Dear Anne,
DCM can certainly be a useful tool to study
memory. Generally, there is nothing wrong with
omitting bilinear terms, but usually these are
the most interesting aspects of studying effective connectivity.
Your warning message about a singular matrix:
I suspect you are using SPM2 code. Make sure you
are using the most recent SPM5 updates, then this
warning should not disappear. If it still does, let us know.
Posterior probabilities & NaN:
The posterior probabilities are computed by
reference to a hard-coded threshold which used to
be ln 2 / 4 = 0.17 in SPM2 and has been changed to zero in SPM5.
The NaN's are nothing to worry about – you can
happily ignore them. They arise because
parameters that are not estimated are fixed to be
zero in the parameter vector DCM.Ep and its
covariance matrix DCM.Cp (which is equivalent to
infinite precision). Therefore, applying the
Cumulative Distribution Function for the Normal
distribution to DCM.Ep and DCM.Cp gives you NaN's
for entries. See spm_dcm_ui around line 344
pp = 1 - spm_Ncdf(DCM.T,abs(DCM.Ep),diag(DCM.Cp));
and around line 79 in spm_Ncdf
if any(~md(:)), F(~md) = NaN;
warning('Returning NaN for out of range arguments'), end
(note that this warning has been switched off in
SPM5 because it kept confusing people).
Best wishes,
Klaas
At 15:22 28/04/2006, Anne Botzung wrote:
>Dear SMPers,
>
>
>I would like to make a DCM analysis. However, I encounter some difficulties.
>My experiment comprised two conditions, a memory task (the subjects
>had to evoke past memories) and a control condition.
>I would like to test, using DCM, the influence of left and right
>frontal regions (FTl + FTr) on two areas of the medial temporal lobes
>(PHl + PHr) during memory evocation.
>In the literature, most of studies have used DCM analysis in the
>context of factorial design, e.g., visual processing modulated by
>attentional processes.
>At first, I would like to know whether the use of DCM is appropriate to:
>1. Basic designs without factorial modulation (in this case, B matrix = 0)?
>2 the study of memory processes?
>
>Secondly, I would like to estimate a model (DCM1) that comprised
>intrinsic connections from FTl to PHl and PHr ; and from FTr to PHl
>and PHr. The memory functions were entered into the DCM through FTl
>and FTr.
>I observed a warning message during the estimation (?the matrix is
>singular?), indeed it seems difficult to distinguish the influence of
>the connection FTl --> PHr from the influence of the connection FTr
>--> PHr. Do I have to separate DCM1 into two parts (DCM2: connections
>from FTl to PHl and PHr / DCM3 model FTr to PHl and PHr)?
>
>Finally, independently of the estimated model (DCM1, DCM2 and DCM3), I
>encountered two problems. In the sequel, values are given for DCM2.
>1. For C, we obtained: DCM.C = [0.0554 0 0] and DCM.pC = [1.0e-03 *
>0.2869 NaN NaN]
>Could you explain why the posterior probability (1.0e-03 * 0.2869) is
>so low? Why do I have NaN for the probabilities of the non estimated
>parameters?
>1. For DCM.A, I obtained the following values:
>[-1 0 0
>0.4867 -1 0
>0.6482 0 -1]
>
>What is the meaning of -1?
>For DCM.pA, I obtained:
>[NaN NaN NaN
>0.9986 NaN NaN
>1 NaN NaN]
>Why do I have NaN for the probabilities of the non estimated parameters?
>
>Thank you very much in advance,
>
>Regards,
>Anne
|