Statistical Bias In Racial and Ethnic Disparity Estimates Using Bayesian Estimation

Bayesian Improved First Name Surname Geocoding (BIFSG) is a common method for inferring race and ethnicity. The assumptions underlying BIFSG are known to fail, but the effects of these failures on downstream estimates are not well understood. In this paper, we combine US administrative tax data with data containing race and ethnicity to assess statistical bias in estimates of differences in tax outcomes by race and ethnicity. We find that BIFSG suffers from majoritarian bias, overstating the probabilities that non-White individuals are White. When using these probabilities to estimate disparities in the earned income tax credit, BIFSG estimates understate differences between White and non-White taxpayers.