Computational models of decision confidence for uni- and multi-dimensional perceptual decisions

Decision confidence plays a crucial role in humans’ capacity to make adaptive decisions in a noisy perceptual world. Often, our perceptual decisions, and the associated confidence judgements, require integrating sensory information from multiple modalities. Empirical investigation of the cognitive processes used for decision confidence under these conditions, however, has been severely limited. To bridge this gap, in this study we investigated the computations used to generate confidence when a decision requires integrating sensory information from both vision and audition and the extent to which these computations are the same when sensory information is solely visual or auditory. Participants (N = 10) completed three versions of a categorisation task with visual, auditory or audio-visual stimuli and made confidence judgements about their category decisions. In each version of the task, we varied both evidence strength, (i.e., the strength of the evidence for a particular category) and sensory uncertainty (i.e., the intensity of the sensory signal). We evaluated several classes of models which formalise the mapping of evidence strength and sensory uncertainty to confidence in different ways: 1) unscaled evidence strength models, 2) scaled evidence strength models, and 3) Bayesian models. Our model comparison approach therefore, provides a compelling specification of the class of algorithms used for decision confidence both when a signal has multiple perceptual dimensions and a single perceptual dimension. Where the signal had multiple perceptual dimensions, we were able to specifically quantify how both evidence strength and sensory uncertainty are integrated across modalities and the extent to which this integration was biased towards a particular modality. Furthermore, by generating predictions from the unidimensional signals and comparing these predictions to behaviour from the multidimensional signals, we determine the extent to which the computations used for decision confidence directly generalise across different decisional contexts.