Speaker: Xiaofei Zhang
Abstract: Nearest neighbor imputation is a hot deck imputation method that is widely used to complete records for a sample with missing records in survey sampling. The direct nearest neighbor imputation estimator suffers from bias that increases as the dimension of the covariate increases. In this article, we give a model-consistent estimator of the mean and a variance estimator of the estimated mean for surveys where the missing probabilities may not be known. When the model is misspecified, the proposed estimator is consistent for the mean under certain conditions. The estimator of the mean is model-superior to the direct nearest neighbor estimator. We also give a model-consistent replication variance estimator which does not require repeated imputation. The simulation results match the theoretical results.