We develop a theoretical framework for analyzing how learning occurs directly from data in the absence of prior information. We interpret learning as the process through which decision makers identify a model, defined as the set of variables that matter for the outcome, and use those variables to form behavioral rules. The comparative statics show that extraction of the optimal rule depends not only on its own value, but also on how clearly it outperforms competing alternatives. Differences in cognitive costs then make agents respond differently to changes in relative rule values, generating systematic and predictable disagreement even when they face the same data. Finally, the framework makes rule-specific costs identifiable from behavior, allowing us to determine which observable measures of complexity can be interpreted as entering the cost function. We test the framework in our experiment and in a reanalysis of data from Kendall and Oprea (2024), finding strong support for its predictions in both settings.

We study, empirically and theoretically, how restitution helps restore cooperation after a breach or an exploratory defection. Restitution strategies propose a return to cooperation by cooperating against defection, and condition actions on the balance between cooperation given and received. We reanalyze experimental data from repeated Prisoner's Dilemma games and find empirical support for restitution strategies in general, and for a strategy we name Payback, in particular.  Besides explaining how subjects deal with conflicts, accounting for restitution strategies helps reconcile discrepancies between theory and experiments—such as the widespread use of non-equilibrium strategies like Tit-for-Tat and the limited predictive power of risk dominance under imperfect monitoring.

Choice behavior is significantly influenced by context. Evidence from cognitive science suggests that this dependency may be driven by associations. In this paper, we propose a choice-theoretic model of contextual associations. We formalize contexts as sets of topics. This allows us to define associations as links between alternatives and topics that directly impact utility. We characterize the model's empirical content, establish conditions for identification, analyze its comparative statics, and propose several extensions. To demonstrate the model's portability, we apply it to a probabilistic voting setting providing a new mechanism based on contextual associations explaining political polarization.

In this paper I study the trade-off between efficiency and strategic risk in the context of a repeated Prisoner's Dilemma with imperfect public monitoring. Because unfavorable public signals and resulting punishments can occur on the equilibrium path, keeping the value of cooperation high requires players to adopt more lenient and forgiving strategies, which leaves them more exposed to opportunistic behavior. For a broad class of one-dimensional public signals, I show that under imperfect monitoring, strategy adoption responds to beliefs about the opponent's strategy, whereas under perfect monitoring, beliefs affect only the extensive margin—that is, the decision of whether to cooperate at all. Finally, I use experimental data from Aoyagi and Fréchette (2009) to document suggestive patterns consistent with subjects choosing strategies to maximize the cooperation value rather than controlling for strategic risk. 

We develop a theoretical framework to study the fundamental attribution error in hiring: the tendency to over-attribute observed performance to worker ability and under-attribute it to circumstances. In our framework, employers learn how observed performance maps into underlying ability and circumstances directly from data, rather than updating pre-existing knowledge withing a given mental model. This allows us to link the emergence and magnitude of attribution errors to measurable features of the environment, including the statistical properties of the data and the relative predictive power of ability and task difficulty. Heterogeneity in individuals’ costs of processing information further generates systematic differences in the extent of the bias. We therefore move beyond treating the fundamental attribution error as a fixed predisposition by linking it to objective features of the environment, thereby identifying potential margins for policy intervention. We test these predictions in a laboratory hiring experiment.