Bayesian model reduction and empirical Bayes for group (DCM) studies.

Area of Science:

• Neuroscience
• Computational Psychiatry
• Statistical Modeling

Background:

• Group studies often involve complex nonlinear models at the individual level (e.g., dynamic causal models) and linear models at the group level.
• Analyzing such hierarchical models can be computationally intensive, limiting the exploration of multiple models or complex effects.

Purpose of the Study:

• To present novel Bayesian procedures for efficiently analyzing group studies with hierarchical nonlinear and linear models.
• To demonstrate the utility of Bayesian model reduction for model inversion and inference in these complex settings.

Main Methods:

• Developed Bayesian procedures integrating nonlinear within-subject models and linear between-subject models.
• Employed Bayesian model reduction for efficient inversion of multiple models or complex hierarchical models.
• Utilized a simulated mismatch negativity study in schizophrenia as a worked example.

Main Results:

• Bayesian model reduction allows for efficient consideration of parametric random effects and group-level inferences (in seconds).
• Demonstrated robustness of Bayesian model reduction to violations of the Laplace assumption in dynamic causal modeling.
• Showcased recursive application for facilitating both classical and Bayesian group difference inference.

Conclusions:

• The proposed Bayesian model reduction offers an efficient and robust approach for analyzing complex hierarchical models in group studies.
• These methods are applicable to classification and prediction tasks within empirical Bayesian frameworks.