[Eeglablist] Critical pitfall of spectral power analysis?

[Cedric Cannard]

Hi Makoto,

I’m really glad this topic came up. I think refining these concepts is critical for the long-term progress of neuroscience, and I hope I can contribute something useful. Here’s my current understanding of the issue and I propose a potential solution. 

As Makoto described to me a while ago and here: https://sccn.ucsd.edu/wiki/Makoto%27s_preprocessing_pipeline#Umbra_Corticalis_.28For_330.2C000_page_views.2C_Added_on_01.2F01.2F2025.3B_corrected_on_06.2F25.2F2025.29 
Broadband 1/f slopes can emerge from periodic sources via amplitude modulation, membrane RC filtering, and spatial summation constraints (Buzsáki et al., 2012; Nunez & Srinivasan, 2006). Scalp electrodes integrate over large, structured cortical regions, with slope changes depending on the extent and synchrony of oscillatory sources. In the spectral fitting tradition (e.g., FOOOF), the aperiodic component is usually treated as a broadband process partially independent from narrowband rhythms.

As Gyurkovics et al. (2021) and Makoto's example point out, if the baseline contains additive broadband components (as is likely), the same absolute oscillatory change can appear larger with a low baseline and smaller with a high baseline. 

One way forward I suggest is to model baseline explicitly rather than remove it by ratio. For oscillatory targets, work in linear power units instead of dB, estimate baseline aperiodic parameters (e.g., offset, exponent via FOOOF or IRASA), and include them (along with baseline band power) as covariates in a GLM or mixed model. This accommodates both additive and multiplicative contributions, yielding a condition effect that reflects rhythmic changes net of broadband shifts (Donoghue et al., 2020; Wen & Liu, 2016; Alday, 2019). 

If one still wanted to focus on broadband state effects rather than oscillations, the same GLM framework could be applied directly to the aperiodic parameters themselves (e.g., baseline offset, exponent). This would allow testing whether these parameters change systematically across conditions, with physiological interpretations such as global power shifts (offset) or potential changes in excitation–inhibition balance (exponent; Gao et al., 2017).

I’ve set up a small MATLAB simulation repo illustrating how this bias arises and how a GLM-based adjustment can remove it: https://urldefense.com/v3/__https://github.com/amisepa/eeg_glm_aperiodic_covariate__;!!Mih3wA!Ce_9hnOFHUAB7C-ZZf4C4Hxmygb-iSFGuxv4cE7TEEml8V0EcFqc-4JQ9slmdryJSv0Ti88G0V1kM_Rb-3jPEJF9pA$ 
 could be a useful starting point for collaborative exploration by the community. 


Cedric Cannard



On Monday, August 11th, 2025 at 11:39 AM, Makoto Miyakoshi via eeglablist <eeglablist at sccn.ucsd.edu> wrote:

> Hi Jinwon and Daniele,
> 
> I've checked that paper recently but haven't read it. Let me guess what the
> main problem is, and let me use a simple example below to share
> understanding of it.
> 
> Subject 1: Baseline-period alpha power magnitude 10 microV^2/Hz, and a task
> increased the power by 10 microV^2/Hz. Thus, the power change is 10
> microV^2/Hz -> 20 microV^2/Hz, which is 3dB.
> 
> Subject 2: Baseline-period alpha power magnitude 100 microV^2/Hz, and a
> task increased the power by 10 microV^2/Hz. Thus, the change is 100
> microV^2/Hz -> 110 microV^2/Hz, which is 0.41dB.
> 
> 
> Thus, even though both subjects showed the same 10 microV^2/Hz power
> increase evoked by the task, dB-conversion showed one is +3dB while the
> other is +0.41dB.
> I guess this is the main point of the problem? I still do not see how the
> source independence issue can relate here, but at least this is a part of
> the problem and is legitimate, right?
> 
> This kind of '1/f slope + peak' conceptualization, together with concepts
> such as 'oscillatory', 'non-oscillatory', reminds me of FOOOF (Donoghue et
> al., 2020; Gao et al., 2017).
> 
> Here is my take:
> If someone makes an assertion that that dB-converted calculation is the
> ONLY VALID way of quantifying it, s/he is wrong.
> Otherwise it is ok to use the dB-converted calculation. It just has
> insensitivity in certain aspects. The calculation itself is valid.
> 
> A practical merit of using dB-conversion is that cross-frequency
> normalization is automatically taken care of.
> For example, if you observe 10 microV^2/Hz power increase in theta and
> gamma bands, which is more prominent? The latter, right? It's because the
> variance of the 'baseline signals' follows 1/f.
> So, if you want, we can publish another paper saying that using microV^2/Hz
> cannot show the significance of the same 10 microV^2/Hz power increases in
> theta and gamma.
> 
> These are just thought experiments. My point is that there are usually
> trade-offs in these approaches, and it is rare to find that one approach
> turned out to be completely wrong. It's usually a matter of distribution of
> sensitivities. Also, it should not be too difficult to use multiple
> calculations and show the results in parallel. However, I do not know what
> EEGLAB developers will do for this issue (will they ever recognize it as an
> issue?)
> 
> I've also seen a criticism that our inter-trial phase coherence is biased.
> https://urldefense.com/v3/__https://onlinelibrary.wiley.com/doi/abs/10.1002/sim.3132__;!!Mih3wA!DDPZ_YphIcjaVH3DU8jtzW_d9dhPFRD0nt95TF-cXHZrIJX7BX9VE_PO2zN4gVthY5GGOkVw0MwTB1uR7e3aMMJ-bQs$
> 
> %%%%%%%%%%%%%%%%%%%%%%%%
> While writing this response, I saw Daniele's post. I'm curious to hear what
> the 'substantial flaw' is in more detail. Looks like my quick and lazy
> problem explained above is different from what he means.
> 
> By the way, I was happy to find that he wrote '1/f can be generated by
> "pure oscillations" with nonuniform amplitude, among other things.' because
> I once said exactly the same thing to express my dissatisfaction to hear
> how the concept of 'aperiodic' had been misused in some communities!
> 
> Makoto
> 
> On Sat, Aug 9, 2025 at 1:25 PM 장진원 via eeglablist eeglablist at sccn.ucsd.edu
> 
> wrote:
> 
> > Hi all,
> > 
> > Recently I found one interesting article that addresses the pitfall of
> > baseline correction that many scientists have used to transform EEG to
> > time-frequency domain. According to this article, power spectrum formation
> > is highly exposed to subject-dependent noise that independently affects
> > power spectrum regardless of signal. Because I am not an engineer who
> > majors signal transformation, I wonder how eeglab could handle this issue
> > in spectral power analysis because this article implies that using alpha
> > (8-13Hz) or theta (4-8)Hz is totally unacceptable in clinical studies.
> > 
> > Reference: Gyurkovics, M., Clements, G. M., Low, K. A., Fabiani, M., &
> > Gratton, G. (2021). The impact of 1/f activity and baseline correction on
> > the results and interpretation of time-frequency analyses of EEG/MEG data:
> > A cautionary tale. NeuroImage, 237, 118192.
> > 
> > https://urldefense.com/v3/__https://doi.org/10.1016/j.neuroimage.2021.118192__;!!Mih3wA!FUy2N9N5bZQJF1IM06-OIaXtDG8YvPWzfrSGxmJE6N_4DPqW9Irqgr9P4PajtadaJV9Jzo1Z9QWJsE2RPNZmbe-Mkw$
> > 
> > Best regards,
> > Jinwon Chang
> > _______________________________________________
> > 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 .