Hi,
I’m having a bit of trouble with implementing the TFCE non-parametric cluster-level significance group-level t-test for my correlation map of 10 participants (with dimension (x,y,z) per participant, based on the voxels in my fMRI Searchlight Analysis).
Tried implementing it via the nilearn.glm.second_level.non_parametric_inference modules (and extracting the TFCE return value).
I want to implement the following via nilearn:
Correlation maps were analysed at group level by conducting voxel-wise one-sample t-tests to determine whether correlations significantly differed from zero. This produced a group-level brain activation map of corresponding z -scores. Statistics were adjusted for multiple comparisons using threshold-free cluster enhancement (TFCE). Statistical significance of these voxel-wise values was then determined via permutation testing. Voxel-wise values were compared to a null distribution of values equally spread around a baseline of zero. Null distribution values were computed over 10,000 iterations by inverting positive and negative signs of the voxel values for a random half of the participants. During each iteration, TFCE values were calculated separately for positive and negative valued voxels, resulting in null distributions for both positive and negative values. The next step involved identifying voxels in which the actual TFCE value was larger than over half of the values generated during the permutations. For these voxels, the TFCE correction procedure took the number of permutations where this was the case, and scaled this to be a value between 0 and 0.5 if the actual TFCE value was negative, or 0.5 and 1 if the actual TFCE value was positive. These values were then computed into a z -score using the inverse normal cumulative distribution function. TFCE-corrected maps were voxel-wise thresholded at z = 1.96, which corresponds to p < .05 after correction for multiple comparisons.
I’d kindly appreciate your help!
Thanks,
Roy.