Klaas, 

Thank you very much for looking into the problem. This is very helpful.
I'll certainly try that.

I have a few more related questions, though. If I understand correctly,
you are suggesting that we use one regressor (including all trials) for
specifying both the "input" and intrinsic connections between the VOI
regions, and use the parametric-modulation regressor (including all 4
force levels) as the modulator. Is that correct? 

Does that mean that, with the current design of the trial conditions we
have, it is not possible to model separately an "input" regressor and
another regressor (e.g., "movement"), in additional to the parametric
modulator "force?" If this is the case, how should the trial conditions
have been constructed such that they would allow a matrix for DCM with
two regressors that would not be highly correlated (if parametric
modulators are not used)? Also, why didn't the "attention" data set
provided at the SPM site have this problem when the regressors are
constructed for the DCM analysis? I couldn't see major differences
between our design and the design of the "attention" data set in terms
of combining trial conditions for building regressors. I'm asking these
questions because we are constructing another experiment which may also
involve DCM analysis.

Thank you very much, again, for your help!

Ben

-----Original Message-----
From: Klaas Enno Stephan [mailto:[log in to unmask]] 
Sent: Tuesday, January 09, 2007 2:22 PM
To: [log in to unmask]
Subject: Re: [SPM] FW: help with design matrix for DCM analysis

Dear Ben,

the problems with the new design matrix may arise from the fact that 
your new regressors are highly correlated.  If you are interested in 
testing your question, i.e. whether a specific connection depends on 
the level of force used, the following design may be easier to use:
1. a regressor encoding the presence of all trial (all stimuli)
2. a parametric modulation of this regressor that represents the 
force level of each trial.  This could either be ordinal [1,2,3,4] or 
encode the actual force used [10,20,30,70].  The latter is probably 
more elegant.

In either case, these two regressors will be orthogonal and should 
not cause you any worries.

Best wishes,
Klaas



At 17:23 04/01/2007, you wrote:
>Hi, Klaas,
>
>Sorry to ask you for help again. I still have the problems  with the
>design for DCM analysis (see below). Would you help me out?
>
>Thanks much!
>
>Ben
>
>-----Original Message-----
>From: Xu, Ben (NIH/NINDS) [E]
>Sent: Thursday, December 28, 2006 10:22 AM
>To: 'Klaas Enno Stephan'
>Subject: help with design matrix for DCM analysis
>
>Dear Klaas,
>
>I'm having trouble building a GLM design matrix for a DCM analysis. I
>posted the problem sometime ago but have not received any response. So,
>here I am, troubling you again for help.
>
>We have constructed a conventional design matrix with SPM2 for a set of
>event-related fMRI data that involves motor tasks and got consistent
and
>significant action across task conditions. Now, we are reconstructing a
>design matrix (with as an implicit baseline as in the conventional
>design, i.e., the resting period between trials) to be used for DCM
>analysis, but running into problems. The main problem is that if a
>condition is included in the "input" regressor, it no longer shows any
>significant action (or any activation at all) when modeled again as a
>separate regressor (needed as modulating factor for DCM). Here's what I
>did:
>
>The data set consists of 4 motor conditions. Conditions 1-3 require
>similar movement force. Condition 4 requires much stronger movement
>force than Conditions 1-3. The conventional SPM analysis (with implicit
>"Rest" as a baseline control) showed consistent action in the expected
>brain regions for all four conditions. In addition, Condition 4 showed
>more action in the same and other regions.
>
>Our interest in the DCM analysis is to see how the increase in motor
>force modulates two specific connections among 4 active regions. Here
is
>the GLM design matrix for the DCM analysis:
>
>Input regressor: includes all 4 conditions
>
>Movement regressor: includes Conditions 1, 2, and 4
>
>Force regressor: includes Condition 4.
>
>(Condition 3 was not modeled as a separate regressor, and "Rest" is
>implicit)
>
>  When I looked at the activation in "Results," it did not show any
>significant voxels (either with FWE or FDR correction at .05) for
>"Movement." The activation for "Force" was mostly in a few unexpected
>regions. Only the "Input" regressor showed relatively similar action
>patterns as seen with the conventional SPM analysis.
>
>Why do activation results change so dramatically with the new design
>matrix? Why are the separate regressors no longer showing significant
>voxels? Is implicit baseline a problem? I can't draw VOIs without
>significant voxels. I assume, without significant voxels, DCM analysis
>will not work either, is that correct?
>
>Your help is greatly appreciated.
>
>Ben