Hi Andrew, 

I ran some graph theory analyses prior to using NBS and the results showed case-control differences in quite a few brain regions (mostly cases having lower values, such as lower nodal efficiency, than controls) after controlling for nuisance covariates like age, sex and site. I then tried using NBS to identify subnetworks, with the expectation that I will find significant subnetworks involving the same regions identified in the graph theory analyses. I used a two-sample t-test with the contrast [-1 1 0 0 0] - the first column for cases, second column for controls, third column for age and fourth column for site. However, there were no significant subnetworks over a range of primary thresholds (p<0.001 to p<0.05) for both extent and intensity. I also tried using the FDR option with 50,000 permutations but likewise, no significant differences were found. Significant subnetworks were identified if I do not control for the nuisance covariates but I do think it is necessary to add them in.

I am a bit puzzled by the results because there were differences at the nodal level and it would make sense to see differences at the connection level. What do you think? I know this is not directly related to NBS methodology, but it would be really great if I can get your opinion on this. 

Thank you very much!

Best Regards
GT

Hi GT,

it is certainly possible for significant effects to be found in nodal metrics but not in connectivity. 

Be sure to correct for multiple comparisons across the set of nodes and measures. So if you have 100 nodes and testing for 3 nodal measures, the total number of comparisons is 300. 

It is difficult to provide meaningful comments without knowing more about your study. 

best
Andrew

 
Originally posted by gthngjy:

Thank you for your quick reply! 

I did correct for multiple comparisons but I did it differently from what you mentioned. I used linear regression to assess case-control differences and did permutation testing 1000 times to determine the empirical p-values. For the 3 global-level network metrics, significance was determined through the empirical p-values with Bonferroni correction of 0.05/3. For each of the 2 node-level network metrics, significance was determined for case-control effects using FDR correction for 85 nodes, with threshold pFDR <� 0.05/2. Does this sound feasible to you? A lot of measures were significant after permutation testing but not FDR correction.

Also, sorry but I still do not really understand why it is possible to see differences in nodal metrics but not connectivity. Isn't it differences at the connection level that will lead to differences at the nodal level? Especially for node-level metrics like nodal efficiency and clustering coefficient?

Thank you again and happy to continue this conversation over email too if it is not too relevant here! 

Best Regards
GT

It is possible to identify differences in nodal or global metrics, without being able to identify effects at specific connections. 

The correction for multiple comparisons sounds reasonable. 
Originally posted by gthngjy: