🤖 AI Summary
This paper addresses the challenge of unobserved individual welfare effects under policy interventions. We propose a distributionally robust evaluation framework grounded in the superquantile (i.e., conditional value-at-risk), which characterizes tail behavior of unobserved welfare impacts without requiring individual-level heterogeneity data. Unlike conventional average treatment effect approaches, our method systematically identifies highly adversely affected subpopulations and patterns of welfare inequality. Innovatively integrating superquantile theory with marginal treatment effects (MTE), the generalized Roy model, and compensating variation analysis, we establish the first framework for *distributional causal inference* of policy welfare impacts. The approach jointly optimizes efficiency and equity considerations. Empirical validation across price-change analyses, self-selected treatment assignment, and social program cost–benefit assessments demonstrates its capacity to rigorously bound welfare gains and losses—thereby enhancing the verifiability and targeting precision of redistributive policy design.
📝 Abstract
This paper proposes a framewrok for analyzing how the welfare effects of policy interventions are distributed across individuals when those effects are unobserved. Rather than focusing solely on average outcomes, the approach uses readily available information on average welfare responses to uncover meaningful patterns in how gains and losses are distributed across different populations. The framework is built around the concept of superquantile and applies to a broad class of models with unobserved individual heterogeneity. It enables policymakers to identify which groups are most adversely affected by a policy and to evaluate trade-offs between efficiency and equity. We illustrate the approach in three widely studied economic settings: price changes and compensated variation, treatment allocation with self-selection, and the cost-benefit analysis of social programs. In this latter application, we show how standard tools from the marginal treatment effect and generalized Roy model literature are useful for implementing our bounds for both the overall population and for individuals who participate in the program.