Hello Dr Zalesky,
If possible, I would like your help in clarifying some questions I have.

My experiment consists of 2 Groups (G1 and G2) and 2 Conditions (C1 and C2).
The 2 conditions are measured for each subject of each group. I have performed a comparison to estimate the GroupxCondition interaction using the following design matrix

1 1 1 0 0 0
-1 -1 1 0 0 0
1 1 0 1 0 0
-1 -1 0 1 0 0
1 -1 0 0 1 0
-1 1 0 0 1 0
1 -1 0 0 0 1
-1 1 0 0 0 1
(col1: Condition, col2: Interaction, col3-6: Within-subjects mean)

and contrast
[0 1 0 0 0 0] with F-test, and
[0 1 0 0 0 0] with T-tests
[0 -1 0 0 0 0]

Based on this comparison, I have the following questions

Q1. Regarding the F-test: Am I right in interpreting that a significant result in this comparison equates to C1>C2 while G1>G2
and C1 < C2 while G1 < G2?

Q2. I want to perform a posthoc analysis to determine which Group x Condition factor "drives" the difference in the interaction e.g., G1xC1 or G2xC1 denote an increase (or decrease) in connectivity. Can I estimate this using my existing design matrix with a different contrast? If not what would be the best way to approach this?

Hi Alexandra, 

To understand the interaction effect, for each individual, consider averaging the connectivity across all the connections in the subnetwork identified. You could then plot the average connectivity for each group and for each condition. This should reveal the nature of the interaction. The NBS manual provides details about how the connectivity values can be extracted from each significant subnetwork. 

I don't know how you have encoded C1 and C2 (i.e. -1 or 1), and so it is difficult to answer your first question. However, note that that F-test will be sensitive to both "positive" and "negative" interaction effects. In contrast, the T-test will only be sensitive to one of the two interactions. 

Plotting the average connectivity within each condition and group should reveal the nature of the interaction. 

It is difficult to determine "driving/causal" effects with a regression model. You may need an alternative model that can infer directional relationships, such as structural equation modelling (SEM) or Bayes nets for this purpose. 

best,

Andrew
Originally posted by Alexandra Anagnostopoulou: