(Machine) Learning what Policymakers Value ∗

This paper develops a method to uncover the values consistent with observed allocation decisions. We use machine learning estimators for heterogeneous treatment effects to identify who benefits from an allocation. We then decompose the objective underlying the allocation into: differential (i) treatment effects, (ii) welfare weights between entities; and (iii) impact weights across outcomes. We apply this approach to Mexico’s PROGRESA anti-poverty program and estimate the preferences consistent with its design. We find evidence of heterogeneous impacts by income and age; accounting for this heterogeneity, allocations imply higher welfare weights on the indigenous, poor, and for families with more children. The implied value of each missed school day and child sick day is estimated imprecisely but does not rule out conventional valuations or preferences reported by Mexican residents. Alternate eligibility criteria could have improved either average consumption, health or schooling outcomes. ∗We thank Luk Yean and Jolie Wei for excellent research assistance. Thank you to Joseph Cummins, Brian Dillon, John Friedman, Yang Xie, and seminar audiences for helpful conversations. †dan@bjorkegren.com ‡jblumenstock@berkeley.edu §samsun knight@brown.edu