About
I am an economics PhD candidate at Cornell University. My research is in econometrics, focusing on partial identification and statistical decision theory. I am advised by Jörg Stoye, Francesca Molinari, and José Luis Montiel Olea.
I am on the 2026–2027 economics job market.
Research
Paying to Sharpen Identification: Optimal Probing of Coarse Data
Job market paper
Abstract
Observations that are rounded, mismeasured, missing, or otherwise coarsened can often be probed through follow-up questions, remeasurement, or linkage across data sources. This paper develops a framework for selecting which observations to probe under a budget constraint. The value of probing is measured by the reduction in identified-set diameter, and the cost of probing is obtained by integrating observation-level costs. For scalar expectations, and for univariate distributions under Wasserstein-1 distance, identified-set diameter decomposes into additive observation-level contributions. This decomposition yields a simple optimal rule for data probing: probe observations in decreasing order of expected diameter reduction per unit cost, until the budget is exhausted. In a counterfactual survey-design application based on Manski and Straub’s (2000) analysis of worker expectations, targeting follow-up questions to the coarsest 25% of initially ambiguous responses captures 70% of the identified-set diameter reduction attainable by questioning all respondents.
Statistical Treatment Choice with Untested Doses
Working paper
Abstract
I extend Manski’s (2025) limited-trial dose-choice model to a finite-sample statistical decision setting. A decision-maker chooses from a finite ordered menu of doses to balance efficacy against adverse effects, but trial data are observed only for a subset of these doses. Outcomes at untested doses are constrained by the assumption that efficacy and adverse effects are weakly increasing in dose, and are hence partially identified. I show that a plug-in rule based on any feasible uniformly mean-consistent estimator is asymptotically minimax-regret optimal. A constrained maximum likelihood estimator satisfies this uniform mean consistency, and the finite-sample suboptimality gap is O(1/√nmin), where nmin is the smallest per-arm sample size in the trial. This plug-in problem can be formulated as a linear programme.
Asymptotic Minimax Optimality of Plug-In Rules
Working paper · with José Luis Montiel Olea and Jörg Stoye
Abstract
We provide finite-sample and asymptotic minimax optimality guarantees for a wide class of plug-in rules. Minimax optimal decision rules are often computationally infeasible in practice. Plug-in, or as-if, rules are an intuitively appealing alternative class of rules that solve the population problem as if an estimate of the payoff-relevant parameter were the true value. We show that when the loss function L(a,θ) is uniformly bounded and the family of mappings θ ↦ L(a,θ) is uniformly equicontinuous, plug-in rules using uniformly consistent estimators are asymptotically minimax optimal. We also derive finite-sample bounds on plug-in maximum risk from bounds on estimator performance. Moreover, we show that under relatively general conditions, these results extend to partially identified settings, in which the plug-in is obtained by first profiling the loss with respect to the unidentified parameter.