DATA-ANALYSIS USING HOT DECK MULTIPLE IMPUTATION

Covariate data which are missing or measured with error form the subject of a growing body of statistical literature. Parametric methods have not been widely adopted, quite possibly due to the necessity of specifying the form of a 'nuisance function' not required for complete data analysis, and the non-robustness of the methods to mis-specification. A non-parametric counterpart of multiple imputation, known as 'hot deck', was proposed by Rubin (1987) and has been used by the Census Bureau to complete public-use databases. However, inference using this method has not been possible due to the distribution theory not being available. Recently, it has been shown that the hot deck estimator has the same asymptotic distribution as the 'mean score' estimator, so that inference using hot deck is now possible. The method is intuitively appealing and easily implemented. Furthermore, it accommodates missingness which depends on outcome, which is an important generalization of many currently available methods. In this paper. the hot deck multiple imputation method is explained, its asymptotic distribution presented and its application to data analysis demonstrated by an example.