At 17:23 11/10/2005, you wrote:
>Dear SPMers and particularly Klaas :-),
>
>Applying DCM in the context of an - admittedly - odd design, during
>estimation I get the following error:
>
>
>repeatedly:
>
>e.g.: E-Step(+X) F: -x.xxe+03 dp: x.xxe-02
>
>followed by:
>
>Warning: Matrix is close to singular or badly scaled.
>Results may be inaccurate. RCOND = -1.0000000000e+00
>
>and the latter by itself until the next E-Step.
>
>As this is only a warning, the estimation runs on. Can I rely on the
>results (as much as one can anyway...)?
>What might this warning tell me about what might be sub-optimal for DCM?
>
>(It's a very basic model with two regions, symettrically connected
>(two ways), testing whether the exp. condition modulates the
>connection from one region to the other (one way)).
>
>Any help very much appreciated,
>
>with best wishes,
>
>Helmut
Dear Helmut,
this points to problems with matrix inversion during the M-step - we
have seen that a few times in the past. Karl has very recently
updated the parameter estimation scheme (spm_nlsi_GN) to overcome
this problem. If you use the most recent version of this routine, it
should no longer occur. I am not sure whether this version is in the
updates directory yet; I will check this with John.
If the same problem still occurs with the new code, please send me
your model and we will have a look at it.
All the best,
Klaas
|