Abstract
A methodological problem in most reaction time (RT) studies is that some measured RTs may be outliers-that is, they may be very fast or very slow for reasons unconnected to the task-related processing of interest. Numerous ad hoc methods have been suggested to discriminate between such outliers and the valid RTs of interest, but it is extremely difficult to determine how well these methods work in practice because virtually nothing is known about the actual characteristics of outliers in real RT datasets. This article proposes a new method of pooling cumulative distribution function values for examining empirical RT distributions to assess both the proportions of outliers and their latencies relative to those of the valid RTs. As the method is developed, its strengths and weaknesses are examined using simulations based on previously suggested ad hoc models for RT outliers with particular assumed proportions and distributions of valid RTs and outliers. The method is then applied to several large RT datasets from lexical decision tasks, and the results provide the first empirically based description of outlier RTs. For these datasets, fewer than 1% of the RTs seem to be outliers, and the median outlier latency appears to be approximately 4-6 standard deviations of RT above the mean of the valid RT distribution.