[Eeglablist] Why average reference for ICA?

[Makoto Miyakoshi]

Dear Andreas,

> Do you mean by this that you should not re-reference data after
preprocessing with ICA at all? Or do I misunderstand this point?

Technically you can still do it as long as you know what you are doing.
However, generally speaking, changing data rank or number of channels after
ICA causes troubles.
In short, performing (standard) average referencing (i.e., make the data
zero-sum across channels for all data points) by definition causes rank
reduction. Then your data will have [number of ICs] > [channel data rank]
which is a strange situation (because ICA cannot yield more components than
data rank).

https://sccn.ucsd.edu/wiki/Makoto's_preprocessing_pipeline#Rejecting_ICs_to_improve_the_next_ICA.3F_Or_what_is_rank_and_linear_algebra_.2807.2F16.2F2018_updated.29

I will disappear until November, and this is my last reply to you.

Makoto

On Thu, Aug 2, 2018 at 12:27 AM Andreas Pedroni <andreas.pedroni at uzh.ch>
wrote:

> Dear Makoto
>
> Thank you very much for your reply. That makes all sense. However, I don’t
> understand why performing average referencing after ICA is not recommended?
> Do you mean by this that you should not re-reference data after
> preprocessing with ICA at all? Or do I misunderstand this point? Thank you!
>
> All the best
> Andreas
>
>
> Am 31.07.2018 um 19:16 schrieb Makoto Miyakoshi <mmiyakoshi at ucsd.edu>:
>
> Dear Andreas,
>
> > Is there a particular reason for this? Is there anything wrong in doing
> ICA on EEG referenced to a single channel?
>
> Average referencing is to impose zero-sum assumption to channel data. ICA
> result does not get affected by this, except DC differences. In other
> words, if you don't apply average reference, some of your IC scalp
> topographies could be all read or all blue.
>
> Overall, I would rate the importance of average referencing before ICA not
> very high. It's just a minor issue even if you forgot to do so, most of
> your IC scalp topos are fine. However, if you perform average referencing
> after ICA, that may cause various problems so not recommended. If you
> forgot to perform average reference and finished lengthy ICA process which
> you can't do again for short of time, just forget it and proceed with what
> you have.
>
> > my idea was to compute the ICA on 1 Hz filtered (and average referenced)
> data and apply the selection of ICs to the same unfiltered (and average
> referenced) data.
>
> If you mean you copy the ICA weight matrices to do so, yes it is a common
> practice.
>
> Makoto
>
>
>
> On Thu, Jul 26, 2018 at 2:29 AM Andreas Pedroni <andreas.pedroni at uzh.ch>
> wrote:
>
>> Dear Makoto, dear EEGLAB list
>>
>> I came across several sources where it is suggested to use average
>> referenced data to perform ICA (including your very helpful discussion
>> about preprocessing
>> https://sccn.ucsd.edu/wiki/Makoto%27s_preprocessing_pipeline
>> <https://sccn.ucsd.edu/wiki/Makoto's_preprocessing_pipeline>). Is there
>> a particular reason for this? Is there anything wrong in doing ICA on EEG
>> referenced to a single channel?
>>
>> In addition, I have the goal to apply ICA on (1 Hz filtered) data with
>> the purpose to get rid of artifacts (using MARA), but apply the clean ICs
>> on unfiltered data to keep lower frequencies than 1 Hz. Based on your
>> comments, my idea was to compute the ICA on 1 Hz filtered (and average
>> referenced) data and apply the selection of ICs to the same unfiltered (and
>> average referenced) data. Is that the way to go, or are there other
>> suggestions?
>>
>> Thanks and all the best
>>
>> Andreas
>>
>> Andreas Pedroni
>> Psychologisches Institut
>> Methoden der Plastizitätsforschung
>> Andreasstrasse 15, AND 4.58
>> CH-8050 Zürich
>> andreas.pedroni at uzh.ch
>>
>>
>>
>> _______________________________________________
>> 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
>
>
>
> --
> Makoto Miyakoshi
> Swartz Center for Computational Neuroscience
> Institute for Neural Computation, University of California San Diego
>
>
>

-- 
Makoto Miyakoshi
Swartz Center for Computational Neuroscience
Institute for Neural Computation, University of California San Diego
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20180802/a4039acb/attachment.html>