Dear Miranda (CC’d Mailing List)

I received some useful feedback from Karl Friston regarding your query. To summarise, you have a working memory paradigm with one experimental factor, which is the feature of the stimulus which is to be maintained over a delay period. This factor has three levels – color, shape and both. You wish to test the hypothesis that a brain region (VOI4) modulates connections between early visual cortex (VOI1) and color and shape selective higher visual regions (VOI2 and VOI3 respectively).

First a couple of general comments on the design of the experiment and analysis:

1. The analysis is slightly complicated by not having a balanced factorial design. I mention this in case it helps for the design of your next experiment. Ideally you would have 2 factors: color and shape. Each of these would have 2 levels: maintain and don’t maintain. Thus, you’d have 4 conditions: maintain nothing (e.g. just fixate), maintain colour, maintain shape, maintain both. You would then have 2 orthogonal contrasts in your GLM and DCM. Instead, we are creating a dummy 2^nd factor by taking the whole trial vs the inter-trial interval (the ‘task’ regressor I suggested previously). Obviously this will not only reflect the delay period, but will also include the encoding and test stages.

2. You may be interested in the difference between onset responses at the beginning of the maintenance period, and the ongoing maintenance of the stimulus. If so, then to model this you would have both a stick regressor (duration 0) at the start of the delay period and a block covering the whole delay period. See previous working memory studies for examples of this.

With regards to the DCM… You are asking for a 3-way interaction, between VOI4, VOI1 and experimental condition. DCM does not allow for 3-way interactions (as you say, it cannot be done with the non-linear DCM). So here are two options to get round this:

- Just include color vs shape as a modulatory input on the connection(s) between VOI1 and VOI2 / 3. You can then give suggestions as to where in the brain this modulation may be coming from (you already know from your SPM results that VOI4 is a reasonably candidate).

- Here’s the slightly more technically demanding option which more closely addresses your question. Create a regressor which is the interaction between the BOLD timeseries in VOI4 and the color / shape difference. This is known as a PPI (psycho-physiological interaction). Include this PPI regressor as a modulatory input in your DCM on the connections between VOI1-VOI2 and VOI1-VOI3.  I think you said you’ve already done a PPI analysis, so you’re already half way there.

Notes:

- For both of these options, don’t use the non-linear option. It doesn’t get at what you’re asking.

- For both of these options, you could compare the evidence for models having the modulation on the connections from VOI1, against having the modulation on the self-connections on VOI2 and VOI3. The first option models attention to different stimulus features as specifically involving VOI1 connections. The second option models attention as more general gain control on VOI2 and VOI3.

I hope that helps.

Best

Peter

I am setting up nonlinear DCM for a task with three conditions: Color, Shape, Color+Shape (referring to the stimuli's dimension(s) that is relevant to a working memory task. A retrocue is used to indicate the relevant dimension(s) after initial encoding of both color and shape dimensions of the stimuli).

I have 4 VOIs: early visual VOI1, color-selective VOI2, shape-selective VOI3, and task rule-selective VOI4. I expect direct effect of task conditions on VOI4 which in turn modulates the VOI1->VOI2 connection and VOI1->VOI3 connection. More specifically, for the Color condition, I predict that VOI4 enhances VOI1->VOI2 connection while reducing VOI1->VOI3 connection. Similarly, during the Shape condition, VOI4 reduces VOI1->VOI2 connection while enhancing VOI1->VOI3 connection.

I have set up and run the model in which effects of task conditions (Color, Shape - I'm ignoring the Color+Shape condition) go into VOI4 which then modulates VOI1->VOI2 and VOI1->VOI3 connections. My question is: should I separate this into "two models" instead, namely 2 models with identical structure but with model 1 receiving data from for the Color condition only and model 2 receiving data from Shape condition only? The reason why I am asking this is because with the current model into which data from both Color and Shape conditions are fed, it is not possible to parse out the effects of VOI4 on VOI1-VOI2 and VOI1-VOI3 connections based on task conditions. As far as I can tell, in setting up the D matrix, it's not feasible to specify task conditions for the modulatory effects. 

I hope this question makes sense. I have consulted Stephan's nonlinear DCM paper (2008), particularly the binocular rivalry task portion but it isn't clear whether the same model was run for each task condition separately. It'd be great if someone could educate me on this.

Best,

Miranda E.