Confidence is influenced by evidence accumulation time in dynamical decision models

Abstract

Confidence judgments are closely correlated with response times across a wide range of decision tasks. Sequential sampling models offer two competing explanations for the relationship between confidence and response time: According to some models, confidence is directly influenced by decision time. Other models explain the correlation by linking the computation of subjective confidence to the dynamics of the decision process. In previous model comparisons, drift diffusion-based confidence models, which do not explicitly consider decision time in the computation of confidence, provided superior model fits compared to race models that directly included decision time in the internal computation of confidence. In the present study, we present theoretical support for the assumption that confidence explicitly takes decision time and post-decisional accumulation time into account by investigating the properties of optimal confidence in a dynamical observer model of decision confidence. We therefore propose the dynamical visibility, time, and evidence (dynaViTE) model, an extension of the dynamical weighted evidence and visibility (dynWEV) model. DynaViTE assumes that confidence is not only based on the final amount of accumulated evidence but explicitly includes time in the computation of confidence. Model comparisons using three previously published data sets with different perceptual decision tasks showed a superior fit of dynaViTE, indicating that the relationship between confidence and response time is not only due to the close link in the accumulation process but also to an explicit inclusion of time in the computation of confidence.

statConfR: An R Package for Static Models of Decision Confidence and Metacognition

“…In metacognition research, the most popular measure of metacognitive performance, the meta-d′/d′ method (Maniscalco & Lau, 2012, 2014, implicitly relies on the independent truncated Gaussian model (Rausch et al, 2023). Finally, confidence models have become a flourishing research topic in their own right (Boundy-Singer et al, 2022;Desender et al, 2021;Guggenmos, 2022;Hellmann et al, 2023Hellmann et al, , 2024Pereira et al, 2021;Rausch et al, 2018Rausch et al, , 2020Shekhar & Rahnev, 2021. However, too few studies have empirically compared different confidence models (Rausch et al, 2018(Rausch et al, , 2020(Rausch et al, , 2023Shekhar & Rahnev, 2021, so there is still no consensus about the computational principles underlying confidence judgments (Rahnev et al, 2022).…”

“…An important limitation of the models implemented in statConfR is that the dynamics of the decision process are not taken into account. This is a problem because confidence judgments are related to the dynamics of decision making (Hellmann et al, 2024;Pleskac & Busemeyer, 2010;Rahnev et al, 2020). However, most previously proposed dynamical models of confidence do not include a parameter to represent metacognitive ability.…”

“…However, most previously proposed dynamical models of confidence do not include a parameter to represent metacognitive ability. There is one proposal for a dynamical measure of metacognitive efficiency, the v-ratio (Desender et al, 2022), which is based on two-stage signal detection theory (Pleskac & Busemeyer, 2010), but two-stage signal detection theory has been outperformed by other models in a number of visual discrimination tasks (Hellmann et al, 2023(Hellmann et al, , 2024Shekhar & Rahnev, 2023). Thus, the static confidence models included in statConfR may still be useful for many researchers.…”

dynConfiR: The R package for sequential sampling models of decision confidence

“…Recently, confidence has gained increasing research interest in the field of cognitive computational modeling (e.g., Adler & Ma, 2018;Aitchison et al, 2015;Desender et al, 2021;Hellmann et al, 2023aHellmann et al, , 2023bKiani et al, 2014;Moran et al, 2015;Pleskac & Busemeyer, 2010;Ratcliff & Starns, 2009, 2013Rausch et al, 2018;Zawadzka et al, 2017). Many experimental tasks and everyday decisions include uncertainty, so the decision-maker can not be entirely sure whether their decision was correct.…”

“…However, as static models ignore the dynamics of the decision process, they do not apply to response time data. But confidence is closely related to decision time in many decision tasks (Hellmann et al, 2023a;Kiani et al, 2014;Rahnev et al, 2020;Vickers et al, 1985). In contrast to static models, dynamical models can explain response time distributions and may thus provide insight into the causal relationship between task difficulty, decision time, and confidence.…”