Modeling Risk and Achieving Algorithmic Fairness Using Potential Outcomes

Abstract

Predictive models and algorithms are increasingly used to support human decision makers, raising concerns about how to ensure that these algorithms are fair. Additionally, these tools are generally designed to predict observable outcomes, but this is problematic when the treatment or exposure is confounded with the outcome. I argue that in most cases, what is actually of interest are potential outcomes. I contrast modeling approaches built around observable vs. potential outcomes, and I recharacterize error rate-based algorithmic fairness metrics in terms of potential outcomes. I also aim to formally model the consequences of using confounded observable predictions to drive interventions.