I decided to test this. Using jspark's OpenSCAD version as a basis, I printed two different bearing mounts (623 bearings on 10mm hardened rod), one with the bearings running on their centres, one with them on their edge, as per the original. I had 400mm hardened chromed rods, which I ran each bearing back and forth on for 5 minutes, with a drill turning the rod at reasonable speed. Testing grip was a bit arbitrary; I did the preload up as tight as I dared, then pulled on the mount until moved. Findings: Grip on the rod seems proportional to the preload you can put on the bearing. With the bearing that ran on their edges, they ran rough if preload was too high. Grip was good, but gave way sharply. With the bearings that ran on their centres, much higher preload could be applied. Grip on the rod seemed equivalent (though preload had to be higher to achieve this), but seemed to lose grip more progressively. Wear on the rod, even a non-hardened drill rod, was negligable. Conclusion: depends on your application. I think running the bearings on their edges will have a shorter life (though this will depend on the quality of your bearings), and wear the rod they run on, but have slightly better grip. I'm going to set up a test to see how repeatable the accuracy is next.