Add that minimum number of channels for defining cluster is 0 in documentation

[kerrencasper]

Proposed documentation enhancement

In mne.stats.permutation_cluster_test there is no way of defining the minimum number of neighbouring points that defines a cluster. This is a major error in the code, which makes the code invalid. This has been discussed by many of the developers for years, but nothing has been done (see issue #10604, #10604). Most people 1) don't know why this is a problem; 2) blindly trust the function. The least you can do if you are not ready to change it is to add to the documentation that this is a shortcoming of the function and that it should not be used until it has been solved. To exclude clusters with less than n number of channels (as has been discussed in the above-linked thread) is not a valid way forward as the t-values for these clusters have already been part of the t-distribution for defining the threshold for what can be considered significant.

[welcome]

Hello! 👋 Thanks for opening your first issue here! ❤️ We will try to get back to you soon. 🚴

[larsoner]

Hmm... I think our minimum number is effectively 1, which should make it so things are valid, right?

[sappelhoff]

I think the answer to this issue lies in this comment: #10604 (comment)

Specifically this part:

For data that are sufficiently smooth I don't think it shouldn't matter much in practice because single-node "clusters" will have very small counts and stat values (which you get/use depends on t_power) so won't be the largest cluster(s) for each permutation used in the end for the maximal statistic.

I agree that in most cases it probably does not matter in practice, because data for which one would use cluster perm tests would usually also be "sufficiently smooth".

However, pre-defining the cluster extent could be a nice feature (currently hardcoded to "0", that is, zero neighbors required ... so one could also say "1" significant node). It exists in fieldtrip, see: https://www.fieldtriptoolbox.org/tutorial/stats/cluster_permutation_timelock/#the-configuration-settings

cfg.minnbchan = 2; % minimum number of neighborhood channels that is
% required for a selected sample to be included
% in the clustering algorithm (default=0).

☝ note how also in fieldtrip, the default is 0.

. To exclude clusters with less than n number of channels (as has been discussed in the above-linked thread) is not a valid way forward as the t-values for these clusters have already been part of the t-distribution for defining the threshold for what can be considered significant.

☝ this is true ... excluding clusters posthoc (AFTER the clustering) through masking would not be valid.