Rumors in social networks
How to cite
charroin, liza; Bloch, Francis; Sarangi, Sudipta. 2021. "Rumors in social networks." AEA RCT Registry. AEARCTR-0008369. https://doi.org/10.1257/rct.8369-1.0
Download: BibTeX · RIS — for reference managers without a browser translator. Zotero and most others pick up this page's citation metadata directly.
* Primary investigator
Abstract
In this online lab-experiment, we study the diffusion of rumors (uncertain information with no clear information on its source) in a social network.
During the experiment, groups of 5 agents are structured in a line network. Some agents (4 out of 5 or 3 out of 5, depending on the treatment) are computerized agents, programmed to behave optimally. The other agents are real participants.
One of the agents in the network receives a signal after a coin flip (Heads or Tails). Heads happens with a probability 0.3, while Tails happens with probability 0.7. This agent can create a message ("The outcome of the coin flip is Heads" or "The outcome of the coin flip is Tails") and transfer it to his direct neighbor(s). The message may then be diffused through the network. More precisely, each agent receiving the message may decide to transfer or block the message. At the end of the period, the group votes for a specific outcome (the outcome implemented by the group depends on a probability function).
There are two types of agents: biased and unbiased agents, with different payoff functions. Biased agents earn points only if the members of the network have implemented thanks to their votes the outcome "Heads", whatever the true outcome of the coin flip. Unbiased agents earn points if the outcome implemented by the group corresponds to the outcome of the coin flip. So while unbiased agents are willing to match the true outcome, biased agents try to enforce "Heads" by influencing the network.
The game is repeated 48 times (varying distribution and location of biased and unbiased agents).
We design two treatments: in the first treatment, only one agents out of five is a real participant, while in the second treatment, two agents out of five are real participants. The objective is to understand the social/human component in the spreading of rumors.
Experimental design
- Online experiment, using the online platform of the University.
- Two treatments: one treatment with one real participant, the second one with two real participants per groups of 5 agents.
- Some agents are biased and others are unbiased. Our real participants are always unbiased.
- In each period, one (computerized) agent receives a signal (Head or Tails, Heads has a probability 0.3) and can then create a (truthful or untruthful) message about the signal received and send it to his direct neighbor(s). The agent(s) who received the message can transfer or block the message. The communication phase ends when every agent received the message or when the message is no longer transferred.
Information: agents do not know the identity of the source of the message but they know where biased/unbiased agents are located in the network (they know their own position too) and they know from which part of the network the message comes from. Thus, they can infer the probability of the veracity of the message.
- At the end of the period, agents vote for their preferred outcome. Based on a probability function, the outcome implemented by the group is either Heads or Tails; the more agents vote for Heads (resp. Tails), the higher the likelihood of Heads (resp. Tails).
- Unbiased agents earn 6 points if the outcome implemented by the group matches the signal (0 point otherwise), while biased agents earn 6 points if the outcome implemented by the group is Heads (0 point otherwise).
- Process of voting: in order to avoid participants beliefs to differ from their transfer/block decisions, the decision to transfer a Heads-message corresponds to a Heads-vote (same for Tails), while the decision to block a Heads-message (resp. Tails-message) corresponds to a Tails-vote (resp. Heads-message).
- In each treatment, we test 48 different networks (varying the number and position of biased/unbiased agents) corresponding to 48 different periods. Our goal is to test whether participants are able to diffuse and infer the veracity of messages with different possible networks.
Design details
Cited by
No citation data available for this registration yet.