Put slightly differently, permutation works well regardless of the
cluster threshold chosen, because it accurately reflects the
properties of the data and gives a valid estimate of the null
hypothesis distribution given that chosen threshold. But permutation
can say nothing about the "correctness" or "appropriateness" of
choosing any particular threshold. And indeed, different chosen
thresholds will yield quite different end results, not because of any
flaw in the permutation procedure, but because the choice of an
apriori threshold is arbitrary - no one threshold is more correct than
another.
TFCE gets around the latter problem by doing away with the selection
of arbitrary thresholds, and instead uses a weighted average of local
voxel activation magnitudes.
So the two techniques are not mutually exclusive.
- Tom (a different one)
On Fri, Apr 18, 2008 at 4:11 PM, Thomas Nichols <[log in to unmask]> wrote:
> Dear Antonios-Constantine,
>
> The Hayasaka & Nichols paper did not look at power and only did null
> hypothesis simulations. Hence by 'permutation behaves well' we meant that
> it had exact or not-too-conservative false positive rates.
>
> The comments in the Smith & Nichols reflect the empirical observation of
> many researchers that small changes to the cluster-defining threshold
> changes size and shape of clusters in unexpected ways, and, as a result,
> changes the significance. This is more a comment about (assumed) true
> signals and hence power.
>
> So the two comments concern specificity vs. specificity of cluster
> inference, and, in the former, are based on extensive Monte Carlo
> simulations, while, in the later, are based on informal observations of many
> years of neuroimaging data analysis. So they're not contradicting as much
> as just commenting on difference facets of cluster inference.
>
> Hope this helps.
>
> -Tom
>
>
>
> On Fri, Apr 18, 2008 at 2:48 PM, Antonios - Constantine Thanellas
> <[log in to unmask]> wrote:
> > Dear fsl users,
> >
> > According to "Validating cluster size inference :random field and
> > permutation methods" by Hayasaka and T.Nichols the cluster size inference
> > based on permutation tests behaves well for any threshold smoothness and
> > even the choice of threshold does not influence the permutation test
> > much.There's also a 0.01 value which is proposed as a "default value".
> >
> > According to "Threshold-Free Cluster-Enhancement Addressing the problem of
> > threshold dependence in cluster inference" by St.Smith and T.Nichols minor
> > adjustments in the cluster forming threshold can result in dramatic
> changes
> > of the resulting inferences.
> >
> > i find these two approaches contradicting, so can you please clarify this
> > cause i probably misunderstood something here.
> >
> > Thank you
> > Antonios-Constantine Thanellas
> >
> >
> >
>
>
>
> --
> ____________________________________________
> Thomas Nichols, PhD
> Director, Modelling & Genetics
> GlaxoSmithKline Clinical Imaging Centre
>
> Senior Research Fellow
> Oxford University FMRIB Centre
--
School of Psychology and CLS
University of Reading
3 Earley Gate, Whiteknights
Reading RG6 6AL, UK
Ph. +44 (0)118 378 7530
[log in to unmask]
http://www.personal.reading.ac.uk/~sxs07itj/index.html
http://beclab.org.uk/
|