Dear Gianpaolo,
There are various differences in the estimation which could explain why you get differences between the two approaches - full inversion versus "optimize" or "post-hoc", now known as Bayesian Model Reduction (BMR). 

One key difference is that when you separately estimated your full and reduced model, it is possible that each model estimation finished in a separate local maxima. Whereas, by using BMR, you only inverted the full model - the evidence for all other models were derived from this, making local maxima problems for the individual sub-models less likely. Larger models such as your full model, which have higher dimensional search spaces, may also be less susceptible to local maxima problems. So, the BMR result should be more reliable than separate estimation of each model. 

If you would like to output all the models from the final iteration of the search (as displayed in the SPM figures), there's an option to do that. See spm_dcm_post_hoc.m . You'll need to call it manually rather than the GUI. Alternatively, if you'd like to specify your models manually rather than using the search, create your models then using the GUI select Search instead of Optimize. This uses the BMR machinery to assign an F value for each model.

Hope that helps,
Peter

-----Original Message-----
From: SPM (Statistical Parametric Mapping) [mailto:[log in to unmask]] On Behalf Of Basso, Gianpaolo
Sent: 19 March 2015 16:30
To: [log in to unmask]
Subject: [SPM] Network discovery using DCM Optimize

I am starting to use the "optimize" option in DCM version 10 to find the best A-matrix to describe the set of areas I found activated during execution or the imagined execution of movements.
After estimating a fully connected A-matrix of 6 areas (leaving empty matrices B and C), I run the optimize option, which starts with the gready algorithm to reduce the parameters and finally computes the best model which fit the data.

To verify that the optimized model is the best fit of the data, I usually specify and estimate a new model selecting only those connections in the A-matrix which were associated with high strength (Hz) and high probability in the optimized results. In many cases, however, if I compute a BMS using FFX and compare the fully connected model against the new model I created based on the optimization results, I find that the log evidence and the posterior probability of the fully connected one are significantly greater than the optimized one. I would expect the opposite. Sometime I am able to discover the optimal network selecting some additional (or sometime fewer) connections respect to those revealed by the otimization routine, which suggests that I am interpreting partially in the wrong way the optimization results. 
What am I doing wrong? Which is the correct way to specify a model using the output from the optimization routine? 

Thanks in advance for any help with this issue.

Gianpaolo Basso 


_____________________________________________
Gianpaolo Basso

Prof. Associato di Neuroradiologia
Università degli Studi di Trento
Dipartimento di Psicologia e Scienze Cognitive Corso Bettini 31, Rovereto (TN)

Fondazione Salvatore Maugeri
Istituto Scientifico di Pavia
Servizio di Radiologia
Via Maugeri 10 - Pavia

Telefono ufficio Rovereto: +39 0464 808714
email: [log in to unmask]
http://www4.unitn.it/People/en/Web/Persona/PER0000685l