Dear Pia,
With VBM, stationary smoothness assumption is unlikely to be met so you
have to be particularly be aware of this for cluster level inference.
See eg section 5 of this:
http://www.sciencedirect.com/science/article/pii/S1053811908000438
Best regards,
Guillaume.
On 06/12/16 06:23, Pia Rotshtein wrote:
> Dear Guillaume and Torben,
> Thank you for your help it was indeed the non stationary default flag
> that had affected the results.
> This made me wonder what is the rationale of non-stationary correction.
> I can see the advantages for small sub-cortical structures. But when
> there is a large true cluster (such as often the cases with
> function-lesion mappings), locally it will always appear more smoothed.
> best
>
> pia
> ------------------------------------------------------------------------
> *From:* Torben Lund [[log in to unmask]]
> *Sent:* 07 November 2016 12:41
> *To:* Pia Rotshtein
> *Cc:* spm Parametric Mapping)
> *Subject:* Re: [SPM] Cluster level correction SPM12
>
> I guess this could be if you have changed the nonstationarity flag, in
> which case the critical cluster size will depend on the loca smoothness
>
>
> Best
> Torben
>
>
>
>
>
>> Den 7. nov. 2016 kl. 13.00 skrev Pia Rotshtein <[log in to unmask]
>> <mailto:[log in to unmask]>>:
>>
>> Dear Experts,
>> We are running a VBM study. We used peak-level cluster of p <.001
>> uncorrected, and 50 voxels.
>> What we found puzzling in the results table (which we never observed
>> before) is that sometime smaller clusters are denoted as more reliable
>> than larger clusters.
>> I took a screen shot to demonstrate the point:
>> <image003.jpg>
>>
>> So you would see that clusters with 200 (highlighted cluster), or with
>> 366 (first cluster) voxels are reliable at FWE and also as uncorr,
>> while clusters with 507 voxels are not (the cluster just above the
>> highlighted one).
>> So Our questions:
>> 1) Has anything changed lately in SPM12?
>> 2) And also how can that be the case; I though SPM corrected for
>> number of expected cluster which should be identical for all clusters
>> in a given comparison/model.
>>
>> Thank you
>>
>> Pia
>
--
Guillaume Flandin, PhD
Wellcome Trust Centre for Neuroimaging
University College London
12 Queen Square
London WC1N 3BG
|