Using REST AlphaSim for whole brain and ROI analysis

Submitted by Karl_1 on Tue, 09/01/2015 - 09:07

Dear developers and users,

I have a dataset with which I want to perform both a whole brain- and ROI analysis and I want to estimate the minimal cluster size using REST AlphaSim. As I understand it, it is correct to estimate cluster size for both analyses separately. For the calculation of the whole brain cluster size, I used the smoothness estimated by SPM (found in SPM.xVol.FWHM from the group analysis), rmm was set as 5 and i used an uncorrected p-threshold of .001. As mask I defined the brain mask from the group analysis. This yields a minimal cluster size of 40 for a corrected alpha value of .05. For the ROI-analysis, I used the same estimated smoothness, the same rmm (although I read in an AFNI forum that rmm should be defined as 1000 when employing relatively small ROIS, my ROI has 1554 voxels) and also an uncorrected p-threshold of .001 (N.B.: the ROI analysis corresponds to a valid a-priori assumption). As mask I used the ROI. This leads to a very small minimal cluster size (even a cluster size 1 is significant at an alpha level of .01) which seems unreasonable to me.  When defining the rmm as 1000 I receive even lower values. 

I looked through the forums and tried a few things, but to no avail. F.ex., when I change the uncorrected p-value to .05 I receive unreasonable high values for both analyses. Furthermore, I tried to estimate the smoothness for the ROI analysis by using the results of an additional 2-level analysis with my ROI as an explicit mask, but the obtained values lead to an even lower cluster threshold. I also tried different ways of estimating the smoothness of the data (using rest_Smoothest.m), using T-maps or residuals as input. Although I observe vastly differing results, the p-values for the ROI analysis are still too small. 

Am I doing anything wrong? Is there any other way to correctly estimate the minimal cluster size for ROI analyses?

Any help/suggestion is appreciated.

Best regards,

Joe

 

Forums:

YAN Chao-Gan

Wed, 09/02/2015 - 07:29

Hi Joe,

p < 0.001 is a pretty strict threshold. When the multiple comparison was performed within a small mask, I expect very small cluster size is needed for such a strict threshold.

To validate the cluster size, you can try 3dClustSim in AFNI, or try GRF correction to see what's the cluster size they computed.

Best,

Chao-Gan

Karl_1

Fri, 09/04/2015 - 08:10

Hi Chao-Gan

Thank you for your reply. 

As I understand it, the "P threshold" value corresponds to the p-value used as an uncorrected cluster-defining thershold, is that correct? Furthermore, when I use a more lenient threshold, f.ex. .05, I receive extremely high minimal cluster sizes for both analyses. These cluster sizes are unrealistic and lead to a "null-finding" in my study.

When I use GRF correction, I receive similar values than with Alphasim. F.ex. the minimal cluster size for my ROI-mask when I use a Voxel-level p of .001 and a Cluster-level p of .05 is 1. When I increase the Cluster-level p to .001, I receive a minimal size of 279, which is extremly big for my ROI and therefore unrealistic. 

I would by grateful for any ideas on how I could proceed in this case...

Thanks a lot!

Joe

YAN Chao-Gan

Mon, 09/07/2015 - 12:17

Hi Joe,

Sometimes if you give a too strict individual p level, the program fails to find the cluster size and report cluster size of 1 (or probably it's real that cluster size 1 is significant under certain conditions). What's the cluster size if you set individual level p at 0.005 or 0.01? What's the estimated smoothness? Cluster-level p at 0.001 is very strict, it's possible to get large cluster size.

Best,

Chao-Gan