Les expériences randomisées sont un outil fondamental pour identifier quelles politiques économiques, interventions éducatives ou traitements médicaux sont les plus efficaces. L’efficacité des traitements peut différer selon les individus, il est donc également important d’apprendre des expériences randomisées comment cibler les traitements en fonction des caractéristiques individuelles. Pour les politiques publiques, nous devons souvent imposer des contraintes à la politique, comme donner la priorité aux personnes à faible revenu. De nouvelles méthodes économétriques existent pour identifier les meilleures politiques qui satisfont à ces contraintes, mais elles demeurent difficiles à utiliser pour les chercheurs.
L'un des objectifs de ce projet est de rendre ces outils plus accessibles à la communauté des chercheurs, en fournissant des codes informatiques et des exemples d'applications. Un autre objectif est de développer de nouveaux résultats théoriques qui faciliteront la résolution du problème du choix de politiques sous contraintes.
Il est également important de comprendre combien d'observations une expérience doit inclure. Ce projet développera de nouvelles méthodes pour calculer les tailles d'échantillon suffisantes pour les expériences randomisées.
Data from randomized experiments contains valuable information that identifies how the effect of treatment varies with individuals’ observed characteristics. Empirical Welfare Maximization (EWM) methods use that information to select a treatment targeting policy (which individuals should/shouldn’t receive the treatment) from a pre-specified constrained policy class by maximizing the estimated welfare gain over that targeting policy class.Constraints on allowed targeting policies distinguishes the EWM approach from a large Statistical/Machine Learning literature on individualized treatment rules. Constraints are important in public policy: e.g., a policy maker may not want to adopt a policy that provides housing vouchers to some higher-income individuals but excludes otherwise similar lower-income individuals. There is no guarantee that a policy derived by ML methods will respect this constraint.The EWM methodology is currently attracting significant theoretical contributions but remains prohibitively difficult for applied researchers to implement. The main hurdle is that maximizing the estimated welfare gain over a seemingly simple constrained policy class like linear treatment rules (which construct a linear combination of covariates and treat individuals if that linear combination is greater or equal to zero) or fixed-depth decision trees is often a very computationally difficult problem.EWM methods need to become more accessible to practitioners. New theoretical advances are needed to overcome the computational difficulty of the problem and to solve it for new classes of policies. Experimenters also need guidance on selecting sample size in experiments used to learn treatment policies, which requires both theoretical advances and software for sample size calculations. While finite-sample bounds on regret of EWM treatment rules are available (Kitagawa and Tetenov, 2018a), they are unduly conservative to serve as a basis for experimental design.This projects aims to:A: Develop and make publicly available general-use code for solving EWM problems for a variety of policy classes, compile a reference collection of publicly available RCT datasets for optimizing the code and demonstrating examples of applications.B: Make new theoretical advances relevant to EWM methods on properties of using surrogate loss functions in problems with constrained policy classes.C: Develop new methods for computing sufficient sample sizes for and extend these methods to RCTs that are designed for treatment targeting using EWM.The results of the project will make it easier for applied researchers to use data from randomized experiments to find effective, simple, implementable, treatment targeting policies.
Last updated:07.06.2022
