Median method for eliminating infrequent artifacts and identifying the signals blurred by latency jitter and uncertain occurrence

Most researchers have employed the average method in the identification of event-related brain potentials (ERPs) because they have assumed that the averaging process results in the cancellation of any type of noise. However, the average method is less effective than is the median method for unexpected, infrequent artifacts. Furthermore (and significantly), the average method does not work well for detecting the endogenous, psychological signals that are noninvariant across every trial. Ruchkin (1988) described two types of possible signal variation in amplitude and in latency. The median method is effective for extreme cases of these variations (i.e., the lack of some signals), even if the amount of data is small and their distribution is non-Gaussian. Importantly, in the latter situation, the trial-to-trial latency jitter of the signal is very difficult to eliminate, which introduces obscure errors into the average ERP measurements. We found that the waveform that was generated by the median method was less affected by this jittering than was that generated by the average method. This effect was demonstrated by simulating the signals of artificial and actual ERP data, when the distribution of latency variation was Gaussian.