|
Abstract:
|
Hot deck imputation is a nonparametric imputation method which replaces missing values with observed responses from "similar" units, often from within imputation cells. The hot deck has long been used by the U.S. Census Bureau and is one of the most common imputation methods for survey data. When the hot deck is used for single imputation, explicit variance formulae exist for simple cases that may not be realistic in practice. Resampling methods have also been proposed, including both jackknife and bootstrap approaches. However, when used as the basis for multiple imputation, hot deck imputation is improper, since simply drawing multiple donors per missing value from an imputation cell does not fully propagate uncertainty. Modifications to the hot deck have been proposed that make it proper, including the Bayesian Bootstrap and the Approximate Bayesian Bootstrap. In this talk I will review and contrast these various strategies for obtaining valid standard errors after hot deck imputation through a detailed example. I will also discuss ongoing work to use the newly proposed maximum likelihood multiple imputation (MLMI) in the context of hot deck imputation.
|
