An Absolute Test of Racial Prejudice

Abstract

Disparities along racial and ethnic lines persist across domains. Distinguishing among the possible sources of such disparities matters. This article introduces an absolute test for identifying prejudice in the presence of statistical discrimination. In the context of police officers deciding whether to conduct vehicle searches, the key intuition of the test is that each officer’s search decisions and search outcomes generate a point on a concave “return possibility frontier,” (RPF) whose slope equals the officer’s search cost, or personal standard of evidence for conducting a search. Variation along a RPF provides information about search costs, and a discrepancy in these costs across drivers of different races constitutes prejudice. The model and test generalize and unify the existing literature, and the test can be partially extended to the setting where officers vary in the quality of their information, or discernment. Higher discernment generates an expansion of the frontier, and a version of the test remains valid for more discerning officers. Empirically, the test finds suggestive evidence of prejudice against Hispanic drivers and of varying discernment among officers of different races and ethnicities. These results are robust to (and not well explained by) officer experience. (JEL C26, K42, J15)

On the Use of Outcome Tests for Detecting Bias in Decision Making

“…In pre-trial release, for example, the analyst considers comparing the average rate of pre-trial misconduct among white defendants released by judge z versus the average rate of misconduct among black defendants released by the same judge. We note that the more general case where ∆ ≡ Y 1 − Y 0 does not affect the derivations 25 See also Alesina and La Ferrara (2014) and Marx (2022).…”

“…Another prominent example of this approach isAlesina and La Ferrara (2014), who study racial bias in death penalty sentencing by exploiting variation in reversal rates by higher courts across combinations of defendant race and victim race Marx (2022). revisitsAnwar and Fang (2006)'s dataset with an absolute, rather than relative, test of bias among police officers of different races.…”

Measuring Racial Discrimination in Bail Decisions

“…arises via non-race characteristics and are infeasible in high-stakes and face-to-face settings such as bail decisions. Outcome-based tests can detect one potential driver of disparate impact-racial bias at the margin of release decisions (e.g., Arnold, Dobbie and Yang, 2018;Marx, Forthcoming)-but cannot detect accurate statistical discrimination or measure the overall extent of disparate impact.…”

Racial Disparities in Motor Vehicle Searches Cannot Be Justified by Efficiency

“…More recently, Arnold et al (2018) and Marx (2020) use variation in behavior among decisionmakers to address inframarginality. Arnold et al (2018) use the quasi-random assignment of defendants to bail judges with varying release tendencies to test whether the rates of pretrial misconduct differ for marginal black and white defendants.…”