Dear Marco,
This is a good question. One possibility would be to create all the models you want for a subject (as actual DCM files), and then use spm_dcm_search (or the 'search' item from the DCM menu) to estimate them all at once using the post-hoc scheme. Then repeat this for each subject. You could now do BMS / BMA on the models in the normal way.
I'll speak to my colleagues at the FIL to see if anyone has any other ideas.
Best,
Peter.
-----Original Message-----
From: SPM (Statistical Parametric Mapping) [mailto:[log in to unmask]] On Behalf Of marco tettamanti
Sent: 27 August 2014 17:09
To: [log in to unmask]
Subject: [SPM] DCM network discovery and comparison between groups
Dear all,
I have a question regarding the correct manner to perform a between-group comparison within the framework of DCM network discovery.
In the framework of BMS, the suggested approach in the presence of more than one experimental group (or family), when the winning models for the different groups differ with respect to parameters/connections, is to use Bayesian Model Averaging (Penny et al. 2010). This provides weighted summary coupling parameters e.g. over the entire model space of each group, and it then allows to perform between-group comparisons, thus avoiding conservative assumptions about any particular model.
In one experiment, I have now used DCM network discovery, instead of BMS, to identify a group-specific optimum model for two different experimental groups.
The resulting optimum models have different parameter/connection configurations between the two groups, and it is therefore difficult to perform between-group comparisons on the coupling parameters.
The problem is that in this case BMA does not seem to be a viable solution since, strictly speaking, I only have one model per group (i.e. the winning model), instead of a model space constituted by several models whose parameters can be averaged.
Are there any recommended approaches to overcome this problem in this case?
I could e.g. manually specify and estimate a set (or even the full set?) of alternative models to the optimum model, then perform BMS to actually verify that the optimum model is indeed the winning model, and, contextually, also calculate BMA. But this does not seem a very efficient solution. Also, I would need to specify families to calculate BMA, and I do not think that there is any particular rationale to do so here.
Another solution that I have explored is, given that for each subject I have both the fully connected model entered in the DCM network discovery and the output optimum model, I can in principle perform BMS entering the two models for each subject and arbitrarily assigning the fully connected model to family 1 and the optimum model to family 2, and in such a way calculate BMA.
But I am not sure that this is a sensible solution.
In addition, in one of my two experimental groups, the optimum model has the same number of connections/parameters as the fully connected model (though e.g.
DCM.F differs between the two models). Therefore, BMS/BMA seems to make even less sense to me.
Any help would be warmly appreciated!
Thank you and best wishes,
Marco
--
Marco Tettamanti, Ph.D.
Nuclear Medicine Department & Division of Neuroscience San Raffaele Scientific Institute Via Olgettina 58
I-20132 Milano, Italy
Phone ++39-02-26434888
Fax ++39-02-26434892
Email: [log in to unmask]
Skype: mtettamanti
http://scholar.google.it/citations?user=x4qQl4AAAAAJ
--------------------------------------------------------------------------
IL MIO 5XMILLE VA AL SAN RAFFAELE DI MILANO PERCHE' QUI LA RICERCA DIVENTA CURA.
CF 07636600962
SE NON QUI, DOVE?
Info: [log in to unmask] - http://www.5xmille.org/
Disclaimer added by CodeTwo Exchange Rules 2007
http://www.codetwo.com
|