Dear James
Yes, you can do parameter-level inference with DCM.
To answer more fully... It's a while since I've read the study of Stefan Frässle (CC'd), but I can give you some general concepts to work with. The parameters you have inferred depend upon the model you have used and the data. The models have changed over DCM versions, in terms of the priors on connection strengths and priors on observation noise. The question you're asking implicitly is whether the model in DCM10+ is better or worse than some other model (e.g. SPM8). The measure of a model's quality is the probability of the data given the model, which is approximated by the free energy. To find out which is the better model for your data, you would need to fit a DCM with SPM8 priors and a model with DCM10 priors and compare the free energies. Note that 'reliability' is whether you'd expect to get the same results if you did the experiment again. To maximise your reliability, you want the model which offers the best tradeoff between accuracy and complexity - this is quantified by the free energy.
In practice, however, I suggest you don't spend time comparing SPM versions, and just use the most recent.
All the best
Peter
-----Original Message-----
From: SPM (Statistical Parametric Mapping) [mailto:[log in to unmask]] On Behalf Of James Trujillo
Sent: 06 February 2018 10:20
To: [log in to unmask]
Subject: [SPM] DCM parameter estimate reliability
Dear SPM experts,
I am working on a DCM analysis (using DCM12) involving two regions and am interested in being able to make inferences at the parameter level. Specifically, I found a winning model with BMS that shows that my experimental input modulates bi-directional coupling between my two regions, and I would like to see if the coupling positively or negatively modulated.
Although I get interesting results from applying BMA to this data, I would like to make sure my interpretation is correct. I am a bit concerned about the reliability of parameter level inference, given the 2015 NeuroImage paper from Fraessle etal suggests that parameter estimates under DCM10 and 12 may be less reliable than older versions. In my case, which is within subjects, so no between-group tests, I find that my manipulation results in positive modulation of one connection, and negative modulation in the other correction. Can I then say that the coupling is increased in one case and suppressed in the other? Or are the parameter estimates not reliable enough for this type of inference?
Thanks in advance!
Best regards,
James Trujillo
|