A note on detecting statistical outliers in psychophysical data

Abstract

This paper considers how to identify statistical outliers in psychophysical datasets where the underlying sampling distributions are unknown. Eight methods are described, and each is evaluated using Monte Carlo simulations of a typical psychophysical experiment. The best method is shown to be one based on a measure of spread known as Sn. This is shown to be more sensitive than popular heuristics based on standard deviations from the mean, and more robust than non-parametric methods based on percentiles or interquartile range. Matlab code for computing Sn is included.