Impact of reduced-dimensionality independent components analysis on event-related potential measurements

Independent components analysis (ICA) is an effective and ubiquitous tool for cleaning EEG. To reduce computation time, many analysis pipelines decrease EEG dimensionality prior to ICA. A 2018 report by Artoni and colleagues detailed the deleterious effects of such reduced-dimensionality ICA (rdICA) on the dipolarity and reliability of independent components. Though valuable for researchers interested in directly analyzing independent components, ICA is more commonly used for cleaning EEG. Thus, a direct examination of the impact of artifact removal via rdICA on EEG data quality is needed. We conducted a registered analysis of 128 electrode recordings of 43 healthy subjects performing an active auditory oddball task. We preprocessed each subject's data under the following conditions: (1) ICA without dimension reduction, (2) ICA with only 64 electrodes included, (3) ICA preceded by PCA retaining 99% of the original data variance and (4) ICA preceded by PCA retaining 90% variance. We then quantified ERP data quality by measuring mean-amplitude, standardized measurement error (SME) of the single-trial mean-amplitudes, and split-half reliability of the N1 and P3 components. We then attempted to replicate our findings in an independent validation dataset. We observed statistically and practically significant changes in the mean amplitude of early sensory components for the 90% condition. Unexpectedly, the SME was only larger for the 64 electrode condition. Also unexpectedly, the effect of rdICA on split-half reliability was inconsistent between datasets. Based on the observed data, we argue that PCA-based rdICA is justifiable when used cautiously.