Hi DCM experts,
I would like to get a better idea of what sanity checks I can do with my first-level DCMs in addition to the notes in spm_dcm_fmri_check() and how to properly interpret the diagnostic plots provided generated by spm_dcm_fmri_check().
I am running a DCM analysis in SPM12/DCM12.5 with 4 fully-interconnected regions (linear, two-state, non-stochastic). I have set up 2 families of models (differing by the driving input region), and the models in each family differ by which connections are being affected by some condition. Just as a side note in case it helps with understanding what I am seeing in the diagnostics, I have a rather complex paradigm with interactions that are of interest and which I have tried to deal with by having a GLM with one regressor of interest for all the trials/the "task" and as parametric modulators on this regressor are my conditions (time and 3 experimental conditions, each with 2-3 levels and not all fully crossed as dependent on participant behavior). I have specified all 4 factors to modulate connections, but the models differ by how a higher interaction of these conditions modulate connections. I know Peter Z. has discouraged this in other posts, though I'm just not sure how to get around it...that will be a subject for another post.
Eventually I would be comparing the connection weights between my patient (N=+/-20) and healthy control groups (N=+/-20) to which I fitted the same models. I am hesitating going to group analysis with PEB, because I am not totally convinced by the reliability of the estimation of the first-level models/that my 1st-level model specification is technically/computationally appropriate/tractable.
Do I understand correctly that in the plots from spm_dcm_fmri_check():
* top, "Responses and Predictions" = the model fitted to the signal -- desirable is when the predicted line is not flat and variance explained > 10% (otherwise text will be in red)
* bottom left, "Intrinsic and Extrinsic connections": posterior expectations where desirable is when bars > 1/8 Hz for one or more extrinsic connections (i.e. away from 0 = prior expectations). By "extrinsic connections", I am assuming it's the modulated connections, i.e. the B matrix and that spm_fieldindices(gcm.Ep,'B') would give you which these are? But how do I figure out what each index falling into the B matrix corresponds to? However, the number of bars displayed in this plot seems to match the number of anatomic/A-matrix connections?
* bottom right, "Posterior Correlations" -- not sure what to look out for here and what # of estimable parameters
I had no convergence issues in any of the models with any participant. However, the percent of variation explained for most of the models for almost everyone is quite low -- less than 1% for most people, though for a very small handful of participants, the % of variance explained by several models is consistently higher (5-20%), which I see tends to also correspond with their higher largest absolute parameter estimates and number of effective number of estimable parameters. I am also disturbed by the fact that the number of estimable parameters for most people across all the models hovers around 1....
Any comments/advice will be very much appreciated!!
Many thanks in advance,
Gina
|