Confidence drives a neural confirmation bias

The study addresses why beliefs can remain entrenched despite contrary evidence, focusing on confirmation bias—the selective incorporation of belief-consistent information. While confirmation bias is well documented behaviorally, its cognitive, computational, and neural mechanisms are unclear, especially in post-decisional processing. Prior work often emphasized complex attitudes (e.g., politics), where motivation and affect confound mechanisms. Recent demonstrations of confirmation bias in low-level perceptual tasks suggest a domain-general shift in evidence integration. The authors hypothesize that confidence in an initial choice modulates post-decision evidence accumulation, promoting confirmatory and suppressing disconfirmatory information, thereby reducing changes of mind. They leverage controlled perceptual decisions to isolate mechanisms and test how confidence shapes post-decisional processing.

Sequential sampling models describe perceptual decisions as noisy accumulation to a bound, supported by parietal and prefrontal populations. Evidence accumulation principles generalize from simple to complex decisions and underlie both choice and confidence. However, mechanisms of post-decision accumulation, and their dependence on prior beliefs and confidence, are less understood. Prior studies of confirmation bias often involve complex beliefs and motivated reasoning, complicating mechanistic inference. Emerging work shows confirmation-like biases in perceptual tasks via selective overweighting of choice-consistent evidence and history biases influenced by confidence. Neural signatures such as EEG centro-parietal positivity track accumulation, and computational accounts link post-choice integration to confidence and changes of mind. This study builds on these literatures to dissociate starting point versus drift-rate mechanisms of confirmation bias and to identify neural correlates using MEG.

Design overview: Across three studies, participants performed a two-stage random dot motion (RDM) task. Each trial presented pre-decision evidence (left/right motion), followed by an initial choice and confidence report, then a second sample of post-decision evidence (same motion direction as pre-decision; i.e., helpful), culminating in a final choice and confidence. The positive-evidence (PE) manipulation dissociated subjective confidence from objective accuracy. Participants: Study 1 (behavioral): n=28 (after exclusions). Study 2 (behavioral optimized for DDM): n=23. Study 3 (MEG): n=25 (after exclusions, including balance criteria for decoding). Ethics approval from UCL REC (#1260-003); informed consent obtained. Stimuli and PE manipulation: RDKs (350 ms) with dots moving coherently both in target and opposite directions; remaining dots random. High positive evidence (HPE): incorrect-direction coherence fixed at 15%, correct-direction coherence staircased to target accuracy. Low positive evidence (LPE): incorrect-direction coherence 5%, correct-direction coherence staircased to match accuracy with HPE. Calibration (180 trials): staircase targeted 60% correct (Study 1) and 71% correct (Studies 2–3). Main task: 360 trials (behavioral) or 352 trials (MEG). After pre-decision stimulus, participants gave initial left/right decision and confidence (7-point 50–100% scale in Studies 1–2; binary high/low in MEG). Post-decision evidence (weak/strong) was always in the same, correct direction as pre-decision evidence; incorrect-direction coherence set to the average of staircased pre-decision values; strong post-decision stimuli derived by multiplying correct-direction coherence by 1.3. Response timing: Study 1 enforced a 300 ms delay before final response. Study 2 allowed immediate response to enable DDM fits to RT distributions. MEG-specific procedure: Responses made with up/down key presses; left/right and high/low confidence mappings randomized per trial and revealed only when responding became possible, decoupling decision from motor preparation. Added 500 ms delays after each stimulus but before mapping reveal to allow abstract decision formation without motor preparation. Incentives: Quadratic Scoring Rule rewarded both initial and final decisions’ accuracy-confidence alignment; performance-based bonuses provided. Exclusions: Predefined criteria included excessive use of same confidence rating (>90%), non-converged staircase (performance <55% or >87.5%), technical issues (MEG triggers), and highly imbalanced responses (>80% one response for decision or confidence in MEG). Behavioral modeling (Study 2): Hierarchical Bayesian drift-diffusion modeling (hDDM) with accuracy coding. Ten models compared via Deviance Information Criterion. Parameters allowed to depend on initial decision (correct=1, incorrect=-1), initial confidence, and their interaction; boundary separation allowed to depend on confidence in all models. Post-decision evidence strength included as a predictor of drift rate. MCMC: 100,000 samples (50,000 burn-in), thinning 25; convergence checked via traces, autocorrelation, and Gelman–Rubin statistics. Neural decoding (Study 3, MEG): Whole-head 273-channel system; data filtered, resampled to 100 Hz. A linear SVM classifier trained at each pre-decision timepoint (10 ms steps; 100 ms windows) on z-scored sensor amplitudes to decode left vs. right choices; classifier reapplied to corresponding post-decision timepoints to yield a continuous decision variable (DV) per trial. The distance to the hyperplane served as graded neural evidence. Within-trial linear fits to DV time series provided intercept (starting point) and slope (drift-rate analogue). For unsigned accumulation strength, slope signs were flipped on trials with leftward motion; same for intercept. Hierarchical regressions linked neural slope/intercept to RT, accuracy, confidence, and to factors of initial decision, confidence, and their interaction. Sensor contribution mapping used repeated training on random 30-sensor subsets. Temporal generalization assessed how pre-decision classifiers generalized across post-decision timepoints; cluster-based permutation tests evaluated confidence effects (p<0.05, corrected).

Behavioral effects of confidence (Study 1): The positive-evidence manipulation selectively increased initial confidence without affecting accuracy or RTs. Confidence increase: mean difference = 0.024, 95% CI [0.008, 0.04], Cohen’s d = 0.21, t(27) = 3.0, p = 0.005. Accuracy unchanged: mean difference = 0.006, 95% CI [-0.022, 0.034], d = 0.02; Bayesian t-test supporting equality: BF01 = 4.61. RT unchanged: mean difference = -0.005 s, 95% CI [-0.029, 0.018], d = -0.04; BF01 = 4.51. Individuals with larger confidence boosts showed stronger reductions in changes of mind (r = -0.69, p < 0.0001), persisting when controlling for accuracy and RT (p = 0.005). DDM results (Study 2): The best model (Model 10) included dependencies of starting point and drift rate on initial confidence, initial decision, and their interaction, with boundary separation dependent on confidence. Starting point ~ 1 + confidence + initial decision + confidence × initial decision; Drift rate ~ 1 + post-decision evidence strength + confidence + initial decision + confidence × initial decision; Boundary separation ~ 1 + confidence. After accounting for main effects, there was an interaction effect of confidence × initial decision on starting point (95% equal-tailed interval = 0.08–0.18) and on drift rate (95% equal-tailed interval = 0.11–0.26). Effects on drift rate were stronger than on starting point. High-confidence correct initial decisions boosted accumulation of confirmatory (veridical) post-decision evidence; high-confidence initial errors reduced accumulation of disconfirmatory evidence, manifesting as lowered drift rate. Neural decoding and accumulation (Study 3, MEG): The neural accumulation slope responded to presented motion direction during the post-decision period (β = 0.07, t(8550) = 6.89, p < 1e-11). Steeper slopes predicted faster RTs (β = -0.007, t(8549) = -2.83, p = 0.005), higher accuracy (β = 0.16, t(8549) = 3.05, p = 0.002), and higher confidence (β = 0.14, t(8549) = 3.53, p = 0.0004). Intercept also related to accuracy (β = 0.10, t(8549) = 2.0, p = 0.045) and confidence (β = 0.12, t(8549) = 3.07, p = 0.002). Sensor contributions were strongest over centro-parietal regions, consistent with known accumulation signals. Critically, neural accumulation slopes showed a main effect of initial decision (β = 0.042, t(8547) = 2.96, p = 0.003) and an interaction with confidence (β = 0.038, t(8547) = 2.64, p = 0.008), indicating that high confidence facilitated accumulation of confirmatory evidence and largely abolished accumulation of disconfirmatory evidence. No significant effects on neural starting point were detected in this slope/intercept analysis (p > 0.05). Temporal generalization revealed earlier reinstatement of a representation of the initial decision in the post-decision phase when confidence was high (cluster p = 0.01, corrected), consistent with a confidence-related shift of the starting point toward the initial decision bound and/or expectation of confirming evidence.

The findings demonstrate that confidence exerts a top-down influence on post-decision evidence processing, selectively enhancing accumulation of choice-consistent evidence while suppressing disconfirmatory evidence, thereby reducing changes of mind. Behavioral modeling dissociated two mechanisms: a likely normative shift in starting point (consistent with continued integration of pre-decision evidence) and a non-normative change in drift rate reflecting confirmation bias. MEG decoding provided convergent neural evidence: confidence strengthened the neural accumulation of confirmatory information and rendered accumulation of disconfirmatory information negligible, while temporal generalization indicated an earlier reinstatement of the initial decision representation under high confidence. By showing these effects in low-level perceptual decisions, the work suggests that confirmation bias can arise from core neural information-processing principles, independent of motivational or social factors that accompany complex decisions. These insights link metacognitive states (confidence) to flexible or inflexible updating, bridging decision neuroscience with research on cognitive biases and potentially informing why highly confident individuals resist counterevidence in broader domains.

This study shows that confidence shapes post-decision evidence accumulation by selectively gating choice-consistent information and diminishing processing of disconfirmatory evidence. Computational modeling and MEG decoding jointly indicate that confidence affects both starting point and, more prominently, drift rate, producing a neural confirmation bias that reduces changes of mind. These results highlight confidence as a central mechanism controlling evidence integration and suggest that metacognitive interventions targeting confidence formation or calibration may help mitigate confirmation bias. Future research could examine how these mechanisms generalize to complex, real-world decisions where motivational and social factors interact with confidence, and whether interventions that shape metacognitive insight can improve openness to disconfirmatory evidence.

The task involved low-level perceptual judgments with post-decision evidence always aligned with the true motion direction, which may limit generalization to real-world contexts where new information can be misleading or noisy. While MEG decoding provided a neural measure of accumulation, the initial slope/intercept analysis could be insensitive to starting point offsets at the onset of post-decision evidence; a temporal generalization analysis was introduced to address this. Additional motivational and social influences, common in higher-stakes decisions, were minimized here and may interact with confidence in more complex settings.