Generalizability with ignorance in mind: learning what we do (not) know for archetypes discovery

When studying policy interventions, researchers often pursue two goals: i) identifying for whom the program has the largest effects (heterogeneity) and ii) determining whether those patterns of treatment effects have predictive power across environments (generalizability). We develop a framework to learn when and how to partition observations into groups of individual and environmental characterstics within which treatment effects are predictively stable, and when instead extrapolation is unwarranted and further evidence is needed. Our procedure determines in which contexts effects are generalizable and when, instead, researchers should admit ignorance and collect more data. We provide a decision-theoretic foundation, derive finite-sample regret guarantees, and establish asymptotic inference results. We illustrate the benefits of our approach by reanalyzing a multifaceted anti-poverty program across six countries.