A Discussion of Uncongeniality for Synthetic Data

• Activity Number: 665
• Type: Contributed
• Date/Time: Thursday, August 2, 2012 : 10:30 AM to 12:20 PM
• Sponsor: Section on Government Statistics
• Author(s): Bronwyn Loong*+ and Carl Morris and Donald B Rubin
• Companies: Harvard University and Harvard University and Harvard University
• Keywords: data confidentiality ; synthetic ; multiple imputation ; uncongeniality ; data utility

Abstract

To limit the risk of disclosing survey respondents' identities or sensitive attributes, statistical agencies may release multiply imputed synthetic data sets in place of the observed data to external users. Given the statistical agency and external user are separate bodies, their sources of input will potentially be different. This is known in the multiple imputation literature as uncongeniality. In this paper, we present a formal definition of uncongeniality for multiple imputation for synthetic data. We use this framework to address common examples of uncongeniality, specifically ignorance of the original survey design in analysis of synthetic data, and when the imputation model conditions upon a different set of records to those analyzed in the analysis procedure. We conclude the formal definition assists the imputer to identify the source of a lack of data utility preservation between observed and synthetic data analytic results. Motivated by our definition of congeniality, we derive and illustrate an alternative approach to synthetic data inference to recover the observed data sampling distribution given the synthetic data.