[FieldTrip] Group-level analyses of single-subject cluster-based statistics

[Schoffelen, J.M. (Jan Mathijs)]

Ciao Andrea,

I have never thought about this in much detail, but naively I would say that what you sketch is not a good way to perform statistics.

Using an inferential statistical test with permutations and clustering for FWER correction, the clusters are just a byproduct of the inferential logic. Strictly, the only thing that can be derived from the test is a binary decision: ‘either or not reject the null hypothesis of exchangeability’. Any downstream stuff done with the clusters (i.e. strong interpretation of effects based on their location, or a second level test based on masked data) is in my view not allowed. See http://www.fieldtriptoolbox.org/faq/how_not_to_interpret_results_from_a_cluster-based_permutation_test for more information (also check out the references at the bottom of that page).

What you could do, however, is to think into the direction of using prevalence statistics to do your inference (to get started you could check Allenfeld et al. 2016, Neuroimage, their code is on github). In this approach, first level permutations are used to estimate a distribution of a test statistic under some null hypothesis, and group inference is done using this prevalence framework.

Perhaps Bruno and Robin might want to chime in here (preferably in Inglese), rumour has it that they have contacted you offline about this, but I think that the community would benefit from being able to read along.

Best wishes,

Jan-Mathijs

J.M.Schoffelen, MD PhD
Senior Researcher, VIDI-fellow - PI, language in interaction
Telephone: +31-24-3614793
Physical location: room 00.028
Donders Centre for Cognitive Neuroimaging, Nijmegen, The Netherlands




On 28 Feb 2019, at 17:37, Andrea Brovelli <andrea.brovelli at univ-amu.fr<mailto:andrea.brovelli at univ-amu.fr>> wrote:

Dear all,

we would like to perform group-level analyses of the results obtained from single-subject cluster-based statistics. In our case, each subject has a statistical map of size (time x channels) thresholded using cluster-based statistics method.

Given you expertise in the cluster-based methods, what do you think would be a good way to perform group-level statistics?

And could you point me to some tools, codes, papers please?

Or simply send a link of previous discussions on the mailing list? I cannot find them....

Thanks a lot.

Best,

Andrea

_______________________________________________
fieldtrip mailing list
https://mailman.science.ru.nl/mailman/listinfo/fieldtrip
https://doi.org/10.1371/journal.pcbi.1002202




-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.science.ru.nl/pipermail/fieldtrip/attachments/20190304/b3cbba52/attachment-0002.html>
-------------- next part --------------
_______________________________________________
fieldtrip mailing list
https://mailman.science.ru.nl/mailman/listinfo/fieldtrip
https://doi.org/10.1371/journal.pcbi.1002202