clusterrandanalysis on scalar values + problem with topoplotER
Eric Maris
maris at NICI.RU.NL
Wed Aug 23 14:53:04 CEST 2006
Dear Marco,
You have posted this message once before. This was during my holidays, so I
did not reply then. When I returned from holidays, I forgot to send you a
reply. My apologies for this.
I am using Cluster Randomization Analysis (CRA) on scalar data and I would
like to be sure that the procedure is correct. Results are plausible, but
since this is not the canonical use of CRA, I would like to have your
feedback.
I have 9 subjects, two experimental condition for each subject, and 39
channels. For every subject, channel and condition, I computed an exponent
alpha of the power law that better fits the temporal long-range
autocorrelations in the data, i.e. a scalar value between 0.5 and 1.5. I
transformed the data in a format suitable to CRA by using eeglab2fieldtrip.m
to obtain the electrode structure, and by transforming the data structure by
hand to mimick the output of the tutorial data, CRA Within-Subjects case.
The result for each condition is a structure like this:
data1 =
label: {39x1 cell}
fsample: 500
elec: [1x1 struct]
cfg: [1x1 struct]
dimord: 'repl_chan_time'
individual: [9x39 double]
time: 1
I then performed CRA with the following parameters:
cfg=[];
cfg.statistic='depsamplesT';
cfg.alphathresh=0.05;
cfg.makeclusters='yes';
cfg.minnbchan=2;
cfg.neighbourdist=1;
cfg.clusterteststat='maxsum';
cfg.onetwo='twosided';
cfg.alpha=0.05;
cfg.nranddraws=500;
cfg.channel={'all'};
[clusrand]=clusterrandanalysis(cfg,data1,data2);
My questions are:
1) Is this procedure correct or are there specific parameters for scalar
data?
As far as I can see, this analysis is correct. However, these are not scalar
data, because you have observed the EEG on 39 sensors. If you had observed
the data on a single sensor only (and a single time point), then the data
would have been scalar (i.e., expressible as a single number for every
experimental condition).
2) Is there a way to compute the null distribution by considering all
possible reassignments of the conditions instead of the Monte Carlo
approximation - in this case it would be feasible because all possible
reassignments should be 2^9, right?
This is not possible with clusterrandanalysis. However, this is possible
with one of the new statistics functions in Fieldtrip that will replace both
clusterrandanalysis and sourcestatistics. The function of interest for you
is timelockstatistics, and you should have a look at the help information
for the lower-level function statistics_montecarlo (which is called by
timelockstatistics). However, provided that you use more than 500 draws from
the permutation distribution, don't expect major changes in the output.
Last point, when I ran the examples on the tutorial data in the Cluster
Randomization Analysis section, Within-Subjects case, I found out a little
error:
clusrand from clusranderfcpFICvsFCorig.mat has two fields with incorrect
dimensions that give errors when plotted:
stats: [151x1x301 double]
raweffect: [151x1x301 double]
while they should be
stats: [151x301 double]
raweffect: [151x301 double]
Yes, this is an inconvenient format for plotting. I will change it.
and a problem with plotting (once the little error corrected): I attach the
figure that I plotted by following the instructions in the tutorial. I am
using Matlab 7 and FieldTrip version 20060731. Can you help me with this?
My hunch is that these strange plotting results are a due to inappropriate
limits of the color axis. You should use the same limits for all plots in a
series. Also, choose the color limits symmetric around zero. By-the-way, I
was surprised to see a time axis in your plots. I thought you only had a
spatial dimension.
Good luck,
Eric Maris
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://mailman.science.ru.nl/pipermail/fieldtrip/attachments/20060823/7fe4b49d/attachment-0002.html>
More information about the fieldtrip
mailing list