[Eeglablist] SIFT RPDC Connectivity

Fri Jun 17 10:00:11 PDT 2022

Dear Sampath,

> 1)Is ersp shown in diagonal RPDC plot similar to ersp estimated using
'newtimef ' from eeglab toolbox??  if yes my results from sift RPDC ersp
and 'newtimef ' ersp are not matching.

No, it is not. Actually, I once asked the same question to Tim, and he said
no.
It is because the input data preprocessed for rPDC, in your case, is
different from what you usually do for preprocessing for calculating ERSP.

> 2) What is the influence of model order in estimating connectivity??-
this sounds more general question but for my data, I get really different
results (image link can be found below) when I use model order 21 (elbow
point) than model order 5 (min point). The consistency,  whiteness, and
stability were fine in all cases.

First of all, a calculation of rPDC and other effective connectivity
measures depends on the choice of sampling rate.  This may be a bit
counterintuitive. This is NOT only because of the Nyquist frequency is
changed and PSD up to the Nyquist  affects the decomposition, but
(probably) because the definition of a single time point is changed. For
example, model order 10 at 500 Hz is 20 ms, but at 100 Hz it is 100 ms. I
do not have much experience in calculating ARmodel, but I predict the same
problem should arise there.

> 3) If the model order makes a drastic change (one can see in images from
the link) which one should I use???

The developer Tim always stuck to the lowest sampling rate possible, which
was typically at 100 Hz, and he used to show the result only up to 20 or 30
Hz. In your case, you want to show 40 Hz, for which downsampling to 100 Hz
may be a bit suboptimal.

One of the merits of choosing 100 Hz is that it is below (American) line
noise frequency at 60 Hz. This helps to keep PSD smooth up to the
Nyquist frequency. If you want to downsample the data to 200 Hz, you can
analyze 40 Hz and its second harmonics (by the way are you using click
trains?) but line noise frequency is included. You can use CleanLine,
ZapLine, or spectral interpolation to remove the sharp peak/dip at your
line noise frequency before calculating rPDC.

Here are some tips:

   - Do not show anything on the diagonal line plots. Turn them off.
   Otherwise, color scales in the off diagonal plots are biased and typically
   you can't see anything as your current example appears.
   - Define a baseline period to subtract the mean value from all the data
   points for each freq bin. This really helps for visualization.
   - Try both 'min' and 'elbow' when you optimize the model order. 'elbow'
   almost always works while 'min' sometimes gets stuck to the right end
   (i.e., no global min found within the specified range) By the way, Tim used
   to use Hannan-Quinn because it is often most balanced.
   - Do not use piecewise detrending. I had a discussion over this method
   with Andreas in the past. He convinced me that using a standard FIR
   high-pass is better. See also Seth's paper for the use of FIR filter for
   preprocessing. Your current choice 'linear' is good.
   - Tim admitted that rPDC sometimes shows a bug near the highest
   frequency range. I also confirmed this several times. If you notice it,
   just ignore these values. This is another reason why I can't recommend
   downsampling to 100 Hz when your target is 40 Hz, it is too close.
   Alternatively, use dDTF08. This is the other good one you can use.
   - There are other things that may be helpful for you. Check out
   https://urldefense.com/v3/__https://eeglab.org/others/Firfilt_FAQ.html*q-for-granger-causality-analysis-what-filter-should-be-used-11212020-updated__;Iw!!Mih3wA!BcITtnngLqfwmvmmsc1I6-V_lJ2Q5dA82rKtocFC7SgQyZbqgE1ww4bWeK8ow60azopdsvTBWtHzB0FpxpKvfJRRqyw$ 
   - When you choose ARfit instead of Vieira-Morf, it sometimes gets stuck
   and does not do anything. I often prefer to use ARfit because of
   simplicity, but when it happens I can't.

Let me know how the above suggestions work for your project.

Makoto

On Fri, Jun 17, 2022 at 9:02 AM Sampath Thoutanahalli Kapanaiah <
sampath.kapanaiah at uni-ulm.de> wrote:

> Hi Everyone,
>
> My name is Sampath, PhD student from the Institute of applied physiology,
> Ulm University, Germany. I am using sift and adapted script provided in the
> script folder to estimate RPDC. My data is recorded from depth single
> multi-site electrodes (PFC, A1..etc.,.) in mice recorded during 40hz ASSR
> 100-stimulation trials and I have 3 questions regarding the results.
>
> 1)Is ersp shown in diagonal RPDC plot similar to ersp estimated using
> 'newtimef ' from eeglab toolbox??  if yes my results from sift RPDC ersp
> and 'newtimef ' ersp are not matching.
>
> 2) What is the influence of model order in estimating connectivity??-
> this sounds more general question but for my data, I get really different
> results (image link can be found below) when I use model order 21 (elbow
> point) than model order 5 (min point). The consistency,  whiteness, and
> stability were fine in all cases.
>
> 3) If the model order makes a drastic change (one can see in images from
> the link) which one should I use???
>
> In all the above cases parameters are kept constant,
>
> WindowLengthSec     = 0.3;
> WindowStepSizeSec  = 0.01;
> algorithm                   = 'Vieira-Morf';
> windowingType         = 'hamming';
> detrend_method       = 'linear';
> normalise_method   = {'time' 'ensemble'};
>
> link to result figures -
> https://urldefense.com/v3/__https://cloud.physio.uni-ulm.de/s/yc7dq5yXW8wQ2Kb__;!!Mih3wA!GfspMDkTjVGCvdSCRo9X8sluGVU3mcFGoLqQAda3Ou5zr2BzJWvqWPX6-TlVtU1cFJpf-ge_MKGsb4LTWZPtNMKEnXvlYzUUc6zt$
>
> Thanks for your time in advance. I will be happy to provide more
> information or a dataset if necessary.
>
> Best Wishes,
> Sampath
>
> _______________________________________________
> 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
>