Philippe Jehiel

We develop a dynamic model where successive, decentralized policy makers must decide whether to implement affirmative action policies aimed at improving the performance of future generations of a targeted group. Employers do not perfectly observe if a worker benefited from affirmative action, but take that possibility into account, resulting in the devaluation of the worker’s credentials and an associated feeling of injustice. We establish that, in equilibrium, affirmative action is implemented perpetually by benevolent policy makers, despite the feeling of injustice that eventually dominates the anticipated benefits. This contrasts with the first best, which requires affirmative action to be temporary.

We study the existence of efficient auctions in private value settings in which some bidders form their expectations about the distribution of their competitor’s bids based on the accessible data from past similar auctions consisting of bids and expost values. We consider steady states in such environments with a mix of rational and data-driven bidders, and we allow for correlation across bidders in the signal distributions about the ex post values. After reviewing the working of the approach in second-price and first-price auctions, we establish our main result that there is no efficient auction in such environments.

Lack of diversity in higher education is partly driven by long-run belief distortions about admission chances at elite colleges. We depart from the rational expectation framework and propose a simple model of expectation formation in which students estimate their admission chances by sampling a pool of given size τ of peers who previously applied to elite colleges. Assuming students consider peers with abil-ity as close as possible to their own, two types of inefficiencies arise in steady state: high-achieving disadvantaged students self-select out of elite colleges, and average students from advantaged families apply to elite colleges even though their true admission chances are null. We then explore the working of the model when students from several possibly dissimilar neighborhoods compete for the same positions, thereby highlighting externalities related to the comparative neighborhood com-positions. Several policy instruments such as quotas or the mixing of neighborhoods are considered.

This study explores selection neglect in an experimental investment game where individuals can learn from others’ outcomes. Experiment 1 examines aggregate-level equilibrium behavior. We find strong evidence of selection neglect and corroborate several comparative static predictions of Jehiel’s (2018) model, showing that the severity of the bias is aggravated by the sophistication of other individuals and moderated when information is more correlated across individuals. Experiment 2 focuses on individual decision-making, isolating the influence of beliefs from possible confounding factors. This allows us to classify individuals according to their degree of naivety and explore the limits of, and potential remedies for, selection neglect.

Families of normal-form two-player games are categorized by players into K analogy classes applying the K-means clustering technique to the data generated by the distributions of opponent’s behavior. This results in Calibrated Analogy-Based Expectation Equilibria in which strategies are analogy-based expectation equilibria given the analogy partitions and analogy partitions are derived from the strategies by the K-means clustering algorithm. We discuss various concepts formalizing this, and observe that distributions over analogy partitions are sometimes required to guarantee existence. Applications to games with linear best-responses are discussed highlighting the differences between strategic complements and strategic substitutes.

We develop a framework for categorization in games, applicable both to multistage games of complete information and static games of incomplete information. Players use categories to form coarse beliefs about their opponents’ behavior. Players best-respond given these beliefs, as in analogy-based expectations equilibria. Categories are related to previously used strategies via the requirements that categories contain a sufficient amount of observations and exhibit sufficient withincategory similarity, in line with the bias-variance trade-off. When applied to classic games including the chainstore game and adverse selection games our framework yields less unintuitive predictions than those arising with standard solution concepts.

We offer an approach to cooperation in repeated games of private monitoring in which players construct models of their opponents’ behavior by observing the frequencies of play in a record of past plays of the game in which actions but not signals are recorded. Players construct models of their opponent’s behavior by grouping the histories in the record into a relatively small number of analogy classes for which they estimate probabilities of cooperation. The incomplete record and the limited number of analogy classes lead to misspecified models that provide the incentives to cooperate. We provide conditions for the existence of equilibria supporting cooperation and equilibria supporting high payoffs for some nontrivial analogy partitions.

We study learning in a decentralized pairwise adverse selection economy, where buyers have access to the quality of traded goods but not to the quality of nontraded goods. Buyers categorize ask prices in order to predict quality as a function of ask price. The categorization is endogenously determined so that outcomes that are observed more often are categorized more finely, and within each category beliefs reflect the empirical average. This leads buyers to have a very fine understanding of the relationship between qualities and ask prices for prices below the current market price, but only a coarse understanding above that price. We find that this induces a price cycle involving the Nash equilibrium price, and one or more higher prices.

We study the existence of e¢cient auctions in private value settings in which some bidders choose their bids based on the accessible data from past similar auctions consisting of bids and ex post values. We consider steady-states in such environments with a mix of rational and data-driven bidders, and we allow for correlation across bidders in the signal distributions about the ex post values. After reviewing the working of the approach in second-price and first-price auctions, we show our main result that there is no e¢cient auction in such environments.