Local Synthesis for Disclosure Limitation via Model-Based Clustering

• Activity Number: 462
• Type: Contributed
• Date/Time: Wednesday, August 6, 2014 : 8:30 AM to 10:20 AM
• Sponsor: Government Statistics Section
• Author(s): Anna Oganyan*+
• Companies: NCHS
• Keywords: synthetic data ; mixture model ; Expectation-Maximization (EM) ; hybrid SDL method ; latent class regression model

Abstract

Medical data records often contain sensitive information about the data subjects. Before releasing such data, e.g. for clinical research purposes, data owners have to apply Statistical Disclosure Limitation (SDL) methods to such data. SDL methods often consist of masking or synthesizing the original data records in such a way to minimize the risk of disclosure of the confidential information and at the same time provide legitimate data users with accurate information about the population of interest. In this paper we propose a new scheme for disclosure limitation which is based on the idea of local synthesis of data. We argue that the procedures of dividing the records in homogeneous groups (the ``local" part) and synthesizing the records in the groups should be carefully chosen, so that clustering and synthesis would ``fit" each other in the best possible way. Our approach to this problem is based on model-based clustering. Our experiments with genuine medical data sets show that local synthesis is superior to other methods considered for comparison including synthetic data generated using the sequential regression approach, because it can preserve complex relatio