Professional Documents
Culture Documents
Because traffic safety studies are not performed in a controlled environment such as a laboratory, but rather in an uncontrolled real-world
setting, traditional analysis methods often lack the capability to evaluate
the effectiveness of roadway safety measures adequately. In recent
years, however, advanced statistical methods have been used in traffic
safety studies to determine the effectiveness of such measures more
accurately. These methods, especially Bayesian statistical techniques,
can account for the shortcomings of traditional methods. Hierarchical
Bayesian modeling is a powerful tool for expressing rich statistical models
that more fully reflect a given problem than traditional safety evaluation
methods could. This paper uses a hierarchical Bayesian model to analyze the effectiveness of raised medians on overall and severe crash frequency in Utah by determining the effect each newly installed median
has on crash frequency and frequency of severe crashes at study sites
before and after installation of the medians. Several sites at which raised
medians have been installed in the past 10 years were evaluated with
available crash data. The results show that the installation of a raised
median is an effective technique to reduce the overall crash frequency
and frequency of severe crashes on Utah roadways, with results showing
a reduction in overall crash frequency of 39% and frequency of severe
crashes of 44% along corridors where raised medians were installed. The
results also show that hierarchical Bayesian modeling is a useful method
for evaluating effectiveness of roadway safety measures.
Background
Because traffic safety studies are not performed in a controlled
environment, such as a laboratory, but rather in an uncontrolled
real-world setting, traditional analysis methods often lack the capability to adequately evaluate the effectiveness of roadway safety
measures. However, safety studies have historically had to rely on
these traditional methods because of the complexity of more effective models. Fortunately, the development of advanced statistical
software and methodologies in recent years has helped overcome
the complexity of the models and made advanced statistical methods more attainable. These advanced statistical methods, especially
Bayesian statistical methods, have the capabilities to account for the
shortcomings of traditional methods and have thus been used in recent
Raised Medians
In this study, the safety data collection and analysis techniques developed are applied to study sites at which raised medians have been
installed throughout the state of Utah. A raised median is a physical
barrier, such as a concrete or landscaped island, in the center portion
of the roadway that separates opposing lanes of traffic and is designed
so that it is not easily traversed. Raised medians are appropriate in
some, although not all, locations and have been found to be most
useful on high-volume, high-speed roadways (2).
Raised medians have frequently been used as an access management technique to improve roadway safety in two primary ways:
(a) raised medians reduce the number of conflict points by allowing
turning movements to be made only at designated openings or at
signalized intersections, and (b) raised medians provide a physical
barrier separating opposing traffic in an effort to eliminate head-on
collisions (3).
Raised medians are not always used to mitigate one specific
type of crash or factor (2). Raised medians have been installed for
beautification purposes, as an access management technique, as a
traffic calming device, for pedestrian refuge, and for various other
purposes (3). Regardless of the purpose, raised medians have had a
Safety Analysis
To determine the safety of a site during analysis, it is essential to
first define what safety is and how it is measured. Roadway safety
is usually defined and evaluated by recorded numbers of crashes or
crash rates. Severity of crashes also plays an important role in the
understanding of roadway safety. For example, one site may experience considerably more crashes than another site. However, the
second site may have a much larger proportion of severe, even fatal,
crashes. Therefore, both frequency of crashes and severity of crashes
are essential in determining the safety of a facility.
One of the key reasons it is difficult to understand why crashes
occur is that crashes are usually not the result of one factor, but instead
the combination of several events, circumstances, and factors. The
combination of multiple events can significantly alter the amount of
risk a driver may face. Even if it were possible to account for all
possible factors that lead to a crash, the ability to predict a crash is
not absolute. The key principle is that understanding the nature of
crashes is a vastly complex and random process when only the known
factors are considered. It is also important to remember that unknown
factors also contribute to crashes. The complexity of known and
unknown contributing factors can be overcome through the development and use of proper statistical tools that can correctly model
crash characteristics and behavior. To properly analyze crash statistics, it is essential to analyze the characteristics of crash statistics to
determine the proper statistical tools to use.
Characteristics of Crashes
One of the key concepts of a crash study is that although there are
trends and factors that increase the likelihood of crashes, the occurrence of a crash is not completely predictable. Crashes, by nature,
are random events. The frequencies of crashes will naturally fluctuate
from year to year. Fluctuations in crash frequency make it difficult
to determine whether a reduction in the number of crashes is a result
of a specific treatment, changes in site conditions over time, or a result
of natural fluctuations from stochastic processes. These fluctuations
present a phenomenon referred to as regression-to-the-mean (RTM)
bias. The RTM phenomenon expects that a value that is determined
to be extreme will tend to regress to the long-term average over
time. This means that a period of high crash frequencies at a site
is statistically probable to be followed by a period of low crash
frequencies (9).
Traditional methods have been used for many years to analyze
crash statistics. Traditional methods use relatively straightforward
before-and-after approaches to compare the crash frequency or rate
of an entity immediately before an improvement was made with the
crash frequency or rate directly after the improvement to determine
the effectiveness of the treatment (9). One of the problems with
97
many traditional analysis methods is that they do not account for the
RTM bias. If the RTM bias is not accounted for, the effectiveness of
a specific treatment is likely to be inaccurately reported.
Predicting Crashes
The RTM bias provides evidence of limitations when short-term
data are used for analysis, which would lead to the assumption
that using data for longer periods provides a better representation
of crash behavior at a site. However, there are problems associated
with this method as well. The characteristics of a site, such as traffic
volume, weather, and pavement condition, change over time. Some
of these characteristics, such as weather, continually fluctuate. Other
factors, such as pavement condition and roadway markings, deteriorate gradually over years of use. These latter factors create a
legitimate limitation when using long-term crash statistics for site
analysis.
Factors contributing to crashes can never be completely controlled
or maintained, providing a level of difficulty in accurately predicting
crashes. Crash data are statistically classified as count data and by
nature are nonnegative integers; therefore generalized linear models
are insufficient because the assumption that the dependent variable
is continuous is not true for crash studies (10). Previous studies have
suggested the use of Poisson models or negative-binomial models
as more appropriate for count statistics (1113).
used because of the nature of crash data classified as count data. The
Poisson distribution also allows for an exposure parameter, which
in this case is the length of the roadway segment. The number of
crashes per mile (or severe crashes per mile), rather than the number
of crashes solely, is then modeled for each roadway segment. Data
from segments with varying segment length may thus be included
in the same model appropriately.
yi ~ Poisson ( i )
(1)
where yi is the number of crashes per mile for each AADT segment
and qi is the mean number of crashes per mile.
The estimation of the mean number of crashes per mile is calculated
using Equation 2.
log ( i ) = A Ai + B Bi + 1AADTi
(2)
where
(3)
The priors, p(bj), were selected with mean zero because previous
analyses of this type have not been performed on the chosen roadway segments to suggest higher or lower mean values for the prior
distributions. Zero mean values allow the posterior distributions of
the coefficients to be positive or negative according to the data without the prior pulling the posterior results in one way or the other. The
posterior distribution, p(bA, bB, b1y), for the b parameters, up to a
constant, is expressed in Equation 4.
( y ) P ( y ) ( A ) ( B ) (1 ) =
n
i =1
e(
Xi )
( e(
n
i =1
yi Xi )
)e
1 2 2 2
2 A + B +1
(44)
where
Xi = matrix with Ai, Bi, and AADTi as its columns;
b = matrix of parameters bA, bB, and b1; and
n = total number of observations.
Because of the complexity of the normalizing constant of the posterior distribution, rather than deriving the distribution theoretically,
Markov chain Monte Carlo methodology is used to sample from
the posterior distribution using the scaled distribution shown in
Equation 4 (12, 13, 22, 23). The results of the algorithm are several
99
Site Selection
Before a detailed crash analysis is conducted, study sites need to
be selected. The study sites that have been selected for analysis where
raised medians have been installed are discussed in this section. For
crash analysis purposes, the study sites will be roadway segments
where raised medians have been installed, as outlined in previous
research in Utah (24). The following study sites were selected for
analysis:
1. University Parkway (SR-265), Mile Point 1.21 to 1.96;
2. Alpine Highway (SR-74), Mile Point 2.40 to 4.29;
3. 400/500 South (SR-186), Mile Point 5.48 to 7.53;
4. St. George Boulevard (SR-34), Mile Point 0.00 to 1.74; and
5. SR-36, Mile Point 59.29 to 60.82.
Listed mile points are current mile points at the time of this study.
More detail on the study sites is available in the literature (25).
Data Collection
Raw crash statistics were provided by the Utah Department of Transportation (DOT) Traffic and Safety Division from the Utah DOT
crash database. The database contains records and statistics obtained
from police reports for crashes that occurred on Utah roadways. At the
time of this study, consistent data were available from 1998 to 2008.
AADT data are used to measure total volume of vehicle traffic of a
highway or road. Previous research has determined that a relationship
exists between crashes and AADT. Although the exact relationship
is still not entirely known, it is known that the relationship is generally nonlinear (9). However, AADT is still an important parameter
in predicting crash frequency and was used as a covariant in the
development of the model.
Locations of crashes are reported as the mile point at which the
crash occurred on the corresponding roadway. However, mile points
on Utah highways have undergone several changes over the past
10 years. Shifts in mileposts are usually the result of either a realignment of the roadway or an extension added on to either end of the
roadway. Although the segments of each roadway of interest were
held constant through each analysis year, corresponding mile points
have changed over the course of the study period. To ensure that
data for the correct segment was used for each analysis year, correct
mileposts were verified through the Utah DOT.
Analysis Results
Individual Site Analysis Results
A hierarchical Bayesian analysis was performed at selected study
sites at which raised medians have been installed. An analysis was
performed on overall crash frequency and severe crash frequency
for each segment. The results of the hierarchical Bayesian model
before and after the raised median installation were used to calculate
a percent change in crash frequency. In addition to the percent change,
the probability that the crash frequency (overall or severe) decreased
was calculated. The probability of reduction helps to identify the
statistical significance of the change.
The analysis of the individual study sites at which raised medians
have been installed showed three of the five study sites experienced
a statistically significant (greater than 95% probability) reduction
(26% to 43%) in the overall crash frequency. The probability of
difference for the remaining two sites (SR-265 and SR-74) was too
low to confidently determine whether a reduction or increase occurred
(although the results for SR-74 were practically significant at a 93%
probability of decrease). In these situations the mean is increasing,
but not at a statistically significant rate. Table 1 provides a summary
of the impact of raised medians on all crashes.
Similar to the overall crash analysis, several of the study sites
also showed a reduction in the frequency of severe crashes. Table 2
provides a summary of the impact of raised medians on severe crashes.
The results indicate that three of the five study sites experienced a
significant (greater than 95% probability) reduction (60% to 67%)
in the frequency of severe crashes along the segment after raised
medians were installed. The analysis indicated an increase (55%)
may have occurred at one of the remaining sites (SR-74); however,
the probability of a difference was too low to confidently determine
if a reduction or increase occurred. The final site (SR-36) showed a
90% probability (practically significant) that a decrease of 43%
occurred. More detailed analysis results for both overall and severe
crash frequencies are provided in the literature (25).
State
Route
County
265
74
186
34
36
Utah
Utah
Salt Lake
Washington
Tooele
The data plot displays the actual data points and the mean of the
posterior predictive distribution, which is a representation of the mean
regression line through the points from a Bayesian perspective. The
reduction is calculated by taking the mean of the posterior distribution
of differences between the two intercepts. The mathematical details
are discussed in the literature (19, 20), but are conceptually equivalent
to taking the after curve and dividing it by the before curve to obtain
the percent reduction.
The second plot produced for the overall analysis results is the
plot of the distribution of the differences between the before and
after periods. The difference plots display the posterior distributions
of differences between the before and after intercepts of the model.
Negative values indicate that the after time period saw a reduction
in crashes. Because the exact form of the posterior distributions is
unknown, the model uses simulated draws from the posterior with
the Markov chain; because those draws represent the actual posterior distribution, the proportion of the draws less than zero represents the probability that there was a reduction in crashes from the
before time period to the after time period.
Figure 1 displays the overall crash frequency for the before and
after periods as a function of AADT. The overall analysis results
indicate a 39% reduction in overall crash frequency after the raised
medians were installed. Figure 2 shows the corresponding probability
distribution of the differences between the before and after periods
for overall crashes. The entire distribution of differences in Figure 2
is less than zero, indicating a 100% probability that a reduction in
overall crash frequency occurred after raised medians were installed.
The analysis results of severe crashes display an even greater
reduction. Figure 3 displays the results for all study sites at which
raised medians were installed. The severe crash analysis results on
all study sites show a 44% reduction in severe crash frequency after
raised median installation. Figure 4 shows the corresponding probability distribution of the differences between the before and after
periods for severe crashes. As with the overall analysis, the entire
distribution of differences shown in Figure 4 is less than zero, indi-
Year of
Installation
Probability of
Reduction (%)
Percent
Change
2002
2002
2001
2006
2005
38
93
100
100
100
3
-19
-29
-26
-43
Conclusions
The analysis in this report was performed with a hierarchical
Bayesian model developed to analyze the effectiveness of various
treatments for improving roadway safety. The model is a valuable
tool with various applications to transportation safety. As part of this
research, the model was applied to study sites at which raised medians
had been installed throughout the state of Utah. This study analyzed
the effectiveness of raised medians on roadway safety by determining the effect each newly installed median has had on frequency of
overall crashes and of severe crashes at the study sites. An analysis
was performed on individual sites at which raised medians have been
installed as well as an overall analysis on all study sites grouped
together.
The results of the raised median analysis indicated a significant
improvement in both crash frequency and frequency of severe crashes
along corridors where raised medians were installed. Results from
all study sites combined show that the overall crash frequency was
reduced by 39% and the frequency of severe crashes was reduced
by 44% after the installation of raised medians along the study sites.
The reduction in frequency of severe crashes is anticipated to be a
result in the change in types of collisions. This study provides further and more statistically significant evidence that installing raised
medians is an effective technique to reduce crash frequency, especially
severe crashes caused by sideswipes or head-on collisions.
The model developed for this research can be expanded to
determine the impact of other safety measures on various crash
types. Additionally, AADT was the only covariate used in this study.
The model has been developed such that additional covariates may
be included in the analysis. Selection of the appropriate covariates to
be used depends on the scope of the study being performed. Finally,
State
Route
County
265
74
186
34
36
Utah
Utah
Salt Lake
Washington
Tooele
Year of
Installation
Probability of
Reduction (%)
Percent
Change
2002
2002
2001
2006
2005
100
41
100
99
90
-60
55
-67
-61
-49
101
50
200
10,000
20,000
30,000
AADT
40,000
50,000
FIGURE 2 Distribution of differences for crashes on all raised median study sites.
15
10,000
20,000
30,000
AADT
40,000
50,000
FIGURE 3 Severe crash frequency for all raised median study sites.
FIGURE 4 Distribution of differences of severe crashes for all raised median study sites.
Acknowledgments
This research was made possible with funding from the Utah DOT
and Brigham Young University. Special thanks to Robert Hull, W.
Scott Jones, Tim Taylor, David Stevens, Mitsuru Saito, Steven Dudley,
and others at the Utah DOT and the university who played key roles
as members of the technical advisory committee and provided input
to the research.
References
1. Highway Safety Manual. AASHTO, Washington, D.C., 2010.
2. Center for Transportation Research and Education. Access Management
Toolkit: Answers to Frequently Asked Questions. Continuous Raised
Medians, No. 18. Iowa State University, Ames. http://www.ctre.iastate.
edu/research/access/toolkit/18.pdf. Accessed May 8, 2010.
3. Access Management Manual. Transportation Research Board of the
National Academies, Washington, D.C., 2003.
4. Gluck, J., H.S. Levinson, and V. Stover. NCHRP Report 420: Impacts of
Access Management Techniques. TRB, National Research Council,
Washington, D.C., 1999.
5. Eisele, W.L., and W.E. Frawley. Estimating the Safety and Operational
Impact of Raised Medians and Driveway Density: Experiences from
Texas and Oklahoma Case Studies. In Transportation Research Record:
Journal of the Transportation Research Board, No. 1931, Transportation
Research Board of the National Academies, Washington, D.C., 2005,
pp. 108116.
6. Gattis, J.L., R. Balakumar, and L.K. Duncan. Effects of Rural Highway
Median Treatments and Access. In Transportation Research Record:
Journal of the Transportation Research Board, No. 1931, Transportation
Research Board of the National Academies, Washington, D.C., 2005,
pp. 99107.
7. Parsonson, P.S., M. Waters, and J.S. Fincher. Georgia Study Confirms
the Continuing Safety Advantage of Raised Medians over Two-Way
Left-Turn Lanes. Presented at 4th National Conference on Access Management, Portland, Ore., Transportation Research Board of the National
Academies, Washington, D.C., 2000.
8. Schultz, G.G., J.S. Lewis, and T. Boschert. Safety Impacts of Access
Management Techniques in Utah. In Transportation Research Record:
Journal of the Transportation Research Board, No. 1994, Transportation
Research Board of the National Academies, Washington, D.C., 2007,
pp. 3542.
103