In this project, we developed methodology and tools for providing researchers, students, and other members of the public access to confidential social science data.  The developments focus on three main concepts.  The first concept is called synthetic data.  A synthetic data file is comprised of entirely simulated data, generated so that the statistical properties of the simulated data mimic those of the confidential data.  This data file can have low risks of unintended disclosures, since the released records do not correspond to actual individuals in the data file.  We developed new methods for generating synthetic data for longitudinal data files with many variables.  We used these methods to develop a synthetic dataset comprising work histories of employees in the federal government.  The second concept is called a verification server.  This is a computer server that holds the synthetic and confidential data files.  Users of the synthetic data can query the server for feedback on the quality of results obtained from the synthetic data; for example, is the value of a regression coefficient estimated from the synthetic data similar to the estimate from the confidential data?  We developed verification measures that satisfy differential privacy, so that releasing verification measures to users does not leak too much information about the confidential data.  We developed software for implementing these measures, as well as a user interface, that could serve as the front end of verification servers developed in other contexts.  The third concept is called secure remote access.   Vetted users can log in to a server at a host institution to access the confidential data.  We developed software that allows the host institution to provide multi-factor authenticated access to users outside the host institution, without having to create accounts for those users.  We demonstrated how to integrate synthetic data, verification, and remote access in a single system.  We use the synthetic data on federal employees to estimate differentials in pay by gender and race in the federal government; we find differences by race and gender.  We verify these results using the differentially private measures.  Finally, we run the analyses on the confidential data, and find that the disparities exist in those data as well.  The integrated system performs as intended.  The synthetic data allow users to get reasonable inferences about pay disparities.  The verification server confirms findings from the analysis of the synthetic data that also exist in the confidential data, and it reveals findings from the synthetic data that do not hold up in the confidential data.   All software from the project is available in a free, public repository on GitHub.  The project trained one post-doctoral associate, two PhD students who finished their degrees, two master's students who finished their degrees, and two undergraduate students who finished their degrees.

Last Modified: 02/21/2019
Modified by: Jerome P Reiter