This research project will develop sound statistical and machine learning techniques for preserving privacy with linked data. Social entities and their patterns of behavior is a crucial topic in the social sciences. Research in this area has been invigorated by the growth of the modern information infrastructure, ease of data collection and storage, and the development of novel computational data analyses techniques. However, in many application areas relevant and sensitive information is commonly located across multiple databases. Data analysis is inherently impossible without merging databases, but at the cost of increasing the risk of a privacy violation. This research will address the problem of how to perform valid statistical inference in the presence of multiple data sources, data sharing, and privacy in the age of "big data." The investigators' new modeling construct for inference and uncertainty quantification will contribute to both statistics and the many disciplines for which statistics is a principal tool. The methods will have a wide range of applications in the social, economic, and behavioral sciences, including medicine, genetics, official statistics, and human rights violations. The investigators will collaborate with post-doctoral researcher and with graduate and undergraduate students. The statistical methods will be encapsulated in open-source software packages, allowing off-the-shelf use by practitioners while facilitating more detailed control and extensions.
This interdisciplinary research project will improve upon methods in record linkage and privacy using state-of-the-art techniques from statistics and machine learning. Record linkage is the process of merging possible noisy databases with the goal of removing duplicate entries. Privacy-preserving record linkage (PPRL) tries to identify records that refer to the same entities from multiple databases without compromising the privacy of the entities represented by these records. The research will focus on three aims: (1) development of new Bayesian methods for PPRL, where the error can be propagated exactly across the entire linkage process and into statistical inference, including new privacy measures to capture a tradeoff between utility and risk of any individual risk in a linked database; (2) development of new robust methods for realizing synthetic data releases post-linkage with differential privacy guarantees and its relaxations to address additional layers of privacy and support broader data sharing; and (3) exploration of "big data" methods such as variational inference to address scalability and latent cluster exchangeability issues existing within linkage and privacy, such that the new methods can scale to multiple and large databases. The new methods will be scalable and assess uncertainty throughout the entire linkage and privacy process and can be evaluated using Bayesian disclosure risk and Bayesian differential privacy. The project is supported by the Methodology, Measurement, and Statistics Program and a consortium of federal statistical agencies as part of a joint activity to support research on survey and statistical methodology.
