- Email: [email protected]
The impact of driver alcohol use on crash severity: A crash specific analysis
Transportation Research Part E 41 (2005) 421–437 www.elsevier.com/locate/tre
The impact of driver alcohol use on crash severity: A crash specific analysis Thomas L. Traynor
*
Department of Economics, Wright State University, Dayton, OH 45435, United States
Abstract This study uses a crash specific data set that is supplemented with location based socioeconomic data to estimate the impact of driver alcohol use on average crash severity. Logit estimates indicate that crashes in which the at-fault drivers had been drinking are more likely to result in a severe injury or death than are crashes caused by sober drivers. Ordered logit estimates indicate that at-fault driver alcohol use increases the expected highest degree of injury resulting from a crash, and Tobit estimates indicate that the number of injuries or deaths per crash increase an average of 0.71 when the at-fault driver has been drinking. Moreover, at-fault driver alcohol use worsens the severity of crashes relative to not-at-fault parties. Collectively, these results indicate that at-fault drinking drivers are involved in more violent crashes and produce more serious injuries to not-at-fault parties than at-fault sober drivers. 2005 Elsevier Ltd. All rights reserved. Keywords: Crash; Alcohol; Impaired driving; Crash severity
1. Introduction There were 42,815 traffic fatalities in 2002, 40% of which were alcohol related. The National Highway Traffic Safety Administration estimated that more than 2,163,000 alcohol involved *
Tel.: +1 937 775 2025/3070. E-mail address: [email protected]
1366-5545/$ - see front matter 2005 Elsevier Ltd. All rights reserved. doi:10.1016/j.tre.2005.03.005
422
T.L. Traynor / Transportation Research Part E 41 (2005) 421–437
crashes occurred in 2002, killing 17,419 people and injuring an estimated 513,000.1 Of the estimated 0.62 alcohol related fatalities that occurred per 100 million vehicle miles traveled (VMT) in 2002, most (0.53 per 100 million VMT) involved high blood alcohol concentrations of 0.08 or above.2 Given this information, it is not surprising that numerous studies have been conducted to identify factors that influence alcohol related fatalities and injuries. This article provides an evaluation of individual crashes to determine the impact of driver alcohol use on crash severity, conveying a unique vantage point to the impact of driver alcohol use. In so doing, this analysis contributes to the existing body of research by exploring whether the fatality and injury levels associated with drunk driving are not only caused by increased numbers of crashes, but by more violent crashes as well. The resulting estimates lend support for the hypothesis that alcohol consumption by at-fault drivers increases the severity of crashes. This is true of both the severity of injury suffered by victims as well as the numbers of injuries and fatalities suffered per crash. These results indicate that drinking and driving policies must not only be evaluated by the reduction in the number of crashes they bring about, but also by the reduction in crash severity they bring about. The remainder of the paper proceeds with a discussion of issues and existing evidence surrounding alcohol consumption and crash severity in Section 2. Sections 3 and 4, respectively, develop the model that is estimated, and describes the data that are used in the analysis. Section 5 reports and evaluates the results, while concluding comments are made in the final section. 2. The impact of driver alcohol use Numerous researchers have studied the incidence of driver alcohol use, identified factors that influence the incidence of alcohol related crashes, and estimated the cost to society of driver alcohol use. Studies of the incidence of alcohol use in crashes primarily focus on the hospital records of crash victims. Perper et al. (1993) evaluated coroner and emergency room records, concluding that passengers face the same risk of fatality as their alcohol-positive drivers. Waller et al. (1997) examined blood samples from patients treated for automobile crash injuries, finding that alcohol was associated with more severe crashes than those with drug involvement or those with no drug/ alcohol involvement. Zador (1991) examined BAC concentrations reported to the Fatality Analysis Reporting System to conclude that each 0.02% increase in BAC doubles the risk of involvement in a fatal single car crash, and that the likelihood of a crash was at least nine times greater at BACs in the 0.05–0.09% range than at 0% BAC for all age groups. Similarly, Robertson and Drummer (1994) evaluated 341 driver fatalities to determine that the relative risk of having a fatal crash increased with increased BAC. Ferrara et al. (1994) reviewed 38 studies of the impact of low alcohol concentrations (below 0.1%) on driving, and found no consensus on whether or not low BAC levels significantly contribute to crash losses. Several investigations have been conducted on the impact of alcohol on the behavior of drivers involved in crashes. Haffner and GrawÕs (1996) examination of 625 crashes caused by alcoholimpaired drivers revealed that speed was the most common cause of crashes involving an alco1 2
Initiatives to Address Impaired Driving, National Highway Traffic Safety Administration, December 2003, pp. 4–5. Initiatives to Address Impaired Driving, National Highway Traffic Safety Administration, December 2003, pp. 7–9.
T.L. Traynor / Transportation Research Part E 41 (2005) 421–437
423
hol-impaired driver at all BAC segments. Peek-Asa and Kraus (1996) found that alcohol-impaired motorcycle drivers were more likely to be speeding and less likely to be wearing a helmet in the event of a crash, while Solnick and Hemenway (1994) discovered that drunk drivers are more likely to flee after hitting a pedestrian. Scholars have also explored the impact of environmental or demographic factors on the risk and severity of alcohol-related crashes. Owens and Sivak (1996) concluded that night time driving did not have a significant impact on alcohol-related traffic deaths. Orsay et al. (1994) used a sample of patients admitted to a Level I trauma center and found that drivers impaired by alcohol and/or drugs were younger, and were more often male than the unimpaired drivers. Muelleman and Mueller (1996) concluded that low population density was correlated with more frequent alcohol use and high intoxication levels among driver fatalities. Ruhm (1996) investigated the relationship between economic growth and drunk driving, finding that drunk driving fatality rates rose as national income rose. McCarthy (2003) empirically evaluated the impact of alcohol availability using the geographic density of alcohol licenses, finding that general alcohol off-site licenses are beneficial while general alcohol on-site licenses are detrimental to highway safety. Estimates of the impact of beer/wine license density did not produce uniform results. In a theoretical examination of the alcohol availability issue, Lee (1997) argued that market factors that determine the geographical distribution of alcohol establishments would influence the incidence of drunk driving. The National Highway Traffic Safety Administration (NHTSA) recently published a general review of alcohol related crash fatalities, noting that the annual total of alcohol related crash fatalities leveled off in the 1990s after declining during the previous decade.3 Miller and Blewden (2001) presented the most recent refinement of a method for estimating the costs of crashes, including costs associated with driver alcohol use. The NHTSAÕs latest update of its Alcohol Cost Fact Sheets (2002) made use of this method to estimate that alcohol related crashes had a total public cost of $114.3 billion in 2000, that the average cost of alcohol-related crash fatalities was $3.5 million, and that the average cost of alcohol-related crash injuries was $99,000. These costs translate to $5.80 per mile driven at a BAC of 0.10 or greater, and $1.00 per drink consumed.4 Traynor (1993) evaluated individual crashes to determine the impact of various factors on the proportion of crash injuries suffered by not-at-fault parties. The results were mixed. While Tobit estimates indicated that at-fault driver alcohol use significantly increases the proportion of crash injuries and fatalities incurred by other parties, ordered logit estimates indicated that the degree of injury suffered by drinking drivers relative to others (no injury–minor injury–severe injury–fatality) is not significantly impacted by at-fault driver alcohol use. The analysis that follows extends this research by estimating the impact of alcohol involvement on a wider array of crash severity measures (the probability that injuries are suffered, total injuries and fatalities, as well as injury/ fatality losses to not-at-fault parties), measures the marginal effects of alcohol involvement on
3
Initiatives to Address Impaired Driving, National Highway Traffic Safety Administration, December 2003, p. 7. Impaired Driving in the United States: State Alcohol Cost Fact Sheets, National Highway Traffic Safety Administration, 2002, p. 1. These costs were measured using data from a variety of sources. A description of the methods used to calculate these estimates is included in Impaired Driving in the United States: State Cost Fact Sheet UserÕs Guide. National Highway Traffic Safety Administration, 2002, pp. 2–5. 4
424
T.L. Traynor / Transportation Research Part E 41 (2005) 421–437
crash severity, and assesses the sensitivity of the relationship with the use of multiple measures of crash costs and alcohol use. The analysis further contributes to the existing literature by combining crash specific and socioeconomic data to evaluate the differential impact of at-fault driver alcohol use on the typical losses suffered per crash. This differs from prior analyses by focusing specifically on the alcohol-crash cost relationship on a per crash basis, while controlling for a variety of socio-economic factors. The resulting estimates provide evidence that alcohol involved crashes are more violent than other crashes.
3. Statistical models To estimate the impact of driver alcohol use on crash severity, an econometric model is developed in which individual crashes are evaluated to observe how varying levels of alcohol involvement impact the injury severity level or the number of injuries/fatalities suffered by involved parties. The hypotheses to be tested in this analysis are: • Crashes caused by drinking drivers result, on average, in more severe injuries and greater numbers of injured parties than crashes caused by sober drivers. • Crashes caused by drinking drivers result, on average, in more severe injuries and greater numbers of injured parties to other victims than crashes caused by sober drivers. The empirical model is a modified form of those that have been used to test the relationship between driving conditions and crash outcomes with similar data. Its general form is LOSS ¼ a þ / ALCOHOL þ bZ þ cF þ e
ð1Þ
where LOSS is a vector of observed sample values that measure crash severity (via injuries or fatalities), and ALCOHOL is a vector of observations that measure at-fault driver alcohol use. Z represents the matrix of control variables used in the analysis, and F is a matrix of fixed effects terms. Lastly, e is the stochastic error term. The parameter / measures the average impact of atfault driver alcohol use upon the severity of the crash, while b is a vector of parameters that estimate the impact of a control factor on crash severity, and c is a vector of parameters that estimate the impact of the fixed effects variables on severity. Since both of the key variables in this analysis (LOSS and ALCOHOL) are inherently difficult to precisely measure, this analysis employs multiple proxies for each variable, each of which is described in Section 4. Estimates are generated using each of the measures developed for LOSS and ALCOHOL. In addition, to determine whether the impact of alcohol on all crashes is not simply the result of changes in the severity of single vehicle crashes, the estimates are calculated both by using all crashes as well as by using only multiple unit (vehicles, cyclists, pedestrians) crashes.5
5
If the estimates using the two different sets of data are not substantively different, this would indicate that the impact of alcohol on crash severity is not driven primarily by single vehicle crashes. Single vehicle crashes do not need to be separately estimated to conduct this test.
T.L. Traynor / Transportation Research Part E 41 (2005) 421–437
425
The vector of control factors, Z, consists of a variety of measures of driving conditions at the time of each crash, demographic information about the at-fault driver, socioeconomic factors for the county of the crash, as well as fixed effect measures. A series of dummy variables measure the driving conditions and location factors at each crash that impact its severity. These variables control for lighting, road, and weather conditions, as well as for crashes that occur in intersections and/or rural locations. Additionally, demographic characteristics of the at-fault driver are included to control for the impact that driver sex and age have upon crash severity. Location factors that impact crash severity are also included. In an attempt to control for socioeconomic factors that may influence crash severity, economic and education data which are aggregated by the county of the crash site are included in the analysis. Variables measuring the mass and age of the at-fault vehicle are also included to control for the impact these factors have upon the severity of crashes to involved parties. Finally, hour and month fixed effects variables were included to improve the efficiency of the estimates.
4. The data The bulk of the data set consists of detailed police reports collected by the Ohio Integrated Traffic Reporting System (OITRS) for a single calendar year. Detailed information regarding the time, location, road type, road conditions, weather conditions, parties involved, crash severity, vehicle type, and driver characteristics at the time of the crash as determined by the reporting officer are collected by OITRS.6 These data are used in conjunction with county level socio-economic data to test the above hypotheses.7 4.1. Dependent variables As mentioned above, LOSS is measured in a variety of ways—seven in all. First, INJURY is a binary variable that is one if the crash resulted in any injury or fatality, and is zero otherwise. Second, SERIOUS is a binary variable that denotes whether or not the crash resulted in an incapacitating injury or fatality. Given the binary nature of INJURY and SERIOUS, estimates are obtained via logit analysis when either of these two dependent variables are used. Third, SEVERITY is an ordered variable which takes on a value from one, if the crash resulted in property damage only, to progressively higher values for crashes resulting in more severe injuries, up to a value of 5 for a fatal crash. SEVERITY is a more comprehensive than INJURY or SERIOUS.8 Also, 6
The OITRS database included a total of more than 330,000 observations, which consists of all crashes that occurred in the 1995 calendar year. Due to extremely lengthy processing times involved in calculating estimates from the complete OITRS database, a random sample of 15% of the crashes (48,828) was drawn from which estimates were calculated. This data set was further reduced to 38,000 observations through the removal of all single vehicle crashes so that the estimates based on multi-unit crashes could be calculated. 7 Although this data set clearly does not allow one to measure the behavior of drivers who do not suffer accidents, the hypothesis that is evaluated in this paper refers only to the effect of driver alcohol use upon relative accident costs. Therefore, only accidents must be observed to test this hypothesis. 8 The values given to an accidentÕs severity are: 1—property damage only, 2—complaint of pain, 3—visible injury, 4— severe injury, 5—fatal.
426
T.L. Traynor / Transportation Research Part E 41 (2005) 421–437
due to its ordinal nature, estimates using SEVERITY as the dependent variable are calculated via ordered logit.9 Two dependent variables were developed to focus on the severity of injury to not-at-fault parties. EXTERNAL is a binary variable that is equal to one if the highest degree of injury suffered by a not-at-fault party is greater than the degree of injury suffered by the at-fault driver, and is zero otherwise. Next, EXTINJ is a binary variable that is equal to one if at least one not-at-fault party suffered an injury while the at-fault driver escaped without injury (and is zero otherwise). Due to the binary nature of these variables, estimates using EXTERNAL or EXTINJ as the dependent variable are calculated using logit analysis. Finally, two dependent variables measure of the number of injuries and fatalities in each crash. NUMINJ is the number of injured and killed parties in a crash. In addition, the variable NUMEX focuses on crash severity to not-at-fault parties by measuring the number of not-at-fault individuals injured or killed in a crash. Since NUMINJ and NUMEX are quantitative variables that have distributions that are censored from below at zero, estimates are obtained for these equations using Tobit analysis.10 4.2. Explanatory variables The two measures of ALCOHOL that are used in the analysis are DRINKING, which is equal to one if the at-fault driver is reported to have been drinking at all (and zero otherwise), and IMPAIRED if the at-fault driverÕs blood alcohol content exceeds the legal limit for the State of Ohio. As mentioned earlier, estimates are obtained both for all crashes and for crashes involving two or more units.11 Thus, all permutations of the Proxies for LOSS and ALCOHOL together with the two data sets create a total of 24 sets of estimates that were calculated for this analysis.12 DARK is one if the crash occurred in darkness, and is zero otherwise. PRECIPIT is one if there was precipitation at the time of the crash (zero otherwise), while SLIP is one if the road was deemed by the reporting officer to have been slippery at the time of the crash (zero otherwise). If the crash took place in a rural area, RURAL carries a value of one and is zero otherwise. HIGHSPD denotes that the crash took place in a high speed zone (posted speed limit of fifty miles per hour or greater), while HIGHWAY denotes that the crash took place on multiple lane divided highway. If the crash took place at an intersection, INTERSEC carries a value of one and is zero otherwise. MALE is a dummy variable that is equal to one if the at-fault driver is male, and is included to account reflect driving behavior differences associated with the at-fault driverÕs gender. Two 9
Most basic econometrics texts provide an intuitive description of which statistical methods are most appropriate for various types of limited dependent variable models. See for example, Gujarati (2003, pp. 580–582). 10 Although Poisson regression is often used to calculate estimates in cases where the dependent variable consists of count data (such as the number of injuries per crash), since the distribution of injuries per crash corresponds to that of the right tail of the normal distribution, Tobit was selected as the estimation method. 11 A ÔunitÕ is any involved motor vehicle (including motorcycles), bicycle, or pedestrian. For a crash to be included in the OITRS data set, at least one of the units must be a motor vehicle. 12 Due to the large number of models estimated, only selected estimates are presented in detail. The remaining estimated models will be summarized in the text and in Table 7. All of the other results are available from the author upon request.
T.L. Traynor / Transportation Research Part E 41 (2005) 421–437
427
dummy variables were created to control for the impact of age on crash severity. YOUNG is one if the at-fault driver is under the age of 20, and OLD identifies crashes in which the at-fault driver is over 65 years of age. INC is the per capita income for the county in which the crash occurred.13 UR is equal to the unemployment rate in the county during the month in which the crash occurred. ED represents the average educational attainment for the county in which the crash happened. TYPE represents the expected losses faced by at-fault drivers based on the crash worthiness of their vehicles as measured by weight class. TYPE carries a value of one through thirteen with a value of one representing the lowest weight class and thirteen representing the highest weight class for the at-fault vehicle.14 YEAR is the year that the at-fault vehicle was manufactured. In addition, fixed effects variables were included in the model to improve the precision of the estimates. First, hour of day dummy variables were included to pick up time of day fixed effects. Second, monthly dummy variables were included to account for time of year fixed effects.15 A variety of interaction variables were also developed to try to improve the overall fit of the model, but only TYPEYEAR, which is the product of the TYPE and YEAR variables, proved to be statistically significant and was included in the model. Table 1 presents basic descriptive statistics for the variables used in this analysis (excluding the fixed effects variables).16
5. Estimates 5.1. The impact of at-fault driver alcohol use on crash severity Estimates of the equations that evaluate the impact of alcohol use on the severity of crashes are presented in Tables 2 and 3. The estimates presented in Table 2 indicate that detected alcohol consumption by the at-fault driver has a statistically significant positive impact on the probability of a crash resulting in serious injury or death at the 99% level of confidence. Additionally, estimates using only multiple unit crashes are similar in magnitude and significance to those of the all inclusive data set, indicating the results are not driven solely by an increase in single car crashes.17 The odds ratio estimates of at-fault driver alcohol use on all the dependent variables used in this study are presented in Table 7. These estimates are measures of the average proportional change in the odds ratio attributable to at-fault driver alcohol use after controlling for the other variables in the 13
Measures of income similar to INC were used in cross-sectional analyses by Peltzman (1975), as well as Traynor and McCarthy (1993). 14 For evidence regarding the relationship between vehicle weight and injury risk, see Kahane (2003) and Frei et al. (1999). 15 County fixed effects were initially included in the models, but were always insignificant, presumably due to the inclusion of the other control factors, and were therefore dropped from the analysis. 16 Due to space considerations, the estimates for the time of day and time of year (month) fixed effects parameter estimates are omitted from the tables. 17 Two of the parameter estimates for the control variables are worth noting. First, the parameter estimates for YOUNG indicate that young at-fault drivers increase the severity of all crashes on average while decreasing the severity of multi-unit crashes, suggesting that single car crashes caused by young drivers are more violent than multi-unit crashes caused by young drivers.
428
T.L. Traynor / Transportation Research Part E 41 (2005) 421–437
Table 1 Descriptive statistics Variable serious injury severity external extinj numinj numex impaired drinking dark precipit slip rural ur inc young old male type ED highspd intersec highway year typeyear
All crashes
Multi-unit crashes
Mean
Std. Dev.
Mean
Std. Dev.
0.026 0.373 4.432 0.158 0.263 0.635 0.402 0.039 0.059 0.227 0.238 0.343 0.396 4.785 22.605 0.245 0.114 0.609 3.598 2.518 0.271 0.534 0.140 86.229 315.035 N = 48,828
0.160 0.484 0.841 0.365 0.440 1.058 0.836 0.193 0.236 0.419 0.426 0.475 0.489 1.279 3.445 0.430 0.318 0.488 1.728 0.267 0.445 0.499 0.347 12.736 155.591
0.020 0.369 4.468 0.191 0.307 0.672 0.474 0.021 0.034 0.165 0.213 0.306 0.307 4.683 23.042 0.234 0.130 0.589 3.589 2.549 0.177 0.647 0.123 85.945 314.381 N = 38,000
0.141 0.483 0.793 0.393 0.461 1.123 0.895 0.143 0.181 0.372 0.409 0.461 0.461 1.179 3.271 0.424 0.337 0.492 1.663 0.252 0.382 0.478 0.329 13.918 149.813
Due to space considerations, the estimates for the time of day and time of year (month) fixed effects parameters are omitted from the tables.
model, where the odds ratio is defined as the probability that the event occurs (such as severe injury) divided by the probability that the event does not occur. For example, the estimate displayed in the first column of the first row of Table 7 indicates that for an at-fault drinking driver (relative to an at-fault non-drinking driver), the odds ratio of a severe injury resulting from the crash increases by a factor of 3.49, on average. The proper interpretation of this estimate is that the odds for a crash resulting in severe injury or fatality when the at-fault driver was drinking is 3.49 times higher, on average, than for crashes where the at-fault driver was not drinking. Similarly, for multi-unit crashes, the odds for a crash resulting in a severe injury is 2.70 times higher, on average, if the at-fault driver had been drinking.18 18
For the logit models, the marginal effect is the expected change in the odds-ratio from a one unit change in the alcohol use status of the at-fault driver (which is a binary 0–1 variable in this study), on average, holding the other independent variables constant. This is not contingent upon drinking being a continuous variable. For the specific formula, see for example, Maddala (1983, p. 23) or SAS/STAT UserÕs Manual, Version 8 (1999, pp. 1952–1955).
T.L. Traynor / Transportation Research Part E 41 (2005) 421–437
429
Table 2 Logit estimates of the probability that a crash results in a serious injury Variable
Intercept drinking dark precipit slip rural ur inc young old male type ED highspd intersec highway year typeyear
All crashes—impact of any drinking by the at-fault driver
Multi-unit crashes—impact of any drinking by the at-fault driver
Parameter estimate
Parameter estimate
Standard error
0.534 0.810 0.086 1.251a 0.042 0.139 0.166 0.124 0.250a 0.110 0.375a 0.079 0.013 0.027 0.004 0.016 0.041 0.074 0.096 0.107 0.245a 0.068 0.841a 0.201 0.397a 0.195 0.665a 0.081 0.152a 0.068 0.215a 0.090 0.032a 0.007 8.2 · 103a 2.3 · 103 2 Log likelihood: 10166.988 Likelihood ratio: v2 = 964.8356 N = 48,828
Standard error
0.207 1.062 0.989a 0.141 0.003 0.179 0.064 0.167 0.118 0.150 0.572a 0.098 0.038 0.039 0.020 0.021 0.035 0.100 0.100 0.125 0.180a 0.086 0.795a 0.258 0.468a 0.262 1.004a 0.103 0.065 0.084 0.353a 0.119 0.032a 0.009 0.008a 0.003 2 Log likelihood: 62873.207 Likelihood ratio: v2 = 1200.2144 N = 38,000
Dependent variable—SERIOUS. Due to space considerations, the estimates for the time of day and time of year (month) fixed effects parameters are omitted from the tables. a Statistical significance at the 95% level of confidence for a one-tailed test.
Also listed in the first row of Table 7 are the estimates for the impact of legal impairment (BAC greater than or equal to 0.10) on the part of the at-fault driver on the odds ratio for severe injury/fatal crashes, which are 2.94 and 2.33 for all crashes and for multi-unit crashes respectively. These estimates are also statistically significant at the 99% level of confidence. The second row of Table 7 reports the estimates for the impact of at-fault driver alcohol use on the odds ratios for the likelihood that a crash will result in an injury (including minor injury) or death. In these cases, the marginal impacts are smaller but are also statistically significant at the 99% level. Table 3 presents parameter estimates for equations in which the scaled variable SEVERITY is the dependent variable. In both cases (all crashes and multi-unit crashes), at-fault driver alcohol use statistically significantly increases the level of crash severity at the 99% level of confidence. The resulting odds ratio estimate for all crashes is 2.45 for detected alcohol use by the at-fault driver. In the case of a scaled variable like SEVERITY, the odds ratio for an ordered logit model represents the cumulative probability that a crash will suffer a specific severity level divided by the probability that it does not. For multi-unit crashes, the odds ratio estimate for SEVERITY is 2.48 for drinking (relative to non-drinking) drivers. For crashes caused by legally impaired
430
T.L. Traynor / Transportation Research Part E 41 (2005) 421–437
Table 3 Ordered logit estimates of the level of severity of a crash Variable
Intcpt 1 Intcpt 2 Intcpt 3 Intcpt 4 drinking dark precipit slip rural ur inc young old male type ED highspd intersec highway year typeyear
All crashes—impact of any drinking by the at-fault driver
Multi-unit crashes—impact of any drinking by the at-fault driver
Parameter estimate
Parameter estimate
a
Standard error
0.320 4.791 2.976a 0.313 0.912a 0.312 0.252 0.312 0.894a 0.040 0.065 0.043 0.053 0.035 0.123a 0.033 0.096a 0.023 0.029a 0.010 0.013a 0.005 0.028 0.023 0.052a 0.030 0.068a 0.020 0.246a 0.064 0.147a 0.063 0.397a 0.027 0.143a 0.020 0.110a 0.029 0.011a 0.003 1.8 · 103a 7.2 · 104 2 Log likelihood: 94221.592 Likelihood ratio: v2 = 1849.0111 N = 48,828
Standard error
a
5.420 0.371 3.672a 0.362 1.557a 0.360 0.185 0.360 0.700a 0.057 0.105a 0.049 0.028 0.042 0.031 0.039 0.107a 0.026 0.028a 0.012 0.009a 0.005 0.005 0.027 0.071a 0.032 0.077a 0.022 0.159a 0.069 0.130a 0.075 0.460a 0.033 0.259a 0.023 0.023 0.036 0.007a 0.003 0.001 7.8 · 104 2 Log likelihood: 70861.16 Likelihood ratio: v2 = 981.9934 N = 38,000
Dependent variable—SEVERITY. Due to space considerations, the estimates for the time of day and time of year (month) fixed effects parameters are omitted from the tables. a Statistical significance at the 95% level of confidence for a one-tailed test.
drivers, the odds ratio estimate is 2.0 when the at-fault driver had been drinking and 2.2 when the at-fault driver was legally intoxicated.19 5.2. The impact of at-fault driver alcohol use on the relative severity of crashes to at-fault versus not-at-fault parties The estimates presented in Table 4 indicate that, for all crashes, at-fault driver alcohol consumption does not significantly alter the likelihood that a crash will result in a not-at-fault party suffering a greater degree of injury than the at-fault driver. Although not reported in Table 4, this finding remains the same for cases of legal impairment. In both cases, the odds ratio estimate is negligible. However, for multi-unit crashes, at-fault driver alcohol use brings about a statistically 19
For details, see Maddala (1983, p. 46) or SAS/STAT UserÕs Manual, Version 8 (1999, pp. 1952–1955).
T.L. Traynor / Transportation Research Part E 41 (2005) 421–437
431
Table 4 Logit estimates of the probability that a crash results in a greater level of injury to a not-at-fault party than to the atfault party Variable
Intercept drinking dark precipit slip rural ur inc young old male type ED highspd intersec highway year typeyear
All crashes—impact of any drinking by the at-fault driver
Multi-unit crashes—impact of any drinking by the at-fault driver
Estimate
Estimate
Standard error
2.841a 0.429 0.019 0.063 0.059 0.058 0.035 0.049 0.088a 0.045 0.104a 0.031 0.003 0.014 0.010a 0.006 0.049a 0.031 0.089a 0.040 0.146a 0.027 0.131a 0.077 0.145a 0.085 0.075a 0.038 0.523a 0.028 0.167a 0.040 0.006 0.004 0.001 8.7 · 104 2 Log Likelihood: 42002.060 Likelihood ratio: v2 = 952.6441 N = 48,828
Standard error
2.302a 0.454 0.174a 0.073 0.110a 0.061 0.050 0.052 0.009 0.048 0.029 0.032 0.001 0.015 0.012a 0.006 0.038 0.033 0.117a 0.041 0.164a 0.028 0.105 0.080 0.156a 0.091 0.182a 0.041 0.241a 0.029 0.027 0.044 0.005 0.004 8.3 · 104 9.1 · 104 2 Log Likelihood: 37079.498 Likelihood ratio: v2 = 316.0174 N = 38,000
Dependent variable—EXTERNAL. Due to space considerations, the estimates for the time of day and time of year (month) fixed effects parameters are omitted from the tables. a Statistical significance at the 95% level of confidence for a one-tailed test.
significant increase in the likelihood that another person suffers a greater degree of injury than the at-fault driver. This is also true for cases where the at-fault driver is legally impaired. The magnitude of the odds ratio estimate is 1.19 in the case of detected alcohol use, and 1.26 in the case of legal impairment. Taken together, the results indicate that while drinking and driving significantly increases the proportion of multi-unit crashes that result in greater injury levels to other parties, there is likely an equivalent increase in the number of injury-producing single car crashes that lead to a negligible aggregate effect.20 The fifth row of Table 7 presents the odds ratio estimates for the impact of at-fault driver alcohol use on the likelihood that a crash will result in no injury to the at-fault driver while causing 20
The parameter estimates for the control variables generally have the signs that would be expected. Noteworthy among the parameter estimates for the control variables are that road conditions are largely not significant determinates of EXTERNAL, and that while young drivers are associated with greater levels of crash injuries among not-at-fault drivers, the opposite is true for elderly drivers. Interestingly, the county based educational attainment variable proved to have a positive correlation with not-at-fault injury likelihood. Also of note is the sign for the parameter estimate for HIGHSPD which is significantly negative for all crashes and significantly positive for multi-unit crashes.
432
T.L. Traynor / Transportation Research Part E 41 (2005) 421–437
Table 5 Tobit estimates of the number of injuries/fatalities suffered in a crash Variable
Intercept drinking dark precipit slip rural ur inc young old male type ED highspd intersec highway year typeyear Scale
All crashes—impact of any drinking by the at-fault driver
Multi-unit crashes—impact of any drinking by the at-fault driver
Parameter estimate
Standard error
Parameter estimate
Standard error
0.609 0.711a 0.070 0.039 0.115a 0.064a 0.059a 0.029a 0.093a 0.056 0.174a 0.202a 0.213a 0.346a 0.371a 0.009 0.009a 1.2 · 103 2.315 N = 48,828
0.417 0.055 0.057 0.047 0.044 0.031 0.013 0.006 0.031 0.040 0.026 0.082 0.084 0.036 0.027 0.039 0.004 9 · 104 0.014
0.801 0.748a 0.129a 0.018 0.014 0.122a 0.073a 0.025a 0.035 0.062 0.168a 0.198a 0.222a 0.502a 0.413a 0.043 0.009a 1.3 · 103 2.492 N = 38,000
0.519 0.084 0.071 0.060 0.055 0.037 0.018 0.008 0.038 0.046 0.032 0.098 0.108 0.048 0.034 0.051 0.005 1.1 · 103 0.018
Dependent variable—NUMINJ. Due to space considerations, the estimates for the time of day and time of year (month) fixed effects parameters are omitted from the tables. a Statistical significance at the 95% level of confidence for a one-tailed test.
injury or death to others. This finding remains substantively unchanged regardless of the specific measure of alcohol involvement or the specific data set (multi-unit or not) used. With respect to all crashes, the odds ratio estimate for detected alcohol use by the at-fault driver is 1.19, while the odds ratio estimate for legal intoxication by the at-fault driver is 1.15. For multi-unit crashes, the odds ratio estimates are 1.47 when the at-fault driver had been drinking and 1.37 when the at-fault driver was legally impaired.21 In all four of these cases, the estimates are significantly different from zero at the 99% level of confidence. 5.3. The impact of alcohol use by at-fault drivers on the numbers of injured or killed parties The Tobit estimates presented in Table 5 indicate that, on average, at-fault driver alcohol use significantly increases the number of injured and killed parties per crash. The estimated magnitudes of this relationship are .71 additional injured/killed parties per crash when any alcohol 21
Regarding the control variables for the estimates presented in Table 6, note that the economic factors, income and unemployment, consistently have a statistically positive impact on EXTINJ, while vehicle characteristics are not important predictors of the dependent variable.
T.L. Traynor / Transportation Research Part E 41 (2005) 421–437
433
Table 6 Tobit estimates of the number of injuries/fatalities suffered by not-at-fault parties in a crash Variable
Intercept drinking dark precipit slip rural ur inc young old male type ED highspd intersec highway year typeyear Scale
All crashes—impact of any drinking by the at-fault driver
Multi-unit crashes—impact of any drinking by impact of any drinking by the at-fault driver
Parameter estimate
Parameter estimate
a
2.592 0.233a 0.036 0.009 0.133a 0.057a 0.058a 0.025a 0.168a 0.022 0.059a 0.016 0.107 0.054 0.731a 0.211a 1 · 104 6 · 104 2.263 N = 48,828
Standard error 0.444 0.062 0.061 0.050 0.047 0.033 0.014 0.006 0.032 0.042 0.028 0.084 0.090 0.039 0.029 0.042 0.004 0.001 0.017
a
1.840 0.483a 0.091 0.002 0.017 0.102a 0.066a 0.014a 0.071a 0.010 0.036 0.019 0.114 0.348a 0.425a 0.081a 0.001 1 · 104 2.194 N = 38,000
Standard error 0.476 0.077 0.065 0.055 0.050 0.034 0.016 0.007 0.035 0.042 0.029 0.088 0.098 0.043 0.031 0.047 0.005 0.001 0.017
Dependent variable—NUMEX. Due to space considerations, the estimates for the time of day and time of year (month) fixed effects parameters are omitted from the tables. a Statistical significance at the 95% level of confidence for a one-tailed test.
consumption is detected for the at-fault driver. When only multi-unit crashes are considered, the estimated coefficients for the alcohol variables are again positive and statistically significant, with point estimates of .74 additional injured/killed parties. Although not reported in Table 5, when the at-fault driver is legally impaired, the estimates are .69 and .81 additional injured/killed parties per crash.22 Again, the estimates are statistically significant. Table 6 summarizes the Tobit estimates for which the number of parties other than the at-fault driver who are injured or killed in a crash was the dependent variable. For all crashes, driver alcohol consumption causes an additional .23 injuries/fatalities to other parties per crash, and legally impaired driving causes an additional .2 injuries/fatalities to other parties per crash. When multiunit crashes are isolated, the estimates are .48 additional injuries/fatalities per crash when the atfault driver was drinking and an additional .57 injuries/fatalities per crash when the at-fault driver was legally drunk.
22 For the Tobit model, the marginal effect represents the marginal impact of at-fault driver alcohol use (or impairment) on the number of injuries/fatalities in a crash. For details, see Maddala (1983, p. 151).
434
T.L. Traynor / Transportation Research Part E 41 (2005) 421–437
Table 7 The marginal effects of at-fault driver alcohol consumption Dependent variable
Probability that a crash results in severe (incapacitating) injury or fatality Probability that a crash results in any injury/fatality Level of severity of worst injury suffered per crash (five point scale) Probability that a crash causes a more severe injury to a not-at-fault party than to the at-fault party Probability that a crash causes an injury or fatality to a not-at-fault party and no injury to the at-fault party Number of injuries/fatalities per crash Number of injuries/fatalities to not-at-fault parties per crash
All crashes
Multi-unit crashes
Marginal impact of any detected alcohol consumption
Marginal impact of legal impairment
Marginal impact of any detected alcohol consumption
Marginal impact of legal impairment
3.49
2.94
2.70
2.33
1.96
1.96
1.75
1.75
2.45
2.48
2.00
2.20
1.01
1.01
1.19
1.26
1.19
1.15
1.47
1.37
0.71 0.23
0.69 0.20
0.74 0.48
0.81 0.57
For the first five rows (which are based on logit and ordered logit estimates), the estimates are odds ratio estimates. These are measures of the average proportional change in the odds ratio attributable to at-fault driver alcohol use after controlling for the other variables in the model, where the odds ratio is defined as the probability that the event occurs (such as a crash resulting in a severe injury or fatality in the first row) divided by the probability that the event does not occur. These estimates are derived by calculating the exponential of the parameter estimates for alcohol use. For the last two rows (which are from the Tobit estimates), the parameter estimates directly measure the marginal impact of at-fault driver alcohol use.
6. Summary and discussion This paper extends the analysis of the impact of driver alcohol use by developing empirical tests of its impact on average crash severity using crash specific datum. In aggregate, the results of the many tests discussed above signify that there is indeed a positive relationship between at-fault driver alcohol use and the violence of the average crash. The estimates presented in this paper lead to the conclusion that, controlling for other influences, crashes in which the at-fault driver had been drinking are relatively more injurious. This is true both as measured by the level of the worst injury suffered by the involved parties as well as by the numbers of injuries and fatalities suffered per crash. The estimates remain statistically significant as the measure of alcohol involvement broadens from legal impairment to any level of alcohol use. This result carries important implications for policy makers wishing to reduce the social costs of driver alcohol use. It appears that drinking and driving not only increases the numbers of motor vehicle crashes that occur, it also increases the injury and fatality losses of a typical crash. Thus, policies should account for this relatively higher crash severity when laws are developed to reduce the negative impact of drinking and driving. Certainly, policies that reduce the frequency of alcohol related crashes will reduce the costs
T.L. Traynor / Transportation Research Part E 41 (2005) 421–437
435
described by this study. Policies which focus on reductions in BACs (as noted in the literature survey) should also be helpful. In addition, with the exception of the relative degree of injury to at-fault drivers versus other victims over all crashes, drinking and driving significantly increases the severity of crashes to other involved parties. This is true both as measured by the impact of alcohol on the degree of injury severity suffered by not-at-fault victims and the number of not-at-fault injuries and fatalities suffered per crash. In addition, alcohol involved crashes are significantly more likely than other crashes to injure innocent parties while leaving the at-fault driver unharmed. The results presented here are by no means a comprehensive analysis of the impact of drinking and driving. These results are only estimates of the impact of alcohol involvement on crashes, and do not attempt to measure the total impact of driver alcohol use on aggregate road safety. Future analyses with alternative data sets will undoubtedly provide more evidence that can help develop appropriate policies regarding drinking and driving. While it is generally assumed that crashes create externalities and that externalities rise as their number and severity rise, precise measures of these externalities are lacking.23 Instead, researchers have tended to use the gross costs associated with property damage, injuries, and fatalities without including measures of compensation provided by at-fault drivers (or their insurers). Future research should account for the impact that driver alcohol use has upon both the frequency and severity of crashes. Additionally, estimates of all externalities created by drunk driving should be pursued in future research efforts by accounting for both the incidence of drunk driving and all the forms of external costs that are created. Finally, future research efforts can assess enforcement effectiveness with crash specific datum by developing interaction terms that measure the relative impact of various enforcement tools on the severity of crashes. With regard to the measurement of the externalities resulting from driver alcohol use, a major difficulty is the estimation of the proportion of drivers on the road at any moment that has been drinking. Levitt and Porter (2001) discovered that the US Fatality Analysis Reporting System provides data which are rich enough, under certain assumptions, to identify the proportion of drivers on the road who have been drinking or who are legally drunk. They estimated that the externality created by drinking and driving in terms of increased fatalities is at least 30 cents per mile. In this study, all costs of alcohol related crash fatalities are included as externalities, which probably does not reflect the true externalities of these fatalities if some of the associated costs are internalized to the drunk driver. In fact, the task of estimating the magnitudes of the various types of losses associated with driver alcohol use has turned out to be a difficult task. Empirical studies have been able to focus only on portions of the issue, such as the fatality specific externalities resulting from drunk driving or the impact of driver alcohol use on aggregate fatalities and aggregate fatal crashes. The losses associated with motor vehicle crashes are ultimately underwritten by the at-fault driver, other involved parties (e.g. other drivers, passengers, pedestrians, motorcyclists, owners of damaged property), and society in general. These include the expected losses from the future stream of productivity net of consumption due to deaths or injuries, the medical costs accrued to the 23
As discussed in a more general analysis of all the external costs of driving by Jones-Lee (1990), when a driver behaves less safely behind the wheel (as is the case when under the influence of alcohol), ceteris paribus, the expected external costs rise as both the driver and other road users face higher expected losses due to crashes.
436
T.L. Traynor / Transportation Research Part E 41 (2005) 421–437
victims, as well as any other costs to society in general from the deaths, injuries, and property losses resulting from crashes. Of these costs, those ultimately underwritten by the at-fault driver through compensatory civil settlements, higher insurance premiums, or their own lost livelihood are internalized.24 The remaining costs are externalities.25 The externalities caused by alcohol related crashes may be underwritten in several different ways. Victims who are not at fault in crashes may be left to underwrite some of their losses themselves through increased insurance premiums, insurance deductibles and co-payments, uncompensated income losses, or through the non-pecuniary losses that may be incurred. The public may also subsidize the losses suffered in crashes through public medical coverage of underinsured victims, the lost net productivity of the victims, or through higher insurance premiums. Clearly, the measurement and evaluation of the externalities of drinking and driving is a very difficult task that is in need of further research.
References Ferrara, S., Zancaner, S., Giorgetti, R., 1994. Low blood alcohol concentrations and driving impairment. International Journal of Legal Medicine 106, 169–177. Frei, P., Kaeser, R., Muser, M., Niederer, P., Waltz, F., 1999. Vehicle structural crashworthiness with respect to compatibility in collisions. Working Group on Accident Mechanics at the Universities of Zurich, Switzerland. Gujarati, D., 2003. Basic Econometrics, fourth ed. McGraw-Hill, New York. Haffner, H., Graw, E., 1996. Changes in the spectrum of alcohol-induced traffic accidents in relation to blood alcohol level. Blutalkohol 33, 78–83. Jones-Lee, M., 1990. The value of transport safety. Oxford Review of Economic Policy 6, 39–60. Kahane, C., 2003. Vehicle weight, fatality risk and crash compatibility of model year 1991–99 passenger cars and light trucks. NHTSA Technical Report No. DOT HS 809 662, National Highway Traffic Safety Administration. Lee, L., 1997. The socioeconomics of drunk driving. Journal of Socio-Economics 26, 95–106. Lemaire, J., 1985. Automobile Insurance: Actuarial Models. Kluwer-Nijhoff Publishing, Hingham, MA. Levitt, S., Porter, J., 2001. How dangerous are drinking drivers? Journal of Political Economy 109, 1198–1237. Maddala, G., 1983. Limited-Dependent and Qualitative Variables in Econometrics. Cambridge University Press, Cambridge. McCarthy, P., 2003. Alcohol-related crashes and alcohol availability in grass-roots communities. Applied Economics 35, 1331–1338. Miller, T., Blewden, M., 2001. Costs of alcohol-related crashes: New Zealand estimates and suggested measures for use internationally. Accident Analysis and Prevention 33, 783–791. Muelleman, R., Mueller, K., 1996. Fatal motor vehicle crashes: variations of crash characteristics within rural regions of different population densities. Journal of Trauma 41, 315–320. National Highway Traffic Safety Administration, Initiatives to Address Impaired Driving, December 2003. National Highway Traffic Safety Administration, Impaired Driving in the United States: State Alcohol Cost Fact Sheets, National Highway Traffic Safety Administration, 2002. National Highway Traffic Safety Administration, Impaired Driving in the United States: State Cost Fact Sheet UserÕs Guide, 2002.
24
Aside from the reduction of externalities per accident, criminal penalties for driving under the influence (DUI) as well as penal damages awarded under civil law for alcohol related crashes are designed to reduce externalities from alcohol related crashes by discouraging drinking and driving. 25 For a more complete discussion of actuarial methods used by the automobile insurance industry, see Lemaire (1985, pp. 39–56).
T.L. Traynor / Transportation Research Part E 41 (2005) 421–437
437
Orsay, E., Doan-Wiggins, L., Lewis, R., Lucke, R., RamaKrishnan, V., 1994. The impaired driver: hospital and police detection of alcohol and other drugs of abuse in motor vehicle crashes. Annals of Emergency Medicine 24, 51–55. Owens, D., Sivak, M., 1996. Differentiation of visibility and alcohol as contributors to twilight road fatalities. Human Factors 38, 680–689. Peek-Asa, C., Kraus, J., 1996. Alcohol use, driver, and crash characteristics among injured motorcycle drivers. Journal of Trauma, Injury, Infection, and Critical Care 41, 989–993. Peltzman, S., 1975. The effects of automobile safety regulation. Journal of Political Economy 83, 677–725. Perper, J., Kuller, L., Shim, Y., 1993. Vehicular crashes and alcohol involvement in drivers at fault, and related fatalities. American Journal of Forensic Medicine and Pathology 14, 177–184. Robertson, M., Drummer, O., 1994. Responsibility analysis: a methodology to study the effects of drugs in driving. Accident Analysis and Prevention 26, 243–247. Ruhm, C., 1996. Alcohol policies and highway vehicle fatalities. Journal of Health Economics 15, 435–454. Solnick, S., Hemenway, D., 1994. Hit the bottle and run: the role of alcohol in hit-and-run pedestrian fatalities. Journal of the Study of Alcohol 55, 679–684. Traynor, T., 1993. The Peltzman hypothesis revisited: an isolated evaluation of offsetting driver behavior. Journal of Risk and Uncertainty 7, 237–247. Traynor, T., McCarthy, P., 1993. Economic regulation and highway safety in the trucking industry: a limited dependent variable analysis. Quarterly Review of Economics and Finance 33, 141–153. Waller, P., Blow, F., Maio, R., Singer, K., Hill, E., Schaefer, N., 1997. Crash characteristics and injuries of victims impaired by alcohol versus illicit drugs. Accident Analysis and Prevention 29, 817–827. Zador, P., 1991. Alcohol-related relative risk of fatal driver injuries in relation to driver age and sex. Journal of the Study of Alcohol 52, 302–310.
Further reading Blomquist, G., 1988. The Regulation of Motor Vehicle and Traffic Safety. Kluwer Academic Publishers, Boston. Greene, W., 2003. Econometric Analysis. Macmillan Publishing Co., New York. Lloyd, C., 1990. Estimating the effect of alcohol on the risk of a fatal road accident. Journal of the Royal Statistical Society 153 (Part 1), 29–52. Robertson, L., 1993. Blood alcohol in fatally injured drivers and the minimum legal drinking age. Journal of Health, Politics, Policy, and Law 14, 817–825.
The impact of driver alcohol use on crash severity: A crash specific analysis Thomas L. Traynor
*
Department of Economics, Wright State University, Dayton, OH 45435, United States
Abstract This study uses a crash specific data set that is supplemented with location based socioeconomic data to estimate the impact of driver alcohol use on average crash severity. Logit estimates indicate that crashes in which the at-fault drivers had been drinking are more likely to result in a severe injury or death than are crashes caused by sober drivers. Ordered logit estimates indicate that at-fault driver alcohol use increases the expected highest degree of injury resulting from a crash, and Tobit estimates indicate that the number of injuries or deaths per crash increase an average of 0.71 when the at-fault driver has been drinking. Moreover, at-fault driver alcohol use worsens the severity of crashes relative to not-at-fault parties. Collectively, these results indicate that at-fault drinking drivers are involved in more violent crashes and produce more serious injuries to not-at-fault parties than at-fault sober drivers. 2005 Elsevier Ltd. All rights reserved. Keywords: Crash; Alcohol; Impaired driving; Crash severity
1. Introduction There were 42,815 traffic fatalities in 2002, 40% of which were alcohol related. The National Highway Traffic Safety Administration estimated that more than 2,163,000 alcohol involved *
Tel.: +1 937 775 2025/3070. E-mail address: [email protected]
1366-5545/$ - see front matter 2005 Elsevier Ltd. All rights reserved. doi:10.1016/j.tre.2005.03.005
422
T.L. Traynor / Transportation Research Part E 41 (2005) 421–437
crashes occurred in 2002, killing 17,419 people and injuring an estimated 513,000.1 Of the estimated 0.62 alcohol related fatalities that occurred per 100 million vehicle miles traveled (VMT) in 2002, most (0.53 per 100 million VMT) involved high blood alcohol concentrations of 0.08 or above.2 Given this information, it is not surprising that numerous studies have been conducted to identify factors that influence alcohol related fatalities and injuries. This article provides an evaluation of individual crashes to determine the impact of driver alcohol use on crash severity, conveying a unique vantage point to the impact of driver alcohol use. In so doing, this analysis contributes to the existing body of research by exploring whether the fatality and injury levels associated with drunk driving are not only caused by increased numbers of crashes, but by more violent crashes as well. The resulting estimates lend support for the hypothesis that alcohol consumption by at-fault drivers increases the severity of crashes. This is true of both the severity of injury suffered by victims as well as the numbers of injuries and fatalities suffered per crash. These results indicate that drinking and driving policies must not only be evaluated by the reduction in the number of crashes they bring about, but also by the reduction in crash severity they bring about. The remainder of the paper proceeds with a discussion of issues and existing evidence surrounding alcohol consumption and crash severity in Section 2. Sections 3 and 4, respectively, develop the model that is estimated, and describes the data that are used in the analysis. Section 5 reports and evaluates the results, while concluding comments are made in the final section. 2. The impact of driver alcohol use Numerous researchers have studied the incidence of driver alcohol use, identified factors that influence the incidence of alcohol related crashes, and estimated the cost to society of driver alcohol use. Studies of the incidence of alcohol use in crashes primarily focus on the hospital records of crash victims. Perper et al. (1993) evaluated coroner and emergency room records, concluding that passengers face the same risk of fatality as their alcohol-positive drivers. Waller et al. (1997) examined blood samples from patients treated for automobile crash injuries, finding that alcohol was associated with more severe crashes than those with drug involvement or those with no drug/ alcohol involvement. Zador (1991) examined BAC concentrations reported to the Fatality Analysis Reporting System to conclude that each 0.02% increase in BAC doubles the risk of involvement in a fatal single car crash, and that the likelihood of a crash was at least nine times greater at BACs in the 0.05–0.09% range than at 0% BAC for all age groups. Similarly, Robertson and Drummer (1994) evaluated 341 driver fatalities to determine that the relative risk of having a fatal crash increased with increased BAC. Ferrara et al. (1994) reviewed 38 studies of the impact of low alcohol concentrations (below 0.1%) on driving, and found no consensus on whether or not low BAC levels significantly contribute to crash losses. Several investigations have been conducted on the impact of alcohol on the behavior of drivers involved in crashes. Haffner and GrawÕs (1996) examination of 625 crashes caused by alcoholimpaired drivers revealed that speed was the most common cause of crashes involving an alco1 2
Initiatives to Address Impaired Driving, National Highway Traffic Safety Administration, December 2003, pp. 4–5. Initiatives to Address Impaired Driving, National Highway Traffic Safety Administration, December 2003, pp. 7–9.
T.L. Traynor / Transportation Research Part E 41 (2005) 421–437
423
hol-impaired driver at all BAC segments. Peek-Asa and Kraus (1996) found that alcohol-impaired motorcycle drivers were more likely to be speeding and less likely to be wearing a helmet in the event of a crash, while Solnick and Hemenway (1994) discovered that drunk drivers are more likely to flee after hitting a pedestrian. Scholars have also explored the impact of environmental or demographic factors on the risk and severity of alcohol-related crashes. Owens and Sivak (1996) concluded that night time driving did not have a significant impact on alcohol-related traffic deaths. Orsay et al. (1994) used a sample of patients admitted to a Level I trauma center and found that drivers impaired by alcohol and/or drugs were younger, and were more often male than the unimpaired drivers. Muelleman and Mueller (1996) concluded that low population density was correlated with more frequent alcohol use and high intoxication levels among driver fatalities. Ruhm (1996) investigated the relationship between economic growth and drunk driving, finding that drunk driving fatality rates rose as national income rose. McCarthy (2003) empirically evaluated the impact of alcohol availability using the geographic density of alcohol licenses, finding that general alcohol off-site licenses are beneficial while general alcohol on-site licenses are detrimental to highway safety. Estimates of the impact of beer/wine license density did not produce uniform results. In a theoretical examination of the alcohol availability issue, Lee (1997) argued that market factors that determine the geographical distribution of alcohol establishments would influence the incidence of drunk driving. The National Highway Traffic Safety Administration (NHTSA) recently published a general review of alcohol related crash fatalities, noting that the annual total of alcohol related crash fatalities leveled off in the 1990s after declining during the previous decade.3 Miller and Blewden (2001) presented the most recent refinement of a method for estimating the costs of crashes, including costs associated with driver alcohol use. The NHTSAÕs latest update of its Alcohol Cost Fact Sheets (2002) made use of this method to estimate that alcohol related crashes had a total public cost of $114.3 billion in 2000, that the average cost of alcohol-related crash fatalities was $3.5 million, and that the average cost of alcohol-related crash injuries was $99,000. These costs translate to $5.80 per mile driven at a BAC of 0.10 or greater, and $1.00 per drink consumed.4 Traynor (1993) evaluated individual crashes to determine the impact of various factors on the proportion of crash injuries suffered by not-at-fault parties. The results were mixed. While Tobit estimates indicated that at-fault driver alcohol use significantly increases the proportion of crash injuries and fatalities incurred by other parties, ordered logit estimates indicated that the degree of injury suffered by drinking drivers relative to others (no injury–minor injury–severe injury–fatality) is not significantly impacted by at-fault driver alcohol use. The analysis that follows extends this research by estimating the impact of alcohol involvement on a wider array of crash severity measures (the probability that injuries are suffered, total injuries and fatalities, as well as injury/ fatality losses to not-at-fault parties), measures the marginal effects of alcohol involvement on
3
Initiatives to Address Impaired Driving, National Highway Traffic Safety Administration, December 2003, p. 7. Impaired Driving in the United States: State Alcohol Cost Fact Sheets, National Highway Traffic Safety Administration, 2002, p. 1. These costs were measured using data from a variety of sources. A description of the methods used to calculate these estimates is included in Impaired Driving in the United States: State Cost Fact Sheet UserÕs Guide. National Highway Traffic Safety Administration, 2002, pp. 2–5. 4
424
T.L. Traynor / Transportation Research Part E 41 (2005) 421–437
crash severity, and assesses the sensitivity of the relationship with the use of multiple measures of crash costs and alcohol use. The analysis further contributes to the existing literature by combining crash specific and socioeconomic data to evaluate the differential impact of at-fault driver alcohol use on the typical losses suffered per crash. This differs from prior analyses by focusing specifically on the alcohol-crash cost relationship on a per crash basis, while controlling for a variety of socio-economic factors. The resulting estimates provide evidence that alcohol involved crashes are more violent than other crashes.
3. Statistical models To estimate the impact of driver alcohol use on crash severity, an econometric model is developed in which individual crashes are evaluated to observe how varying levels of alcohol involvement impact the injury severity level or the number of injuries/fatalities suffered by involved parties. The hypotheses to be tested in this analysis are: • Crashes caused by drinking drivers result, on average, in more severe injuries and greater numbers of injured parties than crashes caused by sober drivers. • Crashes caused by drinking drivers result, on average, in more severe injuries and greater numbers of injured parties to other victims than crashes caused by sober drivers. The empirical model is a modified form of those that have been used to test the relationship between driving conditions and crash outcomes with similar data. Its general form is LOSS ¼ a þ / ALCOHOL þ bZ þ cF þ e
ð1Þ
where LOSS is a vector of observed sample values that measure crash severity (via injuries or fatalities), and ALCOHOL is a vector of observations that measure at-fault driver alcohol use. Z represents the matrix of control variables used in the analysis, and F is a matrix of fixed effects terms. Lastly, e is the stochastic error term. The parameter / measures the average impact of atfault driver alcohol use upon the severity of the crash, while b is a vector of parameters that estimate the impact of a control factor on crash severity, and c is a vector of parameters that estimate the impact of the fixed effects variables on severity. Since both of the key variables in this analysis (LOSS and ALCOHOL) are inherently difficult to precisely measure, this analysis employs multiple proxies for each variable, each of which is described in Section 4. Estimates are generated using each of the measures developed for LOSS and ALCOHOL. In addition, to determine whether the impact of alcohol on all crashes is not simply the result of changes in the severity of single vehicle crashes, the estimates are calculated both by using all crashes as well as by using only multiple unit (vehicles, cyclists, pedestrians) crashes.5
5
If the estimates using the two different sets of data are not substantively different, this would indicate that the impact of alcohol on crash severity is not driven primarily by single vehicle crashes. Single vehicle crashes do not need to be separately estimated to conduct this test.
T.L. Traynor / Transportation Research Part E 41 (2005) 421–437
425
The vector of control factors, Z, consists of a variety of measures of driving conditions at the time of each crash, demographic information about the at-fault driver, socioeconomic factors for the county of the crash, as well as fixed effect measures. A series of dummy variables measure the driving conditions and location factors at each crash that impact its severity. These variables control for lighting, road, and weather conditions, as well as for crashes that occur in intersections and/or rural locations. Additionally, demographic characteristics of the at-fault driver are included to control for the impact that driver sex and age have upon crash severity. Location factors that impact crash severity are also included. In an attempt to control for socioeconomic factors that may influence crash severity, economic and education data which are aggregated by the county of the crash site are included in the analysis. Variables measuring the mass and age of the at-fault vehicle are also included to control for the impact these factors have upon the severity of crashes to involved parties. Finally, hour and month fixed effects variables were included to improve the efficiency of the estimates.
4. The data The bulk of the data set consists of detailed police reports collected by the Ohio Integrated Traffic Reporting System (OITRS) for a single calendar year. Detailed information regarding the time, location, road type, road conditions, weather conditions, parties involved, crash severity, vehicle type, and driver characteristics at the time of the crash as determined by the reporting officer are collected by OITRS.6 These data are used in conjunction with county level socio-economic data to test the above hypotheses.7 4.1. Dependent variables As mentioned above, LOSS is measured in a variety of ways—seven in all. First, INJURY is a binary variable that is one if the crash resulted in any injury or fatality, and is zero otherwise. Second, SERIOUS is a binary variable that denotes whether or not the crash resulted in an incapacitating injury or fatality. Given the binary nature of INJURY and SERIOUS, estimates are obtained via logit analysis when either of these two dependent variables are used. Third, SEVERITY is an ordered variable which takes on a value from one, if the crash resulted in property damage only, to progressively higher values for crashes resulting in more severe injuries, up to a value of 5 for a fatal crash. SEVERITY is a more comprehensive than INJURY or SERIOUS.8 Also, 6
The OITRS database included a total of more than 330,000 observations, which consists of all crashes that occurred in the 1995 calendar year. Due to extremely lengthy processing times involved in calculating estimates from the complete OITRS database, a random sample of 15% of the crashes (48,828) was drawn from which estimates were calculated. This data set was further reduced to 38,000 observations through the removal of all single vehicle crashes so that the estimates based on multi-unit crashes could be calculated. 7 Although this data set clearly does not allow one to measure the behavior of drivers who do not suffer accidents, the hypothesis that is evaluated in this paper refers only to the effect of driver alcohol use upon relative accident costs. Therefore, only accidents must be observed to test this hypothesis. 8 The values given to an accidentÕs severity are: 1—property damage only, 2—complaint of pain, 3—visible injury, 4— severe injury, 5—fatal.
426
T.L. Traynor / Transportation Research Part E 41 (2005) 421–437
due to its ordinal nature, estimates using SEVERITY as the dependent variable are calculated via ordered logit.9 Two dependent variables were developed to focus on the severity of injury to not-at-fault parties. EXTERNAL is a binary variable that is equal to one if the highest degree of injury suffered by a not-at-fault party is greater than the degree of injury suffered by the at-fault driver, and is zero otherwise. Next, EXTINJ is a binary variable that is equal to one if at least one not-at-fault party suffered an injury while the at-fault driver escaped without injury (and is zero otherwise). Due to the binary nature of these variables, estimates using EXTERNAL or EXTINJ as the dependent variable are calculated using logit analysis. Finally, two dependent variables measure of the number of injuries and fatalities in each crash. NUMINJ is the number of injured and killed parties in a crash. In addition, the variable NUMEX focuses on crash severity to not-at-fault parties by measuring the number of not-at-fault individuals injured or killed in a crash. Since NUMINJ and NUMEX are quantitative variables that have distributions that are censored from below at zero, estimates are obtained for these equations using Tobit analysis.10 4.2. Explanatory variables The two measures of ALCOHOL that are used in the analysis are DRINKING, which is equal to one if the at-fault driver is reported to have been drinking at all (and zero otherwise), and IMPAIRED if the at-fault driverÕs blood alcohol content exceeds the legal limit for the State of Ohio. As mentioned earlier, estimates are obtained both for all crashes and for crashes involving two or more units.11 Thus, all permutations of the Proxies for LOSS and ALCOHOL together with the two data sets create a total of 24 sets of estimates that were calculated for this analysis.12 DARK is one if the crash occurred in darkness, and is zero otherwise. PRECIPIT is one if there was precipitation at the time of the crash (zero otherwise), while SLIP is one if the road was deemed by the reporting officer to have been slippery at the time of the crash (zero otherwise). If the crash took place in a rural area, RURAL carries a value of one and is zero otherwise. HIGHSPD denotes that the crash took place in a high speed zone (posted speed limit of fifty miles per hour or greater), while HIGHWAY denotes that the crash took place on multiple lane divided highway. If the crash took place at an intersection, INTERSEC carries a value of one and is zero otherwise. MALE is a dummy variable that is equal to one if the at-fault driver is male, and is included to account reflect driving behavior differences associated with the at-fault driverÕs gender. Two 9
Most basic econometrics texts provide an intuitive description of which statistical methods are most appropriate for various types of limited dependent variable models. See for example, Gujarati (2003, pp. 580–582). 10 Although Poisson regression is often used to calculate estimates in cases where the dependent variable consists of count data (such as the number of injuries per crash), since the distribution of injuries per crash corresponds to that of the right tail of the normal distribution, Tobit was selected as the estimation method. 11 A ÔunitÕ is any involved motor vehicle (including motorcycles), bicycle, or pedestrian. For a crash to be included in the OITRS data set, at least one of the units must be a motor vehicle. 12 Due to the large number of models estimated, only selected estimates are presented in detail. The remaining estimated models will be summarized in the text and in Table 7. All of the other results are available from the author upon request.
T.L. Traynor / Transportation Research Part E 41 (2005) 421–437
427
dummy variables were created to control for the impact of age on crash severity. YOUNG is one if the at-fault driver is under the age of 20, and OLD identifies crashes in which the at-fault driver is over 65 years of age. INC is the per capita income for the county in which the crash occurred.13 UR is equal to the unemployment rate in the county during the month in which the crash occurred. ED represents the average educational attainment for the county in which the crash happened. TYPE represents the expected losses faced by at-fault drivers based on the crash worthiness of their vehicles as measured by weight class. TYPE carries a value of one through thirteen with a value of one representing the lowest weight class and thirteen representing the highest weight class for the at-fault vehicle.14 YEAR is the year that the at-fault vehicle was manufactured. In addition, fixed effects variables were included in the model to improve the precision of the estimates. First, hour of day dummy variables were included to pick up time of day fixed effects. Second, monthly dummy variables were included to account for time of year fixed effects.15 A variety of interaction variables were also developed to try to improve the overall fit of the model, but only TYPEYEAR, which is the product of the TYPE and YEAR variables, proved to be statistically significant and was included in the model. Table 1 presents basic descriptive statistics for the variables used in this analysis (excluding the fixed effects variables).16
5. Estimates 5.1. The impact of at-fault driver alcohol use on crash severity Estimates of the equations that evaluate the impact of alcohol use on the severity of crashes are presented in Tables 2 and 3. The estimates presented in Table 2 indicate that detected alcohol consumption by the at-fault driver has a statistically significant positive impact on the probability of a crash resulting in serious injury or death at the 99% level of confidence. Additionally, estimates using only multiple unit crashes are similar in magnitude and significance to those of the all inclusive data set, indicating the results are not driven solely by an increase in single car crashes.17 The odds ratio estimates of at-fault driver alcohol use on all the dependent variables used in this study are presented in Table 7. These estimates are measures of the average proportional change in the odds ratio attributable to at-fault driver alcohol use after controlling for the other variables in the 13
Measures of income similar to INC were used in cross-sectional analyses by Peltzman (1975), as well as Traynor and McCarthy (1993). 14 For evidence regarding the relationship between vehicle weight and injury risk, see Kahane (2003) and Frei et al. (1999). 15 County fixed effects were initially included in the models, but were always insignificant, presumably due to the inclusion of the other control factors, and were therefore dropped from the analysis. 16 Due to space considerations, the estimates for the time of day and time of year (month) fixed effects parameter estimates are omitted from the tables. 17 Two of the parameter estimates for the control variables are worth noting. First, the parameter estimates for YOUNG indicate that young at-fault drivers increase the severity of all crashes on average while decreasing the severity of multi-unit crashes, suggesting that single car crashes caused by young drivers are more violent than multi-unit crashes caused by young drivers.
428
T.L. Traynor / Transportation Research Part E 41 (2005) 421–437
Table 1 Descriptive statistics Variable serious injury severity external extinj numinj numex impaired drinking dark precipit slip rural ur inc young old male type ED highspd intersec highway year typeyear
All crashes
Multi-unit crashes
Mean
Std. Dev.
Mean
Std. Dev.
0.026 0.373 4.432 0.158 0.263 0.635 0.402 0.039 0.059 0.227 0.238 0.343 0.396 4.785 22.605 0.245 0.114 0.609 3.598 2.518 0.271 0.534 0.140 86.229 315.035 N = 48,828
0.160 0.484 0.841 0.365 0.440 1.058 0.836 0.193 0.236 0.419 0.426 0.475 0.489 1.279 3.445 0.430 0.318 0.488 1.728 0.267 0.445 0.499 0.347 12.736 155.591
0.020 0.369 4.468 0.191 0.307 0.672 0.474 0.021 0.034 0.165 0.213 0.306 0.307 4.683 23.042 0.234 0.130 0.589 3.589 2.549 0.177 0.647 0.123 85.945 314.381 N = 38,000
0.141 0.483 0.793 0.393 0.461 1.123 0.895 0.143 0.181 0.372 0.409 0.461 0.461 1.179 3.271 0.424 0.337 0.492 1.663 0.252 0.382 0.478 0.329 13.918 149.813
Due to space considerations, the estimates for the time of day and time of year (month) fixed effects parameters are omitted from the tables.
model, where the odds ratio is defined as the probability that the event occurs (such as severe injury) divided by the probability that the event does not occur. For example, the estimate displayed in the first column of the first row of Table 7 indicates that for an at-fault drinking driver (relative to an at-fault non-drinking driver), the odds ratio of a severe injury resulting from the crash increases by a factor of 3.49, on average. The proper interpretation of this estimate is that the odds for a crash resulting in severe injury or fatality when the at-fault driver was drinking is 3.49 times higher, on average, than for crashes where the at-fault driver was not drinking. Similarly, for multi-unit crashes, the odds for a crash resulting in a severe injury is 2.70 times higher, on average, if the at-fault driver had been drinking.18 18
For the logit models, the marginal effect is the expected change in the odds-ratio from a one unit change in the alcohol use status of the at-fault driver (which is a binary 0–1 variable in this study), on average, holding the other independent variables constant. This is not contingent upon drinking being a continuous variable. For the specific formula, see for example, Maddala (1983, p. 23) or SAS/STAT UserÕs Manual, Version 8 (1999, pp. 1952–1955).
T.L. Traynor / Transportation Research Part E 41 (2005) 421–437
429
Table 2 Logit estimates of the probability that a crash results in a serious injury Variable
Intercept drinking dark precipit slip rural ur inc young old male type ED highspd intersec highway year typeyear
All crashes—impact of any drinking by the at-fault driver
Multi-unit crashes—impact of any drinking by the at-fault driver
Parameter estimate
Parameter estimate
Standard error
0.534 0.810 0.086 1.251a 0.042 0.139 0.166 0.124 0.250a 0.110 0.375a 0.079 0.013 0.027 0.004 0.016 0.041 0.074 0.096 0.107 0.245a 0.068 0.841a 0.201 0.397a 0.195 0.665a 0.081 0.152a 0.068 0.215a 0.090 0.032a 0.007 8.2 · 103a 2.3 · 103 2 Log likelihood: 10166.988 Likelihood ratio: v2 = 964.8356 N = 48,828
Standard error
0.207 1.062 0.989a 0.141 0.003 0.179 0.064 0.167 0.118 0.150 0.572a 0.098 0.038 0.039 0.020 0.021 0.035 0.100 0.100 0.125 0.180a 0.086 0.795a 0.258 0.468a 0.262 1.004a 0.103 0.065 0.084 0.353a 0.119 0.032a 0.009 0.008a 0.003 2 Log likelihood: 62873.207 Likelihood ratio: v2 = 1200.2144 N = 38,000
Dependent variable—SERIOUS. Due to space considerations, the estimates for the time of day and time of year (month) fixed effects parameters are omitted from the tables. a Statistical significance at the 95% level of confidence for a one-tailed test.
Also listed in the first row of Table 7 are the estimates for the impact of legal impairment (BAC greater than or equal to 0.10) on the part of the at-fault driver on the odds ratio for severe injury/fatal crashes, which are 2.94 and 2.33 for all crashes and for multi-unit crashes respectively. These estimates are also statistically significant at the 99% level of confidence. The second row of Table 7 reports the estimates for the impact of at-fault driver alcohol use on the odds ratios for the likelihood that a crash will result in an injury (including minor injury) or death. In these cases, the marginal impacts are smaller but are also statistically significant at the 99% level. Table 3 presents parameter estimates for equations in which the scaled variable SEVERITY is the dependent variable. In both cases (all crashes and multi-unit crashes), at-fault driver alcohol use statistically significantly increases the level of crash severity at the 99% level of confidence. The resulting odds ratio estimate for all crashes is 2.45 for detected alcohol use by the at-fault driver. In the case of a scaled variable like SEVERITY, the odds ratio for an ordered logit model represents the cumulative probability that a crash will suffer a specific severity level divided by the probability that it does not. For multi-unit crashes, the odds ratio estimate for SEVERITY is 2.48 for drinking (relative to non-drinking) drivers. For crashes caused by legally impaired
430
T.L. Traynor / Transportation Research Part E 41 (2005) 421–437
Table 3 Ordered logit estimates of the level of severity of a crash Variable
Intcpt 1 Intcpt 2 Intcpt 3 Intcpt 4 drinking dark precipit slip rural ur inc young old male type ED highspd intersec highway year typeyear
All crashes—impact of any drinking by the at-fault driver
Multi-unit crashes—impact of any drinking by the at-fault driver
Parameter estimate
Parameter estimate
a
Standard error
0.320 4.791 2.976a 0.313 0.912a 0.312 0.252 0.312 0.894a 0.040 0.065 0.043 0.053 0.035 0.123a 0.033 0.096a 0.023 0.029a 0.010 0.013a 0.005 0.028 0.023 0.052a 0.030 0.068a 0.020 0.246a 0.064 0.147a 0.063 0.397a 0.027 0.143a 0.020 0.110a 0.029 0.011a 0.003 1.8 · 103a 7.2 · 104 2 Log likelihood: 94221.592 Likelihood ratio: v2 = 1849.0111 N = 48,828
Standard error
a
5.420 0.371 3.672a 0.362 1.557a 0.360 0.185 0.360 0.700a 0.057 0.105a 0.049 0.028 0.042 0.031 0.039 0.107a 0.026 0.028a 0.012 0.009a 0.005 0.005 0.027 0.071a 0.032 0.077a 0.022 0.159a 0.069 0.130a 0.075 0.460a 0.033 0.259a 0.023 0.023 0.036 0.007a 0.003 0.001 7.8 · 104 2 Log likelihood: 70861.16 Likelihood ratio: v2 = 981.9934 N = 38,000
Dependent variable—SEVERITY. Due to space considerations, the estimates for the time of day and time of year (month) fixed effects parameters are omitted from the tables. a Statistical significance at the 95% level of confidence for a one-tailed test.
drivers, the odds ratio estimate is 2.0 when the at-fault driver had been drinking and 2.2 when the at-fault driver was legally intoxicated.19 5.2. The impact of at-fault driver alcohol use on the relative severity of crashes to at-fault versus not-at-fault parties The estimates presented in Table 4 indicate that, for all crashes, at-fault driver alcohol consumption does not significantly alter the likelihood that a crash will result in a not-at-fault party suffering a greater degree of injury than the at-fault driver. Although not reported in Table 4, this finding remains the same for cases of legal impairment. In both cases, the odds ratio estimate is negligible. However, for multi-unit crashes, at-fault driver alcohol use brings about a statistically 19
For details, see Maddala (1983, p. 46) or SAS/STAT UserÕs Manual, Version 8 (1999, pp. 1952–1955).
T.L. Traynor / Transportation Research Part E 41 (2005) 421–437
431
Table 4 Logit estimates of the probability that a crash results in a greater level of injury to a not-at-fault party than to the atfault party Variable
Intercept drinking dark precipit slip rural ur inc young old male type ED highspd intersec highway year typeyear
All crashes—impact of any drinking by the at-fault driver
Multi-unit crashes—impact of any drinking by the at-fault driver
Estimate
Estimate
Standard error
2.841a 0.429 0.019 0.063 0.059 0.058 0.035 0.049 0.088a 0.045 0.104a 0.031 0.003 0.014 0.010a 0.006 0.049a 0.031 0.089a 0.040 0.146a 0.027 0.131a 0.077 0.145a 0.085 0.075a 0.038 0.523a 0.028 0.167a 0.040 0.006 0.004 0.001 8.7 · 104 2 Log Likelihood: 42002.060 Likelihood ratio: v2 = 952.6441 N = 48,828
Standard error
2.302a 0.454 0.174a 0.073 0.110a 0.061 0.050 0.052 0.009 0.048 0.029 0.032 0.001 0.015 0.012a 0.006 0.038 0.033 0.117a 0.041 0.164a 0.028 0.105 0.080 0.156a 0.091 0.182a 0.041 0.241a 0.029 0.027 0.044 0.005 0.004 8.3 · 104 9.1 · 104 2 Log Likelihood: 37079.498 Likelihood ratio: v2 = 316.0174 N = 38,000
Dependent variable—EXTERNAL. Due to space considerations, the estimates for the time of day and time of year (month) fixed effects parameters are omitted from the tables. a Statistical significance at the 95% level of confidence for a one-tailed test.
significant increase in the likelihood that another person suffers a greater degree of injury than the at-fault driver. This is also true for cases where the at-fault driver is legally impaired. The magnitude of the odds ratio estimate is 1.19 in the case of detected alcohol use, and 1.26 in the case of legal impairment. Taken together, the results indicate that while drinking and driving significantly increases the proportion of multi-unit crashes that result in greater injury levels to other parties, there is likely an equivalent increase in the number of injury-producing single car crashes that lead to a negligible aggregate effect.20 The fifth row of Table 7 presents the odds ratio estimates for the impact of at-fault driver alcohol use on the likelihood that a crash will result in no injury to the at-fault driver while causing 20
The parameter estimates for the control variables generally have the signs that would be expected. Noteworthy among the parameter estimates for the control variables are that road conditions are largely not significant determinates of EXTERNAL, and that while young drivers are associated with greater levels of crash injuries among not-at-fault drivers, the opposite is true for elderly drivers. Interestingly, the county based educational attainment variable proved to have a positive correlation with not-at-fault injury likelihood. Also of note is the sign for the parameter estimate for HIGHSPD which is significantly negative for all crashes and significantly positive for multi-unit crashes.
432
T.L. Traynor / Transportation Research Part E 41 (2005) 421–437
Table 5 Tobit estimates of the number of injuries/fatalities suffered in a crash Variable
Intercept drinking dark precipit slip rural ur inc young old male type ED highspd intersec highway year typeyear Scale
All crashes—impact of any drinking by the at-fault driver
Multi-unit crashes—impact of any drinking by the at-fault driver
Parameter estimate
Standard error
Parameter estimate
Standard error
0.609 0.711a 0.070 0.039 0.115a 0.064a 0.059a 0.029a 0.093a 0.056 0.174a 0.202a 0.213a 0.346a 0.371a 0.009 0.009a 1.2 · 103 2.315 N = 48,828
0.417 0.055 0.057 0.047 0.044 0.031 0.013 0.006 0.031 0.040 0.026 0.082 0.084 0.036 0.027 0.039 0.004 9 · 104 0.014
0.801 0.748a 0.129a 0.018 0.014 0.122a 0.073a 0.025a 0.035 0.062 0.168a 0.198a 0.222a 0.502a 0.413a 0.043 0.009a 1.3 · 103 2.492 N = 38,000
0.519 0.084 0.071 0.060 0.055 0.037 0.018 0.008 0.038 0.046 0.032 0.098 0.108 0.048 0.034 0.051 0.005 1.1 · 103 0.018
Dependent variable—NUMINJ. Due to space considerations, the estimates for the time of day and time of year (month) fixed effects parameters are omitted from the tables. a Statistical significance at the 95% level of confidence for a one-tailed test.
injury or death to others. This finding remains substantively unchanged regardless of the specific measure of alcohol involvement or the specific data set (multi-unit or not) used. With respect to all crashes, the odds ratio estimate for detected alcohol use by the at-fault driver is 1.19, while the odds ratio estimate for legal intoxication by the at-fault driver is 1.15. For multi-unit crashes, the odds ratio estimates are 1.47 when the at-fault driver had been drinking and 1.37 when the at-fault driver was legally impaired.21 In all four of these cases, the estimates are significantly different from zero at the 99% level of confidence. 5.3. The impact of alcohol use by at-fault drivers on the numbers of injured or killed parties The Tobit estimates presented in Table 5 indicate that, on average, at-fault driver alcohol use significantly increases the number of injured and killed parties per crash. The estimated magnitudes of this relationship are .71 additional injured/killed parties per crash when any alcohol 21
Regarding the control variables for the estimates presented in Table 6, note that the economic factors, income and unemployment, consistently have a statistically positive impact on EXTINJ, while vehicle characteristics are not important predictors of the dependent variable.
T.L. Traynor / Transportation Research Part E 41 (2005) 421–437
433
Table 6 Tobit estimates of the number of injuries/fatalities suffered by not-at-fault parties in a crash Variable
Intercept drinking dark precipit slip rural ur inc young old male type ED highspd intersec highway year typeyear Scale
All crashes—impact of any drinking by the at-fault driver
Multi-unit crashes—impact of any drinking by impact of any drinking by the at-fault driver
Parameter estimate
Parameter estimate
a
2.592 0.233a 0.036 0.009 0.133a 0.057a 0.058a 0.025a 0.168a 0.022 0.059a 0.016 0.107 0.054 0.731a 0.211a 1 · 104 6 · 104 2.263 N = 48,828
Standard error 0.444 0.062 0.061 0.050 0.047 0.033 0.014 0.006 0.032 0.042 0.028 0.084 0.090 0.039 0.029 0.042 0.004 0.001 0.017
a
1.840 0.483a 0.091 0.002 0.017 0.102a 0.066a 0.014a 0.071a 0.010 0.036 0.019 0.114 0.348a 0.425a 0.081a 0.001 1 · 104 2.194 N = 38,000
Standard error 0.476 0.077 0.065 0.055 0.050 0.034 0.016 0.007 0.035 0.042 0.029 0.088 0.098 0.043 0.031 0.047 0.005 0.001 0.017
Dependent variable—NUMEX. Due to space considerations, the estimates for the time of day and time of year (month) fixed effects parameters are omitted from the tables. a Statistical significance at the 95% level of confidence for a one-tailed test.
consumption is detected for the at-fault driver. When only multi-unit crashes are considered, the estimated coefficients for the alcohol variables are again positive and statistically significant, with point estimates of .74 additional injured/killed parties. Although not reported in Table 5, when the at-fault driver is legally impaired, the estimates are .69 and .81 additional injured/killed parties per crash.22 Again, the estimates are statistically significant. Table 6 summarizes the Tobit estimates for which the number of parties other than the at-fault driver who are injured or killed in a crash was the dependent variable. For all crashes, driver alcohol consumption causes an additional .23 injuries/fatalities to other parties per crash, and legally impaired driving causes an additional .2 injuries/fatalities to other parties per crash. When multiunit crashes are isolated, the estimates are .48 additional injuries/fatalities per crash when the atfault driver was drinking and an additional .57 injuries/fatalities per crash when the at-fault driver was legally drunk.
22 For the Tobit model, the marginal effect represents the marginal impact of at-fault driver alcohol use (or impairment) on the number of injuries/fatalities in a crash. For details, see Maddala (1983, p. 151).
434
T.L. Traynor / Transportation Research Part E 41 (2005) 421–437
Table 7 The marginal effects of at-fault driver alcohol consumption Dependent variable
Probability that a crash results in severe (incapacitating) injury or fatality Probability that a crash results in any injury/fatality Level of severity of worst injury suffered per crash (five point scale) Probability that a crash causes a more severe injury to a not-at-fault party than to the at-fault party Probability that a crash causes an injury or fatality to a not-at-fault party and no injury to the at-fault party Number of injuries/fatalities per crash Number of injuries/fatalities to not-at-fault parties per crash
All crashes
Multi-unit crashes
Marginal impact of any detected alcohol consumption
Marginal impact of legal impairment
Marginal impact of any detected alcohol consumption
Marginal impact of legal impairment
3.49
2.94
2.70
2.33
1.96
1.96
1.75
1.75
2.45
2.48
2.00
2.20
1.01
1.01
1.19
1.26
1.19
1.15
1.47
1.37
0.71 0.23
0.69 0.20
0.74 0.48
0.81 0.57
For the first five rows (which are based on logit and ordered logit estimates), the estimates are odds ratio estimates. These are measures of the average proportional change in the odds ratio attributable to at-fault driver alcohol use after controlling for the other variables in the model, where the odds ratio is defined as the probability that the event occurs (such as a crash resulting in a severe injury or fatality in the first row) divided by the probability that the event does not occur. These estimates are derived by calculating the exponential of the parameter estimates for alcohol use. For the last two rows (which are from the Tobit estimates), the parameter estimates directly measure the marginal impact of at-fault driver alcohol use.
6. Summary and discussion This paper extends the analysis of the impact of driver alcohol use by developing empirical tests of its impact on average crash severity using crash specific datum. In aggregate, the results of the many tests discussed above signify that there is indeed a positive relationship between at-fault driver alcohol use and the violence of the average crash. The estimates presented in this paper lead to the conclusion that, controlling for other influences, crashes in which the at-fault driver had been drinking are relatively more injurious. This is true both as measured by the level of the worst injury suffered by the involved parties as well as by the numbers of injuries and fatalities suffered per crash. The estimates remain statistically significant as the measure of alcohol involvement broadens from legal impairment to any level of alcohol use. This result carries important implications for policy makers wishing to reduce the social costs of driver alcohol use. It appears that drinking and driving not only increases the numbers of motor vehicle crashes that occur, it also increases the injury and fatality losses of a typical crash. Thus, policies should account for this relatively higher crash severity when laws are developed to reduce the negative impact of drinking and driving. Certainly, policies that reduce the frequency of alcohol related crashes will reduce the costs
T.L. Traynor / Transportation Research Part E 41 (2005) 421–437
435
described by this study. Policies which focus on reductions in BACs (as noted in the literature survey) should also be helpful. In addition, with the exception of the relative degree of injury to at-fault drivers versus other victims over all crashes, drinking and driving significantly increases the severity of crashes to other involved parties. This is true both as measured by the impact of alcohol on the degree of injury severity suffered by not-at-fault victims and the number of not-at-fault injuries and fatalities suffered per crash. In addition, alcohol involved crashes are significantly more likely than other crashes to injure innocent parties while leaving the at-fault driver unharmed. The results presented here are by no means a comprehensive analysis of the impact of drinking and driving. These results are only estimates of the impact of alcohol involvement on crashes, and do not attempt to measure the total impact of driver alcohol use on aggregate road safety. Future analyses with alternative data sets will undoubtedly provide more evidence that can help develop appropriate policies regarding drinking and driving. While it is generally assumed that crashes create externalities and that externalities rise as their number and severity rise, precise measures of these externalities are lacking.23 Instead, researchers have tended to use the gross costs associated with property damage, injuries, and fatalities without including measures of compensation provided by at-fault drivers (or their insurers). Future research should account for the impact that driver alcohol use has upon both the frequency and severity of crashes. Additionally, estimates of all externalities created by drunk driving should be pursued in future research efforts by accounting for both the incidence of drunk driving and all the forms of external costs that are created. Finally, future research efforts can assess enforcement effectiveness with crash specific datum by developing interaction terms that measure the relative impact of various enforcement tools on the severity of crashes. With regard to the measurement of the externalities resulting from driver alcohol use, a major difficulty is the estimation of the proportion of drivers on the road at any moment that has been drinking. Levitt and Porter (2001) discovered that the US Fatality Analysis Reporting System provides data which are rich enough, under certain assumptions, to identify the proportion of drivers on the road who have been drinking or who are legally drunk. They estimated that the externality created by drinking and driving in terms of increased fatalities is at least 30 cents per mile. In this study, all costs of alcohol related crash fatalities are included as externalities, which probably does not reflect the true externalities of these fatalities if some of the associated costs are internalized to the drunk driver. In fact, the task of estimating the magnitudes of the various types of losses associated with driver alcohol use has turned out to be a difficult task. Empirical studies have been able to focus only on portions of the issue, such as the fatality specific externalities resulting from drunk driving or the impact of driver alcohol use on aggregate fatalities and aggregate fatal crashes. The losses associated with motor vehicle crashes are ultimately underwritten by the at-fault driver, other involved parties (e.g. other drivers, passengers, pedestrians, motorcyclists, owners of damaged property), and society in general. These include the expected losses from the future stream of productivity net of consumption due to deaths or injuries, the medical costs accrued to the 23
As discussed in a more general analysis of all the external costs of driving by Jones-Lee (1990), when a driver behaves less safely behind the wheel (as is the case when under the influence of alcohol), ceteris paribus, the expected external costs rise as both the driver and other road users face higher expected losses due to crashes.
436
T.L. Traynor / Transportation Research Part E 41 (2005) 421–437
victims, as well as any other costs to society in general from the deaths, injuries, and property losses resulting from crashes. Of these costs, those ultimately underwritten by the at-fault driver through compensatory civil settlements, higher insurance premiums, or their own lost livelihood are internalized.24 The remaining costs are externalities.25 The externalities caused by alcohol related crashes may be underwritten in several different ways. Victims who are not at fault in crashes may be left to underwrite some of their losses themselves through increased insurance premiums, insurance deductibles and co-payments, uncompensated income losses, or through the non-pecuniary losses that may be incurred. The public may also subsidize the losses suffered in crashes through public medical coverage of underinsured victims, the lost net productivity of the victims, or through higher insurance premiums. Clearly, the measurement and evaluation of the externalities of drinking and driving is a very difficult task that is in need of further research.
References Ferrara, S., Zancaner, S., Giorgetti, R., 1994. Low blood alcohol concentrations and driving impairment. International Journal of Legal Medicine 106, 169–177. Frei, P., Kaeser, R., Muser, M., Niederer, P., Waltz, F., 1999. Vehicle structural crashworthiness with respect to compatibility in collisions. Working Group on Accident Mechanics at the Universities of Zurich, Switzerland. Gujarati, D., 2003. Basic Econometrics, fourth ed. McGraw-Hill, New York. Haffner, H., Graw, E., 1996. Changes in the spectrum of alcohol-induced traffic accidents in relation to blood alcohol level. Blutalkohol 33, 78–83. Jones-Lee, M., 1990. The value of transport safety. Oxford Review of Economic Policy 6, 39–60. Kahane, C., 2003. Vehicle weight, fatality risk and crash compatibility of model year 1991–99 passenger cars and light trucks. NHTSA Technical Report No. DOT HS 809 662, National Highway Traffic Safety Administration. Lee, L., 1997. The socioeconomics of drunk driving. Journal of Socio-Economics 26, 95–106. Lemaire, J., 1985. Automobile Insurance: Actuarial Models. Kluwer-Nijhoff Publishing, Hingham, MA. Levitt, S., Porter, J., 2001. How dangerous are drinking drivers? Journal of Political Economy 109, 1198–1237. Maddala, G., 1983. Limited-Dependent and Qualitative Variables in Econometrics. Cambridge University Press, Cambridge. McCarthy, P., 2003. Alcohol-related crashes and alcohol availability in grass-roots communities. Applied Economics 35, 1331–1338. Miller, T., Blewden, M., 2001. Costs of alcohol-related crashes: New Zealand estimates and suggested measures for use internationally. Accident Analysis and Prevention 33, 783–791. Muelleman, R., Mueller, K., 1996. Fatal motor vehicle crashes: variations of crash characteristics within rural regions of different population densities. Journal of Trauma 41, 315–320. National Highway Traffic Safety Administration, Initiatives to Address Impaired Driving, December 2003. National Highway Traffic Safety Administration, Impaired Driving in the United States: State Alcohol Cost Fact Sheets, National Highway Traffic Safety Administration, 2002. National Highway Traffic Safety Administration, Impaired Driving in the United States: State Cost Fact Sheet UserÕs Guide, 2002.
24
Aside from the reduction of externalities per accident, criminal penalties for driving under the influence (DUI) as well as penal damages awarded under civil law for alcohol related crashes are designed to reduce externalities from alcohol related crashes by discouraging drinking and driving. 25 For a more complete discussion of actuarial methods used by the automobile insurance industry, see Lemaire (1985, pp. 39–56).
T.L. Traynor / Transportation Research Part E 41 (2005) 421–437
437
Orsay, E., Doan-Wiggins, L., Lewis, R., Lucke, R., RamaKrishnan, V., 1994. The impaired driver: hospital and police detection of alcohol and other drugs of abuse in motor vehicle crashes. Annals of Emergency Medicine 24, 51–55. Owens, D., Sivak, M., 1996. Differentiation of visibility and alcohol as contributors to twilight road fatalities. Human Factors 38, 680–689. Peek-Asa, C., Kraus, J., 1996. Alcohol use, driver, and crash characteristics among injured motorcycle drivers. Journal of Trauma, Injury, Infection, and Critical Care 41, 989–993. Peltzman, S., 1975. The effects of automobile safety regulation. Journal of Political Economy 83, 677–725. Perper, J., Kuller, L., Shim, Y., 1993. Vehicular crashes and alcohol involvement in drivers at fault, and related fatalities. American Journal of Forensic Medicine and Pathology 14, 177–184. Robertson, M., Drummer, O., 1994. Responsibility analysis: a methodology to study the effects of drugs in driving. Accident Analysis and Prevention 26, 243–247. Ruhm, C., 1996. Alcohol policies and highway vehicle fatalities. Journal of Health Economics 15, 435–454. Solnick, S., Hemenway, D., 1994. Hit the bottle and run: the role of alcohol in hit-and-run pedestrian fatalities. Journal of the Study of Alcohol 55, 679–684. Traynor, T., 1993. The Peltzman hypothesis revisited: an isolated evaluation of offsetting driver behavior. Journal of Risk and Uncertainty 7, 237–247. Traynor, T., McCarthy, P., 1993. Economic regulation and highway safety in the trucking industry: a limited dependent variable analysis. Quarterly Review of Economics and Finance 33, 141–153. Waller, P., Blow, F., Maio, R., Singer, K., Hill, E., Schaefer, N., 1997. Crash characteristics and injuries of victims impaired by alcohol versus illicit drugs. Accident Analysis and Prevention 29, 817–827. Zador, P., 1991. Alcohol-related relative risk of fatal driver injuries in relation to driver age and sex. Journal of the Study of Alcohol 52, 302–310.
Further reading Blomquist, G., 1988. The Regulation of Motor Vehicle and Traffic Safety. Kluwer Academic Publishers, Boston. Greene, W., 2003. Econometric Analysis. Macmillan Publishing Co., New York. Lloyd, C., 1990. Estimating the effect of alcohol on the risk of a fatal road accident. Journal of the Royal Statistical Society 153 (Part 1), 29–52. Robertson, L., 1993. Blood alcohol in fatally injured drivers and the minimum legal drinking age. Journal of Health, Politics, Policy, and Law 14, 817–825.