The problem is deaths are so infrequent that it takes many years to get a statistically representative sample. Perhaps at-fault collisions per mile would be a better metric?