Figure 1. Experimental framework for dissociating private and public confidence.

On each trial, subjects made a perceptual group decision with one of four partners. They first decided whether a random dot motion stimulus was moving left or right. We varied the fraction of coherently moving dots in order to manipulate subjects’ internal sense of confidence in their decision. Subjects were then informed about their partner on the current trial and were asked to submit a report of confidence in their initial decision (discrete scale from 1 to 6). Subjects were then shown the partner’s response, after which the individual decision made with higher confidence was selected as the group decision. Finally, subjects received feedback about choice accuracy, before continuing to the next trial. We engineered the partners to have the same choice accuracy as subjects but to differ in mean confidence. Subjects were incentivised to help each group achieve as many correct decisions as possible: they were told that we would randomly select two trials for each group in each session (4 × 2 × 2 = 16 trials) and pay £1 in bonus for every correct group decision (in reality, all subjects received £10 in bonus). In this design, the strategy for maximising group accuracy (reward) is to match your partner’s mean confidence. The structure of the task differed between the behavioural and fMRI sessions as explained in the main text.

Figure 1—figure supplement 1. Confidence matching maximises group accuracy and thereby reward.

The heat map shows expected group accuracy as a function of the mean confidence of two players with the same level of task performance (i.e. sensory noise). The heat map was derived analytically using the sensory noise fitted to an example subject and by assuming maximum entropy confidence distributions (see Bang et al., 2017, for details on calculation). The heat map shows that expected group accuracy is highest along the identity line: that is, when a subject’s mean confidence (y-axis) is matched to that of the current partner (x-axis; the four avatars indicate the four partners’ mean confidence as specified in the task). Under our incentive structure, expected reward is proportional to expected group accuracy: the higher the expected group accuracy, the higher the probability that a randomly selected group decision will be correct.

Figure 1—figure supplement 2. Schematic of study protocol.

Subjects took part in separate behavioural and fMRI sessions on the same day. The prescan session involved four phases. In phase 1, we calibrated four levels of coherence so as to achieve target levels of choice accuracy (60%, 70%, 80% and 90%). In phases 2–4, we trained subjects on the social task. In phase 2, subjects were paired with the partners in a block-wise manner (each partner is indicated by a unique colour and name). There were four cycles of blocks of 10 trials per partner (e.g., A–B–C–D–A–B–C–D–A–B–C–D–A–B–C–D). The context screen was shown before each block of trials but not after a perceptual decision. In phase 3, subjects were paired with the four partners in an interleaved manner, with the current partner’s identity revealed after each perceptual decision. In phase 4, the ‘showdown’ stage was played out in the background. In addition, we introduced a condition where the social context was hidden. The scan session involved four runs, using the same design as in phase 4 of the prescan session. We matched the distribution of conditions (coherence × context) across scan runs in order to facilitate multivariate analysis of the fMRI data.

Figure 1—figure supplement 3. Confidence distributions used to generate four partners who differ in mean confidence.

The distributions were constructed so as to have maximum entropy for a given level of mean confidence (see Bang et al., 2017, for details on calculation).