To follow the herd or break away? Overconfidence and Social Learning
We study the effects of overconfidence in a sequential social learning setting. In a lab experiment, we let subjects form beliefs about their own and others' quality of information by tying the accuracy of their signal to their score on a trivia quiz. Their beliefs about the expected scores allow us to measure and study the effects of confidence on social learning. Our results show two distinct effects of confidence manifesting in their behavior of breaking herds. First, subjects that exhibit more confidence about their relative quiz performance are more likely to follow their signal than the herd. Second, subjects who realize that their absolute performance is better than expected are also more likely to follow their signal. The relative overconfident subjects are likelier to benefit from following their signals in easy quizzes. In contrast, absolute underconfident subjects are more likely to benefit in hard quizzes after knowing their score. These findings can be partially explained by a model of social learning where rational agents have information structures that induce overconfidence about their relative signal accuracy.
Anchored Belief Updating from Recommendations
We study a belief updating behavior in a framework where information is presented as a recommendation from a menu of actions. We introduce a property on belief updating called order independence of recommendations, which is analogous to the Bayes' Rule property of path independence of signals. We show that order independence and the properties that characterize the contraction rule of Ke et al. (2021) lead to an impossibility result on the general domain of recommendations. We then show that such a rule exists if and only if the domain is substantially restricted. Lastly, we propose the anchored contraction rule, which satisfies order independence on the general domain. The anchored contraction rule gives a theoretical ground for the asymmetric confirmation bias in the general domain of recommendations that a belief is updated if and only if information supports an ideal point.
Overinference from Weak and Correlated Signals
We investigate overinference, a bias in processing noisy signals wherein the decision maker treats the signal as more informative. We focus on two distinct sources of overinference: overweighting and correlation neglect. Overweighting is the excessive response from the accuracy of a signal, while correlation neglect is the failure to fully account for the correlation of the signals. We design an experiment on a simple belief-updating task to document the interaction of the two sources of overinference in an environment with conditionally independent and correlated signals. We provide a table of simulated data as evidence for exhibiting one bias and observe the change in their updating behavior for both types of signals. This design allows us to determine if the bias are related or orthogonal to each other. We find that there is prevalent overweighting and correlation neglect. Furthermore, we find that providing evidence for exhibiting one bias reduces the other.