[Eeglablist] Too much drift with just 0.1hz high-pass filtering?

Fri Aug 7 10:39:09 PDT 2015

Dear Kevin, Makoto and list,

I fully agree with Makoto’s comments. Further, I would like to emphasize that there *cannot* be a *general* recommendation or answer to this question (how to filter ERPs). The answer to this question always depends on the data, the data quality and signal-to-noise ratio (SNR), the noise characteristics, the paradigm, the parameter(s) you want to estimate from the data, and other variables.

I would like to argue for two criteria:
(1) Filters should be applied to improve the SNR (reduce the noise) in order to be able to estimate a particular parameter (that is, if there is no noise there is no need to filter!).
(2) Applied filters must not introduce distortions or artifacts systematically biasing your estimated parameter.

Two examples:
* With high frequency noise you might need a lowpass filter to get a valid estimate of component peak latency. However, a lowpass filter might also lead to a systematic underestimation of component onset latency (due to smoothing; VanRullen, 2011). To estimate time window means you usually do *not* need a lowpass filter.
* Many ERP components are not of oscillatory nature but rather introduce a strong DC and low frequency component (in particular P3 and other late/slow components). Highpass filtering will necessarily spread/smear this activity and systematically bias the amplitude of preceding (we use non-causal filters!) and subsequent components. In a paradigm with a P3 in one but not in another condition (e.g. active oddball) the peak/mean amplitude of the preceding N1/N2 might be systematically biased between conditions after highpass filtering (e.g. with a higher 1 Hz cutoff while the effects might be negligible with a 0.1 Hz cutoff).

My recommendation is to systematically explore the effects a filter has on your data. One approach is to inspect and analyze the signal that was removed from the data by filtering, i.e. the difference between filtered and unfiltered data. Another important approach is to apply filters with different parameters and to compare the results (what you already did) and the effects on your parameter estimate. You can also filter test signals modeling your data and explore the observed filter effects/artifacts/distortions.

For your particular question you might want to read the already mentioned Acunzo paper and the recent paper by Tanner and colleagues (2015, Psychophysiology, DOI: 10.1111/psyp.12437). The point is whether your observed „drift" is really noise or evoked activity (I would rather guess that it is the latter) and whether removing it by filtering would bias your estimated parameter (following your description I would guess yes but this is really guessing without seeing the averaged data and knowing the dependent variable).

Hope this helps! Best,
Andreas

> Am 07.08.2015 um 00:58 schrieb Makoto Miyakoshi <mmiyakoshi at ucsd.edu>:
> 
> Dear Kevin,
> 
> > But what I'd really like to know is, does *my* 0.1hz data contain comparable drift to others' 0.1hz data?
> 
> Yes, your data look normal for 0.1Hz high-passed.
> I also recommend you change the time scale so that you can see 10 sec long of data in one screen to visually confirm if there is any wave whose half cycle (i.e. single peak or trough) goes beyond  5 sec.
> 
> Widmann et al. (2014), who advocates the merit of filter, wrote:
> 
> %%%%%%%
> Some authors argue against high-pass filtering (or restrict the applicable high-pass cutoff to frequencies as low as <0.1 Hz; in particular if estimating window mean or peak amplitudes; Acunzo et al.,2012; Luck, 2005) or low-pass filtering (in
> particular if estimating onset latencies; VanRullen, 2011). We certainly want to stress their point–care is needed–but, on the other hand, if the filter applied really increases the signal-to-noise ratio (as it should to motivate its usage) and does not systematically bias the to-be-estimated parameter, these values can be determined with greater precision with than without filtering.
> %%%%%%%
> 
> Exaggerated filter phobia is not useful.
> A common mistake is to think that
> 'Closer to the original, better the signal'
> 'Therefore, less use of filter, better the signal'
> Here, the underlying idea is 'minimizing the use of filter == maximizing signal fidelity' But it does not apply to our case; we are not playing music where the original source necessarily has the maximum signal fidelity. Andreas Widmann's point is 'Hey, what matters is signal-to-noise ratio; if filtering can improve it, why not use it rather than avoiding it?' This makes sense to me.
> 
> However, I'm not sure how to measure the signal to noise ratio.
> In our case, we have been using ICA + mutual information reduction to evaluate the goodness of preprocessing, but this is kind of ICA-centric view and may not be acceptable for others.
> 
> Makoto
> 
> On Thu, Aug 6, 2015 at 12:20 PM, Kevin Tan <kevintan at cmu.edu> wrote:
> Makoto, 
> 
> Thanks so much for the detailed reply. Indeed, given all the caveats, 0.1hz hi-pass produces 'better' ERP waveforms and is in-line with the literature. 
> 
> But what I'd really like to know is, does *my* 0.1hz data contain comparable drift to others' 0.1hz data? It's hard to tell since people don't publish their epoched EEG data. If it is comparable then I can feel much better. 
> 
> And yes, for ICA I use 1hz hipass then copy the resulting weights to 0.1hz data. 
> 
> Mohammed, 
> 
> I've tried detrend on continuous 0.1hz data with little effect. Do you mean detrend on individual epochs? Would that result in consistency issues across epochs?
> 
> Many thanks for the comments!
> 
> Best, 
> Kevin 
> 
> --
> Kevin Alastair M. Tan
> Lab Manager/Research Assistant
> Department of Psychology & Center for the Neural Basis of Cognition
> Carnegie Mellon University
> 
> Baker Hall 434 | kevintan at cmu.edu | tarrlab.org/kevintan
> 
> On Thu, Aug 6, 2015 at 7:50 AM, Mohammed Jarjees <m.jarjees.1 at research.gla.ac.uk> wrote:
> Dear Kevin Tan,
> Have you tried detrend function on 0.1 Hz filtered data. It may be help.
> Best Regards
> Mohammed Jarjees
> 
> ________________________________________
> From: Makoto Miyakoshi [mmiyakoshi at ucsd.edu]
> Sent: 05 August 2015 06:03
> To: Kevin Tan
> Cc: EEGLAB List
> Subject: Re: [Eeglablist] Too much drift with just 0.1hz high-pass filtering?
> 
> Dear Kevin,
> 
> In the Appendix of Rousslet (2012), you can find the shockingly bad effect of 1-Hz high-pass filter compared with 0.5-Hz or below. However, the filter function used here is pop_eegfilt, which is an old generation before Andreas Widmann redesign it, so for 500-Hz sampled data the filter order for 1-Hz is 75000. After Andreas's fix, it is 1651. 75000 is crazy. You need to keep it in mind. FYI, 1-Hz highpass with Hamming window and fiter order 75000 results in the transition bandwidth of 0.022; Andreas's heuristics suggests 1.
> 
> > Is the drift in 0.1hz data ok? I get 'better looking' ERP waveforms & more robust differences between conditions in 0.1hz data – I'm worried this is mostly due to drift.
> 
> Sadly, it is often the case that our eye are trained for something that does not have a good ground. Rousslet (2012) showed 'distortion' of ERP waveforms after 1- and 2-Hz highpass (but again with old function). However, if you know Gibb's phenomenon etc and the exact filter order you are using, you would find nothing wrong. Same goes for your/my impression of the 0.1-Hz high-passed data. I would say the waves are drifting and at least bad for the purpose of ICA. But for the researchers of EEG infraslow oscillations, they would say oh it's a good data.
> 
> So there is no good or bad. After averaging several hundred trials, the apparently drifting signals (to my eyes) will produce 'better' ERP waveforms, thanks to the averaging process. If you say you will run ICA on the 0.1-Hz highpassed data, I'd say you shouldn't.
> 
> Stephan Debener's solution is that you apply 1-Hz high-pass on the data, run ICA, copy the weight matrix to the 0.1-Hz high-passed data.
> 
> Makoto
> 
> On Tue, Aug 4, 2015 at 10:20 PM, Kevin Tan <kevintan at cmu.edu<mailto:kevintan at cmu.edu>> wrote:
> Hi all,
> 
> There are numerous papers that conclude that >0.1hz high-pass filtering distorts ERPs. However, I notice a lot of remaining drift after 0.1hz hi-pass, especially compared to 1hz hi-pass. I'm using a BioSemi Active2 128ch.
> 
> 0.1hz hi-pass:
> https://cmu.box.com/s/1uafw786miveruz85ycj3taxflzg7p7f
> 
> 1hz hi-pass:
> https://cmu.box.com/s/t1dbzntjcwdrsp734m949xnzmycvpw5p
> 
> Is the drift in 0.1hz data ok? I get 'better looking' ERP waveforms & more robust differences between conditions in 0.1hz data – I'm worried this is mostly due to drift.
> 
> The 1hz data has ERP 'distortions': negative slope from start of epoch until P1 & negative deflection of later components. Thus, I'm not comfortable with either of the filters.
> 
> The screenshots show data run only through 1) PREP pipeline 2) high-pass filtering 3) epoching. The final cleaned data shows the same drift.
> 
> My full preproc stream:
> 
> ICA dataset:
> - Load PREP'd data
> - 1hz hi-pass
> - Epoch
> - Epoch rejection
> - Extended ICA (binica)
> - Determine bad ICs
> 
> Final dataset:
> - Load (unfiltered) PREP'd data
> - 0.1hz hi-pass (tried 1hz for comparison too)
> - Epoch
> - Generate ICs from matrices of ICA dataset
> - Remove bad ICs determined from ICA dataset
> - Epoch rejection
> - DIPFIT
> - Make ERPs
> 
> Any input would be much appreciated!
> 
> Many thanks,
> Kevin
> --
> Kevin Alastair M. Tan
> Lab Manager/Research Assistant
> Department of Psychology & Center for the Neural Basis of Cognition
> Carnegie Mellon University
> 
> Baker Hall 434<https://www.google.com/maps/place/40%C2%B026%2729.5%22N+79%C2%B056%2744.0%22W/@40.4414869,-79.9455701,61m/data=!3m1!1e3!4m2!3m1!1s0x0:0x0> | kevintan at cmu.edu<mailto:kevintan at cmu.edu> | tarrlab.org/kevintan<http://tarrlabwiki.cnbc.cmu.edu/index.php/KevinTan>
> 
> _______________________________________________
> Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
> To unsubscribe, send an empty email to eeglablist-unsubscribe at sccn.ucsd.edu<mailto:eeglablist-unsubscribe at sccn.ucsd.edu>
> For digest mode, send an email with the subject "set digest mime" to eeglablist-request at sccn.ucsd.edu<mailto:eeglablist-request at sccn.ucsd.edu>
> 
> 
> 
> --
> Makoto Miyakoshi
> Swartz Center for Computational Neuroscience
> Institute for Neural Computation, University of California San Diego
> 
> 
> 
> 
> -- 
> Makoto Miyakoshi
> Swartz Center for Computational Neuroscience
> Institute for Neural Computation, University of California San Diego
> _______________________________________________
> Eeglablist page: http://sccn.ucsd.edu/eeglab/eeglabmail.html
> To unsubscribe, send an empty email to eeglablist-unsubscribe at sccn.ucsd.edu
> For digest mode, send an email with the subject "set digest mime" to eeglablist-request at sccn.ucsd.edu