An Experiment on Belief Formation in Networks
Posted: 29 Nov 2013 Last revised: 28 Jun 2016
Date Written: July 4, 2014
We analyse belief formation in social networks in a laboratory experiment. Our 3 x 3 design varies the network structure and the amount of information agents have about it. Agents observe an imperfect private signal on the true state of the world and then repeatedly guess the true state, observing the guesses of their network neighbours in each period. Participants' individual choices are well explained by a model of naive learning, but not by Bayesian learning. Comparative statics regarding signal dispersion and the time it takes to reach a consensus are also in line with the naive model. The model predictions regarding whether a consensus is reached and whether the truth is learned are only partially reflected in the data. Changes in behaviour induced by the amount of information participants have about the network structure cannot be explained by the naive model at all. We then estimate a larger class of models and find that the most successful participants do account for correlations in neighbours' guesses (unlike the naive model suggests), place less weight on themselves compared to others, but increase the weight placed on themselves over time. We propose a simple belief formation model that reflects these properties and show that it does well when confronted with the data.
Keywords: Networks, Learning, Beliefs, Experiments
JEL Classification: C70, C91
Suggested Citation: Suggested Citation