Task design and Speed-pressure effects on behaviour.

a) Random-dot motion task with zero-coherence lead-in, before a 1500ms deadline for an initial left/right choice. Moving dots with the same coherence remain on the screen as post-decision evidence, for either 700ms or 3000ms (blocked design) to induce a post-decision speed-accuracy trade-off; participants must rate their confidence before this deadline. b-k) Behavioural results; blue/red dots show overall means with 95% confidence intervals, black dots show individual-participants’ means, and black stars show the significance of the speed-pressure effect, see text for full statistics (*=p<.05, **=p<.01, ***=p<.001, ****=p<.0001). Speed-pressure significantly decreases initial RT (b), confidence-RT (cRT; d), final-accuracy (e), AUROC2 (f), mean certainty (h) and changes-of-mind (CoM) (i). It did not significantly decrease initial-accuracy (c) or mean confidence (f). j-m) Conditional plots showing sliding window analysis of speed-pressure effects. Within each person and condition, we took a sliding window of 20% of cRTs, from the 1st to the 80th percentile, and plot the mean y-axis value at the mean cRT for that percentile-window, across participants, with shading showing SEM. This measures the change in the y-axis measure over increasing percentiles of cRT, while accounting for individual differences in mean and range of cRTs. Slower cRTs were associated with lower final-accuracy (j), lower confidence (k), lower certainty (l) and more changes-of-mind (m), with steeper slopes in the Speed-pressure condition.

Overview of the Candidate Models.

(left) The initial choice was modelled using a drift diffusion model, where noisy evidence accumulates with a certain drift rate (v) until hitting a collapsing boundary (a; collapse rate controlled by u). Accumulation started at the unbiased point a*z. (middle) In Time-based models, the time of stopping was determined by a mean deadline time (τ) with some variability (στ). In boundary-based models, the time of stopping was determined by collapsing boundaries, controlled by parameters related to confidence boundary heights (a2up and a2down for the Single-Process variant, a2 and z2 for the Distinct-Process variant), and corresponding collapse rates (u2up, u2down). Single-Process models evaluate evidence relative to the initial reference frame used during choice formation (i.e., the unbiased starting point), while Distinct-Process models evaluate that evidence from a new reference frame (i.e., the freely estimated confidence starting point z2). All models included a parameter for the rate of accumulation for confidence (v2). (right) Finally, in all models, evidence was translated into a six-point confidence scale using a metacognitive noise parameter (σmeta) and five confidence criteria (c1 to c5). Non-decision components not displayed (see Methods).

Goodness-of-fit metrics for the four classes of models.

Values are mean scores over 14 participants in 2 conditions.

Model Fits to Behavioural Data Across Accuracy and Speed Conditions.

(A) Distributions of confidence response times (cRT) separate for correct (positive RTs) and incorrect (negative RTs) initial-choices in the Accuracy (left) and Speed (right) conditions. (B) final-accuracy, confidence, certainty, and change-of-mind (CoM) probability as a function of confidence-RT (cRT; smoothed using loess regression). (C) Mean cRT, final-accuracy, metacognitive sensitivity (AUROC2), confidence, certainty, and CoM probability for the Accuracy (A) and Speed (S) conditions. Error bars denote the within-subject 95% CI. Black histograms, lines, and points represent observed data, while coloured densities, lines and crosses represent the various model predictions. CoM=Change-of-mind.

Observed CPP & Simulated Decision Variable, and Regression Analysis.

(A) Certainty Ratings. (B) Change of mind (CoM). (C) Speed Pressure. (D) Time-derived regression analyses. In panels A-C, the CPP is presented in the top rows, the simulated decision variable from the Boundary-Single model in the middle rows and the Boundary-Distinct model in the bottom rows. Left columns show choice-locked waveforms and right columns show confidence-locked waveforms. Choice-locked epochs end 250ms after initial response, which is just longer than the average median confidence-RT (243ms). Topography insets show the effects of each factor on mean observed EEG activity -150:-50ms before the initial choice or the confidence response (for Certainty, the difference is between Maybe and Certain ratings);lack dots show the CPP electrodes. (D) Time-derived Regression analyses, showing the beta-coefficients for Certainty, CoM and Certainty*CoM effects for the model DVs (left-axis) and CPP (right-axis). Solid black bars along the bottom show clusters of significant effects on the observed CPP (single-trial time-derived LME regression, with permutation testing and cluster-control of FWER at .05) and Error bars reflect the within-subject SEM.