Research Seminar by Bryan Wilder | Machine learning for public health decision making

Machine learning for public health decision making

Many policy settings, such as public health, require allocating limited resources under uncertainty about which individuals need those resources the most. Machine learning plays an increasing role in such allocations, most commonly to provide predictions of individual-level need. In this talk, I will discuss two recent works that take a more expansive view of the role of ML in resource allocation. First, I will discuss methods that allow a decision maker to optimally balance the immediate goal of targeting high-need individuals and the long-term goal of estimating treatment effects (i.e., does the program help people?). Our proposed method provides a policy for resource allocation which minimizes the variance with which the average treatment effect is estimated, subject to a constraint that at least some fraction of in-need individuals receive the intervention. Second, I will introduce ML as a lens to study the fairness of discretionary allocations by human decision makers, such as a clinician's decision to prescribe a resource-limited drug. Here the challenge is that human decisions are likely to be confounded (based on unobserved variables); we provide nontrivial bounds on allocative equity even in this challenging setting. In both works, we give both strong guarantees for the statistical and computational properties of our proposed methods as well as applications to real-world data from health and social service settings.