A Two-Stage Imputation Procedure to Balance the Risk–Utility Trade-Off

Abstract

There has been little discussion in the literature on how many multiply imputed datasets an agency should release. From the perspective of the secondary data analyst, a large number of datasets is desirable. The additional variance introduced by the imputation decreases with the number of released datasets. For example, Reiter (2003) finds nearly a 100% increase in the variance of regression coefficients when going from 50 to two partially synthetic datasets. From the perspective of the agency, a small number of datasets is desirable. The information available to illintentioned users seeking to identify individuals in the released datasets increases with the number of released datasets. Thus, agencies considering the release of partially synthetic data generally are confronted with a trade-off between disclosure risk and data utility.

Most of this chapter is taken from Drechsler and Reiter (2009) and Reiter and Drechsler (2010).