Non-decision time-informed collapsing threshold diffusion model: A joint modeling framework with identifiable time-dependent parameters

Abstract

Over the past sixty years, evidence accumulation models have emerged as a dominant framework for explaining the neural and behavioral aspects of the process underlying decision making. These models have also been widely used as a measurement instrument to assess individual differences in latent cognitive constructs underlying decision making. A central assumption of most of these models is that decision makers accumulate noisy evidence until a fixed decision threshold is reached. However, both behavioral and neuroscientific findings, along with theoretical considerations related to optimality, have suggested that the decision threshold varies over time. Although time-dependent threshold models often provide a better account of empirical data, a major challenge associated with these models is the unreliable estimation of their parameters. This limitation has led researchers to emphasize model-fitting comparisons rather than interpreting parameter values or accounting for individual differences in the dynamics of the decision threshold. In this work, we address the reliability issue of parameter estimation in time-dependent threshold diffusion models by proposing a joint modeling approach that links non-decision time to external observations. Parameter recovery simulations demonstrate that informing the diffusion model with trial-level noisy measurements of non-decision time substantially improves the reliability of parameter estimation for time-dependent threshold diffusion models. Additionally, we reanalyzed the experimental data from two perceptual decision-making tasks to illustrate the feasibility of the proposed modeling approach. Non-decision time measurements were extracted from electroencephalography (EEG) recordings using the hidden multivariate pattern method. The cognitive modeling results revealed that, in addition to the reliable parameter estimation, constraining non-decision time improves the absolute fit to behavioral data.