thanks very much Klaas for your replies.
I have a couple more questions for when you have a moment...
1) to carry on the thread from your reply to (2) below - if I understand
correctly, then, discrepancies between the GLM and the C matrix can occur
for 2 reasons: either your HRF doesn't look like the canonical HRF, or you
have a further input from another region which might be modulating the
response. But just to be clear, could I give you a specific example?
I have a DCM with 3 regressors and 3 regions. The design is a balanced 2x2
factorial design with stimulus and context factors; the regressors are
thus (1) photic i.e. all visual stimuli (2) stimulus A and (3) context A.
The first VOI is in V1; it is defined as the voxel responding maximally to all visual stimuli
in the GLM and is very statistically robust. The response in this region
is a canonical-looking HRF with a bit of a deep undershoot. However, in
my DCM, when I model a driving input from
'photic' to V1 in the C matrix, this effect - although highly significant
- is on balance negative (in the rfx analysis t is about -7), and the
corresponding 'neuronal responses' are all negative-going curves.
So, is one interpretation of this that although the response
evoked by the stimulus is positive, V1 is receiving 'negative'
(perhaps inhibitory) modulatory input from other regions?
2) my second question is about how to use DCM when your GLM-defined
regions are negative deflections from baseline (i.e. effects of 'less
deactivation' in one condition than another). Imagine I have a study
where context A is my condition of interest, and context B is a
control of some sort, and my VOI is 'less deactivated' for A than B.
Whereas in the GLM analysis
one would typically model stimuli types A and B as separate regessors
(so that they can be compared), in the DCM design matrix - if I
understand correctly - I would simply model the effect of context A (or
stimulus A, or whatever). This means that the timeseries at my VOI is
likely to be negatively correlated with my regressor
(although, paradoxically, my control condition is even more strongly
negatively correlated). In this case, should I do as you say - feed in a
negative neural state to the model? (I presume this means putting a -1 in
the C matrix??). What are the implications of this for my DCM model? Is
it reasonable to assume that, if the contrast is mainly defined by a
deactivation in B rather than an activation in A, then A is unlikely to
drive that voxel (given the caveats you describe)?
Many thanks -
Chris
> are two main reasons. First, remember that the
> hemodynamic model in DCM is not equivalent to
> convolution with a canonical HRF but can cover
> more general forms of hemodynamic responses that
> you would, in a GLM, model as a linear
> combination of basis functions. If you happen to
> have a biphasic BOLD response, for example, which
> resembles the superposition of a canonical and a
> first derivative, then you can model this by
> feeding a negative neural state in the Balloon
> model, and this negative neural state may require
> that the driving input to that region becomes
> negative. Second, the influence of
> back-connections in the model may be able to
> explain parts of the dynamics in the input
> regions, thus changing the numerical value of the driving input.
> Put a different way: the analogy is tight if (i)
> the GLM uses several basis functions and (ii) the
> DCM feeds all inputs to all areas but does not
> use any connections between areas (compare Figure
> 2 in Stephan 2004, Journal of Anatomy).
>
> 3) Ideally, one would want to incorporate
> all experimental manipulations into the
> DCM. Whether or not this is necessary, depends
> on how much of the dynamics can be modelled by
> the unaccounted manipulations. The same kind of
> question arises for any other model, of course, including the GLM.
>
> 4) Not sure I understand this
> question. What do you mean by "there are more
> zeros than non-zero datapoints"? Note it is not
> unusual to find the B values to be rather close
> to zero. After all, they are estimated using shrinkage priors.
>
> Best wishes,
> Klaas
>
>
> At 13:05 28/04/2006, Christopher Summerfield wrote:
> >hi Will & Klaas & others....
> >
> >1) to repeat Amit's recent question - I think he didn't receive a
> >reply - how can one interpret a significant modulatory connection (B) when
> >the intrinsic connection (A) is not significant? is is something like: at
> >baseline, these regions show little or no coupling, but they begin to do
> >so following that particular task peturbation??
> >
> >2) Am I right in assuming that inputs in the C matrix should pretty much
> >track GLM effects - for example, a region defined by its GLM repsonsivity
> >to a regressor coding all visual inputs (like 'photic' in the worked
> >example) should be modulated by that regressor in the DCM analysis (like
> >'connect photic to V1/V2' in the worked example? If this is *not* the
> >case...are the DCM and GLM results contradicting each other?
> >
> >3) In the GLM analysis, variance which is not modelled in the design
> >matrix will either be captured by other regressors, or find its way
> >into the residuals. Does the same apply for DCM? I remember reading
> >something to the contrary on this list, or perhaps in the 2003 paper.
> >Practically, does this mean that it is not mandatory to model all visual
> >events? for example, in the worked example, there were presumably stimulus
> >events such as task instructions - were these modelled in 'photic'? or
> >were they just left out?
> >
> >4) My DCM.B matrix is peppered with zeros (for
> >those modulations specified in DCM.b). Why might the modulation be
> >exactly zero for a given subject/session? is it because the GLM data for
> >that subject/session/voxel don't reach a certain threshold? For some
> >modulations, there are more zeros than non-zero datapoints.
> >
> >many thanks,
> >Chris
> >
> >
> >
> >
> >
> >
> >
> >
> >Christopher Summerfield
> >[log in to unmask]
>
|