Time Granularity, Lag Length, and Down-Sampling Rates for Neurocognitive Data

Time series analysis, nonlinear or otherwise, requires an appropriate lag length between observations. The choice of lag length is contingent to some extent on whether the source data are under- or over-sampled. For neuro-cognitive data, the time granularity should represent psychologically meaningful units. Automatic methods for determining optimal lag length are not readily available, particularly for potentially oversampled data and if the eventual goal is to compare linear versus nonlinear models in large quantities across experimental conditions. The present study examined the interacting roles of down-sampling rate, natural lag rates, task types, real-time lapse, and lag units on the accuracy of linear and nonlinear (exponential structures) autocor-relational models, starting with electrodermal data sampled at 200 obs/sec. Participants were 197 undergraduates organized into groups of 3-7 people in three sequential task conditions: watching a video that explained the problem situation, an individual mental task, and a group problem-solving task. Results showed that the optimal lag structures came from natural rates of 2 obs/sec at 1 sec lag or 3 obs/sec at 1 lag unit. Results varied modestly across the subtasks such that greater stability occurred when participants watched the video, followed by the group task, followed by the individual task. Nonlinear models were more accurate than ARMA generally, although there were specific experimental conditions in which the reverse was true. Future research across disciplines should investigate optimal lags from a perspective of naturally occurring change processes rate rather than rely on automatic computations.