The importance of urgency in decision making based on dynamic information

RESEARCH ARTICLE The importance of urgency in decision making based on dynamic information Lorenzo Ferrucci 1 , Aldo GenovesioID 1 *, Encarni MarcosID 1,2 * 1Department of Physiology and Pharmacology, Sapienza University of Rome, Rome, Italy, 2Instituto de Neurociencias de Alicante, Consejo Superior de Investigaciones Cientı´ficas–Universidad Miguel Herna ´ndez de Elche, Sant Joan d’Alacant, Spain *[email protected] (AG); [email protected] (EM) Abstract A standard view in the literature is that decisions are the result of a process that accumulates evidence in favor of each alternative until such accumulation reaches a threshold and a decision is made. However, this view has been recently questioned by an alternative proposal that suggests that, instead of accumulated, evidence is combined with an urgency signal. Both theories have been mathematically formalized and supported by a variety of decisionmaking tasks with constant information. However, recently, tasks with changing information have shown to be more effective to study the dynamics of decision making. Recent research using one of such tasks, the tokens task, has shown that decisions are better described by an urgency mechanism than by an accumulation one. However, the results of that study could depend on a task where all fundamental information was noiseless and always present, favoring a mechanism of non-integration, such as the urgency one. Here, we wanted to address whether the same conclusions were also supported by an experimental paradigm in which sensory evidence was removed shortly after it was provided, making working memory necessary to properly perform the task. Here, we show that, under such condition, participants’ behavior could be explained by an urgency-gating mechanism that low-pass filters the mnemonic information and combines it with an urgency signal that grows with time but not by an accumulation process that integrates the same mnemonic information. Thus, our study supports the idea that, under certain situations with dynamic sensory information, decisions are better explained by an urgency-gating mechanism than by an accumulation one. Author summary Decisions are the result of a deliberative process that assesses the suitability of each potential option. However, the specific dynamics of such process are still under debate, with two influential views. On the one hand, a standard view on perceptual decision making proposes that sensory evidence in favor of each option is sequentially sampled and accumulated until the accumulation reaches a bound and a decision is made. On the other hand, an alternative view proposes that, once sampled, instead of accumulated, sensory PLOS COMPUTATIONAL BIOLOGY PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009455 October 4, 2021 1 / 17 a1111111111 a1111111111 a1111111111 a1111111111 a1111111111 OPEN ACCESS Citation: Ferrucci L, Genovesio A, Marcos E (2021) The importance of urgency in decision making based on dynamic information. PLoS Comput Biol 17(10): e1009455. https://doi.org/10.1371/journal. pcbi.1009455 Editor: Marieke Karlijn van Vugt, University of Groningen, NETHERLANDS Received: December 11, 2020 Accepted: September 15, 2021 Published: October 4, 2021 Copyright: ©2021 Ferrucci et al. This is an open access article distributed under the terms of the Creative Commons Attribution License, which permits unrestricted use, distribution, and reproduction in any medium, provided the original author and source are credited. Data Availability Statement: The experimental data and scripts have been uploaded to the Open Science Framework. The doi generated by this site is: https://OSF.IO/SK8AJ. Funding: EM, AG and LF were supported by Sapienza University of Rome (“Avvio alla Ricerca 2016” and Ateneo 2018). EM was also supported by Spanish Ministry of Science, Innovation and Universities (Juan de la Cierva-incorporacio ´n scholarship, IJCI-2016-27864) and by the Spanish State Research Agency through the Severo Ochoa Program for Centres of Excellence in R&D (SEV- evidence is low-passed filtered and weighted by an urgency signal, which increases with time. Here using a perceptual decision-making task with dynamic sensory evidence and two computational models, related with each view, we wanted to distinguish whether decisions can be better described by the standard proposal or by the alternative one. Our results show that a model with a low-pass filter and an urgency signal can describe the experimental data better than a model of accumulation, even in a situation where sensory evidence sequentially disappears. Our study contributes to describe the decision-making process by providing experimental and computational support for one of the most influential views in decision making. Introduction When making decisions, one needs to predict which option will lead to the best outcome. To do that, information is gathered from all possible sources and weighted according to its reliability. In laboratory studies, this has been investigated using perceptual decision-making tasks requiring sensory evidence discrimination to correctly select between two options [1,2]. Neurons from the frontoparietal network exhibit a ramping activity that seems to mimic the deliberative process of decision making [2–6]. The general agreement is that, during such deliberative process, information is sequentially sampled until a bound is reached and a decision is made. However, how such samples are incorporated into the decision-making process is still open to debate. Here, we address this issue by studying the accuracy of two widely accepted alternative models–the Evidence Accumulation Model and the Urgency Gating Model—to describe experimental data collected from a decision-making task with information that varied over time. In the last decades, two alternative theories have been proposed to explain decision making. The standard view proposes that decisions are the result of accumulating evidence until a threshold is reached. This view has led to the development of the Evidence Accumulation Model (EAM), which has accounted for a variety of behavioral and neuronal data [7–13]. However, recently, this view has been questioned by an alternative theory that proposes that, rather than being accumulated, sensory evidence is weighted by an urgency signal that grows with time. This could also explain the ramping activity of neurons in the decision-making network as well as behavior in different decision-making paradigms [14–17]. Within this view, the Urgency Gating Model (UGM) proposes that evidence is low-pass filtered and multiplied by a temporally increasing signal [15]. Thus, previous research supports both kinds of models. However, in most tasks, decisions relayed on constant information and, although they have proven to be valid to discriminate between the models in some cases [18], in some others, they have proven to be inadequate [15,19]. In the last years, new perceptual decision-making paradigms involving changes of information in the course of a trial have been proposed [15,16,20–24]. In such tasks, perceptual evidence is sequentially presented in favor of one of two options and humans or animals have to decide which one of the two options is the most favored one. Although these tasks have provided a significant advance towards the description of a general mechanism for perceptual decision making, the question of whether the neuronal dynamics during such decisions follow the mechanism proposed by either the EAM or the UGM remains still unanswered. To distinguish between the models, Cisek et al. [15] used a perceptual decision-making task, called the tokens task, in which visual stimulus sequentially jumped towards a right or left target and stayed there until subjects committed to a choice. In these cases, the EAM failed to explain the PLOS COMPUTATIONAL BIOLOGY The importance of urgency PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009455 October 4, 2021 2 / 17 2017-0723). The funders had no role in study design, data collection and analysis, decision to publish, or preparation of the manuscript. Competing interests: The authors have declared that no competing interests exist. experimental data and, instead, the UGM provided a reliable mechanism by which decisions might be made. The UGM model proposes that, rather than integrated, sensory evidence is modulated by an urgency signal that increases over time, reflecting the increasing need to make a decision as time passes [15,16]. Subsequently, by using the same task, Thura et al. [5] showed that the activity of the neurons in the premotor and primary motor cortex reflected the combination of sensory evidence with an urgency signal. However, the fact that novel sensory evidence remained available until the decision was made might have biased their results, since there was no implicit need for integration. In other words, one could assess the situation just before the decision is made and obtain the same information that one would get by observing for the entire period. Thus, there is still no real consensus on whether the ramping activity of the neurons in the frontoparietal network reflects the integration of sensory evidence or is instead the result of a combination of sensory evidence with an urgency signal that increases over time. Here, we further contrast the predictions produced by the two computational models by using a modified version of the tokens task [5,15], introducing an additional condition in which novel sensory evidence is removed from the screen soon after it is provided. In each trial, fifteen tokens, presented in a central circle, sequentially jumped towards a left or right target, indicated by a circle on the screen. Subjects had to guess which target would have more tokens by the end of the trial. They could make their choice at any time. The trials were divided into blocks that contained only “all-stay” or “all-away” trials (Fig 1A). In all-stay trials the tokens stayed visible during the entire trial whereas in the all-away condition they disappeared soon after they jumped into one target (see Materials and Methods). With this new task design, we can test both models under conditions that might favor evidence integration. Results Overall, the subjects performed the task with an accuracy greater than chance in both the allstay and the all-away conditions, although with a significantly higher accuracy in the all-stay trials (73 ±2%) than in the all-away trials (68 ±2%; paired-samples t-test, p = 0.04, t= 2.25). In addition, their decision times (DTs) were slower in the all-stay than in the all-away condition, at 1.465 ±0.074 s and 1.207 ±0.121 s, respectively (paired-samples t-test, p = 0.04, t= 2.25). The subjects’ mean (±standard error of the mean [SEM]) baseline reaction time (RT) used to calculate the DTs was 0.347 ±0.009 s. Behavior is modulated by context We used the easy and ambiguous trials to investigate whether the subjects’ performance was influenced by the type of trial. DTs and success probabilities (SPs) in these two trial types indicated that the subjects behaved differently in easy and ambiguous trials in both the all-stay and all-away conditions (Fig 2A). Specifically, DTs were faster and SPs higher for easy than for ambiguous trials in both conditions (mean DT and SP ±SEM for all-stay: 1.587 ±0.088 s and 57 ±1% for ambiguous trials, 1.228 ±0.058 s and 93 ±1% for easy trials; mean DT and SP ±SEM for all-away: 1.331 ±0.146 s and 55 ±1% for ambiguous trials, 1.049 ±0.095 s and 89 ±2% for easy trials). Fig 2B shows the behavior of one representative subject during ambiguous and easy trials in all-stay and all-away conditions. Consistent with the group’s behavior, this subject responded significantly faster in easy trials than in ambiguous trials in both experimental conditions (all-stay: 1.578 ±0.080 s for ambiguous trials and 1.251 ±0.044 s for easy trials; all-away: 1.440 ±0.082 s for ambiguous trials and 1.069 ±0.034 s for easy trials) and his/ her SPs were significantly higher for easy trials than for ambiguous trials (all-stay: 61 ±2% for PLOS COMPUTATIONAL BIOLOGY The importance of urgency PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009455 October 4, 2021 3 / 17 ambiguous trials and 95 ±1% for easy trials; all-away: 56 ±2% for ambiguous trials and 92 ±1% for easy trials). Next, we investigated the subjects’ performance during bias-against and bias-for trials. These two trial profiles are the most interesting of the study because, for non-leaky sensory evidence (all-stay condition), the decision-making models make different predictions about DTs for these trial types [15]. While the UGM predicts no differences between them, the EAM predicts that DTs in bias-against trials will be longer than those in bias-for trials. Importantly, using only the all-stay condition, Cisek et al. [15] showed that the subjects’ DTs did not differ between bias-against and bias-for trials, providing strong evidence in support of the UGM. However, one possible explanation for their results was that accumulation or integration of evidence was not required by the task because the information was available on the screen during the entire trial, favoring urgency dynamics. Our all-away condition was designed to control for that possibility. By using a condition in which each token disappeared from the screen after jumping, we could control for the possibility that the previous findings in favor of the UGM over the EAM were not merely a consequence of the limitations of the original experimental design. Fig 1. Experimental design. (A) Temporal presentation of events during a trial. Each trial starts with all the tokens in the central circle. After the participant moves the cursor to inside the central circle, the tokens start jumping successively to the other two (target) circles. In the all-stay trials the tokens remain visible after jumping, while in the all-away trials they disappear soon after they have jumped. The participant has to guess which of the two target circles will contain more tokens at the end of the trial. (B) Success probability profiles for specific trial types. Top panel, success probability for easy (black) and ambiguous (gray) trials. Bottom panel, success probability for bias-against (black) and bias-for (gray) trials. https://doi.org/10.1371/journal.pcbi.1009455.g001 PLOS COMPUTATIONAL BIOLOGY The importance of urgency PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009455 October 4, 2021 4 / 17 We, first, examined the results of the all-stay condition trials. Consistent with Cisek et al. [15][23], we observed no differences in the subjects’ mean DT or SP (±SEM) between bias-for and bias-against trials (bias-against: 1.810 ±0.036 s and 85 ±1%; bias-for: 1.845 ±0.043 s and 86 ±2%; Fig 3A). Then, we examined the results of the new condition that we had introduced: the all-away condition. Interestingly, the subjects’ mean DT and SP differed significantly between the two trial types (bias-against: 1.650 ±0.056 s and 80 ±2%; bias-for: 1.826 ±0.053 s and 86 ±1%; Fig 3A). Notably, contrary to the prediction of the EAM for non-leaky sensory Fig 2. Behavior of subjects during easy and ambiguous trials. (A) Left panel, individual mean decision times (DTs) observed during easy and ambiguous trials for all-stay (gray) and all-away (black) conditions. Inset panel shows a histogram with the difference in DTs between trial types within each condition. The DTs for easy and ambiguous trials were significantly different in both conditions (n = 15; paired-samples t-test, �� p<0.001; all-stay, p = 0.000003, t= 7.48; all-away, p = 0.0007, t= 4.32). Right panel, success probability (SP) at decision time for easy and ambiguous trials. Inset panel shows a histogram with the difference in SPs for the two trial types within each condition. The difference is significant for both conditions (n = 15; paired-samples t-test, �� p<10 −8 ; all-stay, p = 0.4 ×10 −12 , t= 25.58; all-away, p = 0.2 ×10 −9 ,t= 16.38). Error bars indicate SEM. (B) DTs (left panels) and SPs (right panels) of a representative subject, whose mean DT and SP values are indicated by arrows in (A). The subject clearly shows the same behavioral effect as was observed for the group: faster DT and higher SP at decision time for easy trials than for ambiguous trials in both all-stay and all-away conditions (Kolmogorov–Smirnov test, �� p<0.01; n easy = 52, n ambiguous = 37, all-stay: DTs, p = 0.00005, D= 0.4; SPs, p = 0, D= 0.89; n easy = 28, n ambiguous = 22, all-away: DTs, p = 0.002, D= 0.35; p = 0, D= 0.97). https://doi.org/10.1371/journal.pcbi.1009455.g002 PLOS COMPUTATIONAL BIOLOGY The importance of urgency PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009455 October 4, 2021 5 / 17 evidence (see text above), the mean DT was shorter in bias-against trials than in bias-for trials. Indeed, the behavioral difference between both types of trials is related to shortened DTs observed in the bias-against trials compared with the same trials in the all-stay condition. The shortened DTs reduces accuracy but not significantly (paired-samples t-test p = 0.25; all-away: 60 ±6%; all-stay: 68 ±6%). Fig 3B shows the DTs and SPs of the same subject represented in Fig 2B. This subject showed no difference in DTs and SPs between bias-against and bias-for trials in the all-stay condition (bias-against: 1.883 ±0.048 s and 87 ±1%; bias-for: Fig 3. Behavior of subjects during bias-against and bias-for trials. (A) Left panel, DTs observed during bias-against and bias-for trials in all-stay (gray) and all-away (black) conditions. Inset panel shows a histogram with the difference in DTs between the two trial types within each condition. The difference in DTs between bias-against and bias-for trials is significant in the all-away condition but not in the all-stay condition (n = 15; paired-samples t-test, �� p<0.01; p = 0.001, t= 3.96). A Bayes Factor of 16.929 (>3) in the all-away condition and of 0.249 (<1/3) in the all-stay condition confirm the results. Right panel, SPs at decision time for bias-against and bias-for trials. Inset panel shows a histogram with the difference between SPs for the two trial types within each condition. The difference is significant only in the all-away condition (n = 15; paired-samples t-test, �p<0.05; p = 0.01, t= 2.78). A Bayes Factor of 4.063 (>3) in the all-away condition and of 0.277 (<1/3) in the all-stay condition support the result. Error bars indicate SEM. (B) DTs and SPs for the same subject as in Fig 2B, whose mean DT and SP values are indicated with arrows in (A). The subject exhibits the same behavioral effect as was observed for the group: the DT and SP only differed significantly between bias-against and bias-for trials in the all-away condition (n bias-against = 33, n bias-for = 24; Kolmogorov–Smirnov test, �� p<0.001; DTs, p = 0.00002, D= 0.62; SPs, p = 0.0001, D= 0.57). https://doi.org/10.1371/journal.pcbi.1009455.g003 PLOS COMPUTATIONAL BIOLOGY The importance of urgency PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009455 October 4, 2021 6 / 17 1.8883 ±0.050 s and 85 ±2%), but did show a significant difference between trial types in the all-away condition (bias-against: 1.546 ±0.030 s and 78 ±1%; bias-for: 1.841 ±0.063 s and 85 ±2%), with longer DTs and higher SPs in bias-for than for bias-against trials. Consistent with previous research [15,16], subjects were biased towards an urgencylike strategy in the all-stay condition. Next, we wondered whether subjects’ behavior in the all-away condition might be influenced by the order in which they performed the two conditions (see Materials and Methods). In other words, did subjects’ behavior in the allaway condition depend on whether they encountered first the all-stay condition than the all-away condition or vice versa? To answer this question, we analyzed DTs and SPs of subjects sorted by whether they belonged to the first or the second group. Mean DT and SP (±SEM) in bias-for and bias-against trials were close to significance when subjects performed the all-away condition after the all-stay condition (n = 8, paired-samples t-test p = 0.051 for DTs and p = 0.164 for SPs; bias-for: 1.73 ±0.06 and 83 ±2%; bias-against: 1.56 ±0.05 s and 78 ±4%) and significantly different when they performed the all-away condition at first (n = 7, paired-samples t-test p <0.05; p = 0.0152, t = 3.36 for DTs and p = 0.0171, t = 3.26 for SPs; bias-for: 1.94 ±0.07 s and 89 ±2%; bias-against: 1.75 ±0.09 s and 83 ±3%). Nevertheless, in both cases, mean DT and mean SP were longer and higher, respectively, in the bias-for than in the bias-against trials, indicating no influence of the order of blocks in behavior. The urgency-gating model correctly predicts behavior We investigated whether the behavioral results could be better explained by the EAM or by the UGM. To do that, we devised a computational framework in which sensory evidence directly fed the decision-making model, implemented as the EAM or the UGM, or did so through a working memory module (Fig 4A), simulating the all-stay and all-away conditions, respectively. Therefore, the working memory was responsible for monitoring and remembering the sensory evidence that disappeared from the screen. To fit the data with the models, we used a Fig 4. Computational framework and model simulations for correct and error trials. (A) Schematic diagram of the complete network that simulates the observed experimental results. Visual evidence is either provided directly to the decision-making module or stored in a working memory that then provides the information (e leaky ) to the decision-making module. The information (e leaky or e) is used by the EAM or the UGM to make a choice. (B) Distributions of DTs in correct and error trials in the all-stay condition when all individual trials are pooled together for real data and simulated EAM and UGM. (C) Same conventions as in (B) for the all-away condition. In this case, simulations were performed with and without leakage in the sensory evidence. https://doi.org/10.1371/journal.pcbi.1009455.g004 PLOS COMPUTATIONAL BIOLOGY The importance of urgency PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009455 October 4, 2021 7 / 17 differential evolution algorithm with the experimental DTs recorded in correct and error easy, ambiguous, bias-for, and bias-against trials (see Materials and Methods). Table 1 shows the mean and SEM of the best fitting parameters that were obtained for each model and condition. In the all-stay condition, DTs in correct and error trials were better fitted by the UGM than by the EAM (Fig 4B), as indicated by a lower difference between mean experimental and simulated DTs for both kinds of trials (EAM: 135 ms for correct trials, 279 ms for error trials; UGM: 51 ms for correct trials, 41 ms for error trials). Moreover, better performance of the UGM over the EAM was observed when the experimental data was fitted in the all-away condition (Fig 4C), with lower difference in mean DTs that held for sensory evidence without and with leakage (EAM without sensory leak: 79 ms for correct trials, 134 ms for error trials; EAM with sensory leak: 72 ms for correct trials, 115 ms for error trials; UGM without sensory leak: 14 ms for correct trials, 12 ms for error trials; UGM with sensory leak: <1 ms for correct trials, 26 ms for error trials). Indeed, in the two experimental conditions, the shapes of the distributions were better estimated by the UGM in all cases, with the EAM tending to predict shorter DTs resulting in positive skew distributions. We then investigated the data obtained with the models for each trial type separately to assess whether the models showed the same effect as observed in the experimental data. As previously shown [15], in the all-stay condition, the EAM correctly produced shorter DTs and higher SPs in easy than ambiguous trials but predicted longer DTs and higher SPs in biasagainst trials compared with bias-for trials (Fig 5A), despite no difference was observed in the experimental data in those kinds of trials. On the contrary, the UGM correctly replicated the same effects in DTs and SPs for the four kinds of trials, with distributions that were very similar to those from the real data (Fig 5A, green square). Next, we looked at the all-away condition in two cases: when the sensory evidence had no leak (L e = 0) and when it leaked away with time (L e >0) (see Materials and Methods). In the simulations without sensory leak, both models failed to correctly reproduce the experimental data and showed the same results as those obtained for the all-stay condition. In other words, both models correctly reproduced the differences in DTs and SPs in easy and ambiguous trials but failed to produce shorter DTs and lower SPs in bias-against than bias-for trials, with an opposite result in the EAM and no difference in the UGM (Fig 5B). However, a leak in the sensory evidence (L e >0) was sufficient for the UGM to explain the experimental data in the four types of trials, but not for the EAM, which predicted no significant difference in DTs between bias-for and bias-against trials Table 1. Mean and SEM across subjects of the best fitting parameters for the EAM and the UGM in the all-stay and all-away conditions with and without leakage in the sensory evidence. νηθL e all-stay DDM Mean 0,04456 0,00478 0,29062 SEM 0,00183 0,00140 0,01289 UGM Mean 3,76432 7,65377 23022,92488 SEM 0,61932 3,35522 1124,45466 all-away without sensory leak DDM Mean 0,03915 0,01345 0,25270 SEM 0,01519 0,01464 0,06876 UGM Mean 4,12036 12,84759 18791,13380 SEM 2,55487 20,17686 6304,48655 all-away with sensory leak DDM Mean 0,06918 0,01751 0,26127 0,12341 SEM 0,00654 0,00319 0,01986 0,04727 UGM Mean 5,05699 13,03403 17660,64882 0,21371 SEM 0,64309 5,08468 1736,69700 0,04929 https://doi.org/10.1371/journal.pcbi.1009455.t001 PLOS COMPUTATIONAL BIOLOGY The importance of urgency PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009455 October 4, 2021 8 / 17 (Fig 5B). In addition, the shapes of the distributions of DTs and SPs in the real and UGM data were comparable. Thus, the UGM is capable of explaining the data in the two experimental conditions of our task. The fitted parameters of the UGM seem to indicate a difference in the strategy used in the two experimental conditions: higher mean drift rate and lower boundaries were estimated in the all-away condition compared to the all-stay condition (Table 1). This result could be related to an increase in the need to make a decision when sensory evidence disappears, related to the gradual “forgetting” of information. Nevertheless, both experimental conditions support the existence of an urgency signal that modulates the decision-making process, strengthening the idea of decisions being the result of sensory evidence (visual or mnemonic) combined with an internal urge to decide. Discussion In this work we have advanced in the understanding of the general mechanism of decision making. To accomplish this, we used an experimental task in which the available sensory information varied over time. Changing information over time appears to be a critical element of task design, since it allows a more efficient discrimination between computational models of decision making than that using tasks with constant information [19], and therefore between Fig 5. Model simulations for different trial types. (A) Distributions of DTs for ambiguous, easy, bias-for, and bias-against trials in real and simulated data in the all-stay condition. Inset panel, Cumulative distributions of SPs for each trial type. Differences in DTs and SPs in ambiguous and easy trials are fitted correctly by the EAM and the UGM (paired-samples t-test, �� p<0.01). Lack of difference in DTs and SPs in bias-for and bias-against trials is only fitted by the UGM (paired-samples t-test, �� p<0.01, ns: not significant). Dashed lines indicate the mean of the distributions. Green rectangle shows the results that are consistent with the real data. (B) Same conventions as in (A) for the all-away condition with simulations performed without and with leakage in the sensory evidence. Differences between DTs and SPs in easy and ambiguous trials are fitted correctly by the EAM and the UGM without and with sensory leakage (paired-samples t-test, �� p<0.01). Longer DTs and higher SPs in bias-for than in bias-against trials are only correctly fitted by the UGM when sensory evidence leaks away (paired-samples t-test, �p<0.05, �� p<0.01, ns: not significant). https://doi.org/10.1371/journal.pcbi.1009455.g005 PLOS COMPUTATIONAL BIOLOGY The importance of urgency PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009455 October 4, 2021 9 / 17 16. Thura D, Beauregard-Racine J, Fradet C-W, Cisek P. Decision making by urgency gating: theory and experimental support. J Neurophysiol. 2012; 108:2912–30. https://doi.org/10.1152/jn.01071.2011 PMID: 22993260 17. Carland M, Marcos E, Thura D, Cisek P. Evidence against perfect integration of sensory information during perceptual decision making. J Neurophysiol. 2016; 115:015–930. https://doi.org/10.1152/jn. 00264.2015 PMID: 26609110 18. Hawkins G, Wagenmakers E-J, Ratclif R, Brown S. Discriminating evidence accumulation from urgency signals in speeded decision making. J Neurophysiol. 2015; 114:40–7. https://doi.org/10.1152/jn.00088. 2015 PMID: 25904706 19. Trueblood J, Heathcote A, Evans N, Holmes W. Urgency, leakage, and the relative nature of information processing in decision-making. Psychol Rev. 2021; 128:160–86. https://doi.org/10.1037/rev0000255 PMID: 32852976 20. Brunton BW, Botvinick MM, Brody CD. Rats and humans can optimally accumulate evidence for decision-making. Science. 2013; 340:95–8. https://doi.org/10.1126/science.1233912 PMID: 23559254 21. Piet AT, El Hady A, Brody CD. Rats adopt the optimal timescale for evidence integration in a dynamic environment. Nat Commun. 2018; 9:4265. https://doi.org/10.1038/s41467-018-06561-y PMID: 30323280 22. Hanks TD, Kopec CD, Brunton BW, Duan CA, Erlich JC, Brody CD. Distinct relationships of parietal and prefrontal cortices to evidence accumulation. Nature. 2015; 520:220–3. https://doi.org/10.1038/ nature14066 PMID: 25600270 23. Evans N, Hawkins G, Boehm U, Wagenmakers E-J, Brown S. The computations that support simple decision-making: A comparison between the diffusion and urgency-gating models. Sci Rep. 2017; 7:16433. https://doi.org/10.1038/s41598-017-16694-7 PMID: 29180789 24. Winkel J, Keuken M, van Maanen L, Wagenmakers E-J, Forstmann B. Early evidence affects later decisions: Why evidence accumulation is required to explain response time data. Psychon Bull Rev. 2014; 21:777–84. https://doi.org/10.3758/s13423-013-0551-8 PMID: 24395093 25. Britten K, Shadlen M, Newsome W, Movshon J. The analysis of visual motion: a comparison of neuronal and psychophysical performance. J Neurosci. 1992; 12:4745–65. https://doi.org/10.1523/JNEUROSCI. 12-12-04745.1992 PMID: 1464765 26. Britten K, Shadlen M, Newsome W, Movshon J. Responses of neurons in macaque MT to stochastic motion signals. Vis Neurosci. 1993; 10:1157–69. https://doi.org/10.1017/s0952523800010269 PMID: 8257671 27. Shadlen M, Newsome W. Motion perception: seeing and deciding. Proc Natl Acad Sci USA. 1996; 93:628–33. https://doi.org/10.1073/pnas.93.2.628 PMID: 8570606 28. Shadlen M, Newsome W. Neural basis of a perceptual decision in the parietal cortex (area lip) of the rhesus monkey. J Neurophysiol. 2001; 86:1916–36. https://doi.org/10.1152/jn.2001.86.4.1916 PMID: 11600651 29. Roitman JD, Shadlen MN. Response of neurons in the lateral intraparietal area during a combined visual discrimination reaction time task. J Neurosci. 2002; 22:9475–89. https://doi.org/10.1523/ JNEUROSCI.22-21-09475.2002 PMID: 12417672 30. Huk A, Shadlen M. Neural activity in macaque parietal cortex reflects temporal integration of visual motion signals during perceptual decision making. J Neurosci. 2005; 25:10420–36. https://doi.org/10. 1523/JNEUROSCI.4684-04.2005 PMID: 16280581 31. Zhang S, Lee M, Vandekerckhove J, Maris G, Wagenmakers E-J. Time-varying boundaries for diffusion models of decision making and response time. Front Psychol. 2014; 5:1364. https://doi.org/10.3389/ fpsyg.2014.01364 PMID: 25538642 32. Milosavljevic M, Malmaud J, Huth A, Koch C, Rangel A. The Drift Diffusion Model can account for the accuracy and reaction time of value-based choices under high and low time pressure. Judgm Decis Mak. 2010; 5:437–49. https://doi.org/10.2139/ssrn.1901533 33. Drugowitsch J, Moreno-Bote R, Churchland A, Shadlen M, Pouget A. The cost of accumulating evidence in perceptual decision making. J Neurosci. 2012; 32:3612–28. https://doi.org/10.1523/ JNEUROSCI.4010-11.2012 PMID: 22423085 34. Carland M, Thura D, Cisek P. The urgency-gating model can explain the effects of early evidence. Psychon Bull Rev. 2015; 22:1830–8. https://doi.org/10.3758/s13423-015-0851-2 PMID: 26452377 35. Marcos E, Tsujimoto S, Genovesio A. Eventand time-dependent decline of outcome information in the primate prefrontal cortex. Sci Rep. 2016; 6:25622. https://doi.org/10.1038/srep25622 PMID: 27162060 36. Kiani R, Churchland A, Shadlen M. Integration of direction cues is invariant to the temporal gap between them. J Neurosci. 2013; 33:16483–9. https://doi.org/10.1523/JNEUROSCI.2094-13.2013 PMID: 24133253 PLOS COMPUTATIONAL BIOLOGY The importance of urgency PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009455 October 4, 2021 16 / 17 37. Marcos E, Cos I, Girard B, Verschure PFMJ. Motor cost influences perceptual decisions. PLoS ONE. 2015; 10:e0144841. https://doi.org/10.1371/journal.pone.0144841 PMID: 26673222 38. Hagura N, Haggard P, Diedrichsen J. Perceptual decisions are biased by the cost to act. eLife. 2017; 6: e18422. https://doi.org/10.7554/eLife.18422 PMID: 28219479 39. Marcos E, Genovesio A. Determining monkey free choice long before the choice is made: the principal role of prefrontal neurons involved in both decision and motor processes. Front Neural Circuits. 2016; 10:75. https://doi.org/10.3389/fncir.2016.00075 PMID: 27713692 40. Marcos E, Genovesio A. Interference between Space and Time Estimations: From Behavior to Neurons. Front Neurosci. 2017; 11:631. https://doi.org/10.3389/fnins.2017.00631 PMID: 29209159 41. Mazurek M, Roitman J, Ditterich J, Shadlen M. A Role for neural integrators in perceptual decision making. Cereb Cortex. 2003; 13:1257–69. https://doi.org/10.1093/cercor/bhg097 PMID: 14576217 42. Latimer KW, Yates JL, Meister MLR, Huk AC, Pillow JW. Single-trial spike trains in parietal cortex reveal discrete steps during decision-making. Science. 2015; 349:184–7. https://doi.org/10.1126/science. aaa4056 PMID: 26160947 43. Thura D, Cisek P. The basal ganglia do not select reach targets but control the urgency of commitment. Neuron. 2017; 95:1160–70. https://doi.org/10.1016/j.neuron.2017.07.039 PMID: 28823728 44. Laming D. Information Theory of Choice-Reaction Times: London Academic; 1968. 45. Ratcliff R. A theory of memory retrieval. Psychol Rev. 1978; 85:59–108. https://doi.org/10.1037/0033295X.85.2.59 46. Usher M, McClelland J. The time course of perceptual choice: the leaky, competing accumulator model. Psychol Rev. 2001; 108:550–92. https://doi.org/10.1037/0033-295x.108.3.550 PMID: 11488378 47. Bogacz R, Gurney K. The basal ganglia and cortex implement optimal decision making between alternative actions. Neural Comput. 2007; 19:442–77. https://doi.org/10.1162/neco.2007.19.2.442 PMID: 17206871 48. Heathcote A, Brown S, Mewhort D. Quantile maximum likelihood estimation of response time distributions. Psychon Bull Rev. 2002; 9:394–401. https://doi.org/10.3758/bf03196299 PMID: 12120806 49. Ardia D, Mullen K, Peterson B, Ulrich J. DEoptim: Differential Evolution in R [Computer software manual]. Retrieved from http://CRAN.Rproject.org/packageDEoptim (R package version 2.2–2). 2013. https://doi.org/10.32614/RJ-2011-005 50. Mullen K, Ardia D, Gil D, Windover D, Cline J. DEoptim: An R package for global optimization by differential evolution. J Stat Softw. 2011; 40:1–26. https://doi.org/10.18637/jss.v040.i06 51. Rouder JN, Speckman PL, Sun D, Morey RD, Iverson G. Bayesian t tests for accepting and rejecting the null hypothesis. Psychon Bull Rev. 2009; 16:225–37. https://doi.org/10.3758/PBR.16.2.225 PMID: 19293088 52. Jeffreys H. The theory of probability: Oxford: Oxford University Press; 1961. PLOS COMPUTATIONAL BIOLOGY The importance of urgency PLOS Computational Biology | https://doi.org/10.1371/journal.pcbi.1009455 October 4, 2021 17 / 17