[FieldTrip] within-subject partition for between-subject stats

Stenner, Max-Philipp max-philipp.stenner at med.ovgu.de
Thu May 15 17:09:18 CEST 2014

Dear fieldtrippers, 
I have a question regarding permutation statistics in a mixed design with only very few subjects (5 subjects, 2 conditions/subject; this follows up on my email yesterday). Since the number of unique permutations of *trial-averaged* data is limited (2^5=32), would it make sense (is it statistically sound) to re-compute trial-averages for each subject after randomly partitioning trials *within* each of the five subjects for a large number of iterations first (since there are >>5 trials/subject the number of unique permutations is much larger than for trial-averaged data) and then re-compute cluster-statistics (e.g. maxsum) for each of these sets of re-computed trial-averages to obtain a null-distribution?
Thanks very much in advance, 
Von: fieldtrip-bounces at science.ru.nl [fieldtrip-bounces at science.ru.nl]" im Auftrag von "Stenner, Max-Philipp [max-philipp.stenner at med.ovgu.de]
Gesendet: Mittwoch, 14. Mai 2014 16:25
An: FieldTrip discussion list
Betreff: [FieldTrip] (no subject)

Dear fieldtrippers,
I was planning to use cluster-based permutation statistics for group-level inference on the difference between two conditions (depsamplesT) for a fairly small number of subjects (as few as 5). However, this would mean a total of only 2^5=32 possible permutations (=discrete values for the non-parametric null distribution), and therefore a maximum "p-value-resolution" of only 1/32=.0313. Does this mean that group-level permutation statistics are not appropriate for small sample sizes? How do people  correct for MC in these cases (using fieldtrip)? (I am specifically interested in group-level analyses, not primarily in within-subject/across-trials testing).
Thanks in advance,
fieldtrip mailing list
fieldtrip at donders.ru.nl

More information about the fieldtrip mailing list