Nonparametric Multiple Imputation

• Activity Number: 41
• Type: Contributed
• Date/Time: Sunday, August 3, 2014 : 2:00 PM to 3:50 PM
• Sponsor: Survey Research Methods Section
• Author(s): Darryl Creel*+
• Companies: RTI International
• Keywords: multiple imputation ; approximate Bayesian Bootstrap ; decision trees ; prediction model ; response propensity model ; double protection

Abstract

Multiple imputation (MI) approaches like sequential-regression multivariate imputation and MI by chained equations often use a series of generalized linear models to create posterior distributions from which the imputed values are selected. In general, one must specify the correct type and functional form of the generalized linear model at some basic level for each variable imputed. Nonparametric multiple imputation (NMI), by contrast, eliminates much of the problematic specifications. NMI also incorporates a response propensity model in the formation of the posterior distributions and thus provides double protection against model misspecification. A Monte Carlo simulation evaluates the performance of NMI comparing it to MI using the MICE package in R. The evaluation criteria are bias, confidence interval length, and coverage of the 95% confidence intervals.