Event-related potentials (ERPs)
Standard averaging and the additive noise model
To enhance the signal-to-noise ratio of event-related potentials (ERPs), standard-averaging of EEG epochs is certainly the most often used procedure. The basic assumption underlying this procedure is that standard time-domain averaging cancels out the contribution of signals which are not ‘time-locked’ or ‘stationary’ to the onset of the event while it spares ERPs as they occur with a constant time-delay. The fraction of the signal which is cancelled out by standard time-domain averaging, often referred to as ‘additive noise’, includes the ongoing EEG activity as well as artifactual signal changes such as that produced by muscle activity or eye blinks.
A first limitation of this procedure is that it requires that ERPs be perfectly time-locked to the event onset. Therefore, if ERP components are subject to a significant latency-jitter they will not be adequately preserved. This limitation could considerably affect the enhancement of late components whose latency variability is assumed important. Alteration of these components would be especially important if these components are very transient, their higher frequency content making them more sensitive to the ‘dephasing’ introduced by the latency jitter.
Furthermore, recent knowledge has indicated that at least some ERP components may result from a reorganization of the phase of ongoing EEG oscillations (Makeig et al. 2002). It is therefore possible that some ERP components simply result from time-domain averaging procedures not canceling out the ongoing EEG during this transient event-related ‘phase-resetting’ (Makeig et al. 2004a).
More importantly, it is now well accepted that in addition to ERPs, sensory, motor, and cognitive events may induce transient enhancements or attenuations of ongoing EEG oscillations, referred to as event-related ‘synchronization’ (ERS) and ‘desynchronization’ (ERD). However, as these ongoing EEG oscillations are not ‘phase-locked’ to the event onset, they are lost by time-domain averaging.