{ "cells": [ { "cell_type": "code", "execution_count": null, "metadata": { "collapsed": false }, "outputs": [], "source": [ "%matplotlib inline" ] }, { "cell_type": "markdown", "metadata": {}, "source": [ "\n# Tutorial 3: Null models for gradient significance\nIn this tutorial we assess the significance of correlations between the first\ncanonical gradient and data from other modalities (curvature, cortical\nthickness and T1w/T2w image intensity). A normal test of the significance of\nthe correlation cannot be used, because the spatial auto-correlation in MRI\ndata may bias the test statistic. In this tutorial we will show three\napproaches for null hypothesis testing: spin permutations, Moran spectral\nrandomization, and autocorrelation-preserving surrogates based on variogram\nmatching.\n\n
When using either approach to compare gradients to non-gradient markers,\n we recommend randomizing the non-gradient markers as these randomizations\n need not maintain the statistical independence between gradients.
With spin permutations, midline vertices (i.e,, NaNs) from both the\n original and rotated data are discarded. Depending on the overlap of\n midlines in the, statistical comparisons between them may compare\n different numbers of features. This can bias your test statistics.\n Therefore, if a large portion of the sphere is not used, we recommend\n using Moran spectral randomization instead.