Imputation via Triangular Regression-Based Hot Deck

摘要

In principle, hot deck imputation methods preserve means and variances, and can also preserve covariances with other variables included in the allocation matrix. In practice, dimensionality problems arise quickly as predictive variables are added and allocation matrix cells become small, undermining the hot decks theoretical advantages. Predictive-mean nearest-neighbor imputation avoids dimensionality problems, but can reduce the variance. A combination method is described: using the predicted values from a set of sequential, triangular regressions to form hot deck matrices. Triangularity allows the inclusion of predictive variables that are themselves subject to non-response. The method enables the rapid development of allocation schemes, eliminates dimensionality problems,and aids in predicator selection. The implementation of this method in American Housing Survey income data is described and evaluated.