2.3. Choosing the optimal offline reference
Several publications have discussed the effect of different referencing schemes on EEG recordings (e.g. Dien, 1998; Luck, 2104; Li et al., 2018). To date, there is no agreement on an “optimal reference”, however, several facts can inform your choice:
The reference should optimally be influenced by the same noise as the other scalp channels (stemming from the ground circuit, pulse, movements etc.). When subtracting the new reference, this noise is subtracted as well. Just be aware that also the reverse is true: if the reference contains noise that is not shared with the other scalp channels, this noise will be introduced into all other channels during re-referencing. For picking up comparable noise, the channel(s) ultimately forming the new reference should hence be located closely enough to the region of interest (on the head) and provide good data quality (i.e. low impedance).
At the same time, the reference should optimally pick up as little signal of interest as possible. This signal, too, would be subtracted during-re-referencing which might reduce the effect you are looking for. For this reason, the reference should not be located too closely to the region of interest for your specific research question. You see that the choice of the reference hence entails a trade-off.
When choosing your reference, it is also important to consider the comparability to other studies. If study results are to be compared, they should also be referenced in the same way. It is thus useful to find out in other publications which referencing scheme is common for your specific research question or paradigm.
Figure 1 (below) illustrates how the choice of the reference may affect your data. The example data stems from an Oddball paradigm (N=8). The displayed signal at Cz was referenced to A) FCz, B) an average of all 32 recorded channels and C) an average of two mastoid electrodes. The black and blue curves display the event-related potentials (ERPs) in response to frequent and rare stimuli, the red curve reflects their difference.
Figure 1: Event-related potentials (ERPs) from a visual and auditory Oddball task (N=8). Curves display ERPs in response to frequent (black) and rare (blue) stimuli. The maps reflect the topography of the difference wave (red) from 300 to 450 ms (interval marked in light blue).
The largest difference between the rare and frequent condition is expected to appear around 300 ms post-stimulus (P300) and to be located over central-parietal regions. You can see how the choice of the reference affects the signal time course at Cz, while the topographic distribution merely shifts in its overall amplitude (note that the topographies in B) and C) are based on one more channel because FCz was restored). For the detection of the P300 in this paradigm, the average mastoid reference would be preferable over the other two referencing schemes. This might of course differ for other ERPs or for your specific research question.