Nearest-Neighbor--Based Approaches for Multiple Imputation of Unordered Categorical Variables

• Activity Number: 283
• Type: Topic Contributed
• Date/Time: Tuesday, August 5, 2014 : 8:30 AM to 10:20 AM
• Sponsor: Survey Research Methods Section
• Author(s): Florian Meinfelder*+
• Companies: University of Bamberg
• Keywords: missing data ; multiple imputation ; predictive mean matching ; missing at random

Abstract

Multiple Imputation (MI) (Rubin 1978, 1987) is a general-purpose approach to allow for statistical analysis of incomplete data. While survey data sets were among the first data types which MI was applied to, the first MI algorithms had difficulties in dealing with nominal-scale variables which occur frequently in survey data. The currently implemented methods typically use an underlying multinomial logit or probit model, but there is some indication that semi- or non-parametric methods might be more robust to model misspecification. Predictive Mean Matching (PMM) (Rubin 1986; Little 1988) is a nearest-neighbor approach that is used within MI algorithms for metric-scale data. We propose several PMM variants for unordered categorical variable types, and compare them to parametric MI algorithms based on the results of an MC study using simulated and empirical data.