> That makes sense that my problem might have something to do with the
EM algorithm settling on a relatively poor local maximum for the correct
model. How can one mitigate this risk? If the EM algorithm starts at an
arbitrary point, then it would seem that increasing subject or run
number would help. Are there properties of a specific model structure
(or space) that would make one more or less susceptible to this?
This is a good question. I don't know how important this problem in DCM
actually is and (if it is important) how to avoid it. Maybe, there is
already an article about it that I am not aware of.
If you have more detailed pre-assumptions about the real neural and
hemodynamic paramaters of your model it could help to start the EM with
those values. If this is not possible, it can help to run the EM a few
times with different starting positions.
In general, I think, it is necessary to test if model selection works
properly using synthetic data generated with various (!) neural and
hemodynamic parameters. I would like to know if this has been done before.
Best,
Martin
--
SPM for programmers
http://spm.martinpyka.de
|