Chávez De la Peña Adriana F, Vandekerckhove Joachim
Department of Cognitive Sciences, University of California, Irvine, CA, USA.
Department of Statistics, University of California, Irvine, CA, USA.
Psychon Bull Rev. 2025 Jul 25. doi: 10.3758/s13423-025-02729-y.
The EZ-diffusion model is a simplification of the popular drift diffusion model of choice response times that allows researchers to calculate diffusion model parameters directly from data with no need for expensive computations. The EZ-diffusion model is based on a system of equations in which the diffusion model's drift rate, boundary separation, and nondecision time parameters are jointly used to predict three summary statistics (the accuracy rate and the mean and variance of the correct response times). These equations can then be inverted to obtain estimators for the three parameters from these summary statistics. Here, we describe a probabilistic formulation of the EZ-diffusion model that can serve as a hyper-efficient proxy model to the drift diffusion model. The new formulation is based on sampling distributions of summary statistics and consists only of normal and binomial distributions. It can easily be implemented in any probabilistic programming language. We demonstrate the validity of the proxy model through extensive simulation studies and provide multiple examples (via https://osf.io/bzkpn/ ), including an implementation in JASP. We conclude that, although the recovery of some parameters with the proxy model is biased, the recovery of regression parameters is good, making the method useful for cognitive psychometrics (i.e., explanatory cognitive modeling). Casting the EZ-diffusion model in the broad family of Bayesian generative models allows us to benefit from mature implementations, practical workflows, and powerful extensions that are not possible without a probabilistic implementation and not feasible with the regular drift diffusion model. Code and example applications are provided via https://osf.io/bzkpn/ .
EZ扩散模型是流行的选择反应时漂移扩散模型的简化版本，它使研究人员能够直接从数据中计算扩散模型参数，而无需进行昂贵的计算。EZ扩散模型基于一个方程组，其中扩散模型的漂移率、边界分离和非决策时间参数共同用于预测三个汇总统计量（准确率以及正确反应时的均值和方差）。然后可以对这些方程进行反演，以从这些汇总统计量中获得三个参数的估计值。在此，我们描述了EZ扩散模型的一种概率公式，它可以作为漂移扩散模型的超高效代理模型。新公式基于汇总统计量的抽样分布，仅由正态分布和二项分布组成。它可以很容易地在任何概率编程语言中实现。我们通过广泛的模拟研究证明了代理模型的有效性，并提供了多个示例（通过https://osf.io/bzkpn/），包括在JASP中的实现。我们得出结论，虽然使用代理模型恢复某些参数存在偏差，但回归参数的恢复效果良好，这使得该方法对认知心理测量学（即解释性认知建模）很有用。将EZ扩散模型置于贝叶斯生成模型的广泛家族中，使我们能够受益于成熟的实现、实用的工作流程以及强大的扩展，而这些如果没有概率实现是不可能的，并且使用常规漂移扩散模型是不可行的。代码和示例应用可通过https://osf.io/bzkpn/获取。
