Dear Pei Ling,
> How is it possible that contrast 1 at timepoint 2 yielded so much more activation than LR-REST?
Well, it's well a difference whether you test LR + REST – L - R > 0 or LR - REST > 0. You can think of various combinations of beta estimates that would lead to a finding like that - leaving aside whether they are reasonable.
What does REST refer to in your case? Is it a control condition with well-defined trials in which subjects see e.g. a fixation cross followed by some instruction (indicating that they don't have to perform a task), or is it a condition in which you model "everything that is not LR, L, R"?
Looking at the second time point.jpg, based on the pattern I don't recognize typical networks, and there might even be lateral ventricle / white matter "activation". In any case, very different patterns often result from artefacts, e.g. fast head motion, which can lead to effects much stronger than BOLD responses. Thus I would have a closer look at the data / realignment parameters instead of the model, especially as it seems to be a single-subject model based on the many degrees of freedom.
> How do you show under-additive effects?
Under-additive effects, meaning LR < (L + R), should show up for [-1 1 1 0].