Dear Mike,

I think there may be two separate issues here.

At 19:01 02/08/2006, you wrote:
>Dear Dr. Stephan,
>
>I apologize for the unsolicited email, but I
>have a straightforward question that you might
>address. I noticed on the SPM listserv a dozen
>or so posts over the past couple of years
>regarding a NaN error in DCM estimation:
>
> In /usr/local/spm2/spm_Ncdf.m at line 76
> In /usr/local/spm2/spm_dcm_estimate.m at line 72
> In /usr/local/spm2/spm_dcm_ui.m at line 312

This is not a proper error - it is just a warning
which, in SPM2, was *always* given after
estimation because in line 72 of spm_dcm_estimate
(SPM2), spm_Ncdf did not like the NaN values
representing absent connections and inputs. If
you compare the code in SPM5, you will see that
this warning has now been switched off.


>In one of the last posts about a year ago, you
>replied to Dr. Özdemir at Harvard that you would
>try to look into it using one of his datasets.

I can't recall what this was about exactly, but I
suspect this was a different problem.


>I'm hoping I can ask you for the outcome of
>that? I have a similar problem... Only, I have
>a model that works successfully for some
>participants, unsuccessfully for others (it
>works for 42 out of 50 participants). For the 8
>that don't work, the DCM converges after one
>interation, drops the error message, and the
>resultant DCM matrix contains the SAME value for
>all pathways that are included in the model.

This does indeed sound like a problem. If the
error message is the one above (i.e. related to
spm_Ncdf), then ignore this - this simply tells
you that the estimation stopped after one
iteration. This,however, is clearly
pathological. I can currently only think of two cases when this could happen:

1. You failed to define any driving inputs. You
can check this by loading the pathological DCMs
and checking DCM.c - there should at least be one
non-zero value in this matrix. Also check that
the regions receiving direct inputs are not
disconnected from the rest of the network (compare DCM.a).
2. Your data have some odd scaling. Make sure
that you mean-correct your data when you extract
the VOIs. This requires adjusting your data for
the "effects of interest" F-contrast. Note that
the interface will only offer you the option to
adjust your data if at least one F-contrast has
already been specified for that session/subject.

If none of these two cases apply, then
re-estimate the pathological models using SPM5
(the estimation scheme is much more robust than
in SPM2). Make sure you have the most recent updates from the web.
If this does not solve the problem either, then
send me one of the pathological DCMs by email and I will have a look.


>If you can provide any insight, I would
>sincerely appreciate it. My results with this
>technique so far are wonderful. If I can only
>overcome this last analytic hurdle, I think I'm all set.

I am glad you have been enjoying DCM ;-)

Good luck with the trouble-shooting - it would be
nice if you could let the list know what the
problem was, so that others can learn from this example.

All the best
Klaas



>Thanks in advance for any help you may be able
>to provide. Please feel free to refer me to a
>colleague if you are swamped, or if this
>question more aptly suits them. I will appreciate any help.
>
>Sincerely,
>Mike
>
>
>Michael C. Stevens, Ph.D.
>
>Director, Child and Adolescent Research
>The Institute of Living / Hartford Hospital
>
>Director, Clinical Neuroscience & Development Laboratory
>Olin Neuropsychiatry Research Center
>
>Assistant Clinical Professor of Psychiatry
>Yale University School of Medicine
>
>Contact Information:
>200 Retreat Avenue
>ONRC, Whitehall Building
>Hartford, CT 06106
>
>Tel: (860) 545-7552
>Fax: (860) 545-7797