“How Dangerous Are Drinking Drivers?”, Steven D. Levitt, Jack Porter2001-12 (economics, alcoholism, statistical comparison)⁠:

We present a methodology for measuring the risks posed by drinking drivers that relies solely on readily available data on fatal crashes. The key to our identification strategy is a hidden richness inherent in two-car crashes.

Drivers with alcohol in their blood are 7× more likely to cause a fatal crash; legally drunk drivers pose a risk 13× greater than sober drivers. The externality per mile driven by a drunk driver is at least $0.52^$0.3₂₀₀₁.

At current enforcement rates the punishment per arrest for drunk driving that internalizes this externality would be equivalent to a fine of $13,820.65^$8,000₂₀₀₁.

…The ability to identify the parameters arises from a hidden richness in the data due to the fact that crashes often involve multiple drivers. For two-car crashes, the relative frequency of accidents involving two drinking drivers, two sober drivers, or one of each provides extremely useful information. Indeed, given the set of assumptions outlined in Section II, this information alone is sufficient to separately identify both the relative likelihood of causing a fatal crash on the part of drinking and sober drivers and the fraction of drivers on the road who have been drinking. The intuition underlying the identification of the model is quite simple. The number of two-car fatal crash opportunities is dictated by the binomial distribution. Consequently, the number of fatal two-car crash opportunities involving two drinking (sober) drivers is proportional to the square of the number of drinking (sober) drivers on the road. The number of fatal crash opportunities involving exactly one drinking and one sober driver is linearly related to the number of both drinking and sober drivers. Identification of the model arises from these intrinsic nonlinearities. These nonlinearities are not artificially imposed on the problem via arbitrary functional form assumptions, but rather are the immediate implication of the binomial distribution, which relies only on the assumptions to be stated in Section II concerning independence of crashes and equal mixing of the different types on the road.

Applying the model to data on fatal accidents in the United States over the period 1983–93, we obtain a number of interesting results. Drivers identified by police as having been drinking (but not necessarily legally drunk) are at least 7× more likely to cause a fatal crash than drivers with no reported alcohol involvement. Drivers above the blood-alcohol limit of 0.10 are at least 13× more likely to be the cause of fatal crashes. When we apply the model to other observable traits, males, young drivers, and those with bad previous driving records are also more likely to cause crashes. Drinking, however, is far more important than these other characteristics, and much of the apparent impact of gender and past driving record actually reflects differential propensities to drink and drive across groups. The exception is young drivers: sober, young drivers are almost 3× as likely to cause a fatal crash as other sober drivers. The peak hours for drinking and driving are between 1:00 a.m. and 3:00 a.m. when as many as 25% of drivers are estimated to have been drinking. The proportion of drinking drivers appears to have fallen by about 1⁄4^th over the course of our sample. The relative fatal crash risk of drinking drivers, in contrast, appears to have been stable.