Dear Andrew,

That’s not correct. As mentioned, and A(i,j) is the connection from region j to region i. So cell (2,1) is the connection from region 1 to region 2, etc.

Best

From: Andrew Nicholson [mailto:[log in to unmask]]
Sent: 03 October 2016 01:51
To: Zeidman, Peter <[log in to unmask]>
Cc: [log in to unmask]
Subject: Re:

Hello Dr. Zeidman,

Thank you for your response!

Could you possibly clarify the BMA mEp.A matrix output? Given two regions, I have a 4 cell (2x2) matrix.

I am assuming (1,1) cell (negative value), is the self connection of region 1, cell (1,2) is the connection from region 2 to region 1, cell (2,1) is the self connection of region 2, cell (2.2) is the connection from region 1 to region 2?

Thanks!

On Wed, Sep 28, 2016 at 4:13 AM, Zeidman, Peter <[log in to unmask]> wrote:

Dear Andrew

I hope you don’t mind me CC’ing the SPM mailing list again. You have shown that each group has a different winning model, but you have not yet tested your hypothesis that there is a difference between groups. This requires a comparison between groups, rather than a separate analysis within each group.

To compare the parameters (connection strengths) between groups, you need to compare like with like. I.e. you don’t want to compare the parameters of Model 1 in Group 1 against Model 2 in Group 2. So there are a couple of options:

1. As suggested, perform BMA in each group. This is an average over models (weighted by their evidence) AND an average over subjects. You will have one set of BMA parameters per group. More formally, by doing this you are marginalising over models – so you can compare groups without worrying about which model you are drawing data from. You would use the values in mEp for your ANOVA. The mEp.A matrix should be of size [nxn], where n is the number of regions, and A(i,j) is the connection from region j to region i. As your study is resting state, I imagine that this is the only matrix of interest for you.

2. Karl also gave you the alternative of performing a Bayesian Parameter Average (BPA) over subjects, for each model and group of subjects separately. (BPA is an average over one model across subjects, whereas BMA is weighted average of multiple models per subject.) Then you could enter the parameters from all models into a repeated measures ANOVA.

I would suggest option 1 – it seems more elegant to me.

All the best

Peter

From: Andrew Nicholson [mailto:[log in to unmask]]
Sent: 27 September 2016 22:24
To: Zeidman, Peter <[log in to unmask]>
Subject:

Hello Dr. Zeidman,

I am currently conducting a study with 3 groups, in which I ran DCM rfx Bayesian Model Selection separately for each.This was a resting state paradigm, in which there were sets of 2 VOIs. The models I ran characterized either top-down, bottom-up, or bi-directional information flow between the two regions (3 models). Our hypothesis was that each group would differ in terms of the direction of information flow between the 2 regions.

Interestingly, I found that each group had a different winning model, showing that the direction of information flow between the 2 VOIs is unique in each group.

Is it also necessary to then conduct Bayesian Model Averaging (BMA), in which I would run an ANOVA to detect significant differences between the 2 groups? My understanding is that BMA creates a single parameter for each subject, relative to the models inputted. Therefore, how would an ANOVA with BMA confirm that there are significant differences in terms of the winning model for each group? Is it not just telling me that there is a difference in the connectivity strength between the groups? Alternatively, could I take the BMA for just the winning model?

If I did need to do BMA, my understanding is that I would select "compute BMA" in the BMS compare batch for each group, in which I would get a BMA for each participant. However, I am confused about which parameter I would input into the ANOVA. I think it would be the mEp and sEp (mean and standard deviation for the BMA parameter), however there are 4 values in each field, in which I thought there would only be 1.

Thank you for your help in advance!

--

Andrew Nicholson B.Sc.
PhD Candidate in Neuroscience

Schulich School of Medicine and Dentistry
Western University

University Hospital
London Ontario Canada

Andrew Nicholson B.Sc.
PhD Candidate in Neuroscience

Schulich School of Medicine and Dentistry
Western University

University Hospital
London Ontario Canada