When Data Filtering Introduces Bias

By Neuroskeptic | July 7, 2012 1:04 pm

Oh no. Another worrying methods problem for neuroscience, this time for electrophysiologists: Systematic biases in early ERP and ERF components as a result of high-pass filtering.

The event-related potential (ERP) and event-related field (ERF) techniques provide valuable insights into the time course of processes in the brain. Researchers commonly filter the data to increase the signal-to-noise ratio. However, filtering may distort the data, leading to false results. Using our own EEG data, we show that acausal high-pass filtering can generate a systematic bias easily leading to misinterpretations of neural activity… among 185 relevant ERP/ERF publications, 80 used cutoffs above 0.1Hz. As a consequence, part of the ERP/ERF literature may need to be re-analyzed.

The problem in brief: many researchers use a high-pass filter on their electroencephalography (EEG) and magnetoencephalography (MEG) recordings of brain electrical activity. A high-pass filter removes low frequency (i.e. slow) changes from the signal. These slow fluctuations are often considered to be mere “noise”.

The problem is that these filters have side effects: as well as ‘cleaning up’ the data, they can also distort it. There are two main kinds of filter: causal filters are well-known to mutate the signal. Acausal high-pass filters avoid these dramatic artefacts –

But David Acunzo and colleagues point out that acausal filters can actually be more dangerous, because they still distort the data, just in more subtle ways that are harder to spot. In particular, acausal filters can alter the signal at time points before the true signal begins. See the pic above.

That’s not necessarily a problem in all cases, but it’s certainly bad news for researchers interested in measuring exactly when neural responses happen.

The authors highlight an area of neuroscience where this problem could be misleading researchers. The very earliest brain responses to visual stimuli, about 90 milliseconds after the stimulus onset, is called the “C1” response. Classically, it was thought that the size of the C1 wave was purely a ‘bottom-up’ phenomenon, determined only by the brightness etc. of the stimulus. But recently, studies have reported ‘top down’ modulation of C1 by attention, emotional state, etc.

Acunzo et al point out that many of these studies used strong acausal filtering and that what might be happening is that attention actually causes late changes to the visual response, but that due to filtering artefacts, these late changes appear in the data sooner than they really happen. They advise that only weak (low threshold) high-pass filters should be used, and that interesting findings in filtered signals need to be checked against the raw data.

ResearchBlogging.orgAcunzo DJ, Mackenzie G, and van Rossum MC (2012). Systematic biases in early ERP and ERF components as a result of high-pass filtering. Journal of neuroscience methods PMID: 22743800

CATEGORIZED UNDER: bad neuroscience, EEG, methods, papers
  • http://petrossa.wordpress.com/ petrossa

    And given that calibration of the fMRI software used EEG results it's clear, i hope, to all now we now have a serious problem with granularity due to the introduction of errors based on erroneous data calibration.

    Whilst both systems can be used to somewhat accurately measure activity in a given portion of the brain none can be used to make such absurd claims as 'i found the part of the brain that handles philosophical thought' or anything remotely involving higher order processes.

    The whole system is flawed on so many levels (basic premise, execution and interpretation) i foresee in the near future a paper retraction frenzy not unlike the one going on in psychology.

  • Anonymous

    fMRI-based neuroscience better clean up its act and start behaving like a real science – complete with propagation of instrument error – or many of us will be out of jobs.

    If PIs do not stop selling their NIH grants based on wishful thinking and if NIH does not get rid the the reviewers that enable this wishful thinking then a potentially useful tool will wash-out.

  • http://www.blogger.com/profile/08099485960661603080 Matt Craddock

    It's pretty well established that filtering can be a thorny issue in ERP analysis. Steve Luck's book on ERP analysis raises a lot of issues with it, and that was from 2005. There have been a few papers bringing up related issues recently in Frontiers in Human Neuroscience, starting with a general paper on several conceptual as well as technical issues with interpretation of object recognition studies, by Rufin VanRullen:


    There's then commentary from Guillaume Rousselet (http://www.frontiersin.org/Perception_Science/10.3389/fpsyg.2012.00131/full) and Andreas Widmann & Erich Schroeger (http://www.frontiersin.org/Perception_Science/10.3389/fpsyg.2012.00233/full)

    The basic issue they're talking about is how high pass filtering can cause temporal smearing and drastically alter the onset of specific EEG components. I think a broader issue here is actually the study and interpretation of the onset/peak latency/peak amplitude of individual peaks in ERP waveforms anyway. It's something I'm doing a lot at the moment, but sometimes it can be easy forget that there's nothing special about peaks, and we lose a lot of info by focusing on peaks alone. OTOH, that also helps focus on *something*, since there's a lot of info to sift through.

  • DS

    I am really surprised that there is not more discussion here about this particular post.

  • http://www.blogger.com/profile/06647064768789308157 Neuroskeptic

    Mmm, so am I actually. Oh well.



No brain. No gain.

About Neuroskeptic

Neuroskeptic is a British neuroscientist who takes a skeptical look at his own field, and beyond. His blog offers a look at the latest developments in neuroscience, psychiatry and psychology through a critical lens.


See More

@Neuro_Skeptic on Twitter


Discover's Newsletter

Sign up to get the latest science news delivered weekly right to your inbox!

Collapse bottom bar