I’m still skeptical there’s enough data to distinguish from the obvious selection effects. The difference in safety between manual and automatic 911s probably tells you essentially nothing about whether a random driver would be safer driving a random automatic or a random manual.
It is hard to distinguish indeed, since there is no noticeable difference.
Jokes aside, as a data scientist of a long career, I don’t share your skepticism. It would be a hilariously easy task to extract this information from the data with the variables involved and the records insurance companies are able to keep.
