Comparing families of dynamic causal models

Area of Science:

• Neuroscience
• Computational Biology
• Biostatistics

Background:

• Bayesian model evidence is used for comparing mathematical models in scientific data analysis.
• Traditional 'best model' selection in biological sciences can be unreliable with numerous models or subject-specific variations.

Purpose of the Study:

• To propose a novel Bayesian framework combining family-level inference and Bayesian model averaging.
• To enhance the robustness of model comparison and parameter inference in complex biological systems.
• To provide parameter inferences independent of specific model structure assumptions.

Main Methods:

• Implementing family-level inference to address uncertainties in model structure (e.g., system inputs, processing pathways).
• Applying Bayesian model averaging within identified model families.
• Utilizing Dynamic Causal Models (DCM) for brain imaging data analysis.

Main Results:

• Demonstrated a method to reduce uncertainty regarding model structure characteristics.
• Achieved parameter inferences robust to variations in model structure.
• Successfully illustrated the combined approach using Dynamic Causal Models.

Conclusions:

• The proposed Bayesian framework offers a more resilient alternative to 'best model' selection for complex scientific data.
• Family-level inference and Bayesian model averaging enhance the reliability of parameter estimation in biological modeling.
• This approach is particularly valuable for analyzing brain imaging data with Dynamic Causal Models.