> -contr 1.0 -1.0 0.0 0.0 fear_situation_effect
> -contr 0.0 0.0 1.0 -1.0 anger_situation_effect
> -contr 1.0 1.0 -1.0 -1.0 A_main_effect
> -contr 1.0 -1.0 1.0 -1.0 B_main_effect

What you want for all the above tests of interest is a paired t-test type, but in your 3dANOVA script they are treated as two-sample t-test. For example, fear_situation_effect (A1B1 - A1B2) is supposed to be a paired t-test with each subject having both effects of A1B1 and A1B2 available.

> We realize that the assumption of independence is violated to some extent,
> but the violation does not strike me, at least, as serious, although I could be
> wrong. Furthermore, it would also be the case that such an approach would be
> relatively conservative, given that the within-subject variance that exists
> within groups is not being modeled.

Since A1B1 and A1B2 are from two different groups of subjects in your case, the problem with a two-sample t-test is that we can't assess the correlation between the two effects as you've realized. If the two effects are independent with each other, you're fine. If they're positively correlated, a two-sample t-test would be conservative. However, if they are negatively correlated, a two-sample t-test would be too liberal.

> -contr 1.0 1.0 -1.0 -1.0 A_main_effect
> -contr 1.0 -1.0 1.0 -1.0 B_main_effect

Even if you want to go with a two-sample t-test, you probably want to run the above two tests with 3dttest or 3dANOVA with (A1B1 - A2B2) and (A2B1 - A1B2) or (A1B2 - A2B1) from each subject.

HTH,
Gang