Print

Print
Hello,
 
1) This procedure sounds dangerously circular: If your whole brain analysis had shown significant activation in the ROI ( e.g. the thalamus ) would you still have run the ROI-based analysis? It’s always better to decide on the choice of masks etc before running any statistics.

2) Randomise is running a linear regression on your data, generating p-values from an empirically determined null-distribution of a particular test-statistic ( here TFCE ). These could be transformed to partial correlation co-efficients if you want to report those.

3) You are free to choose any mask you want, although as mentioned in 1) the choice of ( e.g. ) a thalamus mask should be made before running a whole brain analysis.

Hope this helps,
Kind Regards
Matthew
--------------------------------
Dr Matthew Webster
FMRIB Centre 
John Radcliffe Hospital
University of Oxford

On 6 Aug 2018, at 21:10, John anderson <[log in to unmask]> wrote:

Dear Dr Anderson Winkler,
I have two groups of subjects (patients n=85 and healthy controls n=33). I would like to use Randomise in a voxel-wise correlation analyses between clinical scales and whole brain signal. For that objective I ran the following command:
randomise -i 4D_data.nii.gz -o vw -m MNI152_T1_2mm_brain_mask.nii.gz -d design.mat -t design.con -n 5000 -T

I threshold the resultant statistical map (vw_tfce_corrp_tstat1.nii.gz) at p<0.05 and this showed significant clusters for some of the analyses in specific region of interest (ROI) in the brain that support the analysis-hypothesis. I mean, I found correlation in whole brain voxel-wise correlation analysis between the clinical scales and brain MR signal in pathologically relevant regions where I really expect a correlation.

Some other analyses were not significant at p<0.05 but I saw clusters of correlation between the clinical scales and brain signal at p<0.3 in specific regions of the brain where I really expect correlation. In order to get around this issue (i.e. correlations are not significant but I can see it in pathologically relevant regions using sub significant-thresholds): To make these correlations significant, I replaced the whole brain mask  in "randomise" by another mask for an ROI where I expect the correlation to be significant. I ran the previous randomise command and indeed the non significant correlation become significant at p<0.05.

My questions are:
1- Is this procedure correct? I mean replacing the whole brain mask by a smaller mask to narrow the voxel-wise analysis from whole brain voxel-wise correlation to an "ROI based voxel-wise correlation". Can I still able to report such analysis in a manuscript under "ROI based voxel-wise correlation"  instead of "whole brain correlation analysis"?

2- What is the difference between this "ROI based voxel-wise correlation" and pearson correlation within the ROI (between the clinical scales and the MR signal in the ROI)?

3- Are there any rules (i.e size of the mask, number of the voxels in the mask, ...) for choosing the mask in the flag "m" in Randomise. In other words-mathematically, should we always use whole brain mask to reduce the false positives after correcting the results to multiple comparison? Are the smaller masks would increase the false postivities? Let's say we expect correlation in a small regions of interest (e.g thalamus). Whole brain correlation would not show the effect significat. Can I replace whole brain mask by a mask for the thalamus to make the correlations significant.

Thank you so much for any clarification,
John

########################################################################

To unsubscribe from the FSL list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=FSL&A=1


To unsubscribe from the FSL list, click the following link:
https://www.jiscmail.ac.uk/cgi-bin/webadmin?SUBED1=FSL&A=1