Dear Li Bei, There are clear answers to both these questions. However, they are more conceptual than practical. 1, Suppose there is an experiment including 2 conditions A, B and 1 baseline. When contrast A with B using t contrast, the activation of A may because of the deactivation of B related to baseline, not the activation of A. In this condition, should I mask t contrast A>B to eliminate the deactivation of B? I found in most of my experiment (similar experiment can be easily found in articles published), baseline > task always got some activations, but almost no one cared about this in their articles, could you tell me why and how to deal with this? 1) I do not think there is much to worry about here. One way of looking at this is to call the baseline condition C. At any voxel, one of the three conditions (A, B or C) will show the lowest signal. The only interesting things are the differences among A, B and C. The only special thing about C is that you assumed, a priori, that C would have the lowest activity. However, this assumption is not necessarily correct. There are many instances of event-related reductions in BOLD and rCBF. 2) If you are only interested in voxels where C has the lowest signal then (1) define regions that are activated in A and B relative to C, using a contrast like 1 1 -2. Then, (2) use the ensuing regions to constrain your search for differences between A and B (e.g., with a contrast -1 1 0). Note you can do this because the contrasts are orthogonal, This is why we sometimes include null events in event-related designs; so that we can define a responsive system that constrains the search for differential responses among the non-null trials. 2, Since the DCM is not exploratory, one must specify a correct DCM model to get a right answer. But because the rule for building a DCM model is vague(anatomic connections may not valid under a certain task, influences can pass through indirect connections between 2 brain regions), how could one know the model he built is correct? The result may be quite different because of a tiny modification of the model, I don't know how much reliance can be put on one model. Could you give me some suggestions? (the model proposed by andrea in 2004 is a very complex one, will a small modification change the result of the model??) J The answer here is yes. Changing the model will change the conditional inferences (that were conditioned on the original model). It is surprising or bad? No. The inference is explicitly conditioned on the model. Can you choose the best model? Yes, one uses Bayesian modal selection to do this. In short, there are two levels of inference in DCM. The first is conditional inference about the coupling parameters of a particular model. The second is inference made in terms of model selection or comparison, averaging over the coupling parameters of each model. With these two levels you can explore a model and the family of models from which your DCM came. People sometimes think that DCM will find the best model automatically and tell you how the brain works. It does not. DCM allows you answer well-formulated and specific questions that are operationally embodied by a model, or by the differences among specified models. It does not perform an exploratory search of model space. If you replace 'DCM' with 'question' then clearly the 'answer' will change with the 'question'. PS There is no such thing as a 'correct' DCM. A DCM can only be more or less likely than another DCM. Does this help? - Karl