The EEGLAB News #4
From the eeglablist
Q. I found some papers recommending not using average reference and using the REST reference (note: these papers mainly talked about ERPs). I will do a power spectral analysis. We are using a Cognionics Mobile 64 device to collect data during a physical task (Isometric exertion). So what should be my re-referencing method ? Initially I was planning for average reference, but now I've gotten confused by those papers.
A. If you use 64 electrodes, you can try both REST and average reference, and decide for yourself. There is no "this is the best reference" (for all purposes) answer, if there was something like,"this is the best referencing method EVER", we wouldn't need to fill pages and papers on that question ;). From my experience: If you use fixed electrode sites (such as linked earlobes) as reference(s), ERP waveforms and total power may be higher in absolute terms than if you use average Reference or REST. This can be good (especially when studying ERP waveforms) or bad (perhaps giving higher noise in the raw EEG). REST and average reference performed similarly in my experiments, so usually I decided to go with average reference, as the REST integration in EEGLAB has some disadvantages (I usually have to re-reference every dataset manually, since I haven't found a way to write scripts that do it).
I hope this doesn't sound rude, but please keep in mind that the paper you linked to is from the authors who "invented" the REST reference, and most of the papers I have read that recommend REST reference are also from them. I think that REST reference is a good concept, but in your case I'd suggest, in practical terms: First, collect the data, then re-reference to any site you like after data collection. Keep in mind that choice of reference can be changed after data collection, as its a linear process, as long as you have recorded the data in a monopolar fashion (all channels having the same reference electrode). That is, choose the referencing method that fits your purposes best. If you find any interesting differences between results using average reference versus REST, publish them in a paper to help other people.
Makoto Miyakoshi
An alternative to the CAR [Common Average Reference] approach is the “infinite reference” one, also known as Reference Electrode Standardization Technique (REST and regularized REST) (Yao, 2001). Both the CAR and REST have been shown to be the extremes of a family of Bayesian reference estimators (Hu et al., 2018b). REST utilizes the prior that EEG signals are correlated across electrodes due to volume conduction [from a crudely modeled cortical source space (see below)], while CAR takes the prior that EEG signals are independent over electrodes (for reviews see Yao et al., 2019; Hu et al., 2019). If the focus of the data analysis is on source space inference (see Section 4.6), re-referencing is, in theory, not necessary but may be useful for comparisons to existing literature. Of note, any linear transform applied to the data (e.g. CAR) should also be applied to the forward matrix used for source space analysis. Such important details are generally taken care of by software tools in the field (and some require data to be in CAR form), but it is worthwhile ensuring that this is done. From https://cobidasmeeg.wordpress.com - Best Practices in Data Analysis & Sharing in Neuroimaging using MEEG.
Some thoughts from Scott Makeig: In this 2017 paper, the REST authors, Yao et al., describe use of the EEGLAB REST tool, which is composed using both a crude source space model and forward head model, a shortcoming the authors acknowledge. As well, in their forward head model they assume a brain/skull conductivity ratio (BSCR) of 80, an early but now outdated estimate for a key value in forward head models that we and others have shown to vary widely across individuals. Yet the precision of their approach (and claim) rests entirely on the accuracy of their model source space and forward head model.
In the paper, they apply Average Reference and REST referencing to data from one 'Oddball' paradigm session and show that the 'P300' peak in the ERP to the 'target' stimuli is larger using REST referencing. They seem to claim that this makes REST the 'more 'correct' and 'effective' method to apply here, but offer no discussion as to how, in this case, a larger P300 (to baseline) ERP peak (with seemingly somewhat larger across-trial variance as well) is more 'correct' than a numerically smaller P300 peak value -- I myself am not convinced.
In particular, their argument seems to rest on their assumption that the cortical source of the EEG is (a crude model of) the entire cortical surface. However, cortical areas of spatially coherent signal should be much more strongly represented in the scalp EEG than cortical areas of random phase activity (e.g., across small area subdomains), these 'effective' source areas possibly differing over task and time. Further, their artifact rejection method in this paper (remove trials with |potential| > 75 uV) will leave in some strong non-brain activities (e.g., from eye movements, etc.) originating in non-brain sources quite far from their assumed source space model.
The average reference method, on the other hand, implicitly makes other assumptions about the source space distribution -- both models are in this sense unrealistic.
ICA decomposition, by contrast, typically (when reasonably applied) models a very large portion of the data from such an experiment session as produced by a session-specific set of effective source projections (many matching a single or dual equivalent dipole source model) from plausible and localizable brain plus readily interpretable non-brain sources. ICA is also not dependent on the referencing method used (so long as it is rank-preserving).
The real test of any analysis method, I would say, is its statistical accuracy, consistency, and sensitivity to true effects -- not the raw values of the measures it delivers. I agree with Makoto that continued progress in EEG research requires continued method comparisons involving applications to actual data, as well as more richly physiologically based theoretical assessments.
Scott Makeig
Note: The REST EEGLAB plug-in (http://www.neuro.uestc.edu.cn/rest/), also available in the EEGLAB plug-in manager, can translate multichannel EEG or ERP data to a new dataset with infinity reference.