Influential listeners: An experiment on persuasion bias in social networks

Abstract

This paper presents an experimental investigation of persuasion bias, a form of bounded rationality whereby agents communicating through a social network are unable to account for possible repetitions in the information they receive. The results indicate that network structure plays a signicant role in determining social inuence. However, the most inuential agents are not those with more outgoing links, as predicted by the persuasion bias hypothesis, but those with more incoming links. We show that a boundedly rational updating rule that takes into account not only agents' outdegree, but also their indegree, provides a better explanation of the experimental data. In this framework, consensus beliefs tend to be swayed towards the opinions of inuential listeners. We then present an eort-weighted updating model as a more general characterization of information aggregation in social networks.

An Experiment on Belief Formation in Networks

“…Equation Since we estimate λ t ij;k separately for each time period, network and network position this means that some estimations (in particular those for the hub in the star and most kite positions) would be based on only 6 observations. For these positions only we estimate instead λ τ ij;k for three or four adjacent time periods, where τ = 1 for t = 2, 3, 4 (τ = 2 for t = 5, 6, 7,.., τ = 6 for t = 17, 18,19,20). As a consequence each λ t ij;k is estimated based on between 18 to 48 observations.…”

The Strength of Weak Leaders - An Experiment on Social Influence and Social Learning in Teams

“…Similarly, Choi et al (2005), Ç elen and, and Ç elen et al (2010) study predictions of social learning models experimentally. A caveat of this theory-testing approach is that the participants are confronted with highly stylized tasks such as guessing an average (or its sign) of randomly drawn numbers (Corazzini et al, 2012;Ç elen and Kariv, 2005;Ç elen et al, 2010) or finding an abstract true state (Choi et al, 2005;Grimm and Mengel, 2016;Chandrasekhar et al, 2016). It is questionable how the investigated learning behavior transfers to settings with real questions.…”

“…For instance, Corazzini et al (2012), Grimm and Mengel (2016), and Chandrasekhar et al (2016) compare sophisticated models of Bayesian learning with simple models of naïve learning in settings in which their predictions diverge. The common conclusion is that the observations are more often consistent with the simple models.…”

“…, i.e., pendants' initial opinions enter the calculation of the consensus with a weight of 20% each, while the center's initial opinion accounts for 40% of the consensus. Corazzini et al (2012) suggest improving the DeMarzo et al Model by increasing the weight of agents who listen to many other agents (and show that this twist improves the model fit to experimental data). The suggested specification is that the weights are 20 In the conservatism models (consisting of the specifications Standard-Plus and Sophisticated-Plus), we make assumptions about higher-order beliefs that close the model in the sense that no agent will expect another agent to behave in a different manner than in the one observed.…”

“…Previous experimental research on social learning has not always found strong support for Bayesian learning, but often suggests that simple rules of updating, such as repeatedly taking averages, fit the data well (Corazzini et al, 2012;Grimm and Mengel, 2016;Battiston and Stanca, 2014;Chandrasekhar et al, 2016). We use their common modeling approach, which is often named after Morris DeGroot, to generate an alternative prediction and to later specify models of more naïve learning.…”