Hi,

So 120 subjects overall. I'd see little case for variance smoothing. I would consider using it with less than maybe about 15 subjects, but this isn't a strict cutoff. Others in the list may have a different opinion.

All the best,

Anderson


On 22 July 2016 at 06:29, neuroimage analyst <[log in to unmask]> wrote:

Thank you Anderson for that quick reply. We have exactly 20 subject in 1 group and around 100 in another. Does this makes the case for variance smoothing? Regardless of our current problem, what should be a good rule of thumb that one should follow for variance smoothing?

I hope sombody could clear up our pcasl smoothing question.

Thank you again for your time.

Regards


On Jul 21, 2016 10:21 PM, "Anderson M. Winkler" <[log in to unmask]> wrote:
Hi,

About smoothing pCASL: I leave for others to comment.

About randomise: Assuming you followed the example from the manual, there is no need for -D as it makes in this case the design rank deficient (randomise may fix this internally but we should enter the design and options correctly anyway).

How many subjects are used? If the sample size is relatively large (e.g., >20) variance smoothing isn't much needed.

About -c: there is no strict rule but probably higher is better. Using 3.1 might be a good idea.

All other options seem fine.

All the best,

Anderson


On 22 July 2016 at 01:40, Neuroimage Analyst <[log in to unmask]> wrote:
Hello FSL Users,

We processed our pCASL dataset using BASIL with the --spatial option.

The original dataset had a resolution of 3.5x3.5x3.5 mm.

All the subjects were then normalised to MNI Space and have a resolution of 2 mm isotropic.

Now, we wanted to perform group level analysis on two groups to see if any regions have statistically significantly different CBF.

So, we smoothed the (already smoothed) normalised  perfusion map using 4mm Gaussian Kernel (using the command:
fslmaths each_subject_perfusion -kernel gauss 1.69866 -fmean each_subject_smoothed_perfusion_map)

Then we used randomise to compare 2 groups and used 4mm variance smoothing inside randomise. Here is the command:

randomise -i my_4d_2grp_concat_files -o comp -d design.mat -t design.con -D -m MNI_152_t1_2mm -v 4 -x -T -R --uncorrp

and found significantly different regions consistent to our hypothesis.

We were just wondering if our statistical analysis pipeline is right?

In addition, if anybody could tell us how do we extract FWE corrected cluster map from tstat* map obtained in the previous step. (Question really is what threshold to input in -c option of randomise?)

We really appreciate the time and suggestion to solve our problem.

Thank you.

Regards