BlueDown algorithm cuts 2020 US Census noise errors by up to 61%

Researchers developed BlueDown, a postprocessing method designed to denoise demographic tables generated under differential privacy protocols, according to a paper published in Science doi.orgSciencedoi.orgOpen source ↗. The 2020 United States Census introduced calibrated noise into summary demographic measurements to protect individual confidentiality before reconciling the counts into official tables. The new algorithm combines techniques from linear regression and constrained optimization while exploiting measurement structure to preserve computational efficiency doi.orgSciencedoi.orgOpen source ↗.

When evaluated on 2020 Census data, BlueDown reduced errors in county-level groupings by race by 21 to 36% and lowered errors across all county-level groupings by 9 to 58% doi.orgSciencedoi.orgOpen source ↗. At the tract level, the method decreased errors in housing type groupings by 38 to 61% doi.orgSciencedoi.orgOpen source ↗. The researchers reported that these accuracy gains required no compromises on individual confidentiality or computational efficiency, supporting demographic analyses that guide more than $1.5 trillion in annual federal funding allocations doi.orgSciencedoi.orgOpen source ↗.