Dear Tali,
a quick reply to your questions below:
1. Yes, although to a lesser degree.
2. The two inputs are not orthogonal, but they
are not perfectly correlated
either. Correlations between regressors/inputs
do not prevent you from estimating the
parameters. When one regressor/input is a
linear combination of some others, then some
contrasts are not estimable but others are - see
previous entries on the helpline for this general issue.
3. I would only adjust your extracted data for
your F "effects of interests" contrast, i.e. simply mean-correct your data.
Best wishes,
Klaas
At 20:27 14/11/2005, Tali Bitan wrote:
>Dear Klaas
>
>Thanks for your fast response.
>please see my followup questions below.
>
>On Mon, 14 Nov 2005 19:35:23 +0000, Klaas Enno Stephan
><[log in to unmask]> wrote:
>
> >Dear Tali,
> >
> >it is usually not a good idea to use the same
> >input as a driving input to one region and as a
> >modulator of an afferent connection originating
> >in that region. The reason is that the
> >conditional estimates of both parameters will be
> >highly correlated and, given that the prior
> >variance for driving inputs is higher than for
> >modulatory ones, this will tend to over-estimate
> >the driving parameters and under-estimate the
> >modulatory parameters. In other words, your
> >estimates are based. (In complex networks this
> >is not always true, but it is a good
> >rule-of-thumb. Individual cases are best checked using simulations.)
> >
> >In your case, it seems best to define an input
> >that includes words and symbols (assuming that
> >they equally activate the input region, let's
> >call it A1) but not fixation (assuming that this
> >does not activate A1) and define a second,
> >modulatory input that only includes words. In
> >this way, you are testing whether a word-related
> >activation in the target region (A2) can be
> >explained by a selective increase of the
> >connection from A1 to A2 during the presentation of words.
>
>1- Would this affect also the modulatory effects on connections that DO
>NOT involve the input region (i.e. in your example a connection from A2 to
>A3)?
>
>2- If we define an input of both 'words+symbols' (mixed) and another input
>of just 'words' in the conventional analysis, they are going to be non-
>orthogonal, and no variance will be left for the 'words' contrast. Is this
>appropriate to do in the conventional model?
>
>
> >
> >All things above are independent of whether you
> >use a blocked or event-related design.
> >
> >Best wishes,
> >Klaas
> >
> >PS. When you say "adjust" to you mean the
> >adjustment in the extraction of data (VOI tool)?
>
>
>3- yes, by "adjusting the ROI" I refer to the extraction of the data in
>the VOI tool. If I only adjust for an F contrast of 'words'
>and 'fixation', does it make any difference? Would I still need to
>use 'symbols' as a driving input?
>
>
>Thanks a lot.
>Tali
>
>
> >
> >
> >At 17:50 14/11/2005, Tali Bitan wrote:
> >>Dear DCM Experts:
> >>
> >>We have a question about how to best disabmiguate
> >>between driving inputs and bilinear effects using
> >>a non-factorial, event-related study.
> >>
> >>We have 3 conditions in the time series: 'words',
> >>'symbols' and 'fixation'. We have adjusted the
> >>ROI to include all 3 effects. We are only
> >>interested in the effect of words, so in the DCM
> >>model we only have 'words' as the driving input
> >>to the input region, and also as exerting a
> >>bilinear effect on all connections, except for
> >>the connections that go back to the input region.
> >>Is this the correct way to go?
> >>
> >>Or - do we need to include all 3 effects (words,
> >>symbols and fixation) as (separate vs. a single
> >>combined) driving input(s) on the input region,
> >>while only the 'words' would have a bilinear effect on all connections?
> >>
> >>If so - would the situation be different if we
> >>only adjust the ROI for 'words' (rather than all
> >>three effects)? Do we still need to include
> >>'symbols' and 'fixation' as driving inputs?
> >>
> >>Is all this true only for an event-related design, or also for a block
> >>design?
> >>
> >>thanks a lot.
> >>Tali Bitan
> >>
> >>-------------------------------------------------------------------------
>--
> >>----------------------
> >>Dear Tali here is my first crack at the question:
> >>
> >>I've looked through the DCM code. As we thought,
> >>what it does is it takes the a, b, and c
> >>parameters and the hemodynamic priors and then
> >>sets up the expected neuronal and then
> >>hemodynamic responses. The c parameters determine
> >>the activity in the system, so no c parameters =
> >>no system activity. I have tried this and
> >>verified it. In a sense it almost seems as if DCM
> >>creates signals in a networks based on the
> >>driving inputs, and then sees if that signal can
> >>be alternatively modeled by some combination of
> >>an interaction and the signals from other
> >>regions ((signal x input) + multi-regional
> >>signal). I think this is why DCM seems best
> >>conceived when the experiment allows different
> >>driving and modulatory effects (i.e., based on an ANOVA model)
> >>
> >>In usual GLM statistics, if your design matrix
> >>leaves out (regressors, covariates, conditions)
> >>that specify important aspects of variance, that
> >>variance ends up in the residuals and affects the
> >>p-values. If this variance substantially
> >>contributes to the residuals and has some defined
> >>structure it may violate some assumptions of the GLM as well.
> >>
> >>I don't know whether the same is true in DCM. If
> >>you include words as driving inputs and as
> >>modulatory effects, but the data is based
> >>on words+symbols+fixation then 1) the design
> >>will only generate activity related to words. 2)
> >>This will potentially work against you when you
> >>then try to use the same input as a modulatory
> >>effect. 3) When DCM tries to compare the expected
> >>signal vs. the actual signal there will be a significant discrepancy.
> >>
> >>What does this all mean about how best to use DCM
> >>given the difficulty of some experimental
> >>designs? I'm not sure but I think you should
> >>probably use words+symbols+fixation as an input
> >>and just words as your modulatory effect. If you
> >>just use words as input then the design would
> >>probably be best for just examining differences
> >>in the intrinsic models for the network. See
> >>Ethofer et al., Cerebral pathways in processing
> >>of affective prosody: A dynamic causal modeling study. Neuroimage, in
> >>press.
> >>
> >>I would be very interested in the thoughts and
> >>commentary from real DCM experts on your questions and my reply.
> >>
> >>Darren
> >>
> >> >----------------------------------------------
> >> >Tali Bitan, PhD
> >> >Department of Communication Sciences & Disorders
> >> >Northwestern University, IL
> >> >Phone (847)467-1549
> >>
> >>-------------------------------------------------------------------------
> >>Darren R. Gitelman, M.D.
> >>Cognitive Neurology and Alzheimer¹s Disease Center
> >>Northwestern Univ., 320 E. Superior St., Searle 11-470, Chicago, IL 60611
> >>Voice: (312) 908-9023 Fax: (312) 908-8789
> >>-------------------------------------------------------------------------
|