Hi - I am not sure what line you are talking about that comments out
the noise evaluation??
Nevertheless, if you do stats in a Bayesian framework, you never make
a point estimate of any parameter (including the std of the noise),
instead you build up a distribution of possible values of that
parameter. Essentially you are solving a bunch of integrals. You can
do this with many techniques, but we use MCMC for almost every
parameter. The exception is the noise parameter, becasue if you assume
Gaussianity, it is possible to perform the relevant integrals
analytically (with a pen and a piece of paper and a friendly MJ),
hence it is never necessary to sample from the noise std - but this
parameter is nevertheless correctly accounted for in the distributions
of all other parameters.
Does this help?
T
On 22 May 2009, at 09:56, SUBSCRIBE FSL Anonymous wrote:
> Hello,
>
> I 've read this paper: Beherens et al, Neuroimage 2007,34:144-55
> "The noise is modelled separetely for each voxel as Gaussian".
>
> But, noise evaluation in program is written as a comment out.
> So, the noise will not be evaluated.
> Why is it a comment out?
>
> Many thanks for your help.
>
|