Hi Keith,

Increasing the number of connections does not require more subjects, but there are two other considerations to bear in mind.

More complex models face a higher penalty in the Free energy estimate of log-model-evidence (F), as this depends on the difference between model accuracy and model complexity. Unlike AIC and BIC (SPM2.SPM5), the F penalty for complexity (SPM8, DCM10) is not fixed by the number of parameters/observations. Rather, it also accomodates the dependencies among parameters (via the KL divergence). Bayesian model selection will then be able help you find the right balance of model complexity and accuracy, for a given data set, for one or more subjects.

The dependencies among parameters need careful consideration, especially if you wish to go on to perform classical statistics on the connectivity parameters, as Marta suggests (and as have been successfully undertaken in many published DCM papers). Relevant
parameter dependecies are more likely for complex models. If the parameters do covary, then the estimates of indidual parameters become unreliable, increasing type II error in the classical stats. Note that this does not affect F-based model comparisons (e.g.
RFX & FFX Bayesian model selection). Klaas offered a clear account of these issues with useful refs in his recent mail #**046065** "DCM-conditional
dependencies". If you frame your hypothesis in terms of model selection (with one or many subjects) then this problem does not arise.

Best wishes,

James

.

Hi Keith,

If you are interested in model comparison, then I'd say you can use whichever number of subjects, irrespectively of the number of connections. In fact, you can make statistical inferences about models in only one individual subject.

However, if you then want to go on and do classical statistical tests on the connectivity parameters, then it might be a good idea to increase the number subjects/parameters per cell.

I'm CCing this to the SPM list in case other people want to comment on it.

Hope this helps!

Marta

--

Dr. Marta I. Garrido

Research Associate

Wellcome Trust Centre for Neuroimaging

University College London

12 Queen Square

London

WC1N 3BG

United Kingdom

Tel: +44(0)20 7833 7472 (ext: 4366)

Fax: +44(0)20 7813 1420

Email: [log in to unmask]

www: http://www.fil.ion.ucl.ac.uk

If you are interested in model comparison, then I'd say you can use whichever number of subjects, irrespectively of the number of connections. In fact, you can make statistical inferences about models in only one individual subject.

However, if you then want to go on and do classical statistical tests on the connectivity parameters, then it might be a good idea to increase the number subjects/parameters per cell.

I'm CCing this to the SPM list in case other people want to comment on it.

Hope this helps!

Marta

On 9 May 2011 10:44, Keith Kawabata Duncan
<[log in to unmask]> wrote:

Hi Marta,

Maria thought you might be able to help me with a DCM question. I was wondering if increasing the number of connections in a model requires increasing the number of subjects - a bit like increasing the number of cells in a factorial design?

Keith

--

Dr. Marta I. Garrido

Research Associate

Wellcome Trust Centre for Neuroimaging

University College London

12 Queen Square

London

WC1N 3BG

United Kingdom

Tel: +44(0)20 7833 7472 (ext: 4366)

Fax: +44(0)20 7813 1420

Email: [log in to unmask]

www: http://www.fil.ion.ucl.ac.uk