Dear Dr Gianelli,
It occurred to me that I used language yesterday which might have
been a bit confusing for understanding the cluster-level statistics,
in that I referred to two different things as the 'threshold'.
You have chosen a threshold above which voxels appear in your spm,
which one might call the 'spm threshold', typically p < 0.001
uncorrected. However, you also have a different threshold above
which you will accept results as significant. If, for example, you
had no prior hypothesis, then you might set this level at p < 0.05
corrected.
The cluster-level statistics refers to the probability of getting at
least the observed number of voxels in the cluster, given the 'spm
threshold' that you have set.
I hope that this makes it a bit clearer!
Best wishes,
Richard.
>X-Sender: [log in to unmask]
>Date: Wed, 4 Oct 2000 15:38:57 +0100
>Subject: Re: Cluster level
>From: Richard Perry <[log in to unmask]>
>To: Marco Giannelli <[log in to unmask]>
>Cc: [log in to unmask]
>X-Unsub: To leave, send text 'leave spm' to [log in to unmask]
>Reply-To: Richard Perry <[log in to unmask]>
>Sender: [log in to unmask]
>
>Dear Dr Gianelli,
>
>>I run SPM to analyze a fMRI-time series.
>>I have a cluster in the Z-map with the result:
>>
>>cluster-level
>>Pcorrected=0.078
>>Ke=6
>>Puncorrected=0.068
>>
>>voxel-level
>>Pcorrected=0.192
>>T=5.46
>>Z=4.31
>>Puncorrected=0.068
>
>Have you used small volume correction? I only ask, because these
>voxel-level results don't appear to have been corrected for the
>whole brain. Very little correction has been applied, suggesting
>that the total number of resels corrected for is very low, whereas
>for a whole-brain acquisition you might expect there to be hundreds
>of resels. The other possibility is that your data is very smooth.
>
>>If I see the voxel-level results I think that the activation is a
>>false positive
>>(Pcorrected=0.192).
>
>There is no reason to think that the result is a 'false positive'.
>It simply hasn't reached the arbitrary level of 5% corrected
>significance, and as a result you would not report it as a
>significant deviation of the data from the null hypothesis (= no
>activation). However, it may well be a real effect that you just
>didn't have enough statistical power to demonstrate conclusively.
>
>>If I see the cluster level I am puzzled: Pcorrected=0.078.
>>My activation is a false positive or a true activation?
>>In Z maps results is more meaningful the voxel-level or cluster-level?
>>What's the difference between cluster-level and voxel-level?
>
>The voxel level and the cluster level statistics are asking
>different questions. The voxel level statistics are to do with the
>height of the effect within a single voxel. This question is asked
>independently of the other voxels (except that all of the data was
>smoothed before the stats was done, of course).
>
>The cluster level statistics is to do with the probability of
>getting, by chance, a cluster of a given size or larger, when the
>data is thresholded at the particular level that you have set.
>
>One is no more meaningful than the other. If you had a large
>diffuse activation, it is possible to imagine that you would have a
>significant result at the cluster level, with lots of voxels just
>managing to exceed the threshold, but a non-significant result at
>the voxel level, because none of these individual voxels exceeds the
>level defined as 'significant'.
>
>On the other hand, we more usually find ourselves in the situation
>of a relatively small cluster with highly significant results at the
>voxel level. The cluster level may in this example not even come
>close to significance, because this only depends on the number of
>voxels in each cluster which exceed your threshold, not on the
>amount by which that threshold is exceeded.
>
>>Thank you in advance for your help.
>
>You're welcome. You might also want to look at the following archived e mail:
>http://www.mailbase.ac.uk/lists/spm/2000-04/0042.html
>
>Best wishes,
>
>Richard Perry.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|