Bias reduction of finite population imputation by kernel methods

Abstract

Missing data is a nuisance in statistics. Real donor imputation can be used with item nonresponse. A pool of donor units with similar values on auxiliary variables is matched to each unit with missing values. The missing value is then replaced by a copy of the corresponding observed value from a randomly drawn donor. Such methods can to some extent protect against nonresponse bias. But bias also depends on the estimator and the nature of the data. We adopt techniques from kernel estimation to combat this bias. Motivated by Pólya urn sampling, we sequentially update the set of potential donors with units already imputed, and use multiple imputations via Bayesian bootstrap to account for imputation uncertainty. Simulations with a single auxiliary variable show that our imputation method performs almost as well as competing methods with linear data, but better when data is nonlinear, especially with large samples.