Faster permutation inference in brain imaging

• DOI: 10.1016/j.neuroimage.2016.05.068
• Publication Date: 2016-06-07T16:57:40Z
• Authors: Anderson M. Winkler; Gerard R. Ridgway; Gwenaëlle Douaud; Thomas E. Nichols; Stephen M. Smith
• Keywords: Gamma distribution; Models, Statistical; Cognitive Neuroscience; Pearson type III distribution; Brain; Reproducibility of Results; Neuroimaging; Permutation tests; Image Enhancement; Sensitivity and Specificity; Article; 03 medical and health sciences; 0302 clinical medicine; Generalised Pareto distribution; Neurology; Data Interpretation, Statistical; Image Interpretation, Computer-Assisted; Humans; Negative binomial distribution; Tail approximation; Low rank matrix completion; Computer Simulation; Algorithms

Permutation tests are increasingly being used as a reliable method for inference in neuroimaging analysis. However, they computationally intensive. For small, non-imaging datasets, recomputing model thousands of times is seldom problem, but large, complex models this can be prohibitively slow, even with the availability inexpensive computing power. Here we exploit properties statistics general linear (GLM) and their distributions to obtain accelerations irrespective generic software or hardware improvements. We compare following approaches: (i) performing small number permutations; (ii) estimating p-value parameter negative binomial distribution; (iii) fitting generalised Pareto distribution tail permutation (iv) p-values based on expected moments distribution, approximated from gamma (v) direct empirical (vi) permuting reduced voxels, completion remainder using low rank matrix theory. Using synthetic data assessed different methods terms error rates, power, agreement reference result, risk taking decision regarding rejection null hypotheses (known resampling risk). also conducted re-analysis voxel-based morphometry study real-data example. All yielded exact rates. Likewise, power was similar across methods. Resampling higher (i), (v). comparable risks, which no permutations done absolute fastest. produced visually maps real data, stronger effects detected family-wise rate corrected by (v), generally results seen set. Overall, uncorrected p-values, found best long symmetric errors assumed. In all other settings, including familywise recommend approximation (iii). The considered freely available tool PALM - Analysis Linear Models.