Dear Kestas,
At 17:10 16/05/2007, you wrote:
>Dear Will and Klaas,
>sorry to bother you personally, but I have a technical question
>about DCM that's really bugging me. So here it goes: I specified six
>models with 3 ROIs, estimated and averaged them. The results seemed
>reasonable, with one of the models dominant over the others, similar
>in every subject (N=10), and in the comparison of averaged models.
As an important aside: one should NEVER compare averaged
models. This is not valid and has been disabled in SPM5 some time ago.
>To make sure the results are not spurious, I ran several checks. I
>first reran the models with the same ROIs but their position in the
>models rearranged (i.e., ROI1 becomes ROI3, ROI3 becomes ROI2 etc.).
>I also reran them with data from the original ROIs randomly shuffled
>(i.e., data in Y and xY.y and xY.u are rearranged at random with the
>randperm function), and also reran the model with the same ROI
>fields filled with white noise.
>Now, the problem is that in all these control steps the individual
>results seem to come out the same as with the original ROIs!
>I thought that perhaps the way I batch the model estimation is the
>problem. I take the (manually) specified models as the template,
>insert new ROIs using spm_dcm_voi and then re-estimate the model by
>submitting it to spm_dcm_estimate.
This is odd - particularly the white noise test - and the only
explanation I can think of is that the original time series in the
manually specified model are not replaced. Did you plot the time
series in DCM.Y.y? Are they the same across all models?
>So, my question is, am I misunderstanding how to batch the model
>estimation (most likely) and there's some data carryover from the
>originally specified models which causes them to come out the
>same? (I checked the individual subject DCMs, and the code is
>inserting the proper ROIs for every subject and every model to be estimated.)
>Or is there some inherent bias in the models that causes them to
>come out the same no matter what the data are?
If your model is totally symmetric, it would give the same results
regardless of the order of VOIs. But the white noise result is not
explained by this. I would recommend that you take a single subject
and compare carefully whether manual specification and VOI
replacement by code give you the same results. If not, set a
breakpoint in spm_dcm_voi and check step-by-step what goes wrong in your case.
Best wishes,
Klaas
|