On Obtaining Correct Variance for Hot-Deck Imputation

Abstract

Hot-deck imputation is popular method in handling nonresponse in survey by filling in for each subject with missing data. A great difficulty associated with hot-deck imputation is how to give a valid inference when the ``completed!.. data is being treated as complete data. Rubin (1978) proposed to use multiple imputation as a way of getting valid inference. Reilly and Pepe (Statistics in Medicine, 1997) showed that the equivalence of mean-score (Reilly and Pepe, Biometrika, 1995) and hot-deck when the fully observed variables are categorical. They suggest the use of variance expression of mean- score method for getting valid inference of multiple hot-deck imputation. On the other hand, Rubin and Schenker (1986) gave a justification of Rubin!|s proposal. In this talk, we will investigate Rubin!|s proposal in more complicated Model than those discussed in Rubin and Schenker (1986).