Transcript Nested Logit slides

Probabilistic Models of
Motorcyclists' Injury Severities in
Single- and Multi-vehicle Crashes
Peter T. Savolainen, Ph.D.
Wayne State University
Fred Mannering, Ph.D.
Purdue University
Overview
 Background
 Research Objectives
 Methodology
 Multi-Vehicle Crash Severity Model
 Single-Vehicle Crash Severity Model
 Conclusions
Background
Motorcycle Cra she s by Ye a r
4,000
Crashes
3,500
3,000
2,500
2,000
1986
1987
1988
1989
1990
1991
1992
1993
1994
1995
Ye a r
Motorcycle Fa ta litie s by Ye a r
140
Fatalities
120
100
80
60
40
1986
1987
1988
1989
1990
1991
Ye a r
1992
1993
1994
1995
Background
2,900
2,700
2,500
2,300
2,100
1,900
1,700
1,500
1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005
Ye ar
Motorcycle Fatalities by Year
120
100
Fatalities
Crashes
Motorcycle Cra she s by Yea r
80
60
40
20
0
1995 1996 1997 1998 1999 2000 2001 2002 2003 2004 2005
Year
Background
 Ridership increasing
 Riding population
changing


Training
Age
Gender
Motorcycle Registrations per Year
150,000
140,000
Registrations

130,000
120,000
110,000
 Bike design
100,000
90,000
 Speed
80,000
1995
 Power
 Safety
 Repealed helmet laws
1997
1999
2001
Year
2003
2005
Research Objective
 To develop probabilistic models of motorcycle
crash-injury severity using Indiana crash data
from 2003 to 2005


Single-vehicle
Multi-vehicle
Single- vs. Multi-Vehicle
Single-Vehicle Multi-Vehicle
Age
38.5
38.3
Female
8%
9%
Helmet Usage
34%
41%
Completed Training Program
6%
7%
Passenger
15%
17%
Bike Age
10.1
7.7
Alcohol Use
2%
5%
Wet Pavement
4%
3%
Percent of Total Crashes
Crashes by Month
20%
18%
16%
14%
12%
10%
8%
6%
4%
2%
0%
JAN FEB MAR APR MAY JUN JUL AUG SEP OCT NOV DEC
Month
Methodology – Multi-Vehicle
 Multinomial logit (MNL) model with,
Sin = βi Xin + εin
 Sin is the function that determines the probability of
severity i for crash n,



Xin is a vector of measurable characteristics
(motorcyclist and roadway characteristics) that
determine the severity level for crash n,
βi is a vector of estimable coefficients, and
εin is an error term accounting for unobserved effects
influencing the injury severity of crash n
Methodology – Multi-Vehicle
 if εin are assumed to be extreme value distributed
(see McFadden, 1981), then a standard multinomial
logit model results,
Pn  i  
EXP  βi X in 
 EXP  β X 
I
I
In
 where Pn(i) is the probability that crash n will result in
severity i and I is the set of possible injury severity
levels (PDO/Possible, Non-incapacitating,
Incapacitating, Fatal).
Methodology – Multi-Vehicle
 Elasticity
E xPnniki 
Pni xnik

 1  P i  ik xik
xnik Pni
 Pseudo-elasticity
 xP 
ni
nik
exp  X ni i  exp xnik i 
I
exp  X ni i  exp xnik i    exp  X ni i 
I In
I In
1
Multi-Vehicle Crash Severity Model
Injury Severity
No Evident Injury
(PDO or Possible)
Non-Incapacitating
Injury
Incapacitating
Injury
Fatal
Injury
Some Multi-Vehicle Crash Severity
Model Findings
Severity level; No injury:
 Factors decreasing no-injury likelihood:



Alcohol use (other motorist) (65%)
Head-on collision (35%)
Motorcycle age
Some Multi-Vehicle Crash Severity
Model Findings
Severity level; Incapacitating injury:
 Factors increasing incapacitating-injury
likelihood:




Motorcyclist speeding (50%)
Motorcyclist age (4.2% per 1% increase in age)
Vertical curve (81%)
Horizontal curve (45%)
Some Multi-Vehicle Crash Severity
Model Findings
Severity level; Fatality:
 Factors increasing fatality likelihood:



Motorcyclist at fault (126%)
Motorcyclist speeding (116%)
Head on collision (566%)
 Factors decreasing fatality likelihood:

Helmet use (right angle) (61%)
Methodology – Single Vehicle
 If εin are correlated (crash severity levels share unobserved
effects):
Pn  i   EXP  i X in  i Lin  /  EXP  i X in  i LSin 
I
Pn  j | i   EXP   j|i X n  /  EXP   j|i X Jn 
J


LSin  LN  EXP   J |i X Jn 
 J

 where Pn(ji) is the probability of crash n resulting in injury
severity j conditioned on the injury severity being in injuryseverity category i, J is the conditional set of outcomes
(conditioned on i), I is the unconditional set of outcome
categories, LSin is the inclusive value (logsum), and i is an
estimable parameter.
Single-Vehicle Crash Severity Model
Injury Severity
PDO or
Minor Injury
No Evident Injury
(PDO or Possible)
Incapacitating
Injury
Non-Incapacitating
Injury
Fatal
Injury
Some Single-Vehicle Crash Severity
Model Findings
Severity level; Minor or No injury:
 Factors increasing minor/no-injury likelihood:


Motorcycle less than 5 years old (20%)
Helmet used (50%)
 Factors decreasing minor/no-injury likelihood:




Motorcyclist age (1.15% per 1% increase in age)
Alcohol use (10%)
Speeding (14%)
Collisions with trees, poles, curbs, culverts, guardrails)
Some Single-Vehicle Crash Severity
Model Findings
Severity level; Fatality:
 Factors increasing fatality likelihood:





Over 2 years since took BRC (171%)
Speeding (212%)
Run-off-road (137%)
Collision with tree (525%)
Collision with pole (344%)
Conclusions
 Critical areas

Poor visibility



Unsafe speed
Risk-taking behavior


alcohol use, not wearing a helmet
Collision type


horizontal curvature, vertical curvature, darkness
Right-angle, head-on, and collisions with fixed objects
Age
Conclusions
 Critical areas (continued)

Rider training (BRC results)

Degradation in skills, self-selectivity, risk
compensation?
 Encouragingly, crashes were found to be less
severe:
 Under wet pavement conditions
 Near intersections
 When passengers were on the motorcycle
Additional Evidence on the
Effectiveness of Motorcycle Training
and Motorcyclists’ Risk-taking
Behavior
Peter T. Savolainen, Ph.D.
Wayne State University
Fred Mannering, Ph.D.
Purdue University
Overview
 Background
 Research Objectives
 Methodology
 Crash Propensity Model
 Top Travel Speed Model
 Helmet Usage Model
 Conclusions
Background
 Rider education and training critical to
motorcycle safety agenda
 Limited research on education/training
programs
 Contradictory results
 Methodological shortcomings
Background
 Methodological shortcomings:

Lack of consideration of variables beyond
violation and crash statistics

Lack of control for exposure

Not fully considering dissimilarity between
trained/untrained riders

Not considering possible risk compensation
Research Objectives
 To provide additional evidence on
effectiveness of motorcycle training courses

Motorcyclist survey
 Using 2005 sample of Indiana motorcyclists
Motorcyclist Survey
 Survey developed to collect data on:
 Demographics
 Training history
 Riding behavior
 Crash involvement
 2 groups of riders
 Trained: ABATE of Indiana – MSF Basic Rider Course
(BRC)
 Untrained: Indiana BMV and ABATE newsletter
 Surveys mailed to 4,000 riders from each group
 Over 1,300 responses obtained
Motorcyclist Survey
 Why ABATE?
 Why combine samples?
 Not statistically different. Proof: likelihood ratio test
X 2  2LL R   LLU 
 LL(βR) = log-likelihood of restricted model
 e.g., BMV only sample
 LL(βU) = log-likelihood of unrestricted model
 e.g., BMV and ABATE sample
 Combining allows more precise parameter estimates
Summary Statistics
 Average age
47.8
 84% male, 16% female
 Completed BRC
 Multiple times
60%
 Completed ERC
 ABATE members
12%
46%
 Annual exposure
 <1000
 1000-5000
 Over 5000
23%
51%
26%
6%
Summary Statistics
 Type of Motorcycle




Sportbike:
Cruiser:
Touring:
Other:
15%
46%
27%
12%
Summary Statistics
 Reasons for not taking BRC




No need to take course:
Could not find time:
Unaware of course:
Could not afford program cost:
47%
34%
15%
4%
Summary Statistics
 Helmet usage frequency



Always/Usually:
Sometimes:
Rarely/Never:
56%
21%
23%
Methodology
 Multinomial logit models developed with,
Rin = βi Xin + εin




Rin is the function that determines the probability of
response i being chosen by motorcyclist n,
Xin is a vector of measurable characteristics
(socioeconomics and rider perceptions) that
determine the response of motorcyclist n,
βi is a vector of estimable coefficients, and
εin is an error term accounting for unobserved effects
influencing the response of motorcyclist n
Methodology
 if εin are assumed to be extreme value
distributed (see McFadden, 1981), then a
standard multinomial logit model results,
Pn  i  
EXP  βi X in 
 EXP  β X 
I
I
In
 where Pn(i) is the probability that motorcyclist
n will choose response i and I is the set of
possible survey responses.
Crash Propensity Model
Cni  i X ni   ni
 Cni is a function that
determines crash
propensity
 Xni is a vector of rider
characteristics
eCni
Pni 
1  eCni
Crash Involvement
0 crashes
1+ crashes
Crash Propensity Model
Crash propensity increases with:







Not wearing helmet (63%)
Ride over 100 mph in past 12 mo. (161%)
Sportbike (54%)
Ride over 10,000 mi/yr (102%)
Age under 35 (59%)
Completed BRC once (44%)
Completed BRC more than once (180%)
Crash Propensity Model
Crash propensity decreases with:


Citing no need for BRC (51%)
Riding experience




Highest during 1st year
Decreases years 2-4 (58%)
Increases slightly year 5+
Riding 500-1000mi/yr (64%)
Crash Propensity Model
Note on BRC findings:



Completed BRC once (increases crash 44%)
Completed BRC more than once
(increases crash 180%)
Cited no need for BRC (decreases crash 51%)
 Evidence that BRC riders may be a self-
selected group of inherently less-skillful riders
Maximum Speed Model
 Binary logit model for
MS ni   i X ni   ni
maximum speed
 MSni is a function that
determines maximum
travel speed
 Xni is a vector of rider
characteristics
e MSni
Pni 
1  e MSni
Maximum
Travel Speed
Less than 90 mph
90 mph or faster
Maximum Speed Model
 Increasing probability of riding over 90 mph:








Motorcycle primary mode of travel (42%)
Usually wear a helmet (39%)
Sportbike riders (128%)
Drank alcohol within 2 hrs of riding (66%)
Licensed at 40+ years old (38%)
Ride 5-10K miles per year (106%)
Ride over 10K miles per year (189%)
Involved in crash/near-miss (30%)
Maximum Speed Model
 Decreasing probability of riding over 90 mph:




Rider age (1.82% per 1% increase in age)
Female riders (61%)
Smaller engines
Usually wear protective clothing/equipment
(30%)
Helmet Usage Model
H ni  i X ni   ni
Helmet Usage
Always/Usually
Sometimes
Rarely/Never
e H ni
Pni 
H nj
e

j
 Multinomial Logit for
helmet usage


Xni rider characteristics
Hni is equal to:
 1 : Always/Usually
 2: Sometimes
 3: Rarely/Never
Helmet Usage Model
 Always wear helmet:





Typically wear other protective equipment or
reflective clothing/equipment
Typical travel speed over 70 mph
Typical travel speed less than 60 mph
Older riders
Number of bikes owned
Helmet Usage Model
 Never wear helmet:
 Motorcycle primary mode of travel
 Never wear other protective equipment
 Larger engine displacement
 Rode over 100 mph in past year
 Involved in near-miss in past year
 Drank alcohol within 2 hrs of riding
 Females
 ABATE members
 Except for those completing BRC
 Self-rated as excellent rider
Conclusions
 Individuals taking BRC are more likely to be crash-
involved
 Inherently less capable riders?
 Overcompensation of risks with learned material?
 Skill-measurement methods must be developed and
research undertaken to understand how skills can be
improved considering:
 Risk compensation
 Self selection of less skilled rider to training
courses
Future Research Directions
 Improvement of crash records system
 Further research into rider training, self-selectivity
into training courses, and risk compensation induced
by course-taught material
 Improvements to Rider Training Program
 Baseline evaluation
 Further application of survey methodology
 Regional/national level
 Focus on other issues