Research
Research
Job Market Paper
Mirrlees Vs The Masses: Optimal Income Redistribution Under a Popular Opinion Constraint
I combine text data from online discussions about Universal Basic Income (UBI) with an online survey experiment to quantify how the general public evaluates the desirability of different income redistribution policy reforms. Public opinion is diverse, and not well described by the standard optimal income taxation model used in economics. Combining estimated policy preferences with structural estimates from a standard income taxation model, I find that the social planner's preferred tax reform is deeply unpopular. However, modest tweaks that echo key features of UBI proposals can achieve majority support while preserving most of the welfare gains.
Peer Reviewed Publications
(Review of Economic Studies, 2023) with John List, Ian Muir, and Devin Pope
Left-digit bias (or 99-cent pricing) has been discussed extensively in economics, psychology, and marketing. Based on observational data from Lyft sessions followed by a field experiment conducted with 21 million Lyft passengers, we provide evidence of large discontinuities in demand at dollar values. Approximately half of the downward slope of the demand curve occurs discontinuously as the price of a ride drops below a dollar value (e.g. $14.00 to $13.99). Our results showcase the robustness of an important behavioral bias for large, modern companies and its persistence in a highly-competitive market.
(Econometric Reviews, 2024) with John List and Ian Muir. Winner of the 2025 Esfandiar Maasoumi Best Paper Award
This study investigates the optimal use of covariates in reducing variance when analyzing experimental data. We show that finding the variance-minimizing strategy for making use of pre-treatment observables is equivalent to estimating the conditional expectation function of the outcome given all available pre-randomization observables. This is a pure prediction problem, which recent advances in machine learning (ML) are well-suited to tackling. Through a number of empirical examples, we show how ML-based regression adjustments can feasibly be implemented in practical settings. We compare our proposed estimator to other standard variance reduction techniques in the literature. Two important advantages of our ML-based regression adjustment estimator are that (i) they improve asymptotic efficiency relative to other alternatives, and (ii) they can be implemented automatically, with relatively little tuning from the researcher, which limits the scope for data-snooping.
(Science, 2025) with Pradhi Aggarwal, Alec Brandon, Ariel Goldszmidt, Justin Holz, John List, Ian Muir, and Thomas Yu
Prior research finds that, conditional on an encounter, minority civilians are more likely to be punished by police than white civilians. An open question is whether the actual encounter is related to race. Using high-frequency location data of rideshare drivers operating on the Lyft platform in Florida, we estimate the effect of driver race on traffic stops and fines for speeding. Estimates obtained across traditional and machine learning approaches show that, relative to a white driver traveling the same speed, minorities are 24 to 33 percent more likely to be stopped for speeding and pay 23 to 34 percent more in fines. We find no evidence that these estimates can be explained by racial differences in accident and re-offense rates. Our approach provides key insights into the total effect of civilian race on outcomes of interest and highlights the methodological import of combining high-frequency data and machine learning to evaluate critical social issues.
Working Papers
The Value of Behavioral Policies
with John List, Matthias Rodemeier, and Sutanuka Roy
Most recent version: 06/16/2026
Behavioral interventions have become central to modern public policy, but their empirical promise remains contested because estimated treatment effects often appear small. We argue that a policy response is economically meaningful only relative to the response generated by alternative policies. We assemble more than 1,200 estimates from over 600 studies comparing "nudges" and traditional price interventions in the markets for cigarettes, alcohol, influenza vaccination, electricity, and residential water. Translating nudge effects into equivalent price changes, we find that behavioral interventions often correspond to enormous fiscal interventions, from an 11% tax on electricity to a 100% subsidy on influenza vaccinations. Nudges are also more cost-effective than price instruments in all markets, but cost-effectiveness does not predict the welfare ranking of policies. Using a behavioral extension of the Marginal Value of Public Funds, we show that nudges have high welfare returns \textit{at the margin}, while price instruments often generate larger total surplus at scale.
AI and the Labor Market: A Worker's Eye View
with Kiara Kim and Nathan Mester
Most recent version: 06/16/2026
We survey ~2,000 U.S. workers to study worker responses to generative AI. Workers reporting that AI complements their effort increase their work effort in response to AI; those reporting the opposite do not. Respondents also report increasing time spent on career-relevant learning; the increase is strongest for heavier AI users and those who find AI a useful learning tool. These patterns survive rich controls and an IV analysis exploiting employer AI provision. Our results combined with economic theory highlight the importance of worker-level responses as well as AI's capability as a learning tool in shaping AI’s labor-market impacts.
Salience and (Non-)Buyer's Remorse: Optimal Nonlinear Pricing with Cognitively Constrained Consumers
with Aaron Bodoh-Creed, Brent Hickman, John List, Ian Muir
Most recent version: 02/19/2026
Nonlinear pricing theory predicts that firms can extract surplus by inducing heterogeneous consumers to self-sort across price contract offers that are ex-post optimal for them. We study subscription pricing when the frictionless sorting assumption fails. Using large-scale subscription experiments conducted by Lyft, we document systematic deviations from optimal self-selection: many high-demand consumers decline subscriptions that would have saved them money, while some subscribers fail to break even. We develop a structural model of intensive-margin demand in which consumers may exhibit salience failures, forecast errors about future demand, or impulsivity. We show that subscription uptake can be recast as one- sided noncompliance in a binary-instrument framework, allowing us to leverage LATE methods to identify counterfactual outcome distributions and a novel “uptake function” linking baseline outcomes to compli- ance behavior. Combining experimental price variation with this identification strategy, we recover utility primitives, demand heterogeneity, and behavioral parameters. Salience failures and forecast errors play quantitatively important roles. Counterfactual analyses show that optimal subscription pricing generates substantial gains relative to linear pricing, but these gains are highly sensitive to consumer deviations from ex-post optimal choice. Implementing nonlinear pricing therefore requires not only optimal contract design for consumer screening, but also coordinated efforts to mitigate behavioral frictions.
with Brent Hickman, John List, Ian Muir
Most recent version: 02/20/2026
We propose methods for empirical market design in adverse-selection settings where identifi- cation based on exogenous price variation is hampered by multi-dimensional unobserved heterogeneity. We derive sharp bounds on counterfactual demand under out-of-sample prices. Bounds arise because plausible DGPs must respect the Law of Demand and observed shift(s) in aggregate demand following an exogenous price change(s). Our bounds use data available in many settings, including our application to rideshare demand. They enable viable, welfare-improving, nonlinear pricing design while achieving robustness against worst-case deviations from baseline model assumptions. Our framework provides novel insights for optimal experimental design of pricing RCTs. We also show that our CDF bounds can be re-interpreted as bounds on reduced-form treatment-effect heterogeneity.
Revealed Preference for Redistribution and the Role of Government
with Matthias Rodemeier
Most recent version: 02/09/2026
How society trades off the welfare of poor versus rich individuals is central to redistributive policy. We study this question using incentivized transfer experiments with a representative U.S. sample of close to 10,000 participants. Combining this data with a large set of estimates of the Elasticity of Taxable Income, we quantify optimal tax rates across the income distribution. Revealed preferences show strong concern for the poor across the political spectrum and imply tax rates far more progressive than existing political agendas. This generates a key puzzle: individuals vote against the tax policies that would reflect their redistributive preferences. This puzzle can be explained by an explicit aversion towards the government. Liberals' support for very progressive tax rates is dampened by misuse of public funds, while conservatives object to taxation on principled grounds tied to coercion and property rights. Our paper illustrates that disagreement over redistribution mostly reflects disagreement over institutions rather than disagreement over helping the poor.
Nonparametric Identification in Optimal Income Taxation Models
Most recent version: 02/20/2026
This paper studies identification questions that arise when taking the canonical Mirrlees (1971) model of optimal income taxation to data. I explore the connection between causal estimands derived from experimental and quasi-experimental variation in tax policy and the underlying structural primitives about agent (taxpayer) prefer- ences that these estimands help to identify. I derive three sets of results, which can be summarized as follows. First, given sufficiently rich exogenous variation in tax policy, agent utility functions are nonparametrically identified with only mild shape restric- tions on utility. Second, the additively separable utility model often used in applied work is nonparametrically identified given as few as two small exogenous shifts in tax policy. Third, counterfactual predictions about small policy reforms are robust to mis- specification insofar as the misspecified model is nonetheless able to fit the data well. In a number of Monte Carlo simulation exercises, I show that the flexibility from using a non-parametric policy can lead to better-informed policy when the econometrician has access to a large enough dataset, but the standard simplified specifications in the literature are reasonable at capturing key effects at smaller sample sizes
Physician Private Information as an Input into Pharmaceutical Innovation: A Proposal
Most recent version: 09/23/2025
How skilled are doctors at predicting what will happen to their patients? And could pharmaceutical companies make use of doctor predictive ability to lower the cost of developing new drugs? In this paper, I provide a simple statistical framework to show how physician-generated predictions about patient health can be directly leveraged to reduce the sample size necessary to conduct a clinical trial. Guided by this framework, I conduct a meta-analysis of studies that quantify the ability of physicians to predict patient outcomes. My results suggest that the information contained in physician predictions can reduce the number of participants needed to run a single clinical trial by 20%. A back-of-the-envelope calculation suggests that this could translate into as much as $7B in cost savings on clinical trial costs annually. If industry regulators are maximizing total surplus, these cost saving amount to $51B in increased economic surplus.
The Generality and Generalizability of Aggregate Ratio Parameters
Most recent version: 10/24/2023
Many parameters in empirical economics are identified by taking the ratio of two average causal effects. A famous example of this strategy, which I refer to as the "aggregate ratio parameter", is the Wald estimator from the instrumental variables (IV) model. I show that this paradigm is more general than the IV example and unifies quantities ranging from the equivalent price metric in behavioral economics to the diversion ratio in merger analysis. I generalize well-known results for IV and show that the aggregate ratio parameter identifies only the average of individual-level ratio parameters within a suitably defined "complier" population. Thus, the concerns raised in the context of IV that estimands may have limited policy relevance/external validity apply more broadly to empirical objects other than treatment effects. I show how to use covariates to empirically probe the importance of these external validity concerns and how to correct estimates to bring the identified complier population closer to the policy-relevant population along observable dimensions. I use my framework to re-analyze data from a diverse range of empirical examples.
Why Do Robust Standard Errors Hardly Change Inference?
Most recent version: 09/21/2022
A folk wisdom in applied econometrics is that adjusting OLS standard errors for heteroskedasticity rarely leads to large changes whereas adjusting for other statistical considerations such as clustering is likely to be more consequential. In this note, I derive a tractable formula for the ratio between robust and conventional standard errors for a given covariate of interest in a multi-variate linear regression. This formula helps clarify when and in which direction we should expect robust standard errors to differ from conventional standard errors. I argue that a motivating example of heteroskedasticity in econometrics pedogogy is ambiguously signed because factors pushing in opposite directions will tend to cancel each other out.
Works in Progress
Identification of Random Coefficients Models without Linearity
Behavioral Nonlinear Pricing
with Matthias Rodemeier and Nicola Gennaioli