Has the great recession and its aftermath reduced traffic fatalities?

Accident Analysis and Prevention 98 (2017) 130–138

Contents lists available at ScienceDirect

Accident Analysis and Prevention journal homepage: www.elsevier.com/locate/aap

Has the great recession and its aftermath reduced traffic fatalities? Robert B. Noland ∗ , Yuhan Zhou Alan M. Voorhees Transportation Center, Edward J. Bloustein School of Planning and Public Policy, Rutgers University, New Brunswick, NJ 08901, United States

a r t i c l e

i n f o

Article history: Received 2 August 2016 Received in revised form 29 August 2016 Accepted 10 September 2016 Keywords: Traffic safety Economy Recession Policy Negative-binomial model Panel data

a b s t r a c t An analysis of state-level data from 1984 to 2014 provides evidence on the relationship between economic recessions and US traffic fatalities. While there are large reductions associated with decreases in household median income, other policy variables tend to have additional and in some cases, larger effects. An increase in the inequality of the income distribution, measured by the Gini index, has reduced traffic fatalities. Graduated licensing policies, cell phone laws, and motorcycle helmet requirements are all associated with reductions in fatalities. Other factors include a proxy for medical technology, and access to emergency medical services (based on the percent of vehicle miles traveled in rural areas); reductions in the latter accounted for a substantial reduction in fatalities and is likely another indicator of reduced economic activity. Changes in the road network, mainly increases in the percent of collector roads has increased fatalities. Population growth is associated with increased traffic fatalities and changes in age cohorts has a small negative effect. Overall, results suggest that there has been a beneficial impact on traffic fatalities from reduced economic activity, but various policies adopted by the states have also reduced traffic fatalities. © 2016 Elsevier Ltd. All rights reserved.

1. Introduction Traffic-related fatalities in the US have declined dramatically since about 2007. Previous research has demonstrated that this effect is partially due to the economic recession that began about that time (Noland and Sun, 2015). The US is not the only country that has experienced substantial declines in traffic fatalities during this time period. Studies have found a correlation in European countries (Yannis et al., 2014) and evidence exists of a similar effect in Great Britain (Lloyd et al., 2016). As the economy has slowly recovered, there is a concern that traffic fatalities may increase to previous levels, and despite the recent reduction the US continues to trail other developed countries that have achieved large reductions in traffic fatalities since the 1970 s (Evans, 2014). Adding additional years of data to the previous analysis in Noland and Sun (2015), the impact of economic conditions is evaluated while controlling for other policies and demographic trends that can affect traffic-related fatalities. These other policies include a range of vehicle safety regulations, safety belt laws, greater enforcement and stiffer penalties for driving while intoxicated, improved emergency management

∗ Corresponding author. E-mail address: [email protected] (R.B. Noland). http://dx.doi.org/10.1016/j.aap.2016.09.011 0001-4575/© 2016 Elsevier Ltd. All rights reserved.

systems, and more recently graduated licensing and cell phone laws (Noland, 2003b; Shope, 2007). Traffic fatalities and crashes have long followed the business cycle, and fatalities have recently increased in line with an improving economy. Early estimates from NHTSA are that 2015 will see an increase in traffic fatalities of 7.7% (NHTSA, 2016a) and the National Safety Council is estimating about a 9% increase for 2016 (National Safety Council, 2016). Despite this, there has been an overall downward trend in fatalities since the peak year of 1972 when 55,000 traffic fatalities occurred. A negative-binomial panel model is developed with key control variables, including measures of the economy, demographic variables, infrastructure variable, proxies for medical technology, and a set of dummy variables for when key policies were implemented in each state. This allows us to examine the effectiveness of policies implemented during this time period and to isolate the effect of the economy on traffic fatalities. This work is similar to the earlier work of Noland (2003b) and also extends the analysis presented in Noland and Sun (2015) with additional years of post-recession data.

R.B. Noland, Y. Zhou / Accident Analysis and Prevention 98 (2017) 130–138

2. Background and hypotheses formulation 2.1. Theoretical formulation Policies designed to reduce traffic fatalities must consider the behavior of drivers and how they interact with the road environment. This can be considered in a utility framework that provides a trade-off between the mobility and safety desires of individuals (Noland, 2013). For a given level of technology and policy enforcement, drivers will maximize their utility with this trade-off in mind, along with other potential costs associated with travel. This may involve decisions on speed choice, whether to travel or not, and the conditions under which one decides to travel (e.g. in adverse weather or when fatigued). There is a large body of literature that examines these issues, as reviewed by Noland (2013) and going back to the work of Peltzman (1975) and Wilde (1982), commonly referred to as either risk compensation or risk homeostasis, but also described as behavioral adaptation. In analyzing trends in traffic fatalities, it is important to consider how these trade-offs might affect actual outcomes of policies. While Peltzman (1975) argued that crash outcomes could worsen in response to various regulatory policies, this view has largely been supplanted by a more nuanced perspective that sees behavioral adaptation as an off-set in risk reduction, from a strictly deterministic and inelastic calculation of the likely benefits. For example, while one might be able to estimate that airbags will save a fixed number of lives, based on the physical properties of the human body and the characteristics of certain crash types, this would not account for how drivers may change their driving behavior because those crashes are now less dangerous, perhaps leading to a slight reduction in the safety benefits of the policy. As indicated by Noland, (2013) there may still be a mobility benefit from safety policies that result in behavioral adaptation, in addition to reduced crashes or crashes with less severe outcomes. This theoretical formulation can help to inform the hypotheses associated with the effect of various policies. In general, it is important to consider how various safety policies may also affect mobility, perhaps leading to less reduction in fatalities than anticipated. Some have also argued that economic recessions will influence risk taking behavior, such as a reduction in speeding and driving while intoxicated (Lloyd et al., 2016). The mix of drivers on the road may also change. This was analyzed by Maheshri and Winston (2016) using insurance data for the state of Ohio and they found that riskier drivers drove less during the recent economic downturn, compared to safer drivers. With these theoretical issues in mind, the key policies and factors hypothesized to be associated with traffic fatalities are discussed below, and linked to measures that can best proxy these effects in a regression analysis. 2.2. Economic factors The relationship between national income and traffic crashes and fatalities is well known and has been shown to be statistically significant in national crash models (Noland, 2003b). Most studies have typically measured economic activity with broad measures of Gross Domestic Product (GDP) or per-capita income (Kopits and Cropper, 2005; Noland, 2003b). However, these may not be good measures for picking up the actual impacts of the economy on individuals and in states, mainly because of increasing disparities in income (Frank, 2009; Piketty and Saez, 2001). While we cannot directly measure how income disparities in the population affect travel behavior there are some possible reasons. Households may defer new vehicle purchases and use and maintain older (and presumably less safe) vehicles for longer periods of time. They may also find it necessary to commute longer distances to find suitable employment if income levels are lower (Maheshri and Winston,

131

2016). We control for income and income disparities using average median income and a measure of state-level Gini coefficients.1 One factor associated with economic activity may be that the type of vehicles using the road network change during a recession. Freight transportation (by truck) decreases; seasonally-adjusted truck tonnage declined by 15% between January 2008 and April 2009, based on the truck tonnage index.2 According to NHTSA (2013) there were 3921 fatalities associated with heavy-duty trucks in 2012; this is about 11.6% of all fatalities, thus one would expect reductions in truck traffic to lead to reduced fatalities. This is controlled for in our data by a rough measure of the percent of trucks registered in each state. 2.3. Policy variables 2.3.1. Graduated licensing Graduated licensing programs have been enacted in almost every state since implementation incentives were put in place in 1998.3 Graduated licensing aims to protect younger drivers by conditionally restricting various riskier driving behaviors. Typical restrictions include not driving with other teenage passengers, forbidding nighttime driving (usually after midnight but with some exceptions if driving to a job), and the requirement to have no citations for at least 12 consecutive months (NHTSA, 2004). States have enacted these with large degrees of variation in the specific rules and conditions. The effectiveness of graduated licensing in reducing traffic fatalities is largely acknowledged. A review by Shope (2007) concluded that these programs have generally led to a reduction of 20–40% in the crash risk of the youngest drivers. A review of some early programs, including in New Zealand, found 7–8% reductions in crash injuries of teenagers (Foss and Evenson, 1999). Restrictions on nighttime driving and carrying passengers reduce the risk of nighttime crashes and those with large groups of teenagers in the car; research has also found a 23-25% percent reduction in injury and fatalities due to night time restrictions (Foss and Evenson, 1999; McKnight and Peck, 2002; Williams, 2007). Graduated licensing can also decrease the benefits that younger people obtain from driving, and some have found that the increased impediment to driving reduces the mobility of teen drivers (Ralph et al., 2014), consistent with the theoretical framework that reduced mobility can improve safety. To control for the implementation of these programs in different states, we use data on when states implemented both passenger and nighttime driving restrictions.4 We do not control for the variation in program design, such as age limits, provisional time limits, or hours in which restrictions are in effect, among others. 2.3.2. Cell phone laws An emerging issue in road safety is the use of cell phones, either for phone calls, texting, tweeting, or email and how this increases the risk of crashes. Research has found a slowing of reaction time associated with cell phone use, whether hand-held or hands-free (Caird et al., 2008; Charlton, 2009). The relative risk of cell phone users compared to those who do not use them is higher (Laberge-

1 The Gini coefficient measures the distribution of income of the residents of a country, state, or other region for which income data is available. It varies between 0 (all incomes are equal) and 1 (complete inequality). 2 Based on US Department of Transportation, Bureau of Transportation Statistics data available at: http://www.transtats.bts.gov/osea/seasonaladjustment/ ?PageVar=TRUCK 3 Incentives were mandated by the Transportation Equity Act for the 21st Century (TEA-21) of 1998. 4 For each law that we use as a policy variable, we count the year implemented if the law was in effect for at least half the year.

132

R.B. Noland, Y. Zhou / Accident Analysis and Prevention 98 (2017) 130–138

Nadeau et al., 2003). Studies have also shown that younger people may be more likely to use cell phones while driving, especially texting (Harrison, 2011). Cell phone use is thus a potential source of increasing crashes (Bhargava and Pathania, 2013). Legislation that forbids cell phone use while driving can help to mitigate these effects; our data includes the year that legislation was enacted in each state to ban the use of cell phones while driving, although we make no distinction in the details of the legislation (i.e., whether this includes talking or texting). 2.3.3. Motorcycle helmet laws In 2014 there were 4586 motorcyclist fatalities accounting for 14% of all traffic fatalities (NHTSA, 2016c). One of the most effective approaches for reducing motorcycle fatalities is the use of a helmet. The effectiveness of comprehensive motorcycle helmet laws is well established (Branas and Knudson, 2001). Research has shown that when these laws are made less stringent, fatalities increase (Kyrychenko and McCartt, 2006). Estimates are that in 2014, 1669 lives were saved from using a helmet and that potentially an additional 661 lives could have been saved with universal helmet usage (NHTSA, 2016b). Currently, 19 states and Washington, DC have universal helmet laws requiring their use while most other states have a partial helmet law (usually applying to just younger riders). Three states still have no helmet requirement (IIHS, 2016), and many states have repealed more restrictive laws at various points in time. Many motorcycle fatalities are due to crashes with other vehicles, thus, while our analysis is not focused on motorcycle fatalities, they are included in our total motor-vehicle fatality statistics, and we control for whether a state has a universal motor-cycle helmet law. 2.3.4. Driving while intoxicated Laws aimed at reducing drunk driving are very effective (Eisenberg, 2003; Noland and Karlaftis, 2005; Voas et al., 2000). These laws include administrative license revocation (that is, the ability to remove the license from an individual at the time of the drunk driving offense) and tighter limits on blood alcohol concentration (BAC) to 0.08% (from 0.10%). Some states with graduated licensing laws have a 0% tolerance for those under 21. We use data on alcohol consumption per capita as a control variable rather than the laws, as this provides a more direct relationship to driving while intoxicated. It may also partially control for any economic effects that reduce the consumption of alcohol during a recession (Lloyd et al., 2016). 2.3.5. Safety-belt laws State laws requiring safety-belt use have been very effective at decreasing traffic fatalities (Rivara et al., 1999). In the US, these have taken the form of both “primary laws” and “secondary laws”. The latter allow enforcement only if the driver is cited for another offense, while the former allow the driver to be pulled over if observed driving without a safety-belt. Over time, most states have adopted primary laws. Safety-belt use in the US nationwide is estimated to be about 88% as of 2015 (NHTSA, 2016d). The lowest rate of any state, at 69.5%, is in New Hampshire, the only state without a safety-belt law for adult drivers. Rates in most states exceed 80%, suggesting that safety-belt laws largely implemented in the 1990 s have been effective at increasing use rates. Our analysis includes a dummy variable for primary safety belt laws. 2.3.6. Trauma and medical care Improvements in emergency medical services (EMS) and trauma care have produced large reductions in traffic fatalities (Nathens et al., 2000; Noland, 2003a; Noland and Quddus, 2004 Noland, 2003a; Noland and Quddus, 2004). Statewide systems of trauma care have been shown to lead to reductions in fatalities (Nathens et al., 2000). Fatalities in rural areas are partly due to lack

of sufficient medical resources and the speed of obtaining emergency care (Maio et al., 1992). Controlling for changes in medical care and technology requires the use of proxy variables. We use the same proxy variable as in We use the same proxy variable as in Noland (2003b) and also used in Noland and Sun (2015), white infant mortality rates by state, as this is a good indicator of the most advanced care available. We also control for the fraction of vehiclemiles of travel (VMT) occurring in rural areas which may partially proxy for access to emergency services.

2.4. Control variables Two demographic trends that likely affect the number of traffic fatalities are the movement of populations towards metropolitan areas from more rural communities and an older population with fewer young people. The latter are well known to face greater driving risks, either due to inexperience or more risk taking behaviors (Jonah, 1986). An unknown question is whether a greater proportion of older people driving vehicles may affect risk in the future. While this age group is generally safer than the youngest age group, the ability to drive is likely to be impaired with age. While some individuals may compensate by driving less or, for example, just during daylight hours, others may not (De Raedt and Ponjaert-Kristoffersen, 2000; Hakamies-Blomqvist, 1994; Noland, 2013). Population data is used to control for both younger and older age groups in each state. The safety of vehicles continues to improve over time. New crash-avoidance technologies are being added to vehicles, airbags are standard equipment, including many cars having side airbags, and anti-lock braking systems are also pervasive. Sport-utility vehicles that tend to have roll-over crashes are less common, many being replaced by cross-over utility vehicles which are built on a car chassis rather than a truck chassis, improving stability. Detailed data on these changes and their prevalence in the vehicle fleet, however, are not readily available. But these and other unmeasured effects can be controlled for using a time trend variable (either a linear year control or year dummy variables). The design of road infrastructure can also have an impact on crash and fatality rates (Albalate et al., 2013). It is well known that controlled access freeways tend to have the lowest fatality rates per VMT, while lower class roads, such as high-speed arterials, tend to have higher risk (Noland, 2003b). Alternatively, traffic calming in urban areas can also reduce the risk of crashes and create a safer pedestrian environment. In a nationwide analysis it is not possible to explicitly control for these detailed effects, but various data on road classes, their lane widths, and the extent of networks can serve as proxies. These tend to be highly collinear with each other as demonstrated in previous analysis (Noland and Sun, 2015). In this analysis we include more selective variables, namely the total lane miles, percent of lane miles classified by functional type, and the average number of arterial lanes. One control variable included in our previous analysis was the variation in climate between states (Noland and Sun, 2015). Severe storms, while making travel more dangerous, can have the effect of reducing travel, and thus potentially lead to fewer fatalities (Eisenberg, 2004), though most studies still show a positive association (Qiu and Nixon, 2008). These studies require detailed micro data to assess correlations between weather events and crash risk; measuring a state-level average is problematic given the large variation in weather patterns within a state. Thus, while our previous work in Noland and Sun (2015) included the average number of days with precipitation exceeding 50 mm, we have opted not to include a climate variable in our analysis. While this variable was

R.B. Noland, Y. Zhou / Accident Analysis and Prevention 98 (2017) 130–138

133

negative and statistically significant in some models, the substantive impact was trivial when examining marginal effects.5

to more easily interpret the impact of various policies on traffic fatalities.

3. Data and modeling approach

4. Results and discussion

The data used for this study is at the state-level. This allows us to estimate cross-sectional time-series models. We have attempted to build as complete a dataset as possible, but there are various omissions and limitations with some of this data. The extent of the data and the source of each data series are listed in Table 1. The dependent variable is total fatalities; these were downloaded from the National Highway Traffic Safety Administration (NHTSA), Fatality Analysis Reporting System (FARS). The data was available for every state up to 2014 for a total of 31 years (1550 observations). We do not include data for the District of Columbia. Two of the variables, the Gini coefficient and white infant mortality are only available through 2013. The Gini coefficient data are based on estimates derived from Internal Revenue Service tax data (Frank, 2009).6 This limits our modeling, when these variables are included to 30 years of data.7 White infant mortality data has omissions for some states, and no data is available for 1998 and 1999. Some missing data was interpolated for 2012. We follow the approach used by (Noland, 2003b) and (Noland and Sun, 2015) to model cross-sectional time-series data. This is a fixed effect negative binomial model, as originally specified by (Hausman et al., 1984). Using fixed effects allows one to control for unmeasured state-specific effects that may also influence traffic fatalities. A negative binomial model is needed when the dependent variable is a count, as traffic fatalities are. These are Poisson distributed and thus do not follow the normality conditions needed for ordinary least squares regression.8 The conditional fixed-effects negative binomial model of (Hausman et al., 1984) has been shown to not be a true “fixedeffects” model (Allison and Waterman, 2002; Guimarães, 2008). To fully remove the individual fixed-effects in this model, there must be equivalence between the individual fixed-effects and the log of the dispersion parameter. Various alternative estimation procedures are recommended by (Allison and Waterman, 2002). These are either the unconditional fixed-effects negative binomial model (i.e., with explicit introduction of dummy variables for each crosssection) or the fixed-effects Poisson model, with corrections to the standard errors to account for over-dispersion. However, both of these may also suffer from the “incidental parameters” problem, leading to biased estimates. A score test developed by (Guimarães, 2008) allows one to test the null hypothesis of equivalence between the individual-fixed effects and the log of the over-dispersion parameter, and has been applied by (Marques and Pinho, 2014). We used this test and could not reject the null hypothesis, thus we present results using the conditional fixed-effects negative binomial model.9 We estimate the model with logs of the independent variables. The functional form of the negative binomial model allows the coefficient estimates to be interpreted as elasticities, providing a means

Model results are shown in Tables 2, 3 and 5. All models only include data through 2013, due to values for the Gini coefficient not being available for 2014.10 An adjusted McFadden 2 is reported to compare models and we estimate the variance inflation factor (VIF) to examine multi-collinearity, as well as individual correlation coefficients. We do not include all the possible infrastructure variables as was done in (Noland and Sun, 2015) due to their high level of multi-collinearity. The score test is also included in each table and we find we cannot reject the hypothesis that the overdispersion parameter is equal to the individual fixed effects, and thus the conditional fixed effects model is appropriate (Guimarães, 2008). Our first set of models are presented in Table 2. Only the median income and Gini coefficient are included, as well as exposure variables (population or VMT). Median income is both positive and statistically significant; that is, a better economy is associated with more traffic fatalities. The Gini coefficient has a negative and statistically significant coefficient. A larger Gini index value means that there is an increased concentration of income (i.e., less evenly distributed). Thus, as incomes become less equal there is a reduction in traffic fatalities. These models also test population and VMT as exposure variables. Both are positive and statistically significant. These cannot be included simultaneously due to their high level of correlation. Our preference is to use population as the exposure variable in the subsequent models as it likely contains less error at the state level. The year trend variable is negative and statistically significant indicating that some unmeasured factors are driving total fatalities down (year dummies are not shown, but would lead to the same conclusion). When using the year trend, the median income and Gini index parameters are inflated, possibly due to some correlation with the year trend. The Gini index has an R = 0.64 with the trend, but median income is only R = 0.23. Our VIF analysis does not suggest a problem and is actually lower than the estimates with year dummies. In any case, we estimate the remaining sets of models both with and without year dummies. Additional variables are added to the models in Table 3. Models 3a and 3b include the percent of the population aged 15–24 years and over the age of 75 years. The former are considered a riskier population group and results confirm that an increase in the younger age group increases total fatalities. The opposite is true for the older age group which has a negative coefficient (in model 3a only significant at the 90% confidence level). Model 3b with year dummies provides stronger results as there is some correlation of both variables with the year trend (R = −0.41 for the 15–24 age group, and R = 0.32 for the over 75 age group). Policy variables are added in Model 4a and 4b. Safety-belt laws were initially passed starting in the late 1980’s. We include only the passage of primary safety-belt laws in the analysis. Surprisingly this variable is not statistically significant. Most prior work has shown these laws to be effective (Noland, 2003b; Voas et al., 2000). One reason for the lack of statistical significance may be that safety-belt usage rates have increased substantially since the 1980’s and are uniformly high (NHTSA, 2016d). Thus, these laws have been very effective at increasing safety-belt usage, and the variation between states and over time is masked by the high usage rates. This does not mean that removing these laws would have no impact on fatalities,

5

An updated dataset of weather patterns at the state level was also unavailable. The Gini coefficient data available at the source listed in Table 1 was updated and corrected since the previous analysis in (Noland and Sun, 2015). The new data series provides more robust results than previously. 7 Models tested without the Gini coefficient did not show any substantive differences in results. 8 We tested a variety of transformations of the dependent variable using both the Shapiro-Wilk and Shapiro-Francia normality tests. These include the log of total fatalities, and fatality rates (by population and by VMT), both logged and unlogged. In all cases, the null hypothesis of normality is rejected. 9 We thank Paulo Guimarães for providing his Stata code for running the score test, fenbtest. This required the use of an older version of Stata (version 9). 6

10 Models estimated without the Gini coefficient and through 2014 provided essentially similar coefficient results.

134

R.B. Noland, Y. Zhou / Accident Analysis and Prevention 98 (2017) 130–138

Table 1 State level data, years available and source. Variables

Year

Source

URL

Dependent Variable Fatalities

1984–2014

NHTSA/FARS

http://www-fars.nhtsa.dot.gov/

Infrastructure Variables Route miles Lane miles Lane widths

1984–2014 1984–2014 1984–2014

FHWA, Highway Statistics/Table HM20 https://www.fhwa.dot.gov/policyinformation/statistics.cfm FHWA, Highway Statistics/Table HM60 https://www.fhwa.dot.gov/policyinformation/statistics.cfm FHWA, Highway Statistics/Table HM53 https://www.fhwa.dot.gov/policyinformation/statistics.cfm

Median income Gini index

1990–2004 2005–2014 1984–2014 1984–2013

CDC WONDER United States Census Bureau FRED Economic Data

http://wonder.cdc.gov/ http://factfinder2.census.gov/ https://research.stlouisfed.org/fred2/ http://www.shsu.edu/eco mwf/inequality.html

Policy Variables Safety belt law Graduate driver license law Motorcycle helmet law Cell phone law

1984–2014 1984–2014 1984–2014 1984–2014

IIHS IIHS NHTSA/FARS

Alcohol consumption

1984–2014

Hands Free Info NIH Surveillance Report/Table 4

http://www.iihs.org/iihs/topics/laws/safetybeltuse http://www.iihs.org/iihs/topics/laws/graduatedlicenseintro http://www.iihs.org/iihs/topics/laws/helmetuse/ http://dx.doi.org/10.1016/j.amepre.2011.02.024 http://handsfreeinfo.com/index-cell-phone-laws-legislation-by-state/ http://pubs.niaaa.nih.gov/publications/surveillance104/CONS14.htm

FHWA, Highway Statistics/Table VM2 United States Census Bureau National Vital Statistics Reports FHWA, Highway Statistics/Table MV-1

https://www.fhwa.dot.gov/policyinformation/statistics.cfm https://www.census.gov/ http://www.cdc.gov/nchs/data/nvsr/nvsr60/nvsr60 05.pdf http://www.cdc.gov/nchs/data/nvsr/nvsr64/nvsr64 09.pdf https://www.fhwa.dot.gov/policyinformation/statistics.cfm

Socio-economic Variables Population by age group

Control variables Vehicle Miles Traveled (VMT) 1984–2014 White infant mortality 1984–2007 2008–2013 Percent truck registered

1984–2014

Table 2 Conditional negative binomial models, with median income and Gini coefficient. Deaths (1984–2013)

Model 1a Coef.

log (median income) log (Gini index) log (total population) log (total VMT) Constant Year Year dummies N initial log likelihood final Log likelihood VIF adjusted ␳2 Score Test Hausman Test

Model 2a z

0.838 −0.660 0.378

18.52 −3.16 11.04

20.991 −0.015 No 1500 −14132.14 −7909.24 1.41 0.44 200.04 −120.41

20.22 −22.69

P = 0.000

Coef.

Model 1b z

0.560 −1.331

12.73 −6.69

0.552 33.249 −0.020 No 1500 −11095.80 −7806.66 1.41 0.30 182.02 −112.02

19.06 27.83 −29.08

but does suggest that drivers are more conditioned to using safetybelts than 30 years ago. Graduated driver’s licensing requirements have mixed results. Night-time restrictions are not statistically significant in the model with a year trend, but are in the model with year dummies; passenger restrictions are statistically significant (and negative) in both models. Both are correlated with the year trend variable (R = 0.83 for nighttime restrictions, and R = 0.70 for passenger restrictions), so our preference is the model with year dummies. More focused analysis has looked at how these laws affect just younger aged drivers and found reduced fatalities when these laws are implemented (Fell et al., 2011a, 2011b). Motorcycle helmet laws are very effective at reducing fatalities, and the coefficient estimate is negative and statistically significant in all models. We only include universal helmet laws. More detailed analysis would likely reveal the differential aspects of different laws, but that is not our primary focus. The same holds for cell phone laws. Almost every state now has some legislation that penalizes cell phone use while driving (47 as of 2014). Coefficient estimates are negative and statistically significant in all the models (one is at a 90% confidence level).

P = 0.000

Coef.

Model 2b z

0.534 −1.333 0.422

10.46 −5.41 12.13

−6.924

−8.60

Yes 1500 −11495.04 −7744.06 1.99 0.32 622.30 −55.95

P = 0.000

Coef.

z

0.384 −1.643

8.18 −7.24

0.653 −5.496

20.09 −9.20

Yes 1500 −9817.02 −7640.12 1.99 0.22 332.37 −136.36

P = 0.000

A variety of policies have been implemented to reduce drunk driving, including increased penalties, reduced blood-alcohol concentration levels, and administrative license revocation laws. These are generally found to be effective policies at reducing fatalities (Noland and Karlaftis, 2005). Instead of explicitly modeling the passage of these laws, we use per capita alcohol consumption in each state as a measure of drunk driving. This is positive and statistically significant in all our models, thus we can conclude that policies that reduce the consumption of alcohol, for example, increased taxes, would be effective at reducing traffic fatalities. Alcohol consumption is also partly correlated with median income (R = 0.33) suggesting that consumption may increase when the economy improves, which may lead to additional traffic fatalities. With the exception of motorcycle helmet legislation, all of these laws have seen additional states adopt them since before the Great Recession in 2006 (see Table 4). This is likely one of the reasons that fatalities have gone down in conjunction with the economic recession. Comparing the coefficient estimates on median income from the prior models (models 1a and 1b) with those in the models with policy variables (models 4a and 4b) one can see a reduction in the coefficient value. While income effects have a statistically sig-

Table 3 Conditional negative binomial models, including policy variables. Deaths (1984–2013)

Model 3a

log (median income) log (Gini index) log (total population) log (population aged 15–24 years (%)) log (population over age 75 years (%)) log (total lane miles) log (interstate lane miles (%)) log (arterial lane miles (%)) log (collector lane miles (%)) Primary Safety Belt Law Graduated Driver License Night Time Graduated Driver License Passenger Motorcycle Helmet Law Cell phone Law log (per capita alcohol consumption) log (number of trucks registered (%)) Constant Year Year dummies N initial log likelihood final Log likelihood VIF adjusted ␳2 Score Test Hausman Test

Model 4a z

0.850 −0.42 0.38 0.11 −0.07

19.241 −0.015 No 1500 −14500.48 −7904.81 1.55 0.45 210.62 −111.70

18.78 −1.74 11.03 2.18 −1.77

15.34 −19.31

P = 0.000

Coef.

Model 5a z

0.744 −0.42 0.56

17.18 −2.11 15.95

−0.002 −0.004 −0.041 −0.072 −0.058 0.498

−0.27 −0.45 −3.97 −6.96 −6.11 8.81

14.331 −0.013 No 1500 −13520.84 −7806.27 1.93 0.42 168.53 −69.95

9.42 −14.70

P = 0.000

Coef. 0.732 −0.53 0.60 −0.08 0.01 −0.09 −0.24 0.410 −0.508 0.001 −0.003 −0.040 −0.071 −0.054 0.578 0.055 17.955 −0.015 No 1500 −13351.33 −7801.13 3.45 0.41 188.85 −162.54

Model 3b z 16.62 −2.43 12.8 −1.37 0.17 −1.5 −1.1 0.79 −1.82 0.07 −0.31 −3.76 −6.79 −5.47 8.06 0.71 7.35 −11.26

P = 0.000

Model 4b

Coef.

z

0.434 −1.200 0.405 0.318 −0.299

8.44 −4.70 12.02 5.02 −7.11

−5.908 Yes 1500 −11499.54 −7712.96 2.02 0.33 614.45 −51.61

−7.26

P = 0.000

Coef.

Model 5b z

0.273 −0.842 0.807

6.13 −4.03 24.53

−0.005 −0.025 −0.024 −0.061 −0.014 1.147

−0.69 −3.06 −2.83 −7.80 −1.77 17.74

−11.109

−16.03

Yes 1500 −10850.26 −7523.26 2.34 0.30 256.81 −114.71

P = 0.000

Coef. 0.237 −0.873 0.792 0.151 −0.161 0.028 −0.138 0.314 −0.639 −0.006 −0.024 −0.021 −0.055 −0.019 1.092 −0.117 −10.688 Yes 1500 −10964.67 −7507.09 3.24 0.31 264.44 −75.22

z 5.29 −4.04 19.98 2.54 −4.01 0.51 −0.66 0.73 −2.60 −0.90 −2.95 −2.49 −6.99 −2.44 16.62 −1.67 −12.16

R.B. Noland, Y. Zhou / Accident Analysis and Prevention 98 (2017) 130–138

Coef.

P = 0.000

135

136

R.B. Noland, Y. Zhou / Accident Analysis and Prevention 98 (2017) 130–138

Table 4 Number of states implementing safety policies, 2006 and 2014.

Primary Safety Belt Law Graduated Driver License – Night Restriction Graduated Driver License – Passenger Restriction Motorcycle Helmet Law Cell phone Law

2006

2014

25 42 30 21 19

33 49 45 20 47

nificant association with traffic fatalities, it is slightly lower when these additional variables are introduced. Models 5a and 5b (see Table 3) introduce some selected measures of road infrastructure, namely the total lane miles in each state and the percent of each functional road category (interstates, arterials, and collector roads). Only the percent collector roads is statistically significant (and negative), implying that a state with more lower-category roads will tend to have fewer traffic fatalities. These variables, however, have some degree of collinearity among themselves, for example total lane miles are correlated with the percent of interstate lane miles (R = −0.35), as are the percent arterials with the percent collector lane miles (R = −0.35). Lane miles are also correlated with population (R = 0.65). The VIF for these models exceeds 3. One additional variable included is the percent of vehicles registered as trucks in a state; in model 5b it is negative but only at a 90% confidence level. This may be a poor proxy for truck traffic in a given state. Our final set of models in Table 5 add two proxy variables for improvements in medical care and technology as well as a reduced set of infrastructure variables. We lose some observations as our data on “white infant mortality” has some omissions, as previously discussed. Results show that as mortality rates increase, traffic fatalities increase, consistent with previous use of this proxy (Noland, 2003b). An additional variable, the percent of VMT in rural areas is also positively associated with more traffic fatalities. Any crashes that may occur in more rural areas would likely have a longer emergency response time, increasing the likelihood of any injuries resulting in a fatality.

Other results in these final models are generally consistent with the prior models. One additional infrastructure variable is added, namely the average number of arterial lanes in a state (representing how wide arterial roads are). This is positive and statistically significant in both models. The percent of collector lane miles is also negative and statistically significant, while the lane mile variable is negative and significant in model 6a and not statistically significant in model 6b. This latter result may be due to correlation with the population variable, as previously discussed. This conflicts with the previous work of (Noland, 2003b) which found a positive effect, using data up to 1997. Some have argued that much of the reduction in fatalities is due to VMT decreases. While population continued to increase, VMT has been relatively flat since about 2006. As a further check on the effect of VMT we estimated the full models with VMT instead of population; these are in Table 5 (models 7a and 7b). The variance inflation factor for both models is similar to the population models, but the model fit (as measured by adjusted ␳2 ) is not as good. VMT is, as expected, positive and statistically significant. Other coefficient estimates vary somewhat in their level of significance but generally have the same directional effect. The coefficient on median income is less than in the population models, so VMT is picking up on some of the income effect. Lane miles is negative and significant in both models 7a and 7b. 4.1. Marginal effects Between 2006 and 2014, encompassing the Great Recession and subsequent recovery, total traffic fatalities have dropped from 42,708 to 32,675, or a drop of 10,033 fatalities per year. While 2015 is forecast to see an increase to about 35,200 fatalities, this is still well below the 2006 level (NHTSA, 2016a). To disentangle the magnitude of these effects, we use the coefficient estimates, i.e., the marginal effects, to determine what factors affected the reduction in fatalities between 2006 and 2014. These changes are estimated using two methods. First, marginal effects are applied to changes in each state and then summed to obtain the national total. Second,

Table 5 Conditional negative binomial models, including medical proxy variables. Deaths (1984–2013)

Model 6a Coef.

log (median income) log (Gini index) log (total population) log (total VMT) log (population aged 15–24 years (%)) log (population over age 75 years (%)) log (total lane miles) log (collector lane miles (%)) log (average number of arterial lanes) Primary Safety Belt Law Graduated Driver License Night Time Graduated Driver License Passenger Motorcycle Helmet Law Cell phone Law log (per capita alcohol consumption) log (number of trucks registered (%)) log (white infant mortality rate) log (rural interstate VMT (%)) Constant Year Year dummies N initial log likelihood final Log likelihood VIF adjusted ␳2 Score Test Hausman Test

Model 6b z

Coef.

Model 7a z

Coef.

0.726 −0.138 0.629

14.98 −0.61 13.71

0.229 −0.514 0.785

4.72 −2.22 18.50

−0.038 0.029 −0.163 −1.134 0.507 0.000 −0.009 −0.028 −0.050 −0.054 0.490 0.251 0.137 1.761 18.890 −0.015 No 1334 −9968.91 −6851.67 3.33 0.31 170.17 43.25

−0.62 0.63 −3.10 −3.80 5.94 0.04 −0.83 −2.58 −4.51 −5.39 6.47 3.12 6.16 5.31 7.14 −11.09

0.169 −0.133 0.007 −0.866 0.310 −0.005 −0.033 −0.014 −0.043 −0.022 1.020 −0.074 0.105 0.743 −10.866

2.70 −3.11 0.14 −3.30 4.34 −0.69 −3.68 −1.56 −5.03 −2.70 14.57 −1.01 5.77 2.48 −11.41

P = 0.000 P = 0.001

Yes 1334 −8619.58 −6610.92 3.28 0.23 195.17 −327.42

P = 0.000 P = 0.000

Model 7b z

Coef.

z

0.412 −0.495

8.48 −2.38

0.175 −0.859

3.48 −3.58

0.773 0.151 −0.019 −0.213 −1.118 0.294 −0.002 −0.017 0.013 −0.037 −0.028 0.525 0.255 0.127 1.744 37.362 −0.022 No 1334 −8433.46 −6755.75 3.38 0.20 189.150 −154.04

20.75 2.67 −0.42 −4.25 −4.06 3.62 −0.22 −1.70 1.28 −3.69 −3.03 7.72 3.43 6.27 5.87 13.90 −16.44

0.707 0.106 −0.086 −0.116 −0.773 0.303 −0.011 −0.038 0.003 −0.029 −0.020 0.568 −0.080 0.090 1.253 −3.671

17.17 1.65 −1.93 −2.26 −2.96 4.02 −1.30 −4.01 0.29 −3.18 −2.35 8.19 −1.03 4.73 3.87 −4.31

Yes 1334 −7832.47 −6621.72 3.32 0.15 206.780 −36.40

P = 0.000 P = 0.000

P = 0.000 P = 0.000

R.B. Noland, Y. Zhou / Accident Analysis and Prevention 98 (2017) 130–138

137

Table 6 Marginal effects, change in fatalities between 2006 and 2014, based on model 6b. Percent change in state average between 2006 and 2014

Marginal effects, state average

Change in Fatalities median income Gini index total population population aged 15–24 years (%) population over age 75 years (%) collector lane miles (%) average number of arterial lanes rural interstate VMT (%) white infant mortality rate per capita alcohol consumption Graduated Driver License Night Time Motorcycle Helmet Law Cell phone Law

−3.19% −1.46% 6.49% −3.61% 1.71% −1.66% 0.51% −8.21% −10.53% 2.87%

−453.7 102.1 2379.4 −210.4 −166.3 778.2 −51.1 −3824.3 −422.9 441.6 −139.2 46.2 −372.3

Percent change in national value between 2006 and 2014

[95% Conf. Interval] −265.4 192.4 2127.3 −57.9 −271.0 1240.3 −28.0 −802.7 −279.3 382.2 −213.4 64.2 −642.0

−642.0 11.8 2631.4 −362.9 −61.5 316.0 −74.1 −6846.0 −566.4 501.0 −65.0 28.2 −102.5

Marginal effects, national average

Change in Fatalities −5.20% 1.22% 6.50% −2.82% 1.64% −1.64% −0.17% −10.63%

−507.6 −267.4 2177.7 −202.7 −93.3 604.6 −22.5 −3373.2

[95% Conf. Interval] −296.9 −504.0 1947.0 −55.8 −152.1 963.7 −12.3 −708.0

−718.2 −30.9 2408.4 −349.6 −34.6 245.5 −32.6 −6038.3

Note: values are only shown for those coefficients at a 95% significance level.

we calculate the national change and apply the marginal effects at a national level. This cannot be done for variables for which we only have rates or normalized values (namely, the white infant mortality rate and per capita alcohol consumption) or for the policy variables (i.e., those with dummy variables) which must be calculated based on state values. Table 6 shows the marginal effects for both state averages and national values. These are calculated for model 6b, which is our preferred model as it has a full set of variables. Both the mean value and the 95% confidence interval are shown; marginal effects are not calculated for those coefficients not statistically significant at a 95% level. As can be readily seen, some changes result in reductions in fatalities while others have had an off-setting increase in fatalities. Reductions in real median income have led to a mean reduction in fatalities of anywhere from 454 to 508 per year.11 The Gini index has declined slightly when calculated as a state average, but increased for the nation as a whole. In the latter case, incomes have become less equitable leading to an estimated reduction in fatalities of about 267 per year. When calculated with state averages there is an increase of about 100 fatalities per year. Population growth, of course, is associated with more fatalities and our estimate is about 2200–2300 more fatalities, and is similar with both calculation methods. Changes in the number of younger and older people have also led to slight declines; in total about 296–376 fewer fatalities a year, again this is similar with both methods. The largest source of reductions is from the reduction in the percent of VMT that occurs in rural areas, leading to over 3000 fewer fatalities (based on both calculation methods). The economic recession likely reduced longer distance travel (in rural areas) where access to emergency trauma services is more limited. Improvements in medical care, proxied by the white infant mortality rate, also led to fatality reductions of about 423. Changes in infrastructure have, for the most part, increased total fatalities. In particular, the percent of lanes miles defined as “collectors” has decreased resulting in increases in fatalities ranging from 604 to 778, fairly similar with both methods. In model 6b,

11 To measure the effect during the recession, one could look just at changes between 2006 and 2009. However, while real median income was $56,958 in 2006 and $54,462 in 2013, it was slightly above the 2013 value in 2009 (at $54,925), so this would not make much difference. The lowest value it reached over the course of the recession’s aftermath was $52,605 in 2012.

lane miles were not statistically significant, so there is no change from an increase in lane miles of about 4% nationally. The average number of arterial lanes is associated with a small reduction in fatalities. Interestingly while there is a small increase in the statelevel average, there is still a reduction (recall that the coefficient estimate is positive). This reduction is driven by changes in California, where there was a 16% reduction in the average number of arterial lanes resulting in a drop of 350 fatalities (in 2006 the state had 4240 fatalities). Other large states, such as Florida and Texas increased the average number of arterial lanes, by 4.0% and 2.6%, respectively, and had small associated increases in fatalities, of 68 and 47. The national effect due to a 0.17% reduction is small, with a reduction of 22 fatalities. Changes in state policies have also contributed to reductions over time as various states have implemented some of the policies between 2006 and 2014 as shown in Table 4, but these effects are smaller than the economic effects. In particular, more states have implemented graduated licensing and cell phone laws. The largest policy effects are from increases in cell phone laws, associated with a reduction of about 372 fatalities. Changes in graduated licensing with night-time restrictions have reduced fatalities by about 139 fatalities per year (passenger restrictions were not statistically significant in model 6b). One state repealed their universal motorcycle helmet law, and the total effect is a small increase in fatalities. Increases in per capita alcohol consumption, of about 2.87%, have increased total fatalities between 212 and 442 fatalities per year.

5. Conclusions The analysis presented here is a comprehensive attempt to determine the source of the large drop in traffic fatalities in the US since 2006. A state-level panel dataset was compiled that accounts for changes in economic and demographic factors, infrastructure, and policy variables. The analysis makes use of variation between states and over time. While there are issues of multi-collinearity between variables and especially with those that vary substantially with time, the results reported here provide evidence for the sources of recent fatality reductions and the effectiveness of policies. As household median income decreases there is a reduction in traffic fatalities, confirming the association with economic effects. Changes in the distribution of income also play a substantive role, with reductions in equality associated with fewer fatalities. While increases in fatalities are associated with a growing population,

138

R.B. Noland, Y. Zhou / Accident Analysis and Prevention 98 (2017) 130–138

changes in population age groups, i.e., fewer young people and more over the age of 75 have also been associated with reduced traffic fatalities. Of particular interest for policy makers is the effectiveness of state-level policies. In general we find that graduated licensing, cell phone laws, and motorcycle helmet laws are effective at reducing fatalities. Safety-belt use is now at high levels in most states and this makes it more difficult to link legal requirements to fatality reductions, but the consensus of other research has shown their effectiveness. Changes in alcohol consumption also affect fatalities and further research should examine what policies are most effective at reducing alcohol consumption. Improved emergency response and trauma care, as measured by various proxy variables also are associated with reduced fatalities. In particular, the reduction in the fraction of VMT on rural roads, likely due to the poor economy, has reduced travel on roads that would have longer ambulance response times, and longer distances to trauma care hospitals. Infrastructure effects are relatively small and not always in a positive direction. While our results suggest that added capacity, i.e., more lane miles seems to reduce traffic fatalities, various characteristics of the road network seem to off-set this. Specifically the fraction of roads classified as collectors has decreased and this has led to a small increase in traffic fatalities, as have the small change in the average number of arterial lanes. Many of these road design features lead to higher speeds, and thus may lead to more crashes. As the economy continues to recover over the next few years, an important issue is whether the reductions in fatalities since 2006 will be sustained. Recent data suggests an upward trend, but the implementation of various policies has had a countervailing beneficial impact. It is likely that maintaining or reducing further fatalities is highly dependent on further implementation of effective policies at the state level. Changes in vehicle technology will also play a role, especially crash avoidance systems and potential full automation of cars and trucks, although this may be many years in the future. References Albalate, D., Fernández, L., Yarygina, A., 2013. The road against fatalities: infrastructure spending vs. regulation? Accid. Anal. Prev. 59, 227–239. Allison, P.D., Waterman, R.P., 2002. Fixed-effects negative binomial regression models. Sociol. Methodol. 32 (1), 247–265. Bhargava, S., Pathania, V., 2013. Driving under the (cellular) influence. Econ. Policy 5 (3), 92–125. Branas, C.C., Knudson, M.M., 2001. Helmet laws and motorcycle rider death rates. Accident Analysis & Prevention 33 (5), 641–648. Caird, J.K., Willness, C.R., Steel, P., Scialfa, C., 2008. A meta-analysis of the effects of cell phones on driver performance. Accid. Anal. Prev. 40 (4), 1282–1293. Charlton, S.G., 2009. Driving while conversing: cell phones that distract and passengers who react. Accid. Anal. Prev. 41 (1), 160–173. De Raedt, R., Ponjaert-Kristoffersen, I., 2000. Can strategic and tactical compensation reduce crash risk in older drivers? Age Ageing 29 (6), 517–521. Eisenberg, D., 2003. Evaluating the effectiveness of policies related to drunk driving. J. Policy Anal. Manage. 22 (2), 249–274. Eisenberg, D., 2004. The mixed effects of precipitation on traffic crashes. Accid. Anal. Prev. 36 (4), 637–647. Evans, L., 2014. Traffic fatality reductions: United States compared with 25 other countries. Am. J. Public Health 104 (8), 1501–1507. Fell, J.C., Jones, K., Romano, E., Voas, R., 2011a. a An evaluation of graduated driver licensing effects on fatal crash involvements of young drivers in the United States. Traffic Inj. Prev. 12 (5), 423–431. Fell, J.C., Todd, M., Voas, R.B., 2011b. b A national evaluation of the nighttime and passenger restriction components of graduated driver licensing. J. Saf. Res. 42 (4), 283–290. Foss, R.D., Evenson, K.R., 1999. Effectiveness of graduated driver licensing in reducing motor vehicle crashes. Am. J. Prev. Med. 16 (1), 47–56. Frank, M.W., 2009. Inequality and growth in the United States: evidence from a new state-level panel of income inequality measures. Econ. Inq. 47 (1), 55–68. Guimarães, P., 2008. The fixed effects negative binomial model revisited. Econ. Lett. 99 (1), 63–66. Hakamies-Blomqvist, L., 1994. Compensation in older drivers as reflected in their fatal accidents. Accid. Anal. Prev. 26 (1), 107–112. Harrison, M.A., 2011. College students’ prevalence and perceptions of text messaging while driving. Accid. Anal. Prev. 43 (4), 1516–1520.

Hausman, J.A., Hall, B.H., Griliches, Z., 1984. Econometric models for count data with an application to the patents-R & D relationship. Econometrica 52 (4), 909–938. IIHS, 2016. Motorcycles. Insurance Institute for Highway Safety (July 2016) http:// www.iihs.org/iihs/topics/laws/helmetuse?topicName=motorcycles. Jonah, B.A., 1986. Accident risk and risk-taking behaviour among young drivers. Accid. Anal. Prev. 18 (4), 255–271. Kopits, E., Cropper, M., 2005. Traffic fatalities and economic growth. Accid. Anal. Prev. 37 (1), 169–178. Kyrychenko, S.Y., McCartt, A.T., 2006. Florida’s weakened motorcycle helmet law: effects on death rates in motorcycle crashes. Traffic Inj. Prev. 7 (1), 55–60. Laberge-Nadeau, C., Maag, U., Bellavance, F., Lapierre, S.D., Desjardins, D., Messier, S., Saı¨di, A., 2003. Wireless telephones and the risk of road crashes. Accid. Anal. Prev. 35 (5), 649–660. Lloyd, L., Wallbank, C., Broughton, J., 2016. A collection of evidence for the impact of the economic recession on road fatalities in Great Britain. Accid. Anal. Prev. 80, 274–285. Maheshri, V., Winston, C., 2016. Did the great recession keep bad drivers off the road? J. Risk Uncertain., http://dx.doi.org/10.1007/s11166-016-9239-6. Maio, R.F., Green, P.E., Becker, M.P., Burney, R.E., Compton, C., 1992. Rural motor vehicle crash mortality: the role of crash severity and medical resources. Accid. Anal. Prev. 24 (6), 631–642. Marques, M., Pinho, C., 2014. Tax-treaty effects on foreign investment: evidence from European multinationals. FinanzArchiv: Public Finance Anal. 70 (4), 527–555. McKnight, A.J., Peck, R.C., 2002. Graduated driver licensing: what works? Inj. Prev. 8 (Suppl. 2), ii32–ii36 (discussion ii36-8). NHTSA, 2004. Graduated Driver Licensing System, Traffic Safety Facts, US Department of Transportation. Natl. Highw. Traffic Saf. Admin., DOT HS 810 888W. NHTSA, 2013. 2012 Motor Vehicle Crashes: Overview, Traffic Safety Facts, Research Note, US Department of Transportation. Natl. Highw. Traffic Saf. Admin., DOT HS 811 856. NHTSA, 2016a. Early estimate of motor vehicle traffic fatalities in 2015. Traffic Saf. Facts, DOT HS 812 269. NHTSA, 2016b. Traffic safety facts: lives and costs saved by motorcycle helmets, 2014, US department of transportation. Natl. Highw. Traffic Saf. Admin., DOT HS 812 295. NHTSA, 2016c. Traffic safety facts: motorcycles, US department of transportation. Natl. Highw. Traffic Saf. Admin., DOT HS 812 292. NHTSA, 2016d. Traffic safety facts: seat belt use in 2015-Use rates in the states and territories, US department of transportation. Natl. Highw. Traffic Saf. Admin., DOT HS 812 274. Nathens, A.B., Jurkovich, G.J., Cummings, P., Rivara, F.P., Maier, R.V., 2000. The effect of organized systems of trauma care on motor vehicle crash mortality. JAMA 283 (15), 1990–1994. National Safety Council, 2016. Motor Vehicle Fatalities Up (Aug 26) http://www. nsc.org/Connect/NSCNewsReleases/Lists/Posts/Post.aspx?ID=134. Noland, R.B., Karlaftis, M.G., 2005. Sensitivity of crash models to alternative specifications. Transp. Res. Part E 41 (5), 439–458. Noland, R.B., Quddus, M.A., 2004. Improvements in medical care and technology and reductions in traffic-related fatalities in Great Britain. Accid. Anal. Prev. 36 (1), 103–113. Noland, R.B., Sun, F., 2015. Why have traffic fatalities dropped in the US? 94th Annual Meeting of the Transportation Research Board. Noland, R.B., 2003a. Medical treatment and traffic fatality reductions in industrialized countries. Accid. Anal. Prev. 35 (6), 877–883. Noland, R.B., 2003b. Traffic fatalities and injuries: the effect of changes in infrastructure and other trends. Accid. Anal. Prev. 35 (4), 599–611. Noland, R.B., 2013. From theory to practice in road safety policy: understanding risk versus mobility. Res. Transp. Econ. 43, 71–84. Peltzman, S., 1975. The effects of automobile safety regulation. J. Pol. Econ., 677–725. Piketty, T., Saez, E., 2001. Income Inequality in the United States, 1913–1998 (series updated to 2000 available), NBER Working Paper No. 8467, http://www.nber.org/papers/w8467. Qiu, L., Nixon, W.A., 2008. Effects of adverse weather on traffic crashes: systematic review and meta-analysis. Transp. Res. Rec.: J. Transp. Res. Board 2055 (1), 139–146. Ralph, K., Blumenberg, E., Taylor, B.D., Smart, M., 2014. Do graduated drivers’s licenses reduce mobility and contribute to the decline in teen driving? The safety mobility trade-off. Work. Pap., 23pp. Rivara, F.P., Thompson, D.C., Cummings, P., 1999. Effectiveness of primary and secondary enforced seat belt laws. Am. J. Prev. Med. 16 (1), 30–39. Shope, J.T., 2007. Graduated driver licensing: review of evaluation results since 2002. J. Saf. Res. 38 (2), 165–175. Voas, R.B., Tippetts, A.S., Fell, J., 2000. The relationship of alcohol safety laws to drinking drivers in fatal crashes. Accid. Anal. Prev. 32 (4), 483–492. Wilde, G.J., 1982. The theory of risk homeostasis: implications for safety and health. Risk Anal. 2 (4), 209–225. Williams, A.F., 2007. Contribution of the components of graduated licensing to crash reductions. J. Saf. Res. 38 (2), 177–184. Yannis, G., Papadimitriou, E., Folla, K., 2014. Effect of GDP changes on road traffic fatalities. Saf. Sci. 63, 42–49.