<!DOCTYPE HTML PUBLIC "-//W3C//DTD HTML 4.01 Transitional//EN">
<meta content="text/html; charset=ISO-8859-1"
<body bgcolor="#ffffff" text="#000000">
<font face="Calibri">Hi Shogo,<br>
Haven't earned that title yet ;). My apologies for my late reply.</font><br>
type="cite">In your 2nd step, I should compute
cluster-level-statistics from random permutation data.
Here, I have an question.
When I define clusters from random permutation data, which should
I define where clusters are the same place (e.g. time, sensor and
so on) as "observed clasters" or should I define newly where
clusters are from random permutation data regardless of the places
of "observed clasters"?
The idea is that for every random permutation of data, you calculate
your statistics and cluster over them. Then, select the largest of
these, whether it occurs at different time-points than your observed
clusters or not. The null-hypothesis that your are trying to refute
is not of the form "this cluster is significant is bigger than
zero", but of the form "the data is interchangeable between
conditions". Therefore, you gather 'largest-clusters' from 1000
random combinations of data. If your biggest cluster is bigger than
95% of the biggest clusters (alpha = 0.05, single-sided test), than
the data is <i>not </i>interchangeable, and thus significantly
different between conditions. This is exemplified by <i>all </i>clusters
that surpass the test based on the <i>same </i>permutation-distribution
of 'biggest-clusters'. Your data is different between conditions,
and all peaks in the mountain-range are representing this (if they
surpass the cluster-level test) (my favorite analogy). <br>
type="cite">I think the latter is right, this is OK?
Second, If I have interests in the cluster that has the second or
third... non first largest cluster-level-statistics from the
experimental hypothesis, how should I test these clusters?
All clusters are tested against the permutation-distribution of
'biggest-clusters'. Any of the smaller observed-clusters that are
still bigger than 95% of your distribution-of-biggest-clusters
(alpha = 0.05, single-sided), are part of <i><b>all</b> </i>the
clusters that <i><b>together </b>show that the data differs
<div class="moz-signature">-- <br>
<font size="3"><font color="darkblue"><font face="calibri">Roemer
van der Meij M.Sc.<br>
Donders Institute for Brain, Cognition and Behaviour<br>
Centre for Cognition<br>
P.O. Box 9104<br>
6500 HE Nijmegen<br>
Tel: +31(0)24 3655932<br>
E-mail: <a class="moz-txt-link-abbreviated" href="mailto:firstname.lastname@example.org">email@example.com</a><br>