Dynamic expressions of confidence within an evidence accumulation framework

Abstract

19Human observers can reliably report their confidence in the choices they make. An 20 influential framework conceptualizes decision confidence as the probability of a decision 21 being correct, given the choice made and the evidence on which it was based. This 22 framework accounts for three diagnostic signatures of human confidence reports, including 23 an opposite dependence of confidence on evidence strength for correct and error trials. 24However, the framework does not account for the temporal evolution of these signatures, 25because it only describes the transformation of a static evidence representation into choice 26 and the associated confidence. Here, we combine this framework with another influential 27 framework: the temporal accumulation of evidence towards decision bounds. We propose 28 that confidence at any point in time reflects the probability of being correct, given the choice 29and accumulated evidence up until that point. This model predicts a systematic time-30 dependence of all diagnostic signatures of decision confidence, most critically: an increase of 31 the opposite dependence of confidence on evidence strength and choice correctness with 32 time. We tested, and confirmed, these predictions in human subjects performing a random 33 dot motion discrimination task, in which confidence judgments were queried at different 34 points in time. We conclude that confidence reports track the temporal evolution of the 35 probability of being correct. 36 Author summary 37Humans are able to report a sense of confidence in the accuracy of a choice. An 38 influential framework states that confidence reflects the probability that a choice is correct. 39We combined human experimenting with computational modelling and extended this notion 40 in the time domain, thus to formally describe the dynamics of confidence. Both human data 41 and our model show that the sense of confidence depends on the point in time, at which it is 42 queried. We conclude that human confidence reports reflect the dynamics of the probability 43 of a choice being correct. 44

Bayesian confidence in optimal decisions

“…These findings can be reconciled by considering the timing of decision and confidence reports (Calder-Travis et al, 2020;Desender, Donner & Verguts, 2020;Kiani et al, 2014…”

Emotional content reduces the cognitive effort invested in processing the credibility of social (mis)information

“…Crucially, confidence can integrate additional information processing after the initial choice was made (e.g. Boldt & Yeung, 2015;Charles & Yeung, 2019;De Martino et al, 2012;De Martino et al, 2017;Desender et al, 2020;Lebreton et al, 2015;Pleskac & Busemeyer, 2010). This way, confidence evaluations can amplify, mute or even reverse the original social judgment (Petty et al, 2007).…”

Dynamic influences on static measures of metacognition

“…We simulated data from a drift diffusion model with additional post-decisional evidence accumulation (see Figure 1A). Decision confidence was quantified as the probability of being correct given evidence, time and choice 32,36,37 . We simulated data for 100 agents with 1000 observations each; for each agent, a different random value was selected for drift rate, non-decision time, decision boundary and post-decision drift rate.…”

“…For example, building on the drift diffusion model it has been proposed that confidence reflects the probability of a choice being correct given evidence, decision time, and the decision that was made 32–34 . Usually, it is (implicitly) assumed that participants learn the mapping between these variables via experience 32,46 , however the precise learning mechanism remains to be unraveled. In the current work, we proposed v-ratio (reflecting the ratio between post-decision drift rate and drift rate) as such a dynamic measure of metacognitive accuracy (following the observation that post-decision drift rate indexes how accurate confidence judgments are 27,28 ).…”

“…Finally, we note that in the modeling efforts reported here, the duration of post-decisional evidence accumulation was simply taken to be the median of confidence RTs. Most previous modeling efforts have likewise assumed that post-decision processing terminates once confidence is externally queried 28 , and only a few studies have explicitly examined different stopping rules for post-decision processing 46,54 . Given that we still lack a clear mechanistic understanding of how post-decisional processing is terminated, we here decided for this pragmatic implementation.…”