> That still leaves the problem that few people look at insurance premium when choosing what car to buy.
It doesn't help that premium calculations are nonlinear and trade secrets. In the real world, it would take a computer and a large database to fuzzily estimate the impact of a particular car purchase on your personal premiums forecasted over the next few years with an error margin any less than a few hundred dollars per year (unless your life is particularly stable and well aligned to some major stereotype you can use to get a closer estimate).
If each insurer just published a table of the incremental impact of a given model of car (or better yet, how linear contributions for theft vs crash-rate vs death-rate-on-crash vs ...) then that'd be easy enough to use during purchasing. If you own a 90s civic in Oakland vs Redwood City though you're much more likely for the defective security measures to be used, and the insurers use a proxy for that information in their calculations, so in practice you have to get a personalized quote for every single car you might be interested in purchasing. Moreover, if you buy the car in a low-car-crime locale and move you can still be surprised by the massive rate hike [0]. And so on; modeling arbitrary risk is complicated, which is (part of) why professionals get paid the big bucks to do it. If there are other workable solutions, I'd prefer most of those to requiring the general public to have to do non-trivial math and statistics for every car purchase, especially above and beyond what they already have to do when estimating the total lifetime costs due to fuel economy or whatever.
[0] My personal solution was just to sell the car in that low-car-crime locale where it had a market value and buy a new vehicle in my destination, but then you're trading premiums for transaction costs, which isn't easy to model if you don't know how often you'll move in 5yrs either (hindsight, definitely worth it by a wide margin).