Disclosure Risk and Data Utility for Partially Synthetic Data: An Empirical Study Using the German IAB Establishment Survey

Statistical agencies that disseminate data to the public must protect the confidentiality of respondents' identities and sensitive attributes. To satisfy these requirements, agencies can release the units originally surveyed with some values, such as sensitive values at high risk of disclosure or values of key identifiers, replaced with multiple imputations. These are called partially synthetic data. In this article, we empirically examine trade-offs between inferential accuracy and confidentiality risks for partially synthetic data, with emphasis oil the role of the number of released datasets. We also present a two-stage imputation scheme that allows agencies to release different numbers of imputations for different variables. This scheme can result in lower disclosure risks and higher data utility than the typical one-stage imputation with the same number of released datasets. The empirical analyses are based oil partial synthesis of the German IAB Establishment Survey.