Exploring the relationship between truck load capacity and traffic accidents in the European Union

Exploring the relationship between truck load capacity and traffic accidents in the European Union José I. Castillo-Manzano a, ⇑ , Mercedes Castro-Nuño a , Xavier Fageda b a Applied Economics & Management Research Group, University of Seville, Spain b Dept. of Economic Policy, University of Barcelona, Spain article info Article history: Received 12 March 2015 Received in revised form 10 February 2016 Accepted 11 February 2016 Available online 2 March 2016 Keywords: Road freight transport Truck types Load capacity Road safety Panel data European Union abstract Applying econometric techniques to EU28 panel data and controlling for explanatory variables such as road types, we find that increased truck load capacity does not necessarily aggravate road traffic safety. Specifically, heavy trucks do not seem to be linked with greater numbers of traffic fatalities/accidents, medium trucks appear to be the worst performers in terms of fatalities, and light trucks seem to be the worst for accidents. In summary, our results clarify the complex relationship between truck load capacity and road safety, pointing to the existence of a negative correlation for accidents per capita and an inverse U-shaped curve for fatalities per capita. Ó2016 Elsevier Ltd. All rights reserved. 1. Introduction Road freight traffic is a key part of the European common market. However, the external costs associated with motorized transportation have prompted the European Union (EU) to develop a more efficient, sustainable and safer transport system within a framework of liberalization and competition (e.g., Koliousis et al., 2013). Additionally, as Sanches Rodrigues et al. (2015) explain, certain features of modern European supply chains (i.e., the high concentration of production seeking to exploit economies of scale, and large volumes of freight transported long distances across the continent) have led practitioners to call for an increase in truck dimensions. The fact that road freight transport seems to be a more flexible mode for door-to-door service has contributed to freight consolidation in larger and heavier vehicles (McKinnon, 2008). However, from a road safety approach, the literature has generated unclear evidence as to the impact of large truck traffic (Carson, 2011). Therefore, the current article explores empirically how different truck configurations impact European road safety, with particular emphasis on the effect of heavy truck traffic compared to other truck categories. Authors such as Stein and Jones (1988) consider that truck accident involvement may be affected by several factors (i.e., truck operating characteristics, such as load or size; truck driver characteristics, such as age or work conditions; weather and road conditions); while others, such as e.g., Jones et al. (1983) and Vallette et al. (1981), state that vehicle weight and loading variables are essential parameters for vehicle dynamics, roll stability and handling characteristics. Thus, we assess the safety http://dx.doi.org/10.1016/j.tre.2016.02.003 1366-5545/Ó2016 Elsevier Ltd. All rights reserved. ⇑ Corresponding author at: Department of Economic Analysis and Political Economy, University of Seville, Spain. Tel.: +34 954 55 67 27. E-mail addresses: [email protected] (J.I. Castillo-Manzano), [email protected] (M. Castro-Nuño), [email protected] (X. Fageda). Transportation Research Part E 88 (2016) 94–109 Contents lists available at ScienceDirect Transportation Research Part E journal homepage: www.elsevier.com/locate/tre differences between three different truck types on the basis of load capacity (i.e., light, medium and heavy trucks) in terms of traffic accidents and fatalities. In general, the earlier literature has broadly shown that truck road accidents are a significant issue. According to Corsi et al. (2014), due to their economic and social costs, road traffic accidents involving trucks have caught the attention of governments, academicians and society in general. In recent years scientific papers have analyzed a vast array of issues. For example, Britto et al. (2010) and Cantor et al. (2008) have investigated the relationship between motor carrier safety and financial performance; Cantor et al. (2009, 2010), Hickman and Hanowski (2011), Mejza and Corsi (1999) and Mejza et al. (2003) have explored the effect of individual driver attributes, technologies and safety best practices on accident likelihood, while other authors have examined the impact of truck drivers’ employment status (Cantor et al., 2013; Corsi et al., 2012). Further studies have tested the influence of external factors, such as infrastructure design (Chang and Chien, 2013); vehicle technologies and maneuvering capabilities (Mooren et al., 2014); environmental conditions (Choi et al., 2014), and specific safety policies (Castillo-Manzano et al., 2015a). Some studies have even analyzed the possibility of internalizing the externalities generated by truck accidents (Anderson and Auffhammer, 2014; Savage, 2011). Less attention has been paid to research into the effects on traffic safety outcomes of trucks disaggregated by size, weight or configuration. Chang and Mannering (1999) explain that it can be difficult to isolate the significance of these effects due to the presence of confounding factors. In addition, Carson (2011) points out that this literature represents a mixture of statistical papers and institutional reports based on different sources, models and methodologies that provide contradictory conclusions. Moreover, since the well-known TRB study (1990), most of studies on the topic have referred to the U.S. (af Wåhlberg, 2008; Braver et al., 1997; Khorashadi et al., 2005; Welki and Zlatoper, 2009, among many others), or discuss the safety of only one truck size category, mainly large and heavy trucks (Anderson and Auffhammer, 2014; Lemp et al., 2011; Zhu and Srinivasan, 2011). More precisely, the previous literature shows that the evidence regarding the relationships between load capacity-based truck categories and traffic accidents and their severity is fragmented or even inconclusive. In fact, the af Wåhlberg (2008) meta-analysis, possibly the broadest review published on this topic to date, states that research on this topic has been characterized by a geographic bias (most of the studies take the US as their case study) and detects certain methodological shortcomings that are basically the result of limited exposure data for holding the road type constant. Therefore, the author concludes that this issue has not been conclusively answered. Our paper therefore attempts to shed light on this still unresolved issue by assessing the safety differences in terms of traffic accidents and fatalities between three different truck types categorized by load capacity (i.e., light, medium and heavy trucks). For this we use a comprehensive database of a large number of heterogeneous EU countries and for a broad recent timeframe, the 1999–2012 period. While the main variables of the empirical analysis are for different categories of truck load capacity, we control for several explanatory factors of safety outcomes, such as country attributes, proxies for the type of roads used by trucks and specific road safety policies. All this enables us to offer robust results on this relevant topic. Specifically, our research aims to answer the following questions: How does truck weight – gauged by load capacity – affect European traffic safety? What differences in safety are there among different truck load capacity categories in the EU? Which truck load capacity category is the most hazardous for EU traffic safety? Are people more likely to be involved in fatal accidents in the EU as truck load capacity increases? The main contribution of our investigation is that it is the first to examine EU road safety issues in relation to this specific breakdown of truck types using a broad set of control variables. We believe that our findings may serve as a starting point for the analysis of different truck types circulating on other geographic levels and in other areas; and, in a practical context, they may also be helpful to governments and policy makers by adding safety issues to the perspectives of sustainability and logistics to orient both logistics and road safety strategies. In this regard, our results could also serve as a guide to adapt technology efforts, regulation and enforcement strategies to different truck configurations. The timeliness of our research is justified by recent EU and individual government initiatives to extend the use of higher capacity freight vehicles (see e.g., Bergqvist and Behrends, 2011; ITF, 2010; McKinnon, 2008; Sanches Rodrigues et al., 2015 for a more in depth analysis). The body of EU legislation that addresses standard regulations on maximum weights and dimensions for heavy goods vehicles, buses and coaches mainly conforms to Directive 96/53/EC (see http://eur-lex.europa.eu/legal-content/EN/TXT/?uri=celex:31996L0053). This norm ensures the unhindered movement in Member States 1 of vehicles within these limits. It also places restrictions on the cross-border movement of vehicles that are heavier, longer or higher than these standards, the so-called Mega Trucks,Long and Heavy Vehicles (LHVs), and EuroCombis or Gigaliners. 2 The European Commission (EC) has attempted to legalize cross-border mega trucks traffic during recent years, but the initial Directive has recently been amended by Directive (EU) 2015/719 (accessible at: http://eur-lex.europa.eu/legal-content/EN/TXT/?uri=uriserv%3AOJ.L_.2015.115.01.0001.01.ENG), which provides for derogations from the maximum lengths to improve the safety and environmental emissions of heavy goods vehicles (through improved aerodynamic performance) and also recommends that no modifications should be made to Directive 96/53/EC’s 1 Today different EU countries have different national regulations regarding maximum dimensions. According to this Directive, the maximum permitted length is 16.50 m for semitrailers and 18.75 m for road trains with a total combined weight of 40 tonnes (the so-called Heavy Goods Vehicles, HGVs). 2 Up to 25 m in length and 60 tonnes in weight. J.I. Castillo-Manzano et al. / Transportation Research Part E 88 (2016) 94–109 95 restrictions on the cross-border movement of mega trucks. 3 A detailed analysis of the impact of mega trucks can be found in e.g., Ortega et al. (2014) and Steer et al. (2013). Following this introduction, the remainder of the paper is organized as follows: Section 2presents the theoretical background to our analysis and the research hypotheses; Section 3describes the data, the variables and the model used; Section 4 analyzes and discusses the results, and Section 5summarizes the main conclusions and implications for policy. 2. Theoretical background and hypotheses Most of literature on the topic agrees that differences in truck weight affect road safety since, from a technical point of view, they lead to different designs, configurations and driving performance. Consequently, as truck traffic tends to change, there are also changes in risk exposure to accidents. A number of scholars, such as Abdel-Aty and Abdelwahab (2004), Bálint et al. (2013), Cantor et al. (2010), Carson (2011), Corsi et al. (2012, 2014), Dong et al. (2014) and Khorashadi et al. (2005), among others, have demonstrated this using a variety of approaches. Given that each truck configuration has its own specific problems, Kahane (2003) highlights physics and intuitive reasons (mass differences, speed differences and a range of other safety factors) that may explain why different weights can be expected to present different accident risks. Nevertheless, from studies such as Braver et al. (1997) that do not find any relevant truck safety differences linked to weight in fatal accidents, the meta-analysis by af Wåhlberg (2008) (possibly the most comprehensive quantitative review published on the topic to date) deduces that this question has not been conclusively answered, as the relationship between truck weight and road safety is really a nonlinear correlation dependent upon the accident category considered. We take as our starting point this relevant contribution by af Wåhlberg (2008), whose outcomes point to accident risk increasing in line with increasing truck weight up to a certain point, except in the case of fatal accidents. However, the author indicates that this finding may only be valid for North America and notes that there are several methodological issues with studies, mainly due to limited exposure data for holding the road type constant. Further analysis of the literature on safety differences among the different truck categories could help to reveal the sign of this relationship. A priori, the association between heavy trucks and road safety outcomes (re: light trucks) seems to be a complex issue because, as Milton et al. (2008) argue, there are two forces depending on the accident category. On the one hand, a higher number of heavy trucks can lead to a decrease in accident risk due to a variety of factors. For example, heavy trucks are mainly used for inter-urban trips (Kanaroglou and Buliung, 2008) and as af Wåhlberg (2008) highlights may have an advantage from circulating on better and safer roads (highways). Moreover, according to Cantor et al. (2008, 2009), Corsi et al. (2012, 2014), Huang et al. (2008) and Welki and Zlatoper (2009), the implementation of the latest safety technologies, specific safety regulations and performance measures for heavy vehicles means that they are more resistant, due to their superior dynamic stability and control. In addition, as Lemp et al. (2011) state, one heavy truck may replace several light trucks, meaning that if heavy trucks are used, fewer truck trips are required to haul any given volume of freight (lower risk exposure). Lastly, as Milton et al. (2008) argue, the presence of heavy trucks may have a slowing effect on traffic speed. On the other hand, although their occupants may be better protected, heavy trucks may also be involved in more severe accidents than light trucks, mainly due to more dangerous maneuvers and as they are less flexible, because of the weight and speed differential (Carsten, 1987; Chang and Chien, 2013; Dong et al., 2014; Khorashadi et al., 2005; Qin et al., 2013; Zhu and Srinivasan, 2011). Accidents specifically involving light trucks have also been explored, although findings have again differed depending on the road safety outcome considered. The accident risk of light trucks seems to be higher than that of heavy trucks, especially in urban settings. According to Kanaroglou and Buliung (2008) the reason for this is that freight distribution mostly requires goods to be transported over short distances by light trucks, especially in the case of the major medium and large European cities where, as Dablanc et al. (2013) state, typical urban morphologies (high traffic density, parking problems, commercial– leisure–tourist activities concentrated in the city center) influence mobility and place restrictions on freight pick-up and delivery by heavier goods vehicles. Friswell and Williamson (2013) state in this regard that the urban environment creates more hazardous exposure for light truck drivers because of the stress that comes from heavy traffic or a lack of parking space. On the other hand, specific light truck features (e.g., a high center of gravity that may increase the risk of rollover, a more rigid structure and worse maneuverability, Anderson, 2008) may also drive up accident risk. Abdelwahab and Abdel-Aty (2004), Abdel-Aty and Abdelwahab (2004), Harb et al. (2007) and White (2004) provide evidence of a greater likelihood of rear-end and multivehicle collisions in urban settings involving light trucks because drivers’ visibility of other smaller vehicles (passenger cars, bicycles, motorcycles) and pedestrians is impeded. Also, Desapriya et al. (2010) suggest that light trucks pose a greater risk of pedestrian injuries and deaths. Other studies analyze factors that also contribute to higher accident risk for light trucks compared to heavy trucks. For example, the number of light trucks has grown considerably in recent years (Welki and Zlatoper, 2009). Certain authors even conclude that the behavior and qualification levels of drivers in these two truck categories are different. For example, a 3 These vehicles are already circulating in some Scandinavian countries with less advanced rail systems (Sweden and Finland), and other States have carried out pilot tests (the Netherlands, Germany, Denmark, Belgium). Mega truck or LHV cross-border traffic is a controversial issue in the EU due to a conflict of interests between the haulage industry, governments, institutions and public opinion (with a powerful campaign opposing mega trucks on the www. nomegatrucks.eu website, with opposition from 243 organizations in 24 countries. Even the European Parliament has shown its reluctance). 96 J.I. Castillo-Manzano et al. / Transportation Research Part E 88 (2016) 94–109 comparison by Huang et al. (2011) finds that heavy truck drivers (with more stringent requirements to obtain the corresponding driving license) are more frequently commercial drivers who are supposed to be better qualified to respond in hazardous situations. Using a Florida accident database the same study also detects that light truck drivers are more aggressive behind the wheel. In sum, the evidence of the net effect of light trucks on traffic fatalities seems to be ambiguous. Depending on the methodological design, we find studies that only focus on the fact that in a fatal collision the greater mass of a light truck may provide better protection for its occupants compared to smaller vehicles, e.g., cars (Evans and Frick, 1993, 1994; Gayer, 2004; Ossiander et al., in press). However, there are also a number of studies such as Brozovic ´and Ando (2009), Gayer (2004) and White (2004) that attempt to split the net fatality effect of light trucks into its internal and external impacts (i.e., also considering the externalities for other road users). Using an alternative approach based on the combination of both accidentlevel data and U.S. state-level data, Anderson (2008) indicates that because of their greater mass and specific features, light trucks may generate an external safety cost to the occupants of other vehicles that could neutralize the initial safety advantage for their own occupants. In keeping with official EU statistics available from Eurostat, our research focuses on three mutually exclusive truck categories classified by load capacity: (1) Light: trucks with a load capacity of under 5000 kg. (2) Medium: with a load capacity of between 5000 kg and 14,999 kg. (3) Heavy: with a load capacity of 15,000 kg or over. Considering all the above evidence (mostly for the U.S. case), we therefore formulate the following hypotheses H1 and H2: Hypothesis 1 (H1). Heavy truck usage is related to fewer traffic accidents compared to light trucks. Hypothesis 2 (H2). Heavy truck usage is related to more traffic fatalities compared to light trucks. Returning to the literature that analyzes the impact of heavy trucks on safety, a number of studies, almost all for the U.S., consider a broad truck category called ‘‘large trucks” for vehicles with a gross vehicle weight in excess of 10,000 lb. For example, Blower et al. (2010) analyze mechanical conditions and crash involvement; Hauer (2001) explores the safety of certain vehicle types of driver groups; Lyman and Braver (2003) consider truck accidents in terms of occupant deaths; and Zaloshnja and Miller (2004) calculate the cost of truck crashes. However, as Park and Pierce (2013) state, considering a separate specific medium trucks category (for example, for trucks between 10,001 and 26,000 lb) would improve knowledge of truck safety. The research that considers a separate category of medium size trucks in road safety is quite scarce. Nevertheless, in general terms the few studies that follow this disaggregation agree that, compared to heavy trucks, there are differential operational characteristics and accident environments for medium trucks that lead to both a higher risk of accidents and more severe accidents: e.g., higher speeds with more severe results in the case of an accident, more aggressive driving, and drivers who are often not required to possess a commercial driver’s license. These studies are for the U.S. (Huang et al., 2011; Park and Pierce, 2013; Pierce and Park, 2014) and for Sweden (Bálint et al., 2013) respectively, but they only evaluate truck types in terms of length and do not take into account other aspects, such as load capacity. However, there is one study of the U.S. (Corsi et al., 2014) that explains how and why vehicles in a truck category equivalent to our study’s medium trucks (10,001–26,000 lb) present worse driver performance and higher traffic violation rates than heavy trucks. Consequently, the previous research and logic lead us to expect that medium trucks might possess a mix of the characteristics of both heavy and light trucks – for example, they can be used for both urban and inter-urban journeys – as a result of which medium trucks might also possess a combination of safety risks. Thus, we formulate the following Hypotheses, H3 and H4: Hypothesis 3 (H3). Heavy truck usage is related to fewer traffic accidents compared to medium trucks. Hypothesis 4 (H4). Heavy truck usage is related to fewer traffic fatalities compared to medium trucks. 3. Empirical framework We apply a country fixed effects model that allows us to consider any omitted constant variable that correlates with the explanatory variables. We use different techniques (ordinary least squares and negative binomial) to verify the robustness of the results to the dependent variable being treated in different ways: as a continuous or a count variable. Specifically, our model takes the following form for country iduring period t: J.I. Castillo-Manzano et al. / Transportation Research Part E 88 (2016) 94–109 97 Y it ¼ a þb k X it þ c k Z it þk k W it þ l i þ m Year t þ e it ð1Þ in which we consider two different Y it dependent variables estimated in two different regressions. First, we consider the log of the total per capita fatality rate (within 30 days of the accident, according to the Vienna Convention definition) as the dependent variable. 4 Second, we consider the log of the total accident rate as the dependent variable. X it contains the vector of the country’s economic and demographic attributes; Z it refers to truck-related variables, and W it represent road safety policy-related variables. l i are country fixed effects that control for omitted time-invariant countryspecific variables; Year is an annual time trend that controls for the common trend in all the countries in the dataset, and e it is a mean-zero random error. Note here that the only data available are annual and aggregated at the country level. This results in some limitations, as we cannot control for seasonal effects or regional differences. We use data for the EU28 countries from 1999 to 2012 and the unit of observation is the country-year pair: 1999 is a relevant year for our scope, as it is the year that the so-called ‘Eurovignette’ Directive 1999/62/EC (modified by Directive 2006/38/EC and Directive 2011/76/EU) was passed that aimed to harmonize the tolls and vignettes associated with the cost of road infrastructure in Member States and charge them to Heavy Goods Vehicles. 2012 is the most recent year for which data were available for the majority of the countries considered. However, data for truck-related variables were not available for some countries (Belgium, Bulgaria, Greece and the United Kingdom) and so these countries were excluded from the econometric analysis. Data were available for Austria, Croatia, Cyprus, the Czech Republic, Denmark (1999–2002), Estonia (2007–2012), Finland, France (2010–2012), Germany (1999– 2008), Hungary (1999–2008, 2010), Ireland (1999), Italy (2004–2012), Latvia, Lithuania (all years except 2008), Luxembourg (1999–2009), Malta (all years except 2003–2007), the Netherlands (all years except 2012), Poland (2002–2012), Portugal (2011–2012), Romania (2008–2012), Slovenia (all years except 2003–2004), Spain (2005–2012) and Sweden (all years except 2010 and 2012). With respect to the countries for which data were not available for all years, we include data for the years for which we do have information, as is generally done in unbalanced panel data regressions. Note here that data for truck variables account for 231 observations and data for other variables, such as fatalities and accidents, are also incomplete. Thus, for some country-pairs we may have information for truck variables but not, for example, for fatalities. Hence, the final sample comprises 222 observations. Note here that any regressions with only one observation for a country fixed effect (as is the case for Ireland, for example) automatically exclude said observation. Tables 1 and 2 give the descriptions, information sources, descriptive statistics and number of observations available for all of the variables used in the analysis. We also specify the variables that capture country attributes, truck numbers and specific safety policy variables. All the variables present adequate variability, given that standard deviation is relatively high in relation to mean values. However, the variability of some variables, such as motorway density, may be low over time. As explanatory variables we include the per capita gross domestic product (GDP) as an indicator of the country’s economic development (Castillo-Manzano et al., 2014a,b, 2015b; Grimm and Treibich, 2013; Yannis et al., 2014). The relationship between a country’s economic development and road safety outcomes could be expected to be an inverted U-shape (Bishai et al., 2006; Kopits and Cropper, 2005; Loeb and Clarke, 2007). Indeed, fatality and accident rates may increase with economic development in very poor countries, due to increased exposure to road traffic accidents. However, the relationship between economic development and traffic safety outcomes may become flat or even invert after reaching a certain wealth threshold. We test the hypothesis concerning non-linearity between GDP and traffic safety outcomes by including the GDP and the square of the GDP as explanatory variables. If the non-linear relationship is confirmed, then the coefficient for the GDP variable will be positive and the coefficient for the GDP 2 variable will be negative. The number of passenger cars per capita (motorization) is also taken into account. This variable relates to the level of development of private transportation. Regarding the expected sign of this coefficient, on the one hand, higher motorization rates may imply higher exposure to road traffic accidents, while on the other, more developed countries may have better infrastructure and vehicles, more advanced policies and more beneficial social attitudes towards road safety (Kopits and Cropper, 2005). By way of example, Albalate and Bel (2012) found a negative relationship between traffic fatalities and the motorization rate. We also take into account the influence of the quality of transport infrastructure by including a variable for motorway density. The sign of the coefficient for this variable is expected to be negative. Indeed, a negative relationship can be expected between the quality of transport infrastructure and road safety outcomes, as was found in Jamroz (2012) and Noland (2003). It should be noted that Malta and Latvia do not possess any motorways, which explains why the sample includes countries with zero motorway density. We also include the density of other roads (state, provincial and district roads) as an explanatory variable. This variable could capture the influence of inter-urban trips on lower quality roads or more frequent short trips. Hence, a priori the sign expected for this variable is not clear in either the traffic fatalities or accidents equations. By including this variable, we seek to take into account the methodological indications in the relevant meta-analysis by af Wåhlberg (2008). Said author identifies the lack of appropriate exposure data regarding road type as the main problem in studies that analyze the safety impact of different truck configurations. Larger trucks usually circulate on larger, safer roads (highways), which could represent a certain advantage and possibly skew the comparison of the findings for different truck configurations. 4 This is the dependent variable typically used to assess road safety outcomes, as the interpretation of policy variables is clearer. 98 J.I. Castillo-Manzano et al. / Transportation Research Part E 88 (2016) 94–109 We also consider two control variables for the percentage of vulnerable population in each country. First, we include a variable for the percentage of the population over 60 years old. Second, we also consider a variable for the percentage of the population aged 15–24 years. While we may expect that risk exposure is higher for the younger population, since younger road users usually take more risks (Langford and Koppel, 2006), the impact of accidents may be greater for senior road users, and so morbidity and mortality rates may be higher for older populations (Braver and Trempel, 2004; Yee et al., 2006). Thus, we expect positive signs in the traffic fatalities equation for the coefficients for these two explanatory variables, while a negative sign is expected in the accidents equation for the old population coefficient and a positive sign for the young population coefficient. Additionally, we include a variable for the population density of the country. We may expect that the proportion of urban trips (short distance trips in intra-urban neighborhoods; Holz-Rau et al., 2014) over total trips will be higher in more densely populated countries; and specifically, following Allen et al. (2012), trips for urban freight activities are associated with city population density. The accident rate should be higher for intra-urban trips than inter-urban or rural trips, but the severity of accidents may be lower for intra-urban trips, due, according to Rakauskas et al. (2009) and Zwerling et al. (2005) among others, to factors such as road design, emergency medical service proximity and lower differential speeds. Hence, the coefficient for this variable is expected to have a positive sign in the accidents equation, while a negative sign might be found in the traffic fatalities equation. We also include additional variables for road safety policies. Following previous studies on road safety (Albalate, 2008; Castillo-Manzano and Castro-Nuño, 2012; Castillo-Manzano et al., 2014a,b), we consider the influence of prevention policies Table 1 Variables used in the empirical analysis. Variables Description Source Type of data Fatalities per capita Number of traffic fatalities per million inhabitants CARE (EU road accident database) Dependent variable Accidents per capita Number of traffic accidents per million inhabitants CARE (EU road accident database) Dependent variable Motorization Number of registered passenger cars per thousand inhabitants UNECE, EUROSTAT (for population) Country attribute GDP per capita Per capita gross domestic product in International Comparable Prices (US$ at 2005 prices and PPP) EUROSTAT Country attribute Motorway density Number kms of motorways divided by km 2 of the country UNECE, EUROSTAT Country attribute Other roads density Number kms of other roads (state, provincial and district) divided by country km 2 UNECE, EUROSTAT Country attribute Old % population aged 60+ EUROSTAT Country attribute Young % population aged 15–24 EUROSTAT Country attribute Population density Density of population EUROSTAT Country attribute Trucks_light Number of trucks (trucks + semi-trailers with load capacity <5000 kg) per million inhabitants EUROSTAT Trucks Trucks_medium Number of trucks (trucks + semi-trailers with load capacity 5000–14,999 kg) per million inhabitants EUROSTAT Trucks Trucks_heavy Number of trucks (trucks + semi-trailers with load capacity P15,000 kg) per million inhabitants EUROSTAT Trucks BAC_05 Dummy variables that take a value of 1 if the permitted maximum BAC rate is under 0.5 g/l European Commission Road Safety Website Road safety policy Points_system Dummy variable that takes a value of 1 if a points driving license system is in force European Transport Safety Council (ETSC) Road safety policy Speed limits Maximum speed limits European Commission Road Safety Website Road safety policy Table 2 Descriptive statistics of the variables used in the empirical analysis. Variables Mean Standard deviation Minimum value Maximum value Number of observations Fatalities per capita 103.46 45.26 21 257 379 Accidents per capita 42798.45 77732.46 633.9 395,689 379 Motorization 430.54 111.61 134 686 388 GDP per capita 25331.75 11622.87 6737 74,128 391 Motorway density 1.73 1.74 0 6.35 388 Other roads density 129.09 134.09 0.49 979.87 388 Old 21.00 2.64 15.1 27.1 389 Young 13.22 1.65 10 17.2 392 Population density 0.16 0.23 0.02 1.37 392 Trucks_light 52007.98 32449.33 2396.3 159,826 231 Trucks_medium 5472.032 2540.95 1297 17,100 231 Trucks_heavy 4140.57 2596.62 31 11,395 231 BAC_05 0.28 0.45 0 1 392 Points_system 0.60 0.49 0 1 392 Speed limits 121.5 13.51 80 112 392 J.I. Castillo-Manzano et al. / Transportation Research Part E 88 (2016) 94–109 99 such as legal blood alcohol limits. A dummy variable is included with a value of one for countries and periods where the maximum blood alcohol concentration allowed for professional drivers is under 0.5 g/L. Most EU28 countries have set the limit at 0.5 or below, so we are able to test whether blood alcohol concentration rates of under 0.5 for professional drivers are effective for improving road safety outcomes. Moreover, we add a variable that captures the implementation of points-based driving licenses through a dummy variable with a value of one if a penalty driving license system is applied. The introduction and application of any type of points system to driving licenses may lead to lower traffic fatality and accident rates, although the long-term effect of this policy has been questioned (see Castillo-Manzano and Castro-Nuño, 2012, for an international meta-analysis of the effects of this type of policy). Another road traffic policy variable that we take into account is the maximum speed limit on motorways. According to previous studies (Elvik, 2012), a positive sign may be expected for the coefficient for this variable, as one of the consequences of a higher speed limit may be a higher number of road traffic fatalities and accidents. The most relevant variables in our analysis are related to truck load capacity. 5 We include three variables for the number of trucks per capita 6 that consider three separate load capacity categories, as explained in the previous section (i.e., Heavy trucks for P15,000 kg; Medium trucks for 5000–14,999 kg; Light trucks for <5000 kg). We opt for load capacity to distinguish between truck types as, following the ITF (2010), load volume determines the number of trucks required per unit of freight, and a higher load capacity can result in fewer vehicle-kilometers traveled. For the U.S. case study, and following both the Federal Highway Administration (FHWA) and the National Highway Traffic Safety Administration (NHTSA), most papers on the topic classify trucks in three groups by Gross Vehicle Weight (GVW) expressed in pounds. However, there are no equivalents to these groups in terms of load capacity in any official statistics available in Europe. Thus, as has been clearly stated above, our grouping criteria is based on available Eurostat data. As was to be expected, the mean value data in Table 2 show that light trucks are much more numerous than trucks in the other categories. In contrast, the numbers of trucks in the other two categories are quite similar. We would like to state that the variables used in the empirical analysis are not proxies for country size. Variables are measured in per capita terms (fatalities, GDP, motorization, truck load capacity variables, population and motorway density – the latter two in per km terms) and are not linked to country size (road safety policy variables, percentage of vulnerable population). 4. Results and discussion Table 3 shows the correlation matrix of the variables used in the empirical analysis. Some variables are highly correlated. Table 3 reports a particularly high correlation between the GDP per capita and motorization variables, and a high correlation between the population density and other roads density variables. These high correlations may cause a multicollinearity issue. Multi-collinearity can exaggerate estimates of the parameter variance and distort its statistical significance or even result in parameter estimates of implausible magnitude in the most extreme cases. The Variance Inflation Factor (VIF) is a measure that is widely used to examine the degree of multicollinearity between explanatory variables. Several rules of thumb for VIF have been considered as a sign of severe multicollinearity. The usual rule of thumb considered in the econometric textbooks is 10, although practitioners may use lower threshold values and a value of 5 is common. Table 4 shows the variance inflation factors of all the variables used in the empirical analysis. The computed VIFs are below 5 for the majority of the variables and below 10 for all of the variables. If we exclude the motorization and other roads density variables, the computed VIFs are below 5 for all variables. It is important to note that the truck motorization variables do not seem to be affected by any multicollinearity issues in our research. Estimates may present problems of heteroscedasticity and temporal autocorrelation in the error term. Running the Wooldridge test for autocorrelation in panel data shows that there may be a problem of serial autocorrelation that needs to be addressed. Additionally, the Breusch–Pagan/Cook–Weisberg test indicates that we have an issue with heteroscedasticity. We also apply the panel unit root test developed by Levin et al. (2002), which can be regarded as an augmented Dickey– Fuller (ADF) test when lags are included. This test indicates that there is no non-stationarity problem with our dependent variable. We also apply the Ramsey Reset Test, which confirms that no relevant variables have been omitted from our model. The results of these tests are given in Tables 5 and 7. Taking the results of these tests into account, we performed the estimation using two different techniques to check for robustness. In Table 5 we report the results of the traffic fatality regressions using the ordinary least squares (OLS) method 5 Overall, we may expect a stronger (negative) impact of trucks on traffic fatality than passenger cars. Castillo-Manzano et al. (2015a) examine the differential impact between the motorization rates for trucks and passenger cars but do not distinguish between different truck categories, which are essential for our analysis. 6 Motorization variables (both for passenger cars and trucks) are in per capita terms, as this is a common indicator in international statistics and the road safety literature. It might be interesting to test our model with an alternative indicator of motorization, such as the ratios between numbers of passenger cars and trucks and total road-kilometers. Unfortunately, Eurostat data for the number of road-kms are very poor and are not available for several countries. Reported data show significant differences among countries (taking load capacities into account) and significant changes within some countries in just one year. 100 J.I. Castillo-Manzano et al. / Transportation Research Part E 88 (2016) 94–109 with country fixed effects. The standard errors are robust to heteroscedasticity and clustered by country to address the autocorrelation problem. In Table 6 we report the results of the traffic fatalities equations using the population-averaged paneldata model with a negative binomial distribution. The standard errors are robust to heteroscedasticity, as they are in the OLS regression. An advantage of these latter regressions is that we can specify an AR(1) within-group correlation structure for panels. Population is used as an exposure variable. In Table 7 we show the results of the accidents equation using both techniques. It should be noted that the negative binomial distribution explicitly models the dependent variable as the number of occurrences. The dependent variable in the OLS regression is total fatalities or accidents per capita, while in the regression that uses the negative binomial distribution the dependent variable is total fatalities or total accidents, with population as an exposure variable. Note that in the OLS regression the dependent variable is expressed in logs, while the implementation of the log-link function in the negative binomial distribution regression means that the log transformation can be used without transforming the dependent variable. The log-link function can be implemented in the framework of generalized linear models assuming a negative binomial distribution of the dependent variable, which is what we consider in our analysis. In Tables 5 and 6 we report the results of five different specifications of the road traffic fatalities equation, while Table 7 shows the results for the accidents equation. We report the results of different specifications. Specification (1) is the baseline regression. Specification (2) adds the lagged value of the dependent variable as an explanatory variable to check whether results with a dynamic estimator differ from the baseline model. Specifications (3) and (4) add two different interaction variables (centering the variables to prevent any multicollinearity issues): interaction between GDP per capita and motorization rate, since the effects of the motorization rate on road safety outcomes may vary depending on GDP per capita; and interaction between trucks per capita and the motorization rate, since mixing higher levels of car and truck traffic may lead to worse safety outcomes. Specification (5) excludes the motorization variables, other roads density and the square of the GDP per capita variables to check whether results for the truck variables are affected by multicollinearity between any of Table 3 Correlation matrix of the variables used in the empirical analysis. Variables 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 11 20.05 1 30.41 0.27 1 40.45 0.03 0.77 1 50.32 0.08 0.53 0.68 1 60.20 0.06 0.10 0.01 0.10 1 70.36 0.36 0.23 0.02 0.16 0.12 1 8 0.38 0.34 0.11 0.02 0.05 0.03 0.43 1 90.21 0.08 0.27 0.17 0.21 0.07 0.32 0.02 1 10 0.42 0.20 0.03 0.17 0.39 0.29 0.09 0.03 0.16 1 11 0.18 0.31 0.33 0.51 0.30 0.27 0.06 0.12 0.01 0.17 1 12 0.36 0.04 0.13 0.10 0.16 0.79 0.07 0.02 0.17 0.24 0.23 1 13 0.01 0.02 0.01 0.06 0.19 0.09 0.18 0.02 0.08 0.15 0.11 0.25 1 14 0.05 0.13 0.23 0.04 0.31 0.08 0.05 0.09 0.15 0.23 0.18 0.31 0.02 1 15 0.14 0.27 0.11 0.21 0.20 0.14 0.04 0.05 0.38 0.31 0.03 0.32 0.02 0.17 1 Note: Denomination of variables: 1 = fatalities per capita; 2 = accidents per capita; 3 = motorization; 4 = GDP per capita; 5 = motorway density; 6 = other roads density; 7 = old population; 8 = young population; 9 = trucks_light; 10 = trucks_medium; 11 = trucks_heavy; 12 = population density; 13 = BAC_05; 14 = points_system; 15 = speed limits. Table 4 Variance inflation factors of the variables used in the empirical analysis. Variables Fatalities per capita Accidents per capita All variables Excluding motorization and other roads density All variables Excluding motorization and other roads density Motorization 4.60 – 4.63 – GDP per capita 6.70 3.18 6.54 2.97 Motorway density 3.51 3.36 3.22 3.13 Other roads density 7.91 – 7.74 – Old 4.51 3.91 4.62 3.65 Young 4.72 4.30 4.94 4.40 Population density 8.73 1.65 8.48 1.65 Trucks_light 2.58 1.77 2.47 1.67 Trucks_medium 2.12 1.86 2.14 1.87 Trucks_heavy 1.97 1.93 1.89 1.84 BAC_05 1.32 1.27 1.31 1.26 Points_system 1.71 1.83 1.59 1.73 Speed limits 2.14 2.30 2.10 2.26 J.I. Castillo-Manzano et al. / Transportation Research Part E 88 (2016) 94–109 101 Table 5 Results of estimates (OLS model). Independent variables Dependent variable: Number of traffic fatalities per capita (1) (2) (3) (4) (5) Number of traffic fatalities per capita (lagged) – 0.43 *** (0.11) – – – Motorization 0.00037 (0.0005) 0.00021 (0.00044) 0.0009 (0.0007) 0.00012 (0.0006) – GDP per capita 0.000039 *** (0.000013) 0.000024 *** (9.31e06) 0.000047 *** (0.000013) 0.000046 *** (0.000013) 0.000028 ** (0.000012) GDP per capita 2 2.48e10 *** (1.48e10) 1.90e10 * (9.89e11) 9.37e10 *** (2.78e10) 3.60e10 ** (1.58e10) – Motorway density 0.003 (0.06) 0.008 (0.05) 0.06 (0.08) 0.01 (0.06) 0.05 (0.06) Density of other roads 0.002 *** (0.006) 0.0022 *** (0.0007) 0.0029 *** (0.0005) 0.002 *** (0.0005) – Old 0.03 (0.02) 0.03 * (0.017) 0.06 (0.03) 0.02 (0.02) 0.06 ** (0.02 Young 0.04 ** (0.02) 0.03 * (0.018) 0.03 (0.02) 0.03 * (0.01) 0.02 (0.02) Population density 1.99 *** (0.53) 1.88 *** (0.55) 2.27 *** (0.59) 1.96 *** (0.53) 0.62 (0.43) Trucks_light 1.63e06 (1.55e06) 6.61e08 (1.14e06) 1.64e06 (1.87e06) 2.49e06 (2.20e06) 1.35e06 (1.82e06) Trucks_medium 0.000026 *** (8.38e06) 0.000021 *** (3.62e06) 0.000,022 ** (9.00e06) 0.000026 *** (8.48e06) 0.000026 ** (0.000011) Trucks_heavy 0.000015 (0.000012) 3.35e07 (5.35e06) 9.20e06 (0.000011) 9.48e06 (0.000011) 0.000022 (0.000013) BAC_05 0.023 (0.07) 0.0017 (0.05) 0.004 (0.07) 0.006 (0.06) 0.006 (0.07) Points_system 0.00006 (0.04) 0.03 (0.03) 0.02 (0.04) 0.004 (0.04) 0.10 (0.06 Speed limits 0.04 *** (0.01) 0.014 (0.010) 0.04 *** (0.01) 0.04 *** (0.01) 0.02 * (0.013) Motorization GDP per capita – – 1.27e07 *** (4.81e08) –– Motorization Trucks per capita – – – 1.25e08 ** (4.97e09) – Time_trend 0.09 *** (0.012) 0.056 *** (0.010) 0.09 *** (0.011) 0.09 *** (0.011) 0.09 *** (0.010) Intercept 0.03 (1.75) 0.64 (1.33) 1.08 (2.15) 0.36 (2.03) 1.59 (2.07) Country fixed effects YES YES YES YES YES R 2 0.96 0.96 0.96 0.96 0.94 Test joint sign (Wald v 2 ) 100.58 *** 120.69 *** 104.84 *** 100.63 *** 86.49 *** Ramsey Reset Test (Ho: No omitted variables) 1.21 0.22 1.83 0.23 1.52 Breusch-Pagan/Cook-Weisberg test for heteroscedasticity (Ho: Constant variance) 7.96 *** 6.79 *** 8.10 *** 6.03 *** 6.26 *** Wooldridge test –autocorrelation (Ho: First-order autocorrelation) 4.88 ** 6.02 ** 5.49 ** 4.88 ** 3.46 * Test: – Trucks_light + Trucks_heavy = 0 1.67 10.98 ** 3.08 * 5.07 ** 5.39 ** Number of observations 222 222 222 222 222 Note 1: Standard errors are given in brackets. Robust to heteroscedasticity and clustered by country. * Statistical significance at 10%. ** Statistical significance at 5%. *** Statistical significance at 1%. 102 J.I. Castillo-Manzano et al. / Transportation Research Part E 88 (2016) 94–109 Stein, H.S., Jones, I.S., 1988. Crash involvement of large trucks by configuration: a case-control study. Am. J. Public Health 78 (5), 491–498. Thompson, R.G., 2015. Vehicle orientated initiatives for improving the environmental performance of urban freight systems. In: Green Logistics and Transportation. Springer International Publishing, pp. 119–129. TRB (Transportation Research Board), 1990. Truck Weight Limits: Issues and Options. Special Report 225. Washington D.C. Accessible at <http://pubsindex. trb.org/view.aspx?type=MO&id=309190>. Vallette, G.R., McGee, H.W., Sanders, J.H., Enger, D.J., 1981. The Effect of Truck Size and Weight on Accident Experience and Traffic Operations, Volume 3: Accident Experience of Large Trucks. Federal Highway Administration Report No. FHWA-RD-80-137. Washington, D.C. Volvo Trucks, 2013. European Accident Research and Safety Report. Accessible at <http://www.volvotrucks.com/SiteCollectionDocuments/VTC/Corporate/ Values/ART%20Report%202013.pdf>. Welki, A.M., Zlatoper, T.J., 2009. How highway safety regulations and enforcement activities affect subcategories of motor vehicle fatalities. Transport. Res. Part E: Log. Transport. Rev. 45 (6), 1030–1038. White, M., 2004. The ‘‘Arms Race” on American roads: the effect of sport utility vehicles and pickup trucks on traffic safety. J. Law Econ. 47 (2), 333–355. Yannis, G., Papadimitriou, E., Folla, K., 2014. Effect of GDP changes on road traffic fatalities. Saf. Sci. 63, 42–49. Yee, W.Y., Cameron, P.A., Bailey, M.J., 2006. Road traffic injuries in the elderly. Emerg. Med. J. 23, 42–46. Zaloshnja, E., Miller, T.R., 2004. Costs of large truck-involved crashes in the United States. Accid. Anal. Prev. 36, 801–808. Zhu, X., Srinivasan, S., 2011. A comprehensive analysis of factors influencing the injury severity of large-truck crashes. Accid. Anal. Prev. 43, 49–57. Zwerling, C., Peek-Asa, C., Whitten, P.S., Choi, S.W., Sprince, N.L., Jones, M.P., 2005. Fatal motor vehicle crashes in rural and urban areas: decomposing rates into contributing factors. Injury Prevent. 11 (1), 24–28. J.I. Castillo-Manzano et al. / Transportation Research Part E 88 (2016) 94–109 109