help > RE: first level threshold
Feb 6, 2019  11:02 AM | Bastien Guerin - Massachusetts General Hospital
RE: first level threshold
Hello -- this is a follow up on the post below. I have used the Matlab code provided by Alfonso to compute FDR-corrected p-value maps on individual subjects. However, they look completely different from the CONN GUI first-level analyses tab display (see screenshot attached: LEFT: 1-tail FDR-corrected p-value from Alfonso's Matlab code previously posted below, MIDDLE: 2-tail FDR-corrected p-value map from Matlab code posted below, RIGHT: CONN GUI first-level display).

This does not inspire confidence and makes me think I have messed up somewhere. Shouldn't the standalone Matlab code and the CONN first-level GUI agree?

Thanks -- Bastien

Originally posted by Alfonso Nieto-Castanon:
Hi Sarah,

One simple general way I could recommend to obtain seed-to-voxel first-level connectivity maps corrected for multiple comparisons would be to use Fisher approximation to the standard error of correlation coefficient estimates to obtain first uncorrected p-values for each voxel, and then apply an FDR-correction to obtain FDR-corrected p-values. The simplest way to approximate the standard error of the seed-to-voxel correlations would be assuming non-colored noise over the (typically small) frequency band defined during the preprocessing step (this is the same approach used by the previous version of the toolbox when using fixed-effect analyses; alternatively you could also use Welch-Satterthwaite approximation for a more precise result but that would require estimating the spectra of the residuals first).

In any way, if you have already run the setup/preprocessing/first-level analyses (and obtained in this way fisher-transformed BETA_Subject*_Condition*_Source*.nii files; one map for each source/condition/subject), you could easily transform these files to the associated uncorrected and FDR-corrected p-values (following the approach described above) using the following script:


nsub=1;       % subject number
ncondition=1; % condition number
nroi=1;       % source/seed number
nanalysis=1;  % analysis number

global CONN_x;
filepath=CONN_x.folders.preprocessing;
filepathresults=fullfile(CONN_x.folders.firstlevel,CONN_x.Analyses(nanalysis).name);
filename=fullfile(filepath,['DATA_Subject',num2str(nsub,'%03d'),'_Condition',num2str(ncondition,'%03d'),'.mat']);
Y=conn_vol(filename);
DOF=max(0,Y.size.Nt*(min(1/(2*CONN_x.Setup.RT),CONN_x.Preproc.filter(2))-max(0,CONN_x.Preproc.filter(1)))*(2*CONN_x.Setup.RT)+1);
filename=fullfile(filepathresults,['BETA_Subject',num2str(nsub),'_Condition',num2str(ncondition,'%03d'),'_Source',num2str(nroi,'%03d'),'.nii']);
a=spm_vol(filename);
Z=spm_read_vols(a);          % fisher-transformed correlation values
z=Z*max(0,sqrt(DOF-3));    % z-scores
p=spm_Ncdf(z);                  % p-values from normal distribution
p=2*min(p,1-p);                 % note: two-sided tests (remove this line for one-sided tests; i.e. only positive correlations)
P=p;
P(:)=conn_fdr(p(:));            % FDR-corrected p-values
a.fname=fullfile(filepathresults,['p_Subject',num2str(nsub),'_Condition',num2str(ncondition,'%03d'),'_Source',num2str(nroi,'%03d'),'.nii']);
spm_write_vol(a,p);
a.fname=fullfile(filepathresults,['pFDR_Subject',num2str(nsub),'_Condition',num2str(ncondition,'%03d'),'_Source',num2str(nroi,'%03d'),'.nii']);
spm_write_vol(a,P);


This will create two files:

   p_Subject1_Condition1_Source1.nii 
   pFDR_Subject1_Condition1_Source1.nii

with the uncorrected and FDR-corrected p-values, respectively, of the seed-to-voxel maps for the first subject/condition/source (edit the first four lines to change the subject/condition/source numbers of interest)

Hope this helps, and let me know if this addresses your question. You might need to do something more complex if you want to look instead at more complex contrasts across conditions or sources, let me know.

Best
Alfonso

Originally posted by Sarah Buetof:
Hi Alfonso,
Thanks for your quick answer.
We are planning to use the first level r-maps in another analysis and they should be corrected for multiple comparisons.
The fixed-effect analysis in 12 does not work and I get following error message:

Step 1/2: Functional data second-level analyses
62.5% ??? Reference to non-existent field 'csources'.

Error in ==> conn_process at 1467
                    csources=SPMall(n1).connvols.csources;

Error in ==> conn_process at 18
        case 'results',         conn_process(14:15,varargin{:});

Error in ==> conn at 1601
                conn_process('results');

Error in ==> conn_menumanager at 112
                            feval(MM.MENU{n0}.callback{n1}{1},MM.MENU{n0}.callback{n1}{2:end});
??? Error while evaluating figure WindowButtonUpFcn


Somebody else already mentioned that error and you provided another conn_process.m file that should solve the problem. But that didn't work for me, too.
We could implement a threshold in the script for our further analysis, but therefore we need the p-values.
Do you have any suggestions how to deal with that issue?
Does it makes more sense to try to fix the problem with version 12 or to do the statitics with matlab with the 13-results? Since I am not very familiar with matlab yet, I would prefer the first way, if that provides the results I am looking for.

Thanks
Attachment: prob_CONN.png

Threaded View

TitleAuthorDate
Sarah Buetof Oct 18, 2011
Alfonso Nieto-Castanon Oct 18, 2011
Sarah Buetof Oct 19, 2011
Alfonso Nieto-Castanon Oct 23, 2011
Kaitlin Cassady Sep 17, 2019
Kaitlin Cassady Oct 4, 2019
RE: first level threshold
Bastien Guerin Feb 6, 2019
Matthew Heard Jan 22, 2019
Erin Mazerolle Jan 17, 2014
Sarah Buetof Oct 31, 2011