The impact of 1/factivity and baseline correction on the results and interpretation of time-frequency analyses of EEG/MEG data: A cautionary tale

Abstract

Typically, time-frequency analysis (TFA) of electrophysiological data is aimed at isolating narrowband signals (oscillatory activity) from broadband non-oscillatory (1/f) activity, so that changes in oscillatory activity resulting from experimental manipulations can be assessed. A widely used method to do this is to convert the data to the decibel (dB) scale through baseline division and log transformation. This procedure assumes that, for each frequency, sources of power (i.e., oscillations and 1/f activity) scale by the same factor relative to the baseline (multiplicative model). This assumption may be incorrect when signal and noise are independent contributors to the power spectrum (additive model). Using resting-state EEG data from 80 participants, we found that the level of 1/f activity and alpha power are not positively correlated within participants, in line with the additive but not the multiplicative model. Then, to assess the effects of dB conversion on data that violate the multiplicativity assumption, we simulated a mixed design study with one between-subject (noise level, i.e., level of 1/f activity) and one within-subject (signal amplitude, i.e., amplitude of oscillatory activity added onto the background 1/f activity) factor. The effect size of the noise level × signal amplitude interaction was examined as a function of noise difference between groups, following dB conversion. Findings revealed that dB conversion led to the over- or under-estimation of the true interaction effect when groups differing in 1/f levels were compared, and it also led to the emergence of illusory interactions when none were present. This is because signal amplitude was systematically underestimated in the noisier compared to the less noisy group. Hence, we recommend testing whether the level of 1/f activity differs across groups or conditions and using multiple baseline correction strategies to validate results if it does. Such a situation may be particularly common in aging, developmental, or clinical studies.