If you are passing correlation coefficients as inputs to 3dANOVA, then you are violating the assumptions of the analysis. In particular, CCs are not even approximately normally distributed, which could explain why your statistics are a little strangely distributed.
You could alleviate this problem by stretching the CCs to be more normal. The simplest way to do this, and the traditional one (going back to the Babylonians) is to take the arc-tanh of each CC. You can do this with 3dcalc, or in whatever program you are using to generate the CC datasets.