Hi Frost,

First thing: make sure you are using the most recent version of FSL (or at least 5.0.8 I think). I remember vaguely that in some old version there was an issue affecting the way as F-stats were converted to z-stats before doing TFCE (only F-stats).

Regardless, what you've found can happen, i.e., nothing on F-tests even though there is "something" in the post hoc t-tests: the relationship between t and F statistics is lost when the spatial signal is considered, such as with TFCE. That classical relationship as we find in stats textbooks are not guaranteed to hold, and very often breaks.

You can safely bypass the F-tests altogether (i.e., don't do F-tests at all). If the idea of the F-test was to check for significance before doing the t-tests, you can instead use the option "-corrcon" in PALM, which will correct across the t-tests (with and without TFCE).

All the best,

Anderson

On 24 September 2017 at 22:51, C.P. Frost <[log in to unmask]> wrote:

Thanks Anderson, that was very helpful!

I ran the analysis with the changes you suggested and T-values were all much more believable. However, this time both the raw and corrected F-tests themselves had a max value of zero. My impression is that this suggests an error, rather than a believable ratio of between-group variance to population variance. However I checked the .mat file and it appeared to have the dummy variables for all three groups coded correctly. What's the most likely explanation?

Below are the TFCE-corrected T- and F-stats in question.

CTQ2_tfce_corrp_tstat1.nii.gz
0.000000 0.699000
CTQ2_tfce_corrp_tstat2.nii.gz
0.000000 0.289000
CTQ2_tfce_corrp_tstat3.nii.gz
0.000000 0.930000
CTQ2_tfce_corrp_fstat1.nii.gz
0.000000 0.000000

Thanks,
C.P. Frost