3dClustsim_ voxelwise p-value

Hi, all

I have question about deciding voxelwise p-value by using 3dClustsim -acf.
According to AFNI handout(Clusterizing), we need to use voxelwise p-value which is less than 0.002 with mixed ACF.
Is it acceptable to use voxelwise p=0.01 or other bigger value? About this issue, if you have a reference, could you let me know?




I don’t believe voxelwise thresholding with p=0.01 (or larger) by itself is acceptable anymore in the FMRI community, given the apparently larger FPRs in some tests. See for example Eklund et al. (2016), as well as Cox et al. 2017:


We estimate the cluster wise threshold using 3dClustsim with ACF options, and look at the results at voxelwise p=0.01, the cluster size threshold is estimated by 3dClustsim with ACF options. In that case, is it okay that we still look at p=0.01?

Hi, Jung-

While what you are saying you did is certainly consistent, I don’t believe it would be accepted in the literature-- there is a general sense that the goodness of false positive control at p=0.01 is not high. That is, one sets the voxelwise p=0.01 and then the clusterwise alpha=0.05 (with the latter meaning that we are asking to have a nominal 5% FPR with our cluster results), but that the results won’t really have good control and 5%FPR-- it might have something more like 10%FPR or higher. That is the idea.

The standard is really to use a voxelwise p=0.001 threshold (though p=0.002 does look reasonable, too), with the thought that when one asks for final cluster results with an alpha=0.05 (and hence nominal FPR of 5%), the output will be much more likely to actually have around a 5% FPR.

Note that there are further thoughts on this, reducing the dependency of a single p-value and still controlling overall FPR, with the ETAC methodology proposed by our veeeeeeeeeery own Bob Cox; you can read more about it here:




See if this approach is feasible for you: https://afni.nimh.nih.gov/afni/community/board/read.php?1,157054,157054#msg-157054

The standard is really to use a voxelwise p=0.001 threshold (though p=0.002 does look reasonable, too)

→ p=0.002 looks reasonable. Everyone knows p=0.001 is fine. But p=0.002 looks unusual, so is there any reference or proof for using p=0.002?