Dear NBS Community,

I have the following study design/question.

2 groups were treated with 2 different treatments. I would like to use their baseline functional connectivity to see if it correlates with their eventual change in the clinical score after treatment completion. I would like to covary for sex, age, and baseline clinical scores.

However, because of the different treatment groups, I would also like to include an interaction term between treatment group and baseline functional connectivity. So something like this:

changeinclinicalscore ~ treatmentgroup*baselinefunctionalconnectivity + age + sex + baselineclinicalscore

However, going over some of previous threads (https://www.nitrc.org/forum/message.php?...), it seems as though such interactions are not recommended in the context of a correlation analysis. I was thinking of doing a "mass univariate" approach, where I apply this model to each connection, but would really like to use the NBS toolbox to investigate this question.

Any suggestions as to whether I can/should attempt to model this interaction term, and if so, how, would be greatly appreciated.

Thank you,
Paul

Hi Paul, 

thanks for your question. 

Just like FSL and SPM, the NBS requires that predictors (independent variables) do not vary between voxels/connections. So a predictor like age satisfies this condition because a person's age is the same for all connections. However, I assume that baseline connectivity will vary between connections and therefore it cannot be included a predictor. Or do you mean that baseline connectivity is an average over all connections in an individual's brain network? 

Also, if your dependent variable is change_in_clinical_score, I don't think it is particularly helpful to covary for baseline_clinical_score. A better model may be to predict the follow-up clinical score and covary for baseline clinical score. 

Check out this post for details: https://stats.stackexchange.com/questions/15713/is-it-valid-to-include-a-baseline-measure-as-control-variable-when-testing-the-e/15759

NBS is suited to a model of the form:

functional connectivity (or change in functional connectivity) ~ treatment_group + change_in_clinical_score + interaction + age + sex + ... 

Originally posted by Paul Dhami:

Dear Dr. Zalesky

Thank you very much for your response.

Ah I see. Apologizes for what is a naive question (I work with EEG primarily, so am not too familiar with FSL and SPM), but why is it that the predictors cannot vary? Is it due to the method of cluster correction applied?

Thank you for the suggestion regarding not using change scores and baseline scores together! Very useful.

So if I am ultimately interested in using baseline functional connectivity to see if it correlated with improvement on the clinical scale (while also including the interaction term between functional connectivity and treatment group), this is doable with the model you specified?

functional connectivity (or change in functional connectivity) ~ treatment_group + change_in_clinical_score + interaction + age + sex

Thank you,
Paul

Hi Paul,

In theory, predictors can indeed vary between connections, but that would mean a separate design matrix is required for each connection/voxel. I think that this is possible with PALM, but not with NBS, FSL Randomise or SPM. It's more just a matter of implementation.

Yes - the new model that you have specified can be implemented straightforwardly. 

best wishes,
Andrew

Originally posted by Paul Dhami: