Learning and Equilibrium under Model Misspecification

ArXiv.org · 2026-01-14

This chapter develops a unified framework for studying misspecified learning situations in which agents optimize and update beliefs within an incorrect model of their environment. We review the statistical foundations of learning from misspecified models and extend these insights to environments with endogenous, action-dependent data, including both single agent and strategic settings.

Learning and Equilibrium under Model Misspecification

arXiv (Cornell University) · 2026-01-14

This chapter develops a unified framework for studying misspecified learning situations in which agents optimize and update beliefs within an incorrect model of their environment. We review the statistical foundations of learning from misspecified models and extend these insights to environments with endogenous, action-dependent data, including both single agent and strategic settings.

Berk-Nash Rationalizability

ArXiv.org · 2025-05-27

We study learning in complete-information games, allowing the players' models of their environment to be misspecified. We introduce Berk--Nash rationalizability: the largest self-justified set of actions -- meaning each action in the set is optimal under some belief that is a best fit to outcomes generated by joint play within the set. We show that, in a model where players learn from past actions, every action played (or approached) infinitely often lies in this set. When players have a correct model of their environment, Berk--Nash rationalizability refines (correlated) rationalizability and coincides with it in two-player games. The concept delivers predictions on long-run behavior regardless of whether actions converge or not, thereby providing a practical alternative to proving convergence or solving complex stochastic learning dynamics. For example, if the rationalizable set is a singleton, actions converge almost surely.

Investment under risk vs ambiguity

AEA Randomized Controlled Trials · 2024-04-22

Investment under risk vs ambiguity

AEA Randomized Controlled Trials · 2024-04-22

Mental Models and Learning: The Case of Base-Rate Neglect

American Economic Review · 2024 · 54 citations

• Computer Science
• Cognitive psychology
• Computer Science

We experimentally document persistence of suboptimal behavior despite ample opportunities to learn from feedback in a canonical updating problem where people suffer from base-rate neglect. Our results provide insights on the mechanisms hindering learning from feedback. Importantly, our results suggest mistakes are more likely to be persistent when they are driven by incorrect mental models that miss or misrepresent important aspects of the environment. Such models induce confidence in initial answers, limiting engagement with and learning from feedback. We substantiate these insights in an alternative scenario where individuals involved in a voting problem overlook the importance of being pivotal. (JEL D83, D91)

Mental Models and Transfer Learning

AEA Papers and Proceedings · 2023-05-01 · 6 citations

Using a laboratory experiment, we investigate the extent to which learning is transferred between related problems in the context of an updating task. The updating principle we study requires updating positively after a positive signal and negatively after a negative signal. In the first environment, most subjects initially fail to satisfy the principle but eventually adjust after feedback. The environment they face subsequently presents the same challenges but with different parameter values. We find weak evidence for transfer learning: only half of the subjects who learn to be consistent with the principle remain so when the parameter values are changed.

Contingent Thinking and the Sure-Thing Principle: Revisiting Classic Anomalies in the Laboratory

Zenodo (CERN European Organization for Nuclear Research) · 2023-09-06 · 9 citations

We present an experimental framework to study the extent to which failures of contingent thinking explain classic anomalies in a broad class of environments, including overbidding in auctions and the Ellsberg paradox. We study environments in which the subject’s choices affect payoffs only in some states, but not in others. We find that anomalies are in large part driven by incongruences between choices in the standard presentation of each problem and a `contingent' presentation, which focuses the subject on the set of states where her actions matter. Additional evidence suggests that this phenomenon is in large part driven by people's failure to put themselves in states that have not yet happened even though they are made aware that their actions only matter in those states.

Contingent Thinking and the Sure-Thing Principle: Revisiting Classic Anomalies in the Laboratory

The Review of Economic Studies · 2023 · 29 citations

• Computer Science
• Economics
• Positive economics

Abstract We present an experimental framework to study the extent to which failures of contingent thinking explain classic anomalies in a broad class of environments, including overbidding in auctions and the Ellsberg paradox. We study environments in which the subject’s choices affect payoffs only in some states but not in others. We find that anomalies are in large part driven by incongruences between choices in the standard presentation of each problem and a “contingent” presentation, which focuses the subject on the set of states where her actions matter. Additional evidence suggests that this phenomenon is in large part driven by people’s failure to put themselves in states that have not yet happened even though they are made aware that their actions only matter in those states.

Seeing What is Representative

The Quarterly Journal of Economics · 2023-05-27 · 20 citations

Abstract We provide evidence for a bias that we call “representative signal distortion” (RSD), which is particularly relevant to settings of statistical discrimination. Experimental subjects distort their evaluation of new evidence on individual group members and interpret such information to be more representative of the group to which the individual belongs (relative to a reference group) than it really is. This produces a discriminatory gap in the evaluation of members of the two groups. Because it is driven by representativeness, the bias (and the discriminatory gap) disappears when subjects are prevented from contrasting different groups; because it is a bias in the interpretation of information, it disappears when subjects receive information before learning of the individual’s group. We show that this bias can be easily estimated from appropriately constructed data sets and can be distinguished from previously documented inferential biases in the literature. Importantly, we document how removing the bias produces a kind of free lunch in reducing discrimination, making it possible to significantly reduce discrimination without lowering accuracy of inferences.