Fixelcfestats outputs

Dear all,

I have a question regarding the outputs of fixelcfestas. I apologize that this issue might be discussed several times, but I’m still confused about what I got so far.

I performed a longitudinal FBA and would like to know whether there is any change across time. I subtracted each time-point 1 image from the time-point 2, and smoothed the outcome images. My design matrix has 2 columns: intercept (1s) and a demeaned covariate. My contrast is 1 0 and -1 0, which I assume reflect global change for the whole sample. I also added a mask to exclude <50 fixels.

I did not find any significant results, however the fwe1-p values are the minimum (1/n), null_contributions are more than 3000, which seem wired. Please see the outputs of the contrast 1 0 as below (similar for -1 0).

abs_effect: -0.057~0.068
uncorrected1-p: 0 ~ 1.17e-38
fwe1-p: 0 ~ 0.0002
null_contributions: 0 ~ 3069

My question is that even if the global changes of my sample are not significant, they couldn’t be zero (the zstats are not trivial), then why 1-p values are so small? I encountered this type of problem every time when I looked at the significance of the global intercepts with demeaned covariates in the models. Which step that I could have done wrong? I would appreciate any help you could provide!

If there is an individual fixel with a value of ~3000 in the null_contributions file, that would be the first point of interest. See some explanation here. Even if your Z-statistics seem reasonable, there could be one fixel that is generating massive enhanced statistics, producing a null distribution with very large values and therefore making it impossible for any fixels in the empirical data to exceed that null distribution. The question is where that fixel is, and why it is repeatedly generating maximal values across such a large fraction of shuffles. The situation where I’ve observed this myself is when you have a fixel that does not have any streamlines connectivity, which is why as of 3.0.0 there is sometimes a warning issued about such fixels being excluded from testing. If you’ve already masked out such fixels, then maybe you’ve found some mechanism by which this problem can arise that I’ve not myself yet seen.

See what you can find by navigating through your raw data; if you can’t find anything I can have a look, but I’d need access to all of the inputs used for the fixelcfestats call.