Dear Siri,
Assuming your design has only a single factor (words vs. letter
strings) and that, except for the input region, you have extracted
your time series from regions which show a differential effect of
words vs. letter strings (or vice versa), you have two options:
1. Drive your input region by separate inputs representing words and
letter strings, respectively. This models how a main effect of words
vs. letters that is induced in the input area is conveyed, via the
specified connections, to other areas. It assumes, however, that all
regions show the main effect in the same direction (i.e. words>letter
strings or vice versa). In this case, which is identical to your
option 2 above, you then focus on the intrinsic connections (A matrix).
2. Drive your input region by a single input vector representing all
stimuli, regardless of whether they are words or letter strings, and
modulate the connections to other areas in the model by inputs
representing words and letter strings separately. This models how
stimulus-bound activity in the input region is conveyed, via the
specified connections, to other areas, but the degree to which
activity is passed on depends on whether the stimuli are words or
letter strings. This model is more flexible because it does not
assume that all regions show the main effect in the same
direction. In this case, you focus on the bilinear parameters (B matrix).
All the best,
Klaas
At 23:24 20/04/2006, you wrote:
>Hi Klaas...
>
>I have another question for you. I'm afraid it is actually getting
>to the point where I might need to fly myself there, or fly you out
>here, and monopolize a day or a week of your time (with full
>compensation, of course). Pardon the informality...
>
>Question: My ROIs comprise of signal timecourses that include two
>different kinds of events (words and letter strings) in
>event-blocks. I have DCMs assembled for two different subject groups
>(patients and controls), and my primary intent is to look for
>disease-related changes in effective connectivity. Darren has
>suggested that I experiment with the approach of examining intrinsic
>connectivity differences, leaving out modulatory effects in the
>model. For this, I would just drive the model with my stimulus onset
>vector, and compare intrinsic network architectures between groups
>by forwarding the A matrices for individual subjects to 2-sample
>t-tests. (This is somewhat similar to the approach used by Ethofer
>et al, 2005, but in many ways, different).
>
>I'm running into the following logistical conundrum: What would be a
>reason stimulus onset vector given that the signal is a response to
>two different stimulus types? If my interest is to look at intrinsic
>connectivity changes when the network is driven by word stimuli
>only, would it be reasonable to:
>1) only include onsets for word stimuli for my input vector, thus
>ignoring any effects letter strings might have on the signal over
>time? On the face of it, this seems like an invalid approach,
>because the signal timecourses include responses to two different
>kinds of stimuli, half of which are not at all being modeled in the DCM.
>2) Or, would it be more reasonable to include separate vector onsets
>for word stimuli and letter stimuli, thus adequately modelling the
>experimental inputs, and examining the intrinsics for the network as a whole?
>3) Or, still yet, would it make sense to include all stimuli in a
>single vector, thus removing any differential effect of words vs.
>letter strings from the model?
>
>Any guidance would be appreciated! Thanks so much!
>
>Siri
>
>
>Sreepadma Sonty
>Cognitive Neurology and Alzheimer's Disease Center (CNADC)
>Northwestern University Feinberg School of Medicine
>320 E. Superior Ave., Searle 11-470; Chicago, IL 60611
>[log in to unmask]
>
>
><http://us.rd.yahoo.com/mail_us/taglines/postman3/*http://us.rd.yahoo.com/evt=39666/*http://beta.messenger.yahoo.com>Yahoo!
>Messenger with Voice. PC-to-Phone calls for ridiculously low rates.
|