Elsevier

European Economic Review

Volume 56, Issue 6, August 2012, Pages 1276-1288
European Economic Review

Influential listeners: An experiment on persuasion bias in social networks

https://doi.org/10.1016/j.euroecorev.2012.05.005Get rights and content

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 repetitions in the information they receive. We find that, after repeated communication within a social network, social influence depends not only on being listened to by many others, but also on listening to many others. We show that persuasion bias can be viewed as an extreme case of a generalized boundedly rational updating rule in which agents receive more or less attention depending on how many other agents they listen to. The results indicate that behavior in the experiment is consistent with an updating rule according to which agents' social influence is proportional to their indegree.

Highlights

► We study experimentally persuasion bias in social networks. ► We find that the most influential agents are those who listen to many others. ► We provide a generalized boundedly rational updating rule. ► Persuasion bias can be viewed as an extreme case of this rule. ► The results support social influence weights proportional to agents indegree.

Introduction

In many social and economic contexts, communication between individuals is determined by social networks (Jackson, 2007, Udry and Conley, 2001). Individuals learn by observing the behavior of those they are connected with in their local environment. In these settings, an important issue concerns how dispersed information held by different individuals can be aggregated over time. In this paper, we investigate experimentally how the communication structure of a social network may affect this aggregation process and, as a consequence, determine social influence.

It is possible to imagine several situations in which different individuals have noisy signals regarding an underlying state of the world. This state of the world may represent the quality of a new product, the returns on a potential investment, the ability of a political candidate to carry out reforms or, more generally, any unknown condition or action that affects the payoff of all individuals in the same way. Consider, for example, a situation where the unknown state of the world is the level of crime in one's neighborhood. Each individual has some information on this issue reflecting her own personal experience. By observing whether other individuals with whom she is in direct contact are installing burglar alarms or purchasing more sophisticated locks or even carrying guns, an individual may draw inferences about the information observed by her direct neighbors. Over time, because of lack of common knowledge about the actions of all individuals in the community, an individual can try to infer her neighbors' knowledge of her neighbors' actions and the private information they reveal. It is apparent that the complexity of the learning problem increases substantially over time.

Whether individuals are capable of rationally processing the information circulating in their social network is an empirical question. If individuals were Bayesian, then the network structure would not affect their ability to reach an efficient consensus.1 However, different network structures may influence the information aggregation process by determining the nature and complexity of the inference problem faced by individuals. In this context, recent studies have attempted to model deviations from rationality. In particular, DeMarzo et al. (2003) model persuasion bias as the outcome of a mechanical updating process, according to which individuals fail to account for repetitions of information when communicating within a network.2 An important implication of the persuasion bias hypothesis is that, after iterated communication, the members of a network converge to a consensus that is biased towards the private signals of the most influential agents, namely those who are listened to by many other agents, either directly or indirectly.

In this paper, we test experimentally whether the evolution of beliefs of individuals communicating through a social network reflects the structure of the network and, in particular, whether consensus beliefs are consistent with the persuasion bias hypothesis. Our results indicate that the structure of the network plays a significant role in determining consensus beliefs. More specifically, we find that social influence depends not only on whether an agent is listened to, directly or indirectly, by many other agents, but also on whether she listens to many other agents: consensus beliefs tend to be swayed towards the opinions of influential listeners. In order to explain this finding, we propose a generalized boundedly rational updating rule to describe social learning that takes into account not only an agent's outdegree (the number of outgoing links) but also her indegree (the number of incoming links). Within this framework, agents who listen to many others may receive relatively more attention. We find that the experimental results are consistent with an updating rule where indegree is weighted proportionally.

The paper is organized as follows. Section 2 briefly reviews the related literature. Section 3 describes the experimental design and procedures. Section 4 discusses the theoretical predictions and hypotheses to be tested. Section 5 presents the experimental results. Section 6 proposes a generalized updating rule to interpret the experimental results. Section 7 concludes.

Section snippets

Related literature

The present work is related to the extensive theoretical literature on social learning. This literature can be generally divided into two main strands: one that focuses on Bayesian learning and the other that deals with myopic or boundedly rational learning.

The literature on Bayesian learning originates from the contributions of Bikhchandani et al. (1992) and Banerjee (1992), who assume an exogenous sequential structure in which each agent, after observing all past actions, optimally updates

The experiment

We consider a repeated learning problem, adapted from DeMarzo et al. (2003), where communication between individuals occurs within a social network. The experiment is designed to test if social influence is affected by the structure of the network. In this section, we describe the experimental task, treatments and procedures. We also briefly discuss the rationale for the specific features of our experimental design. The next section presents the theoretical predictions and hypotheses to be

Theoretical predictions and hypotheses

Let N={A,B,C,D} denote the set of agents, where yi,t represents the guess of individual iN in round t, and yt is the vector of guesses of all individuals within a group in round t. Let xi denote individual signals, x¯ is the average of the four signals within a group and x is the vector of signals. We assume that in each round agents maximize their one-period expected utility.11

Results

In each of the 6 sessions, 6 groups of 4 subjects carry out the experimental task three times, once for each phase, with 12 rounds for each phase. Overall, the sample includes a total of 5184 observations at the individual level, with 432 observations for each round (36 groups×4 subjects×3 phases). The behavior of individual beliefs over successive rounds indicates substantial heterogeneity at both subject and group level. Overall, in the first round, 92.4% of the subjects report their own

A generalized updating rule: influential listeners

The experimental analysis indicates that, relative to a balanced and symmetric network, increasing the number of outgoing links of one node does not lead to a significant increase in social influence for subjects at this node. On the other hand, subjects at another node, where the number of incoming links is increased, become significantly more influential. One possible interpretation of these findings is that, under a boundedly rational updating rule, social influence within a social network

Conclusions

Humans learn most of what they know from others. Starting from this basic premise, this paper addressed a simple but important question: how are agents' opinions affected by the structure of the network through which they communicate? We presented an experimental investigation of the persuasion bias hypothesis, whereby individuals communicating through a social network are unable to properly account for repetitions of information. Under persuasion bias, social influence ultimately reflects the

Acknowledgments

We thank Matthew Jackson, Benjamin Golub and participants at the 2010 ESA World Meeting in Copenhagen for helpful comments and suggestions.

References (25)

  • A. Banerjee et al.

    Word-of-mouth learning

    Games and Economic Behavior

    (2004)
  • P. Bonacich et al.

    Eigenvector-like measures of centrality for asymmetric relations

    Social Networks

    (2001)
  • B. Celen et al.

    Observational learning under imperfect information

    Games and Economic Behavior

    (2004)
  • D. Gale et al.

    Bayesian learning in social networks

    Games and Economic Behavior

    (2003)
  • D. Acemoglu et al.

    Bayesian learning in social networks

    Review of Economic Studies

    (2011)
  • L. Anderson et al.

    Information cascades in the laboratory

    American Economic Review

    (1997)
  • V. Bala et al.

    Learning from neighbours

    Review of Economic Studies

    (1998)
  • V. Bala et al.

    Conformism and diversity under social learning

    Economic Theory

    (2001)
  • A. Banerjee

    A simple model of herd behavior

    Quarterly Journal of Economics

    (1992)
  • A. Bavelas

    Communication patterns in task-oriented groups

    Journal of the Acoustical Society of America

    (1950)
  • S. Bikhchandani et al.

    A theory of fads, fashion, custom, and cultural change as information cascades

    Journal of Political Economy

    (1992)
  • P. Bonacich

    Power and centrality: a family of measures

    American Journal of Sociology

    (1987)
  • Cited by (56)

    • Conformism, social pressure, and the dynamics of integration

      2024, Journal of Economic Behavior and Organization
    • Dissonance minimization and conversation in social networks

      2023, Journal of Economic Behavior and Organization
    • Misinformation due to asymmetric information sharing

      2023, Journal of Economic Dynamics and Control
    • The origins of influence

      2021, Economic Modelling
    View all citing articles on Scopus
    View full text