Trimmed estimators for robust averaging of event-related potentials

Abstract

Averaging (in statistical terms, estimation of the location of data) is one of the most commonly used procedures in neuroscience and the basic procedure for obtaining event-related potentials (ERP). Only the arithmetic mean is routinely used in the current practice of ERP research, though its sensitivity to outliers is wellknown. Weighted averaging is sometimes used as a more robust procedure, however, it can be not sufficiently appropriate when the signal is nonstationary within a trial.Trimmed estimators provide an alternative way to average data. In this paper, a number of such location estimators (trimmed mean, Winsorized mean and recently introduced trimmed L-mean) are reviewed, as well as arithmetic mean and median.A new robust location estimator tanh, which allows the data-dependent optimization, is proposed for averaging of small number of trials. The possibilities to improve signal-to-noise ratio (SNR) of averaged waveforms using trimmed location estimators are demonstrated for epochs randomly drawn from a set of real auditory evoked potential data.

A case-matched study of neurophysiological correlates to attention/working memory in people with somatic hypervigilance

“…Any ERP and behavioural data that were > 3 standard deviations from the mean were replaced with the group mean plus or minus three standard deviations (Leonowicz, Karvanen, & Shishkin, 2005).…”

Power Quality Evaluation using Advanced Spectrum Estimation Methods

“…Averaging is probably the most common basic statistical procedure in experimental science [5]. It is used for estimating the location of data (or "central tendency") in the presence of random variations among the observations, which can be removed by this procedure.…”

“…Numerous publications are devoted to the problem of robustness of estimation of the data location (for details, see [5]). In the presence of outliers, i.e.…”

“…The tails of the distribution of the data are not simply ignored (It can lead to the loss of information and should be avoided when the sample size is small) as with trimmed mean [5].…”

“…As a consequence, measurements of selected power quality parameters show quite high variability. In order to alleviate this problem, the use of robust averaging [5] is proposed.…”