NITRC Network-Based Statistic (NBS) Forum: help

RE: Calculation of component size “Intensity” in NBS

Makoto HAGIHARA — Thu, 01 Aug 2024 23:39:06 GMT

Hi,
Thank you again for your quick and helpful reply. My question has been resolved.
Best regards,
Makoto
 
Originally posted by Andrew Zalesky:
<blockquote>
Hi Makoto,
the subtraction is not essential. It is for convenience, in that any positive t-statistics after the subtraction are above threshold and everything else is below threshold. The result should be the same with or without subtraction. 
I think that some of the early imaging work from the 90's and 00's subtracted the threshold as well, and that might have been my inspiration. 
Andrew
Originally posted by Makoto HAGIHARA:
<blockquote>
Hi Andrew,
Thanks for your quick reply. However, I’m still a little unclear.
I understood from your reply that it is not essential to subtract the threshold from the t-value in the “Intensity” calculation. So why did they choose to subtract the threshold in the NBS calculation? Or is it more common to subtract a threshold? I want to clarify this point.
Excerpt from file “NBS.stats.m (NBS1.2)”
%Compute a test statistic matrix
test_stat_mat=zeros(N,N); 
if ~isempty(STATS.test_stat)
        %Precomputed
        test_stat_mat(ind_upper)=STATS.test_stat(1,:)-STATS.thresh;
        test_stat_mat=(test_stat_mat+test_stat_mat');
else
        %Not precomputed
        test_stat_mat(ind_upper)=test_stat(1,:)-STATS.thresh;
        test_stat_mat=(test_stat_mat+test_stat_mat');
end
 
if Intensity
        %Measure size as intensity
        sz_links(i)=sum(sum(adj(nodes,nodes).*test_stat_mat(nodes,nodes)))/2;
else
 ...
 
Thank you, in advance, for your time,
Best regards,
Makoto
 
Originally posted by Andrew Zalesky:
<blockquote>
Hi makoto, 
I'm not sure if I fully understand your question, but I think that you are referring to the distinction between the cluster mass and cluster extent statistics. For cluster mass, we add the test statistic across connections that comprise the connected component and that yeilds the statistic for the connected component. The threshold can then be substracted from the statistic so that the statistic reflects how much of an effect we have above and beyond the threshold. But it is not essential to subtract the threshold. I hope that I have understood your question correctly. 
Andrew
Originally posted by Makoto HAGIHARA:
<blockquote>
Hi,
 
I’m using the NBS toolbox to analyze differences between the results of two task conditions, in which I have two representative connectivity matrices for each subject. Regarding calculating the component size “intensity”, why does the NBS add up the t-values of each component minus the threshold value? I referred to the calculation of the variable test_stat_mat in the file NBS.stats.m. I would like to know the difference between this method and the calculation method that adds up the t-values of each component as they are, without subtracting the threshold. 
 
Thank you for your time. I look forward to hearing from you soon.
 
Best regards,
 
Makoto
</blockquote>
 
</blockquote>
 
</blockquote>
 
</blockquote>

RE: Calculation of component size “Intensity” in NBS

Andrew Zalesky — Thu, 01 Aug 2024 10:17:06 GMT

Hi Makoto,
the subtraction is not essential. It is for convenience, in that any positive t-statistics after the subtraction are above threshold and everything else is below threshold. The result should be the same with or without subtraction. 
I think that some of the early imaging work from the 90's and 00's subtracted the threshold as well, and that might have been my inspiration. 
Andrew
Originally posted by Makoto HAGIHARA:
<blockquote>
Hi Andrew,
Thanks for your quick reply. However, I’m still a little unclear.
I understood from your reply that it is not essential to subtract the threshold from the t-value in the “Intensity” calculation. So why did they choose to subtract the threshold in the NBS calculation? Or is it more common to subtract a threshold? I want to clarify this point.
Excerpt from file “NBS.stats.m (NBS1.2)”
%Compute a test statistic matrix
test_stat_mat=zeros(N,N); 
if ~isempty(STATS.test_stat)
        %Precomputed
        test_stat_mat(ind_upper)=STATS.test_stat(1,:)-STATS.thresh;
        test_stat_mat=(test_stat_mat+test_stat_mat');
else
        %Not precomputed
        test_stat_mat(ind_upper)=test_stat(1,:)-STATS.thresh;
        test_stat_mat=(test_stat_mat+test_stat_mat');
end
 
if Intensity
        %Measure size as intensity
        sz_links(i)=sum(sum(adj(nodes,nodes).*test_stat_mat(nodes,nodes)))/2;
else
 ...
 
Thank you, in advance, for your time,
Best regards,
Makoto
 
Originally posted by Andrew Zalesky:
<blockquote>
Hi makoto, 
I'm not sure if I fully understand your question, but I think that you are referring to the distinction between the cluster mass and cluster extent statistics. For cluster mass, we add the test statistic across connections that comprise the connected component and that yeilds the statistic for the connected component. The threshold can then be substracted from the statistic so that the statistic reflects how much of an effect we have above and beyond the threshold. But it is not essential to subtract the threshold. I hope that I have understood your question correctly. 
Andrew
Originally posted by Makoto HAGIHARA:
<blockquote>
Hi,
 
I’m using the NBS toolbox to analyze differences between the results of two task conditions, in which I have two representative connectivity matrices for each subject. Regarding calculating the component size “intensity”, why does the NBS add up the t-values of each component minus the threshold value? I referred to the calculation of the variable test_stat_mat in the file NBS.stats.m. I would like to know the difference between this method and the calculation method that adds up the t-values of each component as they are, without subtracting the threshold. 
 
Thank you for your time. I look forward to hearing from you soon.
 
Best regards,
 
Makoto
</blockquote>
 
</blockquote>
 
</blockquote>

RE: Calculation of component size “Intensity” in NBS

Makoto HAGIHARA — Thu, 01 Aug 2024 8:50:49 GMT

Hi Andrew,
Thanks for your quick reply. However, I’m still a little unclear.
I understood from your reply that it is not essential to subtract the threshold from the t-value in the “Intensity” calculation. So why did they choose to subtract the threshold in the NBS calculation? Or is it more common to subtract a threshold? I want to clarify this point.
Excerpt from file “NBS.stats.m (NBS1.2)”
%Compute a test statistic matrix
test_stat_mat=zeros(N,N); 
if ~isempty(STATS.test_stat)
        %Precomputed
        test_stat_mat(ind_upper)=STATS.test_stat(1,:)-STATS.thresh;
        test_stat_mat=(test_stat_mat+test_stat_mat');
else
        %Not precomputed
        test_stat_mat(ind_upper)=test_stat(1,:)-STATS.thresh;
        test_stat_mat=(test_stat_mat+test_stat_mat');
end
 
if Intensity
        %Measure size as intensity
        sz_links(i)=sum(sum(adj(nodes,nodes).*test_stat_mat(nodes,nodes)))/2;
else
 ...
 
Thank you, in advance, for your time,
Best regards,
Makoto
 
Originally posted by Andrew Zalesky:
<blockquote>
Hi makoto, 
I'm not sure if I fully understand your question, but I think that you are referring to the distinction between the cluster mass and cluster extent statistics. For cluster mass, we add the test statistic across connections that comprise the connected component and that yeilds the statistic for the connected component. The threshold can then be substracted from the statistic so that the statistic reflects how much of an effect we have above and beyond the threshold. But it is not essential to subtract the threshold. I hope that I have understood your question correctly. 
Andrew
Originally posted by Makoto HAGIHARA:
<blockquote>
Hi,
 
I’m using the NBS toolbox to analyze differences between the results of two task conditions, in which I have two representative connectivity matrices for each subject. Regarding calculating the component size “intensity”, why does the NBS add up the t-values of each component minus the threshold value? I referred to the calculation of the variable test_stat_mat in the file NBS.stats.m. I would like to know the difference between this method and the calculation method that adds up the t-values of each component as they are, without subtracting the threshold. 
 
Thank you for your time. I look forward to hearing from you soon.
 
Best regards,
 
Makoto
</blockquote>
 
</blockquote>

RE: Calculation of component size “Intensity” in NBS

Andrew Zalesky — Thu, 01 Aug 2024 0:06:27 GMT

Hi makoto, 
I'm not sure if I fully understand your question, but I think that you are referring to the distinction between the cluster mass and cluster extent statistics. For cluster mass, we add the test statistic across connections that comprise the connected component and that yeilds the statistic for the connected component. The threshold can then be substracted from the statistic so that the statistic reflects how much of an effect we have above and beyond the threshold. But it is not essential to subtract the threshold. I hope that I have understood your question correctly. 
Andrew
Originally posted by Makoto HAGIHARA:
<blockquote>
Hi,
 
I’m using the NBS toolbox to analyze differences between the results of two task conditions, in which I have two representative connectivity matrices for each subject. Regarding calculating the component size “intensity”, why does the NBS add up the t-values of each component minus the threshold value? I referred to the calculation of the variable test_stat_mat in the file NBS.stats.m. I would like to know the difference between this method and the calculation method that adds up the t-values of each component as they are, without subtracting the threshold. 
 
Thank you for your time. I look forward to hearing from you soon.
 
Best regards,
 
Makoto
</blockquote>

Calculation of component size “Intensity” in NBS

Makoto HAGIHARA — Wed, 31 Jul 2024 8:25:30 GMT

Hi,

I’m using the NBS toolbox to analyze differences between the results of two task conditions, in which I have two representative connectivity matrices for each subject. Regarding calculating the component size “intensity”, why does the NBS add up the t-values of each component minus the threshold value? I referred to the calculation of the variable test_stat_mat in the file NBS.stats.m. I would like to know the difference between this method and the calculation method that adds up the t-values of each component as they are, without subtracting the threshold.

Thank you for your time. I look forward to hearing from you soon.

Best regards,

Makoto

RE: Rank deficiency error in 2x2 mixed ANOVA with few repeat scans

Andrew Zalesky — Wed, 31 Jul 2024 1:38:26 GMT

Hi Ed,
yes - your design matrix will throw a rank deficient warning. 
You would need to remove the main effect that is not changing between follow up measures. For example, if age group remains the same, logitudinal data is not needed to measure the main effect of age group. Check out some of the other posts on the forum where this issue is discussed in more detail. 
Another option is to move to a full-blown LME model in which individuals are modeled as random effects but that is not supported by NBS due to the computational burden of LME models. 
Andrew
Originally posted by Ed Hutchings:
<blockquote>
Hello,
Just getting in touch regarding my design matrix for a 2x2 mixed anova with few repeat measures. I understand I can’t use exchange blocks here, and my current design matrix (attached) has columns for age group, sex, age*sex, and each subject.
When I run the contrast [1 0 0 0…] to assess main effect of age group, I get a rank deficiency error. Please let me know what I’m doing wrong! Thanks for your time.
Best,
Ed
</blockquote>

Rank deficiency error in 2x2 mixed ANOVA with few repeat scans

Ed Hutchings — Tue, 30 Jul 2024 11:07:36 GMT

Hello,
Just getting in touch regarding my design matrix for a 2x2 mixed anova with few repeat measures. I understand I can’t use exchange blocks here, and my current design matrix (attached) has columns for age group, sex, age*sex, and each subject.
When I run the contrast [1 0 0 0…] to assess main effect of age group, I get a rank deficiency error. Please let me know what I’m doing wrong! Thanks for your time.
Best,
Ed

RE: NBS - Question regarding to one sample t test

Andrew Zalesky — Wed, 24 Jul 2024 2:49:44 GMT

dNBS does not include a one-sample test option. 
Not quite sure what you mean by min and max of 0. 
Note that one-sample is rarely used. 
Originally posted by Eric Wang:
<blockquote>
Dear Professor,
I have a question about how to perform a one-subject t-test in NBS / dNBS. I did find the previous discussion that is similar to this (https://www.nitrc.org/forum/message.php?msg_id=10023) but I don't think I have found the answer to how to set the design matrix.
My case is that I now have one condition and 37 participants' FC under it. I now want to look at the overall group connections under this condition that are significant. How should I set the design matrix? I want to assess if the mean is greater than 0, so I used one sample and tried the design matrix full of 1 and the contrast of [1]. But it always gives me min and max of 0.
(In addition, it seems that you cannot run one sample t-test in dNBS, am I correct?)
Any input is greatly appreciated. Thank you for your time and assistance.
 
Kind regards
Eric
 
 
 
 
</blockquote>

NBS - Question regarding to one sample t test

Eric Wang — Tue, 23 Jul 2024 11:00:56 GMT

Dear Professor,
I have a question about how to perform a one-subject t-test in NBS / dNBS. I did find the previous discussion that is similar to this (https://www.nitrc.org/forum/message.php?msg_id=10023) but I don't think I have found the answer to how to set the design matrix.
My case is that I now have one condition and 37 participants' FC under it. I now want to look at the overall group connections under this condition that are significant. How should I set the design matrix? I want to assess if the mean is greater than 0, so I used one sample and tried the design matrix full of 1 and the contrast of [1]. But it always gives me min and max of 0.
(In addition, it seems that you cannot run one sample t-test in dNBS, am I correct?)
Any input is greatly appreciated. Thank you for your time and assistance.
 
Kind regards
Eric

RE: T-threshold selection

Andrew Zalesky — Sat, 15 Jun 2024 0:31:44 GMT

Hi Hannah,
Rather than setting a threshold, you might want to consider summarazing your results according to the number of signficiant connections implicated for each network. For example, you might have 58% of connections within the DMN are implicated and 45% connection between the DMN and other networks are implicated. I would suggest not to put the threshold too low, since while this may detect statistically signifciant effects, the effect sizes will be low and possibly not meaningful from a practical perspective. Statistical signifciance does not imply practical/clinical signfiicance. 
andrew
Originally posted by Hannah Terborg:
<blockquote>
Hi,
I have a question about-threshold selection.   I know that any choice of component forming t-threshold is rather arbitrary and does not affect specificity. It is often recommended, especially if the type of effect (small and strong vs wide and distributed) is unknown, to try a range of thresholds. I followed this advice for the analysis of my wPLI MEG connectivity data. This led to networks ranging from very dense to consisting of just one or two edges.  Do you have any advice on how to choose one network among all significant networks for interpretation and further analysis?
 
Best,
Hannah  
</blockquote>