0. Abstract
The application of time-resolved multivariate pattern classification analyses (MVPA) to EEG and MEG data has become increasingly popular. Traditionally, such time series data are high-pass filtered before analyses, in order to remove slow drifts. Here we show that high-pass filtering should be applied with extreme caution in MVPA, as it may easily create artifacts that result in displacement of decoding accuracy, leading to statistically significant above-chance classification during time periods in which the source is clearly not in brain activity. This is particularly problematic in paradigms that have long trial durations, such as working memory experiments with long retention intervals, where the signal of interest may reside in low parts of the frequency spectrum and thus is more likely to be affected by high-pass filters. In both real and simulated EEG data, we show that spurious decoding may emerge with filter cut-off settings from as modest as 0.1 Hz. We provide an alternative method of removing slow drift noise, referred to as robust detrending (de Cheveigne & Arzounian, 2018), which, when applied in concert with masking of cortical events does not result in the temporal displacement of information. We show that temporal generalization may benefit from robust detrending, without any of the unwanted side effects introduced by filtering. However, we conclude that for sufficiently clean data sets, no filtering or detrending at all may work sufficiently well. Implications for other types of data are discussed, followed by a number of recommendations.
Footnotes
↵# Author note: CNLO and JJF share senior authorship. JvD and CNLO designed the experiment. JvD and JJF designed and conducted the simulations and analyses. CNLO, JvD and JJF wrote the paper. This work was supported by ERC-CoG-2013-615423 grant from the European Research Council, and NWO Vici grant 453-16-002 awarded to CNLO.