|
Abstract:
|
A growing literature on differential privacy (DP) has developed a suite of noise infusion techniques that provide strong guarantees and allow for a controlled tradeoff between privacy protection and utility of data releases. Nonetheless, despite the large literature there is less information on practical implementation details and how these techniques affect the utility of data products. This presentation discusses potential implications of using DP techniques to protect privacy when constructing average cost of child care measures from the 2012 National Survey of Early Care and Education. We calculate state-level measures of parental out of pocket costs for regular child care for all U.S. states., and then apply a variety of DP techniques to the results of our analysis in order to understand the effects of DP on the accuracy of our statistics. For each algorithm that we investigate, we document differences in results across different values of the privacy loss parameter (epsilon) and across cells of varying sizes. Throughout our analyses, we pay particular attention to various tuning parameters and practical decisions that can be overlooked in high-level discussions of DP.
|
