On confidence in individual and group decision-making

Thesis

This thesis is about the human ability to share and combine representations of the uncertainty associated with individual beliefs – an ability which is called metacognition and facilitates effective cooperation. We distinguish between two metacognitive representations: an implicit confidence variable for oneself and an explicit confidence report for sharing with others. Using visual psychophysics and computational modelling, we address the issues of optimality and flexibility in the formation and the utilisation of these representations. We show that people can compute the confidence variable in an optimal manner (the probability that a given belief is correct as per Bayesian inference). Further, we show that the mapping of this variable onto a confidence report can vary flexibly – with people adjusting their reports according to the history of reports given and feedback obtained. This optimality and flexibility is important for effective cooperation. Being a probability, the optimal confidence variable can be compared across people. However, to facilitate this comparison, people must adapt their confidence reports to each other and develop a common metric for reporting the probability that their belief is correct. We show that people solve this communication problem sub-optimally; they match each other’s mean confidence and confidence distributions, regardless of whether they are equally likely to be correct or not. In addition, we show that, while people can take into account differences in underlying competence to some extent, they fail to do so adequately; they exhibit an equality bias, weighting their partner’s beliefs as if they were as good or as bad as their own, regardless of true differences in their underlying competence. More generally, our results pose a problem for our current understanding of metacognition which assumes that confidence reports are stable over time. In addition, our results show that confidence reports are socially malleable, and thus raise the possibility that well-known biases, such as overconfidence, might reflect particular norms for social interaction.

Authors: Bang, D

• Type of award: DPhil
• Level of award: Doctoral
• Awarding institution: University of Oxford
• Language: English