Dear Jack,
> Could you help me in specifying the definition of effective connectivity
> and in its evaluation by PPI & SEM ?
> Would EFFICIENT connectivity under a cognitive constrain be a good
> approximation of this idea (something that is more easy to translate in
> french) ?
I am not sure why you think that the term effective connectivity should be rephrased
into efficient connectivity.
> What are the classical electrophysiological methods to assess EC ?
> Joint peri-stimulus time histogram ? Cross-correlogram ... ?
You are absolutely right, the origins of effective connectivity stem from
multi-electrode recordings. In this case investigators were faced with an important
problem. If the firing of two cells was correlated, this does not necessarily mean
that they are connected. Another possibility is that the two cells receive a common
input from another cell. The initial idea around that problem was the so called shift
predictor, a technique in which correlations were calculated between different cells
over DIFFERENT trials. If the correlation between two cells is due to the stimulus
(i.e. common input) this should also be present in correlations over different
trials, because the stimulus is kept constant. A JPSTH is a more sophisticated way of
dealing with this problem.
> Using PPI, the decomposition of the data is done on a non orthogonal base.
> Could this lead to important problems due to the use of the least square
> estimator (problem of aberrant value) ? Is there a way to circumvent this ?
In a typical PPI, factor 1 ist a contrast C, factor 2 is the time-series in region B
and the interaction is Contrast x B. Think about the factor "region" as a substitute
for another orthogonal factor in a 2x2 design. Obviously it doesn't make a lot of
sense to define "region" using contrast C, because the time-series and the contrast
will be correlated.
> When performing a SEM analysis, do you have to take into account ALL the
> activated regions, and to perform an 'all brain' scan (otherwise you could
> have false path due to an unscanned area) ?
This is a problem in effective connectivity analyses, namely that you can only
account for common input, if you can take it into account (ie measure it).
> However, couldn't it be possible that a region unmodulated by the task
> modulate two other regions ?
This is certainly possible. Think about slow variations of activity in certain
modulatory mid-brain regions that do not follow a box-car function. A study by Dave
Chawla in Nature Neuroscience is a good example (although without effective
connectivity analyses). These areas can however influence cortical effective
connectivity. Even more important in this context are event-related fMRI designs,
where you can dissociate a cognitive set from certain events.
> My very limited experience in path analysis (I use spm_extract to have the
> different regional activity and use Set path - Statistica - for SEM with
> some convergence problem) lead me to the idea that more path you specify
> (under anatomical constrains), the best model fits you have, even if at the
> end only the interesting path remain statistically significant. An other
> way to say that, is that you HAVE TO specify ALL the known relation between
> area, even if they will not be significant, and removing this insignificant
> regions often lead to model rejection. Is it true ?
> One problem in SEM model acceptance, is that it is against the classical
> 'ho' hypothesis rejection concept. Could it be any possibility to evaluate
> something like the 'power' of the study ?
Model selection is complicated in SEM. There is no simple metric in finding the best
model. Ed Bullmore has worked on this topic. See
How Good Is Good Enough in Path Analysis of fMRI Data?
Ed Bullmore, Barry Horwitz, Garry Honey, Mick Brammer, Steve Williams, Tonmoy
Sharma
NeuroImage, Vol. 11, No. 4, Apr 2000, pp. 289-301 (doi:10.1006/nimg.2000.0544)
> The limited understanding that I have of SEM does not allow me to evaluate
> the possible problem due to autocorrelation in the time series. Is there
> any ?
Yes, there is. The degrees of freedom have to be adjusted because the observations
are not independent.
> At last, since I wanted to reproduce path modification under L-DOPA by a
> PKD patient, I faced the problem of pallidal spontaneous paramagnetism
> (fMRI time series). Thus for one region (in fact two, considering the inner
> and outer part of the pallidum), I have no signal. What would be your
> critics considering the use of latent variables to model its activity ?
If you don't have a signal, it is difficult to create a latent variable Latent
variables are like factors in a factor analysis, so at least you need areas that
would "load" on your latent variable. I cannot see how you would do that in the
absence of a signal.
> Would you have any other advise on the theme 'path modeling for beginners
> in functional imaging' ;-) ?
K. Bollen's book is very helpful and has all the equations.
> In a preceding post, Christian proposed to use the Voltera expansion to
> assess also non linear interaction, and you mention a possible paper on
> this subject (http://www.mailbase.ac.uk/lists/spm/1999-10/0006.html), is it
> published already (or accepted) ?
It is still under review.
> Would it be possible to use more than a first order temporal memory model
> using SEM ?
This is possible. Lagged terms and interactions (like in a Volterra series) could be
incorporated into a (complicated) SEM.
> At last, you proposed a release date around April for your effective
> connectivity toolbox, does it still valid ?
Unfortunately, no, it had to be postponed due to constraints on my time.
-Christian Buechel
--
Dr. Christian Buechel
Neurologische Universitaetsklinik, Haus B
Universitaets-Krankenhaus Eppendorf
Martinistr. 52
D-20246 Hamburg
Germany
Tel.: +49-40-42803-4726
Fax.: +49-40-42803-5086
email:[log in to unmask]
www.uke.uni-hamburg.de/kliniken/neurologie/
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|