Essays on Bounded Rationality and Information Processing

Abstract

This thesis investigates the role of information processing in decision making. Chapter 1 develops a model of bounded rationality where an agent has a limited ability to combine different pieces of information. A sophisticated principal attempts to exploit this limitation by designing a complex contract: a collection of clauses, each providing information about the consequences of different actions. By carefully designing the contract, the principal manipulates the agent into believing that truth-telling is optimal. This model of bounded rationality extends to other settings and provides a theory of what it means for a problem to be difficult for a human to solve.

Chapter 2 develops a revealed-preference model of information disclosure and Bayesian belief updating. One decision maker ranks information sources (Blackwell (1951) experiments) that a second decision maker processes before taking an action that affects them both. The main result is a representation theorem characterizing the decision makers as expected utility maximizers with (possibly) differing priors and utility functions. In addition, the representation characterizes the second decision maker as a Bayesian information processor. The model allows comparisons to be made between the priors and utility functions of the decision makers by examining the first decision maker's preference over Blackwell experiments. Combined with the main representation theorem, this establishes preference for information as a useful new primitive of revealed-preference analysis.

Finally, Chapter 3 characterizes Bayesian subjective expected utility maximization in a model with finitely many states and outcomes. By employing a rich set of signal-contingent preferences over (Savage) acts, beliefs and utilities can be identified and Bayesian belief updating tested. This contrasts two well-known limitations of the standard Savage (1954) framework: (i) it does not explicitly test whether individuals are Bayesian information processors, and (ii) with finitely many states and outcomes, beliefs and utility functions cannot be pinned down.