Neurally-informed modelling unravels a single evidence accumulation process for choices and subsequent confidence reports

Abstract

How is the content of our decision making processes translated into subjective confidence? Current computational models designed to account for choice- and confidence-reporting behaviour disagree on two fundamental questions: do choice and confidence reports arise from shared versus distinct evaluation processes and what stopping-rule is applied to evidence sampling for confidence reports? Here, we employed a neurally-informed modelling approach in which competing models were compared both based on their fits to behaviour and their ability to recapitulate human electrophysiological signatures of evidence accumulation. Participants made self-paced confidence reports after indicating perceptual choices on a random dot motion task. Model comparison based on behavioural data alone showed that boundary-based stopping rules for confidence reporting provided superior fits than time-based stopping rules. However, results were inconclusive when comparing models with a single evidence accumulation process dictating choice and confidence reports compared to models invoking distinct processes. Importantly, the single-process model markedly outperformed the other models in its ability to account for the observed neural evidence accumulation patterns, including the fact that signal amplitudes predicted reported confidence several hundred milliseconds before the initial choice had been reported. Our study demonstrates that choice behaviour, subjective confidence, and associated neural decision signals can be jointly explained by a model invoking a single process of evidence accumulation with separate boundaries for the initial choice and subsequent confidence reports. Significance StatementHow does the brain generate representations of subjective confidence from categorical decisions? While some models suggest that confidence arises from a read-out of the same information that determined the initial choice, others propose that confidence is determined by a distinct computational process. Definitively testing these alternative accounts has proven difficult because they often make similar predictions for behavioural data. Here, we employ a novel neurally-informed modelling method that evaluates models based on their ability to reproduce key neurophysiological signatures of choice and confidence in addition to behaviour. Our results indicate that choices and subjective confidence are determined by a shared sensory evidence accumulation process.