Faster permutation inference in brain imaging.

Area of Science:

• Neuroimaging analysis
• Statistical inference
• Computational neuroscience

Background:

• Permutation tests are crucial for reliable inference in neuroimaging but are computationally intensive.
• The computational burden increases significantly with large, complex neuroimaging datasets, limiting their practical application.
• Existing hardware and software improvements offer generic acceleration but do not address the core computational demands of permutation testing.

Purpose of the Study:

• To develop and evaluate computationally efficient methods for permutation testing in neuroimaging.
• To compare the performance of various acceleration techniques against traditional permutation approaches.
• To identify the optimal method balancing speed, accuracy, and statistical validity for different neuroimaging analysis scenarios.

Main Methods:

• Compared six acceleration approaches: limited permutations, negative binomial distribution, generalized Pareto distribution tail fitting, gamma distribution moment approximation, empirical gamma distribution fitting, and reduced voxel permutation with matrix theory.
• Assessed methods using synthetic data for error rates, statistical power, agreement with reference results, and resampling risk.
• Re-analyzed a voxel-based morphometry study using real neuroimaging data to validate findings.

Main Results:

• All evaluated methods achieved exact error rates and similar statistical power.
• Methods (i), (iii), and (v) exhibited higher resampling risks.
• Method (iv) (no permutations, gamma distribution approximation) was the fastest for comparable resampling risks.
• Methods (iii) and (v) detected stronger effects in family-wise error rate corrected maps for real data.
• Method (iv) is recommended for uncorrected p-values assuming symmetric errors; method (iii) is recommended for corrected p-values and other settings.

Conclusions:

• Several computationally efficient methods can accelerate permutation testing in neuroimaging without compromising error rates or power.
• The choice of method depends on whether corrected or uncorrected p-values are required and assumptions about error distribution symmetry.
• The recommended methods are available in the PALM (Permutation Analysis of Linear Models) tool.