Abstract
Background Neural oscillations are often quantified as average power relative to a cognitive, perceptual, and/or behavioral task. This is commonly done using Fourier-based techniques, such as Welch’s method for estimating the power spectral density, and/or by estimating narrowband oscillatory power across trials, conditions, and/or groups. The core assumption underlying these approaches is that the mean is an appropriate measure of central tendency. Despite the importance of this assumption, it has not been rigorously tested.
New method We introduce extensions of common approaches that are better suited for the physiological reality of how neural oscillations often manifest: as nonstationary, high-power bursts, rather than sustained rhythms. Log-transforming, or taking the median power, significantly reduces erroneously inflated power estimates.
Results Analyzing 101 participants’ worth of human electrophysiology, totaling 3,560 channels and over 40 hours data, we show that, in all cases examined, spectral power is not Gaussian distributed. This is true even when oscillations are prominent and sustained, such as visual cortical alpha. Power across time, at every frequency, is characterized by a substantial long tail, which implies that estimates of average power are skewed toward large, infrequent high-power oscillatory bursts.
Comparison with existing methods In a simulated event-related experiment we show how introducing just a few high-power oscillatory bursts, as seen in real data, can, perhaps erroneously, cause significant differences between conditions using traditional methods. These erroneous effects are substantially reduced with our new methods.
Conclusions These results call into question the validity of common statistical practices in neural oscillation research.
Highlights
Analyses of oscillatory power often assume power is normally distributed.
Analyzing >40 hours of human M/EEG and ECoG, we show that in all cases it is not.
This effect is demonstrated in simple simulation of an event-related task.
Overinflated power estimates are reduced via log-transformation or median power.
Acknowledgements
We thank Thomas Donoghue and Priyadarshini Sebastian for their assistance with data preparation and comments on a draft of this manuscript. Izhikevich is supported by the National Science Foundation Graduate Research Fellowship Program (DGE-1656518) and a Stanford Graduate Fellowship. Gao is supported by the Natural Sciences and Engineering Research Council of Canada (NSERC PGS-D), UCSD Kavli Innovative Research Grant (IRG), and the Katzin Prize. Voytek is supported by a Sloan Research Fellowship (FG-2015-66057), the Whitehall Foundation (2017-12-73), and the National Science Foundation under grant BCS-1736028. The authors declare no competing financial interests.
Author contributions
All authors initiated and designed the study. Izhikevich built the database and analysis pipeline, and all authors analyzed the data. All authors contributed to the manuscript.