Statistical detection of between-group differences in event-related potentials

Objective: Many event-related potential (ERP) situations do not fulfill the multivariate statistics requirement of more cases than measurements. Five simulation studies were carried to select a sensitive test for comparing two independent groups of ERPs. Methods: Simulated signal and noise waveforms were generated using different combinations of parameters: cases or replications per group, correlation between data points, number of points, signal-to-noise ratio (S/N), etc. The false alarm (FA) rate of each method was assessed and their sensitivity compared over sets of 2000 simulated experiments per condition. Results: Study 1 identified the ‘Projection onto Centroids Difference Vectors’ (PCDV) method of Haig and Gordon (Brain Topogr 1995;8:67) as very good, but its FA rate was erratic under several conditions. The following studies served to shape the implementation parameters of a version of PCDV, termed PCDVp, that assesses significance through random permutations of the case labels. The final form is very sensitive. For instance, with two groups of 48 trials of 30-point EEG-like waveforms, its power for α=0.05 is about 50% at S/N 1.0 and 90% at S/N 1.5 (amplitude). Conclusion: PCDVp requires no a priori knowledge and is sensitive to detect differences between independent sets of waveforms, topographies or spatio-temporal data.