Dear Mihaly,
Perhaps you could use stronger priors on h (the log of the observation noise precision). The prior mean on h is M.hE and the prior variance is M.hC.
[you are using an infinitely strong prior on h by not updating it].
To see how M.hE and M.hC affect say the noise SD, sigma, you can
generate some h's from the prior eg. h=rand(1000,1)*sqrt(M.hC)+M.hE,
then sigma=sqrt(1/exp(h)), then hist(sigma) to have a look.
You should then check that this corresponds to what you believe about observation noise. If you have DCM for ERP then the observation noise SD is just stdev of the original trials (the ERP itself is of course just the mean). If you have DCM for fMRI then you could look at how well the GLMs fit in those regions.
Best, Will.
> -----Original Message-----
> From: SPM (Statistical Parametric Mapping) [mailto:[log in to unmask]]
> On Behalf Of Mihály Bányai
> Sent: 11 April 2012 16:56
> To: [log in to unmask]
> Subject: [SPM] Fisher scoring scheme oscillating
>
> Dear SPMers,
>
> I have a problem with estimating the parameters of a dynamical model. I
> have a DCM with 3 areas and 1 input, and I'm trying to estimate
> connection parameters with SPM. However, the Fisher scoring scheme for
> estimating the hyperparameters controlling the observation noise
> covariance is oscillating between zero and one (the log of the
> hyperparams, denoted by h in spm_nlsi_GN.m, are taking values near 0
> and -1500 in successive iterations respectively). I tried with several
> connection patterns, intrinsic and modulatory, I get the same all the
> time.
>
> When I consider noise covariance known (fix the hyperparameters and do
> not update them by the Fisher scheme), the iteration of the E-step of
> the parameter estimation runs without errors, but I don't know how much
> I could rely on such results, I'm assuming not really.
>
> Could you give me a hint about what direction should I look for the
> source of the issue?
>
> Thank you very much,
> Mihály
|