Disentangling the origins of confidence in speeded perceptual judgments through multimodal imaging

Abstract

The human capacity to compute the likelihood that a decision is correct-known as metacognition-has proven difficult to study in isolation as it usually cooccurs with decision making. Here, we isolated postdecisional from decisional contributions to metacognition by analyzing neural correlates of confidence with multimodal imaging. Healthy volunteers reported their confidence in the accuracy of decisions they made or decisions they observed. We found better metacognitive performance for committed vs. observed decisions, indicating that committing to a decision may improve confidence. Relying on concurrent electroencephalography and hemodynamic recordings, we found a common correlate of confidence following committed and observed decisions in the inferior frontal gyrus and a dissociation in the anterior prefrontal cortex and anterior insula. We discuss these results in light of decisional and postdecisional accounts of confidence and propose a computational model of confidence in which metacognitive performance naturally improves when evidence accumulation is constrained upon committing a decision.

Keywords: EEG; confidence; error monitoring; fMRI; metacognition.

Fig. 1.

Experimental paradigm and behavioral results. (A) Experimental paradigm: A participant lying in the fMRI bore equipped with an EEG cap performs (active condition in red) or observes (observation condition in blue) the first-order task and subsequently reports confidence in the committed or observed decision using a visual analog scale. (B) Mixed-effects logistic regression between first-order accuracy and confidence in the active (red) and observation condition (blue). The histograms represent the distributions of confidence for correct (Top) and incorrect (Bottom) first-order responses. (Right) Individual slopes of the mixed-effects logistic regression indicating metacognitive performance. (C) Mixed-effects linear regression between first-order RT and confidence for correct (in green) and incorrect (in red) trials in the active (Left) and observation condition (Right). The histograms represent the distributions of RT and confidence for correct and incorrect first-order responses. Rightmost: Interaction term between first-order accuracy and confidence for RT in the active compared to observation condition. Shaded areas represent 95% confidence intervals.

Fig. 2.

Bounded evidence accumulation model for confidence. (A, Upper) An example trial for which the participant made a first-order error. The violet and blue traces represent accumulators that are incongruent and congruent with a correct response, respectively. A committed first-order decision (D) is taken when the winning accumulator hits the decision bound (dashed horizontal line). Here, the violet trace wins, producing a first-order error. Confidence is assumed to be based on the state of the accumulator corresponding to the first-order choice at the end of the postdecisional period. Confidence in the observed response is based on the state of the accumulator corresponding to the covert decision (cD) at the end of the postdecisional period, except that evidence is “inverted” if the decision cD is incongruent with the observed decision (cD ≠ oD). In both plots, the sigmoid (square box) constrains the result to the [0,100] % interval. t_pd is the postdecisional time. (B) Histogram of the confidence ratings obtained during the experiments, compared to the model simulations (thick line) for error (red) and correct (green). (Upper) Plot for the active condition (second-order model). (Lower) Plot for the observation condition (nondecisional model). Error bars and shaded area represent 95% CIs across subjects. (C, Left) Mixed logistic regression between simulated first-order accuracy and simulated confidence, in the active (red) and observation condition (blue). (C, Right) Individual slopes of the mixed regression model indicating metacognitive performance (see Fig. 1B for the actual behavioral results).

Fig. 3.

EEG-informed correlates of confidence. (A) Event-related potentials time-locked to the first-order response (resp.) are shown for the active condition (Left) and observation condition (Right) for the CPz and FCz sensors. For illustrative purposes, epochs were binned according to three levels of reported confidence: sure error (0 to 33% confidence), unsure (34 to 66% confidence), and sure correct (67 to 100% confidence), although statistics were computed with raw confidence values using mixed-effects linear regression. The shaded areas represent 95% CI. Regions of significance (P < 0.05, few-corrected) are depicted with a gray line, along with topographic maps of the corresponding F values. (B) Leave-one-out decoding performance over time. The plot shows the amount of variance of the reported confidence explained by the decoder (R²) over time in the active (red trace) and the observation condition (blue trace). The shaded areas represent 95% CI, and the horizontal dashed lines the chance level (P < 0.05, computed via nonparametric permutation tests corrected for multiple comparisons). For each participant and condition, the output of the best decoder within an early and late time window was retrained on the whole dataset and used as a parametric regressor to model the BOLD signal. (C) Brain areas coactivated with low decoded-confidence values in the early (Left) and late time window (Right). All displayed BOLD activations are FWE-corrected (P < 0.05) at the cluster level with a threshold at P < 0.001. Not all brain regions are labeled (SI Appendix, Table S4). The coronal view shows significant differences between the active and the observation condition for the labeled region (AI for the early time window and aPFC for the late time window).