Multiple Imputation under Multivariate gh Family of Distributions

• Activity Number: 146
• Type: Contributed
• Date/Time: Monday, August 8, 2005 : 10:30 AM to 12:20 PM
• Sponsor: Section on Survey Research Methods
• Author(s): Yulei He*+ and Raghunathan E. Trivellore
• Companies: University of Michigan and University of Michigan
• Keywords: Missing data ; multiple imputation ; multivariate non-normal data ; Tukey's gh distribution

Abstract

Missing data are a pervasive problem in many scientific investigations involving human populations. A popular approach for analyzing incomplete data is through multiple imputation, which creates "completed datasets" by "filling in" missing items by sets of plausible values. Each completed dataset is analyzed separately; the point and variance estimates are then combined to form a single inference. The imputations usually are drawn from the posterior predictive distribution of the set of missing values under an explicit or implicit model. Many explicit model-based imputation procedures typically use multivariate normal distribution either on the original or transformed scale. However, in many practical applications, these models may not be appropriate, so normal model-based multiple imputation inferences may not be valid. In this paper, we use an extension of Tukey's gh distribution to multivariate nonnormal data to account for skewness and kurtosis. We then develop a multiple imputation method based on this multivariate distribution family. The performance of the proposed method is evaluated through simulations.