Bayesian Workflow for Generative Modeling in Computational Psychiatry

Area of Science:

• Computational Neuroscience
• Cognitive Science
• Psychiatry

Background:

• Generative models hold significant clinical potential but require reliable statistical inference.
• Bayesian workflow is a suggested but underutilized approach in Translational Neuromodeling and Computational Psychiatry (TN/CP).
• Hierarchical Gaussian Filter (HGF) models are used for hierarchical Bayesian belief updating in computational modeling.

Purpose of the Study:

• To demonstrate a practical application of Bayesian workflow in TN/CP.
• To address challenges in statistical inference with univariate behavioral data.
• To introduce novel response models for simultaneous inference from multivariate data.

Main Methods:

• Applied Bayesian workflow to Hierarchical Gaussian Filter (HGF) models.
• Developed and utilized novel response models for multivariate data (binary responses and response times).
• Validated methods using simulations and empirical data from a speed-incentivised associative reward learning (SPIRL) task.

Main Results:

• Models utilizing both binary responses and response times ensure robust statistical inference and parameter identifiability.
• A linear relationship was identified between log-transformed response times and outcome uncertainty in the SPIRL task.
• The study illustrates the benefits of Bayesian workflow for TN/CP applications.

Conclusions:

• Bayesian workflow increases the transparency and robustness of generative modeling in TN/CP.
• Adopting Bayesian workflow is crucial for the long-term success of Translational Neuromodeling and Computational Psychiatry.
• The developed multivariate response models improve inference from limited behavioral data.