Hi NBS team, 

First of all, fantastic work on this and I'm glad to see its still being maintained! 

I have a question regarding the FDR threshold value. It was my understanding that the threshold value wasn't relevant for FDR. However, when running through commands (see below) I get an error when excluding a threshold value. As such, does the threshold value contribute to FDR? 

UI.method.ui='Run FDR';
UI.test.ui='t-test';
UI.size.ui='Extent';
UI.thresh.ui='';
UI.perms.ui='5000';
UI.alpha.ui='0.05';
UI.contrast.ui='[-1,1]';
UI.design.ui=designMatrix;
UI.exchange.ui='';
UI.matrices.ui=connectomes;
UI.node_coor.ui='COG.txt';
UI.node_label.ui='Atlas_label.mat';
% Run with the above UI
NBSrun(UI,[])

Thank you in advance for any guidance you can provide! 

All the best,
Lawrence

Hi Lawrence, 
this threshold (UI.thresh) is irrelevant to FDR. 

However the number of permutations (UI.perms) must be very substantially greater than 5000 for FDR. I suggest using 100,000, but 1 million would be more reliable. This can be slow to run. 

Under FDR, results with 5000 permutations are not reliable and should not be used. 

Andrew
Originally posted by Lawrence Binding:

Thank you for the quick reply. 
I think I'm going to range the permutations from 100k to 500k in steps of 50k to be as transparent as possible. 

Am I right to assume FDR words just as any other FDR algorithm? 

The reason I ask is I have patients preoperative and postoperative raw streamline numbers. I'm working out the percent lost for each patient between nodes and then feeding this into FDR. The idea is this makes loss relative to each patient and reduces multiple-comparisons problem from 380 to around 53. However, this means that the connectome is in negative numbers; should I take the absolute number from these or will FDR suitably handle this?

Thanks in advance.

Hi Lawrence, 

500k permutations will always provide more reliable p-value estimates than 100k. I suggest running twice with 500k and comparing the results, rather than increasing in increments of 50k. 

FDR is implemented in a non-parametric way so that the underlying statistical tests do not make assumptions about the distribution of the data. 

I would need to have more information about the negative numbers questions. In general I suggest to model the pre and post streamline counts. It seems taking absolute values is not ideal because this will fundamentally change a percent loss to a percent gain? 

Andrew
 

Originally posted by Lawrence Binding:

Thanks for getting back to me Andrew.

Running FDR with 500k permutations now, although the severs are fighting back with memory control haha!

for a quick example the raw numbers look something like this:

0 0 -34 0 
0 -2 0 -0.1
0 0 0 0
0 0 0 0
0 0 -4 0

Would FDR still handle this suitably? 

True, I was just wondering wether FDR would handle the matrix better if it was similar to how the original connectomes looked, i.e., with positive values instead of negative. 

Thanks again for the advice! 
Lawrence

Yes - I think that FDR is a reasonable choice, particularly for small networks. 


Originally posted by Lawrence Binding: