Introduction
While the effectiveness of seat belts is well established in research (Blincoe, 1994, Carpenter and Stehr, 2008, Crandall et al., 2001, Evans, 1986, Kahane, 2000, Klein and Walz, 1995, Levitt and Porter, 2001, McCartt and Northrup, 2004, Partyka, 1988, Partyka and Womble, 1989), U.S. seat belt usage rates have been relatively low compared to other developed nations (NHTSA, 2007).1 In 1997, the federal government set targets to increase seat belt usage from 68% in 1996 to 85% by 2000, and then to 90% in 2005, both of which went unmet according to the annual National Occupant Protection Use Survey (NOPUS).2 The National Highway Traffic Safety Administration (NHTSA) of the U.S. Department of Transportation reported that over half of passenger vehicle occupants killed in traffic accidents in 2006 died unbuckled (NHTSA, 2007).
With the rise in federal funding for highway safety initiatives and awareness programs, and more stringent primary and secondary law enforcement at the state level, the failure to meet targeted usage rates is confounding. But more importantly, it points to a need for targeted policies to incentivize usage. Before we can design such policies it is critical to determine the factors that affect vehicle occupants’ decision to wear a seat belt. In the past, seat belt effectiveness studies that used NOPUS data could not address many of these factors because of the lack of certain details in the NOPUS data.
Research on seat belt usage typically utilizes one of two publicly available data sources: (a) National Occupant Protection Use Surveys (NOPUS), and (b) Fatality Analysis Reporting System (FARS). The National Center for Statistics and Analysis of the NHTSA consider NOPUS to be their most reliable data set tracking the trends in seat belt usage by motorists. However, the observational nature of NOPUS data not only subjects it to some limitations due to the probability of human error in the data collection, but also due to lack of reliable data on vehicle occupants’ personal characteristics or nighttime travel behavior.
Fatality Analysis Reporting System (FARS) is the other available database for evaluating the usage rates of occupant’s protection devices.3 An advantage of using the FARS database over NOPUS is that it is more comprehensive in the reported variables, providing, for example, additional data for vehicle occupant characteristics, as well as nighttime data. However, one critical problem with FARS data is that it underestimates seat belts usage when compared to estimates obtained from observational data such as NOPUS4 (Salzberg, Yamada, Saibel, & Moffat, 2002) due to the nature of the reporting system. To be included in FARS, a crash must result in the death of a person (occupant of a vehicle or a non-occupant) within 30 days of the crash. Since it lists only those crashes where there is at least one fatality, there is a potential issue of sample selection given that an individual’s seat belt use affects his or her probability of death, which in turn influences whether the crash is included in the data because of the correlation between seat belt use and fatality. It has been shown that such sample selection leads to biased regression coefficient estimates (Angrist and Krueger, 1999, Heckman, 1979, Heckman et al., 1996). The extent of this sample selection bias becomes even more significant when we consider that only about 0.5% of motor vehicle crashes involve a fatality, and in 90% of the incidents there is just a single death (NHTSA, 1998). Had that death not occurred, the crash would not be included in the FARS database. Therefore, empirically, the impact of sample selection can be substantial, and failing to account for it leads to estimates that systematically understate seat belt usage. Previously, Salzberg et al. (2002) investigated seat belt usage rates by comparing FARS data with observation surveys and concluded that unbelted occupants are over-represented in fatal collisions for two reasons: (a) because of a greater chance of involvement in potentially fatal collisions in the first place, and (b) because they are not afforded the protection of seat belts when a collision does occur. Their model focused on risk but they did not address the sample selection bias.
In our 2009 paper (Islam & Goetzke, 2009), we used an identification method developed by Levitt and Porter (2001)5 to correct for the sample selection bias in FARS data in order to obtain a credible estimate of seat belt usage in the United States. Curiously, but rather insightfully, the sample selection problem that arises because of the exclusion of a majority of non-fatal crash statistics from the FARS data set is countered by further limiting the data. We used two different specifications to correct the sample selection bias in FARS data: (a) correction based on strict independence of seat belt choice (Model 1, details of which are described in Section 3.1 below), and (b) correction based on strict dependence of seat belt choice (Model 2, described in Section 3.2 below). By applying these corrections, we showed that the FARS database can be established as an acceptable and comparable alternative to the observational NOPUS data (Islam & Goetzke, 2009). In this paper, we extend our previous work (Islam & Goetzke, 2009) on correcting the sample selection bias evident in the FARS dataset and using a multivariate regression analysis, address the following critical question: What factors affect the decision to use seat belts? This question is critical in designing effective policies. For example, are there particular socio-demographic groups that policy should focus on? Do specific times of day or regions require heightened enforcement? These specific policy-relevant questions are precisely what the corrected FARS dataset allows us to answer due to the rich cross-sectional detail that it contains on each reported event.
