[Eeglablist] Using PCA before ICA decomposition?

[Simon-Shlomo Poil]

Hi Chang,

Yes. When you do PCA before ICA it reduces your rank. So with your example

> However, If I first do PCA 35 then ICA ( it returns 35 ICs) and remove 20
> ICs, do I only have 15 ranks left or I still have 108 ranks for the cleaned
> data? If I want to do ICA again, what number I should put in ['pca' ...]? Is
> the PCAed data the same as the raw data, but only with a linear transform,
> or EEGlab change the data at PCA step by deleting the small PCs and then
> conduct ICA on the first 35 PCs?

then you have a rank of 15. This means that, if you want to run ICA
again, you have use ['pca', 15] in your call.

Best,
-- 
Simon-Shlomo Poil

Neuronal Oscillations and Cognition Group (NOC)
Department of Integrative Neurophysiology (INF)
Center for Neurogenomics and Cognitive Research (CNCR)
Neuroscience Campus Amsterdam
VU University Amsterdam
De Boelelaan 1085, Room B-435
1081 HV Amsterdam, The Netherlands

E-mail: simonshlomo.poil at cncr.vu.nl
Phone: +31 20 5989408
Webpage: http://www.poil.dk/s and http://www.nbtwiki.net

2012/1/17 Chang Gu <chang.gu at vanderbilt.edu>:
> Hi all,
>
> I'm  conducting ICA to clean my segmented EEG data.   Because the length of
> my data is short, I have to use PCA option (by choosing ['pca', 35] in
> EEGlab, or EEG = pop_runica(EEG,
> 'icatype','runica','dataset',1,'options',{'pca' 35});) before doing ICA
> decomposition.
>
> I wonder what is the number of the ranks for my data after the PCA by
> EEGLAB?
>
> For example, I have 128 channels for my recording, so without PCA, I should
> have 128 ICs corresponding to 128 linear independent time-series. If I
> remove some eye components (say 20 ICs) to clean the EEG, I should still
> have 108 ranks of my cleaned data, and in theory I should be able to conduct
> PCA 108 then  ICA recursively, by choosing ['pca', 108] in runica.
>
> However, If I first do PCA 35 then ICA ( it returns 35 ICs) and remove 20
> ICs, do I only have 15 ranks left or I still have 108 ranks for the cleaned
> data? If I want to do ICA again, what number I should put in ['pca' ...]? Is
> the PCAed data the same as the raw data, but only with a linear transform,
> or EEGlab change the data at PCA step by deleting the small PCs and then
> conduct ICA on the first 35 PCs?
>
> Thanks in advance!
>
> --
> Chang Gu
> Psychology & Human Development
> Vanderbilt University
> Nashville, TN
>
>