The effect of data preprocessing on spike correlation analysis results

In recent years, an increasing number of large electrophysiological data sets have become publicly available, thereby providing researchers with the opportunity to analyze spike train data without conducting their own experiments. While this is undoubtedly a positive development, it increases the need for proper documentation on how the data were collected and what preprocessing was performed on the data, since interpreting analysis results in ignorance of these pieces of information can lead to wrong conclusions. An important preprocessing step is the removal of artifacts from the recordings. Electrophysiological recordings are particularly susceptible to electrical cross-talks between recording channels, resulting in artifact spikes that are coincident in multiple channels on the time scale of the data sampling rate, i.e., 1/30 ms in popular setups. The removal by signal whitening is only possible if also the raw sampled data are available, thus to eliminate this type of artifact is to remove all coincident spikes on the recording time scale to definitely avoid artifact spikes. However, given the lack of the "ground truth", this step has the potential to eliminate, in conjunction with the artifacts, components of the data that are pertinent to the research objective. In this study, we use a modified version of the Unitary Event Analysis and demonstrate that such preprocessing results in significantly lower correlations than expected by chance even on longer time scales. We also propose a method to correct for the bias introduced by this preprocessing. Thus, slight changes in the preprocessing have potentially strong impact on analysis results and methods need to consider these effects.