Thresholding/interpreting TFCE

[clare....@gmail.com]

Hi everyone,

I've recently completed my group-level searchlight analyses using threshold-free cluster enhancement (TFCE), and I've viewed the z-maps using Mango.

I was wondering if I could ask other people's experience/best practice with thresholding and interpreting the results? I'm assuming that as I have applied TFCE that I only need to threshold at voxel-level rather than cluster-level.

How have others decided the z-statistic to threshold at? I'm aware there's likely not a single specific answer, but as I'm quite new to TFCE it would be great to hear more about others' practices with RSA.

Similarly, if anybody has any papers that they could refer me to I'd really appreciate it! I've read up on the principles of TFCE but don't feel confident in applying this to make a decision alone.

Thanks in advance!

Clare

[Nick Oosterhof]

Hi Clare,

On Tue, 6 Jul 2021 at 13:43, clare....@gmail.com <clare....@gmail.com> wrote:

I've recently completed my group-level searchlight analyses using threshold-free cluster enhancement (TFCE), and I've viewed the z-maps using Mango.

I was wondering if I could ask other people's experience/best practice with thresholding and interpreting the results? I'm assuming that as I have applied TFCE that I only need to threshold at voxel-level rather than cluster-level.

Indeed you can threshold at the voxel level.

 

How have others decided the z-statistic to threshold at? I'm aware there's likely not a single specific answer, but as I'm quite new to TFCE it would be great to hear more about others' practices with RSA.

Similarly, if anybody has any papers that they could refer me to I'd really appreciate it! I've read up on the principles of TFCE but don't feel confident in applying this to make a decision alone.

The z-scores correspond to p-values corrected for multiple comparisons using 'standard' null hypothesis statistical testing, for example z=1.65 means p=0.05 one-tailed, and z=1.96 means p=0.05 two-tailed.  A threshold of p=0.05 is to my knowledge the most common threshold for publications in the field, although some have argued for stricter thresholds such as p=0.005. 

Interpretation: a voxel (or more generally a feature) surviving multiple comparison correction does not mean that that voxel itself alone is (statistically) significant, but that the TFCE score of that voxel (which also depends on the values of its neighbouring voxels) is, at the group level,  significant. 

Does that help?

best,

Nick