A low-dimensional approximation of optimal confidence.

Area of Science:

• Cognitive Neuroscience
• Computational Psychiatry
• Decision Science

Background:

• Human decision-making involves a sense of confidence, theoretically linked to the probability of correctness based on available data.
• Optimal Bayesian decision theory suggests confidence reflects learned probabilities, but independent learning for all data combinations is computationally intractable.

Purpose of the Study:

• To present a novel, computationally tractable model for estimating decision confidence.
• To account for individual differences, biases, and deviations from optimal confidence judgments.
• To differentiate between evidence-reliability-based and stimulus-independent confidence biases.

Main Methods:

• Developed a low-dimensional approximation model of optimal Bayesian confidence computation.
• Dissociated confidence biases using parameters α (evidence reliability) and β (stimulus-independent bias).
• Empirically validated the model against choice data (accuracy, response time) and confidence ratings.

Main Results:

• The model accurately fits both behavioral choice data and trial-by-trial confidence ratings.
• Empirically validated two novel predictions: confidence changes independent of performance, and distinct bias patterns from parameter manipulation.
• Demonstrated the model's ability to capture individual idiosyncrasies and biases.

Conclusions:

• The proposed model offers a flexible and tractable framework for understanding confidence computation.
• It provides a method to interpret and resolve various forms of confidence biases in human decision-making.
• The model's dissociation of bias types offers new avenues for research in cognitive and clinical neuroscience.