 |
 |
Dear Nichols, I'm sorry for the delay response. Thanks a lot for your answers, they were extremely useful. Yes, you understand correctly all my doubts and they’re clear for me now. I have a new doubt, sorry!! In the PDF report, I sent you two different results/reports. The first one (the one that you were talking about and attached as picture in your last respond) showed a big cluster FEW-corrected at cluster level, using cluster size inference. But, the last page of this report showed a different result (see the old report that I sent you and/or the one that I’m attaching here). This second result is showing significant clusters at different locations (FDR-corrected). To obtain this second result (we were using the snpm_pp function, instead the batch window) we followed the same steps as the first report until the program ask about ‘results for which img?’ Here we have two options: results for ‘T’ or results for ‘P’ and this time we selected results for ‘P’ (for the first one, we picked ‘T’). In this case the program only gives us the option to used FDR correction. Then, we were assuming that for this second report we were correcting for the whole brain using FDR because the help says that the FDR report is voxel-wise FDR-corrected P-values. My doubt now is about when we picked FDR correction (that is under the ‘P’ img option when we are looking the results) instead of FWE (under ‘T’ img option). Since we selected first in the batch cluster size inference that’s means that the voxel-wise FDR correction option here doesn’t have any sense? Many thanks in advance and I apologize for the delay. Best regards, Lorna. ----- Original Message ----- From: "Thomas Nichols" <[log in to unmask]> To: "Lorna Garcia Penton" <[log in to unmask]> Cc: [log in to unmask] Sent: Wednesday, May 21, 2014 4:12:06 PM Subject: Re: [SPM] FDR results report in SnPM Dear Lorna, Thanks for uploading that information. Tell me if I understand you correctly... Your SnPM results page reported two things that troubled you (first attachment): - The P_FDR-corr values were all the same, 0.0427. -The P_uncorr values were all the same (but different!), 0.0002. Both of these are fine, and both arise from the discrete nature of the permutation P-values. As you performed 5000 permutations, the smallest possible P-value is 1/5000=0.0002, and all of the statistic values in the result had this maximally-significant P-value. This discreteness then also creeps into the FDR corrected P-values (see the FDR plot in the results; 2nd attachment). FDR P-values can have repeated values even if the original uncorrected P-values are continuous. But here, the, the flat line in the FDR plot means that every single one of those voxels (which we know are the minimal, 0.0002 P-value) must have the same FDR P-value. Which happens to be 0.0427. ANYWAY, I'm glad you're looking carefully at your output, but realise that since you selected cluster size inference, the *most* important number is the significant cluster with a P_FWE-cor of 0.0324. That significant blob is thus you're main inferential result, and not the voxel-wise (FWE or FDR) P-values -Tom On Wed, May 21, 2014 at 11:12 AM, Thomas Nichols < [log in to unmask] > wrote: Dear Lorna, Even with parametric (i.e. SPM) inference, FDR corrected P-values can be discrete, with different voxels having the same FDR corrected P. And, of course, in SnPM, with permutation, P-values are always discrete and multiples of 1/NumberOfPermutations. So I can't tell if you've detected a bug or just experiencing typical discreteness. Can you upload a PDF/PS or screenshot of the results, and the SnPM.mat and SnPMcfg.mat file here: https://files.warwick.ac.uk/tenichols/sendto ? -Tom On Fri, May 16, 2014 at 2:23 PM, Lorna Garcia Penton < [log in to unmask] > wrote: Dear SnPM experts, We are using voxel-wise FDR inferences in SnPM () but the FDR report showed the same p(FDR-corr) and p(unc) values at voxel-level for all the significant clusters/coordinates. We performed a 2 group comparison (is a VBM study) using a Two Sample T-test (one scan per subject) with no smoothing variance and 5000 permutations. We want to understand why the same p(FDR) and p(unc) values are reported after thresholding with p(FDR)<0.05 at voxel-wise. We appreciate any help. Best Regards, Lorna. -- __________________________________________________________ Thomas Nichols, PhD Principal Research Fellow, Head of Neuroimaging Statistics Department of Statistics & Warwick Manufacturing Group University of Warwick, Coventry CV4 7AL, United Kingdom Web: http://warwick.ac.uk/tenichols Email: [log in to unmask] Phone, Stats: +44 24761 51086 , WMG: +44 24761 50752 Fax: +44 24 7652 4532 -- __________________________________________________________ Thomas Nichols, PhD Principal Research Fellow, Head of Neuroimaging Statistics Department of Statistics & Warwick Manufacturing Group University of Warwick, Coventry CV4 7AL, United Kingdom Web: http://warwick.ac.uk/tenichols Email: [log in to unmask] Phone, Stats: +44 24761 51086, WMG: +44 24761 50752 Fax: +44 24 7652 4532 -- Lorna Garcia Penton PhD Student BCBL www.bcbl.eu Legal disclaimer/Aviso legal/Lege-oharra: www.bcbl.eu/legal-disclaimer