Optimizing PCA methodology for ERP component identification and measurement: theoretical rationale and empirical evaluation

Objective: To determine how specific methodological choices affect “data-driven” simplifications of event-related potentials (ERPs) using principal components analysis (PCA). The usefulness of the extracted component measures can be evaluated by knowledge about the variance distribution of ERPs, which are characterized by the removal of baseline activity. The variance should be small before and at stimulus onset (across and within cases), but large near the end of the recording epoch and at ERP component peaks. These characteristics are preserved with a covariance matrix, but lost with a correlation matrix, which assigns equal weights to each sample point, yielding the possibility that small but systematic variations may form a factor. Methods: Varimax-rotated PCAs were performed on simulated and real ERPs, systematically varying extraction criteria (number of factors) and method (correlation/covariance matrix, using unstandardized/standardized loadings before rotation). Results: Conservative extraction criteria changed the morphology of some components considerably, which had severe implications for inferential statistics. Solutions converged and stabilized with more liberal criteria. Interpretability (more distinctive component waveforms with narrow and unambiguous loading peaks) and statistical conclusions (greater effect stability across extraction criteria) were best for unstandardized covariance-based solutions. In contrast, all standardized covariance- and correlation-based solutions included “high-variance” factors during the baseline, confirming findings for simulated data. Conclusions: Unrestricted, unstandardized covariance-based PCA solutions optimize ERP component identification and measurement.