A Reputation for Honesty, in Journal of Economic Theory (2022)
Joint with Drew Fudenberg and Harry Pei
A long-lived player can build a reputation for honestly announcing their intended actions against short-run players. Limited records allow the long-run player to replicate their commitment payoff.
Abstract: We analyze situations where players build reputations for honesty rather than for playing particular actions. A patient player faces a sequence of short-run opponents. Before players act, the patient player announces their intended action after observing both a private payoff shock and a signal of what actions will be feasible that period. The patient player is either an honest type who keeps their word whenever their announced action is feasible, or an opportunistic type who freely chooses announcements and feasible actions. Short-run players only observe the current-period announcement and whether the patient player has kept their word in the past. We provide sufficient conditions under which the patient player can secure their optimal commitment payoff by building a reputation for honesty. Our proof introduces a novel technique based on concentration inequalities.
Linear Estimation of Structural and Causal Effects for Nonseparable Panel Data, conditionally accepted at Econometrica (2026)
Joint with Victor Chernozhukov, Ben Deaner, Jerry Hausman and Whitney Newey
Individuals are heterogeneous in an underlying set of parameters. Using panel data, it is possible to estimate average effects across the population by applying a bias corrected ridge estimator on an individual-by-individual basis.
This paper develops linear estimators for structural and causal parameters of nonseparable models using panel data. These models incorporate unobserved, time-varying, individual heterogeneity, which may be correlated with the regressors. Estimation is based on an approximation of a conditional average potential outcome by a linear sieve specification with individual-specific parameters. Effects of interest are estimated by a bias corrected average of individual ridge regressions. We demonstrate how this approach can be applied to estimate causal effects, counterfactual consumer welfare, and averages of individual taxable income elasticities. We show that the proposed estimator has an empirical Bayes interpretation and possesses a number of other useful properties. We formulate Large-T asymptotics that can accommodate discrete regressors and which bypass partial identification in this case. We employ the methods to estimate average equivalent variation and deadweight loss for potential price increases using data on grocery purchases.
Inference from Selectively Disclosed Data (2026)
A cherrypicker with a large dataset optimally drops data to imitate distributions generated by favorable states. Targeted and transparent experimental design can help.
Abstract: A receiver observes a large, multidimensional dataset that is voluntarily disclosed by a strategic sender who cherrypicked it, without knowing how much data were originally obtained. We use a continuous model of big data to characterize equilibrium inferences and give an exact decision-robust welfare order on experiments that describes when, in the presence of cherrypicking, the receiver prefers one experiment over another. A preferable experiment gives stark likelihood-ratio differences in the distribution of data under sender-preferred vs. sender-dispreferred states, so that the sender faces a greater cost when dropping data to mimic more favorable distributions. These predictions formally capture the large-data limit.
Feedback in Selection Tournaments (2026, draft soon)
Joint with Nicole Immorlica, Brendan Lucier, and Markus Mobius
Middle managers have local information about those they monitor. They can use it to inform the executive and provide feedback to agents, but it may not be optimal to do both.
Abstract: Managers in organizations have access to local, divisional information that allows for close comparison of related projects. We consider how this information can be incorporated in the design of meritocratic contests for selection and funding. Using a multi-agent cheap talk model, we ask what managers can truthfully report to a decision-maker or give as feedback to agents. A decision-maker who wants to efficiently allocate a prize while discouraging unnecessary effort should ask for reports of ordinal rankings to be used as nominations, and wishes to, but cannot, block communication of rankings to agents. When managers have richer, cardinal information but agents receive feedback about what was reported, we show that incentive constraints prevent the decision-maker who elicits that information from doing any better than in the ordinal case.
Recommendations with Costly Verification (2026)
Joint with Yingju Ma
A recommender faces consumers who can verify recommendations, but differ in their attention costs or in the salience of their choice of adoption. We characterize the conditions under which a technology that increases transparency by lowering the cost of verifying a product's quality will increase or decrease false recommendations. Without commitment, reducing verification costs never increases false recommendations, but when the recommender can commit to a policy, transparency can worsen both the quality of recommendations and consumer welfare. The necessary and sufficient condition is that the log-elasticity of the distribution of consumer cost types is decreasing. It can hold when the distribution is sufficiently fat-tailed, or when it mixes sophisticated and naive consumers. Our results can be applied to platform-recommended listings, security ratings, and audits.
Model (Non)-Disclosure in Supervisory Stress Tests (2023)
Joint with Marc de la Barrera and Bumsoo Kim
A central bank can choose whether to disclose its stress test model, trading off between providing guidance through transparency and incentivizing the use of private information through nondisclosure.
Abstract: We study the Federal Reserve’s problem of disclosing the models it uses in supervisory stress tests of large banks. Banks argue that nondisclosure leads to inefficiencies stemming from uncertainty, but regulators are concerned that full disclosure can lead to banks gaming the system. We formalize the intuition behind this trade-off in a stylized model where both the regulator and banks have imperfect, private “models” about a risky asset, and the regulator uses its own model to ‘stress test’ the investment. We show that if the regulator uses its model to test the banks’ investment, full disclosure is suboptimal, and the regulator may benefit from hiding the model when the bank’s model is more precise than the regulator’s own model. The key idea is that hiding the regulator’s model forces the bank to guess it using the bank’s own models, effectively eliciting the bank’s private information. We also show that if the regulator can fine-tune disclosure policies, the regulator can approximately enforce the first-best action of banks, as if the regulator fully knew all the private information held by banks. The intuition is closely related to the Cremer and McLean (1988) information rent extraction result.
When to sell a house: Learning about demand in asset markets (2026, with Cyrus Aghamolla)
Incentives and Learning in Platforms with Recommendations (2026, with Nick Wu and Andrew Koh)