I'm new to DCM and BMS, and am attempting to carry out an analysis similar that from to Rowe et al.'s in press NeuroImage paper. I am testing a variety of nested DCMs that all contain three and only three brain structures, but differing intrinsic connections, driving inputs, and modulatory effects. When running random effects BMS on all 32 of my DCMs (for 38 participants), I found three DCMs that fared a lot better than the other 29 in terms of exceedance and expected posterior probabilities (I have not calculated bayes factors yet). Let's call them A, B and C, where A > B > C in terms of both probability types. I then tried running BMS again on only these three DCMs (like Rowe et al., want to look at the reproducibility of DCMs across multiple time points in this sample, though I haven't done this yet). However, when running this smaller model, the rank ordering of probabilities changed... now B > C > A.
My interpretation of this change, stated informally, is that the probabilities associated 29 models must have overlapped in some way with DCMs B and C over and above DCM A. An analogy from multiple regression - it could be that the 29 "predictors" (DCMS) were more correlated with B and C than with A, so that removing them differentially increased the variance associated with B and C.
My question: does this make sense as an explanation, and if not, what's an alternative explanation?
Thanks,
John
|