Hi Arsene,

Please, see below:


On 12 June 2017 at 13:16, Henri Longin <[log in to unmask]> wrote:


Hi Anderson,



Thanks for your help...  Others points to clarify :


1)  In the PALM command below, you suggested to set the output as  -logp  instead of -save1-p, why?

I'm asking this because I'm accustomed in using 1-p result with randomise... What does -logp bring more?


palm   -i    My_Mod_Merg.nii    -d    design.mat     -t    design.con    -m    My_Mask.nii    -eb    design.grp      -T    -n 5000    -logp    -nouncorrected    -approx tail    -corrcon    -o  Myresults


The -logp gives better contrast among the significant regions, which aren't shrank into such narrow interval between 0-0.05 (or 0.95-1). Huge effect size differences may be within the interval, and these differences aren't well reflected in the contrast using p-values (or 1-p). Putting in log-scale improves contrast dramatically, and puts the colours in a (roughly) linear scale with the size of the effect and the test statistic (t-stat).

 


2) When building both the between-subject and within-subject designs with FEAT GUI as we discussed, I received a warning message (see attached)...



I don't see any attachment. However, if the error has to do with rank-deficiency, then when the model was simplified there may have been an error. If the message is about the design having to be separable, in this case the message can be ignored.
 


3)

> It is possible. Either use a separate design with only these timepoints (e.g. t2 for all subjects, ignoring the other timepoints, and then doing a simple, non-paired, two-sample t-test),

> or include in the between-subject design additional EVs that effectively remove the other timepoints (that is, EVs full of zeroes, except for the scan that is to be removed; one such

> EV per scan).


Hummm... This is not simpler to guess...

To turn the question another way round, how could look the between-subject design If I only want to compare the 4 pair-wises using the second strategy? Ex:

  • t1 > t1       (G2>G1, for t1)
  • t2 > t2       (G2>G1, for t2)
  • t3 > t3       (G2>G1, for t3)
  • t4 > t4       (G2>G1, for t4)

I need to compare this as the my second group is using as the control of the first group

You could run 4 separate two-sample t-tests. These are perfectly valid. However, with 4 tests, there is an increased risk for false positives. To correct, run all 4 into the same call to PALM. Something as this:

palm -i data_t1.nii.gz -i data_t2.nii.gz -i data_t3.nii.gz -i data_t4.nii.gz -d design.mat -t design.con -corrmod -corrcon [... other options ...]

The 4 input "modalities" are the data for the 4 timepoints. The subjects must be in the exact same order for all image files. The above assume the same designs for all, but let's say that at each timepoint you have a different nuisance variable (say, systolic blood pressure at the time of the scan). Then each timepoint needs its own design, and the call would be something as this:

palm -i data_t1.nii.gz -i data_t2.nii.gz -i data_t3.nii.gz -i data_t4.nii.gz -d design1.mat -d design2.mat -d design3.mat -d design4.mat -t design.con -designperinput -corrmod -corrcon [... other options ...]

Hope this helps!

All the best,

Anderson

 



Thanks in advance,



Arsene