 |
 |
Hi Anderson, Sorry to bring this thread up again however I still have some queries about FDR. Mainly, I have done FDR 3 atlas-derived masks: amygdala, V1 and STG. When I view each of the outputs individually at the FDR probability threshold all three of these regions are still highlighted. For example: when I look at the othresh output of FDR (with the minimum threshold set in fslview) which used the amygdala mask, areas in the amygdala, V1 and STG were still significant. Doesn’t FDR with a ROI mask only estimate FDR for that ROI? To confuse this even more, when I use a whole brain mask, the probability threshold is zero. Any insight is much appreciated. Kind regards, Isaac From: FSL - FMRIB's Software Library <[log in to unmask]<mailto:[log in to unmask]>> on behalf of "Anderson M. Winkler" <[log in to unmask]<mailto:[log in to unmask]>> Reply-To: FSL - FMRIB's Software Library <[log in to unmask]<mailto:[log in to unmask]>> Date: Wednesday, 6 July 2016 at 5:03 PM To: "[log in to unmask]<mailto:[log in to unmask]>" <[log in to unmask]<mailto:[log in to unmask]>> Subject: Re: [FSL] FDR questions Hi Isaac, Please, see below: On 5 July 2016 at 23:55, Isaac Unger <[log in to unmask]<mailto:[log in to unmask]>> wrote: Hi Anderson, Once again, thank you very much for your reply. Sorry to drag this thread out however I would like to make sure I am 100% correct in my understanding. 1. Regarding the first issue: Perhaps I should have mentioned this before; I am using FDR on the output of a melodic/dual regression pipeline. Given this background should I be using Bonferroni correction after FDR correction (FDR probability threshold)/(2* # of good ICs) or does the answer still remain no? The answer remains no. While it is ok to combine FWER-correction (e.g., across space, then across ICs), with FDR, things don't work like that. In fact, it's even simpler: you can still use FDR at 5% on each IC separately, and yet will at the same time be controlling the FDR overall at a level equal to or below 5%, with no need for Bonferroni. Alternatively, it's possible to correct all ICs at once with FDR, and the overall false discovery rate will be controlled at 5%, but then it will not be possible to tell what the amount of false discoveries will be on each IC separately. 2. Regarding the second issue: given your response, is it correct for me to say: even though I am using FDR on a small region it is still fundamentally FDR? Yes. Also could you please elaborate on what you mean by "as long as the selection of the ROI is completely independent of the results”. Do you mean that I should have hypothesised the ROIs that I specifically FDR correct for? Yes, or derived them from an atlas before doing the experiment/analysis, or from other studies or independent datasets. It is not ok to derive ROIs from the same data being analysed, then use that as a mask to correct for multiple testing. Cheers, Anderson Thank you in advance and kind regards, Isaac From: FSL - FMRIB's Software Library <[log in to unmask]<mailto:[log in to unmask]>> on behalf of "Anderson M. Winkler" <[log in to unmask]<mailto:[log in to unmask]>> Reply-To: FSL - FMRIB's Software Library <[log in to unmask]<mailto:[log in to unmask]>> Date: Tuesday, 5 July 2016 at 6:12 PM To: "[log in to unmask]<mailto:[log in to unmask]>" <[log in to unmask]<mailto:[log in to unmask]>> Subject: Re: [FSL] FDR questions Hi Isaac, Please, see below: On 4 July 2016 at 00:50, Isaac Unger <[log in to unmask]<mailto:[log in to unmask]>> wrote: Hi Anderson, Thank you very much for your reply; This has made FDR a lot clearer for me. If I could ask a few more questions on this topic: - Should FDR corrected values be further corrected via Bonferroni? No. - If I have an a priori hypothesis about a ROI (and use this region as a mask during FDR), can I say something along the lines of: “p-values of ROI “x", were FDR corrected for a small volume at q=0.05…” I don't like this description. It uses jargon that I find extraneous to FDR. Or “given we were interested in ROI “x”, voxels in this region were FDR corrected…” I would go with this one, as long as the selection of the ROI is completely independent of the results. All the best, Anderson Kind regards, Isaac From: FSL - FMRIB's Software Library <[log in to unmask]<mailto:[log in to unmask]>> on behalf of "Anderson M. Winkler" <[log in to unmask]<mailto:[log in to unmask]>> Reply-To: FSL - FMRIB's Software Library <[log in to unmask]<mailto:[log in to unmask]>> Date: Sunday, 3 July 2016 at 6:26 PM To: "[log in to unmask]<mailto:[log in to unmask]>" <[log in to unmask]<mailto:[log in to unmask]>> Subject: Re: [FSL] FDR questions Hi Isaac, Please, see below: On 1 July 2016 at 04:58, SUBSCRIBE FSL Isaac <[log in to unmask]<mailto:[log in to unmask]>> wrote: Dear FSL experts/community, I have a few questions regrading the operation of FDR: 1. Should I use FDR on p_tstat or corrp_tstat output of randomise? p_tstat. 2. If I specify a ROI mask with FDR am I still computing FDR or SVC? If you supply a mask, only the voxels within the mask will be used to find the threshold and will be corrected. The others will be zero out. In a sense this can be called "a" small volume correction (i.e., a correction for a smaller number of voxels), but it is not quite the same as SPM's SVC, which refers to both a correction for small volumes and also a correction that uses random field theory formulas that have been corrected to use lower dimensional terms that are usually negligible in whole-brain analyses, but that become relevant for ROIs that are small or with convoluted shapes. 3. The "Probability threshold" produced by FDR command is my "new" level of significance? In FDR we don't really speak of level of significance in the same way as we define p-values (be them uncorrected or FWER-corrected). The threshold is the value that, once used to threshold the image, will give a 100*q% of false positives. All the best, Anderson Many thanks in advance. Isaac