How to improve the statisitcal significance?

AFNI Message Board
1

Dear all,

    I want to find the difference of voxel-wise connectivity strength between group1 and group2,with 58 subjects in group1 and 24 subjects in group2.

    To do this, i tried 3dttest++ with covariates, 3dttest, and 3dMEMA with covariates, with all the findings failed to pass the cluter size threshold, with p equals 0.001, 0.01 and 0.02 in turn.

    i find a concept called small volume correction by accident, which can improve the significance when we are focusing on a specified region according to prior assumption. I used grey matter mask at the beginning, with subcortical areas followed, coming out smaller threshold, but the cluster size in the result has also been reduced, which means the findings still failed to pass the cluter size threshold.

    However, when i try to focus on a smaller region, for example, hippocampus, it turns out that "mask has only 67 nonzero voxels; minimum allowed is 128.", with recommendation of -OKsmallmask in 3dCluterSim. while it makes me confused that it says "finding results, but your favorite cluster is too small to survive thresholding, so you post-hoc put a small mask down in that region. DON'T DO THIS!" in the help of -OKsmallmask.

    Whether it's a proper way to focus on a small region when the findings failed to pass the cluster size threshold? And do anyone has some advice on how to check the data or getting it cross the threshold?

Best wishes,
Peng

2

Hi, Peng-

Well, I don’t think there is a way to improve the significance. Basically, you should have one test or comparison in mind, and try it once, and the results you get are what you get. Most people set up their tests to be within a whole brain or gray matter tissue mask, and that’s about that.

Something you could check is to make sure that your processing has actually gone well-- was alignment successful for each subject, for example? If you are doing a task-based FMRI study, did you put in the correct stimulus timings? Are you sure you have correctly labelled your subjects and their covariate values? Things like that.

Otherwise, your results are your results. Perhaps gather more subjects for more power?

–pt

3

Dear Paul,

Great thanks for your reply, I will try it again!

Best wishes,
Peng

4

Hey,

I would just like to mention that adding more participants and running the same analysis can inflate your false positive rate and you should therefore properly control your type 1 error if you are planning to do this.

There are quite a few papers out about this but if you wanna start with lighter read you can check these posts:

Hope this helps.

Remi