300 – Innovations in and Applications of Imputation
Imputation in the American Housing Survey: Comparing Multiple Imputation With Current Hot Deck Methods
Stephen Ash
U.S. Census Bureau
Sean Dalby
U.S. Census Bureau
Gregory Mulley
U.S. Census Bureau
Kathy Zha
U.S. Census Bureau
Multiple imputation is an active field of statistical research, encompassing a wide variety of modeling methods with different strengths and weaknesses. This paper provides a theoretical overview and empirical comparisons between multiple imputation – specifically, Fully Conditional Specification – and traditional hot deck imputation in the context of the American Housing Survey. The hot deck method stratifies across demographic and housing-level characteristics to form donor cells of similar housing units. Unlike the hot deck, Fully Conditional Specification methods allow for a wider array of models, and provide well-researched methods for estimating the amount of variance introduced by imputation itself. Both are compared against hurdles present within the American Housing Survey, including numerous structural zeros, thereby highlighting the benefits and trade-offs of each imputation approach.