The neural representation of metacognition in preferential decision-making

Abstract

Humans regularly assess the quality of their judgements, which helps them adjust their behaviours. Metacognition is the ability to accurately evaluate one's own judgements, and it is assessed by comparing objective task performance with subjective confidence report in perceptual decisions. However, for preferential decisions, assessing metacognition in preference-based decisions is difficult because it depends on subjective goals rather than the objective criterion. Here, we develop a new index that integrates choice, reaction time, and confidence report to quantify trial-by-trial metacognitive sensitivity in preference judgements. We found that the dorsomedial prefrontal cortex (dmPFC) and the right anterior insular were more activated when participants made bad metacognitive evaluations. Our study suggests a crucial role of the dmPFC-insula network in representing online metacognitive sensitivity in preferential decisions.

Keywords: confidence; metacognition; neural representation; preferential decision‐making.

FIGURE 1

fMRI task and behavioural results (n = 24). (a) fMRI task: participants were asked to choose one image they preferred (within 3 s) from two images (landscape or food) and rate the level of confidence in their decision (within 3 s). (b) Preference consistency for all individual subjects: participants' choices revealed a high transitivity/consistency, with the percentage of consistent trials equalling 91.6% on average. Yellow line, group‐averaged percentage of consistent trials. Further results are reported only for consistent events. (c) The averaged preference distance (chosen vs. unchosen value differences) on four levels of confidence, with larger distances associated with higher confidence. (d) The averaged choice reaction time (RT) as a function of preference distance and confidence. Distance and confidence ratings are divided at the median point. (e) A strong correlation between preference consistency and the correlational coefficient between distance and confidence. Error bars denote SEM.

FIGURE 2

Trial‐by‐trial quantification of metacognitive sensitivity. (a) Overlapping of the density distribution of reaction time (RT) between large and small distance (value difference) conditions. (b) The number of trials under small and large distances with either fast or slow RTs in which distance and RT are divided at the median point. (c) Quantification of metacognition for each event. Distance, RT, and confidence are divided at the median point. Preference choices with large distance/fast RT or small distance/slow RT are defined as match responses (accuracy = 1), and the opposites are mismatch responses (accuracy = −1). Match responses with high confidence and mismatch responses with low confidence are quantified as good metacognitive sensitivity (metacognition = 1), and the opposite combinations are bad metacognition events (metacognition = −1). (d) Areas of type 2 receiver operating characteristics (ROC) under eight conditions are listed in the table above. Error bars denote SEM.

FIGURE 3

(a) Brain activity in dmPFC [peak (x, y, z) = −9, 15, 45] correlating with decreasing chosen–unchosen value distances (p < .05 with small‐volume FEW‐corrected at the peak level). (b) Brain activity in dmPFC [peak (x, y, z) = −9, 15, 42] correlating with decreasing liking distances (p < .05 with small‐volume FEW‐corrected at the peak level). dmPFC, dorsal medial prefrontal cortex. Left images, sagittal view; right images, coronal view.

FIGURE 4

Neural correlates of confidence. (a) Brain activity in vmPFC [peak (x, y, z) = −6, 30, −12] correlating with linear increases in confidence ratings (p < .05 with small‐volume FEW‐corrected at the peak level). (b) Brain activity in ACC [peak (x, y, z) = −9, 39, 24] correlating with the quadratic extension of confidence ratings (p < .05 with small‐volume FEW‐corrected at the peak level). vmPFC, ventral medial prefrontal cortex; ACC, anterior cingulate cortex. Sagittal views.

FIGURE 5

Brain activations in dmPFC [peak (x, y, z) = 3, 27, 51] and anterior insula [peak (x, y, z) = 42, 12, 9] in the contrast between bad and good metacognitive events (p < .05 with FEW‐corrected at the cluster level). Activations were presented with time‐locked to the onset of the stimulus (blue colour) and to the time of response (hot colour) in GLM models. dmPFC, dorsal medial prefrontal cortex; M1, primary motor cortex. Left image, sagittal view; right image, coronal view.