job market paper


Previous economic models of memory have failed to incorporate one of its critical features: decay. Memories fade over time, losing fidelity and leading to a host of economically interesting behaviors, with recency bias being particularly important. In this paper we present a model of rational memory with decay where players make rational decisions about how much costly cognitive effort to devote to preserving memories of informative events. We apply the model to predict and analyze several phenomena, including recency bias in elections, where events that occur close to elections have disproportionately large impacts on voting behavior. The model also predicts the phenomenon of “insurance cycles” whereby individuals buy insurance after a disaster then let it lapse when no further disasters are forthcoming, contrary to the predicted behavior of perfect learning Bayesian agents. We also apply the model to a game of manipulation through the timing of information release. This application shows how a sophisticated sender can profitably influence the behavior of a rational agent just by changing when they release information.

completed papers

Network Formation

Vying for Dominance: An Experiment in Dynamic Network Formation
Revision requested: Journal of Economic Behavior and Organization

Centrality in a network is highly valuable. This paper investigates the idea that the timing of entry into the network is a crucial determinant of a node's final centrality. We propose a model of strategic network growth which makes novel predictions about the forward-looking behaviors of players. In particular, the model predicts that agents entering the network at specific times will "vie for dominance"; that is they will make more connections than is myopically optimal in hopes of receiving additional connections from future players and thereby becoming dominant. The occurrence of these opportunities varies non monotonically with the parameters of the game. In a laboratory experiment, we find that players do exhibit “vying for dominance” behavior, but do not always do so at the predicted critical times. We find that a model of heterogeneous risk aversion best fits the observed deviations from initial predictions. Timing determines whether players have the opportunity to become dominant, but individual characteristics determine whether players exploit that opportunity.
In many networks, a few highly central nodes have out-sized impacts on the behavior of the system and generate a large amount of value from their position, but what determines which nodes become central? We hypothesize that the timing of entry into the network can play a critical role. In this paper, we present a new dynamic model of network formation with history dependence, growth, and forward looking strategic agents. These features can generate novel strategic behaviors such as “vying for dominance,” whereby an individual makes many connections as he joins the network because he expects doing so will attract more connections from later nodes. We find that all players either vie for dominance or play myopically in equilibrium. Furthermore, if we assume players use a novelty seeking tie-breaking rule, the solution is characterized by periodic vying for dominance separated by periods of low connection, myopic play. Because vying becomes more expensive as the network grows, the time between profitable vying opportunities increases exponentially over time, and the network becomes more stagnant.

Rational Inattention

Estimating Information Cost Functions in Models of Rational Inattention (with Ambuj Dewan)
Revision requested: Journal of Economic Theory

In models of rational inattention, information costs are usually modeled using mutual information, which measures the expected reduction in entropy between prior and posterior beliefs, or ad hoc functional forms, but little is known about what form these costs take in reality. We show that under mild assumptions on information cost functions, including continuity and convexity, gross payoffs to decision makers are non-decreasing and continuous in potential rewards. We conduct laboratory experiments consisting of simple perceptual tasks with fine-grained variation in the level of potential rewards that allow us to test several hypotheses about rational inattention and compare various models of information costs via information criteria. We find that most subjects exhibit monotonicity in performance with respect to potential rewards, and there is mixed evidence on continuity and convexity of costs. Moreover, a significant portion of subjects are likelier to make small mistakes than large ones, contrary to the predictions of mutual information. This suggests that while people are generally rationally inattentive, their cost functions may display non-convexities or discontinuities, or they may incorporate some notion of perceptual distance. The characteristics of a decision-maker’s information cost function have implications for various economic applications, including investment.
We use laboratory experiments to test models of ‘rational inattention,’ in which people acquire information to maximize utility from subsequent choices net of information costs. We show that subjects adjust their attention in response to changes in incentives a manner which is broadly in line with the rational inattention model but which violates models such as random utility in which attention is fixed. However, our results are not consistent with information costs based on Shannon entropy, as is often assumed in applied work. We find more support for a class of ‘posterior separable’ cost functions which generalize the Shannon model.

turkey tail different colors.JPG

papers in progress

Experiments on Learning about Insurable Risk with Rational Memory


Kunreuther et al. 2013 document a phenomenon whereby the demand for disaster insurance increases after a disaster and then falls when no further disasters arise, even when no serial correlation is present. If individuals were perfect Bayesian learners, we would expect these responses to disappear asymptotically, but they appear to be persistent in equilibrium. These “insurance cycles” can be explained by a more general phenomenon called recency bias where more recent events have an out-sized impact on beliefs and behavior. New theoretical work in Neligh (2019) has demonstrated how recency bias and insurance cycles can naturally result from a rational memory model where memories decay over time, but individuals preserve their memories better by expending costly cognitive resources. In this paper, we conduct an experiment testing some predictions of the rational memory model in an insurance purchasing game. We find that players exhibit strong recency bias, as predicted by the model. We also test whether player behavior is better described by a rational memory model or a traditional reinforcement learning model by separately manipulating the value of information and the reward associated with it. We find that both models are valuable, because neither reinforcement nor information value alone is enough to promote substantial learning. Only when information is valuable and reinforcement is present does substantial learning occur.

Re-Encoding for Realistic Memory Models

Modelers have struggled with incorporating the limits of human memory into their models in ways that are tractable, realistic, and make good testable predictions about human behavior across a wide set of contexts. In this paper, we bring together tools and results from neuroscience, coding theory, and the rational inattention literature in order to create a model of memory which matches with stylized facts from the psychology literature. In this model, players encode their memories as strings of bits which decay over time by randomizing with a fixed probability. We show that a simple model in which memories are encoded and left to decay is inconsistent with the observed rate at which individuals forget memorized words. However, by incorporating re-encoding, a process whereby the brain periodically examines existing memories and attempts to correct errors that have arisen, we produce much more realistic predictions. Re-encoding also makes predictions that are broadly consistent with many other features of memory, such as the spacing effect, the benefits of explicit recall, the tenuous connection between confidence and accuracy, and constructed memory.

Rational Inattention

Rational Attention with Perceptual Distances (with Ambuj Dewan)

In recent years, a great deal of work has been done on the idea of rational attention, the concept that people can rationally choose the type of quality of signal they receive about the world with more informative signals being more costly. Thus far, however, most models of rational inattention have not been able to incorporate perceptual distance, ie the fact that some states of nature are easier to distinguish than others. In this paper, we will be proposing and testing a number of models of rational attention which incorporate perceptual distance. We consider models based on normal signals, Markov transition matrices, and several models in which the probability of confusing two states is a fixed function of their distance under some metric. The models are tested and compared using data from incentivized dot counting and angle differentiation tasks.