[FieldTrip] maximum statistic in cluster correction

[van Lier Ben]

I think I just had my aha-moment. Or actually more of a d'oh-moment.

I was already wrong at step 0 which led to a whole cascade of thinking errors. I somehow thought that the t statistics were calculated per channel for a whole trial (so a single comparison) instead of every timepoint within a trial.

every one of those t values that reaches a threshold is considered to be part of a cluster. then sum the t values within each cluster and get the maximum cluster stat. this is your observed statistic.

now mix the trials and randomly assign them to subset1 and subset2. repeat the above to get another max cluster stat. do this many times (monte carlo) to create the permutation distribution of the max cluster stat.

compare all the cluster stats you actually observed originally with your permutation distribution to find their p values. the FA rate is controlled because no other cluster will show a smaller p value than the max stat cluster and the chance on a type I error for the max cluster stat is alpha in "his own" distribution. not completely sure about that last sentence but i think its time for weekend now :)


Thanks Eric

Cheers,
Ben