Dear Alfonso and other CONN users,

I was hoping you might be able to enlighten me regarding the appropriate use of Network Based Statistics in CONN.

I am investigating a network of 40 seeds in 20 individuals and I want to determine if there are changes in intranetwork connectivity that correlate with a continuous behavioural measure. I am most interested in determining whether the network as a whole varies with the behavioural measure, but would also be interested in identifying particular nodes that are most implicated

My questions are:

1. I select all 40 seeds. I threshold ROI-ROI connections by p<0.05 uncorrected. I then threshold SEED ROIs (NBS; by size) by p<0.05 FDR. For the seeds that are found to be significant here, is it appropriate to say that their connectivity to the network significantly correlates with the task performance?

2. I threshold ROI-ROI connections by p<0.05 uncorrected. I then threshold NETWORKS (NBS; by size) by p<0.05 FDR. My results are: 1 significant network with 17 positively correlated edges and 1 negatively correlated edge. What I want to test is whether increased connectivity within the network correlates with the task performance, but it seems that as the graphs can contain negative edges this may not really be testing that?

3. In the examples above the initial ROI-ROI threshold can have a significant impact on results, is there a danger of increasing false positives by using this relatively relaxed threshold (as e.g. occurs with conventional cluster based statistics in the Eklund PNAS paper). It seems to me that we don't have to worry about this due to the NBS permutation testing approach?

4.My understanding is that NBS is testing for the size of graph components. Unless we have shown that there are no differences in connectivity strengths, however, is it possible to say whether a difference as tested by NBS is better understood as representing a change in connectivity strengths or a difference in graph structure?

Many thanks and please let me know if the questions are unclear,

Rob

Hi, Rob.

It seems like you might be able to empirically test your question in #4 by looking at the differences in your data NBS analyzed by size and intensity.  I'm a bit shaky on the whole NBS approach myself, but I thought the NBS intensity was getting at connectivity strength.

Jeff

Dear All,

a follow-up question re NBS: do I understand it correct that the NBS statistics are only appropriate for testing T-contrasts, but not F-contrasts?

I.e. if I have two groups and two conditions and want to look at the 2x2 interaction at the network level, can I test contrast [1 -1; -1 1] (subject effects) with [1 -1; -1 1] (condition effects), or should I rather specify a contrast like [1 -1] (subjects effects) [1 -1] (condition effects)?

Best,
Ekaterina.

Dear Ekaterina,
NBS statistics are appropriate for F-contrasts as well. Both of the contrast specifications ([1 -1;-1 1] or [1 -1]) will work exactly in the same way (assuming that you are specifying "two-sided" effects for the T-contrast if you are using the latter)
Best
Alfonso

Originally posted by Ekaterina Pechenkova:

Dear Alfonso,

sorry for a delayed follow-up. This time I played a little bit with the contrasts and the NBS in another project and found that NBS just hangs with some contrast specifications. I had two groups (Treatment and Control) and one covariate of no interest (Age), and two conditions (Pre and Post). So if I specify between-subject contrast as [1 -1 0] and between-condition contrast as [1 -1], then go to the ROI-to-ROI results explorer, select the ROIs, the two-sided contrast and hit the Enable permutation test button, the results are there in a minute or so. However, if I specify ANOVA as [1 -1 0; -1 1 0] between-subjects and/or [1 -1; -1 1] within-subjects when it comes to the permutations, the toolbox just hangs. Does it mean that only one way of the contrast specification is applicable for the NBS? In other respects the two models seem to be equivalent, do not they?

I use the Conn 17.a with SPM12.

Thank you.

Best,
Ekaterina.
Originally posted by Alfonso Nieto-Castanon: