Dear list
Some others may find Klaas' off-list advice to me useful with respect to
some practical issues about design matrices for DCM - the conversation goes
from top to bottom chronologically! Thanks Klaas.
Happy new year!
Alexa
| -----Original Message-----
| From: Klaas Enno Stephan [mailto:[log in to unmask]]
| Sent: 12 January 2005 20:44
| To: Alexa Morcom
| Subject: Re:
|
|
| At 16:51 12/01/2005, you wrote:
| >Hi Klaas
| >
| >Would you mind if I pick your brains about a DCM design practicality?
| >
| >It's about an event-related study with what I think has been
| >called a 'psuedo-factorial' design, ie word presentation vs fixation
(aiming
| >to use this as perturbing factor) and 2 different classifications of
those
| >words on the basis of a binary behavioural measure (aiming to use as
| >'contextual' factor). In a way it's a continuation of a discussion
between Karl
| >& Darren - see
http://www.jiscmail.ac.uk/cgi-bin/wa.exe?A2=ind0407&L=spm&P=R9545&I=-1
| - and I realise the inference is a bit more limited than with a 'true'
contextual factor
| >
| >The question is how to recode the events in a new design matrix. In the
| >original analysis it's word type A & word type B in separate columns. But
to
| >make A versus B a DCM 'factor' however, this needs to be coded in a
single
| >regressor, right? And parametric modulations can't be used - right? So
how
| >does one create 'A minus B convolved'?
| >
| >best wishes
|
| >alexa
|
|
| Hi Alexa,
|
| Two things concerning you DCM question:
|
| 1. Will's recent comment about the use of parametric modulation
| and DCM on the helpline was not quite right, I think: there should be no
problem
| using a regressor that has been defined using parametric modulation as
| input to a DCM. What does *not* work is to use user-specified regressors
| as inputs to DCM.
|
| 2. If you are interested in showing that event A has a significantly
| different modulatory effect on a particular connection than event B, I
| would suggest to
| (i) represent the two events separately in the design matrix (as you have
| done for your SPM analysis),
| (ii) allow *both* of them to have independent modulatory effects on the
| same connection
| (iii) and, after fitting the model, then the parameter estimates for
| differences using contrasts. This is done at the single-subject level via
| the contrast function in the DCM menu. At the group level you can use a
| paired t-test across subjects (or equivalently a one-sample t-test
| operating on the within-subject differences of the modulatory parameters
| for that connection).
|
| The 'A minus B convolved' option that you refer to above is what we would
| use for SEM, but not for DCM.
|
| Hope this helps - let me know if this is not quite clear yet.
|
| Take care,
| K
|
At 12:43 14/01/2005, you wrote:
>Hello again
>
>you're right about the parametric modulations, so I'll be able to sort
>something out. But why is A-B is an SEM thing *not* a DCM thing, though,
>as a 1 (for events A), -1 (for events B) parametric modulation of the 'A
>and B' regressor *seems* like a good way of having 'A versus B' in as a
>factor...?
>
>
>alexa
Hi Alexa,
I think it is perfectly fine to use parametric modulation to define a
modulatory input in DCM that is 1 for condition A and -1 for condition B
and then test for the significance of the associated parameter to establish
differences in modulation between A and B. The reason why this is not the
standard approach in DCM is that (i) it is less effortful to define A and B
as separate modulatory inputs to the same connection and then use a
contrast to test for differences, and (ii) the latter approach is also more
flexible because you may want to perform additional tests that compare the
effect of A but not B (or vice versa) on one connection with the effect of
A but not B (or vice versa) on another connection. In other words, having
A and B as separate modulators gives you maximally flexibility to perform
any kind of comparison by means of contrasts, particularly if these
comparisons are across connections.
|