Permutation Test in DPABI

Submitted by Bin Lu on

DPABI has implemented a convenient way to perform permutation test in Statistical Analysis, which is based on the PALM package (Winkler et al., 2016).

You can use the default parameters for permutation test, or you can modify the PALM Settings for your own purpose. Of note, if you choose the default Number of permutations (5000), the calculation time may last hours which is depend on your model. If you choose the “Cluster inference” option, a cluster size threshold should be set to perform the cluster-wise permutation test. Setting “Cluster forming threshold (z)” as 2.3 equals to P=0.01 in one-tailed test. Threshold-free cluster enhancement (TFCE, Smith et al., 2009) and False Discovery Rate (FDR) corrections are available as well. You can also choose two-tailed test according to your purpose. A series of acceleration methods for permutation test are embedded in the PALM Setting interface (refer to Winkler et al., 2016 for detailed explanation of acceleration method).

 

 

 

 

A set of files will be generated after running statistical analysis with permutation test. “Temp” folder holds the files necessities for PALM. The content of each file in the result folder is introduced as follows.

 

  1. “_vox_tstat.nii” is the T-value map which is exactly the same as the T value map without permutation test.
  2. “_vox_tstat_uncp.nii” is the p-value map. The p-value corresponds to the rank of the observed T-value within the permutations for a given voxel (the null distribution is the permuted T values of that given voxel). (Under no Family-wise error rate correction)
  3. “_vox_tstat_fwep.nii” is the p-value map. The p-value corresponds to the rank of the observed T value within the permutations of maximum T values across all the voxels (the null distribution is composed by the maximum T value across all the voxels for each permutation, and the p-value for a given voxel is calculated based on its rank in the null distribution).
  4. “_clustere_tstat.nii” is simply the size (in voxels) of the cluster. This number acts as the test statistic.
  5. “_clustere_tstat_fwep.nii” is p-value computed in the same way as 3, i.e., using the distribution of the maximum cluster size.
  6. The TFCE maps are similar to Points 1, 2 and 3, but use TFCE values instead of T values.

 

To visualize the result of permutation test, open the “_vox_tstat.nii” as an overlay in DPABI_VIEW, click the “Apply a Mask for Additional Threshold” button in “Cluster” menu, select proper result file (e.g. “T2_tfce_tstat_fwep.nii”) and modify the threshold. A statistical map after permutation test will be shown in DAPBI_VIEW.

 

  • Smith, S. M., & Nichols, T. E. (2009). Threshold-free cluster enhancement: addressing problems of smoothing, threshold dependence and localisation in cluster inference. Neuroimage, 44(1), 83-98.
  • Winkler, A.M., Ridgway, G.R., Douaud, G., Nichols, T.E., Smith, S.M., (2016). Faster permutation inference in brain imaging. Neuroimage 141, 502-516.