[Eeglablist] Reduction of Data Dimensionality before ICA

[Tarik S Bel-Bahar]

Hi Martin,

Good glad it helped a little.
Note that one thing you can do is eventually select only the the 5 or 10
ICs per participant that
look like real IC-cognitive activity. Just because you get 128 ICs back, it
does not mean
that you need to consider all of them. The most important ones should be
within the the top 30 as you said.

All the best,
Tarik


On Mon, Jan 14, 2013 at 4:23 AM, Krebber, Martin
<martin.krebber at charite.de>wrote:

>  Hello Tarik,
>
> thank you very much for your comments on PCA and ICA. It was all very
> helpful!
>
> I would agree that the ICA results are easier to interpret and more vaild
> without a previous reduction of the dimensionality of the data with PCA.
> For now, I decided to go ahead without the PCA.
>
> When I get the chance I might compare both methods and see for myself. I
> think the advantage of doing a PCA first is that there are fewer ICs to
> look at. In my experience most components beyond 30 (when sorted with
> respect to variance explained) don't look like physiologically plausible
> brain activity, so it's always a pain to look through all of them. But
> apart from that, I don't think there is a good reason to apply PCA in our
> case.
>
> Regards,
> Martin
>
>
>
> On 12.01.2013 01:24, Tarik S Bel-Bahar wrote:
>
> Greetings Martin,
>
> just a one small clarification point:
> I meant that not doing PCA should allow you to retain the true
> dimensionality of your dense EEG data,
> and doing ICA to this non-reduced data should lead to an ICA
> decomposition that more accurately
> reflects the true dynamics within your eeg data. Cheers!
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20130114/55eedcaf/attachment.html>