3dttest++ q values results don't match afni display

AFNI Message Board
1

The output results for q-value of my 3ttest++ on resting state data with two groups (healthy volunteers vs patients), two visits combined, using resting state scans, does not match the results using the afni display.

I used this gen_group_command.py:

Compare all 100s vs 300s, combined visits, run 2

gen_group_command.py -write_script analysis_script.tt.100s_vs_300s_AllV_R2 -command 3dttest++
-prefix ResultsA.tt.100s_vs_300s_AllV_R2
-dsets CorrZfiles/corr_brain.???_1??.Visit?.2.Z+tlrc.BRIK
-dsets CorrZfiles/corr_brain.???_3??.Visit?.2.Z+tlrc.BRIK
-set_labels 100s 300s
-options
-Clustsim -prefix_clustsim Clust.100s_vs_300s_AllV_R2 -mask TSNR_brainmask_files/clip_count_group_brainTissueMask_tSNR10_Visit1.2+tlrc

and got the following results:

++ Smallest FDR q [1 100s-300s_Zscr] = 0.044138 <==========++++++++ @ p = .001, but q = .7157, L Anterior Cingulate Gyrus (BA 24) size 14 voxels
++ Smallest FDR q [3 100s_Zscr] = 1.13601e-10
++ Smallest FDR q [5 300s_Zscr] = 3.00011e-10

  • Added 3 FDR curves to dataset
    ++ Output dataset ./ResultsA.tt.100s_vs_300s_AllV_R2+tlrc.BRIK
    ++ Output dataset ./Clust.100s_vs_300s_AllV_R2.resid.nii
  • 2-sample test: results are 100s - 300s

As noted, the smallest FDR q was 0.044138. However, when I look at the results in afni, there are no significant results at q=.05 (or .04 or .06). The only significant result > 10 voxels at p=.001 (with a q=.7157) is a cluster of 14 voxels which does not meet the clust size min of 38 for alpha = .001 and p= .05.

I’m wondering why the 3dttest++ command found the smallest FDR q = .04413, but I can’t find anything significant at anywhere near that level using the afni display.

2

Hello,

Are you clusting in the GUI? Those FDR results do not consider clustering. When the threshold is dragged up very high, the most strict FDR should be seen in the GUI.

If you are correcting for multiple comparisons at this point, it might be better to pass a group mask in the 3dttest++ command. Otherwise, the correction estimates will be across the entire volume.

  • rick
3

Okay, I see. Without any clustering, there is one voxel that does appear at q= .05. Thanks for the explanation.