Evidence Accumulation: General

Decision models typically assume that reaction time on each trial is the sum of decision and non-decision time, the latter being widely held to capture sensory and motor delays. However, the mathematical assumptions allowing this split can be simplistic, arbitrary and lead to counter-intuitive results. How reaction time should be split between decision and non-decision time is both a theoretical and an empirical question. In this talk, I will explore the boundaries between sensory and decision processes, and between decision and motor processes, to propose a biological definition of when decision time may start and stop. I will then present a model-free empirical approach to estimating non-decision time, directly observable in behavioural data. In contrast to non-decision time parameters from model fits, measures extracted using this approach consistently satisfy widespread selective influence assumptions: they vary predictably with visual and motor factors, and do not vary with higher-level task-demands. Last, using the EZ, DDM and LBA models, we conclude that non-decision time parameter from these models is unlikely to consistently reflect visuomotor delays.

This is an in-person presentation on July 22, 2024 (15:20 ~ 15:40 CEST).

Traditional models of intertemporal preference tend to be static and thus lack an explicit account of the underlying cognitive processes. Several dynamic models of intertemporal choice have been proposed to address this drawback. Most of such models (e.g., Dai & Busemeyer, 2014; Rodriguez et al., 2014) adopt an evidence-accumulation approach and can account for major behavioral regularities regarding both choice responses and corresponding response times. Recently, Dai et al. (2018) proposed an alternative, non-accumulative framework for modeling intertemporal choice using two time-honored concepts in economics and psychophysics, that is, random utility and discrimination threshold. According to this framework, an intertemporal choice is made upon a large enough random evaluation of the relative advantage of one option over the other. When such an evaluation is relatively small in absolute value, another random evaluation will be made without accumulating any information or evidence from previous evaluations. Such models have been shown to be capable of not only qualitatively accounting for major behavioral effects regarding choice responses and response times but also generally outperforming models built upon the DFT framework. More recent development of such models (Zhang et al., 2023) further revealed their power to accommodate intertemporal preferences measured by matching tasks. The general success of such non-accumulative models suggests that evidence accumulation might not be essential for generating intertemporal preference. Further research is in need for a more rigorous test of this novel framework against other preferential decisions.

This is an in-person presentation on July 22, 2024 (15:40 ~ 16:00 CEST).

The EZ-drift diffusion model (EZDDM) consists of an invertible set of equations that relate the drift rate, boundary separation, and nondecision time parameters of the drift diffusion model (DDM) to three summary statistics of choice RT data. The EZDDM is computationally inexpensive, making parameter estimation attainable from simply the accuracy rate and the mean and variance of the correct-trial RTs. We introduce an implementation of the EZDDM within a Bayesian framework, using binomial and normal distributions to model the sampling distributions of these summary statistics. Moving into the realm of Bayesian generative models allows us to implement versatile extensions, such as cognitive latent variable models that capture differences across levels of variation, metaregression structures, and hierarchical models. The resulting “EZ Bayesian hierarchical drift diffusion model” (EZBHDDM) serves as a hyper-efficient proxy model to the hierarchical DDM. We demonstrate, with simulation studies, the efficacy of our proxy model, and present applied illustrations using the graphical Bayesian analysis package JASP. While we find some bias in some of the drift diffusion parameters in recovery studies, we find our proxy model to be robust and highly efficient in recovering critical regression parameters from models that incorporate a metaregression structure. Since the EZBHDDM can be easily implemented in any probabilistic programming language, it can be extended to different Bayesian structures without great computational expense.

This is an in-person presentation on July 22, 2024 (16:00 ~ 16:20 CEST).