Moving forward, we hope our work helps to inform litigation and research in the domain of disparate impact.

If a disparity is justified by business necessity, DI can still be legally established by the existence of an alternative policy with a smaller disparity. Our third proposed approach compares status quo disparities to those estimated under a statistically-optimized policy. 🧵

In settings where risk can be estimated for individual decisions, we propose measuring DI via risk-adjusted regression (Jung et al., 2023). Risk-adjusted regression assesses whether disparities persist across groups after adjusting for individual-level risk. 🧵

We illustrate the second and third proposed approaches using public data from approximately 1.5 million stops recorded by police agencies across California in 2022 pursuant to its Racial Identity and Profiling Act (RIPA). 🧵

The Arnold et al. estimand may suggest evidence of DI even though the disparities can fully be explained by differences in underlying risk factors. In these instances, courts generally don’t hold defendants liable for DI, putting the estimand in direct conflict with the law. 🧵

Arnold et al. (2022) propose measuring disparate impact (DI) as a weighted difference in error rates across groups. However, this estimand is problematic. 🧵

SCOTUS ruled that job performance was not meaningfully related to having a high school diploma, and thus the diploma requirement imposed an illegal disparate impact on Black employees. 🧵

Consider the SCOTUS case Griggs v. Duke Power Co. (1971). Duke Power required employees to have a high school diploma in order to be promoted. For many reasons, Black employees were less likely than white employees to have a diploma, and were thus less likely to be promoted. 🧵

Disparate treatment law prohibits decisions that intentionally condition on protected group status, such as race. Disparate impact law, in contrast, renders illegal policies that do not explicitly account for group membership, yet nonetheless result in unjustified disparities. 🧵

Disparate impact (DI) has been the subject of extensive litigation. However, there is not an agreed-upon method to quantify DI empirically. In new work w/ @5harad and @JulianNyarko, we explore several proposed statistical frameworks for measuring DI. https://t.co/sQi3hRUgKC 🧵

“It’s not affirmative action keeping Asian American students from these selective colleges — it’s things like legacy admissions and geography, sports,” says @ShorensteinCtr researcher @5harad. https://t.co/MXGM9tHW3a @joshdgrossman @soy_beana @linzcpage

Report: Admissions Gap Between White and Asian Applicants https://t.co/4Bvc37SLtK

Op-ed in @GlobeOpinion: Harvard isn’t Hogwarts. Merit, not parentage, should be key to admission.

New @nberpubs paper from Annenberg faculty member @linzcpage along with Joshua Grossman, @soy_beana, and @5harad examining the disparate impacts of college admissions policies on Asian American applicants https://t.co/tlWpEFbs2Q