How to improve parameter estimates in GLM-based fMRI data analysis: cross-validated Bayesian model averaging.

Area of Science:

• Neuroimaging
• Statistical modeling
• Brain activity analysis

Background:

• General linear models (GLMs) are common in fMRI first-level analysis but their quality is often unassessed.
• Previous work introduced cross-validated Bayesian model selection (cvBMS) for group-level GLM analysis.
• A more efficient approach exists when models differ only in nuisance variables.

Purpose of the Study:

• To propose cross-validated Bayesian model averaging (cvBMA) for improving parameter estimates in fMRI GLMs.
• To leverage posterior model probabilities for combining information across models.
• To enhance sensitivity to experimental effects in neuroimaging data analysis.

Main Methods:

• Developed and applied cross-validated Bayesian model averaging (cvBMA).
• Focused on GLM analyses where regressors of interest are common across models.
• Utilized posterior probabilities to average parameter estimates from multiple candidate models.

Main Results:

• cvBMS was shown to prevent the failure to detect established effects.
• cvBMA demonstrated increased sensitivity to experimental effects compared to single-best-model approaches.
• The proposed cvBMA method improves parameter estimation by integrating information across models.

Conclusions:

• cvBMA offers a more efficient and sensitive method for GLM-based fMRI analysis when models vary in nuisance regressors.
• This approach enhances the reliability of detecting experimental effects in neuroimaging studies.
• Bayesian model averaging provides a robust alternative to selecting a single best model.