Consider the normal crash dynamics (I haven't had a look but CAMI may have a video or two of sled tests on the net somewhere). Also run a search on "hic" and "crash dynamics" or similar to find lots of interesting stuff to read ...
(a) if the occupant is sitting normally in a lap belt only restraint, the initial crash load moves the body (in the sitting attitude) forward until the lap belt becomes effective in restricting further movement. The body then jack-knifes about the lap belt. In the typical environment, one expects significant head and lower leg impacts on the structures in front. The dummy test recordings indicate that limb flailing is a significant problem ..
(b) with upper body restraint (sash part of lap sash, or full harness), the upper body restraint slows the torso in its rotation .. with the generally accepted rule of thumb that head impact speed is reduced to around half of the unrestrained value.
(c) (you really need to see some sled test recordings to appreciate this comment) the idea of lowering the head probably reduces the likelihood of significant whiplash style injuries during the impact sequence.
.. me, I'd stick with the head forward, arms tucked into the side ...