An Epistemic Approach to Statistical Fairness

Predictive algorithms are used across a wide variety of settings in order to inform consequential decisionmaking about people’s lives, ranging from pretrial detention decisions in the criminal justice system to surgical intervention decisions in medical contexts. With the increasing deployment of these algorithms comes the risk that they could make unfair predictions along race or gender lines. A popular method to audit these algorithms for such biases employs measures called “statistical fairness criteria.” Strikingly, many of these criteria are impossible to satisfy at once—a result sometimes called the “impossibility of fairness,” for it illuminates an apparently tragic choice between honoring different, intuitive conceptions of fairness. This thesis argues that this interpretation of the impossibility result is, in one sense, too pessimistic, and in another, too optimistic, with both arguments driven by connecting statistical fairness to accuracy-based epistemology. On the one hand, through analyzing its connection to accuracy, I argue one statistical fairness criterion is ill-formulated; reforming it resolves the tension between two of the most prominent statistical fairness criteria. On the other hand, though, the same connection demonstrates how statistical properties give a limited conception of fairness: they only capture salient fairness concerns insofar as accuracy matters to fairness, which in many cases, may not be very much at all.