Hi,
Our default settings (z=2.3 and p=0.05) is perfectly acceptable and
generally
fairly useful - that's why it is the default!
There is no simple relation between uncorrected and corrected p-values.
The relation is given by Gaussian Random Field Theory but it depends
mainly
on the size of the cluster formed after the initial z threshold is
applied. The
actual maths of how it converts this z threshold value and cluster
size into
a corrected p-value is at the heart of the theory and you should read
the relevant
papers if you want to understand this. Note that we have a technical
report on
our website which summarises some of the early papers on this. There
is also
an excellent webpage discussing this and many issues at
http://imaging.mrc-cbu.cam.ac.uk/imaging/PrinciplesRandomFields
As for lowering the z threshold, it is an arbitrary value and as long
as the threshold
is high enough, as otherwise the approximations in the theory break
down.
So we do not recommend lowering it much below 2.3. Your value of 1.6 is
really too low for the inference to be accurate. In this case we
would recommend
you try other methods of inference such as the ones that randomise
offers.
This includes the new Threshold-Free Cluster Enhancement (TFCE) method
which neatly side-steps the problem of having to specify any threshold.
See http://www.fmrib.ox.ac.uk/fsl/randomise/index.html for more
details on
randomise.
All the best,
Mark
On 7 Jan 2009, at 22:26, shih-Wei Wu wrote:
> Dear all,
>
> This is a rather arbitrary question, but I am curious if there is a
> general, accepted criteria
> for setting up cluster z threshold and p threshold for whole-brain
> analysis at the group-level
> (mixed effects FLAME 1). Would the default setting (z=2.3, p=.05) be
> good enough?
>
> A second question is how I should interpret activation when I lower
> cluster threshold to,
> say, z=1.6, but keep p=0.05. In this case, I do get clear
> activations at areas that conform
> to my hypothesis, and it doesn't look like there are cluster
> activations all over the brain. In
> my particular case, there is no cluster in the brain that satisfies
> the default criterion. And
> that is why I lowered the z thresh.
>
> The last question is what is the relation between 'cluster
> threshold' and 'uncorrected'? For
> example, if a paper reports activation p<0.001 (uncorrected), what
> would that normally
> correspond to when using 'cluster threshold'?
>
> thanks
> shihwei
>
|