Problem with data from BESA

Michael Wibral wibral at MPIH-FRANKFURT.MPG.DE
Wed Nov 23 13:26:13 CET 2005

Hi Eric,

thank you for the explanation, things are much clearer now. In the
meantime I have encountered another problem with clusterrandanalysis:
When using maxsum as a test statistic everything works fine, but using
maxsumtminclustersize with the same specified (maximum) alpha the fields
clusrand.posclusters and clusrand.negclusters stay empty - although
there seem to be large enough clusters in posclusterslabelmat and
negclusterslabelmat (I used cfg.smallestcluster=2..) . Is this a bug or
does the use of  maxsumtminclustersize somehow reduce sensitivity (-
from the description in the 2005 tutorial I thought it is similar to
using FDR or 'Holmes' method for computing critical p values?). Maybe I
am also missing something on a conceptual level that makes the
information in posclusters invalid if I use maxsumtminclustersize as a
test statistic ??
Below I pasted the code that produced this behaviour.

%Clusterrandomization analysis
cfg.elec  =elec;
cfg.statistic = 'depsamplesT';
cfg.alphathresh = 0.05;
cfg.makeclusters = 'yes';
cfg.minnbchan = 1; %1 neighbour i.e. 2 channels
cfg.smallestcluster = 2;
cfg.clusterteststat = 'maxsumtminclustersize'; % replace with maxsum to
get lots of entries in clusrand.posclusters
cfg.onetwo = 'twosided';
cfg.alpha = 0.05;
cfg.nranddraws = 1000;
cfg.latency = [0.40 0.65];
[clusrand] = clusterrandanalysis(cfg, AMdata, vAMdata);


Eric Maris schrieb:

> Hi Michael,
>> I have however another question regarding the interpretation of
>> clusteranalysis results. Am I correct in saying that the family wise
>> error rate (alpha) tells me the risk in obtaining a false positive
>> statement of the type that I specify previously with alphatresh? For
>> example if I specify alphathresh of 0.1 (lets calls this trend for
>> abbreviation) in the first pass of the analysis (multiple testing)
>> before clustering then the clusterrandomization using alpha =0.05
>> tells me that I run a risk of 5% of identifying wrongly at least one
>> of these 'trend clusters'.
>> (Or else, if the above is incorrect what is the reason not to use a
>> very lenient criterion in the first pass to feed the
>> clusterrandomization with as many clusters as possible?)
> The issue is statistical power (sensitivity). If you use a very
> lenient criterion (say, alphathresh=0.2) to select candidate cluster
> members, this will result in large clusters purely by chance. If the
> effect in your data is strong but confined to a small number of
> sensors and timepoints, clusterrandanalysis may not pick it up. This
> is because the reference distribution is dominated by these weak but
> large "chance clusters". You will not encounter this problem if you
> put alphathresh higher. On the other hand, a high aphathresh will miss
> weak but widespread effects.
> To sum up, alphathresh determines the relative sensitivity to "strong
> but small" and "weak but large" clusters.
> greetings,
> Eric Maris
> .

More information about the fieldtrip mailing list