Hi Will,
Good questions.
There are many approaches to performing a DCM analysis, but often it is seen a two-step process:
1. Test whether there is evidence that a connection is present (i.e. whether there is significant evidence that is deviates from its prior)
2. If there is such evidence, look at whether the influence of the connection is positive or negative, large or small etc.
The first part can be achieved most straightforwardly with model comparison - compare the evidence for models (or families of models) with and without the given connection. Then, for the second part, take the BMA over all the models or the winning family of models to inspect the parameter estimates. It's fine to use posterior probabilities or confidence intervals here, this just expresses the precision of the posterior estimate at the group level.
Let me know if anything is still unclear.
Best,
Peter
-----Original Message-----
From: SPM (Statistical Parametric Mapping) [mailto:[log in to unmask]] On Behalf Of Will
Sent: 06 February 2015 13:50
To: [log in to unmask]
Subject: [SPM] reporting DCM results
Hi all - quick question about reporting DCM results. After going through model selection and model averaging, I'm finding that the parameters for the winning model, when the standard deviation (sEp) values are taken into account, yield some of these parameters as not significantly different from zero. In this case, is it good practice to report means and standard deviations of these parameters? Also - how does this affect the interpretation of the model selection? For example, say the model predicts that input X has a modulatory effect from A to B. (Yay! It makes sense!) But then, when I look back at the B matrix parameter values (mEp) and (sEp) for X that are calculated by BMA, and it comes out as something like (95% CI: -0.01 to + 0.04) - then what does this mean? I.e. does modulatory effect X sometimes increase, but sometimes decrease coupling between A and B? Or are the results too vague to really say anything at all?
I've noticed that some papers just report the average parameter values, whereas others report the parameters as confidence intervals. Is there a consensus among experts as to what is the best way to go? Thank you for your kind guidance.
-Will
|