Forming global estimates of self-performance from local confidence

Abstract

Metacognition, the ability to internally evaluate our own cognitive performance, is particularly useful since many real-life decisions lack immediate feedback. While most previous studies have focused on the construction of confidence at the level of single decisions, little is known about the formation of "global" self-performance estimates (SPEs) aggregated from multiple decisions. Here, we compare the formation of SPEs in the presence and absence of feedback, testing a hypothesis that local decision confidence supports the formation of SPEs when feedback is unavailable. We reveal that humans pervasively underestimate their performance in the absence of feedback, compared to a condition with full feedback, despite objective performance being unaffected. We find that fluctuations in confidence contribute to global SPEs over and above objective accuracy and reaction times. Our findings create a bridge between a computation of local confidence and global SPEs, and support a functional role for confidence in higher-order behavioral control.

Fig. 1

Experimental design, Experiment 1. a Learning blocks were composed of randomly alternating trials from two “tasks”. Each task was either easy or difficult and provided feedback or no feedback (b), resulting in six possible task pairings (a). At the end of each learning block, subjects were asked to choose which task should be used to calculate a monetary bonus based on their performance at the chosen task. They were also asked to rate their overall ability at each task on a continuous rating scale. A new block ensued with two new color cues indicating two new tasks. c Trial structure. Each trial consisted of a perceptual judgment as to which of two boxes contained a higher number of dots. Each judgment was either easy or difficult according to the dot difference between the left and right boxes. Following their response, subjects either received veridical feedback (correct, incorrect) about their perceptual judgment or no feedback

Fig. 2

Behavioral dissociation between objective performance and SPEs in Experiment 1 (N = 29). a Performance (mean percent correct) was better for easy than difficult tasks, but was not different in tasks with and without feedback. Global self-performance estimates (SPEs) as measured directly by task choice (b) or indirectly via task ratings (c) were higher in the presence than in the absence of feedback, despite objective performance being unchanged. Error bars represent S.E.M. across subjects. Black dashes are individual data points (there are fewer task choice data points due to the limited number of blocks per subject)

Fig. 3

Effects of task difficulty and feedback on global SPEs in Experiment 1 (N = 29). a, b Task choice frequency (a) and task ability ratings (b) were visualized for the six task pairings. a Task choice frequencies could only take on the values 0, 0.5, or 1 due to the limited repetitions of pairing types per subject; pie charts display the fractions of subjects for whom these values were 0, 0.5 or 1 (for the right-hand bar of each plot). b Black dashes are individual data points. c Chosen tasks (Ch.) were rated more highly than unchosen tasks (Unch.), indicating consistency across our two measures of SPEs. ^***p < 0.000001, paired t test. Error bars represent S.E.M. across subjects

Fig. 4

Effects of performance and learning duration on global SPEs in Experiment 2 (N = 29). a Experimental design. Block duration varied from 2 to 10 trials per task. b Task choice frequency in the six types of learning blocks. The central circle of each subplot represents the average task choice frequency over all blocks [“A”]. The left and right circles display the same data split into blocks with smaller [“S”] and larger [“L”’] difference in objective performance between both tasks, indicating that fluctuations in local performance influenced SPEs over and above objective difficulty. ~p = 0.076, ***p < 0.0005 indicate the significance of the regression coefficient regarding the effect of the difference in task performance on task choice. c Task choice frequency as a function of block duration for the six task pairings. **p < 0.01 (resp. NS) denotes whether block duration had a significant (resp. not significant) influence on task choice (statistical significance of the logistic regression coefficient, see Methods). Error bars indicate S.E.M. across subjects. See also Supplementary Fig. 2

Fig. 5

Relating local confidence to global SPEs in Experiment 3 (N = 46). a Subjects rated their confidence higher when they were correct than incorrect, and this difference was greater when the task was easier, indicating that local confidence ratings were sensitive to changes in objective performance. Error bars indicate S.E.M. across subjects. b Logistic regression indicating that the difference in confidence level between tasks (confDiff) explained subjects’ task choices over and above a difference in accuracy (accDiff) and RTs (rtDiff). Error bars indicate S.E. over regression coefficients. ***p < 0.0001 indicates statistical significance of regression coefficient. c Fluctuations in local confidence are predictive of task choices made at the end of blocks. When the difference in mean trial-by-trial confidence between both tasks on the block was larger (“L”) (respectively smaller, “S”), the difference in task choice was larger (resp. smaller). Error bars indicate S.E.M. across subjects. d Between-subjects scatterplot revealing that subjects with higher metacognitive efficiency were better at selecting the easiest task in end-of-block task choices. Purple dots are subjects’ data, dotted lines are 95% CI. Note that task choice data are clustered in discrete levels due to a limited number of blocks per subject. See also Supplementary Fig. 3 and 5