Dear Christopher,
hence I am far away from being an expert in DCM i will give you my
point of view to some of Your questions, hoping that Your image of
DCM will get sharper.
1) Of course event related has a lower SNR ratio. Therefore the
Parameter estimation with an blocked of epoch design will be more
reliable. So, if You are interested in examining structures wich are
sending a relatively low signal - it is better to use blocked or epoch
design.
2) This is because of the nature of the DCM model. In difference to
the GLM model used in a normal contrast analysis, DCM tries to model
the BOLD response, assuming a model of neuronal responses. This
is mainly done by a set of nonlinear diffential equations, the parameter
estimation for the DE sets is done by the EPI images. So the DCM and
GLM (polynominal linear equations) models have very different
ancestors in mathematical theory. Because a GLM analysis is based
on common classical test theory statistics, it is valid to code Your
regressors like in a factorial design of CTT. DCM, on the other side
needs Your regressors (system pertubing variables) isolated.
3) Please look in the SPM-List Archive, recently there where some
posts about group analysis and DCM.
4+5+6) I'll leave that for the experts.
7) This reflects the neuronal activity assumed by the DCM model.
Since we can only measure the BOLD response, but there is good
evidence that it is strongly connected with the neuronal activity
of brain areas, we needed a model. One part of DCM models the
neuronal responses on the basis of Volterra kernels.
Sincerely,
Ferenc
------------------------------------------------------------
Ferenc Acs
Lehrstuhl Prof. Dr. M. W. Greenlee
Institut für Psychologie
Universität Regensburg
93040 Regensburg
Tel. +49 (0)941 943 3582
Fax +49 (0)941 943 3233
http://www.psychologie.uni-regensburg.de/Greenlee/team/Acs/acs.html
>>> Christopher Summerfield <[log in to unmask]> 12/15/04 11:36 >>>
dear listmembers,
I have a number of [naive] questions about dynamic causal modelling. any
input would be very much appreciated.
1) are there obvious advantages/disadvantages to different types of design
(event, blocked) and trial parameters (ISI, ITI) in DCM analyses, other
than having a rapid TR?
2) It seems from Ollie Hulme & Barrie Roulston's powerpoint presentation
(very useful, thanks: www.fil.ion.ucl.ac.uk/spm/doc/mfd/dcm_practical.ppt) that
it's often best to use a different design matrix for DCM analysis than you
might use for a normal SPM analysis. in the example they give, a 2x2
factorial motion/no motion x attention/no attention is reduced to 3
regressors: no motion, motion and attention. why is this?
3) this has been asked before, but there seem to be conflicting accounts
about how to deal with it. What is the best way to use DCM for random
effects analysis? Would it be reasonable to exactly match VOIs across
subjects/sessions by using an absurd P threshold (for example, p<1) such
that exactly the same VOI could be extracted, by using the activation
cluster from
the group as an inclusive mask? assuming the scans are normalised, would
this be a reasonable approach?
4) I am struggling to conceptualise what is meant when the connectivity
values in the output matrices A,B or C are negative. Does this mean that
an area is exhibiting 'less than zero' modulatory influence on another
region?
5) any thoughts on why would I get a warning: 'Returning NaN for out of
range arguments'
(these seem to occur in the probability matrices when the connectivity is
zero).
6) are there any constraints on how you define your VOIs? for example, if
I have 2 trial types, and my first VOI is centered on the peak voxel for
A>B, and the other is the same for B>A, and another is the peak voxel
responding to A *and* B, am I not introducing artificial correlation into
my timeseries...should I be accounting for this in my model?
7) when I display 1st order kernels, what are 'neuronal responses'? (left
graphs).
thank you very much.
chris
Christopher Summerfield
[log in to unmask]
|