SpinozaWiki:Permutation testing

Permutation testing

A 4D skeletonised FA image or a GM_mod_merg_s??.nii.gz can be tested with randomise on a voxel per voxel basis (voxelwise statistics). Randomise is a FSL program (command line) that performs permutation testing.

Syntax

randomise -i <input name> -o <outputname> -m <mask> -d design.mat –t design.con –n 5000 –V -x

Example syntax VBM

randomise -i GM_mod_merg_s4 -m GM_mask.nii.gz -o VBM_output -d design.mat –t design.con –n 5000 –V -x

Example syntax DTI

randomise -i all_FA_skeletonised -o DTI_output -m mean_FA_skeleton_mask.nii.gz -d design.mat –t design.con –n 5000 –V -x

• The option –n indicates how many iterations will be made. By default, however, the randomise tool will run 5000. For a publication grade analysis this value should be in the 25000.
• The option –V makes randomise run in Verbose mode, e.g. indicate where it is with the analysis.
• You can use the GLM GUI to generate design.mat and design.con. Note that the order of the entries (rows) in your design matrix must match the alphabetical order of your original FA or GM images because this is order the files are placed in the all_FA_skeletonised or GM_mod_merg images.
• The option –m gives the name of the maskfile. Since permutation is a time consuming process it is sensible to only calculate it for locations were data is actually located. The mask name of a VBM analysis is GM_mask.nii.gz and for TBSS analysis is all_FA_skeletonised_mask.nii.gz.
• The option -x must be specified to obtain files with p values (the p-values wil be 1-p meaning that for instance 0.01 will become 0.99).
• If you have fewer than 20 subjects then you will usually see an increase in power by using variance smoothing, as the following example which smoothens the variance with a 10 mm HWHM (see below).

randomise -i <Input, e.g. all_FA_skeletonised or GM_mod_merg_s> -o <outputname> -d design.mat –t design.con –n 5000 –v 10 -V

The following options exist for multiple comparison correction:

• -c <c_thresh> carry out cluster-based thresholding.
• -C <cmass_thresh> carry out cluster-mass-based thresholding (as well as max and voxelwise).

The following options exist for the sharpening of statistical values.

• -T threshold free cluster enhancement for VBM.
• -T2 threshold free cluster enhancement for DTI.

Statistical output

Randomise will create several statistical maps:

• <output>_tstat? t-values.

In addition the following thresholded statistics will be calculated depending on the specified options

• <output>_max_tstat? Voxel-based thresholding corrected for multiple comparisons (bonferroni).
• <output>_vox_tstat? Voxel-based thresholding uncorrected for multiple comparisons.
• <output>_max_tstat? Voxel-based thresholding corrected for multiple comparisons.
• <output>_maxc_tstat? Cluster-based thresholding corrected for multiple comparisons.