This might sound a very naive question but I will ask it anyway since I
want to have it clear in order to avoid affecting all my analysis:
When calculating the average signals in mne_browse_raw or mne_process_raw
I have the option of turning off the projection (usually the average EEG
reference). The basic idea of an average EEG reference seems to be a
reasonable thing to do. Nevertheless, doing this changes, sometimes very
drastically, the waveforms.
How strong would anyone recommend me to use the average EEG reference?
Is there any situation in which using the average EEG reference is not the
appropriate path to follow?
In physics the reference point for electric potentials is usually
chosen to be in the infinity, which is a nice choice, because it is
unique and makes sense otherwise. However, choosing reference point on
the head surface is necessary to remove electrical environmental noise
from EEG. To simplify matters, let's assume that EEG activity recorded
on the scalp reflects only the neuronal activity in the (small)
underlying part of cortex. By using referencing you effectively
compare the activity under a given electrode with activity elsewhere
in cortex. What is better for your purpose: to compare it with average
cortical activity, or with cortical activity in the vicinity of your
reference electrode(s)? It's up to you to decide, really. In practical
terms, referencing to one or two electrodes means that signal on
nearby electrodes will be weaker and, thus, less informative than on
remote electrodes. Also, keep in mind that the particular choice of
reference is immaterial for source localization: the result will be
the same.
I agree with both Yury and Daniel, just a couple of additional comments:
(1) Actually, clinical neurophysiologists not only look at different
references but also derivations, i.e., differences or more complex
linear combinations (surface Laplacian) between the EEG channels.
Effectively, this is like constructing a coarse inverse solution from
EEG data. I am always very impressed how well-informed predictions
about the sources underlying, e.g., epileptic activity a trained
physician can make with this simple method.
(2) Theoretically, the choice of the the reference electrode is
immaterial. However, with the average EEG reference the value
predicted by the forward model at the reference electrode does not
affect the solution, which may be important if the reference electrode
location is such that the forward modeling error is larger than for
other electrodes. I do not know whether this really counts in practice.
(3) The MNE inverse operator calculation enforces the average EEG
reference to be used.