Handling Multiplicity in Neuroimaging Through Bayesian Lenses with Multilevel Modeling

Area of Science:

• Neuroimaging analysis
• Statistical modeling in neuroscience

Background:

• Massively univariate approaches in neuroimaging are inefficient and over-penalize results.
• Current methods for correcting multiple testing issues can lead to over-penalization and disadvantage smaller brain regions.

Purpose of the Study:

• To propose a more efficient Bayesian multilevel (BML) modeling approach for neuroimaging analysis.
• To control for type S (sign) and type M (magnitude) errors, which are more relevant than the false positive rate (FPR).
• To improve modeling efficiency by integrating information across brain regions and dissolving the multiple testing issue.

Main Methods:

• Utilized Bayesian multilevel (BML) modeling to pool and share information among brain regions.
• Applied the BML approach at the region of interest (ROI) level, ensuring all regions are treated equally.
• Demonstrated the approach's performance and validity using an experimental dataset.

Main Results:

• The BML approach improves modeling efficiency and spatial specificity compared to massively univariate methods.
• It effectively controls type S and type M errors, enhancing the quality of statistical inference.
• Results reporting is promoted in totality and transparency, avoiding arbitrary p-value thresholding.

Conclusions:

• Bayesian multilevel (BML) modeling offers a more efficient and robust alternative to traditional neuroimaging analysis methods.
• This approach enhances statistical inference quality by focusing on error types more relevant to scientific conclusions.
• The BML methodology, available in the AFNI suite, promotes transparent and comprehensive reporting of neuroimaging results.