Dear Andrew, 

I am comparing a patient and control group, while accounting for effects of age, gender and years of education. 
My design matrix looks then something like this: 

1 0 41 1 12 
1 0 34 0 17
1 0 27 1 10
0 1 55 1 12
0 1 58 0 15
0 1 22 0 20

I want to assess for both decreased and increased connectivity in the patient group, hence my contrasts using a t-test are [-1,1,0,0,0] and [1,-1,0,0,0]. I do find significant subnetworks for both contrasts. 
I now want to check whether there is a correlation with clinical scores. I already searched on the forum, and I came across different possibilities: 

1) Add an extra column to the design matrix and then set the contrast as [0,0,0,0,0,1] to test for positive correlations and [0,0,0,0,0,-1] for negative correlations with connectivity strength.

However, how exactly do you interpret the result? In case of a significant result, can you conclude something on whether the correlation was found for instances of decreased or increased connectivity in the patient group? Is it checking then the correlation with the same subnetworks that come out to be significant using the previously mentioned contrasts? Or could potentially be that totally different subnetworks correlate with clinical variables? 

2) Go for a within-group comparison, with contrasts [0,0,0,0,1] to test for positive correlations and [0,0,0,0,-1] for negative correlations, basically kicking out the control group.

Is there a difference with approach (1)?

3) Extract the edge weights of the significant subnetworks from the original analyses and correlate those with clinical scores. 

Can I expect different results and does it make sense to do this, if I do not find any difference with approach (1) or (2)?

Lastly, if I have more than one clinical score, should I have different design matrices for each clinical score that I want to correlate, or can they be combined in one design matrix? 

Sorry for the many questions and I hope it is clear!

Thanks in advance,

Kind Regards
Samantha

Hi Samantha, 

Subnetworks that associate with a clinical variable do not necessarily need to correspond with subnetworks associated with a between-group difference. They can be mutually exclusive or fully overlap. 

In many cases, the clinical variable (e.g. symptom severity) would not be available in the control group and thus the correlation analysis would be constrained to the sample of patients. 

If you have measured the clinical variable in controls, then you may want to consider including the controls when testing for association, but it will be important to control for the effect of diagnosis. 

The three options are all valid but test different hypotheses. Option (3) asks whether there is an association with the clinical variable within the subnetwork associated with the between-group difference that has already been found, whereas Option (1) and (2) test for an association across all connections in the brain network.

Combining the clinical scores into one design matrix ensures that you control for the other scores when testing a particular score. In contrast, if a separate design matrix is used for each score, the effect of other scores is ignored. In any case, it will be important to correct for multiple comparisons across the scores that are tested.

Best wishes, Andrew 




Originally posted by Samantha Baldi:

Hi Andrew, 

Thanks for your quick and very clear reply. Now everything makes sense. 

I just have one last question about approach (3). Would you correlate the clinical scores with the single edge of each subnetwork or with the sum/average of the edges of the whole subnetwork (if the subnetwork is made of e.g. 3 edges)? 

Is one of these two options more correct than the other?

Hi Samantha, 

Both approaches are reasonable and valid for post-hoc testing. Using the average may be easier to interpret since you will only have a single value to consider per subject. This would be a good starting point.  

On the other hand, you may want to make inference about one of the three specific connections in your subnetwork, in which case considering each connection separately would clearly be advantageous. 

Best wishes, 
Andrew

(PS. My name is Andrew.)

Originally posted by Samantha Baldi:

Dear Andrew, 

Thanks again for the clear answer and so so sorry, I have no idea what happened to my brain while I was typing these questions. 

Best