Bayesian confidence in optimal decisions

Abstract

The drift diffusion model (DDM) provides an excellent account of decisions and response times. It also features the optimal property of tracking the difference in evidence between two options. However, the DDM struggles to account for human confidence reports, because responses are triggered when the difference in evidence reaches a set value, suggesting confidence in all decisions should be equal. Previously considered extensions to the DDM fall short of providing an adequate quantitative account of confidence. Possibly because of this, much confidence research has used non-normative models of the decision mechanism. Motivated by the idea that perceptual decision-making will reflect optimal computation, we consider 9 variants of the DDM. Motivated by the idea that the brain will not duplicate the representation of evidence, in all model variants confidence is read out from the decision mechanism. We compare the models to benchmark results, and make 4 qualitative predictions which we verify in a preregistered study. Modelling confidence on a trial-by-trial basis, we find that a subset of model variants provide an excellent account of the precise quantitative effects observed in confidence data. Specifically, models in which confidence reflects a miscalibrated Bayesian readout perform best. These results support the claim that confidence is based on the decision mechanism, which is itself optimal. Therefore, there is no need to abandon the idea that the implementation of perceptual decision-making will reflect optimal computation.

Stage 2 Registered Report: Metacognitive asymmetries in visual perception

“…Across experiments, asymmetry in mean confidence (Hypothesis 1) and in response time (Hypothesis 4) were mostly aligned. This is consistent with previous reports of a negative correlation between response times and confidence across trials within participants (Calder-Travis et al, 2020;Henmon, 1911;Moran et al, 2015;Pleskac & Busemeyer, 2010). To test if this was the case across participants too, and not only across experiments, we fitted a mixed-effects regression model to data from all seven experiments with experiment as a random effect ( ∼ + (1 + | )).…”

“…First, detection 'yes' responses are faster than detection 'no' responses (Mazor et al, 2020). And second, when participants are not under strict time pressure, reaction time inversely scales with confidence (Calder-Travis, Charles, Bogacz, & Yeung, 2020;Henmon, 1911;Moran, Teodorescu, & Usher, 2015;Pleskac & Busemeyer, 2010). Based on these findings, if 1 and 2 responses are similar to detection 'yes' and 'no' responses not only in explicit confidence judgments, but also in response times, we should also expect a response time difference for these stimulus pairs.…”

Metacognitive asymmetries in visual perception

“…First, detection “yes” responses are faster than detection “no” responses ( Mazor et al 2020 ). And second, when participants are not under strict time pressure, reaction time inversely scales with confidence ( Henmon 1911 ; Pleskac and Busemeyer 2010 ; Calder-Travis et al 2020 ). Based on these findings, if and responses are similar to detection “yes” and “no” responses not only in explicit confidence judgments, but also in response times, we should also expect a response time difference for these stimulus pairs.…”

Filling the gaps: Cognitive control as a critical lens for understanding mechanisms of value-based decision-making

“…While only signals preceding the initial decision can be used to trigger such adaptation of the ongoing response, monitoring continues beyond this initial decision and dynamically updates estimates of the goodness of the decision as novel information becomes available (Fig. 5A) (Calder-Travis, Charles, Bogacz, & Yeung, 2020;Charles & Yeung, 2019;Desender, Donner, & Verguts, 2021;Shapiro & Grafton, 2020;Steinhauser & Yeung, 2010, 2012Ullsperger et al, 2014;Yeung, Botvinick, & Cohen, 2004;Yeung & Cohen, 2006;Yeung & Nieuwenhuis, 2009;Yeung & Summerfield, 2012).…”