They looked at fatal crashes only, which is presumably a very small share of all crashes. They also normalised to per mile driven using a sample of data they have - presumably some data on miles driven by car type.

Could be sketchy, could just be a much smaller sub-population.