Optimizing principal components analysis of event-related potentials: matrix type, factor loading weighting,...

Area of Science:

• Neuroscience
• Cognitive Science
• Computational Neuroscience

Background:

• Event-related potentials (ERPs) are crucial for understanding neural dynamics.
• Principal Component Analysis (PCA) is widely used for ERP data reduction.
• Conflicting recommendations exist for optimal PCA parameter selection in ERP analysis.

Purpose of the Study:

• To establish a standardized protocol for applying PCA to ERP datasets.
• To resolve discrepancies in existing literature regarding PCA methodology for ERPs.
• To provide evidence-based guidelines for maximizing the utility of PCA in ERP research.

Main Methods:

• Simulated 100 ERP datasets to evaluate PCA parameter effects.
• Compared covariance vs. correlation matrices.
• Assessed Kaiser normalization vs. covariance loadings.
• Examined truncated vs. unrestricted solutions and Varimax vs. Promax rotations.
• Investigated the influence of component size on parameter effects.

Main Results:

• Correlation matrices led to significant variance misallocation.
• Promax rotation demonstrated superior accuracy over Varimax rotation.
• Covariance loadings were less effective than Kaiser Normalization or unweighted loadings.
• Unrestricted solutions did not substantially enhance results when optimal parameters were used.

Conclusions:

• A covariance matrix, Kaiser normalization, and Promax rotation are recommended for PCA in ERP analysis.
• Optimized PCA procedures can significantly improve ERP source localization.
• Further development of PCA methods holds promise for advancing ERP data analysis.