Robust Multiple Imputation for Discrete Data in Missing-by-Design Settings

• Activity Number: 115
• Type: Topic Contributed
• Date/Time: Monday, August 1, 2011 : 8:30 AM to 10:20 AM
• Sponsor: Section on Survey Research Methods
• Author(s): Florian Meinfelder*+ and Trivellore Raghunathan
• Companies: Universität Bamberg and University of Michigan
• Keywords: multiple imputation ; missing data ; statistical matching ; data fusion

Abstract

Multiple Imputation (Rubin 1978, 1987) is a generally accepted method for analyzing incomplete data sets. Missing values are 'filled-in' or imputed m>1 times, thus creating m completed different data sets which are identical in the observed part, but vary over the imputed part. Most models and applications still focus on the imputation of continuous variables, and are usually based on normal distribution assumptions. However, survey data typically feature a large percentage of discrete data with non-normal distributions. This work addresses the multiple imputation of such variables in missing-by-design patterns (e.g. data fusion), where 'blocks' of data are missing. We propose Predictive Mean Matching (Rubin 1986) for vectors of variables as described by Little (1988) in combination with a Bayesian Bootstrap to create multiple imputations. An additional imputation step is needed, if the donor parts have missing values as well. This data situation can be seen as a mixture of missing-by-design with an overlaying ordinary item-nonresponse.