Hello NBS users,

I was hoping to get some feedback on my design matrices/contrasts for my study. 

I have several behavioural measures (BM1, BM2, BM3, BM4, BM5), and three independent variables (patients vs controls, men vs women, genetic risk vs no genetic risk).  I'd like to see how the functional networks underlying the behavioural measures compare between patients and controls, between men and women, and between genetic risk vs no genetic risk.  I'm *not* interested in interactions (i.e., interaction between patient status vs gender). 

For BM1, I'm looking at patient vs control.
For BM2, I'm looking at men vs women.
For BM3, BM4, BM5 I'm looking at genetic risk vs no genetic risk.

From my understanding, I would need to create 5 separate design matrices for the 5 behavioural measures. 

Design matrix example for BM1 looking at patient vs control for five participants:

The 1st column is intercept; 2nd is patient or control; 3rd is behavioural measure
1   1   1.5
1   1   0.78
1   0   3.2
1   0   1.1
1   1   0.5

F-test + Contrast [0 0 1] to test for either a negative or positive association

***OR***

Would my design matrix need to have an interaction term between the behavioural measure and diagnosis if I'd like to see how the networks underlying behavioural performance compare between patients and controls?

The 1st column is intercept; 2nd is patient or control; 3rd is behavioural measure; 4th is the interaction between diagnosis and behavioural measure
1   1    1.5     1.5
1   1    0.78   0.78
1   -1   3.2    -3.2
1   -1   1.1    -1.1
1   1    0.5     0.5

F-test + Contrast [0 0 0 1] to test for either a negative or positive association

Or is there a different way that I should be setting up my design matrices and contrasts for the type of analysis I'd like to do?
Would there be a difference between running an F-test or t-test (i.e., t-test contrasts [0 0 0 1] and [0 0 0 -1] for two-tailed)?
I'm also wondering if the design matrices/contrasts used for NBS would be the same ones to use if I were to try to run analyses in NBS-Predict?

Thank you for your help!

Alica

Hi Alicia, 

both of the design matrices that you have specified are technically correct. However, given that you emphasize that the interaction term is not of interest, there is probably no need to include the interaction, as you have done in the second design. 

It seems a bit odd that you are not interested in the interaction, since this means that the main effect of patient vs control is effectively treated as a confound in the first design. Of course there is nothing wrong with that, but it could be asked why bother acquiring patients in the first place. 

You can also run two separate t-test with a contrast of [0 0 1] and [0 0 -1]. This would allow you to disentangle positive and negative associations with the behavioural measure. 

Yes - you can create five different define matrices. You may need to use a more stringent p-value threshold to account for multiple comparisons (i.e. 0.05/5)

best,
Andrew
Originally posted by rogojin:

Hi Andrew,

Thank you very much for your quick response.

Sorry, I meant I'm not interested in interactions between my independent variables (i.e., diagnosis and sex, or diagnosis and genetic risk).  I am, however, interested in the interaction between the independent variables and behavioural measure. 

1) For example, I'd like to see if the functional networks correlated with BM3 differ between patients and controls - would I then need to use my second design matrix structure?  If so, could I use t-tests with contrasts set to [0 0 0 1] and [0 0 0 -1]?

2) In terms of the p-value, would I still need to use a more stringent value even if each design matrix is looking at a different behavioural measure and only one of my independent variables?  E.g., for BM1, I'm only looking at the comparison between patients vs controls (I'm not looking at sex or genetic risk). 

3) Would there be any disadvantages to using these design matrices in NBS-Predict instead of NBS?  I believe in NBS-Predict identified a large-scale subnetwork associated with general intelligence.  I was wondering if I could do the same with my data using my behavioural measures, except with the added diagnosis/sex/genetic risk interactions.  Essentially, identify subnetworks underlying each behavioural measure and comparing if they differ between my groups.  I'm guessing I would need to use my second design matrix for this?

Thank you again for your help,
Alica

Hi Alicia, 

1. Yes - your second design matrix is correct in this case. And yes you could test two separate contrasts with a t-test. 

2. Yes - you would still need a more stringent p-value because each new behavioural measure is a new comparison and it is important to correct for multiple comparisons. So if you consider five behavioural measures, you probably want to use a p-value threshold of 0.05/5, rather than 0.05.

3. Not quite sure if I understand your question here. NBS-Predict enables prediction of behaviours and the method is a little more complex to use (advanced users). Of course you are welcome to give it a go!

best,
Andrew

Originally posted by rogojin:

Hi Andrew,

Thank you very much! 

Sorry, I wasn't very clear in my last message - I just meant if there are any situations in which you might *not* recommend someone use NBS-Predict over NBS?

Since I wouldn't be using NBS-Predict for its classifier machine learning functions, would the results of a regression run in NBS differ from a regression run in NBS-Predict?  I.e., would it be bad to run the analysis in NBS instead, since I believe my research question can be investigated using both NBS and NBS-Predict?  Would the outputs from both be similar in terms of interpreting the results? 

All the best,
Alica

Hi Alicia, 

The choice between NBS and NBS-Predict depends on your research question and aims. NBS-Predict allows for out-of-sample prediction using cross-validation, however, it is a more complex and computationally demanding method. NBS enables statistical inference. 

Using NBS is perfectly fine. 

Andrew

Originally posted by rogojin:

Hi Andrew,

Thank you very much, I appreciate your help with this.

Best,
Alica
Originally posted by Andrew Zalesky: