Pëllumb Reshidi

Research Associate, Duke University

Photo of Pellumb Reshidi

Contact Information
Department of Economics
Duke University
Social Sciences Building, Office 210B

Curriculum Vitae

Working Papers

In complex environments—where carrying out Bayesian updating is computationally unfeasible—the DeGroot model has emerged as a reliable, and heavily utilized alternative. An assumption present in practically all versions of this model is that agents receive information simultaneously. We relax this assumption by allowing for sequential arrival of information. We find that the final beliefs can be altered by varying only the sequencing of information, keeping the information content unchanged. In this setup the \textit{wisdom of crowds} typically fails: as the number of group members grows, the sequential arrival of information compromises the group’s beliefs, in all but particular cases, beliefs converge away from the truth. We identify the optimal and pessimal information release sequences that yield the highest and lowest attainable consensus, respectively. In doing so, we bound the variation in final beliefs that can be attributed to the variation in the sequencing of information. Groups in which all members are equally influential turn out to be most susceptible to information sequencing.

We test whether the order and timing of information arrival affect beliefs formed within a group. In a lab experiment, participants estimate a parameter of interest using a common and a private signal, as well as past guesses of group members. By varying the sequencing of information arrival, at odds with the Bayesian model, we find that the order and timing of information affect final beliefs, even when the information content is unchanged. Although behavior is non-Bayesian, it is robustly predictable by a model relying on simple heuristics. We explore ways in which the network structure and the timing of information help alleviate correlation neglect. Finally, we document an important heuristic—the influence of private information on participants' actions is time-independent.

We study asymptotic learning when the decision-maker is ambiguous about the precision of her information sources. She aims to estimate a state and evaluates outcomes according to the worst-case scenario. Under prior-by-prior updating, ambiguity regarding information sources induces ambiguity about the state. We show this induced ambiguity does not vanish even as the number of information sources grows indefinitely, and characterize the limit set of posteriors. The decision-maker's asymptotic estimate of the state is generically incorrect. We consider several applications. Among them we show that a small amount of ambiguity can exacerbate the effect of model misspecification on learning, and analyze a setting in which the decision-maker learns from observing others' actions.

Many committees—juries, political task forces, etc.—spend time gathering costly information before reaching a decision. We report results from lab experiments focused on such information-collection processes. We consider decisions governed by individuals and groups and compare how voting rules affect outcomes. We also contrast static information collection, as in classical hypothesis testing, with dynamic collection, as in sequential hypothesis testing. Several insights emerge. Static information collection is excessive, and sequential information collection is non-stationary, producing declining decision accuracies over time. Furthermore, groups using majority rule yield especially hasty and inaccurate decisions. Nonetheless, sequential information collection is welfare enhancing relative to static collection, particularly when unanimous rules are used.

Work in Progress

Speed, Accuracy and Caution: the Timing of Choices Under Risk Aversion

With Joao Thereze

A large empirical literature on the timing of binary choices documents that quicker decisions are often more accurate, even when subjects know the quality of both available options. This evidence suggests individuals decrease the standards with which they make decisions over time, at odds with the classic sequential testing model in which optimal standards are time-independent. We use a novel approximation technique to show that incorporating risk aversion can account for time-dependent standards. Our technique sidetracks many of the difficulties in solving non-stationary optimal stopping problems and allows us to partially characterize the optimal boundaries.

The Economics of Information Intermediaries

With Can Urgun

An information intermediary, such as a news outlet, observes multiple independently developing news topics and must decide which one to use as its headline and when to release the story. The information intermediary weighs the population demand for this information, which decreases in time, and the precision of the information, which in expectation increases in time across all topics. We model each topic's precision as an independent Brownian motion and assume that eventually, the demand for information in the topic drops to zero. We translate the optimization of this problem to an optimal stopping problem over the multi-dimensional process with discounting and a deadline. We find that optimal information release standards decrease in time. For a specific demand function, we find an explicit boundary that determines the news intermediary's optimal behavior.