Constructing Confidence Intervals for BIFSG Disparity Estimates

Abstract

This paper describes a method to construct confidence intervals for estimates of disparities in outcomes between racial/ethnic groups when race and ethnicity are estimated using Bayesian Improved First Name and Surname Geocoding (BIFSG). Standard errors produced by regression packages understate the uncertainty in these estimates because they fail to incorporate noise arising from estimating race and ethnicity. This can produce false confidence when assessing the direction and magnitude of disparities. Conversely, the bootstrapping approach introduced in this paper incorporates uncertainty arising from the use of BIFSG. However, this procedure does not correct for any statistical bias associated with using BIFSG.