Dear Sascha,
So, as expected, the peak-level FWE-corrected p-value does not change
with the height threshold. In contrast, topological FDR looks at peaks
whose statistic is above a given threshold and therefore depends on that
initial height threshold (hence the different qFDR-corr values you
observe), see:
https://doi.org/10.1016/j.neuroimage.2009.10.090
At last, cluster-level inference relies on the specification of a
(height) cluster-forming threshold (to define clusters) so not only are
the clusters of different sizes (kE equal to 6 vs 46 voxels) but their
associated cluster-level FWE-corrected p-values differ (the lower the
threshold the larger a cluster will be and the more likely it will be to
be observed under the null).
Remember also that an assumption of the random field theory used for
cluster-level inference (and topological FDR) is that the cluster
forming threshold is sufficiently high (eg p<0.001 uncorrected).
Have a look at Tom Nichols' presentation on the subject for more details:
https://mediacentral.ucl.ac.uk/Player/2938
http://www.fil.ion.ucl.ac.uk/spm/course/slides17-oct/05_ParamThresh.pptx
Best regards,
Guillaume.
On 14/11/17 15:36, Sascha Frölich wrote:
> Dear Guillaume,
>
> thank you for your reply! However, then I do not understand the following:
>
> If I choose "p value adjustment to control" -> FWE=0.2 I get the
> following result for one specific cluster and contrast:
>
> At peak-level
> puncorr=0.000
> qFDR-corr=0.238
> pFWE=0.052
>
> At cluster-level
> puncorr=0.293
> kE=6
> pFWE=0.063
>
> Then, when I choose "p value adjustment to control" -> none, "threshold
> {T or p value}" = 0.001, I get a cluster at the same position (mm mmm
> mm), but with different values:
>
> At peak-level
> puncorr=0.000
> *qFDR-corr=0.054*
> pFWE=0.052
>
> At cluster-level
> puncorr=0.1
> *kE=46*
> *pFWE=0.745
>
>
> *
> Both results have pFWE at peak-level well below 0.2, so why are the
> resulting clusters, and thus their cluster-level significance, so
> different? Why don't I get the same cluster twice?
>
> Thanks!
>
> Best,
> Sascha
>
> 2017-11-14 15:15 GMT+01:00 Guillaume Flandin <[log in to unmask]
> <mailto:[log in to unmask]>>:
>
> Dear Sascha,
>
> You are right. The choice you make at that stage defines the height
> threshold to be applied on the statistical map. The FWE option will
> ensure a control of the family-wise error rate at the peak level for the
> entire search volume.
>
> Best regards,
> Guillaume.
>
>
> On 13/11/17 16:08, Sascha Frölich wrote:
> > Dear all,
> >
> > After performing a two-sample t-test and estimating the model, I
> want to
> > visualize the results by clicking the "Results" button and
> choosing the
> > corresponding SPM.mat-file. After SPM asks if I want to apply masking,
> > it asks "p value adjustment to control" with the possible option
> "FWE".
> > What exactly is this option? I suppose it displays only those results
> > where pFWE at peak-level is less than the entered threshold. Is that
> > correct?
> >
> > Thanks in advance!
> >
> > Best,
> > Sascha
>
> --
> Guillaume Flandin, PhD
> Wellcome Trust Centre for Neuroimaging
> University College London
> 12 Queen Square
> London WC1N 3BG
>
>
--
Guillaume Flandin, PhD
Wellcome Trust Centre for Neuroimaging
University College London
12 Queen Square
London WC1N 3BG
|