 |
 |
Dear Mark, That makes perfect sense. Thanks for the additional information. Best regards, Stuart On 2013-11-26, at 6:57 PM, Mark Jenkinson <[log in to unmask]> wrote: > Hi, > > The "size" is the number of voxels. The "mass" is the sum of the values within the cluster (and so depends on cluster size and magnitude of the values within the cluster). > > All the best, > Mark > > > On 26 Nov 2013, at 18:07, Stuart Fogel <[log in to unmask]> wrote: > >> Dear Mark, >> >> Thanks for the explanation. This gives me a much clearer picture on how to advance with this set of analyses. >> >> I am however, not familiar with the distinction between the -c and -C cluster options, other than what is described in the FSL guide. It sounds like -c corrects for multiple comparisons cluster-wise for cluster "size", whereas -C uses cluster "mass". Are you able to clarify, or could you point me in the direction of a good resource for these details? >> >> Many thanks, >> Stuart >> >> On 2013-11-26, at 3:03 AM, Mark Jenkinson <[log in to unmask]> wrote: >> >>> Dear Stuart, >>> >>> It is fine to use another option for multiple comparison correction rather than TFCE. >>> However, you should not run cluster on the _tstat1 image, as this isn't telling you about the statistical significance. It is only the *corrp* files from randomise that you should use, as these contain p-values that are corrected for multiple comparisons. >>> >>> So you should just look at the corrp outputs from randomise when using -x or some of the other options (e.g. -c, -C). Note that you will need to choose arbitrary thresholds for many options (e.g. -c, -C) and there is no right or wrong answer to what threshold to choose, as it is an arbitrary choice. TFCE was developed to avoid the need for choosing such arbitrary thresholds (i.e. it is "Threshold Free …"). Also, I wouldn't consider a t-value of 4 to be that high in an imaging experiment, since the null distribution on its own, when there are lots of voxels, can generate such values. Therefore you may not see much with the -x option, as I suspect that you will need to exploit the clustered nature of the results to achieve significant (i.e. using -c or -C). >>> >>> All the best, >>> Mark >>> >>> >>> >>> On 25 Nov 2013, at 18:37, Stuart Fogel <[log in to unmask]> wrote: >>> >>>> Hello VBM experts and non-experts (like me), >>>> >>>> I am in the process of running a VBM analysis, comparing young vs. older adults, and I am interested in how grey matter intensity relates to a number of covariates of interest differentially in these two groups. >>>> >>>> The issue I am having is that when I run the TFCE procedure, very little/nothing remains significant, despite having t-values that are > 4 (and the pattern of results makes a lot of sense in terms of my hypotheses). There are however, many small clusters, but several larger clusters above t=3.5, for example. There are 13 subjects in one group and 15 in the other, so I suspect that I have enough statistical power for this analysis. >>>> >>>> Here's the parameters entered into the randomise command using TFCE: >>>> >>>> randomise -i GM_mod_merg_s3 -o GM_mod_merg_s3 -m GM_mask -d design.mat -t design.con -n 1000 -T -V >>>> >>>> I used 1000 permutations for this preliminary look at the data to save time (rather than 5000 or 10000 permutations for the finalized, "proper" analysis). >>>> >>>> Since the t-maps make sense, and the t-values are high, rather than using the TFCE option, I was thinking of: >>>> 1) generating the t-maps using the voxel-based thresholding method (option -x): >>>> randomise -i GM_mod_merg_s3 -o GM_mod_merg_s3 -m GM_mask -d design.mat -t design.con -x >>>> >>>> 2) then defining the clusters at a particular threshold (taken more-or-less from the FSLwiki guide, in this example t-threshold = 4): >>>> cluster -i GM_mod_merg_s3_tstat1.nii.gz -t 4 -o cluster_index --osize=cluster_size > cluster_info.txt >>>> >>>> 3) then using fslmaths to extract a binary mask of the clusters of interest (taken more or less from the FSLwiki guide, in this example, cluster ID #1): >>>> fslmaths -dt int cluster_index -thr 1 -uthr 1 -bin cluster_mask1 >>>> >>>> So my questions are: >>>> 1) Is this appropriate? >>>> 2) How do I determine what t-threshold to use (in step 2, above)? >>>> 3) Am I missing something, or is anything ill-advised? >>>> 4) Is there a better way? >>>> >>>> There was a similar post dating back to 2009, but I could not find a message with a solution to this problem. >>>> >>>> Thanks! >>>> Stuart >>>> >>>> == >>>> Stuart Fogel | PhD >>>> Research Scientist >>>> Brain & Mind Institute | Department of Psychology >>>> Western University | London | Ontario | Canada >>>> www.BMIsleeplab.uwo.ca >>>> ==