Identifying Inequity in Treatment Allocation

Yewon Byun, Dylan Sam, Zachary Chase Lipton, Bryan Wilder

Keywords: ML for Healthcare, Fairness, Resource Allocation

TL;DR: A principled approach to identifying inequity in treatment allocation

Abstract:

Disparities in resource allocation, efficacy of care, and patient outcomes along demographic lines have been documented throughout the healthcare system. In order to reduce such health disparities, it is crucial to quantify uncertainty and biases in the medical decision-making process. In this work, we propose a novel setup to audit inequity in treatment allocation.
We develop multiple bounds on the treatment allocation rate, under different strengths of assumptions, which leverage risk estimates via standard classification models. We demonstrate the effectiveness of our approach in assessing racial and ethnic inequity of COVID-19 outpatient Paxlovid allocation. We provably show that for all groups, patients who would die without treatment receive Paxlovid at most 53\% of the time, highlighting substantial under-allocation of resources. Furthermore, we illuminate discrepancies between racial subgroups, showing that Black patients who would die without treatment receive Paxlovid at most 32\% and 65\% lower than White and Asian patients, respectively.

Submission Number: 105