[Eeglablist] ICA

Arnaud Delorme arno at sccn.ucsd.edu
Tue Jul 15 15:12:04 PDT 2003

>Hello, I have two (rather basic) questions about ICA, i hope anyone can help me out. I am about to use ICA for artifact-correction in a discourse-comprehension paradigm. Because our critical words follow each other closely, there's quite some overlap in the segments that i feed the ICA. Does anyone know whether this constitutes any kind of problem for the (interpretation of) the resulting components? 
If there is a large overlap of data segments, ICA will weight more 
duplicated data points. In practice, it usually does not prove to be a 
problem. However, you might want to extract non overlapping sub-epoch 
from your long data epochs, apply ICA and then export the ICA 
decomposition back to the dataset with the longer epochs. See the end of 
paragraph I.9.4 of the tutorial for how to do that


Also, since ICA is a statistical algorithm, you must have more different 
data points than values in the ICA weight matrix (as a rule of thumb, 
for 64 electrodes, the ICA weight matrix contains 4096 points so you 
need more than 4096 points in your EEG data (multiply this number at 
least by 3 to account for the high correlation in time of the EEG 
activity ...)).

>The other question is about rereferencing. We have online referenced to M1, and want to offline rereference to (M1&M2)/2. My intuition tells me it's better to rereference after ICA-based correction, since otherwise you introduce a location-independent (non-informative) contribution into your data, but im not sure about it. 
>If someone could help me with these issues, i would be very gratefull,
>yours sincerely,
Theoretically, it does not matter to rereference before or after ICA. 
Since rereferencing is a linear operation, it does not affect the result 
of the ICA algorithm. Note it might if you rereference and then subtract 
the baseline from each trial. We believe it is better to have a more 
accurate baseline before running ICA (so components do not have to 
account for artifactual DC shifts in the baseline), so if you think 
rereferencing will provide a better baseline, then do so before running ICA.


>Drs. Mante S. Nieuwland
>University of Amsterdam
>Dept. of Psychology (PN)
>Roetersstraat 15, kamer 614                     
>1018 WB Amsterdam                                    
>The Netherlands 
>mnieuwland at fmg.uva.nl
>phone:   +31-20-5256808
>fax:   +31-20-6391656
>Eeglablist mailing list Eeglablist at sccn.ucsd.edu
>Eeglablist page: http://www.sccn.ucsd.edu/eeglab/eeglabmail.html
>To unsubscribe, send an empty email to eeglablist-unsub at sccn.ucsd.edu


* Arnaud Delorme <http://www.sccn.ucsd.edu/%7Earno>, Ph.D.* , SCCN, 
UCSD, San Diego, USA

*SCCN main page:* www.sccn.ucsd.edu <http://www.sccn.ucsd.edu/>
*EEGLAB main page:* www.sccn.ucsd.edu/eeglab/ 
*EEGLAB documentation:* www.sccn.ucsd.edu/eeglab/eeglabdocs.html 
*EEGLAB FAQ page:* www.sccn.ucsd.edu/eeglab/eeglabfaq.html 
*EEGLAB mailing lists:* www.sccn.ucsd.edu/eeglab/eeglabmail.html 

-------------- next part --------------
An HTML attachment was scrubbed...
URL: http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20030715/5c6fdce8/attachment.htm

More information about the Eeglablist mailing list