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

[Arnaud Delorme]

A word of caution here "Is it inappropriate to include more than one component from the same subject in the analysis".

Yes and no.
No because your Null hypothesis is modified. When you include more than one component from the same subject, you are not making inference about the general population of subjects any more but instead about components of the specific subjects you are studying. It is all a matter of how many components you have per subject compared to the number of subject. For example, if you have on average 1 component per subject (some subjects having 0, some other 2 component in the cluster), and you have 200 subjects, then the original null hypothesis (which allows to make inference about the general population of subject) is mostly preserved. If you have 10 subjects and 10 components per subject, it is not.

In general, I prefer either (1) to use 1 component per subject per cluster because this avoids having to compromise with the statistics or (2) merge components from the same subject in the cluster before performing any statistics. In (2), if you have for example 2 components for one subject, the ERP of both components will be pooled (pondered by the ICA inverse weight matrix). Then each subject will have either 1 or 0 ERP for each condition (if you have different conditions) and you may perform group statistics on these ERPs. This way you also have at most 1 component per subject - although when you pool components, it is not really an ICA component anymore. Unfortunately, there is no automated way to perform (2) in EEGLAB at the moment.

Using several components per subjects is fine as well as long as you are aware that you are slightly compromising your null hypothesis (see above).

Best,

Arno

On 27 Feb 2013, at 17:17, James Schaeffer wrote:

> Dear Eeglablist,
> 
> After clustering components using k-means, some of my clusters contain more than one component from a single subject.  I want to compare ERSPs using permutation analysis.  Is it inappropriate to include more than one component from the same subject in the analysis? If so, what is a good method for selecting which components to use?
> 
> Thanks for your help,
> James
> 
> _______________________________________________
> 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