Behavioral and cognitive neuroscience findings regarding assumptions of the evidence accumulation model
摘要: 证据积累模型(evidence accumulation models, EAM)是关于人类决策过程的主要认知模型之一，其假定决策者不断搜集信息并将信息整合成与决策有关的证据，当累积证据量达到某个阈值时做出决策并反应。虽然EAM在研究中得到广泛应用，甚至有研究者认为其已经达到了理论的高原期，但EAM的理论预设并未被严格检验。以最具有代表性的EAM计算模型——漂移扩散模型(drift diffusion model, DDM)——为例，其存在5个模型预设：（1）证据积累的普适性、（2）证据积累的选择性、（3）证据以存在噪音的线性方式积累、（4）决策标准恒定、（5）决策与运动执行过程独立。回顾对这五个基本预设进行检验的实证研究，可以发现：尽管DDM被广泛应用于知觉决策、记忆和基于价值的决策任务，但研究者仅验证了证据积累是否存在于知觉决策任务中；证据积累的选择性目前较缺乏实证研究；证据以存在噪音的线性方式积累的预设得到了较多知觉决策实验数据的支持，但在基于价值的决策中其是否成立仍然存在争议；决策标准恒定的预设则存在较大争议；决策独立于反应执行的预设近年来受到关注，但较多实证研究质疑了这一预设。总之，对证据积累模型的预设进行验证的实证研究并不均衡，部分预设的实证证据有限，亟需更多的实证研究进行验证。研究者们需要在解释DDM的结果时保持谨慎。同时，通过对EAM预设进行清晰表述并回顾其实证证据，本研究表明清晰地表述模型预设有助于全面而系统地检验模型，从而不断地推动模型的更新与理论的发展，以更好地理解人类认知过程。
Abstract: The evidence accumulation model is a widely used cognitive model of human decision-making, which assumes that decision-makers continuously gather and integrate information into evidence relevant to the decision and make a decision once the accumulated evidence reaches a predefined threshold. With the increasing popularity of evidence accumulation model, some researchers claim it has reached a theoretical plateau and can be considered as the standard model for analyzing response time and choices. However, the theoretical assumptions underlying these models lack rigorous testing. As an example, the drift-diffusion model (DDM) is an instantiation of evidence accumulation and has five underlying assumptions: (1) the universality of evidence accumulation; (2) the selectivity of evidence accumulation; (3) linear integration of evidence with noise; (4) a constant decision criterion; and (5) decision-making is independent of motor execution. DDM has been widely used in cognitive tasks, such as value-based decision-making, and social decision-making, probably due to the availability of user-friendly software for parameter estimation. However, only a few studies systematically examined to what extent these five assumptions of DDM were supported by empirical studies. To fill the gap, we reviewed studies that tested these five assumptions.
For the first assumption of DDM, the universality of evidence accumulation, we only found direct evidence from studies that employed perceptual decision-making tasks. For other studies that used DDM for modeling, such as value-based decision-making or social decision-making, we found few studies that directly tested the existence of evidence accumulation. The second assumption, the selectivity of evidence accumulation, suggested that only information related to the goal would contribute to evidence accumulation. We did not find empirical data supporting this assumption except for O’Connell et al. (2012). However, evidence from conflict tasks (e.g., flanker task) suggested that information irrelevant to the goal may also be incorporated into the evidence accumulation. Data from conflict tasks inspired new models related to evidence accumulation model and called for further investigation into the mechanism behind the selectivity of evidence. The third and fourth assumptions constitute the core assumptions of DDM, i.e., “evidence accumulate-to-bound”. Regarding the third assumption, which posits that evidence with noise is accumulated linearly, supporting data were found from animal studies and human EEG studies that employed perceptual decision-making. However, human EEG data from value-based decision-making tasks has challenged the validity of this assumption. The fourth assumption, that the decision criterion is constant, is controversial and has been challenged by several other evidence accumulation models, such as collapsing boundary models. The last assumption, that decision-making is independent of motor execution, has also been questioned by empirical data from both animal studies and human behavioral and electromyography data, despite support from EEG recording.
In summary, we found that, while the standard DDM is commonly used in many sub-fields of psychology and neuroscience, empirical studies that directly tested five assumptions of DDM were mainly from perceptual decision-making tasks. Also, we found that challenging these assumptions often resulted in new computational models. These findings call for studies to test these assumptions and develop new models. Besides, these findings suggest that researchers should be cautious when interpreting the parameters estimated from standard DDM. Finally, our review suggests that increasing transparency in model assumptions will accelerate the revision of models and theories, and ultimately deepen our understanding of human cognitive processes.