Typically in mechanism design it is assumed that the designer has a prior over the agents’ preferences. This is often unrealistic, especially in multi-dimensional settings. For example, in a combinatorial auction, the prior for just a single bidder would have a number of support points that is doubly exponential in the number of items. To address this, Sandholm and Likhodedov introduced the idea of mechanism design with just samples from the prior. This raises the question, how many samples are sufficient to ensure that a mechanism that is empirically optimal on the samples is nearly optimal in expectation on the full, unknown prior? I will discuss structure shared by a myriad of pricing, auction, and lottery mechanisms that allows us to prove strong sample complexity bounds: for any set of agents’ preferences, the objective is a piecewise linear function of the mechanism's parameters. We prove new bounds for mechanism classes not previously studied in the sample-based mechanism design literature and match or improve over the best known guarantees for many classes. The functions we study are significantly different from well-understood functions in machine learning, so our analysis requires a sharp understanding of the interplay between mechanism parameters and agents’ preferences. We strengthen our main results with data-dependent bounds when the distribution over agents’ preferences is well-behaved. We investigate a fundamental tradeoff in sample-based mechanism design: complex mechanisms often have higher objective value than simple mechanisms, but more samples are required to ensure that empirical and expected objective are close. We provide techniques for optimizing this tradeoff. I will then present an even more general recent sample complexity theory, which we have used also for voting and redistribution mechanism design. I will then discuss how similar techniques can be used to estimate the approximate incentive compatibility of non-incentive-compatible mechanisms. Finally, I will discuss new work on learning within an instance, applied to combinatorial auctions and to revenue preservation in a shrinking multi-unit market.

Big data, machine learning and AI inverts adverse selection problems. It allows insurers to infer statistical information and thereby reverses information advantage from the insuree to the insurer. In a setting with two-dimensional type space whose correlation can be inferred with big data we derive three results: Frist, a novel tradeoff between belief gap and price discrimination emerges. The number of contracts offered is small. Second, we show that forcing the insurance company to reveal its statistical information can be welfare reducing. Third, we show in a setting with naïve agents that do not perfectly infer statistical information from price of offered contracts, price discrimination significantly boosts insurers’ profits.

There is increasing regulatory interest in whether machine learning algorithms deployed in consequential domains (e.g. in criminal justice) treat different demographic groups “fairly.” However, there are several proposed notions of fairness, typically mutually incompatible. Using criminal justice as an example, we study a model in which society chooses an incarceration rule. Agents of different demographic groups differ in their outside options (e.g. opportunity for legal employment) and decide whether to commit crimes. We show that equalizing type I and type II errors across groups is consistent with the goal of minimizing the overall crime rate; other popular notions of fairness are not.

We use a large-scale personalized field experiment on Wikipedia to examine the effect of motivation on the contributions of domain experts to public goods. Our baseline positive response rate is 45%. Furthermore, experts are 13% more interested in contributing when we mention the private benefit of contribution, such as the likely citation of their work. In the contribution stage, using a machine learning model, we find that greater matching accuracy between a recommended Wikipedia article and an expert’s paper abstract, together with an expert’s reputation and the mentioning of public acknowledgement, are the most important predictors of both contribution length and quality. Our results show the potential of scalable personalized interventions using recommender systems to study drivers of prosocial behavior.

This paper presents an empirical framework to study the assignment of children into foster homes and its implications on placement outcomes. The empirical application uses a novel dataset of confidential foster care records from Los Angeles County, CA. The estimates of the empirical model are used to examine policy interventions aimed at improving placement outcomes. In general, it is observed that market thickness tends to improve expected placement outcomes. If placements were assigned across all the administrative regions of the county, the model predicts that (i) the average number of foster homes children go through before exiting foster care would decrease by 8% and (ii) the distance between foster homes and children’s schools would be reduced by 54%.

We examine the potential of healthy food subsidies for reducing nutritional inequality through demand-side interventions. Using a pre-registered field experiment with low-income grocery shoppers, we show that low-cost, scalable behavioral interventions make subsidies substantially more effective. Our unique design allows us to elicit choices and deliver subsidies both before and during a shopping trip. We examine two novel interventions: giving shoppers greater agency through a choice between subsidies and introducing waiting periods designed to prompt deliberation about food purchases. The interventions increase healthy purchases by 61% relative to choiceless healthy subsidies, and 199% relative to a control group.

Of the 650,000 individuals released annually from prisons, two-thirds are re-arrested within three years. Many struggle to fulfill basic needs or associate with new peers. Programs that offer "aftercare services", a comprehensive menu of supportive services (e.g. housing, transportation, job counseling, peer support groups, etc), are potentially important. We partner with an aftercare service provider in Pittsburgh and conduct an RCT where participants are incentivized to use their services at various intensities, i.e. 3 or 5 visits, to receive a reward. The intervention increased visits relative to the control group, and the composition of services used evolved from basic needs to employment and group support as the number of visits increased. However, meeting the 3-visit goal is attainable, whereas participants give up trying to visit 5 times. This is reflected in preliminary results on re-arrests. Modest encouragement is more successful at reducing recidivism.

The University of California recently suspended through 2024 the requirement that California applicants submit SAT scores, upending the major role standardized testing has played in college admissions. We study the impact of this decision and its interplay with other policies (such as affirmative action) on admitted class composition. We develop a market model with schools and students. Students have an unobserved true skill level, a potentially observed demographic group membership, and an observed application with both test scores and other features. Bayesian schools optimize the dual-objectives of admitting (1) the "most qualified" and (2) a "diverse" cohort. They estimate each applicant's true skill level using the observed features and potentially their group membership, and then admit students with or without affirmative action. We show that dropping test scores may exacerbate disparities by decreasing the amount of information available for each applicant. However, if there are substantial barriers to testing, removing the test improves both academic merit and diversity by increasing the size of the applicant pool. We also find that affirmative action alongside using group membership in skill estimation is an effective strategy with respect to the dual-objective. Findings are validated with calibrated simulations using cross-national testing data.

We estimate a model of a college admissions contest with affirmative action (AA) where students compete for seats at better schools by choosing pre-college human capital (HC) investments. We are able to identify and flexibly estimate the contest structure in the college admissions market, including the tendency for AA to affect admissions profiles on different segments of the US college quality spectrum. We identify the effects of college quality, pre-college HC, and unobserved student characteristics on post-college income using a control function derived using methods from the auctions literature. College quality is the primary income determinant, while unobserved student characteristics play a secondary role. Pre-college HC affects income indirectly through its influence on graduation probability and enrollment. We also estimate the fraction of pre-college HC “wasted” by the rat-race nature of the admissions contest, and run counterfactuals of admissions, graduation rates, and post-college income race gaps under alternate admissions schemes not observed in the data.

Affirmative action policies provide a balance between meritocracy and equity in a wide variety of real-life resource allocation problems. We study choice rules where meritocracy is attained by prioritizing individuals based on merit, and equity is attained by reserving positions for target groups of disadvantaged individuals. Focusing on overlapping reserves, the case where an individual can belong to multiple target groups, we characterize choice rules that satisfy maximal compliance with reservations, elimination of justified envy, and non-wastefulness. When an individual accommodates only one of the reserved positions, the horizontal envelope choice rule is the only rule to satisfy these three axioms. When an individual accommodates each of the reserved positions she qualifies for, there are complementarities between individuals. Under this alternative convention, and assuming there are only two target groups, such as women and minorities, paired-admissions choice rules are the only ones to satisfy the three axioms. Building on these results, we provide improved mechanisms for implementing a variety of recent reforms, including the 2015 school choice reform in Chile and 2012 college admissions reform in Brazil.

Randomized Controlled Trials (RCTs) enroll hundreds of millions of subjects and involve many human lives. To improve subjects’ welfare, I propose a design of RCTs that I call Experiment-as-Market (EXAM). EXAM produces a Pareto efficient allocation of treatment assignment probabilities, is asymptotically incentive compatible for preference elicitation, and unbiasedly estimates any causal effect estimable with standard RCTs. I quantify these properties by applying EXAM to a water cleaning experiment in Kenya (Kremer et al., 2011). In this empirical setting, compared to standard RCTs, EXAM improves subjects’ predicted well-being while reaching similar treatment effect estimates with similar precision.

Subsidy programs are typically accompanied by large costs due to the difficulty of screening those who should receive the program from those who would have purchased the good anyway. We design and implement a platform intended to increase the take-up of improved sanitation services by targeting the poorest households for subsidies and using purchases by the wealthy households to increase available subsidies to the poor. We develop a theoretical model designed to isolate the key factors of concern in designing the pricing system. The field project then proceeds in two stages: we first create a demand model based on market data and a demand elicitation experiment, and use the model to predict prices that will maximize take-up subject to an expected budget constraint. We then test the modeled prices on a new sample of households. The treatment led to an increase in market share of mechanical desludging of 4.4 percentage points. The decreased probability of purchasing a manual desludging among those with the largest subsidies was 7.6-8.2 percentage points leading to a market share increase of mechanical desludging of 7.9-9.6 percentage points in that group. The health impacts among neighborhoods with many poor households were large: a 10% increase in the number of poor households in a treatment neighborhood meant that there was a 2.2 percentage point larger decrease in diarrhea. We compare the outcomes of the pricing treatment with alternative targeting methods and pricing structures and show that the pricing treatment outperforms proxy means testing, auctions with perfect pass-through of costs, and straight subsidies on the basis of take-up and/or budget sustainability.

We study a setting in which dynamically arriving items are assigned to waiting agents, who have heterogeneous values for distinct items and heterogeneous outside options. An ideal match would both target items to agents with the worst outside options, and match them to items for which they have high value. Our first finding is that two common approaches – using independent lotteries for each item, and using a waitlist in which agents lose priority when they reject an offer – lead to identical outcomes in equilibrium. Both approaches encourage agents to accept items that are marginal fits. We show that the quality of the match can be improved by using a common lottery for all items. If participation costs are negligible, a common lottery is equivalent to several other mechanisms, such as limiting participants to a single lottery, using a waitlist in which offers can be rejected without punishment, or using artificial currency. However, when there are many agents with low need, there is an unavoidable tradeoff between matching and targeting. In this case, utilitarian welfare may be maximized by focusing on good matching (if the outside option distribution is light-tailed) or good targeting (if it is heavy-tailed). Using a common lottery achieves near-optimal matching, while introducing participation costs achieves near-optimal targeting.

A designer wants to implement a social objective that depends on an unobserved state of the world. Unlike the standard approach, agents do not have ex-ante private information but can acquire costly information on their preferences. The choice of the mechanism generates informational incentives as it affects the information that players choose to acquire before play begins. A mechanism is informationally simple if agents have no incentive to learn about others’ preferences. This endogenizes an “independent private value” property of the interim information structure. We show that a mechanism is informationally simple if and only if it is dictatorial. This holds for generic environments and any smooth cost function which satisfies an Inada condition. Hence even strategy-proof mechanisms incentivize agents to learn about others and require to hold beliefs about opponents’ play. We then show that lack of informational simplicity may induce a new type of rent—even when the cost of information is vanishing—due to incomplete information aggregation. Informationally Simple implementation which could restore strategic-simplicity and avoid this rent is limited. Finally we show that even though agents’ information is endogenously correlated full surplus extraction à la Crémer-McLean is impossible, and we investigate how sequential mechanisms can sometimes restore informational simplicity.

We study the classical sequential hypothesis testing problem (Wald, 1950), but add the memory constraint modelled by finite automata. Generically, the constrained optimum by Bayes rule is impossible to implement with any finite-state automata. We then introduce stochastic finite-state automata with memory constraint and study the constrained optimal rule. Two classes of information structure are considered: the model of breakthroughs in which one signal fully reveals the state of nature but not others, and the more symmetric model where two signals are of similar strengths. In the first, randomization is strictly optimal whenever the memory constraint is binding and the optimum requires some learning. In the second, randomization is not optimal but the optimal finite-automaton uses qualitative probabilities.