Are you joking? It is a rate. It is cited in per 100,000 because that is the characteristic rate that results in low integer-scale values. Rates are scale-invariant so you could just as well cite it in per 1,000 if you desire, but then you get a fire fatality rate of 0.0085 per 1,000 for the Pinto which is a much more annoying number to understand and compare for regular people.
Frankly, it would have been better to just cite the inverse of 1 fire fatality per ~8609.5 Cybertrucks (I already discounted the suicide) versus the 1 fire fatality per ~117,536 Pintos. Does that help you understand the usage and comparison of rates?
Are you? The smaller the data set the more likely it's anecdotal and down to luck,
good or bad.
If a self-driving car drives one mile with zero accidents, would we extrapolate from there and say they have a perfect driving record? Because thats just how the number work out? No! We'd want to see how this hypothetical self-driving car does over a million or a billion miles, over a wide range of conditions, before drawing any conclusions.
Oh I see. You appear to be unfamiliar with how to evaluate failure rates.
You appear to be under the mistaken assumption that your sample size needs to be in proportion to the failure rate you hope to achieve to draw any conclusions. That is untrue. Your sample size only needs to be in proportion to the actual failure rate to draw conclusions.
If your self-driving car drives one mile with zero crashes, that provides almost no evidence to support a claim that your self-driving car has a crash rate of 1 per billion miles. However, if your self-driving car drives one mile and crashes 10 times, that provides tremendous amounts of evidence against the claim that your self-driving car has a crash rate of 1 per billion miles. This is despite the fact that both instances only have a sample of a single mile.
Proving, failure to prove, and disproving are not symmetrical at all. But do not worry, that is a common mistake made by people with no training in scientific or statistical analysis; just take this as a learning experience.
You’re making a valid point about the asymmetry in proving vs. disproving failure rates, but the condescending tone is unnecessary. That might be acceptable behavior back on Reddit or Kiwi farms or wherever you come from, but we strive for a higher level of discourse here.
I never said a sample size needs to match the target failure rate—only that proving an extremely low failure rate requires significant data. Yes, a high failure rate can be detected quickly, but that doesn’t mean a small sample is enough to confirm an ultra-low failure rate. Just because a car doesn’t crash in one mile doesn’t mean it won’t in the next billion. You’re oversimplifying the problem while assuming I don’t understand statistical inference.
Frankly, it would have been better to just cite the inverse of 1 fire fatality per ~8609.5 Cybertrucks (I already discounted the suicide) versus the 1 fire fatality per ~117,536 Pintos. Does that help you understand the usage and comparison of rates?