Bootstrapping fMRI Data: Dealing with Misspecification.

Area of Science:

• Neuroimaging
• Statistical Modeling

Background:

• The General Linear Model (GLM) is widely used for analyzing functional magnetic resonance imaging (fMRI) time series.
• Recent studies have questioned the validity of GLM-based inference due to unmet assumptions.
• Bootstrap procedures offer a potential alternative by reducing reliance on strict modeling assumptions.

Purpose of the Study:

• To empirically compare two voxelwise GLM-based bootstrap approaches for fMRI data analysis.
• To evaluate the inferential properties and data characteristic reproduction capabilities of semi-parametric versus fully parametric bootstrap methods.
• To assess the performance of these methods under various noise structures and signal generation mechanisms, including model misspecification.

Main Methods:

• Simulation of fMRI data with different noise structures and signal characteristics.
• Implementation of a semi-parametric bootstrap approach using independent blocks of residuals.
• Implementation of a fully parametric bootstrap approach assuming independent whitened residuals.
• Empirical comparison based on inferential accuracy and data reproducibility.

Main Results:

• The fully parametric bootstrap approach performed well in both inference and data reproduction when the GLM and noise models were correctly specified.
• In cases of model misspecification, the fully parametric approach benefited from additional blocking of residuals.
• The semi-parametric approach showed weaker inferential performance but comparable image reproducibility to the blocked fully parametric approach.
• GLM-based bootstrapping demonstrated an advantage over classical parametric inference when the expected signal model was misspecified.

Conclusions:

• Bootstrap methods can enhance the validity of inference in fMRI analysis, particularly when GLM assumptions are not met.
• The choice between semi-parametric and fully parametric bootstrap depends on the specific data characteristics and the degree of model specification.
• GLM-based bootstrapping provides a robust alternative to classical inference for fMRI time series analysis, especially under model misspecification.