It is always a little difficult to comment on specific DCMs without knowing the results of the corresponding SPM analysis. I thus only briefly summarize a few general rules.
SPM and DCM have the same goal in that they try to explain locally measured fMRI responses, but they do so in different ways. In a general linear model, each region (voxel) is modeled as a linear combination of all experimental influences (represented as inputs or regressors), but there are no interactions (connections) amongst the regions. In DCM, regional activity also depends on experimental inputs, but their effects may not only affect the regions directly (in terms of driving inputs), but can also be mediated indirectly via connections. That is, any input that enters the modelled system somewhere will have some impact on any other region in the system, but the degree of this impact depends on where it enters and
what strengths the intermediate connections have. Therefore, if you are including a region that shows a specific experimental effect (e.g. a main effect for a given experimental factor), then your DCM must include inputs representing this differential effect somewhere in the network.. Often, these inputs will be on the afferent connections to the region showing a main effect, but it could also be on more remote connections. If you are trying to explain the interaction that you observed in a particular region, then your DCM must have some way of changing/modulating the effect of one experimental factor by another experimental factor. For example, if you drive one particular region with two inputs representing the two levels of factor 1, and you then modulate the efferent connection originating from this area with inputs representing the two levels of factor 2, you might be able to model an interaction in the target region.
Of course, there is no guarantee that your parameter estimates will actually indicate that the interaction of the target region is arising through this particular mechanism (it could be mediated through other mechanisms that you did not include in the model; this can be assessed by model comparison), but in any case, you must ensure that your model structure could potentially accommodate the experimentally induced effect you are trying to explain. You could always ask yourself: How would I have to choose parameter values (qualitatively) such that, under a given input function, I could reproduce the type of main effect or interaction that I observed in my general linear model. There is a short discussion on this issue in Stephan et al. 2007, Journal of Biosciences (http://www.fil.ion.ucl.ac.uk/spm/doc/papers/Stephan_JBiosci_32_129_2007.pdf), but only in the context of a simple two area model. For larger models, there are many
alternative ways how a particular main effect or interaction could be explained by different DCMs; this then becomes an issue of model selection.
Von: Ian Ballard <[log in to unmask]>
An: [log in to unmask]
Gesendet: Mittwoch, den 8. Juli 2009, 17:38:42 Uhr
Betreff: [SPM] DCM Design
I have been reading the posts about factorial DCM design and I am still a little confused as how to optimally design my DCM
In my block-design experiment, there are two main trial types, self and charity, and two conditions, high value and neutral. There are two questions I am interested in:
1. the structure of the network in the self trials and the bilinear effects from the high value condition
2. how the network for these trials compares to the network for the charity trials.
I can think of two possible ways to proceed:
- Construct a model where self-high and self-neutral cues are combined and modeled as the driving input and the self-high condition acts as contextual, modulatory variable. The t-contrast I would use for extracting VOIs would be (high – neutral) self. Then I construct a second set of models exactly the same way, except replacing self with charity.
- Construct one model which has all stimuli as driving inputs and two modulatory contexts: self high and charity high. In this case, it is not clear to me which contrast I should use for extracting VOIs. I’ve read that one should include all experimental conditions in a single DCM, but it seems like this is ignoring the interesting possibility that the network structure may change according to the type of task.
Thanks for any help,