Abstract
When researchers are interested in the time course of data which consists of one-sample-per-trial (e.g. accuracy as a function of reaction time), they usually split the data into time bins (also known as Vincentizing). In some cases an arbitrary number of bins is used, which, as we demonstrate, can significantly misrepresent the underlying signal. Several researchers circumvented this problem by smoothing the time course using a Gaussian kernel. A major problem for this method is missing data, as well as inability to perform proper statistical analysis on the smoothed time course. In the analysis of neural time series there is a long history of using permutation testing for determining statistical significance. However, this method assumes a complete time series for each trial and therefore cannot be directly applied to one-sample-per-trial data. Here we present a novel method that combines the improved smoothing of one-sample-per-trial data with permutation testing. We show that this method can be used to visualize the time course and to perform statistical testing for differences between conditions, as well for differences against a predetermined baseline. The method is demonstrated using eye-tracking data (saccade curvature), as well as psychophysical data (accuracy) and compared against two common methods of binning data. The advantages and disadvantages of this method are discussed.
Meeting abstract presented at VSS 2018