Adjusting the neuroimaging statistical inferences for nonstationarity

Area of Science:

• Neuroimaging
• Statistical inference
• Brain imaging analysis

Background:

• Cluster-based inference in neuroimaging is generally more powerful than voxel-wise methods.
• Standard cluster methods assume stationarity, leading to spatially variant false positive risks under nonstationarity.
• Existing Random Field Theory (RFT) adjustments control family-wise error rate but haven't been assessed for local false positive risk homogeneity.

Purpose of the Study:

• To propose and evaluate a novel empirical cluster size adjustment for nonstationary neuroimaging data.
• To introduce a new metric for assessing the homogeneity of local false positive risk.
• To compare the proposed empirical method against RFT-based adjustments and unadjusted methods.

Main Methods:

• Development of a new cluster size adjustment using local empirical cluster size distributions and a two-pass permutation method.
• Proposal of a novel approach to quantify the homogeneity of local false positive risk.
• Application and comparison of methods to both cluster-based inference and threshold-free cluster enhancement (TFCE) using simulated and real neuroimaging data.

Main Results:

• Unadjusted cluster inference exhibits expected heterogeneity in local false positive risk.
• RFT-based adjustments partially reduce, but do not eliminate, this heterogeneity.
• The proposed empirical adjustment significantly enhances the homogeneity of local false positive risk.
• Threshold-free cluster enhancement (TFCE) demonstrates robustness to nonstationarity.

Conclusions:

• The novel empirical adjustment method offers improved control over local false positive rates in nonstationary neuroimaging data.
• Homogeneity assessment is crucial for evaluating cluster inference methods under nonstationarity.
• TFCE represents a robust alternative for cluster inference in the presence of nonstationarity.