Social Learning in Networks: Theory and Experiments

Abstract

This paper presents a non-Bayesian model of social learning in networks in an environment with a finite set of actions. We conduct a laboratory experiment in which participants play an urn-guessing game over several decision rounds while observing the previous choices of the network members to whom they are connected. We identify three properties of individual choice revision: consistency, monotonicity and identity independence. We consider the class of revision functions satisfying such properties and establish that consensus occurs in arbitrary strongly connected networks if and only if the revision functions of all agents are identical and preference-based. Thus, consensus is hard to achieve, which is supported by evidence from our experiment. The theoretical prediction differs sharply from the existing results in the Bayesian and non-Bayesian literature.