Você está na página 1de 10

Transportation Research Record: Journal of the Transportation Research Board,

No. 1897, TRB, National Research Council, Washington, D.C., 2004, pp. 1827.
18
A study was done to develop macrolevel crash prediction models that can
be used to understand and identify effective countermeasures for improv-
ing signalized highway intersections and multilane stop-controlled high-
way intersections in rural areas. Poisson and negative binomial regression
models were t to intersection crash data from Georgia, California, and
Michigan. To assess the suitability of the models, several goodness-of-t
measures were computed. The statistical models were then used to shed
light on the relationships between crash occurrence and traffic and geo-
metric features of the rural signalized intersections. The results revealed
that traffic ow variables signicantly affected the overall safety perfor-
mance of the intersections regardless of intersection type and that the geo-
metric features of intersections varied across intersection type and also
inuenced crash type.
Transportation is vital to both economic success and quality of life
in urban and rural areas. However, the rapid growth of city popula-
tions and corresponding vehicle miles of travel, commerce, and
transportation infrastructure has generated such negative effects as
congestion, deterioration of air quality, noise, and motor vehicle
crashes. According to NHTSA, there were 6,322,795 traffic crashes
in 2001 (1). A fatal crash occurred every 13 min on average. These
statistics emphasize why traffic safety remains a major concern of the
traveling public and why traffic safety improvements are needed.
To improve traffic safety, each element of the roadway trans-
portation system should be examined. Representing conict points
in the road network, intersections have received and should continue
to receive considerable attention, since intersections continue to rep-
resent crash-prone locations on a transportation network. When
approaching intersections, drivers are confronted with a complex
driving task that includes observing and responding to traffic control
devices, reacting to these devices by stopping, reducing speed, or
proceeding without delay, executing turns, observing and reacting to
pedestrians and bicyclists, and avoiding conicts with other vehicles.
Adding further complexity are drivers that disobey traffic control
devices, most notably signicant red-light-running violations in many
states. The complex vehicle movements at intersections lead to traf-
c conicts. A subset of conicts lead to crashes. Thus, intersection
safety is a high priority in local, regional, and state traffic agencies
and among transportation researchers and academicians.
Considerable research has concentrated on identifying the safety
effects of accident countermeasures. Inconsistency of ndings across
studies, however, presents difficulties in understanding and esti-
mating the safety effects of countermeasures. Furthermore, little
research has been done to identify and understand factors that con-
tribute to accidents on multilane stop-controlled and signalized
highway intersections in rural areas. Several accident prediction
models (24) have provided limited knowledge on the safety effects
of traffic ow and geometric designrelated countermeasures for
rural multilane stop-controlled and signalized intersections. These
crash prediction models, furthermore, may be inaccurate across
jurisdictions because the estimation data were limited in number of
explanatory variables, geographical diversity of the intersections,
and overall sample size.
This research presents the estimation results of statistical crash
prediction models based on larger sample sizes, greater geographi-
cal diversity, and a larger set of explanatory variables than in previ-
ous research efforts. Statistical models are discussed for the following
three intersection types in rural areas:
Three-legged intersections with major-road four-lane and
minor-leg two-lane stop-controlled intersections,
Four-legged intersections with major-road four-lane and
minor-leg two-lane stop-controlled intersections, and
Signalized intersections with both major- and minor-road two-
lane intersections.
The accident prediction models were calibrated for total accidents
and injury (fatal and nonfatal injury) accidents within 250 ft of the
intersection center and were used to identify traffic ow and design
countermeasure safety effects for these rural intersections.
DATA DESCRIPTION
The data for the study were based on a total of 136 three-legged and
124 four-legged stop-controlled multilane and 100 four-legged sig-
nalized highway intersections in rural areas. The data used for model
calibration were obtained from two sources. The calibration data were
obtained from the California and Michigan Highway Safety Informa-
tion System. The crash and other data were available from 1993 to
1998. The second source of data was the state of Georgia. Four data
sources were used to study Georgia intersections: crash les, roadway
characteristic (RC) les, aerial photographs, and geographic infor-
mation system (GIS) maps. Crash and RC les were available for
1996 and 1997. RC les provided detailed information on road char-
acteristics. Digital orthophotography quarter-quadrangles (DOQQs)
aerial photos were used from 1994 and 2000 to extract information
about intersection angle and degree of horizontal curvature of selected
intersections by overlapping with GIS roadmaps. Table 1 summarizes
the sources of data used for the models.
Development of Accident Prediction
Models for Rural Highway Intersections
Jutaek Oh, Simon Washington, and Keechoo Choi
J. Oh, Korea Transportation Institute, Ilsan 411-701, Republic of Korea. S. Wash-
ington, Department of Civil Engineering, University of Arizona, Tucson, AZ 85721-
0072. K. Choi, Department of Transportation Engineering, Ajou University, Suwon
442-749, Republic of Korea.
Oh, Washington, and Choi 19
Number of Sites
Number of
Total/Injury Accidents
State
Years of
Data
3-
Legged
4-
Legged Signalized
3-
Legged
4-
Legged Signalized
California 1991-1998 60 54 18 427/196 478/268 507/200
Michigan 1993-1997 24 18 31 248/63 277/92 1262/159
Georgia 1996-1997 52 52 51 124/56 222/104 489/118
TABLE 1 Data Sources
On the basis of underlying theories of crash causation and to
establish a suitable statistical model that enabled the examination of
possible relationships among accident frequencies, geometric, and
traffic characteristics of intersections, 53 possible explanatory vari-
ables were considered. Traffic volumes, including left-, right-, and
U-turn volumes, were estimated from annual average daily traffic
(AADT). Geometric elements included vertical and horizontal
curves, sight distances, hazard ratings around the intersection, ter-
rain, presence of exclusive left- and right-turn lanes, lane and shoul-
der widths, median types and widths, driveway intensity, and signal
control types. Driving characteristics, such as peak period left-turn,
right-turn, and through traffic percentages and peak period truck
percentages, were also included.
Table 2 describes the variables available in the analysis, their
units, and the abbreviation used in model estimation results. Tables
3, 4, and 5 give the summary statistics of the variables discussed in
this paper. Summary statistics of all variables considered for modeling
are available elsewhere (5).
MODEL CALIBRATION METHODS
This section briey presents the statistical methods used for cali-
brating the crash prediction models. The basic statistical background
and general model forms of Poisson and negative binomial regres-
sion models are provided. Additional information on count models
is available elsewhere (6).
Poisson Model
Since the occurrences of accidents at intersections are relatively rare
discrete events, the Poisson model is a natural choice. On the assump-
tion that the number of accidents, Y

i
, follows a Poisson distribution, the
general form of the expected number of accidents occurring at the ith
intersection with mintersection parameters, X
i1
, X
i2
, X
i3
, . . . , X
im
, is
where
j
are estimated regression coefficients.
In the case of a Poisson regression model, the model coefficients
are estimated by using the maximum likelihood method (7). The
likelihood function for the Poisson regression model is given by
The maximum possible value of the likelihood function for a
given data set occurs if the model ts the data exactly. This occurs
if
i
is replaced by y
i
, as shown in the likelihood equation
L f Y
X X
Y
i i
i
n
i
Y
i
i i
n i


( ) ( )
( ) [ ] ( ) [ ]


1 1
, exp ,
!

exp exp Y X X X X X
i m m j j
j
m
+ + + + ( )

_
,


0 0 1 1 2 2
0
L
In practice, crashes do not t a Poisson process exactly because
(a) the exact relationships between explanatory variables and crashes
are unknown and approximated, (b) crashes are stochastic, (c) some
insignificant but real explanatory variables may be omitted, and
(d) some signicant explanatory variables may be omitted (but it is
hoped not). For these reasons, a maximum value of the likelihood
function is sought.
Negative Binomial Model
The negative binomial distribution can be used to relax the limita-
tion of the Poisson distribution requirement that the mean equals the
variance of the distribution. In Poisson regression models used for
modeling crash occurrence, if the variance of the data exceeds the
estimated mean of the data, then the data are said to be overdispersed.
Overdispersion occurs in practice because the crash means vary
across segments (road, intersections, ramps, etc.) as a consequence
of unobserved heterogeneity. In these cases, the negative binomial
provides an appealing alternative for dealing with overdispersion
in crash data. Similar to the Poisson model, the negative binomial
regression model relates the expected number of accidents occur-
ring at the ith element with q explanatory variables, X
i1
, X
i2
, . . . , X
iq
,
as follows:
However, the negative binomial regression model results in a
quadratic term added to the variance repressing overdispersion. The
negative binomial model takes the form
where K is the overdispersion parameter and the variance is given
by
As a result, the negative binomial regression model accommo-
dates extra-Poisson variation because of unobserved heterogeneity.
If overdispersion, K, equals 0, the negative binomial model reduces
to the Poisson model. The larger the value of K, the more variabil-
ity there is in the data beyond that associated with the mean
i
. For
the Poisson model, the coefficients
j
are estimated by maximizing
the log-likelihood log
e
L(). For the negative binomial distribu-
tion, the estimated coefficient vector and y
i
, along with an estimate
K

for K, are obtained by maximizing log


e
L(, K):
Var y E y KE y
i i i
( ) [ ] + [ ]
2
P y
y K
y K
i
i
i
y
y K
i
i
( )
+ ( )

( )
+ ( )
+
1
1 1
!
! !

Function
i i i q iq
X X X ( ) + + + +
0 1 1 2 2
L
( )

+ ( )

log
e
i i i
i
L
y X

0
20 Transportation Research Record 1897
Variable
Abbreviation
Variable Description
AADT1 Annual average daily traffic on major road (vehicles per day)
AADT2 Annual average daily traffic on minor road (vehicles per day)
COMDRWY1 Commercial driveways on major road within 250 feet of the
intersection center
HAU Intersection angle variable defined where the angle between the
major and minor roads is measured from the far side of the minor
road:

3-Legged Intersections
I angle-90 if minor road is to the right of the major road in the
increasing direction
I 90-angle if minor road is to the left of the major road in the
increasing direction
4-Legged Intersections
I (right angle - left angle)/2
HAZRAT1 Roadside hazard rating on major road within 250 feet of the
intersection (from 1, least hazardous case, to 7, most hazardous case)
HI Sum of degree of curve in degrees per hundred feet of each horizontal
curve on major road any portion of which is within 250 feet of the
intersection center divided by the number of such curves
HEI1 Sum of degree of curve in degrees per hundred feet of each horizontal
curve on major road within 800 feet of the intersection center divided
by the number of such curves
HEI2 Sum of degree of curve in degrees per hundred feet of each horizontal
curve on minor road within 800 feet of the intersection center divided
by the number of such curves
HEICOM (1 / 2) (HEI1 + HEI2)
LIGHT Light at intersection (0 = no, 1 = yes)
MEDTYPE Median type on major road (0 = no median, 1 = painted, 2 = curbed, 3
= others)
MEDWDTH1 Median width on major road (feet)
PKLEFT Left-turn percentage during the peak hour (%)
PKLEFT1 Left-turn percentage on major road during the peak hour (%)
PKLEFT2 Left-turn percentage on minor road during the peak hour (%)
PKTHRU2 Through percentage on minor road during the peak hour (%)
PKTRUCK Truck percentage passing through the intersection during the peak
hour (%)
PKTURN Peak turning percentage (%)
SDR2 Right-side sight distance on minor road (feet)
SPD1 The average posted speed on major road in vicinity of the intersection
(mph)
SPD2 The average posted speed on minor road in vicinity of the intersection
(mph)
VEI1 Sum of absolute change of grade in percent per hundred feet for each
curve on major road any portion of which is within 800 feet of the
intersection center, divided by the number of such curves
VEI2 Sum of absolute change of grade in percent per hundred feet for each
curve on minor road any portion of which is within 800 feet of the
intersection center, divided by the number of such curves
VEICOM (1 / 2) (VEI1 + VEI2)
TABLE 2 Analysis Variables
Model Validation Methods
Note that only through the assessment of several goodness-of-t cri-
teria can an objective statement be made about the performance of
log ,
log log log
log log !
e
e j
j
y
e i i e i
i i i
i
L K
K Ky y
y
K
K y
i

( )
+ ( )
( )
+ ( ) +
+

_
,
+ ( ) ( )

1
]
1
1
1

1 1
1
1
0
a particular model or set of models. The goodness-of-t measures
used in this research are Pearson product-moment correlation co-
efficients, mean prediction bias (MPB), mean absolute deviation
(MAD), and overdispersion.
Pearson Product-Moment Correlation Coefficients
The Pearson product-moment correlation coefficient, usually denoted
r, is an example of a correlation coefficient. It is a measure of the
linear association between two variables Y
1
and Y
2
that have been
measured on interval or ratio scales and is given by
Variables Frequency Mean Median Minimum Maximum
TOTAL accidents per year 136 1.35 0.80 0.00 10.60
INJURY accidents per
year 136 0.55 0.33 0.00 4.00
AADT1 136 13011 12100 2360 33333
AADT2 136 709 430 15 9490
MEDTYPE on major road
No median
Painted
Curbed
Other
136
69(50.7%)
45(33.1%)
14(10.3%)
8(5.9%)
MEDWIDTH1 136 12.6 6 63
HAU 136 1.3 0
0
-65 90
HAZRAT1
1
2
3
4
5
6
7
136
16(11.8%)
58(42.6%)
26(19.1%)
25(18.4%)
8(5.9%)
2(1.5%)
1(0.7%)
COMDRWY1 136 1.5 0 0 14
SPD1 136 52.5 55 30 65
LIGHT
0
1
136
97(71.3%)
39(28.7%)
HEI1 136 2.01 0.73 0 26.63
VEI1 136 0.9 0.6 0.0 6.7
PKTRUCK 84 9.15 7.79 1.18 28.16
PKTURN 84 6.68 4.28 0.27 53.09
PKLEFT 84 3.28 2.16 0.13 25.97
PKLEFT2 84 55.31 60.29 0.00 100.00
SDR2 136 1428 1555 80 2000
TABLE 3 Summary Statistics for Three-Legged Stop-Controlled
Multilane Intersections
Variables Frequency Mean Median Minimum Maximum
TOTAL accidents per year 124 2.0 1.4 0.0 10.8
INJURY accidents per year 124 0.9 0.5 0.0 5.7
AADT1 124 12881 11496 3150 73799
AADT2 124 621 430 21 2990
MEDTYPE on major road
0: No median
1: Painted
2: Curbed
3: Other
124
70(56.5%)
27(21.8%)
22(17.7%)
5(4.0%)



MEDWDTH1 124 16.1 6.5 0 60
HAZRAT1
1
2
3
4
5
6
124
24(19.4%)
43(34.7%)
32(25.8%)
21(16.9%)
2(1.6%)
2(1.6%)
COMDRWY1 124 0.6 0 0 12
LIGHT 124
0
1
87(70.2%)
37(29.8%)
VEI1 124 0.87 0.35 0.00 12.50
HEI 124 3.28 0.60 0.00 233.33
HAU 124 1.5 0 -50 55
SPD1 124 55.6 55 25 65
SPD2 124 34.7 35 25 55
PKTRUCK 72 10.95 8.36 1.75 37.25
PKTURN 72 9.47 6.56 0.00 48.52
PKLEFT 72 4.80 3.08 0.00 25.26
PKTHRU2 72 15.69 10.82 0.00 68.09
PKLEFT2 72 38.89 36.66 0.00 100.00
SDR2 124 1329 1354 215 2000
TABLE 4 Summary Statistics for Four-Legged Stop-Controlled
Multilane Intersections
where Y

is the mean of the Y


i
observations.
MPB
The MPB statistic provides a measure of the magnitude and direction
of the average model bias compared to validation data. The smaller
the average prediction bias the better the model is for predicting
observed data:
where n is the validation data sample size and Y

is the tted value Y.


MAD
MAD provides a measure of the average misprediction of the
model. It differs from MPB in that positive and negative predic-
tion errors will not cancel each other. A value close to zero sug-
gests that on average the model predicts observed data well. It is
given by
MPB
( )

Y Y
n
i i
i
n

1
r
Y Y Y Y
Y Y Y Y
i i
i i
12
1 1 2 2
1 1
2
2 2
2
1 2

( ) ( )
( ) ( )
[ ]


22 Transportation Research Record 1897
Overdispersion (K)
Overdispersion in a Poisson model causes underestimation of the
variance of the model coefficients. This results in overstating the
signicance of the coefficients. The deviance of the model contain-
ing all the parameters (including the intercept) divided by its degrees
of freedom, n-p, provides a test for over- or underdispersion and a
measure of t of the model. Evidence of underdispersion or over-
dispersion indicates inadequate t of the Poisson model. For the
Poisson regression model, the model deviance is as follows:
MODEL CALIBRATION AND INTERPRETATION
The crash prediction models were estimated by using LIMDEP
econometric software. The models include many theoretically appeal-
ing variables to explain as much variation in crash occurrence as pos-
sible, given knowledge of crash causation and the available set of
DEV . . log X X X Y
Y
Y
p i e
i
n
i
i
i i
i
n
0 1 1
1 1
2 , , . ,


( )

_
,
( )

1
]
1


MAD

Y Y
n
i i
i
n

1
Variables Frequency Mean Median Minimum Maximum
TOTAL accidents per year 100 5.9 5.3 0.0 26.5
INJURY accidents per year 100 1.8 1.5 0.0 6.5
AADT1 100 9126 8700 430 25132
AADT2 100 3544 3100 420 12478
MEDTYPE on major road
0:No median
1:Painted
2:Other
100
87(87%)
12(12%)
1(1%)
MEDWDTH1 100 1.3 0 0 13
HAZRAT1
1
2
3
4
5
6
7
100
12(12%)
29(29%)
27(27%)
16(16%)
13(13%)
3(3%)
0(0%)
COMDRWY1 100 2.64 2 0 11
LIGHT
0
1
100
29(29%)
71(715)
SDR1 51 822 798 103 2000
VEI1 100 1.45 1.19 0.00 11.97
HEI 100 3.95 0.61 0.00 94.87
HEICOM 100 2.56 0.58 0.00 32.54
HAU 100 0.07 0.00 -45.00 40.00
SPD1 100 45.2 45 25 65
SPD2 100 40.9 40 20 55
PKTRUK 49 8.96 7.71 2.69 45.43
PKTURN 49 35.64 34.48 7.07 72.66
PKTHRU2 49 43.90 41.99 8.45 84.09
PKLEFT 49 18.17 17.97 4.20 37.07
PKLEFT2 49 28.21 24.88 2.59 75.73
TABLE 5 Summary Statistics for Signalized Intersections
potential explanatory variables. The best models discussed in the fol-
lowing section were selected by comparing several candidate models.
When all candidate models were compared, the goodness-of-t mea-
sures discussed previously were used, in addition to inspection and
theoretical appeal of model coefficients and their associated statistical
signicance. A detailed discussion of these comparison and selection
activities was provided by Washington et al. (5).
During model selection an alpha (probability of a Type I error)
equal to 0.10 was used. Model results are presented in Tables 6 and 7.
The model ndings agree with results of previous studies and also
shed new light on the effect of countermeasures on intersection
safety. The following sections describe the implications of the
modeling results for each of the explanatory variables.
Traffic Flow and Safety
In agreement with numerous studies, traffic volumes are the most
stable and reliable predictor of crash occurrence at intersections.
Results showed that crash frequency increased with higher log
(AADT) on major and minor roads for the three intersection and two
accident types investigated. Coefficients relating crash frequencies
to the major-road AADT ranged between 0.5 and 0.9 and reect the
nonlinear relationship between traffic volume and crashes with the
log transformation. The estimated coefficients are between 0.2 and
Oh, Washington, and Choi 23
0.3 for minor-road AADT, reecting the relative inuence of major
and minor-road AADT on crashes at intersections. The results
reveal that the turning volume percentages are signicant for three-
legged, or T, intersections for injury crashes.
Driveways and Safety
The number of commercial driveways on the major within 250 ft of
the intersection is associated with higher crash frequencies for three-
legged stop-controlled multilane intersections and signalized inter-
sections. To interpret the safety effect of commercial driveway
density, it is important to understand what types of accidents occur
at commercial driveways. Numerous research studies have been
conducted into the nature of traffic accidents that occur near drive-
ways. Although the results of previous studies vary considerably
(24), a common conclusion is that entering and exiting turning
vehicles are involved in most driveway-related crashes. Thus, it is
the presence of additional conict points near intersections that is
associated with crashes. The modeling results revealed that three-
legged stop-controlled multilane intersections and signalized inter-
sections had significant commercial driveway effects, whereas
four-legged stop-controlled multilane intersections did not. A pos-
sible reason for the difference is the high correlation between peak
turning movements and commercial driveway density. For four-
Variables
3-legged
Intersection
(s.e., p-value)
4-legged
Intersection
(s.e., p-value)
Signalized
Intersection
(s.e., p-value)
Intercept
-10.1914
(1.5232,0.0000)
-7.4713
(1.8930,0.0001)
-5.1527
(1.8653,0.0057)
LOG of AADT1
0.8877
(0.1666,0.0000)
0.7350
(0.1849,0.0001)
0.4499
(0.1968,0.0223)
LOG of AADT2
0.3228
(0.0585,0.0000)
0.2390
(0.0926,0.0099)
0.2699
(0.0767,0.0004)
COMDRWY1
0.0681
(0.0281,0.0154)
0.0539
(0.0304,0.0757)
VEI1
0.1081
(0.0556,0.0519)
HAU
0.0101
(0.0059,0.0861)
MEDWDTH1
-0.0106
(0.0060,0.0760)
MEDTYPE1
-0.3209
(0.1771,0.0700)
SDR2
-0.0003
(0.0001,0.0403)
PKTRUCK
-0.0479
(0.0110,0.0000)
PKLEFT
0.0229
(0.0118,0.0525)
SPD1
0.0177
(0.0090,0.0482)
HEICOM
-0.0288
(0.0153,0.0597)
LIGHT
-0.2938
(0.1837,0.1098)
Overdispersion, K
0.4229
(0.1064,0.0001)
0.4001
(0.0958,0.0000)
0.4019
(0.0765,0.0000)
Pearson product-moment
correlation coefficients 0.70 0.77 0.77
MPB/year 0.09 0.12 -0.02
MAD/year 0.84 1.16 2.90
TABLE 6 Parameter Estimates for Total Accident Main Models
legged stop-controlled multilane intersections and signalized inter-
sections, correlations between peak turning movements and com-
mercial driveway density were statistically signicant, whereas there
was no signicant correlation for three-legged stop-controlled multi-
lane intersections. This indicates that peak turning movements have
greater explanatory power than commercial driveway density for four-
legged stop-controlled multilane intersections, and vice versa for
signalized intersections. Another possible reason is the observed
negative correlation between driveway density and channelization
at intersections. Compared to three-legged stop-controlled multilane
intersections (64% of the sites had left-turn lanes onto a major road),
four-legged stop-controlled multilane intersections (77% of the sites
had left-turn lanes onto a major road) were relatively well channel-
ized, which suggests that high access management decreases poten-
tial conict points near intersections, which in turn decreases left-turn
crash frequencies. Since access management involves managing
traffic movements into and out of commercial driveways, these nd-
ings support the notion that effective access management can improve
safety near intersections.
Presence of Trucks and Safety
The modeling results suggest that the inuence of truck volume per-
centages on safety is different for stop-controlled intersections and
24 Transportation Research Record 1897
signalized intersections. For three- and four-legged stop-controlled
multilane intersections, higher peak truck percentages are associated
with lower frequency of crashes, and the opposite is true for signal-
ized intersections. A possible explanation for this lies in the operation
of stop-controlled versus signalized intersections. At stop-controlled
intersections, trucks enter an intersection while vehicles on alterna-
tive approaches wait until the intersection is clear before proceeding.
Truck drivers are spared gap-acceptance maneuvers at stop-controlled
intersections and simply wait their turn. At signalized intersections,
truck drivers are routinely forced to determine an acceptable time
gap in traffic to execute turns, which results in potential conicts
with oncoming traffic. In addition, commercial trucks can take longer
to execute a maneuver (e.g., a left turn) than the clearance time pro-
vided by yellow and all-red phases, which results in potential conicts
with perpendicular traffic.
For both three- and four-legged stop-controlled intersections,
the injury accident models included the peak truck percentage
variable, whereas the variable was signicant only in the total acci-
dent model for four-legged stop-controlled intersections. Peak truck
percentage is negatively correlated with traffic flow on three- and
four-legged stop-controlled multilane intersections, suggesting per-
haps that truck drivers in general seek rural routes with low traffic
volumes on average.
Finally, the model coefficient of the peak truck percentage vari-
able for the total accident model is lower than the same coefficient
Variables
3-legged
Intersection
(s.e., p-value)
4-legged
Intersection
(s.e., p-value)
Signalized
Intersection
(s.e., p-value)
Intercept
-10.6443
(2.0474,0.0000)
-7.3927
(2.1279,0.0005)
-9.0707
(1.9064,0.0000)
LOG of AADT1
0.8498
(0.2097,0.0001)
0.5008
(0.2186,0.0220)
0.6697
(0.1899,0.0004)
LOG of AADT2
0.2188
(0.0949,0.0212)
0.3027
(0.1341,0.0240)
0.2509
(0.0929,0.0069)
COMDRWY1
0.0627
(0.0353,0.0756)
HAZRAT1
0.1889
(0.0923,0.0407)
HAU
0.0163
(0.0053,0.0021)
PKTRUCK
-0.0253
(0.0135,0.0605)
-0.0520
(0.0127,0.0000)
PKTURN
0.0254
(0.0135,0.0592)
PKLEFT1
0.0523
(0.0128,0.0000)
SPD1
0.0397
(0.0093,0.0000)
SPD2
0.0289
(0.0145,0.0465)
HEI2
-0.0284
(0.0126,0.0244)
LIGHT
-0.3985
(0.1702,0.0192)
Overdispersion, K
0.5102
(0.1426,0.0003)
0.4671
(0.1296,0.0003)
0.2360
(0.0958,0.0138)
Pearson product-
moment correlation
coefficients 0.71 0.68
MPB/year 0.05 0.00
MAD/year
0.66
-0.05
0.43 0.65 0.98
TABLE 7 Parameter Estimates for Injury Accident Main Models
for the injury accident model. These ndings agree with results of
research by Vogt and Bared (3) and Miaou et al. (8), who found that
higher percentages of truck traffic are associated with fewer truck
crashes and fewer crashes on rural roads. However, this nding does
not apply for signalized intersections. Relating to the earlier discus-
sion about truck maneuvers, about 40% of the total maneuvers
at signalized intersections were turning movements, compared to
about 6% and 10%, respectively, for three- and four-legged stop-
controlled multilane intersections.
Vertical Curves and Safety
Researchers have examined the effect of vertical curves on safety
but did not establish a clear relationship between vertical curves and
accident frequency (Hauer, unpublished draft, 2001). In addition,
signicant safety effects of vertical curves on total crashes were
established only at three-legged stop-controlled multilane rural
intersections. A possible explanation is that vertical curves at rural
highway intersections in the data set consist of minor initial grades,
and adequate sight distance to the intersections is provided. An
interesting observation is the correlations among the roadside haz-
ard rating, posted speed, and vertical curvature on major road. For
four-legged stop-controlled multilane intersections and signalized
intersections, vertical curves are signicantly negatively correlated
with sight distance and posted speed limit. Because the models for
these intersection types include these variables, sight distance and
posted speed limit appear to have greater explanatory power for
crash frequency than vertical curves for four-legged stop-controlled
multilane intersections and signalized intersections. In short, the
effect of vertical curvature may be accounted for in these models
indirectly.
Intersection Angle and Safety
Inspection of the final negative binomial regression models re-
veals that an intersection angle that departs from a 90 angle is, in
accord with expectation, detrimental to safety at three-legged stop-
controlled multilane intersections. Engineering judgment suggests
that sight-distance restrictions, difficulty in maneuvers, and incon-
sistency with driver expectation because of skew angles lead to
reduced safety. However, skew angle was not found to be a signif-
icant factor at four-legged stop-controlled multilane intersections
and signalized intersections. A possible explanation is that traffic
maneuvers at four-legged stop-controlled intersections are more
complex than at three-legged intersections, and drivers are more
cautious. Another likely explanation is that skew angles are, on
average, more severe at three-legged intersections and are more
likely to affect safety adversely.
Finally, for signalized intersections, intersection angle has a rel-
atively smaller effect on safety than for stop-controlled intersections
because the traffic signal manages conicting vehicles and hence
drivers are provided greater traffic control.
Horizontal Curves and Safety
Numerous studies have examined the relationship between hori-
zontal curves and safety (Hauer, unpublished draft, 2001). Unlike
Oh, Washington, and Choi 25
crashes on roadway segments, intersection crashes are defined as
those crashes occurring within 250 ft of an intersection center.
This relatively small crash region suggests that safety effects of
horizontal curves differ from those for roadway segments because
most crashes at intersections occur from turning or stopping
maneuvers. The analysis presented here revealed that for three-
and four-legged stop-controlled multilane highway intersections,
a horizontal curve is not a significant contributor to the number of
observed accidents. Similar to that for vertical curves, this finding
may be an indication that horizontal curves at existing rural high-
way intersections are constrained to be mild in that they possess
small central angles, large radii, and little to no superelevation, and
in general they do not represent significant safety problems.
According to the summary statistics for three- and four-legged
intersections, the average horizontal curves of major roads are 2
(2,870 ft for radius) and 3.3 (1,740 ft for radius) per 100 ft,
respectively. These statistics indicate that horizontal curves at
most of the three- and four-legged intersections examined in this
research are larger than the minimum horizontal curve radius
guidelines provided by AASHTO (9). At signalized intersections,
horizontal curves have positive safety effects for both total and
injury crashes. It is hypothesized that on average, drivers reduce
speeds when encountering sharp horizontal curves near intersections,
and driver attentiveness is relatively high.
Speed and Safety
Although the role of speed in crashes is a primary concern in
safety, the effect of average speed on safety is murky. Numerous
studies have demonstrated positive and negative safety effects of
speed (1012). Some studies indicate that fast speeds are not haz-
ardous and perhaps are safer because turning or lane-changing
maneuvers occur more frequently when the average speed of traf-
fic is lower. Other studies revealed that higher speeds considerably
increase crash involvement because speeding reduces the distance
in which drivers can react and avoid crashes and lengthens stop-
ping distances, which in turn increases both the likelihood of crash-
ing and the severity of crashes. The ndings in this research suggest
that the safety effects of speed are different across intersection and
accident types.
For both three- and four-legged stop-controlled multilane inter-
sections, the frequency of total accidents is not consistently related
to speed. This is probably because turning maneuvers at stop-
controlled intersections are controlled by a stop sign, so that a driver
on a minor road must stop or at least reduce vehicle speed before
entering the intersection, which indicates that the driver may have
time to avoid possible conflicts with oncoming traffic on a major
road. When three- and four-legged intersections are compared, higher
speed increases injury accidents only at four-legged stop-controlled
intersections, presumably because of increased turning maneuvers
at four-legged stop-controlled intersections. For four-legged stop-
controlled intersections, turning activities are one of the factors that
increase accidents, and these intersections experienced about twice
the proportion of turning movements as three-legged intersections
on average. For signalized intersections, the research revealed neg-
ative safety effects of speed for both total and injury crash frequen-
cies. For both intersection types, the speed of the oncoming vehicles
may be a factor in crashes because higher speeds of oncoming traf-
fic make it more difficult for other vehicle drivers to judge the
amount of time and distance required to safely negotiate an inter-
section. The inconsistency of the effect of speed across crash and
intersection type is also related to the incongruence between posted
and actual speeds and the importance of variance in speeds in crash
occurrence.
Medians and Safety
The analysis suggests that painted medians are associated with lower
frequencies of accidents for three-legged intersections and total
crashes. Similarly, median widths are negatively related to crash fre-
quencies for three-legged stop-controlled intersections. No safety
effects were observed for curbed medians for all intersection and
accident types. Curbed medians have a trade-off safety effect: plac-
ing a curb in the median will largely reduce cross-median accidents
but may increase sideswipe crashes and crashes that result from vehi-
cles being deected back into the traffic stream. Hence, the net effect
of placing a curb in the median is unclear for rural highway intersec-
tions. The effect of medians in this research is unclear and is proba-
bly the consequence of inconsistencies in median treatment across
statesfor example, medians installed in response to crash histories,
access management policies, or traffic volumes.
Turning Maneuvers and Safety
Turning maneuvers have been shown to have signicant effects on
intersection safety because turning vehicles result in conicts with
oncoming traffic, and vehicles slowing to turn cause conicts with
following vehicles. This analysis supports that turning maneuvers
signicantly affect safety at three- and four-legged stop-controlled
multilane intersections. For three-legged intersections, vehicle turn-
ing percentage was signicant for injury accidents, and left-turning
maneuvers were signicant contributors for both total and injury
accidentswith increased percentages associated with increased
crash frequencies. However, turning maneuvers did not reveal sig-
nicant safety effects for signalized intersections. These ndings are
consistent with engineering expectationsuperior control of left-
turn movements is often part of the justication for signalization of
intersections.
Sight Distance and Safety
It is generally believed that intersection sight distance is one of the
important safety factors at intersections. This study revealed that
safety effects of sight distance varied across intersection types. For
four-legged stop-controlled multilane intersections, sight distance
was a signicant effect, whereas no signicant effects were revealed
for three-legged multilane intersections and signalized intersections.
For three-legged intersections, sight distance shows only minor pos-
itive safety effects, which are also insignicant. At signalized inter-
sections, traffic is controlled by signals such that sight distance
should be less of a concern than at stop-controlled intersections.
When left and right sight distances were compared, right sight dis-
tance revealed itself as a more important safety consideration than
other sight distances (longitudinal and left sight distance) at stop-
controlled multilane intersections. This is explained by observing that
right sight distance inuences potential crashes between left-turning
vehicles and through vehicles entering from the rightconicting
vehicles that must be observed across the median. In other words,
26 Transportation Research Record 1897
left turns are more protracted intersection movements than right
turns, and left-turn movements require right sight distance, whereas
right turns require left sight distance.
Lighting and Safety
This research found a positive relationship between the presence of
intersection lighting and safety for signalized rural intersections, in
agreement with previous research ndings. At signalized intersec-
tions, the absence of lighting contributes signicantly to the fre-
quency of both total and injury crashes. However, for three- and
four-legged stop-controlled multilane intersections, no signicant
safety effects of lighting were observed. For three- and four-legged
stop-controlled intersections, fewer than 30% had installed lighting,
which may suggest that lighting was installed because of high acci-
dent conditions. If this is true, the presence of lighting may be associ-
ated with higher numbers of crashes after the installation of lighting.
This phenomenon could explain the positive correlation signs
between crashes and lighting and also the positive correlations
between lighting and commercial driveways and turning move-
ments, which also were associated with the frequency of crashes at
stop-controlled intersections.
CONCLUSIONS
The difficulties and challenges faced during the model development
activities in this study were related primarily to the data, the collec-
tion of variables available for analysis, and the intended end use of
the crash prediction models. These included the following:
Interactions among variables that contribute to accident expe-
rience, most notably the interactions between geometric variables
and AADT, had to be carefully considered in model estimation.
The need to forecast crashes across states posed significant
difficulties. Hence, interstate versus intrastate variability in crash
occurrences leads to inconsistencies in countermeasure effects
across states.
Observational data also limit the amount of variation in inde-
pendent variables observed. The lack of sufficient variation in some
variables of interest made it difficult at times to estimate their effects
with the highest reliability, resulting in the possibility that some
important effects were omitted from the models.
Resource and data reliability restrictions prohibited assembly
of a data set that was sufficiently large, was randomly selected,
and contained all the pertinent variables that contribute to accident
experience.
Incongruencies between data sets across states and across peri-
ods required the assembly of unique sets of data to support each of
the three intersection models.
Despite these challenges, it was possible to develop the accident
prediction models for the three types of rural intersection. After
careful assessment of the modeling results, and a detailed look at
aspects of the intersection crash prediction models, the following
general conclusions regarding model calibration were drawn.
Traffic ows were the most reliable explanatory factor in acci-
dent occurrence. This indicates that obtaining reliable AADT esti-
mates is very important for improving or estimating crash prediction
models. However, the collection and reporting of reliable AADTs
is practically difficult, largely because of the limited number of per-
manent traffic ow count stations and limits in the frequency and
duration of counts (e.g., frequency of once yearly and duration of
24 h). Most of the state AADT data used in this research were esti-
mated AADTs. That is, all three states used in this research col-
lected short-term traffic counts (for example, for a 48-h period) for
many of their highway segments and adjusted the traffic flow for
day of week and seasonality to derive AADT estimates. Hence, it
can be assumed that the use of such estimated AADTs may cause
some bias in model estimation for safety. Furthermore, the AADT
estimates used in this research are from highway segments nearest
the intersection, not necessarily at the intersection. Unlike crash
models for highway segments, the crash prediction models for
intersections use only 250 ft for the crash location such that traffic
ows affecting intersection accidents may be different from those
for the highway segments. Furthermore, measurement errors in
turning maneuver counts also inhibit the ability to estimate reliable
models. The conclusion arising from these observations is that
improvements in intersection traffic volume counts, including turn-
ing movement counts, will yield commensurate improvements in
crash models estimated by using these data.
To test for individual state effects, state indicator variables were
tested, and no signicant effects were retained in the nal models.
However, several preliminary candidate models included the state
indicator variable for the Michigan data for total accidents at four-
legged intersections and for injury accidents at signalized intersec-
tions. [More detailed discussions about these results are available
elsewhere (5).] That state indicator variables can reveal themselves
as signicant explains why many existing fully parameterized mod-
els (models with many explanatory variables) are not transferable
across jurisdictions and illustrates that high intercorrelation of explana-
tory variables with traffic volumes may render isolation of the safety
effects of individual variables difficult at best, leading to inconsis-
tent predictions. Therefore, one cannot realistically expect that mod-
els with many variables will predict crashes equally well across
states. These state effects arise because states, and jurisdictions within
states, apply safety countermeasures in different ways. Some inter-
sections receive countermeasures because of political pressures,
some because volume warrants are met, some because crash war-
rants are met, and some because safety audits support their applica-
tion. These differences in countermeasure application result in
inconsistent magnitudes and directions of effects in crash models,
as revealed by the effect of the lighting variable in this research
effort. Caution should be exercised, therefore, in generalizing too
broadly the results of models estimated by using data of this nature.
It is recommended that national standards for estimating and
reporting crash, traffic ow, and geometric variable information at
intersections be established. There are no national standards for the
measurement and reporting of these variables, and reporting prac-
tices vary across states and jurisdictions. Furthermore, the variables
measured across states and jurisdictions are inconsistent and
updated too infrequently to be wholly reliable for crash prediction.
In this research effort, considerable time was spent assembling data
from each state, and difficulties were encountered in trying to obtain
missing information (such as peak turning percentage movements
in Georgia and minor-road geometric information in California and
Michigan) required for model development. Hence, developing a
national standard for measurement and reporting of the variables
Oh, Washington, and Choi 27
will yield major improvements in the ability to estimate models for
national applications.
Finally, the study results developed strong support for the inter-
active highway safety design model (IHSDM) approach: use of base
models with accident modication factors. Support for the approach
included inconsistency of geometric explanatory variables across
states, strong interstate crash variability, and the dominance of AADT
as the prime explanatory factor in crash occurrence. After detailed
examination of the data obtained across several states and periods,
it became apparent that the most defensible approach for forecast-
ing crashes, given current data availability, is the approach pro-
posed by IHSDM. As a corollary, statistical models that rely on an
expanded set of predictor variables are not reliable across states and
should be applied with caution.
ACKNOWLEDGMENTS
The authors acknowledge FHWA sponsorship of the core work that
led to the paper. The authors thank B. N. Persaud and Craig Lyon
for their support and contributions during the course of the study and
their parallel efforts and thank Joe Bared of FHWA, Andrew Vogt
of Pragmatics, Inc., Yusuf Mohamedshah and Forrest Council of the
Landis Corporation, and the Georgia Department of Transportation
for their assistance in acquiring data needed for this study.
REFERENCES
1. Traffic Safety Facts. NHTSA, U.S. Department of Transportation, 2001.
2. Vogt, A. Crash Models for Rural Intersections: Four-Lane by Two-
Lane Stop-Controlled and Two-Lane by Two-Lane Signalized. FHWA-
RD-99-128. FHWA, U.S. Department of Transportation, 1999.
3. Vogt, A., and J. Bared. Accident Prediction Models for Two-Lane Rural
Roads: Segments and Intersections. FHWA-RD-98-133. FHWA, U.S.
Department of Transportation, 1998.
4. Harwood, D. W., F. M. Council, E. Hauer, W. E. Hughes, and A. Vogt.
Prediction of the Expected Safety Performance of Rural Two-Lane
Highways. FHWA-RD-99-207. FHWA, U.S. Department of Trans-
portation, 2000.
5. Washington, S., B. Persaud, J. Oh., and C. Lyon. Validation and Recal-
ibration of Accident Prediction Models for Rural Intersections. FHWA,
U.S. Department of Transportation, 2003.
6. Washington, S., M. Karlaftis, and F. Mannering. Statistical and Econo-
metric Methods for Transportation Data Analysis. Chapman and
Hall/CRC, Boca Raton, Fla., 2003.
7. Neter, J., M. H. Kutner, C. J. Nachtsheim, and W. Wasserman. Applied
Linear Statistical Models, 4th ed., McGrawHill, New York, 1996.
8. Miaou, S. P., P. S. Hu, T. Wright, S. C. Davis, and A. K. Rathi. Devel-
opment of Relationship Between Truck Accidents and Geometric
Design. FHWA-RD-91-124. FHWA, U.S. Department of Transporta-
tion, 1993.
9. A Policy on Geometric Design of Highways and Streets, 4th ed.,
AASHTO, Washington, D.C., 2001.
10. Parker, M. R. Effects of Lowering and Raising Speed Limits on Selected
Roadway Sections. FHWA-RD-92-084. FHWA, U.S. Department of
Transportation, 1997.
11. Navon, D. The Paradox of Driving Speed: Two Adverse Effects on
Highway Accident Rate. Accident Analysis and Prevention, Vol. 845,
2002, pp. 17.
12. Shinar, D. Speed and Crashes: A Controversial Topic and an Elusive
Relationship. Traffic Engineering, Vol. 41, 1999, pp. 5255.
Publication of this paper sponsored by Safety Data, Analysis and Evaluation
Committee.

Você também pode gostar