First, I want to know when you use the 3dttest, Did you use the option -paired?
If you aren't using it, you are doing the usual two sample t-test. For each subject, you collect two sets of data for each condition. In this case, you are assuming that all the data are independent, even for the data from the same subject. In another other, you are ignoring the subject effect. If you compare the results from ANOVA, it will differ a lot if subjects have much difference.

If you are using -paired option, you are doing the matched paired t-test. It consider the subject effects. It is actually equivalent to the 2-way fixed effect ANOVA (Block design, look at subjects as block, assuming no interaction in this case). Unfortuately, you can't confirm this using 3dANOVA2 -type 1, because in 3dANOVA2, it always assumes the interaction term. Then for n=1, no replication data, it will say "sample size is too small for fixed effects model".

The ANOVA command you are using is 3dANOVA2 -type 3, you are considering subjects effect as random. In this case, with n=1, we will have the enough data to estimate the parameters.
For example B~N(0,sigmab^2), the only parameter is sigmab^2.
In this case you include not only factor B(subject effect) in the model, but alse the interaction of subject and factor A in the model.


So, if you aren't using -paried, then subject effect and interaction of subject and factor A will make the difference.
If you are using -paired, the interaction and assumption of randomness make the difference.
But I guess the big difference you said might due to the missing option -paired. Which one is reasonable to use? It depends on your assumption and the model you use.