I think it is better to chose your PE direction in such a way that the signal from the area you care the most is stretched, as opposed to compressed. If the signal is stretched, after you correct for the distortions using the field map, you combine the intensity from several voxels into a single one in the corrected image. Therefore, the intensity in the corrected image would be the correct one. Following Jesper's analogy, it would be equivalent to extracting the sample mean from the sample: you can always do it. The reverse, you can't: if the signal has been compressed into a single voxel, the best you can do is to split it evenly into the corresponding voxels in the corrected image. In that case, as Matt says, you correct for the geometrical distortions, but not for the intensity.
The short answer: yes, you would be better off using P>>A. (This is if you care about frontal cortex: there might be areas where the distortions results in a compression.)
PS: About the registration to a T1-weighted target: for functional data (T2*-weighted), note that on top of the distortions, there is a substantial dropout in orbito-frontal cortex. Say that, for example there will be 3-4 mm less of brain on that inferior part of frontal cortex, compared with the T1 image. Many, many times, automated registration methods will give you a resulting image tipped forward so that there are 2 mm "missing" on top and 2 mm on the bottom. This is especially true for lower resolution functional images (3 mm isotropic), since the image is very blurred and there are less anatomical landmarks. To check for this effect in the registration we recommend to check in the corpus callosum. I imagine you can get a more accurate registration if you use a mask generated from the T2* image as "inweight".