Christian Ruff wrote:
> Dear SPM-list,
>
> I am currently in the process of analysing the data from an experiment for
> which I would expect considerably small, but widespread activation in
> several regions of the brain. An exploratory analysis using the
> voxel-level statistics and a threshhold of p<0.05 corrected for multiple
> comparisons across the whole brain as well as in my prespecified ROIs
> revealed only some significant SINGLE voxels. However, a look at the
> results threshholded with p>0.001 uncorrected revealed that these voxels
> were indeed peaks of large clusters with smaller Z-values, while not many
> other clusters emerged. I hence feel justified to analyze the data with
> respect to the spatial extent of the activations.
>
> I have two questions in this context:
>
> First, is the calculation of the (un)corrected p-values for the clusters
> in SPM99 based on the size of the clusters only, or on a combination of
> their spatial extent and peak activation? I have already come across both
> explanations on the help list, so a clarification would be greatly
> appreciated.
>
The latter.
You set the height threshold to, say, p=0.001 and then specify the
extent eg. k=30 (SPM will not report statistical results for clusters having
fewer above threshold clusters than this).
Under the clusters column in the results printout you will then be
given the number of above threshold voxels in each cluster, V,
and a p-value which gives the (corrected and) uncorrected p-values for
each cluster (this is the probability of getting V voxels in that cluster
under the null hypothesis that there's no activation there).
>
> My second question in this context is which INITIAL height threshhold for
> the SPM(Z) should be used for this analysis (in other words: which p-value
> should be entered when prompted for the height threshhold of the
> statistical map, before having a look at the p-values for the clusters
> present in the map)? I understand that this choice considerably modifies
> the cluster-level inference (since it alters the size of the clusters),
> and that the initial height-threshhold should be theoretically
> pre-specified instead of "fishing around" at different height
> tresholds. As a default I would assume p<0.001 uncorrected is OK, but I
> wonder if there are actually papers dealing with the issue, or if anybody
> has any experience/suggestions.
>
If your signals are focal then a high threshold is optimal (we guard against
false positives).
If your signals are diffuse then a low threshold is optimal (we guard against
false negatives). If this sounds like ducking the issue then thats because it
is !
I'd welcome others to comment.
See you,
Will.
>
> Thanks a lot,
> Christian
>
> - - - - - - - - - - - - - - - - - - - - -
> Christian Carl Ruff
> Center for Cognitive Science
> Albert-Ludwigs-University Freiburg
> Friedrichstrasse 50
> D-79098 Freiburg
> Germany
>
> MAIL : [log in to unmask]
> PHONE: +49 761 2034937
> FAX : +49 761 2034938
> - - - - - - - - - - - - - - - - - - - - -
--
William D. Penny
Wellcome Department of Cognitive Neurology
University College London
12 Queen Square
London WC1N 3BG
Tel: 020 7833 7478
FAX: 020 7813 1420
Email: [log in to unmask]
URL: http://www.fil.ion.ucl.ac.uk/~wpenny/
|