Multiple Comparisons Corrections

AFNI Message Board
1

Hello,

I am wondering if you could clarify how to do multiple comparisons corrections using afni. Here are the steps I am currently using:

  1. afni_proc.py to get the blur estimates for each subject
  2. average each of the first three parameters from err_reml ACF blur estimates (blur_est.1D) across subjects
  3. Use 3dMean on each subjects group.mask file created by afni_proc.py
  4. 3dClustSim -mask group.mask.all -acf 0.813135017 3.86235345 16.96651552 -prefix ClustSim.acf.all

2-sided thresholding

Grid: 104x125x104 1.75x1.75x1.75 mm^3 (366992 voxels in mask)

CLUSTER SIZE THRESHOLD(pthr,alpha) in Voxels

-NN 1 | alpha = Prob(Cluster >= given size)

pthr | .10000 .05000 .02000 .01000

------ | ------ ------ ------ ------

0.050000 1379.5 1689.5 2091.0 2437.0
0.020000 488.1 580.0 728.0 866.0
0.010000 275.1 323.4 401.7 461.7
0.005000 175.8 205.2 247.3 286.0
0.002000 105.7 124.2 150.5 174.1
0.001000 75.6 90.2 108.8 125.3
0.000500 55.0 66.2 79.9 92.7
0.000200 36.9 44.7 56.0 64.3
0.000100 27.1 33.8 42.1 50.3

Does this look right/reasonable so far?

Say I use pthr=0.01 and alpha=0.05, does the 324 voxel cluster threshold then get applied to all group-level analyses or do I need to re-run 3dClustSim using different parameters/mask for each different analysis? If the latter, which parameters and mask am I supposed to use?

Thanks!!

2

Hello,

That looks reasonable, though using an uncorrected p=0.01
might be a little high to be accepted these days.

As for the group mask, that is probably identical across
subjects, as it is simply a mask from the group template.
If you want a tighter mask, consider making it from the
EPI datasets (full_mask*). For example, one could make a
dataset which has coverage at each voxel by at least 70%
of the subjects:

3dmask_tool -input full_mask.*+tlrc.HEAD -prefix mask_overlap.7 -frac 0.7

Computing minimum cluster sizes via the average blur_est.1D
values allows you to use that size across all group tests.

  • rick
3

Excellent! Thank you for your response. I will definitely try this with the full_mask and see if I can get “better” results.