How dangerous are drinking drivers now?: replicating and updating Levitt and Porter

There were 10,322 alcohol-impaired driving fatalities in the United States during 2012, accounting for roughly 31% of all motor-vehicle related fatalities (NHTSA, 2013). In an influential study, Levitt and Porter (2001) estimate that alcohol-impaired drivers (those with a BAC level above the contemporary per selegal limit of 0.1) were 13 times more likely to cause a fatal motor vehicle crash. Although their findings continue to be widely cited in the drunk-driving literature, their analysis utilizes data that is now over two decades old. During that time period, there have been significant advances in automobile technology, improvements in road design, greater usage of consumer electronics while driving, and a trend toward decriminalizing marijuana for personal use. To the extent that these changes asymmetrically influence to probability of causing or surviving a motor vehicle crash, the relative risk drunk-drivers impose to sober drivers could be very different today.

Moreover, the data available to drunk-driving researchers has changed significantly. When Levitt and Porter conducted their analysis, a large percentage of drivers involved in fatal motor vehicle accidents were not administered a test to determine BAC. This introduces a serious concern about potential selection of drivers into their sample. Since that time, the proportion of drivers killed in a fatal motor vehicle accident who are tested to determine BAC has increased from 54% in 1982 to 68% in 1997 to 76% in 2008 (Hedlund, Ulmer, Northrup, 2004; Cassanova, Hedlund and Tilson, 2012). For surviving drivers, the proportion has increased from 16% in 1982 to 26% in 1997 to 29% in 2008 (Hedlund, Ulmer, Northrup, 2004; Cassanova, Hedlund and Tilson, 2012). Furthermore, researchers at the Department of Transportation have developed a multiple imputation strategy to accommodate missing BAC values in FARS that can be applied to all available incident reports from 1982 onward (Subramanian, 2002).

Given these changes in both driving technology and the available data, this paper conducts a thorough replication analysis of the original Levitt and Porter exercise and then applies their methodology to provide updated estimates of the relative risk and prevalence of drinking drivers. In a preview of our preliminary results, we find that it is not possible to replicate the results of Levitt and Porter as they describe their approach. Nevertheless, we find that despite using a different set of sample restrictions than are actually described, the results obtained by Levitt and Porter are not meaningfully affected by this choice. We also find that across a number of model specifications and definitions of alcohol-status, the relative risk of alcohol-impaired drivers has increased significantly over the past three decades: by approximately one-third from 1983-1993 to 2004-2013. Despite an increase in the relative-risk of alcohol-impaired driving, we find that the incidence and external cost of alcohol-impaired driving have both fallen.