Hello FSL list,
The problem outlined below (previous posts on this topic) remains a mystery to me... I can get by without solving it but I'm also curious to understand what is going on. Every input is welcome: speculations as well as complete explanations.
Best wishes,
Emma
------------------------------------------------------------------------------
Dear all,
Apologies to bother you again with Ftest questions but I'm still puzzled (see previous posts "randomise interaction F-test") but now with a different issue.
After Ftests on 3 group comparisons gave me very different results from what I expected given previously performed planned T-contrasts, I got a little suspicious. So I ran both an F-test and a T-test on 1 contrast using randomise with TFCE (on TBSS data) and I get corrected and uncorrected p-values that are completely different. The T-test results in uncorrected p-values as low as 0.0014 and corrp values of 0.0772, while the Ftest gives a minimum uncorrected p of 0.0088 and a minimum corrected p of 1 (fslstats -R on this image results in 0.000000 for both maximum and minumum). All of this despite fslmaths -sqr of the raw Tstat image ensured compatible raw T and F statistics in each voxel (F=T²). Given I'm testing only 1 contrast with the Ftest, I think they should also have the same DoF.
So I'm wondering what is going on... Am I consistently doing something wrong? I wouldn't know how because both analysis just ran with the same design files and 1 command. Could it be that because Fstats are non-directionally specific, random permutations may result in larger clusters and thus higher thresholds? Or does TFCE change because of this non-directionality of F stats? I'm not sure how to compare the TFCE stats between the two methods, but it could be meaningful if I could test their correspondence in a similar way as I did with the raw stats. Finally, could it be the version of randomise I'm using? It's randomise v2.1 in FSL 4.1.1.
Best wishes,
Emma
--------------------------------------------
Hi Jeanette,
I wasn't expecting a perfect match, and not even a perfect match between 2*p(Ttest) and p(Ftest). I just wasn't expecting as big a difference as I got (see summary measures below). Especially the corrected p values differ no end: minima of 1(?!) versus 0.07. With respect to raw p-values, the robust minumum raw p in the Ftest is more than 70 times higher than the Ttest's.
Although the F fits T² perfectly in every voxel, the TFCE Tstats are higher than the TFCE Fstats. I was just wondering if this has something to do with the bi-directional quality of F statistics: could it make the data less spatially consistent? And secondly, could an Ftest lead to larger clusters in permutations / by chance?
Many thanks for your help,
Emma
---------------------------------------------
SUMMARY MEASURES:
Corrp: fslstats -R
Ftest: 0.000000 to 0.000000
Ttest:. 0.000000 to 0.923800
Corrp: fslstats -r
Ftest: 0.000000 to 0.000000
Ttest: 0.019400 to 0.897010
Raw p: fslstats -R
Ftest: 0.000000 to 0.991200
Ttest: 0.000000 to 0.998600
Raw p: fslstats -r
Ftest: 0.098129 0.363770
Ttest: 0.567205 0.991610
TFCE: fslstats -R
Ftest: 0.000000 160431.156250
Ttest: 0.000000 183206.484375
TFCE: fslstats -r
Ftest: 3850.347656 56792.628906
Ttest: 366.412964 153343.828125
|