[Eeglablist] What to do with more than one IC per subject in a cluster

[Makoto Miyakoshi]

Dear Chadwick,

I appreciate you are serious about it.

> I assume that the source of error used in the statistical test is the
between subject error, unless of course the authors explicitly state
this is a single-subject, in which case the source of error is the
between-trial error. This is convention, isn't it?

Yes you said it. I thought ICA-clustering is another convention since it's
been used for more than 10 years. There the errors are between-IC.

> Do you have any other suggestions?

Just let the data self-organized and see what's there. If you don't see
anything significant then I can help you.

>Thus, I have been manually editing my clusters so each cluster has
exactly 1 IC from each subject,
> Should I merge ICs from a subject when they are over-represented in a
cluster? Should I exclude subjects that are not represented in a cluster
and explicitly state that along with the result?

I would not do any of these. Just be confident.

For your information, I wrote a plugin (unreleased yet) for backprojecting
cluster ICs to each individual's channel(s) (e.g. Cz) so that I can have
one data per subject (Jason Palmer, a developer of AMICA, actually
recommended it). If you are interested in it let me know.

Makoto

2013/3/4 Chadwick Boulay <boulay at bme.bio.keio.ac.jp>

> Dear Makoto,
>
> I've been trying to do IC clustering but I guess I just don't get it.
>
> On 3/1/2013 8:42 AM, Makoto Miyakoshi wrote:
> >
> > If it were inappropriate then most of our publications would be bad :-)
> >
> > ICA is data driven and generates subspaces where ICs are
> > intradependent. It also reflects individual differences. This is like
> > cost for this nice method. However, what actually matter is
> > convention, isn't it?
> >
> >
> In general, reporting a statistically significant difference in a group
> of subjects implies that the source of error is between-subject and each
> subject is represented exactly once (for a given condition). If, in a
> peer-reviewed article, I see a statement such as, "P300 amplitude was
> greater after attended stimuli than unattended stimuli (p<0.05)" then I
> assume that the source of error used in the statistical test is the
> between subject error, unless of course the authors explicitly state
> this is a single-subject, in which case the source of error is the
> between-trial error. This is convention, isn't it?
>
> Since IC clusters may, at the extreme, represent a single subject then
> the above statement could really mean a single-subject difference even
> though many subjects were measured. To me, this breaks convention and is
> unintuitive.
>
> Thus, I have been manually editing my clusters so each cluster has
> exactly 1 IC from each subject, even if sometimes that means excluding
> an IC that fits my model if it comes from a subject with multiple
> clusters that fit the model or including an IC that does not fit the
> model if a given subject has no clusters that fit the model. I see no
> other way to report my results in a manner that follows convention.
> Editing clusters this way is biased but I'm not sure what else I can do.
>
> Do you have any other suggestions?
> Should I merge ICs from a subject when they are over-represented in a
> cluster? Should I exclude subjects that are not represented in a cluster
> and explicitly state that along with the result? i.e., "Left M1 cluster
> mu-rhythms were significantly more desynchronized during task A compared
> to task B among the 6 (of 10) subjects with an IC that localized to left
> M1 and had similar ERSP profiles."
>
> I read the MPT paper and tried using it but I ended up with an
> interesting domain that contained only 1 dipole (from a set of 13
> subjects x 64 ICs). I can't report that. I still don't understand this
> method fully so I do not admit defeat yet.
>
> Thank you,
> Chadwick Boulay
> JSPS Postdoctoral Fellow
> Keio University, Yokohama, Japan
> _______________________________________________
> Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
> To unsubscribe, send an empty email to
> eeglablist-unsubscribe at sccn.ucsd.edu
> For digest mode, send an email with the subject "set digest mime" to
> eeglablist-request at sccn.ucsd.edu
>



-- 
Makoto Miyakoshi
JSPS Postdoctral Fellow for Research Abroad
Swartz Center for Computational Neuroscience
Institute for Neural Computation, University of California San Diego
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20130307/3c617dea/attachment.html>