An Exploratory Multinomial Logit Analysis of Single-Vehicle Motorcycle Accident Severity

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 12

Joumal of Safety R-h, Vol. 27, No. 3. pp. 183-194.

19%
CopyrightQ 19% NationalSafety council and Elscvier Science Ltd
Pergamon Ptinted in the USA. Au rights reserved
00224375/X $15.00 + .OO

PI1 S00224375(%)00010-2

An Exploratory Multinomial Logit Analysis of


Single-Vehicle Motorcycle Accident Severity

Venkataraman Shankar and Fred Manner&

Most previous research on motorcycle accident severity has focused


on univariate relationships between severity and an explanatory variable of
interest (e.g., helmet use). The potential ambiguity and bias that univariate
analyses create in identifying the causality of severity has generated the
need for multivariate analyses in which the effects of all factors that
influence accident severity are considered. This paper attempts to address
this need by presenting a multinomial logit formulation of motorcycle-
rider accident severity in single-vehicle collisions. Five levels of severity
are considered: (a) property damage only, (b) possible injury, (c) evident
injury, (d) disabling injury, and (e) fatality. Using S-year statewide data on
single-vehicle motorcycle accidents from the state of Washington, we
estimate a multivariate model of motorcycle-rider severity that considers
environmental factors, roadway conditions, vehicle characteristics, and
rider attributes. Our findings show that the multinomial logit formulation
that we use is a promising approach to evaluate the determimmts of motorcycle
accident severity.

INTRODUCTION spatial, and nonrenewable (e.g., fossil fuels)


resource concerns. However, despite advantages
Motorcycles provide an efficient form of in these areas relative to passenger cars, motor-
transportation with respect to environmental, cycles have declined in popularity since their
peak years in the late 1960s. To some extent, this
decline can be explained by the comparatively
Venkataraman Shankar is currently a PhD candidate in
high accident risk and high risk of severe injury
the Civil Engineering Department at the University of
Washington. He received his Masters degree from the associated with motorcycle riding. To this end,
University of Washington in 1990. He has authored a statistical analyses of the likelihood of motorcy-
number of papers on vehicular safety and the focus of his cle accidents and their severity have the poten-
dissertation is the application of new methodological
tial to uncover underlying causality and clarify
approaches to vehicle-safety reseatch.
Fred Mannering is a professor in the Civil Engineering possible misunderstandings relating to motorcy-
Department at the University of Washington. He received his cle safety. Such analyses could help identify fac-
PhD from the Massachusetts Institute of Technology in 1983. tors that one can control to keep motorcycle
Dr. Mannering has authored numerous papers in the ateas of
risks at an acceptable level, thereby saving lives,
travel demand, economics, vehicular safety, automobile
demand, statistical methodologies, and transportation preventing injuries, and making motorcycling a
network analysis. more competitive mode of travel.

Fall 19%Nolume 27hkmber 3 183


Recent studies have shown that motorcyclists In contrast, multivariate studies of automo-
arc generally aware of the key factors that con- bile accident severity have been more common,
tribute to an increased risk of an accident. For and such studies have the potential to provide
example, Mannering and Grodsky (1994) found some general insight into the factors that deter-
that factors such as driving exposure, riding at mine motorcycle accident severity. Examples of
speeds that exceed posted speed limits, improper automobile-severity studies include the work of
lane changes, and passing on double yellow Jones and Whitfield (1988) in which logistic
lines, were all recognized by motorcyclists to regression was used to model the severity risk as
increase the likelihood of an accident. However, a function of anthropometric measures, vehicle
accident severity (as opposed to the likelihood mass, age of rider, and restraint system use.
of an accident occurring) presents another, Logistic regression was also employed by Lui,
arguably less understood, phenomenon. The McGee, Rhodes, and Pollock (1988) to model
study of the determinants of the conditional dis- the probability of fatalities conditioned on the
tribution of accident severity (i.e., conditioned occurrence of an accident using variables such
on the occurrence of an accident) has the poten- as driver age, gender, impact points, vehicle
tial to provide additional insights into the true crash severity, restraint system use, and vehicle
consequences of motorcyclists’ behavior (e.g., mass. Other methodologies such as multivariate
speeding, helmet use, etc.) and the interactions time-series (Lassarre, 1986). double-pair com-
of this behavior with environmental, roadway, parison (Evans, 1986), headway-based severity
and vehicle characteristics. (Glimm & Fenton, 1980). bivariate probit
Most previous research on motorcycle accident (Hutchinson, 1986), and discriminant analysis
severity has focused on issues relating to head (Shao, 1987) have been successfully applied.
injuries and fatalities (see Chenier & Evans, 1987; Finally, Shankar, Mannering, and Barfield (in
Gabella, Reiner, Hoffman, Cook, & Stallones, press) have shown the potential of the nested
1995) and on helmet-related issues such as pat- logit specification by using environmental, geo-
terns of head injuries and effectiveness of helmets metric, weather, vehicular, and human factors to
in reducing both fatalities and severity of head develop a predictive model of automobile acci-
injuries (Bachulis, Sangster, Gorrell, & Long, dent severity.
1988; Evans & Frick, 1986; Foldvary & Lane, The present study attempts to extend the
1%; Watson Zador, & W& Unfortunately, empirical and methodological contributions of
toward a previous work (drawing on past findings from
univariate examination of accident severity (with studies of both automobile and motorcycle acci-
respect to such as use or alcohol dent severity) by developing a model of motor-
consumption). little research cycle-rider accident severity (for single-vehicle
safety accidents) that can be used to understand the
using a true multivariate examination of the impacts of helmet use, alcohol-impaired riding,
minants of accident severity (i.e., and other factors, on accident severity. The focus
on single-vehicle accidents allows us to concen-
possi- trate on accidents in which the motorcyclist was
ble that a systematic result as well as likely to have made an error in judgment, and
ambiguity thus we avoid the variability in accident severity
(1986) and Weiss introduced when multiple-vehicles are involved
has made important contributions by modeling (i.e., cases where automobiles may significantly
influence the overall accident severity).
(i.e., a true multivariate approach) on motorcy- Our paper begins with a discussion of the
proposed methodological approach. This is fol-
lowed by a description of data used in model
probit estimation. We then present model estimation
model to examine the severity results and a detailed discussion of our findings
(Hurt, Ouellet, Thorn, and their implications for accident severity anal-
198 1). These studies have provided ysis. The paper concludes with a brief discus-
sion of model specification issues, a summary,
and directions for future research.
184 Journal of Safety Research
METHODOLOGY P,(i) = P(piX” - /31X,2 Em - Ein) v I f i (4)

We begin by developing a conditional model With equation 4, an estimable severity model


of motorcycle accident severity, that is a model can be derived by assuming a distributional form
conditioned on the fact that a single-vehicle for the error term. A natural choice would be to
motorcycle accident has occurred (see Shankar, assume that this error term is normally distribut-
Mannering, & Barfield, 1995, for a study of the ed. Such an assumption results in a probit model.
likelihood of an accident occurring). Severity of However, probit models are computationally dif-
a single-vehicle motorcycle accident is specified ficult to estimate (see Ben-Akiva & Lerman,
to be one of five discrete categories: (a) property 1985). A more common approach for models of
damage only, (b) possible injury, (c) evident this type is to assume that Ein’s are generalized
injury, (d) disabling injury, and (e) fatality. The extreme value (GEV) distributed. The GEV
injury classification is made on the basis of rider assumption produces a closed form model that
injuries only, the injuries to passengers (if any) can be readily estimated using standard maxi-
are not considered (this will be discussed again mum likelihood methods. It can be shown
later in the paper). Given these five discrete (McFadden, 1981) that the GEV assumption
severity categories, a statistical model that can produces the simple multinomial logit model,
be used to determine the probability of an acci-
P,(i) = exp[&l
dent having a specific severity level can be I I$ exp[P&l (5)
derived. We start the derivation with the follow-
ing probability statement, where all variables are as previously defined
and the vector pi is estimable by standard maxi-
P,(i) = P(S, 2 SI,) tl I # i mum likelihood methods.

where P,(i) is the probability that accident n is


severity i, P denotes probability and Si, is a EMPIRICAL SETTING
function of covariates that determine the likeli-
hood of accident n being severity i (I is the set Data on single-vehicle motorcycle accidents
of possible severities). To estimate this probabil- (i.e., accidents involving one motorcycle) were
ity, a function defining the severity likelihoods obtained from the Washington State Department
must be specified. We use a linear form such of Transportation accident files. These files con-
that, tain data that can be classified into five cate-
gories: (a) general accident data, (b) weather
data, (c) pavement surface data, (d) vehicle data,
and (e) rider-related data. General accident data
where X,, is a vector of measurable characteristics included information on rider injury in terms of
that determine the severity (e.g., rider age, rider the five accident-severity categories (i.e., prop-
gender, roadway attributes, prevailing weather erty damage only, possible injury, evident injury,
conditions, vehicle type, use of helmets, and so disabling injury, and fatality, as previously dis-
on), pi is a vector of estimable coefficients, and cussed); rider injury classifications (such as
I+,, is an error term that accounts for unobserved head, torso, limb, or multiple injuries); time of
factors influencing accident severity. The term day of accident; accident location with respect
PiX, in this equation is the observable component to the traveled way (on or off the roadway,
of severity determination because the vector X,, whether the accident occurred on a curve or
contains measurable variables (e.g., roadway straight section or a grade, roadway illumination
attributes at the location of accident n), and in is information, types of roadside objects involved
the unobserved portion. in collision, and accident type); and roadway
Given equations 1 and 2, the following can functional classification. Weather data contained
be written, information on atmospheric conditions at the
time of the accident (e.g., rain, snow, fog) and
P,(i) = P(fliX” + I+*2 &X, + Eh) tr I * i C3) pavement-related data contained information on
pavement surface conditions at the time of the
or, accident (e.g., ice, snow, wet, or dry). Vehicle
Fall 19%.44olume 27LVwnber 3 185
data included information on motorcyclists’ hel- the extent of severity misjudgment, in our case,
met use, ejection status of occupants (i.e., is less likely to be a serious problem because
whether or not the motorcyclist was ejected we consider only the injuries of the motorcycle
from the vehicle), motorcycle headlight operat- operator and not passengers. However, our
ing condition at the time of accident (whether findings should be viewed with this potential
headlights were on), and whether or not a pas- bias in mind.
senger was present. Motorcycle rider-related Of the 1,594 single-vehicle motorcycle acci-
data included information on sobriety at the time dents, 944 had missing data and thus the sam-
of accident, age, gender, and rider action leading ple was reduced to 650. An analysis of this
to the accident (e.g., inattention, not granting reduced sample, by statistically comparing the
right of way, improper passing, excessive speed, means of variables contained in both the reduced
and so on). sample and the entire sample, revealed no sys-
Our accident data was drawn from a 5-year tematic bias resulting from the elimination of
period between 1989 and 1994. Over this period, observations with missing data. This suggests
there were a total of 3,469 reported accidents that missing data seems to be random (i.e., per-
involving a motorcycle, 1,594 (45.95%) of haps due to random data recording mistakes and
which were single-vehicle accidents (there were omissions) and not correlated with the severity
1,715 two-vehicle accidents and 160 accidents or other accident-related attributes. Of the 650
involving 3 or more vehicles). Before proceed- accidents included in our sample, 32 resulted
ing, it is important to discuss the consequences in property damage only, 86 resulted in possi-
of using only reported accidents as a basis for ble injuries, 176 resulted in evident injuries,
estimating our severity model and the conse- 322 resulted in disabling injuries, and 34 result-
quences of using police officer judgments of ed in fatalities. As stated previously, the injury
severity instead of actual medical records. The classification is made using rider injuries only
concern with using only reported accidents is - injuries to passengers (if any) are not con-
that it is likely that many accidents (particularly sidered (only 18.31% of the 650 accident-involved
those that are minor in severity) may go unre- motorcycles had a passenger at the time of the
ported. This means that our accident sample is accident). It should also be noted here that pos-
not a random sample of all accidents. Fortunately sible injury accidents (which may seem to be a
this under-reporting will have a minimal impact somewhat vague category) are determined at
on multinomial logit model estimation results. In the scene by Washington State troopers using
fact, all coefficients will be correctly estimated well-defined, uniformly taught identification
with the exception of the constant terms. If the procedures. Our testing of various model struc-
number and severity of unreported accidents tures suggests that this is a unique severity cat-
were known, the constant term reported in this egory and must be considered separately (i.e.,
paper could be adjusted by a simple calculation even though the accident will eventually be
and no additional estimation would be necessary classified as an injury or a property damage
(see Ben-Akiva & Lerman, 1985, for details on only accident).
such stratified-sample adjustments). Table 1 provides information on the distribu-
With regard to our using police officer judg- tion of accident severity by important conditioning
ments of severity as opposed to actual medical variables such as helmet use, sobriety, functional
records, some caution should be exercised in classification of roadway, engine displacement,
interpreting our findings. This is because sever- and whether or not the accident occurred at an
al studies have shown discrepancies between intersection. An interesting point in this table is
officer judgments and medical records. For the comparatively high number of riders that were
example, Barancik and Fife (1985) found that nonhelmeted at the time of the accident (12.15%).
15% of patients showing up at emergency Although Washington State had a helmet law in
departments for vehicle-related injuries had effect throughout this period, it was declared
police reports that indicated no injury. Barancik unconstitutional for much of 1993 (during which
and Fife go on to note that officer reports are time many law-enforcement agencies refused to
much more likely to be accurate when estimat- enforce it) until it was rewritten. Thus it appears
ing the severity of the injury to the vehicle driv- that quite a few riders reverted to nonhelmet use
er relative to passengers. It could be argued that during this period. As will be shown in forth-
186 Journal of Safety Research
TABLE 1
SINGLE-VEHICLE MOTORCYCLE ACCIDENT SEVERITY DISTRIBUTION BY KEY VARIABLES

Severity Frequency

Property Possible Evident Disabling


Accident Conditioning Variable Percent of Entlre Sample Damage InjUrY Injury Injury FataMy

Helmet use................... 07.65 28 80 149 265 23

Helmet not used ................ 12.15 4 6 27 37 5

Alcohol-impaired riding ........... 21.08 13 52 55 16

Sober riding ................... 76.92 31 73 124 267 18

Interstate location .............. 23.65 9 37 40 100 6

Arterial location ................ 53.06 15 43 100 167 20

High displacement (greater than 500~~) 60.92 17 43 113 193 24

Intersection location ............. 17.36 4 18 23 67

Non-intersection location .......... 82.62 28 66 153 255 33

coming model estimation, having 12.15% of the accident severity (as measured by f-statistics).
sample being nonhelmeted is a large enough Finally, Table 2 shows the frequency and per-
percentage to produce highly significant find- centage of accidents conditioned on several of
ings with regard to the impact of helmets on these variables.

TABLE 2
ACCIDENT FREQUENCY AND PERCENTAGE BY CONDITIONING VARIABLE

Primary Accident Conditioning Variable Interaction Variable Frequency Percentage

Helmet Fixed-object Collision 201 35.20

Non-fixed-object Collision 370 64.60

No helmet Fixed-object Collision 34 43.03

Non-fixed-object Collision 45 56.97

Wet pavement Dry Weather 28 49.1

Rainy Weather 23 50.9

Exceeding speed limit None 53 9.06

No helmet Low-speed Riding (less than 72 km/h) 28 35.44

High-speed Aiding (72 km/h or htgherl 51 64.56

Alcohol-impaired riding Helmeted 107 78.10

Non-helmeted 30 21.90

Inattention None 60 12.31

Interstate location None 194 23.65

Rider ejected None 464 71.36

High-displacement lgreater than 500 cc) None 396 60.92

Young rider (less than 2 1 Years old) High Displacement (greater than 500 cc) 64 66.09

Low Displacement 1500 cc or less) 30 31.91

Note: Percentages for the exceeding speed limit, inattention, interstate, rider ejected and high-displacement variables are given in
relation to the entire sample. Percentages for other variables shown are shown by prunary class, i.e. helmet, no-helmet, alcohol-
impaired riding, and young rider.

Fall 1996NoIume 27hkunber 3 187


FIGURE 1
ACCIDENT SEVERITY MODEL STRUCTURE

MODEL ESTIMATION wear a helmet could be a self-selected group that


has riding habits and skills that make them less
Estimation of the multinomial logit specifica- susceptible to fatal injuries in fmed-object acci-
tion (with severity alternatives as shown in dents. Unfortunately, our sample of nonhelmeted
Figure 1) was carried out using standard maxi- fatalities in fixed-object collisions is too small to
mum likelihood methods. Results of the estima- statistically explore this third possibility in
tion are shown in Table 3. This table shows that detail. It must be pointed out that, although our
the signs of all model coefficients are of plausi- finding relating to this interaction term is highly
ble sign and that the model has good overall significant (r-statistic = 4.57), further research
convergence, with the log-likelihood converging into this somewhat surprising finding is warrant-
from -1046.1 to -779.22, giving a p2 of 0.256. ed with a larger data set that includes a greater
An interpretation of the model findings is pro- number of nonhelmeted riders.
vided below.
Variable: No-helmet/fixed-object interaction.
Variable: Helmeted-rider/fixed object interaction. Finding: Increases the likelihood of evident injury.
Finding: Increases the liielihood of fatality.
This variable indicates that nonhelmeted rid-
It is generally believed that helmet use tends ers face a greater risk of evident injury, relative
to decrease the occurrence and severity of head to other types of accident severities, when they
injuries. It is also widely believed that helmets, are involved in fmed object collisions. It is possi-
by themselves, will be most effective in reducing ble that nonhelmeted riders may be cognizant of
fatalities where head injuries are the primary the risks of riding without a helmet and maintain
cause of death. However, the effectiveness of lower speeds, and/or they could be a self-selected
helmet use in reducing fatalities when other fac- sample of more experienced, safer riders. These
tors (e.g., speed, roadway conditions) are consid- possibilities could explain their increased likeli-
ered, is not well understood. Our finding that hood of evident injury as opposed to the more
helmeted riders have an increased likelihood of serious severity categories of disabling injury
being fatally injured when colliding with fixed and fatality.
objects, while seeming improbable at first Variable: No-helmet/alcohol-impaired riding
glance, could be explained in at least three ways. interaction.
First, the increased likelihood of a fatal injury Finding: Increases the likelihood of fatality.
could be due to some physiological factors asso-
ciated with helmeted riders hitting fixed objects This variable captures both the lack-of-aware-
in single-vehicle collisions. Second, the increased ness and risk-taking behavior of al~hol-impaired
likelihood could be an outgrowth of helmeted riders (those riders with their abilities impaired by
riders tending to ride more recklessly in response alcohol, as determined by the officer at the acci-
to that added sense of security that a helmet pro- dent scene). Several studies (Murdock & Waxman,
vides (i.e., risk compensation as previously 1991; Nelson, Sklar, Skipper, & McFeeley, 1992)
explored by msearchers such as Peltzman, 1975, have found a strong association between alcohol
and Adams, 1983). Third, riders choosing not to use and nonhelmeted riding. Although this could

188 Joumal of Safety Research


TABLE 3
ESTIMATION OF THE MULTINOMIAL LOGIT MODEL OF MOTORCYCLE-RIDER SEVERITY

Estimated
Variable Coefficient t-statistic

Constant (specific to fatalityl -3.13 -8.06

Helmeted-rider-fixed object interaction indicator (1 if rider was helmeted and hrt a


fixed object, 0 otherwise; specific to fatalityl 2.02 4.57

No-helmet, fixed object interaction indicator (1 If rider was non-helmeted and hit a
fixed object. 0 otherwise; specific to evident injury) 0.81 2.18

No-helmet, alcohol-impaired riding interaction indicator (1 if rider wes non-helmeted


and alcohol impawed, 0 otherwise; specific to fatality) 2.36 3.49

No-helmet, low-speed riding interaction indicator (1 if rider was non-helmeted and


estimated speed at time of accident was less than 72 km/h, 0 otherwise: specific to
evident and disabling injury) 1.41 1.84

Alcohol-impaired riding indicator I1 if riding ability was impaired by alcohol, 0


otherwise; specrfic to fatality, evident injury and disabling injury1 0.78 2.49

Motorcycle displacement indicator (1 if displacement is greater than 500 cc, 0


otherwise; specific to fatality, evident injury and disabling injurv) 0.31 1.49

Age-displacement interaction indicator (1 if rider age is less than 21 years old and
displacement of motorcycle is greater than 500 cc, 0 otherwise; specific to property
damage, disabling injury and possible infuryl . 0.77 2.46

Motorcycle rider age es a continuous variable (specific to fatality and disabling


injuryl . . . . . . . . . . . . . . . . . 0.014 4.57

Ejection of rider indicator (1 if rider was ejected. 0 otherwise; specific to fatality,


evident injury, disabling injury and possible injuryl 1.05 3.17

Speeding indicator (1 if exceedrng speed limit was the primary cause of the
accident, 0 otherwise; specific to fatality, evident injury and disabling injury) 0.88 1.95

Vehicle speed at the time of accident as a continuous variable (specific to fatalrty,


evident injury and disabling injury) 0.014 3.24

Rider inattention indicator I1 if inattention was the primary cause of the accident, 0
otherwise; specific to evident and disabling injury) 0.85 2.36

Roadway type indicator 1 (1 if accident occurred on an interstate freeway. 0


otherwise; specific to disabling injury) 0.41 2.22

Roadway type indicator 2 (1 if accident occurred on an mterstate freeway, 0


otherwise; specific to possible injury) 0.87 3.22

Wet pavement/not-raining interaction indicator (1 if pavement surface was wet and


it was not raining at the time of accident, 0 otherwise; specific to property damage
and possible injury) . 1.56 3.77

Number of observations . 650


Log-likelihood et zero . -1046.1
Log-likelihood at convergence . -779.22
p*,...,,.,,,,,,..,,,,..,...,....,.,......,................,.................,........,.. 0.256

be true in a general context, the fmding on alco- This finding has important ramifications in
hol-impaimd riding without a helmet suggests that unraveling the true effectiveness of helmets
riding without a helmet while alcohol-impaired is when speed is taken into account. Intuitively
indeed a deadly combination. one might expect that riding without a helmet
would predispose riders to a greater risk of fatal-
Variable: No-helmet/low-speed interaction. ity. However, our finding here suggests that
Finding: Increases the likelihood of evident and when the interaction with speed is taken into
disabling injury. account, this expectation is tempered by the
Fall I994Nohme 27/Number 3 189
presence of several mitigating factors such as severe forms of injury including fatalities, when
low speed riding, less-risky riding behavior, and aggravating factors such as speed and alcohol-
the possibility that nonhelmeted riders may be a impaired riding are controlled for, the risk of
self-selected group of safer riders. While riders evident injury seems to be reduced to milder
traveling at speeds less than 72 km/h without a forms while the risk of fatality seems to be
helmet are less likely to be involved in an acci- reduced to disabling injuries.
dent in which they are fatally injured, our results
show that they are also less likely to be involved Variable: Motorcycle rider age.
in property damage only and possible injury Finding: Increases the likelihood of fatality and
accidents. Thus such behavior increases the like- disabling injury.
lihood that the accident will be in the middle of
the severity range. This variable captures the simultaneous effect
of several factors. Although older riders may
Variable: Alcohol-impaired riding. tend to ride at lower speeds and be less likely to
Finding: Increases the likelihood of fatality, evi- be in an accident, once in an accident they tend
dent, and disabling injuries. to have severe injuries. This may be due to phys-
iological factors associated with advanced age.
This finding is consistent with several studies
(Luna, Maier, Sowder, Copass, & Oreskovich, Variable: Ejection of rider.
1984; Waller et al., 1985) that have reported that Finding: Increases the likelihood of any form of
alcohol has a strong association with an increase injury relative to property damage.
in traumatic injury in vehicular accidents. This
variable captures inherent risk-taking tendencies From a statistical perspective, this variable
in alcohol-impaired riders and also the potential may be endogenous because riders with more
of alcohol to increase the severity of head injuries severe injuries (i.e., injuries incurred as the acci-
once a crash has occurred (see Waller et al.). dent evolves) may be more likely to be ejected.
This means that a left-side (dependent) variable,
Variable: Motorcycle displacement. injury severity, can change the value of the
Finding: Increases the likelihood of fatality, evi- right-side variable, rider ejection. If the model
dent, or disabling injuries. were to be estimated without considering this
possible endogeneity, model coefficient esti-
This variable, used as an indicator for dis- mates would be biased and inconsistent. To
placements exceeding 500 cc, suggests that high- resolve this potential specification problem, we
er displacement vehicles pose a greater risk of estimate a binary logit model of ejection proba-
severe forms of injury to riders, perhaps as a bility (using all other exogenous variables as
result of the effects that such vehicles may have predictors) and use the estimated probability of
on rider behavior and/or the higher vehicle weight being ejected from this model as the variable in
associated with larger displacement motorcycles. our severity model. This approach ensures unbi-
ased and consistent model estimation.
Variable: Age-displacement interaction. The resulting coefficient estimate confirms
Finding: Increases the likelihood of property earlier expectations - that ejection from the
damage, possible injury, and disabling motorcycle tends to result in some form of
injury. injury primarily due to the impact with some
hard object or the pavement, at speed.
This variable captures an interesting causal
process behind the interaction of age and dis- Variable: Speeding.
placement. Young riders (aged 20 or less) when Finding: Increases the likelihood of fatality, evi-
riding motorcycles of displacement 500 cc or dent injury, and disabling injury.
higher are either prone to be involved in colli-
sions of mild severity such as property damage This finding indicates that if the collision
and possible injury or severe injury such as dis- occurs due to the rider exceeding the speed
abling injuries. While this variable by itself limit, the consequence is likely to be some
would appear to predispose young riders to severe form of injury. This underscores the dan-
190 Journal of Safefy Research
gem of riding at speeds that exceed the design This variable suggests that if the accident
limits of the highway. occurred on wet pavement with no rain falling, the
severity of the crash tends to be limited to proper-
Variable: Vehicle speed at time of accident. ty damage and possible injury. Wet pavements
Finding: Increases the likelihood of fatality, evi- may be acting as visual deterrents to speeding,
dent injury, and disabling injury. which may encourage motorcyclists to maintain
longer headways and lower riding speeds.
This finding shows, as expected, that the like- To examine the marginal effects of some key
lihood of injury increases to more severe forms factors, elasticities of the variables included in
as vehicle speed increases. This variable captures the model were estimated. The elasticity of vari-
the gradual increase in urgency and complexity ables, in our case, will measure the responsive-
of the riding task as riding speed increases, and ness of the probability of some accident-severity
the increase in kinetic energy that must be dissi- category to changes in variable values. In gener-
pated (often on the body) as speed increases. al, elasticity is computed as

Variable: Rider inattention. P 6) _ aP,(i). x,


EX”” (6)
Finding: Increases the likelihood
disabling injury.
of evident and ax, P,(i)
were E represents the elasticity, x is the value of
When rider inattention is identified as the pri- the variable being considered, and P,(i) is the
mary factor for accident occurrence, it indicates probability of accident n being of severity i.
that speed was a less significant factor. As a Applying this equation to the multinomial logit
result, the rider tends to be less prone to fatal formulation (equation 5) gives
injury but is more prone to evident or disabling
injuries as a result of improper evasive action. E~(i)=(l-FPncj))Pix~
(7)

Variable: Interstate as a roadway type.


where J is the set of alternatives that have the
Finding: Increases the likelihood of disabling
variable x, in their severity function (Si” in equa-
and possible injury.
tion 2). Equation 7 is only valid for continuous
variables (x) such as rider age and vehicle speed.
This variable indicates that, given that the
For noncontinuous variables (i.e., those variables
accident occurred on an interstate freeway, its
that take on values of zero or one, such as riding
severity is more likely to be in the possible and
when alcohol-impaired or riding a motorcycle
disabling injury categories. Interstates, although
over 500 cc in displacement) we estimate a
high-speed facilities, are the safest class of road-
“pseudo-elasticity” that captures the approximate
ways in the United States, with adequate radii
elasticity of the variable. In the multinomial logit
on curves, ample sight distance on grades, and
formulation this pseudo elasticity is given as
limited access by design. In addition, interstate
riders seem to have higher helmet use (perhaps E p”(i) =
due to riders’ increased perception of danger ‘n
due to higher speeds and more stringent helmet
law enforcement), with less than 5% of inter- exp[A(PiXg)lCexPIPjX,l
state accidents involving nonhelmeted riders as J 1
opposed to more than 12% in the overall sam-
ple. The combination of these factors seems to exp[A(~iX,)I~eXP[~jX~I +CexP[&l -(8)
J J+i
concentrate single-vehicle motorcycle accident
severity probabilities toward possible and dis- where all variables are as previously defined.
abling injuries and away from property damage Table 4 presents the elasticity computations
only, evident injury, and fatality. for some select variables. As an intuitive exam-
ple of what these values mean, the elasticity of
Variable: Wet pavement/not-raining interaction. rider age (the value of which is 0.198 from
Finding: Increases the likelihood of property Table 4) means that a 1% increase in a rider’s
damage and possible injury. age results in a 0.198% increase in the probability

Fall 1996Nolume 27hVumber 3 191


TABLE 4
ELASTICITIES OF KEY DETERMINANTS OF MOTORCYCLE-RIDER ACCIDENT SEVERITY

Variable (see corresponding variable definitions in Table 3) Elasticity/Pseudo-Elasticity

Helmeted-rider-fixed object interaction (specific to fatality) 5.669

No-helmet, fixed object interaction (specific to evident injury) . 0.694

No-helmet. alcohol-impaired riding interaction (specific to fatality) 6.949

No-helmet, low speed ridmg interaction (specific to evident and disabling injury) 0.235

Alcohol-impaired riding (specific to fatality. evident injury and disabling injury) 0.483

Motorcycle displacement (specific to fatality, evident injury and disabling injury) 0.223

Age-displacement interaction (specific to property damage, disabling injury and possible injury) 0.322

Motorcycle rider age (specific to fatality and disabling injury1 . 0.196

Ejection of rider (specific to fatality, evident injury, and disabling injury and possible injury) . 0.714

Speeding (at time of accident, specific to fatality, evident injury and disabling injury) 0.531

Vehicle speed at the time of accident (specific to fatality, evident injury and disabling injury 0.112

Rider inattention (inattention at time of accident, specific to evident and disabling injury) 0.172

Roadway type indicator 1 (interstate freeway accident, specific to disabling injury) 0.216

Roadway type indicator 2 (interstate freeway accident, specific to possible injury) . 1.062

Wet pavement/not-raining interaction (specific to property damage and possible injury) 2.021

that the rider’s accident will be either a disabling only and possible injury accidents may share
injury or fatality. Other values in this table can unobservables such as internal injury or effects
be interpreted in a similar manner. The results in associated with lower-severity accidents. In the
Table 4 show that only helmeted-fixed object, presence of shared unobservables, the logit for-
alcohol-impaired riding while without a helmet, mulation will erroneously estimate the model
and pavement surface/weather interactions as coefficients. The problem of shared unobserv-
well as the interstate freeway indicator (with ables is referred to as an independence of irrele-
possible injury) were found to be elastic (i.e., an vant alternatives (IIA) specification error. To test
elasticity value greater than 1). This underscores for the possibility of this error, the Small-Hsiao
the comparative importance of these four vari- test of the IIA assumption was used (Small &
ables in estimating single-vehicle motorcycle Hsiao, 1985). The test results showed that IIA
accident severity. violations were not significant in our model at
the 95% confidence level. We thus conclude that
our model is properly specified with regard to
MODEL SPECIFICATION ISSUES this important IL4 concern (see Shankar et al., in
press, for an alternative model formulation when
the IIA property is violated).
The multinomial logit model used in this paper
(as shown in equation 5) can potentially be
afflicted with a serious specification problem
because the derivation of this model requires us SUMMARY AND DIRECTIONS FOR
to assume that the unobserved terms @,,‘s) are FURTHER RESEARCH
independent from one severity type to another.
Intuitively, it is possible that some severity types This paper provides an important method-
could share unobserved terms and have a corre- ological framework (i.e., the use of a multinomi-
lation that violates the assumption made during al logit specification) for estimating motorcycle
model derivation. For example, property damage rider accident severity likelihood conditioned on
192 Journal of Safety Research
the occurrence of an accident. In association the methodological approach demonstrated in
with the extant body of research on motorcycle this paper needs to be applied to larger and more
safety, the findings of this study confirm some detailed databases. In the past, studies that have
previous conclusions, and more importantly used large and detailed motorcycle-accident
unravel some causal processes underlying sin- databases have tended to apply comparatively
gle-vehicle motorcycle accident severity. By simplistic statistical approaches that have great-
developing a probabilistic model that contains ly limited the usefulness of their findings.
several important variables relating to environ- Finally, it would be interesting to conduct the
mental factors, roadway conditions, vehicle severity analysis proposed in this paper using
characteristics, and rider attributes, we have actual medical records as compared to the police
shown that possible ambiguity and bias stem- off%zerjudgments used herein. The use of such
ming from confounding effects in a partially records would permit a quantification of any
specified model (i.e., a model that does not potential biases that may result from using
include all relevant variables, which produces police officer judgments that are commonly
an omitted variables specification error) can be available in national accident databases.
eliminated. In addition, this research provides
suggestive results by its use of variables such as
helmet-sobriety interaction and helmet-fixed REFERENCES
object interaction. Specifically, it suggests that
helmeted-riding may be an effective means of Adams, J. G. (1983). Public safety legislation and the risk
reducing injury severity in some types of colli- compensation hypothesisz The example of motorcycle helmet
sions, however, the benefit of helmet use may legislation. Bnkwtmenf and Pkznning C: Governmen? and
Policy, I, 193-203.
be offset in fixed-object collisions where the Bachulis. B. L.. Sang&r, W., Gorrell. G. W., & Long, W. B.
risk of fatality was found to increase. This (1988). Patterns of injury in helmeted and nonhelmeted
increase could be due to the possibility that rid- motorcycles. American Journal of Surgery, 155,108-l 11.
ers not wearing a helmet represent a self-select- Barancik, J. I., IQ Fife, D. (1985). Discrepancies in vehicular
crash injury reporting: Northeastern Ohio trauma study
ed sample with riding behavior/experience that IV. AccidentAnalysisand Prevention, 17(2), 147-154.
makes them less prone to fatal injuries, could be Ben-Akiva, M., & Lerman S. R. (1985). Discrete choice
the outgrowth of some physiological factors analysis: Theory and applicarion to travel demand.
associated with helmeted riders hitting fixed Cambridge, MA: The MlT Press.
Chenier, T. C.. & Evans, L. (1987). Motorcyclist fatalities
objects in single-vehicle collisions, or could be and the repeal of mandatory helmet wearing laws.
the result of some form of risk compensation Acciaknr Analysis and Prevention, 18, 133-I 39.
among helmeted riders. In general, our findings Evans, L. (1986). Double pair comparison a new method to
with regard to helmet use and accident severity determine how occupant characteristics affect fatality risk in
traftic crashes. Accident Analysis and Prevenrion, 18(3),
suggest a rather complex interaction and one 217-228.
that could benefit from an analysis similar to Evans, L., & Frick, M. C. (1986). Helmet eflecriveness in
that conducted in this paper, but on a more preventing mororcycle driver and passenger fatalities
extensive, and perhaps national, database. (Research Publication GMR-5602). Warren, MI: General
Motors Research Laboratories.
Our analysis also uncovered many important Foldvary, L. A., & Lane, J. C. (1964). The effect of
relationships between accident severity and compulsory safety helmets on motorcycle accident fatalities.
motorcycle displacement, rider age, alcohol- Austmlian Road Research, 2,7-24.
impaired riding, rider ejection, speed, rider Gabella, G., Reiner, K. L., Hoffman, R. E., Cook, M., &
Stallones, L. (1995). Relationship of helmet use and head
attention, pavement surface, and type of injuries among motorcycle crash victims in El Paso
highway. The wide diversity of variables found County, Colorado, 1989-I 990. Accident Analysis and
to influence accident severity suggests a number Prevention, 27(3), 363-370.
of important directions for future motorcycle Glimm, J., & Fenton, R. E. (1980). An accident-severity
analysis of a uniform-spacing headway policy. IEEE
accident severity research. First, a study that Transactions of Vehicular Technology, 29(l), %-103.
considers multivehicle accidents should be Goldstein, J. R (1986). The effect of motorcycle helmet use
undertaken. Such a study would require a large on the probability of fatality and the severity of head and
data set because of the additional variability neck injuries. E&&ion Review, IO, 355-375.
Hutchinson, T. P. (1986). Statistical modeling of injury
introduced by the many different vehicle types severity, with special reference to rider and front seat
that could influence severity, but additional, passenger in single-vehicle crashes. Accident Analysis
important conclusions could be drawn. Second, and Prevention, 18(2), 157-167.

Fall l996/Volume 27LWumber3 193


Hurt, H. H., Ouellet, J. V., & Thorn, D. R. (1981). Shao, S. P. (1987). Estimating car rider injury severity in
Mororcycle accidenr cause factors and iden@icarion of car/tractor-trailer collisions. Accident Analysis and
countermeasures. Volume I: Technical Report (Contract Prevention, N(3). 207-217.
No. DOTHS-5-01160). Washington, DC: National Small, K. A., & Hsiao, C. (1985). Multinomial logit
Highway Traffic Safety Administration. specification tests. International Economic Review, 26.
Jones, I. S., & Whiffield, R. A. (1988). Predicting injury risk 619-627.
with “new car assessment program” crashworthiness ratings. Wrdler,P. F, Stewart J. R. Hansen A. R, Stutts J., Popkin C. L.,
AccidentAnalysk and Prevention, 20(6), 411-419. & Rodgman. E. (1985). The potentiating effects of alcohol
Lass-, S. (1986). The introduction of the variables “traftlc on driver injury. American Association of Automotive
volume,” “speed” and “‘belt-wearing” into a predictive Medicine, Proceedings of 29th Annual Meeting, Morton
model of the severity of accidents. Accidenr Analysis and Grove, IL.
Prevention. 18(2), 129-134. Watson, G. S., Zador, P. L., & Wilks. A. (1980). The repeal
Lui, K. J., McGee D., Rhodes P., & Pollock, D. (1988). An of helmet use laws and increased motorcyclist mortality in
application of a conditional logistic regression to study the the United States. American Journal of Public Health. 70,
effects of safety belts, principal impact points. and car weights 579-585.
on riders’ fatalities.Journal of Safm Research, 19.197-203. Weiss, A. A. (1992). The effects of helmet use on the severity
Luna, G. K, Maier R. V., Sowder L., Copass M. K., & of head injuries in motorcycle accidents. Journal of rhe
Oreskovich, M. R. (1984). The influence of ethanol American Sbtistical Association, 87,48-56.
intoxication on the outcome of injured mot~yclists. Journal
of Trmrmq 24.695-699.
Mannering, F. L., & Cirodsky,L. (1994). Statistical analysis of
motorcyclists’ perceived accident risk. Acciaknr Analysis NOMENCLATURE
and Prevenrion, 26(6), 2 l-3 1.
McFadden, D. (1981). Econometric models of probabilistic E elasticity.
choice. In Manski & D. McFadden (Eds.), Structural
analysis of discrete data with economerric applicarions.
I index of the set of possible accident
Cambridge, MA: The MIT Press. severities.
Murdock, M. A., & Waxman, K. (1991). Helmet use improves P,(i) probability that accident n is severity i.
outcomes after motorcycle accidents. Western Journal of P denotes probability statement.
Medicine, 155.370-372.
Nelson, D., Sklar, D., Skipper, B., & McFeeley, P. J. (1992). sin is a function of covariates that deter-
Motorcycle fatalities in New Mexico: The association of mine the likelihood of accident n being
helmet nonuse with alcoholic intoxication. Annals of of severity i.
Emergency Medicine, 21(3), 279-283. the vector of measurable characteristics
Peltzman. S. (1975). The effects of automobile safety xn
legislation. Journal of Political Economy, 83,667-725. that determine the accident severity of
Shankar, V. N., Mannering, F. L., & Barfield, W. (1995). accident n.
Effect of roadway geometries and environmental conditions Pi the vector of estimable coefficients for
on rural accident frequencies. Accident Analysis and accident severity i.
Prevention, 27(3), 371-390.
Shankar, V. N., Mannering, F. L., & Barfield, W. (in press). Ein an error term that accounts for unob-
Statistical analysis of accident severity on rural freeways. served factors influencing accident n
Accident Analysis and Prevention. with severity i.

194 Journal of Safety Research

You might also like