Guassian-Random-Field-Correction

Submitted by Loose on

Hi, I want to apply GRF-correction on my data and am insecure what to type in the GRF-box in DPARSF:

I calculated a two-sample-t-test and now open the resulting t-test-image. I then click on GRF-correction don't add a mask, calculate the smoothness and set p=0.01 for voxel and p=0.05 for cluster. Then I press compute. Is that the correct procedure?
It works without adding a mask but I am wondering whether and especially why I should add a mask?
Furthermore, when I read into articles using GRF-correction, they often only report a z-value, for instance 2.3 and a p-value, e.g. 0.05. To which values are the papers referring since in the GRF-box in DPARSF I can't find a z-value.
Also, since I don't have a directed hypotheses and only expect an alteration in functional connectivity I want to do a two-tailed investigation. However, I am not sure whether the button "two-tailed" in the GRF-box is referring to the same matter or whether describes something else?

Best

Markus

 

YAN Chao-Gan

Tue, 08/30/2016 - 01:27

Permalink

Hi Markus,

When you click the checkbox "Two-Tailed", you will get a note:

'The "two-tailed" option is doing as following (e.g., if set VoxelPThreshold = 0.01 and cluster level p<0.05): 1) Get Group A > Group B, set Z>2.576 (two-tailed p <0.01) and cluster level p<0.025  2) Get Group A < Group B, set a Z<-2.576 (two-tailed p <0.01) and cluster level p<0.025.  3) Add 1 and 2 together, which could ensure the total p<0.05.'

As most studies using FSL doing two one-tailed tests, each is Z>2.3, p<0.05, for better comparison, you could input:

Two-Tailed: VoxelP: 0.0214, ClusterP: 0.1.

Though you have to keep in mind that recent Eklund paper (PNAS, 2016) notes p<0.001 for VoxelP should be used.

Dear Chao-Gan,

Thanks for the response. However, I thought that Eklund talked about the threshold for the cluster-p-value? At least Eklund always mentions a Cluster-Defining-Threshold (CDT) or does he mean the voxel-p-value with that?

So following Eklund, I would set the voxelP at 0.001 and the clusterP at 0.05, is that right?

Best
Markus

 

Hi Markus,

As one-tailed tests are somehow accepted by lots of studies, and you may want to use a "looser" threshold. You can click "two-tailed" and input 0.002 and 0.1. That's equivalent to two one-tailed tests: each is 0.001 and 0.05.

Best,

Chao-Gan

Hi all,

Is it the case that a negative Z value for cluster-level p is theoretically inappropriate because it is not possible to have a cluster size less than 0? This seems like it would justify having a one-tailed alpha for cluster-level p, but not voxel-level p. Is this correct?

Thanks in advance,

Felix

Hi Dr. Yan and DPABI experts,

Thanks for your response, I now understand the reasoning for this.

As a follow-up:

1. Is it appropriate to divide the cluster-level alpha (ie, Bonferroni correction), in order to control for having conducted separate mixed-effects analyses using multiple seed ROIs? A dissertation committee member pointed out GRF may not be sufficient at voxel-p<.01, cluster-p<.05, 2-tailed, because I have 10 different seeds (and therefore 10 different mixed-effects models). Can I address this by simply using voxel-p<.01, cluster-p<.005, 2-tailed, to allow for 10 separate seeds? What about the fact that each mixed-effects analysis yields 2 T-values and an F-value, does this mean I need to additionally divide alpha by 3?

2. To mask or not to mask: I was told I use too many masks! Previously, for postpro and standardization, I created and used GroupMask90%; for group stats and display (that is, for GRF smoothness estimation), I used aal because I had used aal seeds for rsFC postpro. Since then, I repeated all analyses using no mask for postpro group stats, and GroupMask90% for standardization (Mean reg with SD div, no GSR) and display. Unfortunately, doing it this way changed my findings considerably and now there's a lot of "activation" in WM regions. Now I am considering using (GroupMask90% * SPM's a priori GM mask), but this seems like cheating and I'm concerned it will disrupt standardization to do it this way. Can you please advise?

Can't give enough thanks for all the help I've been getting through the forums, and for the fact that DPABI even exists, I'd be lost without it! I look forward to the day that I am no longer a grad student and can contribute something substantial towards that mickle ; )

All the best,

Felix

Forums
DPABI/DPABISurf/DPARSF