Algorithmic Fairness, Decision Thresholds, and the Separateness of Persons

Can an algorithmic procedure for allocating a medical treatment such as chemotherapy be fair, if it entails that the therapy should be administered to candidates, who are better off, in expectation, if they do not to receive it? In this article I argue (i) that the answer is “no” and (ii) that popular statistical fairness criteria must answer “yes.” I then argue that fairness requires that the procedure allocates the therapy for the right reason, and that the right kind of reason must respect the separateness of persons. I then anchor my proposed individual-level approach to algorithmic fairness in John Broome's theory of fairness and conclude that algorithmic fairness requires individual decision thresholds.