[Eeglablist] re referencing to average after ICA

[Raquel London]

Dear Tarik,

Thank you for your detailed reply and reading advice, I appreciate it. I
have a follow-up question if that's OK. I have some experience in EEG and
eeglab / matlab, but relatively new to ICA. I've just decided to reassess
my pre-processing protocol and try to come to a deeper understanding of
what are good practices and why.

You write; "The usual step is apply the weights to the continuous data that
is essentially the same as the data that went into ICA."
With *the same data* you mean the same time points? If I understand that
correctly, then I should run ICA on continuous 1Hz HP filtered data and
then apply the weights to the unfiltered continuous data?

Thanks!
Raquel

On Fri, Sep 2, 2016 at 7:47 PM, Tarik S Bel-Bahar <tarikbelbahar at gmail.com>
wrote:

> Greetings Raquel, some quick notes below, best wishes.
>
>
>
> *****************************************************
> Some drift will be attenuated with 1hz filter, but that in itself does not
> require dropping of any time periods, unless you are referring to
> abnormally extreme drift within a trial. However in my experience, removing
> artifactual ICs does attenuate some slow artifacts.
>
> If you haven't had a chance to yet, please google past eeglablist
> discussions about filtering, ICA, and applying the weights to unfiltered
> data.
> In particular, i don't think there was a satisfactory conclusion on
> whether it's okay to apply the weights from ICA to unfitlered data.
> The usual step is apply the weights to the continuous data that is
> essentially the same as the data that went into ICA.
>
> Note that ICA won't necessarily catch slow-drift spatial maps, and that
> the usual thing before ICA is to get rid of low-frequency "slow" drifts as
> artifacts.
>
> I would not leave drifts in, and/or be very clear about how I am
> controlling them, or eliminating their effects (when for example, using a
> baseline, or creating contrasts within-subjects).
>
> Consider several possible strategies for drift artifact control, such as
> 1hz lowpass, demeaning, or detrending. These are all mentioned in past
> eeglablist discussions, and often used in published papers.
>
> If you're just starting up with EEG, just use the "basic" steps from
> eeglab tutorial and makoto's pipeline. After that, use methods from
> high-impact journals by respect EEG researchers focused on a protocol that
> is similar to yours.
>
> However, you can't hurt things by processing the data with and without the
> drifts, and seeing for yourself the effects.
>
> Remember that baselining your epochs/trials and similar things like that
> are also filters for the data. However a 1hz lowpass filter does help ICA
> get better results.
>
> To learn more about various details and caveats.....
> Google-Scholar some of the recent methods articles about baselining EEG
> (for ERPs) from Luck and others.
> Google-scholar the  "usual" methods in recent papers that use ICA for EEG
> with eeglab. This should give you a good idea of what happens with drifts
> and ICA usually.
> See also Mike X. Cohen's book on time-frequency analyses for extensive
> discussions of multiple points you are asking about (e.g., effects of
> preprocessing on time-frequency results, etc...). I believe there is a
> one-day pre-conference in Minnessota at the upcoming SPR by Mike that one
> can also attned.
> ************************************************************
> ***************************
>
>
>
>
>
>
>
>
>
>
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20160903/110ec0e4/attachment.html>