Behavioral and Structural Barriers to Information Aggregation in Networks

We investigate the impact of network architecture on learning dynamics in medium-sized networks using urn-guessing laboratory experiments. Participants are incentivized to guess the true state of nature based on two sources of information: (i) a private, noisy signal received prior to the start of the game and (ii) past guesses of their immediate neighbors. Our findings reveal that subjects consistently under-react to new information and are reluctant to imitate neighbors with strictly better information. Additionally, we identify the "Royal Family'' and "Bottleneck'' structural barriers predicted by theory. These insights help explain three key network-level observations. First, imperfect learning occurs in all networks, including the fully connected one. Second, networks with clusters struggle to process highly noisy yet informative signal distributions. Finally, networks with a single central individual connected to everyone perform surprisingly poorly when the distribution of private signals strongly favors one action. We also demonstrate that the behavioral frictions can be partially alleviated by reducing the amount of information disseminated in the network. Overall, we argue that our behavioral approach offers a novel perspective on information aggregation in networks, complementing standard Bayesian and Naive (a la DeGroot) models.

(with Marina Agranov and Ben Gillen)

Dotan Persitz's google scholar page: https://scholar.google.com/citations?user=Kevii0IAAAAJ&hl=en