Biased Opinions: An Experiment in How We Use Estimates of Other People

Abstract

We propose an experiment to test how subjects use information, in the form of a probability estimate, when this comes from another subject who has a well-known bias. Subjects are asked to estimate the probability of drawing a red ball from an urn which has two components: a fixed component which has a known distribution of balls (Y), and a state component (X), which with probability p has distribution A and probability 1-p has distribution B, both of which are known. In the first stage of the experiment subjects have access to a sample before they make their estimate. In the second stage they have no sample but instead receive the estimate of another subject (the Sender) who has seen a sample and was subject to the same state (either A or B). Hence, his estimate is potentially informative about the state. We operationalise the idea of a bias as a difference in the fixed components (Y) of the two subjects. Hence, Receivers and Senders are subject to the same state but may have different prior probabilities due to different fixed components. This difference is easily adjusted for and clearly visualised in the experiment. The results show that even though differences are easily adjusted for, Receivers place higher weight on the estimates of more "similar" Senders. In a third stage of the experiment we allow Receivers to choose the Sender, knowing only his fixed component (Y). In this stage, Receivers generally give higher weight to Senders' estimates than in stage 2 but "punish" dis-similarity more severely, even though Senders in this stage come from the exact same subject pool as in stage 2."