Abstract
Event related potentials (ERP) have been an essential part of EEG analysis since its early days. Common practice is to average over many trials to get an estimate of the underlying brain response. However, many experiments contain events of variable length (e.g. due to differences in reaction times, fixation duration, stimulus duration, etc.). These varying durations are rarely considered, be it due to a lack of analysis tools or plain unawareness, in the worst case leading to biased or even nonsensical inferences about the nature of the brain. Even worse, the varying event durations often co-occur with temporal overlap of different ERPs (e.g. responses to stimulus onsets and button presses) adding further bias. We applied regression methods to simulated and real-world data and systematically explored how event duration affects the resulting ERPs and how to adequately model them. To account for the temporal overlap, we used deconvolution based overlap correction as implemented in the unfold-toolbox (https://www.unfoldtoolbox.org/) and investigated its additional influence on the ERP estimation. We find that modelling event durations as binned or linear predictors performs poorly. However, non-linear effects using spline-regression seem to be able to capture the main patterns and are thus a promising candidate for further study.