Dear Qizhu,
Here is a suggestion on what may be going on: the 1st eigenvariate is
computed through SVD, which has a sign indeterminacy. This is not an
error, but something intrinsic to the method. SPM, it seems, resolves
the ambiguity by multiplying the 1st eigenvariate by the sign of the sum
of the elements in 1st right-hand singular vector, i.e., being the SVD
of Y given by USV, return U(:,1)*sign(sum(V(:,1))) [and scale
appropriately taking the corresponding variance and number of columns in
Y into account].
If you have enough evidence that the sign is flipped, you may safely
flip it manually again for your subsequent analyses, or consider using a
different (but slower) method to try to solve the indeterminacy. See the
link below for a function available in MATLAB Central that uses a
different (and slower) strategy:
http://www.mathworks.com/matlabcentral/fileexchange/22118-sign-correction-in-svd-and-pca
Hope this helps,
Anderson
On 01/11/2011 04:19 PM, Qizhu Wu wrote:
> Dear all,
>
> Actually this a topic has been posted by Stefan Teipel one year ago (038560: Eigenvariate SPM8: disagreement between statistical output and raw values?).
> That is: I analyzed my FA maps from 2 groups (A and B) of subjects using a simple 2-sample t test. With a T contrast [1 -1], I got significant cluster. When plotting the peak voxel over scan/time (actually subject here), the graph looks Ok that Group A higher than Gruop B; however, when plotting the eigenvariate, I got opposite results, in which Group A are lower than group B.
> I have searched through the email archive, no response to Stefan's question was found.
> I wonder how how does this happen?
>
> Qizhu Wu
> Huaxi MR Research Cneter
> West China Hospital of Sichuan Uni
>
|