Hi Anderson,Thanks for your information. Just another question about how randomise works for bad voxels in GLM. If one voxel in one subject from a group has zero intensity, while others have non-zero intensities for the same location, will randomise still consider that zero value for permutation, or simply skip that voxel for analyses? If the former is true, any way to take advantage of the information from others except that one have zero value? Thanks again.
ShengweiOn Fri, Aug 1, 2014 at 7:20 PM, Anderson M. Winkler <[log in to unmask]> wrote:AndersonAll the best,A suggestion is to instead compute the global volume of WMH, and relate that to cognition. To gain a bit in spatial specificity, the volume of WMH can be computed per tract, using, for instance, the regions of the JHU atlas, and then relate these regional, tractwise measures, to the cognitive measurements.Hi Shengwei,Thanks for explaining. Yes, randomise can handle this and it's still valid, but I'm afraid just reversing the sides of the equation (with voxelwise EVs, etc) won't solve the problem, that is, the fact that the lesions don't have matching topography across subjects.
On 1 August 2014 15:55, Shengwei Zhang <[log in to unmask]> wrote:
I'm not very familiar with the statistical process behind randomise. Thanks again for your information!So my question is: can randomise still handle this? It's suggested that the cognition and WMH can be inter-changed in the model (i.e. model cognition as a function of WMH plus others). In that case I just have a score instead of an image for each subject as input for randomise, and I'm filling the corresponding score to the brain mask of the subject. Would this still work and would the p-value after correction still valid?The goal was to model WMH from a group of subjects in the same space as a function of cognition, controlling for age, sex, etc. in GLM using randomise. The WMH map is binary, and the cognition was given by a score for each subject. What I'm worry about is that the model may fail in voxels where most subjects (e.g. 95%) have or don't have WMH.Hi Anderson,Thank you for your reply. Just want to provide more details for clarification.
ShengweiOn Thu, Jul 31, 2014 at 9:35 PM, Anderson M. Winkler <[log in to unmask]> wrote:Not sure if it helps. Maybe if you give more details.The spatial statistics (cluster extent, cluster mass and TFCE) are, however, only meaningful for actual imaging data. If these quantities that you define as scalars aren't truly imaging data (i.e., not voxelwise), and were stored as NIFTI just for some convenience, these statistics aren't meaningful.Hi Shengwei,I'm not sure if I understand. The images we model in randomise are all scalars. Do you mean something that is constant across space and/or constant across multiple subjects? If yes, randomise can still be used. FWE results are valid, and if all voxels across space are identical, it will give the same result as the uncorrected (and both are fine then).
All the best,
Anderson
On 31 July 2014 17:14, Shengwei Zhang <[log in to unmask]> wrote:
Any help is appreciated.Can I just fill that scalar quantity in the brain mask for each participant and use it as input for randomise? Will it affect the way that randomise is supposed to work to correct FWE or TFCE?Hi FSL experts,I have a question about the input for randomise. I'm interested in modelling a scalar quantity as a function of voxel-dependent EV plus other EVs, and using randomise for the analyses.
Shengwei