Dear Andre,

When looking at the families of models that you compare, it is clear that family 1 is best in all 3 groups, and also that model is the best model for all your groups. I would therefore recommend that you compare the connectivity parameters in model 1. However, as you say, there are quite striking differences between your groups, and you would probably want to report these. Especially in the patient group, your confidence that any of these models is really better than the rest, is very low. One of the reasons could be, that in the patient group, the data is a lot noisier, and that none of your models explains very much. Some of the diagnostics that I recommended may tell you more about this.

For more detailed guidance on how to proceed, I would recommend reading "Identifying abnormal connectivity in patients using dynamic causal modeling of FMRI responses." (Seghier et al. 2010), which is an excellent article discussing various approaches to DCM in patients, and making clear best-practice recommendations.

best,
Hanneke

On Tue, Apr 17, 2012 at 3:02 AM, <[log in to unmask]> wrote:

Dear Hanneke,

Thanks for you reply. I will try Karl Friston' spm_dcm_fMRI_check and also Jean Daunizeau' script to explore my DCMs.

In more detail, I have doubts whether my inference approach is valid. It would be very nice when you can quickly comment it:

After BMS (including Family Inference), the BMS results for healthy controls (HC) and patients at high-risk (ARMS) look quite consistent and I suggest that I could extract the parameter of model 1 to compare model 1 among HC and ARMS. However, the BMS results for the "first-episode" group seem to be a little bit strange (see attachment)=. Based on these inconsistent "FE" models I've decided to do BMA over all 12 models to do group level inference on the coupling parameters. Now I got significant differences in one backward connection (HC: mean= 0.0018145, SD= 0.004749127; ARMS: mean = -0.002866667, SD = 0.004528113; FE: mean = -0.004558333, SD = 0.006528324).

Any suggestion would be very appreciated.

Best regards,
andré

André Schmidt, MSc ETH
Psychiatric University Hospital Zurich
Division Neuropsychopharmacology
and Brain Imaging & HRC Zürich
Lenggstr. 31
CH-8032 Zurich
Switzerland
Mail: [log in to unmask]
Office: +41 44 384 26 16
Mobile: +41 78 759 40 18

-----Hanneke den Ouden <[log in to unmask]> schrieb: -----

An: [log in to unmask]
Von: Hanneke den Ouden <[log in to unmask]>
Datum: 16.04.2012 23:38
Betreff: Re: [SPM] DCM coupling parameters

Dear Andre,

These are very small numbers, but 'plausible' is a difficult question to answer. Do you mean that you want to know whether there is any evidence for these connections to 'exist'? You can have a look at the distribution of the parameters (i.e look at the variance as well as the mean) to check whether these also very small. You can compute the probability of a parameter being larger (or smaller) than 0, and see whether you have any confidence that these parameters are different from 0.

To more generally diagnose your model, have a look at the email that Karl Friston sent on March 11, which included a short matlab routine to provide some simple diagnostics of DCMs to ensure that everything is okay.

Hanneke

On Fri, Apr 13, 2012 at 10:58 AM, Schmidt André <[log in to unmask]> wrote:

Dear DCm experts,

I've computed DCMs during a n-back working memory task in healthy and several groups of schizophrenic patients using fMRI. Is it plausible to get positive and negative coupling parameters of 1.0-4.0E-03?

Thanks for your suggestions.

best wishes,
andré

--
Hanneke den Ouden, PhD

Center for Neural Science
New York University
4 Washington Place, room 873
New York NY 10003

--
Hanneke den Ouden, PhD

Center for Neural Science
New York University
4 Washington Place, room 873
New York NY 10003