Geometry-derived statistical significance: A probabilistic framework for detecting true positive findings in MRI data

Area of Science:

• Neuroimaging analysis
• Statistical methodology
• Brain imaging

Background:

• Standard false discovery rate (FDR) procedures lack statistical power in neuroimaging due to limited participant numbers and failure to incorporate random field geometry.
• Existing methods like topological FDR and threshold free cluster enhancement (TFCE) improve power by considering local geometry but have limitations like required threshold or weight specifications.

Purpose of the Study:

• To introduce and evaluate the Geometry-Derived Statistical Significance (GDSS) procedure for enhanced statistical power in neuroimaging.
• To compare GDSS performance against existing multiple comparison correction methods using synthetic and real-world data.

Main Methods:

• Developed the Geometry-Derived Statistical Significance (GDSS) procedure, combining voxelwise p-values with local random field geometry probabilities.
• Compared GDSS performance against topological FDR and TFCE using synthetic and real-world neuroimaging datasets.

Main Results:

• GDSS demonstrated substantially greater statistical power compared to existing methods, with less variability related to participant numbers.
• GDSS was more conservative than TFCE, requiring higher effect sizes for null hypothesis rejection.
• Observed a decrease in Cohen's D effect size with increasing participant numbers, suggesting potential underestimation in small study sample size calculations.

Conclusions:

• GDSS offers significantly improved statistical power for detecting true positives while controlling false positives in neuroimaging.
• GDSS is particularly effective in small imaging cohorts (fewer than 40 participants).
• The study recommends presenting effect size maps alongside p-value maps for accurate interpretation of neuroimaging findings.