[Eeglablist] about selecting AMVAR model order

[Baris Demiral]

Dear Hui-Bin,

I am also using SIFT, and I think model fitting is the most difficult
part of the process. I use almost the same values for those parameters
you mentioned, and even though I can "fit" the model with lower model
orders, the whiteness test generally fails or cannot reach good values
in the most of the tests (Box-pierce, Ljung etc.) Overall, the process
of finding the best model fit is a bit cumbersome and time consuming,
and there is not a consistency between the model fit tests, one says
good other says not good.

My understanding of MVAR modeling is that this approach needs "very
clean" and "stable", and "stationary" data... Well, what do these all
mean ? I think this means that for the model fit very well, you need
to detrend, take  out the slow waves etc. Another very important fact
is: The number of components or electrodes you want to use! If you
have many electrodes or ICs, the model will have difficulty to fit
into the data. Also, I THINK,if the ERPs have bi-phasic behaviour, the
model fitting gets worse. If you have only P3 or N400 it might work
ok, but if you have two or more components and a lot of variability
within the epoch or across the epochs, things get very hard to handle.

So far, all we read about connectivity is related to the simulation
studies (NOT real data!!), or sensor-based connectivity, where people
generally do things for the sake of doing connectivity and ignoring
problems like volume conduction. I am still positive though, and I
strongly favor SIFT and source based approaches (please see sLORETA,
which also provides source based connectivity measures), but this
time, individual differences in ICs and sources become a big problem,
unless we agree to accept to have a population ICs and find the
clusters/domains common for the population first and then finding the
individual ICs (similar to GIFT approach, or similar to
Microstates-Cartool approach). Tim is working on to apply Bayesian
approach to solve the problem of non overlapping IC and dipoles in
connectivity analysis, and I think we will one day be able to do those
great stuff by starting and finding individual dipoles extending to
the population.
Baris



On Mon, Apr 23, 2012 at 8:55 AM, Hui-bin Jia <420247417 at qq.com> wrote:
> Hi  everyone,
>
>     In these days, I'm trying to do effective connectivity using the
> datasets included in the EEGLAB tutorial (e.g. the animal categorization
> task).
>
>    When I was processing these datasets, I used SIFT. Unfortunately, in
> order to make sure most of the windows pass the ACF test, I had to use a
> very large model order (e.g. more than 70). But even the model was that
> high, very few windows passed the portmanteau tests ( e.g. 1 out of 108
> windows can pass these test).
>
>    The parameter preferences were the same as the parameter preferences in
> the SIFT manual.
>
>    The value of Option "EpochTimeRange" was [-1 1.25], the value of Option "
> Normalization method" was "ensemble" and "time", the value of Option "Select
> MVAR Algorithm" was " Vieira-Morf ", the value of Option "Window
> length(sec)" was "0.4", and the value of Option "Step Size(sec)" was "0.01".
>
>    I have tried my best to conquer this problem. But my efforts were
> useless. So I sincerely hope that I can get some suggestions from you. Thank
> you!
>
>
>
> Yours, Hui-bin Jia
>
>
> _______________________________________________
> 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



-- 
Ş. Barış Demiral, PhD.
Department of Psychiatry
Washington University
School of Medicine
660 S. Euclid Avenue
Box 8134
Saint Louis, MO 63110
Phone: +1 (314) 747 1603