Multiple Imputation Using Sequential Hierarchical Regression Models

• Activity Number: 414
• Type: Contributed
• Date/Time: Wednesday, August 6, 2003 : 2:00 PM to 3:50 PM
• Sponsor: Section on Survey Research Methods
• Author(s): Recai Murat Yucel*+ and Nathaniel Schenker and Trivellore E. Raghunathan
• Companies: Institute for Health Policy and National Center for Health Statistics and Institute for Social Research
• Keywords: cluster ; general location model ; item nonresponse ; random effect ; multiple imputation

Abstract

Multiple imputation is an increasingly popular method for handling item nonresponse in surveys. When using multiple imputation, it is beneficial to reflect the sample design in the imputation model. If the sample design involves clustering, one way to account for cluster effects is via random effects in the imputation model. Although this idea has been developed in detail for imputing continuous variables, it is less well-developed for imputing mixtures of categorical and continuous variables. We propose two approaches to producing multiple imputations for such variables. The first approach uses the general location model with random effects on both continuous and categorical variables, extending the work by Raghunathan and Grizzle (1995). This approach follows the fully model-based paradigm for multiple imputation. In highly multivariate problems, however,this approach becomes problematic. For such situations, we extend the methods given by Raghunathan et al.(2001), in which we produce imputations by fitting sequential hierarchical models and by drawing missing values variable-by-variable from these models. We illustrate and compare these techniques using simulated data.