Pitfalls in multivariate classification analysis of EEG data
Multivariate pattern analysis (MVPA) in EEG/MEG has really taken flight over the past decade, offering increased sensitivity and novel analytical approaches such as temporal generalization. However, this improved sensitivity also comes at a price. In particular, MVPA is more sensitive to adverse effects of confounders and interpretation issues, some of which are particular to MVPA, but also general confounders that merely have a larger influence in MVPA due to its higher sensitivity. In this talk, I will highlight issues and confounders that I have come across in recent years which researchers should be aware of, such as the adverse effect of high-pass filtering on long time series, the influence of (small) eye-movements on classification accuracy, the unwanted effects of unbalanced designs when attempting to compare conditions in train-test regimes (and more generally the effect of trial numbers on MVPA effect sizes), as well as the adverse effects of post-hoc sorting approaches in EEG decoding. General recommendations will be provided whenever possible.