In principle yes, but I'm reluctant to say "always" and be forgetting some scenario... In any case, Jeanette Mumford has an excellent page that explains mean centering and similar issues here: http://mumford.fmripower.org/mean_centering/
2012/12/3 Tseng Mark <[log in to unmask]>Thanks Anderson.So, since I only have one EV (behaviour scores) in the correlation map (I am interested in brain regions correlated with the scores), should I always add another EV with ones (representing group mean) in order to absorb the mean?
Thanks again.Mark2012/12/3 Anderson M. Winkler <[log in to unmask]>
Hi Mark,2012/12/2 Tseng Mark <[log in to unmask]>
Ah, thanks for your reminder of the bias, Anderson. I then used the mask in fsl atlas thresholded at 50% probability to do randomise, and the results were good. Is that ok?
Sounds ok.
Still one point confuses me. In the correlation map that I want to run randomise, there is only one EV, that is, the behaviour scores, which has been demeaned. (I think that's why the warning message in randomise said "All design columns have zero mean") Since I have done demean, why do I have to demean it again while running randomise?
The reason is that with all regressors having zero mean, there won't be any one in the matrix to absorb the mean that is almost certainly present in the image. It's as if fitting a line to the data and forcing it to cross the y-axis at zero, when the best fit probably would be somewhere else.
By mean-centering the data, all points are shifted so that the intercept will indeed be at zero and you'll get the best fit for the slope, which is what you care about. If a column with ones is included, it will absorb the mean while the slope will be captured by the other regressor (the one with the behavioural scores), such that both ways are equivalent.
Hope this helps!
All the best,
Anderson2012/12/2 Anderson M. Winkler <[log in to unmask]>
Hi Mark,
I find interesting that the input file is called "filtered_func_data.nii.gz", which immediately suggests it's a time series. In any case, if this file contains group-level data, and if the columns of the design matrix have all zero mean, then yes, you can run randomise with the option -D. You can also add a column with ones to the design.
Having said this, however, it sounds to me that there may be some selection bias in your study, i.e., using the peak voxels from one analysis to further run a second analysis on the same data. Note also that a sphere surrounding the peak isn't a safe FWER (or even FDR) procedure.
All the best,
Anderson2012/12/2 Tseng Mark <[log in to unmask]>
Hi Anderson,Thanks for your reply.No, not for 1st-level results.I have a hypothetical area, and this area was activated in a group activation map (map 1) of our subjects. I want to prove further that this area is correlated with a behaviour score. So, I did another group analysis to see if a COPE is correlated with the score (let's called it correlation map). I then created a spherical roi centred at the local maximal coordinate of map 1 and used it to do small volumn correction by randomise command in the correlation map.Mark2012/12/2 Anderson M. Winkler <[log in to unmask]>Dear Mark,
It sounds as if you were trying to run randomise for 1st level results (i.e., FMRI time series for a given subject), is this right? If yes, then randomise isn't the tool you need. Instead, in the Feat GUI, in the Post-stats tab, there are two options that may interest you:
- Pre-threshold masking, in which you can supply the mask you have (in the same space as the subject's data, which I believe in your case is not the standard, but the native space).
- Contrast masking, in which you can specify other contrasts, computed for the same subject, for masking.
In case you'd like to use a mask from a group analysis (then probably in standard space) to a single subject, then you'll need to warp it to the subject's native space before using it for the Pre-threshold masking.
Hope this helps!
All the best,
Anderson2012/12/2 Mark <[log in to unmask]>
Hi,
I want to do a small volume correction in a spherical ROI defined from another group contrast in my study. I created a spherical ROI centred in a certain coordinate with 5-mm radius (called sphere_roi.nii.gz), enter the cope directory that I want to analyse, and then run:
randomise -i filtered_func_data.nii.gz -o output_directory -d design.mat -t design.con -m sphere_roi.nii.gz -T -c 2.3
The permutation then started but, before that, there appeared a message:
Warning: All design columns have zero mean - consider using the -D option to demean your data.
Anywhere wrong? Should I do anything, such as adding -D?
Thanks in advance.
Mark