Print

Print
Shashwath, & co,

Indeed, this statistician finds nothing wrong with voxel-wise FDR and notably, to my knowledge, cluster-wise and peak-wise FDR hasn't been implemented in any other packages.  The voxel-, cluster- and peak-wise FDR are all distinct methods with different strengths and weaknesses.  

One weakness already mentioned is that peak-wise FDR depends on the 'feature defining threshold'.  Another not mentioned is that, in my experience, there is little additional power gained in moving from cluster-wise FWE to cluster-wise FDR, thus defeating the purpose of FDR (which is to obtain greater power while still controlling false positives in some formal manner).

The strength of voxel-wise FDR is the (generally) great power gains over voxel-wise FWE.  Its weakness is that it is a voxel-wise and not 'cluster aware', and so risk of false positives (as an expected proportion of number of detected voxels) cannot be linked to any particular cluster.

So, my take on it is that all are valid and suitable methods methods as long as they are appropriately interpreted. 

-Tom


On Mon, Jul 13, 2015 at 7:37 PM, Shashwath Meda <[log in to unmask]> wrote:
Hi Elena :) - Yes the field seems to be shifting more toward cluster based estimates more recently (more so after the Chumbley papers that Helmut refers to). But again the answer might be different based on which statistician you ask.  I would also take a look at the TFCE correction employed in FSL which does not bias one towards the initial cluster forming threshold. In my experience it also improves sensitivity towards finding a significant result. Christian Gaser has a TFCE toolbox for SPM that you might find useful as well.

Best
Shashwath



On Mon, Jul 13, 2015 at 12:36 PM, Elena Ivleva <[log in to unmask]> wrote:
Dear Experts,

I have a question re: FDR correction in SPM.

I have been using voxel-wise FDR correction: i.e, using 'Result' function, I choose FDR for 'p value adjustment to control', then indicate a specific p value (e.g., 0.01), and then choose a minimum extent cluster size: e.g., 50 voxels.

Now, I am being asked by a reviewer to use cluster-wise FDR correction. Am I understanding correctly that to use, for example, 0.01 FDR cluster-wise correction, I need 1) display this between-group contrast at 0.01 uncorrected, 2) get the minimum FDR-corrected cluster size from the spm output, i.e., 'FDRc' value, 3) re-run this contrast using 0.01, FDR, and plug in the suggested cluster size (=FDRc)?

Is the first approach incorrect in principal, or is it just an alternative way to look at the data?

Please advise, thanks,
Elena       



UT Southwestern

Medical Center

The future of medicine, today.



--
Best,

Shashwath



--
__________________________________________________________
Thomas Nichols, PhD
Professor, Head of Neuroimaging Statistics
Department of Statistics & Warwick Manufacturing Group
University of Warwick, Coventry  CV4 7AL, United Kingdom

Email: [log in to unmask]
Tel, Stats: +44 24761 51086, WMG: +44 24761 50752
Fx,  +44 24 7652 4532