Rationalizing Dynamic Choices

2019-03-29

Social Science Research Network

AI Summary

This paper studies how an observer determines whether a sequence of observable actions can be rationalized by a Bayesian agent with endogenous information acquisition and updating, under a known utility function. Method: The authors derive the first necessary and sufficient condition for dynamic rationalizability: behavior is irrationalizable if and only if a universally dominant deviation exists—a deviation that strictly improves expected utility across all possible information structures. This condition is information-structure-free, greatly enhancing testability. Contribution/Results: The framework extends to stochastic choice, enabling monotonic rationalization under risk aversion, empirical falsification of Bayesian models, feasibility characterization in dynamic information design, and partial identification of utility parameters. Its core innovation is a verifiable dominance-based rationalizability criterion, which reveals that stronger risk aversion weakens predictive power of behavior and permits preference identification without assuming any specific information structure.

Technology Category

• Reasoning under Uncertainty: Decision/Utility Theory
• Game Theory and Economic Paradigms: Imperfect Information
• Multiagent Systems: Mechanism Design

Application Category

• Economics, Online Markets and Human Computation: Incentives in network design for Web infrastructures and ecosystems
• User Modeling, Personalization and Recommendation: Metrics for user behavior and evaluating success
• Responsible Web: Machine-in-the-loop, human agency and autonomy

Abstract

Consider an analyst who observes an agent taking a sequence of actions. The analyst ponders whether the sequence of actions observed could have been taken by a rational, Bayesian agent. Although the analyst observes the chosen actions, he does not have direct access to the agent’s information and must therefore consider a multitude of possibilities. Could some gradual release of information have led the agent to optimally take that sequence of actions? We show that a sequence of actions cannot be rationalized by any information structure if and only if it can be proved to be dominated via a deviation argument. This argument prescribes a way of deviating that would leave the agent better off in any possible scenario, regardless of the information she might have. As an application of this characterization, we show that an increase in the agent’s risk-aversion leads to less predictive power—more sequences of actions can be rationalized. We also show results that simplify the analyst’s search for a deviation argument and demonstrate how these arguments can be used to partially identify utility parameters without making assumptions on the agent’s information.

Problem

• Rationalizing dynamic choices without agent's information
• Testing justification via Bayesian model with utility function
• Extending result to distributions over action sequences

Innovation

• Rational Bayesian model justification check
• Single deviation argument for agent actions
• Extension to action sequence distributions