Online Program

Formally Private Microdata at Scale: Reducing the Magnitude of Upward Bias (308096)

*Philip Leclerc, United States Census Bureau 

Keywords: statistical disclosure control, data privacy

For the 2020 Decennial Census, formally private noise-infusion mechanisms have been combined with nonnegative optimization to generate formally private microdata. Generating microdata at scale requires that all estimated counts be consistent and nonnegative. As a result, post-processing can increase upward bias relative to simpler unconstrained estimators. In this talk, we consider this problem and discuss several approaches for addressing it.