[Eeglablist] scree test for ICA components?

[Siri-Maria Kamp]

Dear all,

I have one question about an ICA that I am currently working on. I have
subject averages from 24 participants and 2 experimental conditions, each
including 163 data points. The data are from 129 electrodes. There are two
problems this ICA: First, I do not have enough data points to get 129
reliable ICs. Second, the activations of my ICs of interest are not
significantly different between conditions when I extract all 129
components (I DO get significant differences when preprocessing the data
with PCA and keep, for example, 10 components). Therefore, I would like to
reduce the number of components to extract.

The question is: How do I determine how many components I should keep? I
know that this is a frequently asked question, but I still did not find an
answer that actually helped me.

One possibility seems to be conducting a scree test, which is commonly
used in PCA, and which has been used in ICA as well (Dien, J., Khoe, W.,
Mangun, G. R. (2007). Evaluation of PCA and ICA of Simulated ERPs: Promax
vs. Infomax Rotations. Human Brain Mapping, 28, 742-763). I wonder how
suitable this test is for ICs, because in contrast to PCA, ICA extracts
all components in parallel. Thus, if I do a scree test on an ICA with 129
components, does the information from this scree test "transfer" to ICAs
with a lower number of components? In other words: can I really decide
from an ICA with lots of components, how many of these components are
"meaningful", "interpretable", or however one wants to call it?

Thank you so much for your help. I've been discussing this issue with
several people for some time, but nobody seems to know a good answer.

Siri Kamp





-- 
Siri-Maria Kamp, BSc
Graduate Student
USF EEG laboratory
Department of Psychology
University of South Florida