Reply-To: | | [log in to unmask][log in to unmask]<mailto:[log in to unmask]>> To: "Patricia Bracer" <[log in to unmask]<mailto:[log in to unmask]>> Cc: "[log in to unmask]<mailto:[log in to unmask]>" <[log in to unmask]<mailto:[log in to unmask]>> Subject: RE: [SPM] DCM in two groups Dear Patricia, Let’s step back for a moment. You said you want to do an exhaustive analysis of the model space (this could be using the post-hoc search, or you could do a truly exhaustive analysis by specifying each model individually). However, if you have a large number of models, the ideal situation is to split these models into a small number of groups, which we call families. All models within a family share something in common, e.g. the presence of a particular connection. For the hypotheses you wish to test, is there any way you could split your model space into families?
Best, Peter
From: Patricia Bracer [mailto:[log in to unmask]] Sent: 16 November 2015 13:46 To: Zeidman, Peter <[log in to unmask]<mailto:[log in to unmask]>> Cc: [log in to unmask]<mailto:[log in to unmask]> Subject: Re: [SPM] DCM in two groups
Dear Peter,
Thanks for this.
So just to be sure. If I have 4 ROI's fully and reciprocally connected if I want to exhaust the model space will be quite hard I guess. So I suppose that for each of the 40 subjects (20 controls and 20 patients) I do invert only the full model, right? I meant to use a post hoc DCM for each subject after I fit the full model in order to know for each individual which is the winning model. But if I understood correctly this is not the way to proceed?
Then assuming that I have a winning model for each subject, I will the pull all the models for each group seperately and use BPA for each of the two?
Finally you suggested specifying my models manually, would you please point to a post that this is explained? It seems I cannot find it
Thank you for your help
Regards, Patricia
Sent: Monday, November 16, 2015 at 2:29 PM From: "Zeidman, Peter" <[log in to unmask]> To: [log in to unmask] Subject: Re: [SPM] DCM in two groups Dear Patricia, As DCM for fMRI is a model of BOLD timeseries data, so you’ll need to fit a model to each subject’s BOLD data individually.
Ideally to compare between groups, you want the same model space for each group, and to marginalise over models in each group using Bayesian Model Averaging. You can then compare the parameters from each average model. The standard post-hoc code doesn’t facilitate this. However, in coming days / weeks we have a technical note coming out in NeuroImage describing new software tools for doing exactly as you ask.
For the time being, I would recommend specifying your models manually for each group (see previous posts on how to script this). If speed is an issue, you can use the same technology as post-hoc DCM to estimate these rapidly (click ‘Dynamical Causal Modelling’ then ‘Search’ in the GUI – do this for each subject individually, or call spm_dcm_search from a script‚,ih |