search

andersonwinkler / PALM

PALM: Permutation Analysis of Linear Models
69 stars 28 forks source link

very high TFCE z-statistics #30

Closed marinkakoenis closed 3 years ago

marinkakoenis commented 3 years ago

Dear Anderson, Palm-team

I'm running palm to test for group differences in an fMRI task. I use the -tfce1D option and I want to report the z-statistic. However, the TFCE z-stat is huge (7-11 digits), which I guess overlaps with the t-stats in the range of [100-1200] I found in earlier analyses (using default TFCE settings, not the 1D option). The voxel-wise statistic is in a more normal range [-3 5], but that's not what I base my significance on. Even more so, the voxelwise and cluster wise approaches show hardly any significance results, whereas the TFCE is significant all over the brain (after FWE correction).

Can you help me understand these test-statistics?

Thank you! Best, Marinka

andersonwinkler commented 3 years ago

Hi Marinka,

Could you give more details about why using -tfce1D? Also, what version of Matlab and/or Octave are you using?

All the best,

Anderson

marinkakoenis commented 3 years ago

Hi Anderson,

I thought -tfce 1D was the right way to go because the data is based on timeseries, but I realize now that the input is not timeseries themselves, but the COPE contrast. I'm using matlab 2018b.

Best, Marinka

marinkakoenis commented 3 years ago

Hi Anderson,

I've rerun my analyses with the default -T option (no extra tfce flags). The t and zstat TFCE maps are exactly the same and range from [0 ~1200] (with significance >~800). The voxelwise test statistic does change slightly when the -zstat option is used, and is in the range of [-5 5].

How do you advice to report the test statistic?

Many thanks. Best, Marinka

andersonwinkler commented 3 years ago

Hi Marinka,

The TFCE statistic can indeed have very high values, with or without -zstat. However, it isn't typical that everything becomes significant. Could you post the data somewhere and send me the link via email?

Thanks!

All the best,

Anderson

marinkakoenis commented 3 years ago

Hey Anderson,

With the default -T the results make more sense, they're not whole brain anymore. I'm sorry was my previous message was not clear. The -zstat option doesn't seem to do anything for the TFCE-tstat maps: both tstat and zstat maps are exactly the same.

Best, Marinka

andersonwinkler commented 3 years ago

Hi Marinka,

The spatial statistics are always based on z-stats, hence the option -zstats have no extra effects. The reason for this is to allow TFCE comparisons across contrasts that may have different degrees of freedom.

All the best,

Anderson

PS: I'm closing the issue, but happen to open again if needed.