Hi,

On 24 Nov 2008, at 06:26, Emma Bendall wrote:

> Hi Steve and FSLers,
>
> I have run into a problem that I hope someone can shed light on. I  
> refer to
> the design matrix in my previous email (below) with additional  
> contrasts to
> allow for testing for negative as well as positive correlations. To
> reiterate, I am using a paired t-test on FA data obtained for a  
> group at two
> separate time points with an additional EV (behvioural difference  
> scores,
> demeaned and orthogonalised wrt EV1). The FA data was processed  
> using TBSS.
> The design matrix is similar to:
>
> Inputs        Group  EV1  EV2  EV3  EV4  EV5  EV6
> SessA_s1     1        1       1      0      0      0      -2.25
> SessA_s2     1        1       0      1      0      0      -2.25
> SessA_s3     1        1       0      0      1      0       0.75
> SessA_s4     1        1       0      0      0      1       3.75
> SessB_s1     1       -1       1      0      0      0       2.25
> SessB_s2     1       -1       0      1      0      0       2.25
> SessB_s3     1       -1       0      0      1      0      -0.75
> SessB_s4     1       -1       0      0      0      1      -3.75
>
> With contrasts:
>                              EV1  EV2  EV3  EV4  EV5  EV6
> Con1  A-B              [  1      0       0       0       0       0 ]
> Con2  B-A              [ -1      0       0       0       0       0 ]
> Con3  A-Bmod+ve  [  1      0       0       0       0       1 ]
> Con4  B-Amod+ve  [ -1      0       0       0       0       1 ]
> Con5  A-Bmod-ve   [  1      0       0       0       0      -1 ]
> Con6  B-Amod-ve   [ -1      0       0       0       0      -1 ]
>
> Firstly, is it correct to simply include Con5 and Con6 to obtain  
> negative
> correlations, or should I alter the design matrix in some way?

The design matrix is fine but contrasts 3-6 aren't right. A sensible  
contrast would be
[0 0 0 0 0 1]
which tests for positive correlation with the behavioural difference.  
It doesn't really make sense to include EV1 and EV6 in the same  
contrast - the results will be hard to interpret, as you found.

Cheers.



>
>
> Secondly, assuming the above matrix and contrasts are correct, is it
> possible that a small n (e.g. n = 5) could be responsible for FSL  
> randomise
> outputting contradictory results that survive cluster thresholding and
> additional Bonferroni correction? For example, I ran randomise using  
> the
> above design and contrasts with cluster threshold of 3 and 10000
> permutations. A cluster resulting from the Con4 test overlapped  
> (almost
> completely, though it was about 40% the size of) a cluster resulting  
> from
> the Con3 test. I checked all_FA_skeletonised voxel values within each
> respective cluster (using fslmeants with -showall option and  
> standard stats
> software) and conclude that the Con4 result is incorrect. The other  
> results
> seem reasonable, however I should probably abandon the test procedure
> altogether given that randomise has output what seems to me to be at  
> least
> one nonsense result.
>
> A similar situation occurred recently with a test group of slightly  
> larger
> n, where a cluster from Con6 (negative correlation) overlapped with a
> cluster from Con4 (positive correlation). Again I suspect the small  
> n is the
> source of my problem, but I'd be grateful for any further insight or
> suggestions regarding how randomise might be coming up with these
> contradictory results.
>
> Thanks,
>
> Emma
>


---------------------------------------------------------------------------
Stephen M. Smith, Professor of Biomedical Engineering
Associate Director,  Oxford University FMRIB Centre

FMRIB, JR Hospital, Headington, Oxford  OX3 9DU, UK
+44 (0) 1865 222726  (fax 222717)
[log in to unmask]    http://www.fmrib.ox.ac.uk/~steve
---------------------------------------------------------------------------