 |
 |
Hello again Anderson, Thank you for the many explanations you gave me! It seems that i am being thick headed, and can’t understand a particular point that i would like to clarify before commencing my research. I am sending you this to your email directly because i will include precise measures that i am using and prefer not to get this publicly! First i compared the brain scan of patients and control using ICV and age as covariate in TBSS. From that i got some significant brain regions. So when i do this the below, there are no 2 tests to be circular at least till # 2, is that correct? There is only the corrected tstat file that i got from TBSS from which i thresholded and then extracted the FA in those areas! So until here it is ok I’m guessing since there is no second statistical test here! >>> 1- fslmaths tbss_tfce_corrp_tstat1.nii.gz -thr 0.95 -bin sig_results_mask >>> 2- fslmeants -i all_FA_skeletonised -m sig_results_mask -o sig_FA.txt Now when i want to correlate with clinical data (# 3) here is when circular analysis might occur! Is that correct? >>> 3- Got the results and correlated with clinical data. I want to pose 2 scenarios: 1- If i have clinical data that is not in anyway linked to FA values until now! lets say behavioural test score. Even though behaviour is of course might be affected by white matter integrity, but this behavioural test has never been proved to be linked to FA values before, would it be circular to test if it is intact linked? 2- Ok so what i want to check here is whether the ages of the parents at birth of patients leads to their significant decrease in FA or MD etc.. Until now, there is no confirmation that a correlation exists (In the literature i mean)! Would that be considered circular? Thank you again and again, Walid > On Jan 30, 2017, at 10:12 AM, Anderson M. Winkler <[log in to unmask]> wrote: > > Hi Walid, > > On 27 January 2017 at 01:38, Walid Yassin <[log in to unmask] <mailto:[log in to unmask]>> wrote: > Hello, > > Thanks again for your comments! > > I’m a little confused now. Previously i used the significant regions resulting from my TBSS analysis as a mask (as shown below), but you said that it is not valid! > Now you are saying that it is fine to extract the values within significant regions? > > If you will do two tests, where the first is used to define ROIs that are subsequently used in a second analysis, then: > > - If the data in the two tests are not independent, the test is invalid and circular. > - If the data in the two tests are independent, then it is fine. > > The ROI can come from an atlas (as opposed to from a test) in which case it is also fine. > > > I would like to extract the FA in the significant regions i got instead of a ROI for further analysis with clinical data! > > Is the clinical data in any way correlated with the variables used to define the significant regions? If yes, the test is invalid. > > What is being circulatory here? The clinical data were not entered in the initial analysis in TBSS as covariates or something! > > The clinical data may not have entered but it may still be correlated. For example: you find the regions of FA difference between subjects with diabetes and controls, then run a second analysis in an ROI to see if FA is associated with fasting levels of blood sugar. Obviously blood sugar levels are correlated with the diagnosis (diabetes) such that although blood sugar levels didn't enter in the first test, the analysis is still circular. Sometimes the association is more insidious and it isn't that easy to perceive, so care must be taken. If unsure, use an independent dataset. > > Could you please just clarify this point so that i understand what is circular about the below script? and maybe guide me on how to make it un-circular using my significant voxel from TBSS as mask? > > > 1- i used the significant voxels from TBSS results > 2- i made a mask of them to extract the FA from > 3- i used the FA values extracted to correlated with clinical data that are used here for the first time! > > >>> >>> 3- Got the results and correlated with clinical data. > > > > The above will be circular if the clinical variable is associated with the variable used in the initial TBSS analysis. > > Hope this helps. > > All the best, > > Anderson > > > > Thank you in advance, > > > Walid > >> On Jan 26, 2017, at 6:48 PM, Anderson M. Winkler <[log in to unmask] <mailto:[log in to unmask]>> wrote: >> >> Hi Walid, >> >> It is common to extract the average values within an ROI (e.g., if there is a pre-defined hypothesis affecting such regions). It is also fine to extract values within a significant region to further study the association with another variable that is unrelated to the one used in the first test. >> >> It is not correct, however, to run a statistical test using data that is the same or correlated with the data used in an initial statistical test. The results are not valid. >> >> All the best, >> >> Anderson >> >> >> On 25 January 2017 at 10:27, Walid Yassin <[log in to unmask] <mailto:[log in to unmask]>> wrote: >> Hello Anderson, >> >> Thank you very much for answering all my questions! >> I had a good look through the forum and this is how I saw that others are extracting their FA values! So are the recommendations: to check which regions in the brain are your significant voxels in and then use those regions from an atlas in fslmeants as mask instead of the significant_fa_mask I used? >> >> Though after reading your comment, I really don't want to be double dipping. >> Is there another way that is more proper/trustworthy? >> >> Thank you in advance, >> >> Best Regards, >> >> >> Walid >> >> On Jan 25, 2017, at 17:47, Anderson M. Winkler <[log in to unmask] <mailto:[log in to unmask]>> wrote: >> >>> Hi Walid, >>> >>> Please see below: >>> >>> On 24 January 2017 at 00:27, Walid Yassin <[log in to unmask] <mailto:[log in to unmask]>> wrote: >>> Dear FSLers, >>> >>> >>>Sorry for double posting, but i didn't include a title for my post yesterday for archiving. >>> Also it wasn't forwarded. Please check below.<<< >>> >>> >>> I would like to ask you for a confirmation and some help please! >>> >>> >>> I- I did all TBSS steps between 2 groups and got the results: >>> >>> tbss_tstat1.nii.gz, tbss_tstat2.nii.gz, tbss_tfce_corrp_tstat1.nii.gz, tbss_tfce_corrp_tstat2.nii.gz, etc.. >>> >>> I want to check if there is a relationship between the significant areas that resulted and my clinical data so i did: >>> >>> 1- fslmaths tbss_tfce_corrp_tstat1.nii.gz -thr 0.95 -bin sig_results_mask >>> 2- fslmeants -i all_FA_skeletonised -m sig_results_mask -o sig_FA.txt >>> 3- Got the results and correlated with clinical data. >>> >>> Is this OK? Should i have used tbss_tstat1.nii.gz instead of tbss_tfce_corrp_tstat1.nii.gz ? >>> >>> >>> >>> This isn't ok I'm afraid. This is what has been known as "voodoo" correlations: it's a circular analysis that shouldn't be used. >>> >>> >>> II- I wanted to check which areas are my significant voxels in. So i did: >>> >>> 1- atlasquery -a "Juelich Histological Atlas" -m significant_results_mask.nii.gz >>> >>> Let’s say these are my only results: >>> >>> WM Optic radiation R:16.9093 >>> WM Superior longitudinal fascicle R:2.41935 >>> WM Uncinate fascicle R:0.792453 >>> >>> Now, if i want to do the same analysis as before, meaning if i want to check significant voxels in the "WM Optic radiation R” ROI only, and correlate those to my clinical data. should i do: >>> >>> Get the "WM Optic radiation R” ROI from the atlas. Let’s say i call it optic_rr.nii.gz So it will be: >>> >>> 1- fslmaths tbss_tfce_corrp_tstat1.nii.gz -thr 0.95 -bin sig_results_mask >>> 2- fslmeants -i optic_rr -m sig_results_mask -o sig_FA_oprr.txt >>> 3- Got the results and correlated with clinical data. >>> >>> -Is this correct? >>> -Also, which atlas is better to use? Should i try several white matter atlases and see which ones my voxels hit most? >>> >>> >>> For the reason above, don't do this analysis. >>> >>> However, if you were do to it, you could use the region from the atlas directly as the mask with the command fslmeants. >>> >>> >>> >>> III- Cluster >>> >>> I am wondering after i have done the above whether there is merit in doing cluster analysis! >>> >>> I already did the following to report cluster size: >>> >>> cluster -i tbss_tfce_corrp_tstat1.nii.gz -t 0.95 --mm > cluster_95_all >>> >>> and Got 8 clusters ranging from voxel# 1 to 750. >>> >>> Is doing further cluster analysis meaningful? such as z- or t- stats per cluster? >>> >>> Last question. Can clusters have labels about the regions the contain them? >>> For example cluster 8 is contains 750 voxels, can i know which regions in a JHU atlas are in this cluster? >>> >>> >>> This is fine. It will list the coordinates of the significant regions. For the region names, you can overlay on a structural image and compare with the actual brain anatomy, or use a tool such as the command autoaq (shipped with FSL). >>> >>> All the best, >>> >>> Anderson >>> >>> >>> >>> >>> >>> >>> Thank you in advance, i highly appreciate it. >>> >>> Walid >> > >