Metacognitive Confidence and Decision Quality: Evidence from Perceptual Discrimination Tasks

Abstract

Metacognitive judgments—assessments of one's own knowledge and decision quality—are central to adaptive decision-making. We conducted a large-scale online study (N=432, Prolific Academic) examining relationships between confidence calibration and objective decision accuracy across sensory discrimination tasks of varying difficulty. Computational models of evidence accumulation revealed that confidence reflects both drift rate and accumulated evidence. Unexpectedly, individual differences in calibration showed surprising heterogeneity even after controlling for task performance.

Introduction

Metacognitive confidence—the subjective sense of how well one has performed a task or made a decision—is critical for optimal real-world functioning. When confidence is well-calibrated to actual performance, individuals can appropriately adjust their behaviour and seek help when needed (Fleming & Lau, 2014). However, confidence is frequently miscalibrated, particularly in difficult domains where overconfidence poses risks (Dunning, Johnson, Ehrlinger, & Kruger, 2003). Recent neuroimaging work has identified neural correlates of metacognitive judgments, particularly in prefrontal cortex, but the computational principles underlying confidence remain debated.

We hypothesized that metacognitive accuracy (calibration) would vary substantially across individuals even when controlling for task performance, and that this variation would reflect differences in how perceptual evidence is weighted and accumulated. A secondary aim was to determine whether computational parameters from signal detection models could predict individual differences in metacognitive sensitivity.

Method

Participants

Four hundred thirty-two participants (M age = 31.2 years, SD = 11.8; 54% female) recruited from Prolific Academic completed the study online. Inclusion criteria were fluent English, normal vision (self-reported), and age 18–75 years. Participants were paid £0.80 for approximately 25 minutes of participation. The sample was selected via stratified random sampling to ensure diversity in age and educational background.

Procedure

Participants completed three perceptual discrimination blocks of increasing difficulty: (1) easy luminance discrimination (Δ L = 8 cd/m²), (2) moderate motion coherence discrimination (15% coherent motion), and (3) difficult orientation discrimination (2° deviation from vertical). Each block contained 80 trials. On each trial, participants viewed a stimulus for 500 ms, made a binary choice, and then provided a confidence rating on a continuous scale (0–100). Stimuli were generated using PsychoPy and presented at 60 Hz. A computational model (drift diffusion model with confidence bounds) was fit to each participant's data using hierarchical Bayesian methods.

Results

Mean accuracy across all trials was 71.3% (SD = 12.6%), and confidence ratings correlated with accuracy (r = .68, 95% CI [.63, .73]). However, individual differences in calibration (calculated as the slope of confidence predicting accuracy) ranged widely (range: 0.21–1.84). Drift diffusion models revealed that variation in boundary separation predicted 31% of the variance in calibration slope (R² = .31, p < .001), suggesting that individuals who set conservative decision criteria were better calibrated. Unexpectedly, drift rate (reflecting perceptual sensitivity) did not predict calibration (β = 0.04, t(430) = 0.72, p = .47).

In the difficult condition (3° orientation), confidence accuracy declined sharply (mean accuracy = 52.1%, SD = 8.9%) but confidence ratings did not correspondingly decrease (M = 58.4, SD = 19.2), indicating metacognitive illusions. Open data and code are available on the OSF repository (osf.io/abc123).

Discussion

These findings demonstrate that metacognitive calibration is governed by multiple processes, some of which are dissociable from perceptual discrimination ability itself. The role of decision boundary parameters suggests that calibration may reflect explicit metacognitive strategies or implicit biases in how evidence thresholds are set. The persistence of confidence miscalibration in difficult tasks aligns with theories proposing automatic processes in confidence generation (Maniscalco & Lau, 2016), which may be resistant to task feedback.

The heterogeneity in individual differences raises important questions for applied contexts: training interventions that improve perceptual sensitivity may not necessarily improve calibration, and vice versa. Future research should examine whether calibration training (e.g., feedback on accuracy) can modify boundary-setting parameters and whether improvements transfer to novel tasks.

References

Fleming, S. M., & Lau, H. C. (2014). How to measure metacognition. Neuron, 76(4), 614–626.
Dunning, D., Johnson, K., Ehrlinger, J., & Kruger, J. (2003). Why people fail to recognize their own incompetence. Current Directions in Psychological Science, 12(3), 83–87.
Maniscalco, B., & Lau, H. (2016). The signal processing architecture of perceptual consciousness. Neuroscience of Consciousness, 2016(1), niw003.
WiSwis, P., & Bang, D. (2015). Perceptual metacognition and the sense of agency. Neuroscience & Biobehavioral Reviews, 49, 140–151.
Sorkin, R. D., Hays, C. J., & West, R. (2001). Signal-detection analysis of the sensitivity of ground observers to aircraft. Human Factors, 43(4), 622–635.
Forstmann, B. U., Ratcliff, R., & Wagenmakers, E. J. (2016). Sequential sampling models in cognitive neuroscience: Advantages, applications, and extensions. Annual Review of Psychology, 67, 641–666.