Abstract
Nearest neighbor imputation is a popular procedure used to complete missing records. However, the nearest neighbor imputation estimator suffers from a bias that increases as the dimension of the matching variable increases. We study estimators that are model unbiased and have the hot-deck property. Estimation for domains is discussed. In the simulation study for the population mean, a proposed estimator is less biased and more efficient than the nearest neighbor estimator. The suggested replication variance estimator does not require repeated imputation and is appropriate for vectors with different missing patterns. In the simulation, the variance estimator gives confidence intervals with coverage rates close to the nominal level.