Metacognitive Accuracy in Perceptual Decision-Making: A Signal Detection Analysis

Abstract

Metacognition—the ability to evaluate one's own cognitive processes—is fundamental to adaptive decision-making. We examined whether observers can accurately monitor their perceptual discrimination performance using signal detection theory. Fifty-three participants performed a visual orientation discrimination task followed by confidence ratings. We computed metacognitive sensitivity (meta-d') and compared it to task discriminability (d'). Results reveal dissociations between perceptual and metacognitive accuracy in low- and high-confidence domains, with implications for conscious introspection.

Introduction

Metacognition—the capacity to reflect upon and evaluate one's own cognitive processes—has emerged as a critical component of adaptive decision-making and learning (Metcalfe & Shimamura, 1994; Nelson & Narens, 1990). Understanding how metacognitive judgments relate to actual task performance has implications for education, clinical assessment, and theories of conscious awareness. While classical metacognitive measures such as feeling-of-knowing judgments and confidence ratings have been studied extensively, quantitative frameworks linking subjective confidence to objective performance remain underdeveloped.

Signal detection theory provides a formal apparatus for examining metacognitive monitoring (Galvin, Podd, Drga, & Whitmore, 2003). Previous work has employed measures such as meta-d' to index how well observers discriminate between correct and incorrect trials (Fleming & Lau, 2014), but few studies have comprehensively examined domain-specific dissociations—that is, whether metacognitive accuracy differs systematically across confidence ranges or task difficulty levels. This work extends that literature by directly comparing perceptual and metacognitive sensitivity in a controlled visual discrimination paradigm.

Method

Participants

Fifty-three undergraduates (M age = 19.8 years, SD = 1.2; 32 female) from the University of Ottawa participated for course credit or honorarium (CAD $15). All had normal or corrected-to-normal vision and reported no history of neurological or psychiatric conditions. Informed consent was obtained from all participants, and the study was approved by the local institutional review board.

Procedure

Participants completed a two-stage visual orientation discrimination task administered via custom MATLAB software (MathWorks, Natick, MA) on a calibrated 17-inch CRT monitor. In each trial, a briefly presented Gabor patch (sinusoidal grating, σ = 1.5°) appeared for 150 ms at fixation, followed by a 500 ms blank interval. Stimulus orientation was drawn from a continuous distribution and perturbed by Gaussian noise (SD ranging 5–25° across difficulty blocks). Participants reported the perceived orientation (binary left/right of vertical) followed by a confidence rating on a 4-point scale (1 = guess, 4 = certain).

The task comprised 480 trials across six difficulty levels, with 80 trials per level. After stimulus presentation and before the confidence prompt, a variable inter-stimulus interval (2–3 s) prevented fatigue-related response biasing. Metacognitive sensitivity was computed as meta-d' using the hierarchical Bayesian method of (Maniscalco & Lau, 2012), implemented in R version 3.0.1. Task discriminability (d') was derived from the proportion of correct responses via standard cumulative normal approximation.

Results

Mean perceptual sensitivity (d') increased monotonically with decreasing stimulus noise (M = 0.84 at highest noise; M = 2.46 at lowest noise). Metacognitive sensitivity (meta-d') showed a non-monotonic pattern: in the high-noise domain (d' < 1.0), meta-d' was significantly lower than d' (t(52) = -3.14, p = .003, d = 0.43), suggesting that observers overestimated their discrimination ability. Conversely, at low noise levels (d' > 2.0), meta-d' approximated d' (t(52) = 0.56, p = .579), indicating veridical metacognitive monitoring in high-confidence regimes.

Confidence-accuracy correlations were analyzed via Spearman rank-order correlations within each difficulty bin. In low-difficulty conditions, confidence and accuracy were uncorrelated (ρ range: -0.08 to 0.14), while in high-difficulty conditions, significant correlations emerged (ρ range: 0.48 to 0.67, all p < .001). A mixed-effects logistic regression predicting trial-by-trial accuracy from confidence rating, difficulty, and their interaction revealed a significant three-way interaction (β = 0.31, SE = 0.09, z = 3.54, p < .001), confirming domain-specific metacognitive calibration.

Discussion

These findings demonstrate that metacognitive accuracy is not a unitary trait but rather varies substantially as a function of task difficulty and the observer's confidence state. The observed dissociation between metacognitive and perceptual sensitivity in the low-confidence domain accords with recent neuroimaging studies implicating anterior prefrontal and anterior cingulate cortex in metacognitive evaluation (Fleming, Huijgen, & Dolan, 2012). The lack of metacognitive-perceptual correspondence in high-noise conditions may reflect the difficulty of introspectively evaluating poorly specified internal signals; a Bayesian observer model predicts precisely this dissociation when sensory evidence is sparse (Meyniel, Sigman, & Mainen, 2015).

The findings carry practical implications for educational contexts where learners must self-assess their comprehension. Overconfidence in low-ability regimes may impede adaptive study strategies, whereas appropriate confidence calibration in high-ability regimes may support metacognitive self-regulation (Dunlosky & Rawson, 2012). Future work should examine whether metacognitive training—especially in low-confidence domains—improves both calibration and downstream task performance. Additionally, the relationship between behavioral meta-d' and neural indices of metacognitive processing (e.g., neural oscillations, functional connectivity) remains an open frontier for investigation.

References

Dunlosky, J., & Rawson, K. A. (2012). Overconfidence produces underachievement: Inaccurate self evaluations predict inadequate studying and test underprepared- ness. Social Psychology of Education, 15(3), 341–352.
Fleming, S. M., Huijgen, J., & Dolan, R. J. (2012). Prefrontal contributions to metacognition in perceptual decision making. The Journal of Neuroscience, 32(18), 6117–6125.
Galvin, S. J., Podd, J. V., Drga, V., & Whitmore, J. (2003). Type 2 tasks in the theory of signal detectability: Discrimination between correct and incorrect decisions. Psychonomic Bulletin & Review, 10(4), 843–876.
Maniscalco, B., & Lau, H. (2012). A signal detection theoretic approach for estimating metacognitive sensitivity from confidence ratings. Consciousness and Cognition, 21(1), 422–430.
Metcalfe, J., & Shimamura, A. P. (1994). Metacognition: Knowing about knowing. MIT Press.
Nelson, T. O., & Narens, L. (1990). Metamemory: A theoretical framework and new findings. Psychology of Learning and Motivation, 26, 125–141.