Confidence Judgments Reflect the Standard Error of Noisy Evidence Samples Across Domains

Abstract

Confidence judgments play a critical role in guiding behavior by shaping information-seeking, learning, and decision strategies. These functions are most effective when confidence is well calibrated, that is, when subjective uncertainty aligns with the objective uncertainty in the presented evidence. Motivated by this, we investigated how people form confidence judgments from noisy samples of information, and whether they use statistically grounded strategies or rely on heuristics. Participants performed two categorization tasks, one with visual orientation stimuli and one with number stimuli. In each task, participants saw sequentially presented observations and made a decision about the generating category and simultaneously reported their confidence in that decision. We independently manipulated the number of observations and standard deviation of the sample to assess whether confidence reflected an integrated estimate of both sources of statistical uncertainty. Behaviorally, confidence and accuracy both increased with larger sample sizes and lower variability. Furthermore, confidence and accuracy were equivalent in samples matched for standard error, suggesting that participants relied on a statistically grounded strategy. Computational modeling further supported this interpretation: a model that scaled confidence according to the standard error of the sample mean provided the best fit to the data, outperforming more heuristic and Bayesian alternatives. This pattern generalized across the orientation and number tasks, suggesting a domain-general strategy for uncertainty estimation. Together, these findings demonstrate that people use structured, statistically grounded strategies to compute their confidence, supporting well-calibrated decision-making even in the absence of full Bayesian inference.