I think as a general rule one should avoid using floating point math all together (if at all possible) on micro-controller applications, just too much code bloat and speed penalty to pay. For an applications like reading a load cell and converting to units of measurement it can all be done in integer math.
Agree on these 
However the point I wanted to make is that averaging the float values might be more accurate than the averaging the raw integers.
And sometimes one wants to trade speed for precision ...