1-vertex clusters in permutation cluster test with TFCE

[okapi (Viktoria)]

• MNE version: 1.6.0
• operating system: Arch Linux, Linux 6.0.11-arch1-1 #1 SMP PREEMPT_DYNAMIC Fri, 02 Dec 2022 17:25:31 +0000 x86_64 GNU/Linux

Dear all,

I’m trying to run permutation cluster test, on 5124 brain vertices and 121 time points. I’m interested in how the effect I observe at the sensor level is localized in the brain at different time points over the specified 121-length time interval. I use TFCE in this test to have more or less ‘objective’ measure of threshold.

This is an example of my code below:

threshold_tfce = dict(start=0, step=0.2)
adjacency = mne.spatial_src_adjacency(src)
T_obs_tfce, clusters, p_values, H0 = clu = \
    spatio_temporal_cluster_1samp_test(X_arr, 
                                        n_jobs=None,
                                        threshold=threshold_tfce,
                                        adjacency=adjacency,
                                        buffer_size=None,
                                        tail=1, n_permutations=1024)

I noticed that when I run such a test I get the output that there are 620004 clusters specified, which probably means that all these clusters are 1 vertex in size, as 5124*121=620004. I also observed that if I remove time dimension, there are still only 1-vertex clusters (5124 clusters in this case). My adjacency matrix should be fine, it worked with the same test without TFCE.

Sorry, I’m not an expert in this test, but I decided to ask as it looks a bit strange for me… Shouldn’t the clusters be represented not as 1 vertex clusters, but rather as a group of vertices?
When I run the same function without TFCE on the same data, I get only 22 clusters specified, and their size is greater than 1. And I can see how the clusters change in time in the brain when I visualize them. On the opposite, with TFCE, when I visualize all of the 1-vertex ‘significant clusters’ in time, despite all together they look like a cluster, at each individual time point there’s only one significant vertex. So at individual time points the picture seems meaningless…

I’m very sorry, maybe I misunderstand the TFCE concept of something. I would be very grateful for any clarification.

All the best,
Viktoria

[drammock (Dan McCloy)]

This is expected behavior. There is a (short) explanation in this tutorial section: Statistical inference — MNE 1.6.0 documentation. Basically, with TFCE you get a t-value per vertex/timepoint, instead of a t-value per cluster. The clusters are then defined as “contiguous groups of vertices/timepoints that are suprathreshold”. The adjacency and time_step params determine what counts as “contiguous”.

@okapi

with TFCE, when I visualize all of the 1-vertex ‘significant clusters’ in time, despite all together they look like a cluster, at each individual time point there’s only one significant vertex.

I’m not sure I quite follow this, but when you say “despite all together they look like a cluster” then maybe they are? This other tutorial has an example of how to find the significant clusters after TFCE (it’s sensor-space data, not brain, but the idea is the same): Visualising statistical significance thresholds on EEG data — MNE 1.6.0 documentation — significant_points = cluster_pv.reshape(t_obs.shape).T < 0.05

[okapi (Viktoria)]

Thank you a lot for your answer! Now it’s much more clear, probably I used another function to form the cluster after the permutation test. Maybe my visualization was a bit strange because of that. I’ll check with what you suggested, thanks again!