Dear Sam
Sorry for the delay in getting back to you. You are quite right that there are (at least) two ways of modelling a two-factor design in DCM. These options could be compared using Bayesian model selection. I will make this answer generic
by calling your two factors A and B, and treating each factor as having 2 levels (i.e. a 2x2 design). There are therefore four experimental conditions: A1_B1, A1_B2, A2_B1 and A2_B2.
Option 1: Drive the network with only level 1 of factor A (A1_B1 and A1_B2), leaving the other level unmodelled. Modulate particular connections using factor B, e.g. by computing an input vector (parametric regressor) which is
the simple main effect of A1_B1-A1_B2.
Option 2: Drive the network by all four conditions (these could be combined into a single ‘Task’ input to capture the mean of the experimental conditions). Modulate particular connections either by the individual conditions,
or contrasts of them. For example, the main effect of factor A: (A1_B1+A1_B2)-(A2_B1+ A2_B2) and the main effect of factor B: (A1_B1+A2_B1)-(A1_B2+A2_B2).
The second option is typically used when there’s an implicit extra factor, induced by having a baseline condition, inter-trial intervals or null trials. In other words, when in addition to the 2x2 design, there’s an extra factor of Task
vs baseline. In this case, Task can be used as the driving input, leaving the baseline unmodelled. The main effects of each factor (A and B) can then be used as modulatory inputs on particular connections. There is no general answer as to whether Option 1
or Option 2 is best, hence my suggestion to do a Bayesian model comparison to help make the decision.
Regarding how to implement your 2 x 3 factorial design with factors condition (active vs passive) and valence (positive, negative, neutral). I will assume you have some unmodelled null trials or inter-trial intervals serving as a baseline.
Under option 2, you could create two conditions in your GLM: active and passive. For each of these, have two parametric regressors: emotion vs neutral and positive vs negative, coded with 1s and -1s. These split your 3 levels of valence into 2 differences.
Set orthogonalisation to false in the GLM specification, so the software doesn’t alter the regressors. In the DCM, use the active and passive conditions as driving inputs, and use each of the four parametric regressors as modulatory inputs where you think
is appropriate in your network. (I suggest you answer yes to mean-centering in the DCM specification, i.e. DCM.options=true).
I hope that helps – do let me know if anything’s unclear.
Peter
-----Original Message-----
From: Sam W. <[log in to unmask]>
Sent: 17 February 2019 02:44
To: [log in to unmask]; Zeidman, Peter <[log in to unmask]>
Cc: Sam W. <[log in to unmask]>
Subject: Re: DCM questions
Hi Peter,
thank you very much for your (always) very helpful response!
I think I understood what you meant but just to be sure, can I ask you two other questions?
In your response you wrote "I would drive the DCM using just the active condition (leave the passive unmodelled)", but I thought it is not a good idea to leave conditions unmodelled. In this post
https://www.jiscmail.ac.uk/cgi-bin/webadmin?A2=spm;555aca73.1803, you seem to recommend modelling everything in the DCM,
or did I misunderstand?
If I decide to go for the more efficient modelling, should the my design matrix look like this for the active condition?
emot vs neu active PM(emot vs neu active) pos vs neg active PM(pos vs neg active)
Pos 1 Pos 1
Neg 1 Neg -1
Neg 1 Neg -1
Pos 1 Pos 1
Neu -1
Neu -1
Pos 1 Pos 1
where Pos,Neg and Neu are the onsets for those valence types, and PM the parametric modulator. I would have 4 columns in total (emot vs neu active, pos vs neg active and their corresponding parametric regressors). Or should I still include
the passive condition in the design matrix (emot vs neu passive, pos vs neg passive and their corresponding parametric regressors), in which case I would have a total of 8 columns?
Thank you for your help!
Best regards,
Sam
On Thu, Feb 14, 2019 at 11:34 AM Zeidman, Peter <[log in to unmask]> wrote:
Dear Sam
This is a good question and helpful for others as it's quite generic. You've got a 2 x 3 factorial design: condition (active vs passive) and valence (positive, negative, neutral). You hypothesise an interaction: a connection from region A to region B should be boosted by positive valence, specifically in the active condition.
There are various ways to model this, and when in doubt, you can try different options and compare them using Bayesian Model Comparison. Typically, one factor is used as the driving input and another factor is used as modulation. So in this case, I would drive the DCM using just the active condition (leave the passive unmodelled). And place valence as a modulatory input on the connections of interest. This expresses your hypothesised interaction - valence will only have an effect when stimuli are in the active condition.
A little complexity is added by having three levels to your valence factor. One possibility would be to rework this into two modulatory inputs: emotional vs neural and positive vs negative. To do this, you'd need to go back to your GLM and add parametric modulators for these (e.g. 1 for emotional and -1 for neutral). Alternatively, you could keep all three valences separate, and place them on the connections of interest. I think this is a bit less elegant, efficient and interpretable, but easier to implement.
Hope that helps
Peter
-----Original Message-----
From: SPM (Statistical Parametric Mapping) <[log in to unmask]> On Behalf Of Sam W.
Sent: 10 February 2019 07:16
To: [log in to unmask]
Subject: [SPM] DCM questions
Hi,
I've just started to learn DCM and I was wondering what would be the best way to specify my model.
My experimental design consists of a 2 condition x 3 valence. Condition has two levels (active and passive), and valence has three levels (positive, negative and neutral). The example in the SPM manual uses a visual input regressor consisting of every visual stimulus, but this doesn't seem to be the recommendation for factorial designs, so I'm wondering how to use the SPM design I already specified.
My hypothesis is that there will be forward connectivity from region A to region B for positive emotional words only in active condition but not in passive condition. So is condition my modulatory input and all words my driving input? Or should I specify only positive words as my driving input?
In the SPM design matrix I have the following regressors ActivePos, ActiveNeg, ActiveNeu, PassivePos, PassiveNeg, PassiveNeu
With only two regions, and the driving input to region A, my C-matrix should look like 1 1 1 1 1 1; 0 0 0 0 0 0, correct?
If active condition has a modulatory effect on connection A to B during positive words, then my B-matrix should be b(:,:,1)=[0,0;0,1] b(:,:,2:6)=[0,0;0,0] is that correct?
Thank you!
Sam