Dear Tali,
>Dear SPMers
>
>I have a number of DCM questions-
>
>SMOOTHING:
>Is there any reason to use a small smoothing kernel in DCM analysis, more
>than in the cnventional analysis (I saw 3.75 and 5 mm in the published
>papers)?
DCM itself does not care what spatial smoothing kernel was used during the
initial SPM. All DCM needs are some time series that represent distinct
brain regions. Thus, your spatial smoothing should be chosen such that it
matches the desired anatomical interpretability of your data. For example,
if you wanted to include a sulcal region in your model and wanted to make
sure that the associated time series mainly represents sulcal voxels, you
should choose a small kernel. The same applies to mesial areas (e.g.
anterior cingulate cortex) if you want to model corresponding regions from
both hemispheres and want to avoid overlap of hemispheric time series. If,
however, all your areas are broadly defined regions on the lateral surface
of the brain, a larger kernel is likely to be fine.
BTW, similar thoughts apply to the question what sphere radius should be
chosen for data extraction by the VOI tool in SPM.
>INTERPRETATION:
>1- How are the 'intrinsic connnections' calculated? Is it the connections
>during baseline (the tails of the hrf)?
No. The strength of the intrinsic connections (which are stored in the A
matrix) reflects coupling strength in the absence of contextual modulation,
computed across the entire time series. Take a look at equation 2 in the
DCM paper (Friston et al., NeuroImage 2003): a given parameter from A
describes that component of the change in the neuronal state of a region
which depends on the neuronal state in other region(s).
>2 - What is the meaning of the 'b' parameters: Is it the STRENGTH of the
>connection during the external input, or is it the CHANGE in strength
>during the external input relative to baseline?
For a given contextual input u, the parameters from the associated B matrix
describe how coupling strengths between regions are changed (scaled)
whenever u is present.
>In other words: Are the 'b' parameters comparable across different models
>of the same subject?
No - see below.
>across subjects?
Yes - if the model is identical. This forms the basis for second-level
analyses (see below).
Parameters are not directly comparable across different models. Changing a
model by adding or removing parameters leads to different dynamics, and the
values of the remaining parameters will change accordingly. By means of
Bayesian model selection procedures that evaluate both model fit and
specification costs you can, however, determine whether or not a given
model is better than another one (such tools are currently under
development for DCM). For example, if you have a hypothesis about a
specific parameter P in a given model M1, you could compare M1 to a reduced
model M2 where P has been removed. If the decrease in model fit outweighs
the advantage of having one parameter less, this would speak to the
importance of the mechanism (e.g. contextual modulation of a connection)
that was represented by P. This approach is conceptually similar to the
chi-square-diff tests in structural equation modelling.
>GROUP ANALYSIS:
>Is there a way to do group analysis in DCM2? Which '.img' files can be
>imported to the second level?
The second-level analysis simply operates on the parameter estimate(s) that
represent your hypothesis. For example, after having analysed each subject
separately, you can make group inferences by simply performing one- or
two-sample t-tests on the coupling parameter(s) you are interested in. See
also:
http://www.jiscmail.ac.uk/cgi-bin/wa.exe?A2=ind0304&L=spm&P=R4776&I=-1
http://www.jiscmail.ac.uk/cgi-bin/wa.exe?A2=ind0309&L=spm&P=R18908&I=-1
Best wishes,
Klaas
___________________________________________
Dr Klaas Enno Stephan
Wellcome Dept. of Imaging Neuroscience
12 Queen Square, WC1N 3BG, London, UK
phone: +44-20-78337485
e-mail: [log in to unmask]
web: http://www.fil.ion.ucl.ac.uk/~kstephan
|