Irene,
Such a small difference in p-value is not surprising for two main reasons:
** The LME model you adopted using lmer() is close to, but not exactly the same as, the one with 3dMVM. The MVM model is elaborated in the following paper: [
www.sciencedirect.com]
** Extracting and the averaging the voxels within an ROI would change the data structure to some extent. In other words, the two processes of averaging and modeling are not commutative.
Gang