Bayesian Improved Surname Geocoding (BISG) is the most popular method for proxying race/ethnicity in voter registration files that do not contain it. This paper benchmarks BISG against a range of previously untested machine learning alternatives, using voter files with self-reported race/ethnicity from California, Florida, North Carolina, and Georgia. This analysis yields three key findings. First, when given the exact same inputs, BISG and machine learning perform similarly for estimating aggregate racial/ethnic composition. Second, machine learning outperforms BISG at individual classification of race/ethnicity. Third, the performance of all methods varies substantially across states. These results suggest that pre-trained machine learning models are preferable to BISG for individual classification. Furthermore, mixed results at the precinct level and across states underscore the need for researchers to empirically validate their chosen race/ethnicity proxy in their populations of interest.

No abstract

BACKGROUNDTimely, accurate, and precise demographic estimates at various levels of geography are crucial for planning, policymaking, and analysis. In the United States, data from the decennial census and annual American Community Survey (ACS) serve as the main sources for subnational demographic estimates. While estimates derived from these sources are widely regarded as accurate, their timeliness is limited and variability sizable for small geographic units like towns and neighborhoods. OBJECTIVEThis paper investigates the potential for using nonrepresentative consumer trace data assembled by commercial vendors to produce valid and timely estimates. We focus on data purchased from Data Axle, which contains the names and addresses of over 150 million Americans annually. METHODSWe identify the predictors of over-and undercounts of households as measured with consumer trace data and compare a range of calibration approaches to assess the extent to which systematic errors in the data can be adjusted for over time. We also demonstrate the utility of the data for predicting contemporaneous (nowcasting) tract-level household counts in the 2020 Decennial Census. RESULTSWe find that adjusted counts at the county, ZIP Code Tabulation Areas (ZCTA), and tract levels deviate from ACS survey-based estimates by an amount roughly equivalent to the ACS margins of error. Machine-learning methods perform best for calibration of countyand tract-level data. The estimates are stable over time and across regions of the country. We also find that when doing nowcasts, incorporating Data Axle estimates improved prediction bias relative to using the most recent ACS five-year estimates alone.

No abstract

No abstract