Imputation and Unbiased Estimation: Use of the Centered Predictive Mean Neighborhoods Method

• Activity Number: 28
• Type: Contributed
• Date/Time: Sunday, August 8, 2004 : 2:00 PM to 3:50 PM
• Sponsor: Section on Survey Research Methods
• Author(s): Avinash Singh and Eric Grau*+ and Ralph Folsom
• Companies: RTI International and RTI International and RTI International
• Keywords: predictive mean matching ; nearest neighbor imputation ; multivariate imputation ; predictive mean neighborhoods ; unbiased estimation

Abstract

Methods for determining the predictive distribution for multivariate imputation range between two extremes, both of which are commonly employed in practice: a completely parametric model-based approach, and a completely nonparametric approach such as the nearest neighbor hot-deck (NNHD). A semiparametric middle ground between these two extremes is to fit a series of univariate models and construct a neighborhood based on the vector of predictive means. This is what is done under the predictive mean neighborhoods (PMN) method, a generalization of Rubin's predictive mean matching method. Because the distribution of donors in the PMN neighborhood may not be centered at the recipient's predictive mean, estimators of population means and totals could be biased. To overcome this problem, we propose a modification to PMN which uses sampling weight calibration techniques such as the GEM (generalized exponential model) method of Folsom and Singh to center the empirical distribution from the neighborhood. Empirical results on bias and MSE, based on a simulation study using data from the 2002 National Survey on Drug Use and Health, are presented to compare the centered PMN with other methods.