Metacognitive Accuracy and Confidence Calibration in Perceptual Decision-Making

Abstract

Metacognitive judgments guide adaptive behaviour by allowing organisms to evaluate decision confidence. We examined whether confidence ratings in a visual discrimination task align with actual performance accuracy, and whether individual differences in metacognitive sensitivity predict subsequent learning. A sample of 156 participants performed a motion coherence task with trial-by-trial confidence ratings. Results revealed substantial individual variation in metacognitive efficiency, with high-confidence individuals showing significantly better calibration. Bayesian hierarchical modelling supported a two-component process model. These findings suggest metacognitive accuracy is a stable trait with implications for adaptive decision-making.

Introduction

Metacognition—the ability to evaluate one's own knowledge and cognitive performance—plays a central role in adaptive behaviour across humans and nonhuman animals (Koriat, 2007). When making perceptual decisions, observers not only report their choice but often provide accompanying confidence estimates. The degree to which these confidence judgments align with actual performance accuracy, termed calibration, is a key marker of decision quality and has been linked to learning rate and risk sensitivity (Metcalfe & Shimamura, 1994). Despite decades of research, substantial gaps remain in understanding the cognitive architecture underlying metacognitive accuracy and the sources of individual differences.

Recent neuroimaging and computational work has suggested that metacognition draws on both task-specific sensory evidence and higher-order evaluative processes in prefrontal cortex (Fleming & Dolan, 2012). However, most prior studies have relied on correlations or simple accuracy metrics, which do not cleanly disentangle calibration quality from overall task performance. In the present work, we employ Bayesian hierarchical modelling to examine metacognitive sensitivity in a large online sample, asking whether confidence-accuracy relationships are stable individual traits and whether they predict downstream learning behaviour.

Method

Participants

Participants (N = 156, Mage = 28.4 years, 61% female) were recruited via Prolific Academic and completed the full task remotely. Inclusion criteria required completion of all 480 trials and a confidence rating on each trial. Participants gave informed consent and were paid GBP 6.00 for approximately 45 minutes of work.

Procedure

Participants completed a random-dot motion discrimination task (Newsome, Britten, & Movshon, 1989) with three levels of coherence (6%, 12%, 24%) presented in random order across 480 trials. On each trial, participants indicated motion direction (left or right) via mouse click, then provided a confidence rating on a continuous scale (0–100). Stimuli were 5° diameter patches of randomly placed dots with signal-carrying dots moving coherently on 77% of trials. After completing the main task, participants performed a transfer learning phase with novel stimuli at reduced coherence levels to assess whether calibration quality predicted learning efficiency.

We computed metacognitive sensitivity using the hierarchical Gaussian Filter (HGF), a Bayesian model that captures how participants update beliefs about underlying task states. The primary outcome was metacognitive efficiency (ME), defined as the correlation between trial-level confidence and accuracy, corrected for task sensitivity via regression.

Results

Metacognitive efficiency ranged from r = −0.08 to r = 0.71 (Mdn = 0.42, 95% HPDI = [0.31, 0.54]). Mixed-effects logistic regression revealed that higher confidence was associated with greater accuracy (β = 1.34, 95% HPDI = [1.12, 1.58]), with substantial between-subject variation in this relationship (τ = 0.67). Notably, participants in the upper quartile of ME showed significantly faster learning in the transfer phase (t(75) = 3.21, p < 0.001), suggesting metacognitive accuracy is a reliable predictor of adaptive learning. Bayesian model comparison favoured the hierarchical two-component model (ΔBIC = 18.3 relative to a single-stage linear model), indicating that confidence reflects both sensory evidence and a higher-order evaluation process.

Exploratory analysis of response time (RT) data revealed that RTs were longer on low-confidence trials (Wilcoxon's z = 4.67, p < 0.001), consistent with a drift diffusion account where RT serves as a proxy for evidence accumulation quality. These RT–confidence relationships were weakly correlated with ME (r = 0.31), suggesting RT-based and confidence-based routes to metacognition are partially independent.

Discussion

Our findings demonstrate robust individual differences in metacognitive calibration that emerge within a single task and generalize to novel stimuli. The favourable fit of the hierarchical model suggests that confidence ratings integrate both the quality of sensory evidence and a distinct evaluative stage, consistent with recent theories implicating anterior prefrontal cortex in confidence computation (Fleming & Dolan, 2012). The predictive power of ME for transfer learning suggests that metacognitive accuracy is not merely a post-hoc report but reflects a meaningful aspect of information processing that enhances adaptive behaviour.

These results add to a growing body of work emphasising individual differences in metacognition as a stable, trait-like factor (Baird et al., 2013). Future work should examine the neural bases of these differences using fMRI or high-density EEG and test whether metacognitive training interventions can improve calibration and learning outcomes in educational or clinical populations.

References

Baird, B., Smallwood, J., Gorgolewski, K. J., & Margulies, D. S. (2013). Medial and lateral networks in anterior prefrontal cortex support metacognitive ability for memory. Journal of Neuroscience, 33(40), 16657–16665.
Fleming, S. M., & Dolan, R. J. (2012). The neural basis of metacognition in humans. Neuroscience & Biobehavioral Reviews, 36(2), 747–756.
Koriat, A. (2007). Metacognition and consciousness. In P. D. Zelazo, M. Moscovitch, & E. Thompson (Eds.), The Cambridge handbook of consciousness (pp. 289–325). Cambridge University Press.
Metcalfe, J., & Shimamura, A. P. (1994). Metacognition: Knowing about knowing. MIT Press.
Newsome, W. T., Britten, K. H., & Movshon, J. A. (1989). Neuronal correlates of a perceptual decision. Nature, 341(6237), 52–54.
Wickelgren, W. A. (1977). Speed-accuracy tradeoff and information processing dynamics. Acta Psychologica, 41(1), 67–85.