Balancing Learning and Targeting in Predictive Allocation

This letter provides an overview of our recent work on "Learning Treatment Effects While Treating Those in Need" (published at the 2025 ACM Conference on Economics and Computation) as well as a more general perspective on design goals for algorithmic systems that are used to allocate limited resources in policy settings. Our motivation is the kind of algorithms that are used widely at present to prioritize candidates for various kinds of social interventions: public housing assistance, drop-out prevention programs in education, unconditional cash transfers in development, or a variety of other social services. By far the most common way of constructing such systems is the lens of predictive allocation : the algorithm designer identifies an outcome that the program seeks to alter (long-term home-lessness, dropping out of school, etc) and constructs a predictive model for that outcome [Vaithianathan and Kithulgoda 2020; Aiken et al. 2022; Pan et al. 2017; Toros and Flaming 2017]. Candidates are ranked by predictions of risk so that, e.g., limited spots in a housing program might be offered to those at greatest predicted risk of long-term homelessness.