 |
 |
Hi Kavous, The intercept is a column full of ones. All the best, Anderson On 11 August 2017 at 02:12, Kavous Salehzadeh <[log in to unmask]> wrote: > Thanks a lot Anderson, > > I tried in your way with and without PALM. Results are different from > mine. So I'll back and redo my FSL analyses. If I find any point I will > update! > > Just could you please clarify for me what do you mean by intercept here? > > Sincerely, > > Kavous > > ---------- > > *PhD Candidate* > > *Center for Human Engaged Computing <http://xrenlab.com/>* > > *782-8502**, Kochi University of Technology, Kami City, Japan* > > > *Skype: kavus.salehzadeh* > > On Thu, Aug 10, 2017 at 10:25 PM, Anderson M. Winkler < > [log in to unmask]> wrote: > >> Hi Kavous, >> >> I'm not sure I'll be able to look carefully to check the proposed >> solution (it seems correct on a first glance, tough I'm impressed with >> these huge scatter plot differences). Here is a different approach: to >> residualise one variable A with respect to a set of others B = [var1 var2 >> ...], compute: >> >> % Residual forming matrix: >> R = eye(size(B,1)) - B*pinv(B); >> >> then the residualised variable à is simply: à = R*A. >> >> In your case, B = [Age Sex Intercept] and A1 = GMVoL and A2 = Covariate >> (you'd do the procedure separately for each). This would help in making the >> scatter plot, but this isn't good for testing (it leads to invalid results). >> >> Since you have PALM installed, you can benefit from the model >> partitioning that appeared in Beckmann et al (2001). Say M = [Covariate Age >> Sex Intercept], and you'd test Covariate with a contrast C = [1 0 0 0]'. >> Then do: >> >> % Partion the model: >> [X,Z,eC] = palm_partition(M,C,'beckmann'); >> >> X is now the EV of interest, already residualised, and Z is the nuisance >> already orthogonal to X. Then compute the residual forming matrix as above: >> R = eye(size(Z,1)) - Z*pinv(Z), and residualise the dependent variable: Y = >> Rz*GMVoL. >> >> Finally plot with scatter(X,Y), and check the correlation with corr(X,Y). >> >> Hope this helps. >> >> All the best, >> >> Anderson >> >> >> On 10 August 2017 at 01:00, Kavous Salehzadeh <[log in to unmask] >> > wrote: >> >>> It seems fMRIB ML has a size limit. So I uploaded them here: >>> https://drive.google.com/drive/folders/0B3NU_Dg9ZBP0Z0hLVkJO >>> Vk5wQms?usp=sharing >>> >>> Sincerely, >>> >>> Kavous >>> >>> ---------- >>> >>> *PhD Candidate* >>> >>> *Center for Human Engaged Computing <http://xrenlab.com/>* >>> >>> *782-8502**, Kochi University of Technology, Kami City, Japan* >>> >>> >>> *Skype: kavus.salehzadeh* >>> >>> On Thu, Aug 10, 2017 at 10:45 AM, Kavous Salehzadeh < >>> [log in to unmask]> wrote: >>> >>>> Hi Anderson, >>>> >>>> Thank you very much. I really appreciate your help. >>>> >>>> I have almost confidence about the steps which I've done on FSL. >>>> However, I did not regress out group membership before correlation. >>>> >>>> Frankly, I did not have any idea why and how should I regress out group >>>> membership. So, I followed a discussion here: >>>> https://stats.stackexchange.com/questions/117840/how-to-regr >>>> ess-out-some-variables >>>> And applied the same method by regressing out the effect of age and >>>> gender on my data. >>>> >>>> My data structure is: DATA=[Age Gender GMVoL Covariate] >>>> >>>> So I first calculated correlation: [r,p] = corr(DATA) >>>> And then did as mentioned in discussion: >>>> %% >>>> f1=r(:,1); >>>> R1=f1*f1'; >>>> sub_R1=r-R1; >>>> f2=sub_R1(:,2); >>>> f21=f2/(sqrt(f2)); >>>> R21=f21*f21'; >>>> R12=sub_R1-R21; >>>> %% >>>> Finally I applied residuals matrix (R12) to my data: new_DATA=DATA*R12; >>>> And run new correlation: [new_r,new_p] = corr(new_DATA); >>>> >>>> After all, I got a significant correlation. Please see attached scatter >>>> plots before and after this action. >>>> >>>> Could you please let me know whether I did right? >>>> >>>> Thanks again for your support. >>>> >>>> >>>> Sincerely, >>>> >>>> Kavous >>>> >>>> ---------- >>>> >>>> *PhD Candidate* >>>> >>>> *Center for Human Engaged Computing <http://xrenlab.com/>* >>>> >>>> *782-8502**, Kochi University of Technology, Kami City, Japan* >>>> >>>> >>>> *Skype: kavus.salehzadeh* >>>> >>>> On Wed, Aug 9, 2017 at 10:21 PM, Anderson M. Winkler < >>>> [log in to unmask]> wrote: >>>> >>>>> Hi Kavous, >>>>> >>>>> TFCE benefits from voxels that may be outside the cluster, from the >>>>> "support region", whereas just getting the data from within the cluster >>>>> ignores that. >>>>> >>>>> Even so, I'm surprised that there were no signs of correlation. Are >>>>> you sure there were no missteps along the way? Are the results of the >>>>> correlation completely off? >>>>> >>>>> Also, remember you need to regress out group membership from X and Y >>>>> before doing the correlation in Matlab. >>>>> >>>>> All the best, >>>>> >>>>> Anderson >>>>> >>>>> >>>>> On 6 August 2017 at 09:47, Kavous Salehzadeh Niksirat < >>>>> [log in to unmask]> wrote: >>>>> >>>>>> Dear FSLers, >>>>>> >>>>>> I used FSLVBM for a single-group average with a behavioral measure >>>>>> for over 600 subjects, exactly as explained here: >>>>>> >>>>>> https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/GLM#Single-Group_Aver >>>>>> age_with_Additional_Covariate >>>>>> >>>>>> Fortunately, I got a significant result (FWE-corrected). I would like >>>>>> to report and visualize my results. As I read from the previous >>>>>> discussions, I used cluster to report it: >>>>>> >> cluster -i tmp_tfce_corrp_tstat4 -t 0.95 --scalarname="1-p" >>>>>> --minclustersize --mm > my_cluster.txt >>>>>> >>>>>> I found only one cluster. Later I masked my cluster as follows: >>>>>> >> fslmaths tmp_tfce_corrp_tstat4 -thr 0.95 -bin >>>>>> mask_tmp_tfce_corrp_tstat4 >>>>>> >>>>>> and extracted its volume by this: >>>>>> >> fslstats mask_tmp_tfce_corrp_tstat4 -V >>>>>> >>>>>> I used fslmeants to extract mean GM volume for each subject and also >>>>>> eigen variates: >>>>>> >> fslmeants -i GM_mod_merg_s3 -m mask_tmp_tfce_corrp_tstat4 -o >>>>>> mean_GM_volume.txt >>>>>> >> fslmeants -i GM_mod_merg_s3 -m mask_tmp_tfce_corrp_tstat4 -o >>>>>> eig_GM_volume.txt --eig >>>>>> >>>>>> I multiplied volume and each subject's GM_volume. However >>>>>> surprisingly when I tried to compute correlation in MATLAB between >>>>>> extracted GM results and my behavioral data, I could not see any >>>>>> significant correlation. I tried corr(X,Y) either between mean_GM and >>>>>> behavioral data or eig_GM and behavioral data. Both results showed no >>>>>> significant correlation. >>>>>> >>>>>> Could you please help me to understand what I am doing wrong here? >>>>>> >>>>>> Sincerely, >>>>>> Kavous >>>>>> >>>>> >>>>> >>>> >>> >> >