a rounding method.