Reply-To: | | [log in to unmask][log in to unmask]>
> I had exactly the same problem and, as I understood, there is not univocal > solution. > Changing "onset" and "sd", it is hugely better and now everything works. > At the beginning I had set the onset too early and the sd too narrow. I > would suggest to play a bit around with those two settings, obviously being > sure that they are biologicaly meaningful for your data. > > I computed everything again using SPM12 and I have the feeling that it > works better on this aspect too. Anyway, I did not perform a scientific > comparison between the results obtained from SPM8 and SPM12, therefore I > could be wrong. > > Giovanni > > > 2013/7/26 Marlene Bönstrup <[log in to unmask]> > >> Dear DCM-developers and users, >> >> during the estimation procedure, the EM-step, of my DCM of induced >> responses for EEG-data, the free energy doesn't start to increase. It is >> trapped in a local minimum, so that the model converges after a couple of >> iterations and the conditional covariances and expectations (Ep, Eg, Cp, >> Cg) are not updated. The result is an empty B-matrix, or the connections in >> the B-matrix being numerically very small, without the connection strengths >> being estimated. >> >> Also the conditional responses in source space, i.e. the predicted >> time-frequency evolution, is numerically so small that it doesn't resemble >> the observed response at all. >> >> This phenomenon I observe more or less constantly over a wide range of >> different peristimulus time windows, model orders (2 to 10 tested) and >> other slight variations like hanning window, detrending options, frequency >> window etc. Also over several subjects and differently arranged B-Matrices >> as input parameters. >> >> Strikingly, in a different data set which is almost identical in the >> experimental design except for a different mode of stimulus presentation, >> the DCM-iteration starts normally and goes over several EM-steps until it >> converges with numerically sensible output matrices in DCM.Ep. >> I would be very thankful for any comments, ideas about possible reasons >> for a DCM to get trapped in a local minimum and not starting to iterate >> over several EM-steps (i.e., [dF < 1e-2] after 2 or 3 steps). >> >> Marlene >> > >[log in to unmask] |