Dear Delin Sun



(1) Following SPM12 manual, I modeled two events: "neutral+emotional" and "emotional" in DCM's GLM. However, should I just model "neutral" and "emotional" seperately (as what I did in common fMRI GLM)?



Yes this is an ok way of modelling it - so neutral+emotional (which I'll call Task for convenience) will form the driving input to your DCM and emotional will form a modulatory input on one or more connections. Note that this isn't optimally efficient - the two conditions aren't independent and so there'll be covariance between the parameters - however it is the best you can do with this experimental design (ideally you would have at least 2 factors).



(2) Following SPM12 manual, I made the F contrast for "Effects of interest" (i.e. neutral+emotional) and two T contrasts (i.e. "neutral+emotional" and "emotional"). However, should I also get the T contrast for "emotional-neutral"?



These contrasts are simply there to help you choose ROIs (effects of interest has a special role in addition to this, for regressing out nuisance effects like motion). So the question of which contrasts you use depends on the hypotheses you wish to test. The "emotional" contrast will test for any effects of emotion trials relative to the unmodelled time in the GLM (i.e. the average of the neutral condition and inter-trial intervals). So that includes paying attention, viewing task instructions, etc. The "emotional-neutral" contrast will presumably control for things like viewing trial instructions / cues, so is probably a better contrast for identifying regions specifically modulated by emotion. 



(3) I extracted time series from each ROI per session per subject. However, should I concatenate all sessions per subject before extracting time series, given that the number of blocks per session are unbalanced between emotional (1-2 blocks/session) and neutral (3-4 blocks/session) conditions?



In general I would concatenate sessions - instructions here https://en.wikibooks.org/wiki/SPM/Concatenation 



(4) Following SPM12 manual, I defined the VOI of L amygdala by logically AND two images: [i1] the SPM results of "neutral+emotional" (height thresholded at p < 0.5) and [i2] the sphere covering the predefined ROI of L amygdala. On the other hand, I defined the VOI of vmPFC by logically AND two images: [i1] the SPM results of "emotional" (height thresholded at p < 0.5) and [i2] the sphere covering the predefined ROI of vmPFC. However, can I just extract signals from my predefined ROIs? Or, can I locate the VOI through using the SPM outputs corresponding to the contrast "emotional-neutral", which is more accurate than "neutral+emotional" or "emotional"?



Sorry I didn’t understand this question - perhaps you could rephrase it?



(5) I modified the "dcm_spm12_batch.m" in SPM official sample dataset of attention for my data. However, for the section of "Experimental inputs", I do not understand why there is a "33" in "DCM.U.u    = [SPM.Sess(j).U(1).u(33:end,1) ..."



SPM adds 32 time bins before the first trial (for historical reasons). You should keep this as shown in the example script.



(6) for the section of "MODEL DEFINITION", may I know whether I am wrong to model as below:

DCM.a = [1 1; 1 1]; % the bi-directional intrinsic connections between L amygdala (ROI 1) and vmPFC (ROI 2) 

DCM.b = zeros(2,2,2); 

DCM.b(1,1,2) = 1; 

DCM.b(2,2,2) = 1; 

DCM.b(2,1,2) = 1; % task modulations on L amygdala, vmPFC and L amygdala --> vmPFC 

DCM.c = [1 0;

                0 1]; % input to both L amygdala and vmPFC 

DCM.d = zeros(2,2,0); % There is no nonlinear modulations



The A and B matrices look correct, but the C-matrix does not look correct. I interpret your B-matrix as: emotion modulating the self-connection on region 1, the self-connection on region 2 and the connection from region 1 to region 2. The C-matrix is one row per region and one column per input. So your C-matrix says "Task is driving region 1 and emotion is driving region 2". Whereas, I think you meant to have Task driving both regions, which would be: [1 0; 1 0].



(7) For FFX method, is it correct that the best model is at least larger than the alternative models in 3 unit of log-evidence? On the other hand, I can't find log-evidence outputs in RFX method. How can I judge which model is the best using RFX method?



Correct. For the RFX, the convention is to look for a posterior probability or an exceedance probability of at least 95%.



(8) Am I right to extract the "mean DCM parameters per subject" from BMS.DCM.ffx.bma.mEps or BMS.DCM.rfx.bma.mEps for between-group comparison? If I understand it correctly, the DCM parameter, e.g. 0.123, of task modulation on region A-->B means that if there is a 1 unit neural activity increase in region A, there is a 0.123 unit neural activity increase in region B. Am I right? Is there any other DCM parameters or outputs for between-group comparison?



This procedure is fine, although you may wish to use the new Parametric Empirical Bayes (PEB) framework which is designed comparing groups of subjects. Temporary instructions at https://en.wikibooks.org/wiki/User:Peterz/sandbox .



Best

Peter