Print

Print
Hi Matthieu,

Please, see below:


On 30 May 2016 at 10:29, Matthieu Vanhoutte <[log in to unmask]> wrote:
Hi Anderson,

Please see below :

2016-05-28 10:16 GMT+02:00 Anderson M. Winkler <[log in to unmask]>:
Hi Matthieu,

Please, see below:


On 27 May 2016 at 10:30, Matthieu Vanhoutte <[log in to unmask]> wrote:
Hi Anderson,

Thanks for you clear answers ! 

Please see below :


2016-05-27 11:03 GMT+02:00 Anderson M. Winkler <[log in to unmask]>:
Hi Matthieu,

Please, see below:


On 27 May 2016 at 09:58, Matthieu Vanhoutte <[log in to unmask]> wrote:
Dear FSL experts,
I am planning to use PALM for statistical analysis at the surface level.  Could you please explain me your choice of "palm" different parameters (below) and anwser to me about some questions ? :
For completeness, here is my PALM setup: I used palm version palm-alpha95version
Here are my commandline parameters:
-i All_inputs.mgh
-s fsaverage/surf/lh.white
-n 10000
-m mask.mgh (freeSurfer's mask to exclude subcortical surface)
-Cstat extent
-C 2.3
-d Xg.csv
-t contrast/mycontrast.csv
1) How do you compute the z-threshold (-C parameter)  and which kind of values are valuable ?

There isn't a strict rule for the cluster-forming threshold. Higher is probably better, and a recommendation maybe is to use something as 3.1. Please see this paper for an interesting discussion:

Woo C-W, Krishnan A, Wager TD. Cluster-extent based thresholding in fMRI analyses: pitfalls and recommendations. Neuroimage. 2014 May 1;91:412-9.


OK, I will look at this paper. In the pvalues FWER-corrected within contrast 1 file ("palm.clustere_tstat_fwep_c1.mgz"), I got corrected clusters but some of them have pvalue > 0.05. Is this necessary in this corrected file to threshold the resulting clusters at a pvalue <= 0.05 ?


The file with the p-values shows all clusters that are formed after the cluster-forming threshold (2.3 in your case). Some of these might be considered significant (fwep < 0.05) or not. So yes, need to threshold this image to keep only the smaller p-values.

Alternatively, consider using the option -logp, such that you threshold the p-values map at -log(0.05)=1.301. It makes things easier, and helps when generating figures too.

Ok thank you for the tips concerning figures. What the "pmethodp" and "pmethodr" parameters concretely correspond to ?

These refer to the method used for partitioning of the model into effects of interest and nuisance effects. The partitioning occurs for each contrast, and can use one of three methods (Guttman, Beckmann, or Ridgway), all described in the randomise paper.

This partitioning can happens in two different places: (1) to define the set of permutations and (2) to actually do the regression, i.e., the model fit. To choose the method for (1) use -pmethodp, whereas for (2) use the -pmethodr. These settings, however, rarely need to be changed, except in very special cases.

 

Another point : once permutation and uncorrected results computed, is it possible to launch different multiple comparisons corrections without re-computing all the permutations ?

Unfortunately not. The correction uses the distribution produced with all permutations, and these are thrown away internally at each iteration. While it's possible to save them (-saveperms), using them would require a custom script. It's probably simpler to run again.

All the best,

Anderson

 

Best regards,
Matthieu


All the best,

Anderson

 

 

2) Is this necessary to manually mean-centered covariates in statistical model with two or more groups and is this design below coherent (for example 3 groups with 2 patients and 2 mean-centered covariates) ? :

EV1 EV2 EV3 EV4 EV5
1 0 0 6.2346573213 -10.578125
1 0 0 -3.8105172167 -3.578125
0 1 0 -6.9508320696 3.421875
0 1 0 6.8835279578 1.421875
0 0 1 -1.4942954508 3.421875
0 0 1 -0.1171565049 4.421875

It depends on the contrast. If testing the last two covariates, not necessary, but if testing either of the first three alone (not their differences) then yes.

See Jeanette Mumford guide on this matter: http://mumford.fmripower.org/mean_centering/ 

All the best,

Anderson

Best regards,
Matthieu