Hi Anne,

I think this is because you've over specified your design (it says the beta is not uniquely specified). If you were to subtract from regressor 2 and 3 from regressor 1  [2 -1 -1] then you would arrive at regressor 4 with a linear combination of the first 3 regressors in the present design.

In general, your design is valid (without regressor 4), but might not be very sensible depending on your DCM analysis. Note that there's only one difference between conditions (and comparing to the unmodeled fixation cross might not make very much sense especially for DCM). If the difference between A,B,C is parametric (for instance, reward of 1,5,10 $) it would make more sense to use one regressor. In case it is not and these are different conditions read below:

In a normal design with two condition you would do DCM as follows:

I would extract VOIs from a Condition 1 AND Condition 2 AND Condition 3 > unmodelled fixation cross [1 1 1] as you do here and then extract one or several areas that are activated by condition 1 > condition 2 & 3 (or any permutation of these numbers) to see whether that condition 1 > condition 2(&3) regressor modulates a connection between those VOIs. There is only one difference between condition 1, 2 and 3 and so having several regressors to modulate activity means that you over parameterize your model.

I don't know about your particular design, but say you have houses and faces and a fixation cross in your design. Then the best design would be to extract V1 from a visual contrast that is Houses AND faces > fixation cross [1 1 -2] or [1 1] if the fixation cross is not modelled. Then you would extract the fusiform face area from the Faces > Houses contrast. And the modulator should be Faces > Houses and not faces [1] and houses [0 1]. I have the feeling what you're doing right now is analogous to looking at Faces > Fixation cross and Houses > Fixation cross in your design and then you have connections modulated by these single condition regressors ([1 0] for condition 1 and [0 1] for condition 2 in case your fixation is an implicit baseline). And then you want see how connections are modulated by both these conditions and compare them somehow. In this case you probably want to get at the difference in conditions directly with a parametric modulator that compares condition 1 > 2 (& 3) ([2 -1 -1]) and then you can say that a connection is more modulated as you move from condition 1 to 2, if the modulation is positive. Crucially, as you know, in DCM it's only sensible to extract active areas. If you find and extract a blob from contrast of condition 1 > 2, then that should also be the modulator.

I hope this helps.

Hauke :)