 |
 |
Dear Chris, I have updated the TFCE toolbox that now allows to select an additional mask image for small volume correction. Please keep in mind that the mask image has the same dimensions such as your data. Please find the new versions here or simply use the update function: http://dbm.neuro.uni-jena.de/tfce/ Best, Christian On Sun, 29 Sep 2013 22:37:49 +0100, Christian Gaser <[log in to unmask]> wrote: >Dear Chris, > >On Wed, 25 Sep 2013 22:36:34 -0600, Chris Madan <[log in to unmask]> wrote: > >>Thanks Christian! >> >>Two more questions, if you don't mind: >> >>(1) If a cluster is significant at p<.05-FWE, without TFCE, is it fair to >>assume that it should still be significant with the TFCE? From my >>understanding of the TFCE, anything that is already significant with TFCE >>should still be significant, but previously subthreshold, but broad, >>clusters can be 'enhanced' by the TFCE to also be significant (e.g., Smith >>& Nichols, 2009, Fig 1, seems to show this). However, I have a contrast >>that has some suprathreshold clusters at p<.05-FWE (no TFCE), but with the >>TFCE nothing is significant at even p<.001-uncorrected. Does this sound >>plausible...? (It may be of relevance, I did not scan the whole brain in >>this study, but only a slab.) >If you use very low initial voxel thresholds (e.g. P<0.05 uncorrected) it often happens that you obtain a few very large clusters that are significant at the cluster level after correction for multiple comparisons. In that case TFCE may not result in any significant results, because the initial voxel threshold is too low. > >> >>(2) Is it possible to do the TFCE with a small volume correction? I >>understand that this may not be currently implemented since the TFCE >>distribution is based on the scanned area, and this distribution would then >>need to only be based on the SVC area. >You can apply SVC if you replace your mask.img/hdr file with a new mask defining your SVC-ROI. I will try to implement this in the new TFCE toolbox release. > >Best, > >Christian > >> >>~ Chris >> >> >>On Wed, Sep 25, 2013 at 4:06 PM, Christian Gaser < >>[log in to unmask]> wrote: >> >>> Dear Chris, >>> >>> On Wed, 25 Sep 2013 12:44:28 -0600, Chris Madan <[log in to unmask]> >>> wrote: >>> >>> >Thanks for reply, Christian! >>> > >>> >To be clear, the TFCE toolbox is basically doing two different procedures >>> >simultaneously then, the TFCE correction and a nonparametric T-test, >>> right? >>> >>> Yes, both tests are non-parametric and the sample distribution is >>> estimated using a permutation scheme. However, non-parametric t-test is >>> especially useful for small samples where the degrees of freedom is small >>> (where violations from Gaussian distribution have large impact on validity >>> of the statistic). For large samples with sufficient degrees of freedom >>> (and/or large smoothness filter) the differences between parametric (GLM) >>> and non-parametric tests are rather small. See: >>> http://www.ncbi.nlm.nih.gov/pubmed/14599004 >>> >>> In contrast, the TFCE statistic has the advantage of combining voxel and >>> cluster statistic regardless the degrees of freedom. Because the TFCE >>> distribution is not known a permutation scheme is necessary to compute the >>> sample distribution. There is rather an upper limit for sample size because >>> of computational and memory demands. >>> >>> > >>> >If I do use the TFCE statistic, should I also be using an FDR or FWE >>> >correction, or is that not needed since the toolbox is already combining >>> >voxel and cluster statistics? >>> The rules for applying a correction for multiple comparisons are the same >>> as for the GLM. Always use a correction (both FWE or FDR are valid). There >>> are some rare circumstances where a correction for multiple comparisons can >>> be skipped (e.g. if you have a clear and convincing anatomical hypothesis >>> about the expected effects). However, this should be rather the exception... >>> >>> Best, >>> >>> Christian >>> >>> > >>> >Thanks, >>> > >>> >~ Chris >>> > >>> > >>> >On Wed, Sep 25, 2013 at 8:23 AM, Christian Gaser < >>> >[log in to unmask]> wrote: >>> > >>> >> Dear Chris, >>> >> >>> >> the TFCE toolbox allows to estimate both TFCE statistic as well as T >>> >> statistic. The latter is the non-parametric approach to GLM in SPM and >>> >> should be close to the SPM results (the more degrees of freedom the >>> closer). >>> >> I have offered this options to have an alternative to SnPM. However, >>> TFCE >>> >> has some advantages because you can combine both voxel and cluster >>> >> statistic. Maybe I should simply remove that option for the T statistic >>> in >>> >> the TFCE toolbox to prevent further confusion... >>> >> >>> >> Best, >>> >> >>> >> Christian >>> >> >>> >> On Tue, 17 Sep 2013 09:12:49 -0600, Chris Madan <[log in to unmask]> >>> >> wrote: >>> >> >>> >> >Hi all, >>> >> > >>> >> >I am currently trying out Christian Gaser's TFCE toolbox, and I have a >>> >> >fairly basic question about interpreting results outputted from it. >>> >> > >>> >> >When going to view the TFCE results, one of the prompts is "type of >>> >> >statistic: TFCE or T?". How are these options different? The T option >>> >> >definitely produces images that are different than those from the >>> >> 'default' >>> >> >SPM results menu, so this clearly is not just showing the 'original', >>> >> >non-TFCE results. >>> >> > >>> >> >Thanks! >>> >> > >>> >> >~ Chris >>> >> > >>> >> >>> >> >>> >> >>> > >>> >>> >>