Dear John,
The implementation of smoothest requires the input volume
to contain values that are distributed as a Gaussian with
zero mean and unit standard deviation (i.e. z stats).
If residuals are input then they are normalised internally
so that they are mean zero and stdev one - but this code is
only run if the input is 4D, not 3D. So you can only run the
smoothest code on a 3D volume if it contains normalised
Gaussian-distributed values. This means that you can use
the parameter estimates, as they are Gaussian-distributed,
but you will need to normalise them before passing them in.
However, t and f stats should not be used as they are not
Gaussian-distributed.
Note that zstats (or param estimates) are not recommended
though, as they contain signal as well as noise. What you
really want is just as estimate of the smoothness in the
noise. Within Feat, smoothest is only ever run on the
residuals, and this is what we would strongly recommend
unless you have a very special reason not to use them.
All the best,
Mark
On 5 Oct 2005, at 14:33, John Herrington wrote:
> The command-line help for the smoothest program indicates that the
> inputs to the program should be either 'residual fit' or zstat images.
> Is this truly the case, or will the program's output be legitemate if
> one were to use some other time of statmap input, like parameter
> estimate, f or t-stat maps?
>
> Thanks,
>
> John Herrington
|