Hi,

I am using NBS on a small group of patients and I wanted to try FDR analysis.
I wonder what represent the nbs.NBS.test_stat values when running a FDR analysis (ttest permutation 100000) .
Indeed, I am usually using the NBS analysis and so I am using the matrix as t-value (or f-value). But this time when I am extracting the value related to the significant links I am getting a range of value going from -2.2 to 5.8.

Did I misunderstood this matrix?
Am I doing something wrong?
Thanks a lot for your help.
Cc

Hi Cecile,

I don't think that it is possible to have significant effects in cases when the test statistic is negative.

It is impossible for the F-statistic to be negative.

You may need to provide further details if you would like more detailed advice.

Andrew


Originally posted by Cecile Bordier:

Hi,

Sorry I should have given more details. But yes, you comfort what I was thinking about negative values.

So more details. I have a group of 61 subjects with acquisition at 3 time points (6months, 36months and 60months respectively M6, M36 and M60).

So I use the fsl (feat) website to create my design matrix. Here is an example for 2 subjects:

 1  1 1 0
 1  1 0 1
-1  0 1 0
-1  0 0 1
 0 -1 1 0
 0 -1 0 1

With the two first columns coding the timepoints and the 2 last for the subjects
So one of my contrast to get M6-M60 is [1 2 0 0].

Basically,  I ran that design on NBS (with nbs stat, a t-test, 5000 permutations, block exchange 1:61,1:61,1:61 (for my 61 subjects)) and I got stat at 0.054. So close of significance that I decided to try the FDR analysis. I increased the number of permutation to 100000, changed the stat to FDR  (kept the exchange blocks is it right?) and ran it.
It took a long time and I got significant results (for my 3 contrasts M6-M60; M6-M36 and M36-M60).
So my idea to represent them was to multiply nbs.NBS.con_mat with nbs.NBS.test_stat and plot the connection which were not equal to zeros. But here I am getting positive and negative values (even low values) and that for the results of my three contrasts.

I think I must do something wrong, but I  have trouble figuring where.
Thanks a lot for your patience and help.

Cecile

Hi Cecile,

If you use the f-stat, it is impossible to get a negative test statistic.

I have some concerns with your design matrix. if you want test the contract M6 - M60, I would use the following design:

1 0 0 1
1 0 1 0
0 1 0 1
0 1 1 0
0 0 0 1
0 0 1 0

And use the contrast of: [1 0 0 0] or [-1 0 0 0]

The first column is for M6 and the second column is for M36. The last two columns are for subjects 1 and 2.

Andrew




Originally posted by Cecile Bordier:

Hi Andrew,

Thanks for your quick and clear answer.
I will try your design matrix (I get mine from the fslwiki website (https://fsl.fmrib.ox.ac.uk/fsl/fslwiki/G... in the feat details). I will run it and come back to you but it usually take around 48h to compute the FDR with 100000 permutations.

But do you think that can explain the results I am getting? I mean getting negative value doing a ttest with FDR? I am using ttest because I am interested in the side of the effect (M6>M60 for example).

Thank you so much for your help!
Cecile

Hi Andrew,

After, few days of running the FDR analysis with the model you advised me to use, I am finding the same kind of results than before.
More annoyingly, I am also getting negative value of f-test when I am looking at the significant links.

I really do not get it! I checked my subject matrix, they are sparse and with nothing special (no NaN or Inf).
Any idea/ suggestion?

Thanks a lot for your patience.
Cecile

Hi Cecile,

It is not possible to get a negative F statistic.

If you feel comfortable doing so, I suggest that you send me your data, design matrix and contrasts. I can then try to replicate the error locally.

My email is: andrewzalesky@gmail.com


Andrew


Originally posted by Bordier Cecile:

Hi,


Thanks a lot for your help!!!
Sorry it took me quite a bit of time to re-run my test.
So just in case people are curious. My mistake was that I was running a ttest with fdr correction instead of ftest.

Ftest is taking even longer to compute. But I am getting significant results with a minimum fvalue (so the nbs.NBS.con_mat of the significant links) of 10^-4. It is really close of zero but it is not negative anymore!

Thanks very much!

Cc