Models of costly information acquisition have grown in popularity in economics. However, little is known about what form information costs take in reality. We show that under mild assumptions on costs, including continuity and convexity, gross payoffs to decision makers are non-decreasing and continuous in potential rewards. We conduct experiments involving simple perceptual tasks with fine-grained variation in the level of potential rewards. Most subjects exhibit monotonicity in performance with respect to potential rewards, and evidence on continuity and convexity of costs is mixed. Moreover, subjects' behavior is consistent with a subset of cost functions commonly assumed in the literature.

We use laboratory experiments to test models of rational inattention, in which people acquire information to maximize utility net of information costs. We show that subjects adjust their attention in response to changes in incentives, in line with the rational inattention model. However, our results are qualitatively inconsistent with information costs that are linear in Shannon entropy, as is often assumed in applied work. Our data are best fit by a generalization of the Shannon model, which allows for a more flexible response to incentives and for some states of the world to be harder to distinguish than others.

Historically, economic models of memory have neglected a pivotal aspect: decay, the established notion that memories deteriorate over time, resulting in a loss of fidelity. In this paper we show that a framework of rational memory with decay can produce the recency effect and other intuitively appealing memory phenomena. We apply the framework to both finite time and dynamic programming settings, deriving a wide range of results with a focus on comparative statics and asymptotic behaviors. Many of our results can also be applied to multi-dimensional rational inattention settings where individuals have to split their attention across several information sources.

We consider what happens to a memory if, during the decay process, an individual periodically attempts to reset their memory to its original pre-decayed state based on their current beliefs. We find that the resulting “posterior-modal re-encoding” is optimal and improves the performance of the memory system but also introduces some notable features including a rehearsal/testing effect, a spaced rehearsal effect, exponential decay in performance, false/constructed memory, and a limited correlation between memory accuracy and confidence. These feature all arise human memory. Therefore, we suggest that something resembling posterior-modal re-encoding may be taking place in the human brain during recall and neural replay.

Many modern economic games use endogenous information as a core feature, including models of communication, disclosure, persuasion, and rational inattention. In simple endogenous information environments, the mapping between signals and actions can often be assumed to be straightforward such that the player takes the action corresponding to the signal received, because signal labels do not matter. However, in more realistically complex information environments, particular those with multiple sources of noise, straightforwardness is not guaranteed. We show that straightforwardness is equivalent to a specific matrix product being diagonally row-maximal. We then provide a sufficient condition for a sequence of garblings to result in a straightforward signal. The condition is based on a class of diagonally row-maximal matrices that is closed under multiplication. Intuitively, this condition forces errors, priors, and utility penalties to be symmetric and quasi-concave in a circular space. Using this condition, we can easily solve complex information design problems.

Previous economic models have generally treated information processing as a one-step process (as in models of rational inattention) or as a consistent dynamic process (as in drift diffusion models). However, from work in psychology and computer science we know that information processing has multiple distinct phases. At minimum, information must be acquired, stored, and retrieved before it is used. We model this dynamic process as a three-part effort allocation problem reminiscent of a dynamic input selection problem for a firm. Based on the theory, we run an experiment which can separately identify attention, memory, and retrieval errors. Results show that in the laboratory setting, most errors are made during the attention phase while very few errors are made during storage.

We propose novel generalizations of two prominent theories of learning - Bayesian updating and operant conditioning - that provide differential, non-overlapping testable implications, based purely on which observables people learn from, and allow for separation and identification of learning types. We propose an experiment designed to generate maximal conflict between the theoretical predictions of these theories, to cleanly distinguish between them. This incentivized experiment will allow us to classify observed behavior in various contexts, across both laboratory and remote settings, according to our generalizations of the two theories. It will also allow us to identify the distribution of learning modalities in the subject population.