Which model better predicts brain-behavior relationships during decision-making? A simulation-based comparison of single-trial directed and integrative drift-diffusion models

The field of model-based cognitive neuroscience uses (neuro)cognitive models to link human behavior to functional brain recordings. Many researchers have used these methods to uncover the neural mechanisms underlying decision-making. However, there are a variety of ways to connect different data types, and it remains unclear which model best yields reliable inferences from empirical data, particularly when neural recordings are contaminated by noise. In this study, we compared two neurocognitive models that use single-trial neural measures (e.g., single-trial EEG potentials), namely: Directed models that treat single-trial neural data as univariate independent variables for cognitive parameters, and Integrative models that describe single-trial neural data in addition to choice response time data through shared latent parameters. Both use drift-diffusion models (DDMs) as the base cognitive models. We simulated decision-making data with varying neural signal-to-noise ratios (SNRs) as well as varying neural-behavior coupling strengths to assess how well each model recovers the true latent cognitive parameters. We found that Integrative models are better able to detect true nonzero relationships, but suffer from a problem of false positives, at least as currently implemented in simulation-based inference methods. Directed models, on the other hand, are more robust to false positives but struggle to identify any relationships in the presence of noise. This simulation-based approach highlights each model’s strengths and limitations for real-world applications, particularly in addressing noise-contaminated EEG recordings.