Dear Carole
Sorry for the delay in replying. You asked how to calculate an interaction between factors at the within-subject level. (Sorry, in my previous reply I was thinking you were referring to between-subjects effects, in which case hand-calculating the interaction is very straightforward using PEB, with the code I posted below. The answer to your question is a bit more subtle.)
If you wanted to include the interaction between conditions as a modulatory input to certain connections, you could specify the DCM in the normal way using the GUI, and then manually edit the design matrix within the DCM (DCM.U.u), to add an extra regressor with a hand-calculated interaction term. You would also have to adjust the size of certain other matrices to account for this, and delete any old priors from your previous analysis (DCM = rmfield(DCM,'M').
I can help you with that if it were needed. However, I would suggest against doing this if possible. If you have one experimental factor driving, and the other modulating, you are naturally getting the interaction in the dynamics of the model - i.e., activity will only flow down a connection if there is driving input, and the activity can be gated by the modulatory input. That's a regionally specific interaction.
All the best
Peter
-----Original Message-----
From: carole guedj <[log in to unmask]>
Sent: 29 April 2022 15:05
To: Zeidman, Peter <[log in to unmask]>
Cc: [log in to unmask]
Subject: Re: [SPM] DCM model interaction
⚠ Caution: External sender
Dear Peter,
Many thanks for you answer and precious advices.
Sorry for my mistake, when I mentioned the D matrix: ‘D = [1 0 0; 1 0 0; 1 0 0; ] ‘, in fact I wanted to talk about the C matrix.
So yes, for the D matrix I used the default parameters which are 0s.
Just to be sure, if I would like to try the interaction regressor, I need to create a new GLM with this time 4 regressors, and run another DCM, right? What you said about: 'Mean-subtract each regressor and then multiply them to generate the interaction’.
> X(:,2) = X(:,2) - mean(X(:,2));
> X(:,3) = X(:,3) - mean(X(:,3));
> X(:,4) = X(:,2) .* X(:,3);
You mean that I can directly create the interaction in the existing DCM?
Sorry for my naive questions…
Thanks again for your help,
Carole
> Le 29 avr. 2022 à 11:22, Zeidman, Peter <[log in to unmask]> a écrit :
>
> Dear Carole
> All good questions!
>
>> I have a 2x2 factorial design where I would like to run an DCM.
>>
>> Briefly, it is a simple attentional task where subjects need to find a target amoung distractors.
>> In half of trials the target location is cued (others the cue is uninformative), and in half of trials a salient distractor appears.
>> So we have 4 conditions:
>> - Cue - with Salient Distractor
>> - Cue - without Salient Distractor
>> - No Cue - with Salient Distractor
>> - No Cue - without Salient Distractor
>
> Excellent, we like 2x2 designs.
>
>> I create a GLM specifically for my DCM analysis where I have 3 regressors (+ nuisance variables):
>> - regressor 1: onsets of cue display (including cue and uncued trials) > use as driving input for the DCM (matrix C)
>> - regressor 2: onsets of trials with Cue (the one which is informative) > use as modulatory input for the DCM (matrix B1)
>> - regressor 3: onsets of trials with Salient Distractor > use as modulatory input for the DCM (matrix B2)
>
> Perfect.
>
>> I extracted the time series of my 3 regions of interest (adjusted with F-contrast modelling my 3 regressor of interest) QUESTION 1: Is my GLM and time-series extraction correct?
>
> Yes.
>
>> Then I specified a full DCM : full A matrix, full B matrices, and matrix D = [1 0 0; 1 0 0; 1 0 0; ].
>
> By using matrix D you switched on non-linear DCM, which enables regions to mediate connections between other connections (https://eur01.safelinks.protection.outlook.com/?url=https%3A%2F%2Fdx.doi.org%2F10.1016%252Fj.neuroimage.2008.04.262&data=05%7C01%7Cpeter.zeidman%40ucl.ac.uk%7Cc68ea473a6ee4539aaaf08da29e94cad%7C1faf88fea9984c5b93c9210a11d9a5c2%7C0%7C0%7C637868379244249045%7CUnknown%7CTWFpbGZsb3d8eyJWIjoiMC4wLjAwMDAiLCJQIjoiV2luMzIiLCJBTiI6Ik1haWwiLCJXVCI6Mn0%3D%7C3000%7C%7C%7C&sdata=b8eKpOz7d1Lbhib%2Bht%2FeEW8X8GzDLmeSkyJ060Vw8DQ%3D&reserved=0). I'm just checking that's what you intended? It's a more difficult model estimation than standard DCM, so it should only be used when it's really needed.
>
>> QUESTION 2: Then I saw in the PEB tutorial that I could try two options: automatic search of nested models, or specified myself some reduced models to test.
>> If I understand correctly, if I don't have any covariate to add in the model and I do not want to compare between groups, but I just want to find the best model for all my subjects, I can simply use a design matrix M=ones(N,1) ?
>
> Yes.
>
>> QUESTION 3: If I use the automatic search, can I consider that the parameters exhibiting a high posterior probability are important to model my task effects?
>
> Yes. In general, I recommend using hand-specified hypotheses and models where possible. That's because DCM is mainly intended as a tool for scoring the evidence for different hypotheses. The automatic search is valid when all reduced models are equally likely a priori - in other words, if you would be equally happy with any mixture of connections showing task effects. If that's not the case, and you have particular hypotheses, I'd hand-craft some candidate models (i.e., mixtures of connections).
>
>> QUESTION 4: How can I evaluate the interaction effect between my factors? Do I need to add a 4th regressor in my GLM that specified the onsets of trials where both Cue and Salient Distractor appear?
>
> Technically, yes you can do this. Mean-subtract each regressor and then multiply them to generate the interaction. Assuming that the first regressor is the constant (all ones), which is expected by the software, and 2nd and 3rd regressors are your main effects, then:
>
> X(:,2) = X(:,2) - mean(X(:,2));
> X(:,3) = X(:,3) - mean(X(:,3));
> X(:,4) = X(:,2) .* X(:,3);
>
> However, note that interpreting the interaction may be tricky - there will be lots of positive and negative signs in play.
>
> All the best
> Peter
|
