The Geography of Racially Polarized Voting: Calibrating Surveys at the District Level

Abstract

Voting in the United States has long been known to divide sharply along racial lines, and the degree of racially polarized voting evidently varies across regions, and even within a state. Researchers have further studied variation in racially polarized voting using aggregate data techniques, but these methods assume that variation in individual preferences is not related to geography. This paper presents estimates based on individual level data of the extent and variation in racially polarized voting across US Congressional Districts. Leveraging large, geocoded sample surveys, we develop an improved method for measuring racial voting patterns at the Congressional District-level. The method overcomes challenges in previous attempts of survey modeling by allowing survey data to inform the synthetic population model. This method has sufficient power to provide precise estimates of racial polarization even when survey data are sparse. We find that variation across districts but within states explains roughly 20 percent of the total variation; states explain a further 20 percent of the total variation, and 55 percent of the variation is simply national differences between races. The Deep South still has the highest racial polarization between White and Black voters, but some Midwestern congressional districts exhibit comparably high polarization. The polarization between White and Hispanic voters is far more variable than between Black and White voters.

Recalibration of Predicted Probabilities Using the “Logit Shift”: Why Does It Work, and When Can It Be Expected to Work Well?

“…Furthermore, recent developments can help correct some of the limitations we have highlighted. For instance, minimizing differences with respect to multiple aggregated targets can help resolve issues raised by heterogeneity in prediction errors among subgroups, and provide more information about the shape of the distribution of true probabilities (e.g., Kuriwaki et al 2022). A fruitful avenue for future research would explore whether these attempts can also be justified as approximations to a posterior update that conditions on multiple totals, highlighting the connections between the logit shift and the problem of ecological inference (e.g., King, Tanner, and Rosen 2004;Rosenman 2019).…”

“…A common heuristic solution to the recalibration problem is the so-called "logit shift" (e.g., Ghitza and Gelman 2013;Ghitza and Gelman 2020;Hanretty, Lauderdale, and Vivyan 2016;Kuriwaki et al 2022). 1 To motivate this approach, consider a simple scenario in which the p i are generated from a logistic regression model.…”

“…Executing the update at higher levels of aggregation, however, can imply more heterogeneity in prediction errors-heterogeneity that may not be correctable using the rank-preserving logit shift with a single target total (Kuriwaki et al 2022). To see why, consider an example in which there are two groups of voters: Black voters and White voters.…”

Individual and Differential Harm in Redistricting

“…The mean support µ (s) receives an informative prior based on the racially polarized voting estimates of Kuriwaki et al (2021), with µ (s) ∼ N logit((0.18, 0.90, 0.35)), (0.04 2 , 0.04 2 , 0.2 2 ) . These estimates were derived through multilevel regression and poststratification from survey data, calibrated to actual election results.…”

Embedded multilevel regression and poststratification: Model‐based inference with incomplete auxiliary information

“…Practical applications solicit population information from either census records or large studies with minimal errors. For example, across MRP application studies, Si et al 3 obtain the joint population control distribution from the American Community Survey (ACS); Wang et al 9 use aggregated exit polls; Zhang et al 10 use census records; Yougov 11 uses the Current Population Survey; Kuriwaki et al 12 use the ACS and election official turnout statistics; and Ghitza and Gelman 13 turn to large‐scale voter registration databases to directly obtain such information for the poststratification adjustment.…”