Thursday, July 20, 2017

Is My Driver Tracking App Giving Me Bad Advice?

I have an app on my iPhone that tracks my driving behavior. It's called "Drive Well", but there are numerous incarnations of the same concept. It uses my phone's accelerometer and location services to figure out where I'm driving, how I'm driving, how fast I'm driving, if I'm making too many hard breaks, and so on

The concept is fundamentally sound. These kinds of devices have been tested in the market for almost a decade now, and insurance companies and data scientists have been pouring over the collected data. The data truly are predictive of future auto accidents. If you make a lot of hard breaks, you're probably a careless driver. You aren't leaving enough space between your vehicle and the one in front of you, you respond too late to risks that more observant drivers would anticipate, etc. If you speed (above and beyond normal "speeding" that everyone does), you are more likely to cause accidents, and the accidents that you cause are likely to be more severe. If you corner too sharply, if you play with your phone while driving, if you accelerate too quickly, then you are more likely to have an accident. And, of course, the more miles you drive and the more time spent on the road, the more likely you are to get into an accident through sheer exposure to the risk. Someone who had analyzed this kind of data told me that it can replace an insurance company's entire rate plan. That is to say, directly measuring driving behavior is as predictive as measuring everything knowable about the person's demographics (age, gender, marital status, credit history, state and zip code, prior accident history, etc.). The predictive power of this information can be validated against enormous, statistically credible datasets. The signal is real. These things are predictive of future auto accidents.

Still, I think that the app sometimes gives me bad advice. I keep getting dinged for "rapid acceleration," and my app tells me I should cool it. It's true that I rapidly accelerate to get up to speed on the interstate, even in town. It's probably true that people who accelerate rapidly statistically have more accidents, and this generalization probably applies to me whether I'd like it to or not. Please don't think I'm special pleading that "In my case it's different." But in those particular situations where I am accelerating rapidly to match the flow of traffic, I would be less safe if I were to slow it down. In other words, the app is correctly telling me that I have a higher-than-average accident frequency (on this one count, not in total, as my other driving behaviors are good), but possibly giving me bad advice on how to improve my driving.

Someone familiar with this driving behavior data once told me that driving too slow is actually predictive of accident frequency. This makes sense. If someone is driving too slow, they are probably in an unfamiliar place, possibly looking for their exit or house number. Such a distracted driver is likely to cause an accident. But it would be wrong to say, "You should drive fast because slow driving correlates with high accident frequency." No, driving fast while you're distracted and confused probably makes it even more likely that you'll get into an accident. The statistical methods used in predictive models are not capable of teasing out what is causal and what is mere correlation.

I don't want to over-encourage this kind of second-guessing. Statistical methods often surprise us. And people have funny ways of convincing themselves that their vices are actually good for them. "My speeding is fine, it's the slow drivers who drive the actual speed limit who are causing the accidents!" (Loud annoying buzzer sound!) "I've driven home drunk many times and never had a problem. I'm actually safer, because I drive extra careful to avoid police scrutiny." (Loud annoying buzzer sound!) Don't be that guy. But a touch of skepticism is warranted. No statistical method will tell you which behaviors are merely correlational, such that changing them won't affect your accident risk (or will affect it adversely). Only human ingenuity and a cleverly designed test will be able to establish what is a truly causal effect.

No comments:

Post a Comment