[Eeglablist] Effect of anti-aliasing low-pass filter on connectivity analysis

Tue Aug 25 10:35:15 PDT 2015

Dear Andreas,

Thank you. I found it in Barnett and Seth (2011) page 416 'We have shown
that a primary cause is the large increase in empirical model induced by
filtering; high model orders become necessary in order to properly fit the
modified aspects of the power spectrum (low power in stop band, steep
roll-off, etc.).'

Actually, this made me wonder whether ICA eases or complicates the solution
of Granger Causality analysis; ICA-decomposed signals tend to have unique
frequency spectra and thus increase the variance across signals. I remember
when I cluster ICs just by spectra at the group-level analysis, the result
was surprisingly well clustered spatially as well. Judging from this, I
guess ICA complicates Granger Causality analysis. Maybe that's why I can
never get a satisfactory results in model validation in SIFT.

Thank you very much Andreas! Your input is very precious all the time.

Makoto

On Tue, Aug 25, 2015 at 2:50 AM, Andreas Widmann <widmann at uni-leipzig.de>
wrote:

> Dear Makoto,
>
> > I have been wondering about this but keep forgetting to ask you about it:
> >
> > > no stopband (below Nyquist)
> >
> > Why do you need to avoid having a stopband below Nyquist?
> > I appreciate your kind help.
> This refers to the notion by Barnett and Seth (2011; quoted below) that
> „low power in stop band“ is one primary cause for the large increase in
> required empirical model order and thus, responsible for the problems
> introduced by filtering in Granger causality analysis.
>
> This only applies to GC analysis. For standard ERP-analysis it might be
> recommended to have a considerably lower anti-aliasing filter cutoff and a
> stopband below Nyquist to improve signal fidelity. In the repaired
> pop_resample function anti-aliasing cutoff and transition band/roll-off can
> be adjusted on the command line.
>
> Hope this helps! Best,
> Andreas
>
> > Makoto
> >
> >
> > On Wed, Jun 24, 2015 at 4:02 AM, Andreas Widmann <widmann at uni-leipzig.de>
> wrote:
> > Dear Makoto,
> >
> > to my understanding filter causality and Granger causality are not
> directly related. The output of a causal linear filter is identical to the
> output of a non-causal linear filter but shifted on the time axis (delayed;
> but equally across channels and bands!). Non-linear (here min phase)
> filters distort the phase spectrum and should, to my understanding, not be
> used for GC analysis.
> >
> > Barnett and Seth (2011, J Neurosci Meth) show that GC is in theory (but
> not in practice) invariant under filtering. They do recommend filtering to
> achieve stationarity (e.g., drift, line noise; also in recent 2015 J
> Neurosci paper). The main problem with filtering and GC is the increase in
> required model order ("We have shown that a primary cause is the large
> increase in empirical model induced by filtering; high model orders become
> necessary in order to properly fit the modified aspects of the power
> spectrum (low power in stop band, steep roll-off, etc.).“).
> >
> > That is, to my understanding for a carefully designed anti-aliasing
> filter (linear, zero-phase) the impact should be limited. The anti-aliasing
> filter as it is implemented in the repaired pop_resample function (in
> develop but not yet in eeglab13 branch) will have no stopband (below
> Nyquist) and a rather shallow roll-off (and low order) with default cutoff
> (fc = 0.9 * Nyq) and transition band width (df = 0.2 * Nyq). The cutoff and
> transition band width can be manually defined by the user, so you can try
> to apply a more shallow roll-off, e.g. with fc = 0.8 and df = 0.4. This
> conclusion should, however, be actually tested with a simulation. From a
> practical perspective any M/EEG signal has been filtered with an
> anti-aliasing filter.
> >
> > > As the ERP handbook by Luck (or his other book) recommends,
> anti-aliasing should better have the margin of 4-5 times of the new
> sampling rate e.g. if you downsample signlas to 250 Hz, anti-aliasing
> low-pass at 125 Hz is the standard, but recommendation is 75 Hz or even 50
> Hz. Well, I haven't tested it myself so I am not sure what bad it would do
> if I use 125 Hz (any comment on this, anyone?) but in this case, I guess
> the anti-aliasing low-pass filter does affect the subsequest connectivity
> analysis--am I correct (assuming that I analyze EEG up to 50 Hz)?
> > To my understanding this conservative oversampling ratio is intended to
> improve signal fidelity (resolution, noise) rather than anti-aliasing
> alone. Given the result demonstrated by Barnett and Seth I would not
> recommend applying a lowpass filter with a stopband below Nyquist.
> >
> > Best,
> > Andreas
> >
> > > Am 24.06.2015 um 02:59 schrieb Makoto Miyakoshi <mmiyakoshi at ucsd.edu>:
> > >
> > > Dear Iman,
> > >
> > > > Using causal filter may adversely effect the direction of information
> > >
> > > flow in the GC analysis. It is recommended that one use a
> > >
> > > non-causal filter (for example, finite impulse response filters) with
> > >
> > > zero phase lag
> > >
> > >
> > > Really? The impulse response of the non-causal FIR filter spreads in
> both ways in the time domain, which means info of future events leak to
> past... I thought using causal filter with minimum phase makes more sense.
> > >
> > > Makoto
> > >
> > > On Tue, Jun 23, 2015 at 4:29 PM, Iman Mohammad-Rezazadeh <
> irezazadeh at ucdavis.edu> wrote:
> > >
> http://journal.frontiersin.org/article/10.3389/fnhum.2015.00194/abstract
> > >
> > >
> > >
> > > Using causal filter may adversely effect the direction of information
> > >
> > > flow in the GC analysis. It is recommended that one use a
> > >
> > > non-causal filter (for example, finite impulse response filters) with
> > >
> > > zero phase lag (Mullen et al., 2012, Coben and Rezazadeh, 2015)
> > >
> > >
> > >
> > >
> > >
> > > From: eeglablist-bounces at sccn.ucsd.edu [mailto:
> eeglablist-bounces at sccn.ucsd.edu] On Behalf Of Makoto Miyakoshi
> > > Sent: Tuesday, June 23, 2015 2:07 PM
> > > To: Vito De Feo
> > > Cc: EEGLAB List
> > > Subject: Re: [Eeglablist] Effect of anti-aliasing low-pass filter on
> connectivity analysis
> > >
> > >
> > >
> > > Thank you Vito for your response. Forgive me to ask you one more
> question.
> > >
> > >
> > >
> > > As the ERP handbook by Luck (or his other book) recommends,
> anti-aliasing should better have the margin of 4-5 times of the new
> sampling rate e.g. if you downsample signlas to 250 Hz, anti-aliasing
> low-pass at 125 Hz is the standard, but recommendation is 75 Hz or even 50
> Hz. Well, I haven't tested it myself so I am not sure what bad it would do
> if I use 125 Hz (any comment on this, anyone?) but in this case, I guess
> the anti-aliasing low-pass filter does affect the subsequest connectivity
> analysis--am I correct (assuming that I analyze EEG up to 50 Hz)?
> > >
> > >
> > >
> > > Makoto
> > >
> > >
> > >
> > > On Mon, Jun 22, 2015 at 9:31 AM, Vito De Feo <
> vito.defeo at zmnh.uni-hamburg.de> wrote:
> > >
> > > Dear Makoto,
> > > this will not affect the connectivity analysis if the frequency of
> interest are far from the Nyquist frequency. For example if you downsample
> to 500 Hz (Nyquist freq = 250 Hz) you will have no problem in the band
> 0-100 Hz.
> > > Best
> > > Vito
> > >
> > >
> > > Il giorno 20/giu/2015, alle ore 00:28, Makoto Miyakoshi ha scritto:
> > >
> > > > Dear List,
> > > >
> > > > If I use zero-phase low-pass filter for anti-aliasing, does it
> affect the subsequent connectivity analysis? I ask this because EEGLAB
> pop_resample() automatically applies it. If it does, is there a workaround?
> Should I use minimum phase causal filter for anti-aliasing?
> > > >
> > > > --
> > > > 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
> > >
> > >
> > >
> > >
> > > --
> > > Pflichtangaben gemäß Gesetz über elektronische Handelsregister und
> Genossenschaftsregister sowie das Unternehmensregister (EHUG):
> > >
> > > Universitätsklinikum Hamburg-Eppendorf
> > > Körperschaft des öffentlichen Rechts
> > > Gerichtsstand: Hamburg
> > >
> > > Vorstandsmitglieder:
> > > Prof. Dr. Burkhard Göke (Vorsitzender)
> > > Prof. Dr. Dr. Uwe Koch-Gromus
> > > Joachim Prölß
> > > Rainer Schoppik
> > >
> > >
> > >
> > >
> > >
> > >
> > > --
> > >
> > > 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
> >
> >
> >
> >
> > --
> > 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
>
>

-- 
Makoto Miyakoshi
Swartz Center for Computational Neuroscience
Institute for Neural Computation, University of California San Diego
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://sccn.ucsd.edu/pipermail/eeglablist/attachments/20150825/11a51b79/attachment.html>