LINDA NG BOYLE2,*
1Dept. of Industrial and Systems Engineering
University of Wisconsin-Madison
3217 Mechanical Engineering Building
1513 University Ave
Madison, WI 53706
Phone: (319) 430-7583
2 Associate Professor, Dept. of Industrial &
Dept. of Civil & Environmental Engineering
University of Washington
G5 Mechanical Engineering Building
Seattle, WA , USA 98195-2650
Phone: (206) 616-0245
Fax: (206) 685-3072
* Corresponding author
KEYWORDS: injury severity, crash data, General Estimates System (GES), Midwestern crashes, rural areas, crash type
This study used binary logit models to examine the crash factors that contribute to severe injuries to the drivers across four Midwestern states: Iowa, Kansas, Missouri, and Nebraska. The findings regarding the association between many crash factors (e.g., driver's age, gender, seat belt use, and alcohol use) and severe injuries are consistent with previous studies. However, the magnitude of the associations varies greatly with some outcomes not even significant in some states (e.g., adverse weather). Findings were then compared to those from regional crash estimates using the General Estimates System (GES) and differences were obtained for rural road crashes. The outcomes bring up issues on the appropriateness of implementing safety countermeasures based on geographical regions and underscore the need for standard crash reporting procedures.
Many traffic regulations and countermeasures are aimed at reducing the risk of driver fatalities and injuries. However, traffic safety is still a major concern in the United States. U.S. crash data from the year 2008 show that over 37,000 people were killed and about 2.5 million were injured from motor vehicle crashes (NHTSA 2009). In one study, highway crashes were estimated to be about 3.2% of the total medical cost in the United States, and over 14% for those in the 15–24 years range (Miller, Lestina, and Spicer 1998). Although several studies haveprovided some insights on the driver, vehicle, and road and environmental factors associated with these motor vehicle crash injuries and fatalities (e.g., Bedard et al. 2002; Connor et al. 2004; Evans and Frick 1994; Huelke and Compton 1995; Kim et al. 1995; O'Donnell and Connor 1996), there are differences that exist across states and many of these differences correspond to the data used as well as the analytic techniques employed.
Driver characteristics related to elevated crash risks include age and experience (Zhang et al. 1998; Kweon and Kockelman 2003), weather conditions (Khattak, Kantor, and Council 1998; Khattak and Knapp 2001), alcohol impairment (Zador, Krawchuk, and Voas 2000; Keall, Frith, and Patterson 2004), and driver distraction (Klauer et al. 2006; Neyens and Boyle 2008; Violanti 1998). However, the patterns of injury risk do differ across regions. For example, a model of injury severity based on data from Hawaii showed no significance differences for age and gender (Kim et al. 1995), while studies on Wisconsin (Tavris, Kuhn, and Layde 2001) and Iowa do reveal differences in age and gender. Estimates from the Iowa crash data also differed from the national estimates (Hill and Boyle 2005). These findings demonstrate the impact of aggregating data to too high a level. That is, a model based on national data may not be able to capture patterns specific to a state or region.
The Midwestern states in the United States do have common characteristics, including many rural roads and sparsely populated areas. These rural areas also contribute to a large proportion of crash fatalities in the United States (NHTSA 2008b). A study on four Midwestern states, Kansas, Nebraska, South Dakota, and North Dakota, showed that there is an inverse relationship between motor vehicle crash fatality rates (per 100,000 persons) and population density (Muelleman and Mueller 1996). That is, the more sparse the population in these rural areas, the higher the fatality rates. A 5-mph increase in roadway speed limit increases the odds of fatalities and injuries (Baum, Lund, and Wells 1989; Renski, Khattak, and Council 1999). Although many studies tend to group this region into one cluster, there may be differences between these states with respect to traffic patterns.
The present study examines different factors surrounding traffic crashes and the severity of driver injuries within four Midwestern states: Iowa, Kansas, Missouri, and Nebraska. The analyses attempt to estimate the likelihood of severe injuries, given a crash has occurred. In other words, exposure data is not incorporated into the analysis, as the goal is not to estimate the general crash likelihood of driver groups or driving conditions, but instead to estimate the odds of severe outcomes when a crash occurs. Each state is examined individually and then compared to estimates from the regional model. It is hypothesized that the injury trends will be similar to those previously observed in other studies using Midwestern states, but the magnitude of such associations may differ. Comparisons will then be made to the representative sample at the region level of all 12 Midwestern states, extracted from the General Estimates System (GES) (NHTSA 2008a). Conclusions and the impacts on policies are then considered relative to the results of the analyses and comparisons.
Data for this study was obtained from the Departments of Transportation and Roads for Iowa, Kansas, Missouri, and Nebraska. The four databases contained information on crashes for the years 2001 to 2006. The scope of this study was limited to passenger vehicles, so data from other vehicle types (e.g., buses, trucks, and motorcycles) were eliminated prior to analyses. Moreover, only data related to drivers (not passengers or pedestrians) were included in the analysis in order to achieve consistency for model comparisons. That is, passengers could be situated in different locations within the vehicle, but the driver is always in the driver seat. Each state also had different formats for their crash data. Hence, the databases were standardized and reformatted to facilitate comparisons. The usable crash records available for analysis (i.e., records with sufficient crash information) encompassed 78.33% of Iowa, 84.49% of Kansas, 85.68% of Missouri, and 70.53% of Nebraska's reported crashes. The differences noted are based on the amount of available information extracted from each state's database.
A model of the Midwestern region of the United States based on data from the National Automotive Sampling System (NASS)-GES for the same years (2001 to 2006) was then used as a comparison to the individual findings (NHTSA 2008a). The GES data is a stratified sample of crashes weighted to represent national crash patterns. The GES obtains its data from a national representative probability sample that is extracted from police accident reports (PARs). The sampling from PARs is accomplished in three stages: 1) sampling of geographic areas, which provides the Primary Sampling Units (PSUs), 2) sampling of police jurisdiction within each PSU, and 3) selection of crashes within the sampled police jurisdictions (NHTSA 2005).
Injury Level Classification in Crash Data
The classification of injury level in crash reports is based on the KABCO scale, which was introduced by the National Safety Council in the late 1960s (Compton 2005). This rating system, also used in GES (NHTSA 2005), categorizes occupant injuries into five groups: fatal (K), incapacitating (A), non-incapacitating (B), possible injuries or complaint of pain (C), and not injured (O). In addition, categories such as "unknown" and "not reported" are included for some states because discerning the level of injury may not always be possible. All four states examined in this study employ the KABCO scale; although there were some differences in definitions (table 1).
Four separate models were developed to examine the factors that may increase the likelihood of a severe injury for each state. Although an ordering to the severity level may initially seem obvious, there are two general problems with employing ordered models in the context of crash injury severity as noted by Savolainen and Mannering (2007): 1) non-injury crashes may be underreported in crash data and this can lead to biased coefficient estimation by the model, and 2) ordered models restrict variable influences. In other words, the hypothesis that the parameters from an ordered logit model are equal across all the levels of the dependent variable was rejected. Rather, many researchers have used multinomial logit models to examine the severity of occupant injuries (Awadzi et al. 2008; Bedard et al. 2002; Khorashadi et al. 2005; Watt et al. 2006).
TABLE 1 Detailed KABCO Injury Categories in Crash Databases
|Crash database||Iowa||Kansas||Missouri||Nebraska||GES national sampled data|
|KABCO injury levels||K||Fatal||Fatal injury||Fatal||Fatal||Fatal injury|
|KABCO injury levels||A||Incapacitating||Disabled (incapacitating)||Disabling injury||Disabling||Incapacitating injury|
|KABCO injury levels||B||Non-incapacitating||Injury not-incapacitating||Evident injury (not disabling)||Visible||Non-Incapacitating injury|
|KABCO injury levels||C||Possible||Possible injury||Probable injury (not apparent)||Possible||Possible injury|
|KABCO injury levels||O||Uninjured||Not injured||Not apparent||No injury||No injury|
Died prior to crash
Unknown if injured
KEY: na = not available
The primary goal of this study was to compare outcomes across multiple states and the use of the "KABCO" scale was different in each one (see table 1). Hence, a more simplistic binary logistic regression model (or logit model), as used by Al-Ghamdi (2002), is employed to provide insights on injuries while also allowing a clear comparison across the four states, which can then be generalized to the regional level. The injury severity levels were therefore grouped into two general categories of "severe" (including codes K and A) and "non-severe" injuries (including codes B, C, and O) and examined using simultaneous binary logit models developed with SAS (Statistical Analysis System) version 9.1 and the CATMOD procedure (Allison 1999). The CATMOD procedure is used to estimate the likelihood of a driver sustaining severe injuries when compared to non-severe injuries. The model is represented in equation 1.
Where Xir is the value of the explanatory variable r for driver i, and βjr is the coefficient associated with the rth variable (r = 1, …, R) for the jth injury severity level. Yi is a random variable whose value (j = 1 or 2) indicates the severity level of the injuries sustained by driver i. The CATMOD procedure uses maximum likelihood estimation (MLE) and outputs logarithmic ratio estimates of the likelihood of severe (versus non-severe) injuries, based on the levels of each explanatory variable. By exponentiating the logarithmic ratio estimates, odds ratios for sustaining severe (versus non-severe) injuries were obtained. The adjusted odds ratios [AORs] are odds ratios that have been adjusted for the other explanatory variables in the model because they are calculated based on a multivariate model that controls for other factors. Wherever the logarithmic ratio estimate is positive, exponentiating this estimate will give a value greater than 1, and thus the odds of sustaining severe injuries are higher than non-severe injuries. Conversely, when this estimate is negative, exponentiating will give a value less than 1, and the odds of having a severe injury are less than a non-severe injury. The likelihood ratio test was used to compare the goodness of fit (Cochran 1952) of the fitted model to a saturated model (i.e., backward elimination) (Ananth and Kleinbaum 1997). A high p-value would suggest that the fitted model was a good fit and that no significant terms were omitted.
(1) log((Pr(Yi=m)/Pr(Yi=k)) = R∑r=1 (βmr-βkr)Xir
The statistical models included explanatory variables shown to have an impact on the likelihood of a crash or a severe injury in a crash. Drivers were categorized into three age groups: 24 years old and younger (younger drivers), aged between 25 and 65 (reference group), and drivers older than 65 (older drivers) as similarly done in other studies (Zhang et al. 1998; Khattak, Kantor, and Council 1998; Farmer, Braver, and Mitter 1997). Weather conditions were divided into two categories, normal and adverse. The adverse weather category encompassed situations where rain, snow, freezing rain, fog/smoke, mist, sleet, severe winds, blowing sand/soil/dirt, or combinations of these conditions were present. If none of the above conditions were present, then weather was labeled as normal. Lighting was considered in two categories, daylight and non-daylight situations with the latter consisting of night, dawn, and dusk. Roadway speed limit was set up into three groups: less than 35 mph, between 35 and 55 mph, and higher than 55 mph. The categories used for weather and lighting conditions are consistent with those used in similar studies (Khattak, Kantor, and Council 1998; Zhang et al. 1998; Abdel-Aty 2003). The point of impact variable (the first point that produced damage or injury) was examined using five categories; front, driver side, passenger side, top/under, and rear of the car.
Five crash types were considered: rear-end, head-on, angular, sideswipe, and single-vehicle crashes. Angular, rear-end, sideswipe, and head-on crashes are the four categories of "collision with motor vehicle in transport" used by US DOT, while single-vehicle crashes correspond to "collisions with fixed object" and "collision with object not fixed" (NHTSA 2009). In addition to crash type, the (initial) crash point of impact was included for states whose crash database supported this variable (i.e., Iowa and Nebraska).
Two driver-related factors were also of interest given the abundance of literature demonstrating increase crash risk, driver distraction and blood-alcohol content (BAC). However, the crash databases did not include sufficient information regarding these two factors for the years examined. More specifically, the proportion of crashes that included any details about the distraction-related factors encompass only 1.27% in Iowa, 1.33% in Kansas, 1.18% in Missouri, and 0.81% in Nebraska. Surprisingly, driver BAC information was not available in any of the states' databases. Those states that did include this variable had a large proportion of non-reporting (e.g., about 51% of Iowa crashes with drivers under the influence of alcohol lacked BAC level). Considering these limitations, only the more general factor of "being under the influence of alcohol or drugs" (yes or no) was used in the analyses. It should be noted that several other factors could contribute to the severity of injuries sustained by driver. For example, vehicle size and mass are known to influence driver fatality (Evans and Frick 1992, 1994). However, this information was not available in the datasets that were used for this study.
A separate model was developed for each state, with each state's databases including the majority of variables of interest. The Iowa crash database included all variables of interest. Kansas did not have sufficient information regarding point of impact, while Nebraska lacked data on air bag deployment. Missouri did not have information on drug use, air bag deployment, and point of impact. All models fitted well based on the likelihood ratio test. The significance level was set to 0.0001.
There were similar demographic patterns across the four states (table 2). Drivers' mean age ranged from 36.64 (in Nebraska) to 37.90 (in Missouri). The proportion of female drivers ranged from 43.23% (Kansas) to 45.09% (Nebraska). Among crash types, angular and rear-end crashes were the most common in each of the five databases, comprising 65–85% of crashes (table 3).
Driver and Vehicle Characteristics
The four binary logit models are shown in table 4, with a general finding that female drivers were more susceptible to serious injuries in the four Midwestern states examined. There were similar estimates between Iowa and Kansas (adjusted odds ratios [AORs] = 1.07 and 1.08, respectively) and between Missouri and Nebraska (AORs = 1.21 and 1.24, respectively). With respect to driver age, younger drivers (younger than 25) were less likely to sustain serious injuries when compared to the middle-aged group (aged 25-65). Older drivers were more likely to be severely injured. There was also an age and gender interaction in Missouri only, with young females being less likely to sustain severe injuries compared to middle-aged male drivers (AOR = 0.96).
TABLE 2 Descriptive Statistics of State Crash Data
|State||Number of crashes||Mean age (SD)||Gender (%)||Seat belt use (%)||Drug/alcohol use (%)|
|Nebraska||271,445||36.64 (17.23) 54||54.91||45.09||79.1||1.52|
TABLE 3 Frequencies of Crash Types in Crash Databases
|Crash type||Iowa||Kansas||Missouri||Nebraska||GES (Midwest)|
Passengers were shown to have a protective effect in Iowa and Kansas, with drivers being less severely injured driving with passengers when compared to driving alone. In contrast, drivers with passengers in Missouri were more likely to sustain severe injuries. No significant association was observed between injury severity and passengers in Nebraska.
As expected, there was a higher likelihood of a severe injury when the driver did not use a seat belt, and this was consistently observed in all four states with Nebraska and Missouri being more similar in odds (AORs = 2.70 and 2.74, respectively) and Iowa and Kansas having higher odds ratios (3.59 and 4.24, respectively). Air bag deployment data was available in Iowa and Kansas only, and drivers in these two states were more likely to be severely injured with an airbag deployment. In each state, drivers under the influence of alcohol or drug were significantly more likely to sustain severe injuries compared to sober drivers. The magnitude of this effect varied slightly from 1.32 (Kansas) to 1.74 (Nebraska).
Crash Types and Points of Impact
The odds of sustaining severe injuries were higher for head-on crashes when compared to rear-end crashes, and ranged from 3.18 in Iowa to 4.92 in Kansas. Drivers in sideswipes were less likely to sustain severe injuries in all four states, with quite similar odds ratios (from 0.38 to 0.50). Observations for single-vehicle crashes were consistent for Iowa and Missouri, indicating higher likelihoods of serious injuries (AORs = 1.29 and 1.93, in Iowa and Missouri, respectively). However, the odds of having severe injuries were not significantly different between single-vehicle and rear-end crashes in Kansas and Nebraska. No significant difference was observed between angular and rear-end crashes in any of the states in terms of severe injury odds. In Iowa and Nebraska, drivers in crashes with impacts on the driver side were 1.16 and 1.82 times, respectively, more likely to sustain severe injuries compared to those whose vehicles were impacted on the rear side. No other significant differences were observed with respect to crash types and points of impact.
In all four states, drivers involved in crashes in rural settings were more likely to sustain severe injuries when compared to those having crashes in urban areas (AORs ranged from 1.71 (Iowa) to 2.55 (Missouri)). Non-dry surfaces were associated with lower likelihoods of severe crashes in all four states, i.e., the odds of sustaining severe injuries on non-dry surfaces were between 0.81 (for Nebraska) and 0.92 (for Missouri) compared to dry surfaces. There was an interaction effect between crash location and crash type, with drivers more likely to be severely injured if they were involved in head-on crashes in rural settings (AORs = 1.26, 1.48, and 1.20 in Iowa, Kansas, and Missouri, respectively), compared to those involved in rear-end crashes in urban settings. By contrast, drivers in single-vehicle crashes in rural settings were less likely to sustain severe injuries (AORs = 0.72, 0.48, and 0.83 in Iowa, Kansas, and Missouri, respectively). In Missouri, two additional contrasts were significant as well; drivers in angular crashes in rural settings were 0.88 times less likely and those in sideswipes were 1.22 times more likely to have severe injuries.
Findings regarding lighting conditions were not consistent across the states. In Iowa and Kansas, drivers were slightly less likely to sustain severe injuries in crashes occurring in non-daylight situations, i.e., during night, dawn, or dusk (AORs = 0.94 and 0.92, respectively). In Nebraska, contrary to Iowa and Kansas, the odds of sustaining severe injuries were higher in daylight hours (AOR = 1.11). The Missouri model showed no significant association between lighting and injury severity. Weather condition at the time of crash was a significant factor only in Missouri, where drivers were slightly less likely to be severely injured in crashes occurring in adverse weather conditions (AOR = 0.94).
The likelihood of driver's sustaining severe injuries also increased on roads with higher posted speed limits. The odds ratios for severe injuries on roads with lower speed limits (less than 35 mph) compared to the reference speed limit (35–55 mph) were very similar for Iowa, Missouri, and Nebraska (0.62, 0.56, and 0.58, respectively), while Kansas revealed a slightly lower odds ratio (0.39). For roads with higher speed limits, the odds ratios ranged from 1.34 in Iowa to 2.20 in Kansas.
TABLE 4 State Models for the Likelihood of Severe Injuries
|Estimate||SE||χ2||Adjusted OR||Estimate||SE||χ2||Adjusted OR|
|Age < 25||-0.36||0.02||346.3||0.7||-0.26||0.02||180.5||0.77|
|Age > 65||0.4||0.02||279.7||1.49||0.33||0.03||164.7||1.39|
|Passenger(s) present in the car||-0.1||0.01||64||0.9||-0.07||0.01||30.3||0.93|
|Under influence of alcohol/drug||0.5||0.02||753.7||1.64||0.28||0.02||250.9||1.32|
|No seat belt in use||1.28||0.02||2813.8||3.59||1.44||0.03||2105.1||4.24|
|Air bag deployed||0.76||0.02||1254||2.13||0.27||0.03||115.7||1.32|
|Speed limit < 35 mph||-0.48||0.03||361.5||0.62||-0.93||0.03||1348.6||0.39|
|Speed limit > 55 mph||0.29||0.03||110.5||1.34||0.79||0.03||837.6||2.2|
|Point of impact: front||-0.03||0.02||ns||0.97||-||-||-||-|
|Point of impact: driver side||0.15||0.03||30||1.16||-||-||-||-|
|Point of impact: passenger side||-0.11||0.03||ns||0.9||-||-||-||-|
|Point of impact: top/under||0.1||0.06||ns||1.1||-||-||-||-|
|Head-on crashes in rural settings||0.23||0.03||44.6||1.26||0.39||0.05||57.1||1.48|
|Angular crashes in rural settings||0.04||0.02||ns||1.04||0.01||0.04||ns||1.01|
|Sideswipes in rural settings||0.07||0.04||ns||1.07||0.14||0.07||ns||1.15|
|Single-vehicle crashes in rural settings||-0.33||0.02||217.9||0.72||-0.73||0.03||546.5||0.48|
|Female drivers younger than 25||-||-||-||-||-||-||-||-|
|Female drivers older than 65||-||-||-||-||-||-||-||-|
|Number of observations||370,428||598,070|
NOTE: All parameters are significant at p ≤ 0.0001 unless otherwise noted (ns). For variables not found statistically significant, no contrast estimate is reported.
Comparisons Across States
There were some common and consistent findings across all four Midwestern states for various driver characteristics (gender, age, alcohol and drug use, and seat belt use), as well as environmental conditions including surface condition, posted speed limit, and rural/urban settings. However, differences were observed for crash type. Single-vehicle crashes significantly impacted the likelihood of a severe injury in Iowa and Missouri, but in Kansas and Nebraska, there was no difference between single-vehicle and rear-end crashes.
Similarly, the interaction between crash type and location (rural/urban) was significant for all states but Nebraska. The interaction between age and gender, on the other hand, was only significant in Missouri. Results pertaining to weather condition showed significant differences only in Missouri. Driving in non-daylight conditions was associated with a decrease in injuries in both Iowa and Kansas, but increased injuries in Nebraska. No significant difference was observed in Missouri. Passengers were found to be similarly associated with a protective effect for drivers in Iowa and Kansas, but an increase in severe injuries in Missouri. No significance was observed in Nebraska.
Point of impact information was only available for Iowa and Nebraska, where it produced patterns in the same direction (although the magnitude was different). Air bag deployment information was also available in Iowa and Nebraska only. Here again, results indicated associations in the same direction but different magnitude.
Figure 1 summarizes the results (point estimates and confidence intervals) for variables that showed a significant association with severe injuries across the four states examined. Note that in many cases, the confidence intervals overlap heavily (e.g., age, under the influence of alcohol/drug) which indicates great similarities across the four states examined. In other cases, however, the estimates are more diverse (e.g., seat belt use, head-on crashes) which indicates considerable differences among the states. These findings motivate developing a Midwestern crash injury severity model, extract driver injury patterns from it, and compare them to the four state models to assess the extent to which sampled crash databases can describe injury patterns across the region.
The Midwestern region of the United States was examined using the sampling region established by the National Automotive Sampling System (NASS) and collected as part of the GES data (NHTSA 2005). This region consisted of 12 states (i.e., Ohio, Indiana, Illinois, Michigan, Wisconsin, Minnesota, North Dakota, South Dakota, Nebraska, Iowa, Missouri, and Kansas). The goal of this analysis was to assess the level of agreement between the outcomes observed from the states' crash databases and the outcomes from a sample of crashes in the same region (GES).
FIGURE 1 State Estimates and Confidence Intervals for the Association Between Crash Variables and Severe Injuries
Midwestern crash data used in the injury severity model included 97,070 weighted records, representing 12,252,262 crashes. The binary logit model demonstrated a good fit (based on the likelihood ratio test). Results obtained from this regional model are summarized in table 5 (only estimates pertaining to significant factors are included in the table) and discussed in the forthcoming section.
Driver and Vehicle Characteristics
Driver gender was found significant in the regional model with female drivers being 8% more susceptible to severe injuries than males (AOR = 1.08). Driver age was a significant factor as well. Younger drivers were less likely to be severely injured while older drivers were more likely to sustain severe injuries when compared to those aged between 25 and 65 (AORs = 0.71 and 1.47, respectively). The interaction between age and gender was also significant; younger female drivers were less and female drivers aged more than 65 were more likely to be severely injured in car crashes (AORs = 0.95 and 1.03).
Restraint use was also significant. The likelihood of having severe injuries for drivers with no restraint was 5.29 times more than drivers wearing seatbelts. As expected, air bag deployment was associated with severe injuries (AOR = 3.42). Drivers under the influence of alcohol or drugs were found to be 2.47 times more likely to have severe injuries compared to sober drivers. Conversely, drivers with passengers in their cars were slightly less likely to be seriously injured compared to drivers who traveled alone (AOR = 0.97).
Crash Types and Points of Impact
Drivers involved in head-on crashes were 2.82 times more likely to have severe injuries compared to those in rear-end crashes. Drivers in single-vehicle and angular crashes were also more likely to sustain severe injuries (AORs = 1.86 and 1.10, respectively). As expected, sideswipes were mainly associated with minor or no injuries (AOR = 0.34).
Vehicles impacted on the driver side were more likely to have a severely injured occupant than those vehicles impacted on the back (AOR = 1.82). Other areas of the vehicle (i.e., front, passenger side, top, or undercarriage) were associated with lower likelihoods of severe injuries, compared to rear of the vehicle (AORs = 0.92, 0.90, and 0.75, respectively).
The regional model revealed that drivers were more likely to sustain severe injuries in crashes occurring in non-daylight (i.e., dark, dawn, or dusk) conditions compared to crashes during daylight hours (AOR = 1.08). By contrast, crashes on non-dry surfaces (e.g., snow covered, icy, wet, dirty) were associated with less likelihood of severe injuries (AOR = 0.82). Weather conditions (i.e., adverse versus normal weather) were found insignificant.
Crashes in rural settings were significantly less injurious for drivers (AOR = 0.92). The interaction between crash type and rural or urban setting was also significant; drivers who had been in angular and single-vehicle crashes in rural settings were more likely to be seriously injured (AORs = 1.08 and 1.09), and those in sideswipes were less likely to have severe injuries (AOR = 0.71). This interaction was insignificant for head-on crashes.
Roadway speed limit was found significant: drivers involved in crashes on roadways with speed limits lower than 35 mph were 0.55 times less likely to have severe injuries compared to those in crashes on roads with a 35 to 55 mph speed limit. Crashes on roadways with posted speed limits higher than 55 mph were 1.47 times more likely to result in severe injuries than those on roadways with speed limits between 35 and 55 mph.
State and Regional Level Comparison
The goal of comparing the state outcomes with the sampled data collected as part of GES is to assess the capability of gaining insights on the Midwestern states when aggregated to the regional level. It should be noted that the GES data for the Midwest does cover 12 states within the region. The additional eight Midwestern states are Ohio, Indiana, Illinois, Michigan, Wisconsin, Minnesota, North Dakota, and South Dakota. GES does not provide data at the state level and as such, it was not possible to isolate the four states for which the individual analyses had been done.
The odds ratios (and corresponding confidence intervals) for the parameter estimates common across the four states and at the regional level are listed in table 6 and are graphically depicted in figure 2. The greatest similarities are for driver age and roadway surface condition, where the odds ratios estimated by the four state models are close and the odds ratios calculated by the GES-based model fall in their range. The same pattern is evident for the contrast between lower (less than 35 mph) and reference (35–55 mph) speed limits. For driver gender, the odds ratio calculated for the contrast between female and male drivers (1.08) is equal to the odds ratio for the same contrast in Kansas and very close to that of Iowa (1.07); however, the value of the odds ratio for this contrast is higher for Missouri and Nebraska (1.21 and 1.24, respectively). For higher speed limits (above 55 mph), the odds of sustaining severe injuries is 1.47 based on the GES model, which is between the odds ratios calculated for the states of Iowa and Missouri (1.34 and 1.60, respectively). However, the odds ratios estimated for the same contrast in Kansas and Nebraska are considerably higher (2.20 and 2.09, respectively).
In many cases, there was general agreement among the state models and the GES-based model on the association between certain levels of a factor (e.g., no restraint in use by driver) and severe injuries, but as expected, the strength of such association was not always similar. For crash type, the likelihood of a head-on crash sustaining greater injuries when compared injuries associated with rear-end crashes (OR = 2.82) was lower at the regional level than at the state level. The same is observed for sideswipes where the odds ratio of sustaining serious injuries (0.34) is lower at the regional level than at the individual states' models (range of 0.38 to 0.50). On the contrary, the GES-based odds ratios for alcohol and drug use and restraint use are higher than the highest odds ratios found in the individual states' models, indicating stronger associations between being under the influence of alcohol or drugs and having no restraint in use, and sustaining severe injuries by drivers. No confidence interval overlap is evident between the GES-based Midwestern model and the individual states models.
The air bag deployment factor could only be incorporated in the models of Iowa and Kansas due to the unavailability of precise data for the other two states. The odds of having serious injuries for cases in which air bags had been deployed were 3.42 times the cases without air bag deployment, based on the GES Midwestern model. Iowa and Kansas models showed weaker incompatible contrasts; i.e., odds ratios of 2.13 and 1.32, respectively. The comparison of confidence intervals revealed no overlap between the results of the three models. Therefore, the observations for air bag deployment yield no consensus for the Midwestern states considered in this study.
As noted earlier, point of impact was only available in for Iowa and Nebraska. The contrast between driver side and rear side of the vehicle was significant in predicting driver injury severity for both states, indicating higher likelihoods of serious injuries for drivers whose cars were impacted on driver side versus those involved in crashes in which the rear of the car was affected (odds ratios of 1.16 for Iowa and 1.82 for Nebraska). While the pattern observed in Nebraska was exactly the same as that calculated based on the GES data (with a wider confidence intervals for Nebraska), the contrast was smaller for Iowa, depicting a weaker difference between the levels of injury sustained by drivers for the two points of impact.
For rural versus urban settings, the directions of findings were completely opposite. The GES Midwestern model depicted slightly lower likelihoods of severe injuries for drivers in crashes occurring in rural settings (odds ratio of 0.92), whereas all the individual state models predicted higher likelihoods of such injuries in rural regions compared to urban settings, with odds ratios in the range of 1.71 to 2.55. This is the only contradiction between the states data and GES data-based models.
Table 5. Regional Model for the Likelihood of Severe Injuries
|Age < 25 years old||-0.34||0.006||3,820.60||0.71|
|Age > 65 years old||0.38||0.007||2,809.70||1.47|
|Passenger(s) present in the car||-0.03||0.004||62.5||0.97|
|Under influence of alcohol/drug||0.91||0.023||1,557.20||2.47|
|No restraint in use||1.67||0.009||32,223.50||5.29|
|Air bag deployed||1.23||0.005||62,029.30||3.42|
|Speed limit < 35 mph||-0.59||0.006||10,441.20||0.55|
|Speed limit > 55 mph||0.39||0.007||2,872.10||1.47|
|Point of impact: front||-0.09||0.009||93.5||0.92|
|Point of impact: driver side||0.6||0.01||3,307.10||1.82|
|Point of impact: passenger side||-0.11||0.011||97.3||0.9|
|Point of impact: top/under||-0.29||0.032||82||0.75|
|Head-on crashes in rural settings||-0.03||0.011||9.1||0.97|
|Angular crashes in rural settings||0.08||0.006||151.1||1.08|
|Sideswipes in rural settings||-0.35||0.016||465.6||0.71|
|Single-vehicle crashes in rural settings||0.08||0.007||155.5||1.09|
|Female drivers aged less than 25||-0.05||0.005||89.9||0.95|
|Female drivers aged more than 65||0.03||0.007||17.8||1.03|
|Number of unweighted observations||97,070|
|Number of weighted observations||12,252,262|
Table 6. Parameters Related to the Findings Across the Four Midwestern States and the
|Parameter||Logit Models (AOR and Confidence Intervals)|
|Crash Type||Head-on crashes||3.18||4.92||3.44||4.26||2.82|
|(compared to rear-end)||(2.83, 3.58)||(4.30, 5.63)||(3.18, 3.72)||(3.20, 5.67)||(2.72, 2.92)|
|(compared to rear-end)||(0.44, 0.56)||(0.34, 0.49)||(0.34, 0.43)||(0.40, .64)||(0.32, 0.36)|
|(compared to urban)||(1.62, 1.80)||(1.75, 2.10)||(2.43, 2.67)||(2.00, 2.30)||(0.90, 0.93)|
|(compared to males)||(1.02, 1.12)||(1.04, 1.13)||(1.18, 1.25)||(1.17, 1.31)||(1.06, 1.09)|
|Driver age||< 25 years||0.7||0.77||0.81||0.73||0.71|
|(compared to 25-65)||(0.65, 0.74)||(0.73, 0.82)||(0.79, 0.84)||(0.66, 0.80)||(0.70, 0.72)|
|> 65 years||1.49||1.39||1.28||1.42||1.47|
|(compared to 25-65)||(1.38, 1.61)||(1.27, 1.51)||(1.23, 1.34)||(1.27, 1.60)||(1.43, 1.50)|
|(compared to dry surface)||(080, 0.90)||(0.77, 0.91)||(0.89, 0.95)||(0.74, 0.89)||(0.81, 0.84)|
|Alcohol/ drug impairment||1.64||1.32||1.36||1.74||2.47|
|(compared to sober driving)||(1.55, 1.74)||(1.25, 1.40)||(1.31, 1.40)||(1.54, 1.97)||(2.29, 2.67)|
|No restraint in use||3.59||4.24||2.74||2.7||5.29|
|(compared to seat belt in use)||(3.31, 3.88)||(3.82, 4.70)||(2.60, 2.89)||(2.35, 3.10)||(5.13, 5.45)|
|Speed limit||< 35 mph||0.62||0.39||0.56||0.58||0.55|
|(compared to 35-55 mph)||(0.57, 0.67)||(0.36, 0.43)||(0.53, 0.58)||(0.51, 0.66)||(0.54, 0.56)|
|> 55 mph||1.34||2.2||1.6||2.09||1.47|
|(compared to 35-55 mph)||(1.22, 1.47)||(2.01, 2.41)||(1.55, 1.66)||(1.84, 2.38)||(1.44, 1.51)|
|Air bag deployed||2.13||1.32||NA||NA||3.42|
|(compared to no air bag deployment)||(1.99, 2.29)||(1.21, 1.43)||(3.36, 3.47)|
|Point of impact: Driver side||1.16||NA||NA||1.82||1.82|
|(compared to rear side)||(1.06, 1.27)||(1.44, 2.31)||(1.76, 1.89)|
The goal of this paper was to investigate the factors associated with severe (fatal or incapacitating) injuries sustained by drivers in crashes, with a focus on the Midwestern states in the central part of the United States. The majority of the findings from each state were consistent with the literature. For example, our findings showed that females and older drivers were more susceptible to severe injuries in car crashes, and this has been observed in previous studies (O'Donnell and Connor 1996; Bedard et al. 2002). Seat belt use had an even greater effectiveness at the state level when compared to estimates from other studies (Martin, Crandall, and Pilkey 2000; Malliaris, Digges, and DeBlois 1995; Evans 1993; Bedard et al. 2002). Alcohol and drug use is another explanatory variable in our model that has consistently shown to increase the likelihood of severe injuries (Evans and Frick 1993; Evans 1990; Keall, Frith, and Patterson 2004; Mayhew et al. 1986; Sjogren et al. 1997; Zador, Krawchuk, and Voas 2000). Head-on crashes were associated with the highest odds of sustaining severe injuries, and this is consistent with the findings of O'Donnell and Connor (1996).
FIGURE 2 Comparison Between State and Regional Estimates
(Lines Indicate Confidence Intervals of the Estimates)
There were differences that are worth noting. The four Midwestern states were not consistent with respect to crash type, with severe injuries more likely in single-vehicle crashes compared to rear-end crashes in Iowa and Missouri, and equally likely in the two crash types in Kansas and Nebraska. Crashes in rural settings were more likely to cause severe injuries than those occurring in urban crashes at the state level. However, the opposite was observed at the regional level, underscoring the impact of potential information loss when aggregating to the general region. Although this study had crash data for only four states, it clearly demonstrates that differences do exist from the state to regional level. Population distribution differences and geographical properties of different regions of the Midwest are influential in the disparity observed, even though the same modeling technique was used in all the models developed.
The four states examined may also have more rural characteristics when compared to other Midwestern states such as Illinois, Michigan, and even Indiana, with much larger metropolitan areas (e.g., Chicago, Detroit, and Indianapolis). Research has shown that differences in injury patterns in rural and urban settings may be due to the variations in availability of trauma care systems and distance from these facilities (Brodsky and Hakkert 1988; Bentham 1986). Therefore, with more crashes occurring in areas with access to advanced medical facilities, these differences may lessen with other factors influencing urban crashes, e.g., roadway geometry, type of vehicle, distractions, etc., playing a larger role in the severity of injuries.
There are many data quality issues with using crash data at the state and national level related to underreporting, misclassification, and omitted data. At the national level, crashes are sampled and reported as four separate regions: Northwest, Midwest, South, and West. There is no information to the analyst to connect the data back to a specific state. Given that GES data is a weighted sample, it is not actually possible to have a direct comparison of the models developed at the state level to the sampled data state using GES data.
The state crash databases used in this study had several shortcomings that resulted in the need to exclude many crash records from the data used in statistical analyses. There were also missing crash, vehicle, and driver attributes for some state models (i.e., point of impact in Kansas, air bag deployment, point of impact, and drug use in Missouri, and air bag deployment in Nebraska). The same problem was identified by Ghazizadeh and Boyle (2009) in their study of driver distraction. Crash factors were not as comprehensive at the state level as initially expected. Missouri had the highest percentage (over 85%) of reported crashes that included the explanatory variables needed for the statistical model.
The crash report forms for each state for the years studied provide some insights on the relatively low numbers for some explanatory factors. In all four states, there was no specific callout for the various types of distraction, but instead all four states had "contributing circumstances" as a variable with distraction or inattention as a category. The Missouri and Nebraska forms did include distraction as a check box, whereas in Iowa and Kansas, categories of distraction were to be entered under a generic contributing circumstances area. The form used in Missouri included check-boxes for factors with several potential categories, which eliminated the need to refer to code sheets (as was the case in Iowa). Standardization of information could provide researchers better insights on safety issues and also allow better comparisons across states, which can have implications at the regional and national level. It is recognized that some improvements may actually lengthen the already cumbersome task of data entry, but could actually decrease the chance of non-reporting and even misclassification. Prioritizing information based on the findings in the literature of injury severity may also help officers in collecting the most critical information surrounding a crash.
Non-reporting and misclassification of conditions surrounding a crash is another potential issue that can impact the reliability of the estimates. However, past studies have shown that even though the crash reporting systems might not be ideal, estimates driven based solely on crash databases still offer valuable insights. For example, Cummings (2002) compared estimates of fatalities based on seat belt use for police-reported data and data based on trained crash investigators' reports and found no substantial difference, and Guo, Eskridge, Christensen, Qu, and Safranek (2007) showed that misclassifications of seat belt and alcohol use in Nebraska biased the odds ratio estimates of injury only slightly.
Recent studies have explored more rigorous statistical methods to predict crash injury severity outcomes (for a recent review of the methodologies, see Savolainen et al. 2011). For example, Anastasopoulos and Mannering (2002) examined the utility of the random-parameter logit model that used a less detailed crash profile (including injury outcomes, roadway geometrics, pavement condition, general weather, and traffic characteristics only) relative to the fixed-parameter model. Although the models based on individual crash-data provide better fit, their findings suggest that these models would be difficult to use for assessing the changes in injury severities caused by safety countermeasures because of the large number of variables that need to be determined for each crash. The random-parameter model, on the other hand, provides reasonable accuracy while also being easier to build. In other studies, Chang and colleagues used a non-parametric classification tree model in analyzing traffic injury severity (Chang and Chien 2013; Chang and Wang 2006). These methods and others can aid researchers in appropriately connecting the right model to the data being examined.
Crash data clearly has limitations in terms of exposure and standardization of information, but they do provide useful information on traffic, vehicle, and environmental factors that can be examined further in other test settings (e.g., simulator, test track, naturalistic studies). However, it is important to recognize the differences that exist, even within states in one geographic region. Future studies should examine the differences in rural/urban areas and crash type over a larger portion of the Midwest, and over a longer time period. It would also be of great interest to examine the underlying reasons for the disparity observed in injury trends across various states. Research in this direction can help provide insights for more effective crash countermeasures that can guide safer driver behaviors and driving environments.
A version of this paper was presented at the 3rd International Conference on Road Safety and Simulation (in Indianapolis, IN), September 2011. We would like to acknowledge the Midwest Transportation Consortium for sponsoring this project, and Iowa Department of Transportation (DOT), Kansas DOT, Missouri DOT, and Nebraska Department of Roads for providing us with crash data. We would also like to thank the editor and two anonymous reviewers for their helpful comments on previous versions of this article.
Abdel-Aty, M. 2003. "Analysis of driver injury severity levels at multiple locations using ordered probit models." Journal of Safety Research no. 34 (5):597-603.
Al-Ghamdi, A.S. 2002. "Using logistic regression to estimate the influence of accident factors on accident severity." Accident Analysis & Prevention no. 34 (6):729-741.
Allison, P. D. 1999. Logistic regression using the SAS system: Theory and application. Cary, NC: SAS Institute Inc.
Ananth, C.V., and D.G. Kleinbaum. 1997. "Regression models for ordinal responses: a review of methods and applications." International Journal of Epidemiology no. 26 (6):1323-1333.
Awadzi, K. D., S. Classen, A. Hall, R. P. Duncan, and C. W. Garvan. 2008. "Predictors of injury among younger and older adults in fatal motor vehicle crashes." Accident Analysis & Prevention no. 40 (6):1804-1810.
Baum, H. M., A. K. Lund, and J. K. Wells. 1989. "The mortality consequences of raising the speed limit to 65 mph on rural interstates." American Journal of Public Health no. 79 (10):1392-1395.
Bedard, M., G. H. Guyatt, M. J. Stones, and J. P. Hirdes. 2002. "The independent contribution of driver, crash, and vehicle characteristics to driver fatalities." Accident Analysis & Prevention no. 34 (6):717-727.
Bentham, G. 1986. "Proximity to hospital and mortality from motor vehicle traffic accidents." Social Science & Medicine no. 23 (10):1021.
Brodsky, H., and A. S. Hakkert. 1988. "Risk of a road accident in rainy weather." Accident Analysis & Prevention no. 20 (3):161-176.
Chang, L.Y., and J.T. Chien. 2013. "Analysis of driver injury severity in truck-involved accidents using a non-parametric classification tree model." Safety Science no. 51 (1):17-22.
Chang, L.Y., and H.W. Wang. 2006. "Analysis of traffic injury severity: An application of non-parametric classification tree techniques." Accident Analysis & Prevention no. 38 (5):1019-1027.
Cochran, W.G. 1952. "The χ2 test of goodness of fit." The Annals of Mathematical Statistics:315-345.
Compton, C. P. 2005. "Injury severity codes: A comparison of police injury codes and medical outcomes as determined by NASS CDS Investigators." Journal of Safety Research no. 36 (5):483-484.
Connor, J., R. Norton, S. Ameratunga, and R. Jackson. 2004. "The Contribution of Alcohol to Serious Car Crash Injuries." Epidemiology no. 15 (3):337.
Cummings, P. 2002. "Association of seat belt use with death: a comparison of estimates based on data from police and estimates based on data from trained crash investigators." Injury Prevention no. 8 (4):338-341.
Evans, L. 1990. "The fraction of traffic fatalities attributable to alcohol." Accident Analysis & Prevention no. 22 (6):587-602.
Evans, L. 1993. "Restraint effectiveness, occupant ejection from cars, and fatality reductions." Accident Analysis & Prevention no. 22:167-175.
Evans, L., and M. C. Frick. 1993. "Alcohol's effect on fatality risk from a physical insult." Journal of Studies on Alcohol no. 54 (4):441-449.
Evans, L., and M. C. Frick. 1994. "Car mass and fatality risk: has the relationship changed?" American Journal of Public Health no. 84 (1):33-36.
Farmer, C. M., E. R. Braver, and E. L. Mitter. 1997. "Two-vehicle side impact crashes: The relationship of vehicle and crash characteristics to injury severity." Accident Analysis & Prevention no. 29 (3):399-406.
Ghazizadeh, M., and L. N. Boyle. 2009. "Influence of driver distractions on the likelihood of rear-end, angular, and single-vehicle crashes in Missouri." Transportation Research Record no. 2138:1-5.
Guo, H., K. M. Eskridge, D. Christensen, M. Qu, and T. Safranek. 2007. "Statistical adjustment for misclassification of seat belt and alcohol use in the analysis of motor vehicle accident data." Accident Analysis & Prevention no. 39 (1):117-124.
Hill, J. D., and L. N. Boyle. 2005. Analyzing Severe Injury Risk for Crashes Nationally and Within Iowa. Paper read at Mid-Continent Transportation Research Symposium, at Ames, Iowa.
Huelke, D. F., and C. P. Compton. 1995. "The effects of seat belts on injury severity of front and rear seat occupants in the same frontal crash." Accident Analysis & Prevention no. 27 (6):835-838.
Keall, M. D., W. J. Frith, and T. L. Patterson. 2004. "The influence of alcohol, age and number of passengers on the night-time risk of driver fatal injury in New Zealand." Accident Analysis & Prevention no. 36 (1):49-61.
Khattak, A. J., P. Kantor, and F. M. Council. 1998. "Role of Adverse Weather in Key Crash Types on Limited-Access: Roadways Implications for Advanced Weather Systems." Transportation Research Record no. 1621 (-1):10-19.
Khattak, A. J., and K. K. Knapp. 2001. "Interstate Highway Crash Injuries During Winter Snow and Nonsnow Events." Transportation Research Record no. 1746 (-1):30-36.
Khorashadi, A., D. Niemeier, V. Shankar, and F. Mannering. 2005. "Differences in rural and urban driver-injury severities in accidents involving large-trucks: An exploratory analysis." Accident Analysis & Prevention no. 37 (5):910-921.
Kim, K., L. Nitz, J. Richardson, and L. Li. 1995. "Personal and behavioral predictors of automobile crash and injury severity." Accident Analysis & Prevention no. 27 (4):469-481.
Klauer, S. G., T. A. Dingus, V. L. Neale, J. D. Sudweeks, and D. J. Ramsey. 2006. The impact of driver inattention on near-crash/crash risk: an analysis using the 100-car naturalistic driving study data. Washington, DC: NHTSA.
Kweon, Y. J., and K. M. Kockelman. 2003. "Overall injury risk to different drivers: combining exposure, frequency, and severity models." Accident Analysis & Prevention no. 35 (4):441-450.
Malliaris, A. C., K. H. Digges, and J. H. DeBlois. 1995. "Evaluation of Airbag Field Performance." SAE transactions no. 104 (6):1513-1534.
Martin, P. G., J. R. Crandall, and W. D. Pilkey. 2000. "Injury trends of passenger car drivers in frontal crashes in the USA." Accident Analysis & Prevention no. 32 (4):541-557.
Mayhew, D. R., A. C. Donelson, D. J. Beirness, and H. M. Simpson. 1986. "Youth, alcohol and relative risk of crash involvement." Accident Analysis & Prevention no. 18 (4):273-287.
Miller, T. R., D. C. Lestina, and R. S. Spicer. 1998. "Highway crash costs in the United States by driver age, blood alcohol level, victim age, and restraint use." Accident Analysis & Prevention no. 30 (2):137-50.
Muelleman, R. L., and K. Mueller. 1996. "Fatal Motor Vehicle Crashes: Variations of Crash Characteristics within Rural Regions of Different Population Densities." The Journal of Trauma: Injury, Infection, and Critical Care no. 41 (2):315.
Neyens, D. M., and L. N. Boyle. 2008. "The influence of driver distraction on the severity of injuries sustained by teenage drivers and their passengers." Accident Analysis & Prevention no. 40 (1):254-259.
NHTSA. 2005. National Automotive Sampling System (NASS) General Estimates System (GES): Analytical user's manual 1988-2005. Washington, DC: U.S Department of Transportation.
———. 2008. General Estimates System (GES) 2001-2006 2008a [cited June 2008]. Available from ftp://ftp.nhtsa.dot.gov/GES.
———. 2008b. Traffic Safety Facts 2007 Data: Rural/Urban Comparison. Washington, DC: U.S. Department of Tranportation.
———. 2009. Traffic Safety Facts 2008: A compilation of motor vehicle crash data from the fatality analysis reporting system and the general estimates system. Washington, DC: U.S Department of Transportation.
O'Donnell, C. J., and D. H. Connor. 1996. "Predicting the severity of motor vehicle accident injuries using models of ordered multiple choice." Accident Analysis & Prevention no. 28 (6):739-753.
Renski, H., A. J. Khattak, and F. M. Council. 1999. "Effect of speed limit increases on crash injury severity: Analysis of single-vehicle crashes on North Carolina Interstate highways." Transportation Research Record no. 1665 (1):100-108.
Savolainen, P., and F. Mannering. 2007. "Probabilistic models of motorcyclists' injury severities in single- and multi-vehicle crashes." Accident Analysis & Prevention no. 39 (5):955-963.
Savolainen, P.T., F.L. Mannering, D. Lord, and M.A. Quddus. 2011. "The statistical analysis of highway crash-injury severities: A review and assessment of methodological alternatives." Accident Analysis & Prevention no. 43 (5):1666-1676.
Sjogren, H., U. Bjornstig, A. Eriksson, U. Ohman, and A. Solarz. 1997. "Drug and Alcohol Use Among Injured Motor Vehicle Drivers in Sweden: Prevalence, Driver, Crash, and Injury Characteristics." Alcoholism: Clinical and Experimental Research no. 21 (6):968-973.
Tavris, D. R., E. M. Kuhn, and P. M. Layde. 2001. "Age and gender patterns in motor vehicle crash injuries: importance of type of crash and occupant role." Accident Analysis & Prevention no. 33 (2):167-172.
Violanti, J. M. 1998. "Cellular phones and fatal traffic collisions." Accident Analysis & Prevention no. 30 (4):519-524.
Watt, K., D. M. Purdie, A. M. Roche, and R. McClure. 2006. "Injury severity: role of alcohol, substance use and risk-taking." Emergency Medicine Australasia no. 18 (2):108-117.
Zador, P. L., S. A. Krawchuk, and R. B. Voas. 2000. "Alcohol-related relative risk of driver fatalities and driver involvement in fatal crashes in relation to driver age and gender: An update using 1996 data." Journal of Studies on Alcohol no. 61 (3):387-395.
Zhang, J., S. Fraser, J. Lindsay, K. Clarke, and Y. Mao. 1998. "Age-specific patterns of factors related to fatal motor vehicle traffic crashes focus on young and elderly drivers." Public Health no. 112 (5):289-295.