[Eeglablist] Advice on filtering strategy before ICA (EEGLAB pipeline; FAA, beta/alpha ratio/ PSD)

Mon May 18 20:03:51 PDT 2026

Hi Kathryn,

If you have 64-ch datasets, use an average reference.
In the case of Biosemi, if I were you, I would choose Cz as the initial
reference. This makes Cz continuous zeros, and your data rank is nbchan-1.
Then apply average reference including Cz. This does not change the data
rank, and now Cz has data back. If you want to make your data full rank
again, reject ANY one channel (e.g., one of EOG channels!) If you do not
use ASR, GEDAI etc., it is ok to have rank deficient data. Even for infomax
ICA, such clear rank deficiency will be detected and handled appropriately.
The only problem confirmed is the case of spline channel interpolation.

If considering these things is too much, you can use your original code you
pasted. Rejecting 2 out of 66 electrodes (and they are mastoids) must be
totally ok.

Makoto

On Mon, May 4, 2026 at 9:53 AM Kathryn Bolton <kathryn.bolton at torontomu.ca>
wrote:

> Hello Makoto,
>
> Thank you so very much for your advice and sharing those resources with
> me! I have a follow-up question regarding referencing to the mastoids: in
> my script I re-reference to the mastoids, but then remove them from before
> the ICA (script pasted below). I do all this before the high-pass filter
> and downsampling. I'm using a 64-channel BioSemi, so I could try your
> suggestion of using Dezhong Yao's REST, but what might that look like in my
> script?
> % 3. Re-reference to channels 65 and 66
> EEG = pop_reref(EEG, [65 66]);
> % 4. Keep only scalp channels 1 through 64
> % This removes the reference channels and any other extra sensors
> EEG = pop_select(EEG, 'channel', 1:64);
> % Save channel locations for later interpolation
> original_chanlocs = EEG.chanlocs;
> save([filename '_chanlocs.mat'], 'original_chanlocs');
> With many thanks, Kathryn
>
> *Kathryn Bolton, MA*
>
> Pronouns: she/her
>
> PhD Candidate | Clinical Psychology
>
> Department of Psychology | Cognitive Aging Lab (CAL)
>
> Toronto Metropolitan University
>
> *E-mail:* *kathryn.bolton at t <kathryn.bolton at ryerson.ca>orontomu.ca
> <https://urldefense.com/v3/__http://orontomu.ca__;!!Mih3wA!DvoqGZaNMAX6PWBKD1OkS_noSQPTe78PYfHQe_Ywu0CqcnwcUuNMfyeIxPWt6rVImWOUKmw3RB3xwHQOYYZEgi7CLWMfEhs$>*
>
> *Website*: *https://urldefense.com/v3/__https://psychlabs.torontomu.ca/cal/portfolio/kathryn-bolton/__;!!Mih3wA!Hh0u3SPT8LQUEV97Rkxfw9llZ4XtmaOUPFxereKz5F7uZBgp5HDHgHw_uTrM4SWT3OP4Kr5n62DcyRxWHNg2MkoUtYw$ 
> <https://urldefense.com/v3/__https://psychlabs.torontomu.ca/cal/portfolio/kathryn-bolton/__;!!Mih3wA!DvoqGZaNMAX6PWBKD1OkS_noSQPTe78PYfHQe_Ywu0CqcnwcUuNMfyeIxPWt6rVImWOUKmw3RB3xwHQOYYZEgi7C3kS363U$>*
>
>
> *Due to the volume of emails I receive, I will do my best to respond
> within 2 business days. If you require an immediate response or if it is an
> emergency, please put "urgent" in the subject line. Thank you for your
> understanding. *
>
>
> On Wed, Apr 29, 2026 at 8:01 PM Makoto Miyakoshi via eeglablist <
> eeglablist at sccn.ucsd.edu> wrote:
>
>> Hi Kathryn,
>>
>> But my question about filtering is whether you recommend applying only a
>> high-pass
>> filter prior to ICA (as above), and then applying the full bandpass (e.g.,
>> including a low-pass) later, prior to PSD analysis?
>>
>>
>> Andreas Widmann answered that question once and for all in this 2015
>> paper.
>> 10.1016/j.jneumeth.2014.08.002
>>
>> The answer is, apply high-pass and low-pass separately, with an
>> (inevitably) short transition band width (TBW) to high-pass filter, and a
>> gentle TBW for low-pass filter.
>> For example, if you want to set a band-pass filter 0.5-55 Hz @-6dB, then
>>
>> Highpass: cutoff 0.5 Hz (TBW 0.5 * 2 = 1.0)
>> Lowpass: cutoff 55 Hz (TBW 5, which makes pass-band edge at 55-5/2 = 52.5
>> Hz, and maximum suppression starts at 55+5/2 = 57.5 Hz)
>>
>> For confirming definitions of terminologies, see this page. This page is a
>> summary of my questions and his answers.
>>
>> https://urldefense.com/v3/__https://eeglab.org/others/Firfilt_FAQ.html__;!!Mih3wA!HPZkz51i3HC_zP7sK-ZAnbc3e7bwYY3dwxbgLZn8v25vAVwmBxhdk6YfF87fgdUWoKXpUNspBYhMnKmFvicAeN5kvEY$
>>
>> The reason why you want to do this is that you do not want to use an
>> unnecessarily steep slope in the low-pass end, which makes the filter
>> order
>> unnecessarily long.
>>
>> Re-reference to linked mastoids prior to ICA (BioSemi recommendations)
>>
>> My advice: do not use linked mastoids. Physically linked mastoids/earlobes
>> is out of the question, this is the worst because electrically your
>> subject's head is no longer shaped like a human head (heads are connected
>> bilaterally)!!! Digitally linked mastoids/earlobes is also problematic for
>> the following reason (from *Electric Fields of the Brain* p. 294, Nunez
>> and
>> Srinivasan 2006).
>>
>> The mathematical linked-ears or linked-mastoids reference may seem to
>> provide a "solution" to these problems associated with the physical linked
>> reference. However, does the original goal of using the average potential
>> of the two ears as the reference make sense? The motivation for this
>> procedure is not based on any physical properties of head volume
>> conduction. The measured potential difference between a pair of electrodes
>> depends on sources located near both electrode positions. With the
>> mathematical linked-ears reference, the recorded potential depends on
>> sources at three different locations. This approach may further complicate
>> the interpretation of scalp potentials and possible source locations
>> rather
>> than simplifying it. One rationalization often argued to support the
>> linked-ears or linked-mastoids reference is its purported tendency to be a
>> "symmetric" reference with respect to both brain hemispheres, thereby
>> providing a tool for characterizing hemispheric asymmetries in EEG
>> studies.
>> However, this is not generally true; the effect of the average mastoids
>> reference is artificially to correlate data from recording electrodes near
>> the two mastoids (Srinivasan et al. 1998), potentially reducing estimates
>> of dynamic hemispheric source asymmetries. We will consider several
>> simulations in the next section where we introduce the average reference
>> and compare it to the mathematical linked-ears reference.
>>
>>
>> Instead, use either average reference (if you have > 64 ch) or Dezhong
>> Yao's REST (if you use the low number of electrodes, you might want to use
>> this, but it is conceptually much more complicated than average
>> reference).
>>
>> Makoto
>>
>> On Tue, Apr 28, 2026 at 11:35 PM Kathryn Bolton via eeglablist <
>> eeglablist at sccn.ucsd.edu> wrote:
>>
>> > Hi EEGLAB mailing list,
>> >
>> > I hope you're having a good week so far! I’m a PhD student working on an
>> > EEG study examining affective responses to music in older adults.
>> > Participants listen to multiple 1-minute music samples, and I plan to
>> > compute frontal alpha asymmetry (FAA) and beta/alpha ratio (PSD-based
>> > measures) and relate these to valence/arousal ratings.
>> >
>> > I’ve structured my pipeline into three stages, partly to distribute work
>> > across research assistants:
>> >
>> >    - *Script 1:* preprocessing + ICA (run by RAs; outputs ICA-ready
>> >    datasets)
>> >    - *Script 2:* manual ICA inspection/cleaning (topography, spectrum,
>> >    timecourse), followed by channel interpolation
>> >    - *Script 3:* epoching, PSD computation, and extraction of FAA and
>> >    beta/alpha ratio
>> >
>> > In my original pipeline, I noticed that the component maps used for
>> > inspection in Script 2 appeared a bit off (e.g., nearly uniform colour
>> > across components). I’ve since revised Script 1 to the following
>> workflow:
>> >
>> >    - Re-reference to linked mastoids prior to ICA (BioSemi
>> recommendations)
>> >    - High-pass filter at 0.5 Hz only, removing a 50 Hz low-pass filter
>> at
>> >    this stage
>> >    - Downsample to 512 Hz
>> >    - Automated detection and removal of bad channels (logged), prior to
>> ICA
>> >    - Run ICA using rank-adjusted PCA
>> >
>> > This seems to have resolved the issue with the component maps. But my
>> > question about filtering is whether you recommend applying only a
>> high-pass
>> > filter prior to ICA (as above), and then applying the full bandpass
>> (e.g.,
>> > including a low-pass) later, prior to PSD analysis?
>> >
>> > I want to ensure that the filtering approach is appropriate both for ICA
>> > decomposition and for downstream frequency-based analyses (FAA and
>> > beta/alpha ratio).
>> >
>> > Thank you very much for your time and advice, I really appreciate it!
>> >
>> > Best regards,
>> > Kathryn
>> >
>> >
>> > *Kathryn Bolton, MA*
>> >
>> > Pronouns: she/her
>> >
>> > PhD Candidate | Clinical Psychology
>> >
>> > Department of Psychology | Cognitive Aging Lab (CAL)
>> >
>> > Toronto Metropolitan University
>> >
>> > *E-mail:* *kathryn.bolton at t <kathryn.bolton at ryerson.ca>orontomu.ca
>> <https://urldefense.com/v3/__http://orontomu.ca__;!!Mih3wA!DvoqGZaNMAX6PWBKD1OkS_noSQPTe78PYfHQe_Ywu0CqcnwcUuNMfyeIxPWt6rVImWOUKmw3RB3xwHQOYYZEgi7CLWMfEhs$>
>> > <
>> >
>> https://urldefense.com/v3/__http://orontomu.ca__;!!Mih3wA!HEf3RTkd7gvWzdftBOsrCvgP_ahawX4XGp6D3OMOOK2ck_zqZyrcOpLjBJdpl7-leggZto82bqKspbCJ49hb7OlEv45uNvcNR5M$
>> > >*
>> >
>> > *Website*: *
>> >
>> https://urldefense.com/v3/__https://psychlabs.torontomu.ca/cal/portfolio/kathryn-bolton/__;!!Mih3wA!HEf3RTkd7gvWzdftBOsrCvgP_ahawX4XGp6D3OMOOK2ck_zqZyrcOpLjBJdpl7-leggZto82bqKspbCJ49hb7OlEv45uwX-37m8$
>> > <
>> >
>> https://urldefense.com/v3/__https://psychlabs.torontomu.ca/cal/portfolio/kathryn-bolton/__;!!Mih3wA!HEf3RTkd7gvWzdftBOsrCvgP_ahawX4XGp6D3OMOOK2ck_zqZyrcOpLjBJdpl7-leggZto82bqKspbCJ49hb7OlEv45uwX-37m8$
>> > >*
>> >
>> >
>> > *Due to the volume of emails I receive, I will do my best to respond
>> within
>> > 2 business days. If you require an immediate response or if it is an
>> > emergency, please put "urgent" in the subject line. Thank you for your
>> > understanding. *
>> > _______________________________________________
>> > To unsubscribe, send an empty email to
>> > eeglablist-unsubscribe at sccn.ucsd.edu or visit
>> > https://sccn.ucsd.edu/mailman/listinfo/eeglablist  .
>> _______________________________________________
>> To unsubscribe, send an empty email to
>> eeglablist-unsubscribe at sccn.ucsd.edu or visit
>> https://sccn.ucsd.edu/mailman/listinfo/eeglablist .
>
>