Hi Tony,

If this is not 1st level fMRI, my suggestion is to use permutations. Say X is the 4D image (the 4th dimension being the subjects), and Y is the other variable that you're correlating with the image (Y can also be another image of the same size as X). Then write a script in Matlab to shuffle Y randomly, recompute the Spearman's correlation (let's call it r), checking the largest value of r across the statistical map. For each voxel in the unpermuted image, increment a counter if that voxel has larger or equal r than the maximum across the image for that permutation. Repeat this procedure, in a for-loop, many times (a few thousands), and at the end, divide the counter by the number of permutations. Make sure that the 1st permutation in the non-permuted (i.e., without any shuffling). The output will be an image with the corrected p-values. This gives strong, familywise error rate control (FWER), already taking into account, implicitly and without modelling, the spatial interdependence between voxels.

If Y is another 4D image, each volume needs to be shuffled as a whole.

The same idea applies to corrected p-values for clusters. Have a look at Matlab's function bwconncomp, which labels and calculates the sizes of connected voxels (it's part of the Image Processing toolbox; if you don't have it, it's available in recent Octave versions in the "image" package; another option is Jesper Andersson's spm_bwlabel funcion, that comes with SPM). You'd need to run the labelling for each permutation, and increment the counter just in the same way as for the voxelwise image.

Alternatively, and much simpler (or if this is 1st level fMRI), is if you are calculating the Spearman p-values using some parametric approximation. In this case, you can use false discovery rate (FDR). It's not the same as FWER, but it's an assumption free method to address the same multiple testing problem.

Attached it goes a little function that does FDR correction in Matlab. You can save as NIFTI and use FSL's "fdr" command.

Hope this helps!

All the best,

Anderson


On 25 September 2014 20:32, Tony Jiang <[log in to unmask]> wrote:

Lets say I did rank-based correlation,(spearman R) what can we do when finding the clusters of significance.

 

Thanks,

Tony

 

From: FSL - FMRIB's Software Library [mailto:[log in to unmask]] On Behalf Of Anderson M. Winkler
Sent: Thursday, September 25, 2014 11:26 AM
To: [log in to unmask]
Subject: Re: [FSL] multiple comparisson correction on stat maps

 

Hi Tony,

This is far from a simple topic. What nonparametric test are you using?

All the best,
Anderson

 

On 25 September 2014 15:14, Tony Jiang <[log in to unmask]> wrote:

Dear FSL community,

 

How do you deal with stat maps created from other software (for example: created from some nonparametric tests coded in matlab) regarding to multiple comparisons and clustering? Any general idea?

 

Thanks,

Tony

 

Kessler Foundation Signature

 

Kessler Foundation Facebook PageKessler Foundation You TubeKessler Foundation Research Twitter Page


The information in this transmission is intended for official use of the Kessler Foundation. It is intended for the exclusive use of the persons or entities to which it is addressed. If you are not an intended recipient or the employee or agent responsible for delivering this transmission to an intended recipient, be aware that any disclosure, dissemination, distribution or copying of this communication, or the use of its contents, is strictly prohibited. If you received this transmission in error, please notify the sender by return e-mail and delete the material from any computer.   ญญ