Fw: [FIELDTRIP] cluster randomization

[Eric Maris]

Dear Olga,


Am I right, that after cluster randomization I will everytime get smaller
amount of significant results, than after traditional t-test?
I mean if set cfg.alphathreshold = 0.05, and choose dependent t-test my
clusters can only include the channels and time-points that is significant
(p<0,05)
when I just use dependent t-test to compare the two condition time by time
on every channel. And only after this first step analysis the cluster is
formed based
on this t-test. Probably I am missing something, but after
clusterrandomization analysis I've got a cluster (even significant, but it
should not matter) with differential activation at posterior left site,
although t-test (made in eeglab) on this paticular channel does not show any
significant points.


Your reasoning is completely correct. As I see it now, either
Clusterrandanalysis or EEGlab performs its dependent samples t-tests
incorrectly. Without having your data, it is not possible for me to check
this numerically. However, you can do this very easily yourself: take the
data of the (channel,time)-pair of interest, and calculate the dependent
samples t-statistic on it using Matlab's statistics toolbox or by means of
the formula in an elementary statistics book. Compare the result with the
statistic for this (channel,time)-pair that is given in the output of
clusterrandanalysis (in the .stats-field). Let me what you have found.


Kind regards,

Eric Maris