Dear Dr. Smith
We followed your suggestion of making a prethreshold mask with our a
priori ROI's (Your comment from below: "but if you have a priori ROIs
then you can use these in prethreshold masking to reduce the effects
of multiple comparison correction".). As a test, we ran this on a
few contrasts where we had activations on a whole brain basis and
find that we actually find no activations using the mask with the
same cluster correction as the whole brain analyses. This is rather
odd and we cannot understand why this is happening.
When we run the mask uncorrected, it works as would be expected -
showing the same voxels as in the z-stat images but only if they were
in the mask. In checking the log files, we do not see any errors or
points of difference between the whole brain cluster corrected
analyses and the ROI masked cluster corrected analyses.
any thoughts?
On Feb 21, 2009, at 8:28 AM, Steve Smith wrote:
> Hi
>
> --------------------
> Stephen M. Smith, Professor of Biomedical Engineering
> Associate Director, Oxford University FMRIB Centre
>
> FMRIB, JR Hospital, Headington,
> Oxford. OX3 9 DU, UK
> +44 (0) 1865 222726 (fax 222717)
> [log in to unmask]
> http://www.fmrib.ox.ac.uk/~steve
> ----------------------
>
> On 20 Feb 2009, at 07:49, Sarah Master <[log in to unmask]> wrote:
>
>> Dear FSL users,
>> We have run a study with 17 participants, each with 3 runs. After
>> running
>> our models on each run (lets say 3 EV’s per run), we combined the
>> 3 runs per
>> subject under a fixed effects framework. We then combined the
>> resulting
>> COPE files using FLAME 1&2 and threshold at Z=2 and cluster p<.05.
>>
>> In reviewing the results, we find very few, if any, significant
>> voxels in
>> the thresh_zstat images. However, when reviewing the zstat
>> images, there
>> are quite a few voxels (and only in a priori, hypothesized
>> regions) with z’s
>> above 2.0. We assume that the cluster correction is the sole or
>> primary
>> source of the lack of significance. Is this correct?
>
> Yes unless there are problems in the analysis such as data
> artefacts, misregistrations or misspecification of the model.
>
>
>
>>
>>
>> We have tried running the analyses using FLAME 1 with the same
>> results. We
>> also tried running the data under a Fixed effects framework for
>> exploration
>> purposes. This obviously led to far more identified voxels, but
>> running
>> Fixed effects analyses are not appropriate at this stage.
>>
>> We have several questions.
>> Is it correct to interpret this pattern as resulting from just sub-
>> threshold
>> activation/effects?
>> Might there be a problem we are missing?
>
>
> See above.
>>
>> Would it be appropriate to report any statistics or findings from the
>> uncorrected zstat images in a write-up of results? (for example,
>> voxels with
>> z-values above a certain number?)
>
> Not in general, but if you have a priori ROIs then you can use
> these in prethreshold masking to reduce the effects of multiple
> comparison correction.
>
> Cheers
>
>
>>
>> Short of collecting more data (which is not an option), do you
>> have any
>> suggestions?
>>
>> We have been working on analyzing the data using ROI analyses to
>> work around
>> the relatively stringent correction in whole brain analyses but
>> suggestions
>> will be appreciated.
>>
|