Welfare at Risk: Distributionally Robust Bounds for Policy Evaluation

Introduction to the show: ident: Economics. The day in economics: theory, mechanisms and the evidence beneath them.

Ruth: We're looking at the paper: "Welfare at Risk".

Warren: Here's the gist of the paper. This paper develops a distributional welfare framework that accounts for the heterogeneity in welfare generated by policy interventions by deriving quantile-based bounds that allow analysts to learn about the distribution of individual welfare changes even when this distribution is unobserved.

Ruth: Distributional Welfare Primitives The approach builds on the perturbed utility model (PUM) of Allen and Rehbeck

3: and McFadden and Fosgerau

3: , which is a latent choice model that accommodates general additive unobserved heterogeneity.

Warren: The PUM encompasses a broad class of latent utility models with additively separable unobserved heterogeneity, including additive random utility models (ARUMs), bundling models, and matching models.

Ruth: The conditional average welfare W(x) is defined as E max y∈B U(y; x, ε) X = x.

Paper summary: Iris: The paper builds this framework on the perturbed utility model, or PUM, of Allen and Rehbeck and McFadden and Fosgerau, which is a latent choice model that handles general additive unobserved heterogeneity.

Ruth: That PUM structure encompasses a broad class of models like additive random utility models or ARUMs, which are used for their tractability in counterfactual analysis.

Warren: The paper focuses on settings where the analyst only observes aggregate market data or conditional average welfare changes, like average treatment effects.

Iris: They mention that because an agent is only observed in one of two potential-outcome states—treated or untreated—the individual welfare change and its distribution can't be recovered <ref:2512.20918#pg2>.

Ruth: That’s the core problem they are addressing, so their approach aims to learn about that distribution even when it’s unobserved.

Warren: They introduce quantile-based bounds, specifically superquantiles, to get this information about individual welfare changes without needing the full distribution.

Iris: The key result here is Theorem one which shows that the superquantile of individual welfare changes—which is unobservable to the analyst—is bounded above by the superquantile of conditional average welfare changes that they can identify <ref:2512.20918#pg1>.

Ruth: And they also bound the share of people whose welfare loss from a policy exceeds any given threshold, which is described as a distribution-free counterpart to average cost–benefit analysis >

Warren: They also extend both these bounds to allow for ambiguity in the analyst’s knowledge of the distribution of observable characteristics, which means they can handle situations where an estimate based on one population might be wrong for another <ref:2512.20918#pg1>.

Iris: The robust version, which they call S R beta(Z), is defined as the infimum over a phi-divergence neighborhood where the center is given by the nominal covariate distribution >

Ruth: And they show that for any beta between zero and one, and for any phi function and any rho greater than zero, this upper bound on the superquantile of individual welfare changes still holds even when the marginal distribution of X is itself ambiguous, with the ambiguity being controlled by rho <ref:2512.20918#pg3>.

Warren: It seems like they're building a tool to get a valid guarantee for policy effects even when you don't know exactly what the individual welfare changes look like Ruth So, this whole setup is designed to help analysts learn about the distribution of individual welfare changes even when that distribution isn't observable or identifiable from the data >

Iris: And they apply this to generalized Roy models with subjective participation costs, deriving bounds for treatment on the treated that are new in that literature <ref:2512.20918#pg3>.

Ruth: Corollary four provides a lower bound on the buffered probability of welfare loss among the treated, which can be computed just from the identified function W TT(X, Z) alone >

Warren: It's interesting that they also address instrument sensitivity by taking an infimum over an ambiguity set instead of over instrument choices, meaning the robust welfare risk is a valid bound no matter which instrument you use to estimate transition-specific propensity scores >

Ruth: The price for this robustness is quantified by Inf(rho*), which measures the distance above the trivial floor relative to the non-robust bound, and that quantity can be computed directly from estimated propensity scores >

Conclusion: Ruth: So wrapping up this paper, "Welfare at Risk: Distributionally Robust Bounds for Policy Evaluation," it’s about creating a distributional framework for analyzing welfare heterogeneity across individuals using superquantiles and observable-group average welfare effects >

Warren: The authors show that you can construct informative bounds on individual welfare changes without relying on strong structural assumptions or detailed micro-level data >

Iris: The core results established in Section four extend naturally to this setting, even though the relevant welfare concepts and identification strategies differ <ref:2512.20918#pg3>.

Ruth: This framework extends naturally to richer multi-sector environments studied by HUV, providing distributional welfare guarantees that standard instrumental variable methods alone cannot give >

Warren: It’s a way to get a picture of how policy effects are spread across the population, going beyond just the average effect >

Iris: The paper gives us tools that help analysts move from just knowing the mean effect to understanding where gains and losses actually fall for different groups >