Bayesian model selection for group studies.

Area of Science:

• Neuroimaging
• Computational Neuroscience
• Statistical Modeling

Background:

• Bayesian model selection (BMS) is crucial for hypothesis testing in data analysis.
• Current group-level BMS in neuroimaging, like dynamic causal modelling (DCM), often uses fixed-effects metrics (e.g., group Bayes factor) that overlook inter-subject variability and outliers.
• There's a need for more robust methods to combine BMS results across subjects.

Purpose of the Study:

• To compare the traditional group Bayes factor (GBF) with two random-effects methods for group-level BMS.
• To introduce and evaluate a novel hierarchical Bayesian random-effects model for group BMS.
• To assess the robustness and informativeness of these methods, particularly in the presence of outliers.

Main Methods:

• Compared group Bayes factor (GBF) with classical (frequentist) and Bayesian random-effects approaches for group BMS.
• Developed a novel hierarchical Bayesian model treating the model itself as a random variable, estimating a Dirichlet distribution for model probabilities.
• Utilized variational Bayes optimization to estimate model probabilities and associated exceedance probabilities.

Main Results:

• The novel hierarchical Bayesian random-effects method provides more informative inference on model space than GBF and frequentist approaches.
• This Bayesian approach is significantly more robust to outliers in group data.
• Optimizing the conditional density of model probabilities proved superior to existing methods.

Conclusions:

• The proposed hierarchical Bayesian random-effects model offers a superior alternative for group-level BMS in neuroimaging and other fields.
• This method enhances the reliability of model selection in group studies by accounting for heterogeneity and outliers.
• The approach is applicable to diverse modeling endeavors beyond DCM, including EEG/MEG source reconstruction and computational modeling of cognition.