On 02/21/2017 02:26 PM, Shane Schofield wrote:

Hi,

I have 3 beginner questions about the use of mri_glmfit-sim

1. Under what type of circumstances should I run mri_glmfit-sim with

permutation simulations?

This is a little bit of an open question as there is a big debate going on now in neuroimaging as to how to correct for multiple comparisons. It was recently shown that Guassian assumptions in fMRI results in a very high false positive rate (FPRs), making permutation more attractive, at least at low voxel-wise thresholds. Thickness is not nearly as bad, but still usually have elevated FPRs. If you have continuous covariates, then you cannot use mri_glmfit-sim (but you can use a tool called palm).

2. The voxel threshold of 0.01 (cache 2) is used a lot. Is this too

lenient in cortical thickness comparisons?

It is not bad in terms of FPR. If the applied smoothing level is 6mm or above, you are probably at an FPR of 7-10% instead of 5%. If the cluster significance is .02 or better, then the FPR will be at 5% or better (based on my simulations).

3. About the sign of the test (abs, neg, pos): is a one-tailed sign

justified if I have an hypothesis?

Yes, if it is an a priori hypothesis.

If the uncorrected (mri_glmfit) test shows a significant difference in one direction but not the other, am I justified to use pos/neg (instead of abs) in the mri_glmfit-sim stage and explicitly state that in the manuscript?

This is a tricky question. In general, I would say no because you're not really controlling for false positives at the rate you think you are. Having said that, I've seen lots of publications where people test both postiive and negative and do not further correct for testing both signs (which probably elevates the FPRs by almost a factor of 2).

Thanks. Hope to learn more. Have a good week.

Cheers, Shane

---

Freesurfer mailing list Freesurfer@nmr.mgh.harvard.edu https://mail.nmr.mgh.harvard.edu/mailman/listinfo/freesurfer