Collision Prediction for Two Lane Rural Roads Using IHSDM: Abstract

Collision Prediction for Two Lane Rural Roads Using IHSDM:  Abstract
Collision Prediction for Two Lane Rural Roads Using IHSDM:
A Canadian Experience
Matthew Marleau, Eric Hildebrand
University of New Brunswick Transportation Group
Abstract
A study at the University of New Brunswick was performed on the transferability of the collision
prediction capabilities of Interactive Highway Safety Design Model (IHSDM) within a rural
Canadian context. Collision prediction models are generally created with data that intrinsically
reflect geographic, environmental, and operational characteristics unique to a particular region.
The transferability of a model to a different area can sometimes be problematic. The IHSDM
model has been developed with two levels of calibration that are intended to facilitate its
widespread application.
This study evaluated both the 2008 and 2009 (beta) versions of the collision prediction models
for two lane rural roads embedded in IHSDM. A sample of seventy-five rural two-lane road
segments from the province of New Brunswick was analyzed to evaluate the model’s
performance. The analysis compared the predicted collision frequencies with the empirical
collision data provided by the New Brunswick Department of Transportation. Results of the study
have shown that overall the model performs poorly when compared with the average number of
observed collisions. Predicted collision frequencies for all test sites combined were overestimated by 38 to 78 percent (depending on the level of calibration employed). Goodness-of-fit
testing including mean absolute deviation, mean prediction bias, and R2 showed that the model
does not perform well within a Canadian context and that the calibration methods need refining.
Model fits described by coefficient of determination, R2, ranged from only 0.001 to 0.255.
Résumé
Une étude a été réalisée à l’Université du Nouveau-Brunswick sur la portabilité du potentiel de
prédiction des collisions du modèle interactif de conception de la sécurité routière (IHSDM) dans
un contexte rural canadien. Les modèles de prédiction des collisions sont généralement créés à
l’aide de données qui reflètent intrinsèquement les caractéristiques géographiques,
environnementales et opérationnelles propres à une région donnée. La portabilité d’un modèle à
une région différente peut parfois poser des problèmes. Le modèle IHSDM a été conçu avec
deux niveaux d’étalonnage dont le but est de faciliter la généralisation de son application.
Proceedings of the 20th Canadian Multidisciplinary Road Safety Conference,
Niagara Falls, Ontario, June 6-9, 2010
e
Compte-rendu de la 20 Conférence canadienne multidisciplinaire sur la sécurité routière,
1
Cette étude a évalué à la fois les versions 2008 et 2009 (bêta) des modèles de prédiction des
collisions sur les chemins ruraux à deux voies enchâssés dans l’IHSDM. On a analysé un
échantillon de 75 tronçons de chemins ruraux à deux voies dans la province du NouveauBrunswick pour évaluer l’efficacité du modèle. L’analyse a comparé les fréquences prévues des
collisions aux données empiriques sur les collisions fournies par le ministère des Transports du
Nouveau-Brunswick. Les résultats de l’étude ont démontré que, dans l’ensemble, le modèle a
un comportement plutôt médiocre par rapport au nombre moyen de collisions constatées. Les
fréquences prévues des collisions à tous les sites expérimentaux confondus ont été surestimées
de 38 % à 78 % (selon le niveau d’étalonnage utilisé). Les tests de validité de l’ajustement,
notamment l’écart absolu moyen, le facteur moyen de correction de la prévision et R2 ont révélé
que le modèle se comporte mal dans le contexte canadien et que les méthodes d’étalonnage
doivent être fignolées. Les ajustements du modèle décrits par un coefficient de détermination R2
variaient d’à peine 0,001 à 0,255.
INTRODUCTION
Collision prediction modeling has many uses in road safety engineering and planning including
black spot identification, road safety audits, and benefit/cost analysis of road improvements. In
recent years, collision prediction models have been developed using data from many different
geographic areas. These models are typically designed to be applied in specific areas and
problems arise when attempts are made to use these models in different regions. This is partly
due to the fact that collisions are random events which can have many regional specific factors
influencing them.
With the development of the Interactive Highway Safety Design Model (IHSDM) by the Federal
Highway Administration (FHWA), a standardized approach to safety analysis including collision
prediction has been sought [1]. Since IHSDM was originally released, several upgraded versions
have been developed with each new version generally introducing a new element or an
improvement to other principal components. The latest full version of IHSDM was released in
October 2008. The initial focus of the IHSDM was on rural two lane roads; however, a beta
version released by the FHWA in November 2009 includes capabilities to predict collisions on
urban, suburban, and multilane facilities [2].
Since the release of IHSDM, few studies have been performed on the transferability of the
model. This report synthesizes a study performed at the University of New Brunswick to test the
transferability of the collision prediction module within IHSDM, as well as the level of calibration
required to produce optimal results. If this design suite is to become widely adopted in the
United States, it is important to understand its limitations and whether it can accurately be
employed in the Canadian context.
Proceedings of the 20th Canadian Multidisciplinary Road Safety Conference,
Niagara Falls, Ontario, June 6-9, 2010
e
Compte-rendu de la 20 Conférence canadienne multidisciplinaire sur la sécurité routière,
2
IHSDM BACKGROUND
In an attempt to improve the level of safety included in road design and planning, the FHWA has
been developing a suite of tools collectively known as the IHSDM since the mid-1990s. It is
widely expected that these tools will establish a new standard for highway design in the United
States. The IHSDM software was initially developed when a deficiency was recognized in
checking road compliance with federal, state, and local policies. A need was also recognized to
determine the road users’ comprehension of road designs in their driving practices [1]. The
current full release of the IHSDM software suite contains six analysis modules. The components
are [2]:






Collision prediction
Driver/vehicle
Policy review
Design consistency
Traffic analysis
Intersection review
The collision prediction model embedded in the October 2008 version was created in the late
1990’s by Harwood et al. [3]. This collision prediction model was only developed to prediction
collisions on two-way, two-lane rural highways. The model was created by using negative
binomial regression using data taken from Minnesota and Washington. The model applies the
following basic structure:
Nrs = Nbr (AMF1 AMF2 ...... AMFN)
(1)
Where Nrs is the total number of predicted collisions for a given road segment. Nbr represents
the number of collisions under nominal conditions, or a base condition. This is calculated by
using several factors including lane width, shoulder width, segment length, average daily traffic,
etc. Each AMF represents a multiplication factor which positively or negatively affects the
number of collisions at a location. The ideal AMF factor is set to 1.0, where no change will occur
to the model.
In order to facilitate transferability of the collision prediction model, a calibration method was
derived by Harwood et al. The calibration model contains two levels for which calibration can be
performed, named “level 1” and “level 2.” Both of these levels of calibration require a minimum
set of road segment specific data including the total length of two-way, two-lane highway,
accident data, alignment and grade data, and require this information to be sorted into several
traffic volume groups. For a level 2 analysis, the same data is required but it is further
segregated into sub-groups by shoulder and lane widths [3].
The beta version of a newer IHSDM collision prediction model was released in November 2009.
This new model is set to conform with Part C of the upcoming Highway Safety Manual [2], which
Proceedings of the 20th Canadian Multidisciplinary Road Safety Conference,
Niagara Falls, Ontario, June 6-9, 2010
e
Compte-rendu de la 20 Conférence canadienne multidisciplinaire sur la sécurité routière,
3
has been scheduled for release sometime later in 2010 [4]. Since the Highway Safety Manual
has not been released at the time of publication, it is unknown what level of calibration will be
included, or whether there is a major fundamental difference beyond the 2008 version. An
important item of interest with the 2009 version of IHSDM is the inclusion of urban and suburban
roads, as well as the inclusion of multi-lane facilities in rural areas.
PREVIOUS STUDIES
Since the initial development of IHSDM there have been few studies on the accuracy and
transferability of the model. Several projects have been undertaken using the suite of software to
evaluate the potential for collision reductions in realignment projects, but no retrospective
reports on the accuracy of the model have been published [2].
Saito and Chuo [1] performed a study in 2008 of the viability of using IHSDM in road safety
audits. Their study focused on three road segments, the US-10, SR-150, and US-6. The study
utilized the built-in Empirical-Bayes (EB) modification utility by using previous collision history on
the three highway segments. The study by Saito and Chuo showed a large variation between
the actual collision data and the estimated collision rates.
A study by Donnell et al. [5] tested the IHSDM collision prediction model on two highway
segments in the state of Pennsylvania. The Donnell et al. study tested IHSDM using the level 1
and level 2 calibration techniques over three geographic areas: county, district and state. The
study, along with the Saito and Chuo study found there was a large variation between the actual
and estimated collision data.
A third study was performed at the University of New Brunswick in 2006. The study compared
three collision prediction models to historical data for the province. The three models tested
included IHSDM, the Transportation Association of Canada model, and MicroBENCOST. The
study found that of the three models, IHSDM produced estimates that more closely represented
observed values, however, it was noted that the prediction error was approximately 46 percent
for the total number of collisions [6].
MODEL EVALUATION
In order to evaluate the IHSDM collision prediction model, 75 random test segments across the
province of New Brunswick were selected. Criteria for these test segments were taken and
modified from research published by Ye [6] in which each test segment was required to be over
2.0 km in length, and more than 80 m from any intersection in order to reduce the number of
intersection related crash data from being included.
Proceedings of the 20th Canadian Multidisciplinary Road Safety Conference,
Niagara Falls, Ontario, June 6-9, 2010
e
Compte-rendu de la 20 Conférence canadienne multidisciplinaire sur la sécurité routière,
4
Since there were two versions of the IHSDM collision prediction model available for use, both
were tested. The 2008 version of IHSDM, the latest full release version, was tested at three
application levels. The tests included evaluating the model in a base scenario over five years
using an un-calibrated model and comparing the estimated collision rates to the actual observed
collision rates for the test segments. The second and third evaluations of the 2008 model used
the level 1 and level 2 calibration methods and compared their respective estimated collision
rates to empirical crash data. The 2009 beta version of the IHSDM was tested in an uncalibrated state only as the documentation on the calibration process has not been made
available at the time of study.
The model was not evaluated the accuracy of the severity of collisions predicted. This is
because the IHSDM model uses set values for each type of collision, such as property damage
only, injury, and fatality. It is possible to set these values to reflect the conditions for the
geographic area including the configuration of collisions that occurred. The calibration for
collision configuration does not affect the overall number of collisions predicted as each
modification factor for the type of collision sums to 100% of the total number of collisions.
Both versions of IHSDM have a built-in Empirical-Bayes calibration method. This method utilizes
previous collision history and will have the predicted number of collisions conform more to the
historical values. This method was not tested in this study as the effects of the Empirical-Bayes
method will sway the predicted number of collisions closer to the actual historical values and
would undermine the test of goodness-of-fit of the model under actual conditions. The intended
application of this method is more related to estimating the impacts of design modifications to an
existing road which is outside the objective of this study.
Data Collection and Model Calibration
Test segment data were acquired from the New Brunswick Department of Transportation and
included basic characteristics such as average annual daily traffic, speed limits, road and
shoulder widths, degrees of curvature, grade severities, intersection locations, and collision
history. Once the data were collected, the information was entered into the IHSDM Highway
Editor in order to model the highway segment within the software.
In order to perform the level 1 and level 2 calibrations, more road segment descriptive data were
required for each test segment. All of the data collected for the calibration methods were placed
into different annual daily traffic (ADT) groups as well as subdivided by lane and shoulder
widths. The data collected for the level 1 and level 2 calibrations included the following:





Total length of road
Total length of road on grade and on a curve
Average grade for those sections on grade
Average curvature for those sections of curves
Total collisions
Proceedings of the 20th Canadian Multidisciplinary Road Safety Conference,
Niagara Falls, Ontario, June 6-9, 2010
e
Compte-rendu de la 20 Conférence canadienne multidisciplinaire sur la sécurité routière,
5


Shoulder and road widths
Percentage of segment length that is flat, rolling, or mountainous
The calibration method for IHSDM consisted of using the Administration Tool provided with the
software. The tool creates a spreadsheet where the collected data for the road network is
compiled. Once the data entry is complete, the spreadsheet calculated a calibration ratio (Cr).
Depending on the various geometry and collision history the value will be either above or below
zero. The calibration ratio is then entered into a configuration file and used in the analysis of the
section. In the case of this project, the Cr values were 1.291 and 1.283 for the level 1 and level 2
calibrations, respectively, which basically means that the base model was estimated to underpredict the total collisions by 28-29% given the site-specific characteristics included in the
calibration algorithms.
Data Analysis
The model evaluation techniques used in this research consisted of testing the goodness-of-fit of
the estimated collision rates from IHSDM to the historical data provided by the New Brunswick
Department of Transportation. Three tests of goodness-of-fit were undertaken during this project
including the mean prediction bias, the mean absolute deviation, and linear regression.
The mean prediction bias (MPB) test provides a measure of the goodness-of-fit of the model by
comparing the overall difference between the test data and the historical data, as well as
showing what direction the output is from the historical data. A low MPB value indicates the
model performs well in comparison to the historical data, where a high MPB value indicates poor
conformance. Positive MPB rates show the model over-predicts the number of collisions, while
negative MPB rates show the model under-predicts. MPB is calculated using the following
formula:
∑
(2)
Where Y is the predicted value, Yi is the actual value and n is the number of samples [7].
The mean absolute deviation (MAD) provides a similar goodness-of-fit comparison as the MPB
test does; however, the MAD model gives the average difference in prediction of the model in an
absolute format, meaning negative and positive differences in prediction will not cancel each
other out. Similar to the MPB, values closer to 0 show that the model performs well when
compared to historical data. MAD is calculated using the following formula:
∑
|
|
(3)
Where Y is the predicted value, Yi is the actual value and n is the number of samples [7].
Proceedings of the 20th Canadian Multidisciplinary Road Safety Conference,
Niagara Falls, Ontario, June 6-9, 2010
e
Compte-rendu de la 20 Conférence canadienne multidisciplinaire sur la sécurité routière,
6
Linear regression can be used to show whether or not there is a direct linear relationship
between the model output and the observed collision data when plotted against one another.
The model takes a general form of:
Y = a + b*X
(4)
Where Y is the predicted number of collisions, X is the historical number of collisions, “a” is the
y-intercept and “b” is the slope of the linear line. Generally, the model will want to have an
intercept close to 0 and a slope close to 1 in order to show a strong relationship between the
model output and empirical data. A test of the strength of the linear relationship between the two
sets of data is reflected by the R2 coefficient. The value of R2 is always between 0 and 1. A value
of R2 that is closer to 0 shows that there is very little or no linear relationship between the two
variables and the model is not a good fit [6].
All of the IHSDM output data were compared against the observed average number of collisions
per year, and the average number of collisions per million vehicle-kilometres. These data were
delineated into eight sub-groups which were selected based on the road classes and sub-groups
required to perform the level 1 and level 2 calibrations provided by Harwood et al. [3]. The subgroups used to perform the data analysis were as follows:








All test sections
Arterial roads
Collector roads
Average daily traffic (ADT) under 1,000
Average daily traffic (ADT) between 1,001 and 3,000
Average daily traffic (ADT) between 3,001 and 5,000
Average daily traffic (ADT) between 5,001 and 10,000
Average daily traffic (ADT) over 10,0001
Table 1 gives a breakdown of how many test segments are located within each sub-group. The
sub-group with average daily traffic over 10,001 has a very small sample size. This is due in part
to the low volume of many of the roads in rural New Brunswick and the data requirements for
road segment length between intersections.
Table 1 – Road Segment Sample Sizes
Arterial
Roads
Collector
Roads
24
51
ADT Volume Bins
Under 1,000
1,001 –
3,000
3,001 –
5,000
5,001 –
10,000
Over 10,001
16
24
21
12
2
Proceedings of the 20th Canadian Multidisciplinary Road Safety Conference,
Niagara Falls, Ontario, June 6-9, 2010
e
Compte-rendu de la 20 Conférence canadienne multidisciplinaire sur la sécurité routière,
7
RESULTS
All Test Segments Combined
The data in Tables 2 and 3 synthesize the results for the model results for all 75 road test
segments. The data in Table 2 represent the results for collisions per year while the data in
Table 3 represents the results for collisions per million vehicle-kilometres (mvkm).
Table 2 - All Test Segments Results (collisions/year)
IHSDM Model Results
Avg. of all
segments
MPB
MAD
R2
Actual
2008
No Calibration
2008
(Level 1)
2008
(Level 2)
2009
(beta)
1.279
1.760
2.276
2.255
2.089
-
0.481
0.843
0.184
0.997
1.183
0.184
0.976
1.167
0.184
0.810
1.034
0.177
Table 3 - All Test Segments Results (collisions/mvkm)
IHSDM Model Results
Avg. of all
segments
MPB
MAD
R2
Actual
2008
No Calibration
2008
(Level 1)
2008
(Level 2)
2009
(beta)
0.503
0.587
0.759
0.753
0.691
-
0.085
0.282
0.001
0.257
0.379
0.001
0.250
0.375
0.001
0.189
0.330
0.005
The results in Tables 2 and 3 show that, on average, the 2008 model with no calibration is the
most accurate in predicting the number of collisions in terms of both collisions per year and
collisions/mvkm. The difference between the averages for collisions per year and
collisions/mvkm are 38% and 17%, respectively. It is noteworthy that the calibration (levels 1 and
2) exercise actually adjusted the base model output in the wrong direction. According to the
calibration algorithms, the values for variables which further describe test segment geometry and
characteristics result in correction factors that expect the base model to be under-estimating
collisions. This was found to be in error since the model always over-predicted actual collision
experience. It is also of note that the 2009 beta model generates results that are worse than the
earlier 2008 base model.
Proceedings of the 20th Canadian Multidisciplinary Road Safety Conference,
Niagara Falls, Ontario, June 6-9, 2010
e
Compte-rendu de la 20 Conférence canadienne multidisciplinaire sur la sécurité routière,
8
When the data are compared with MPB and MAD statistics along with the R2 value, it was found
that the models do not accurately predict the total number of collisions. This was reflected most
with the MAD values which range from 0.843 to 1.183 for annual collision frequencies. When
evaluating the data based on the number of collisions/mvkm travelled, the MPB values are
closest to 0 for the model when it is applied as un-calibrated. The R2 values ranging from 0.001
to 0.184 indicate that the collision prediction models contribute little beyond using a simple
across-the-board average,
The data synthesized in Figure 1 represent the estimated collision frequencies plotted against
the observed frequencies for the 2008 IHSDM model with no calibration. The top line in the
figure represents the 45-degree line that would occur if the model results perfectly reflected the
actual collision frequencies. The second line represents a best-fit linear regression line. It is
noteworthy that there are several large values predicted by the model compared to the relatively
small values representing the actual collision rates (bottom-right corner of plot area).
All Test Segments: 2008 (No Calibration)
6.0
5.0
Actual
4.0
y = 0.3043x + 0.7435
R² = 0.1844
3.0
2.0
1.0
0.0
0.0
1.0
2.0
3.0
4.0
5.0
6.0
Model
Figure 1: 2008 Model - All Test Segments (collisions/year)
Proceedings of the 20th Canadian Multidisciplinary Road Safety Conference,
Niagara Falls, Ontario, June 6-9, 2010
e
Compte-rendu de la 20 Conférence canadienne multidisciplinaire sur la sécurité routière,
9
Arterial Roads
The data in Tables 4 and 5 synthesize the average, MPB, MAD and R2 values for the 24 arterial
road segments that were tested. The ADTs for these roads ranged from under 1,000 vehicles
per day to over 10,000 vehicles per day.
Table 4 - Arterial Roads (collisions/year)
Avg. of all
arterials
MPB
MAD
R2
IHSDM Model Results
2008
2008
(Level 1)
(Level 2)
Actual
2008
No Calibration
1.481
1.871
2.417
2.391
2.192
-
0.391
0.887
0.010
0.937
1.144
0.010
0.910
1.125
0.010
0.711
0.979
0.015
2009
(beta)
Table 5 - Arterial Roads (collisions/mvkm)
IHSDM Model Results
Avg. of all
arterials
MPB
MAD
R2
Actual
2008
No Calibration
2008
(Level 1)
2008
(Level 2)
2009
(beta)
0.466
0.491
0.633
0.627
0.498
-
0.025
0.210
0.024
0.167
0.249
0.015
0.161
0.245
0.016
0.032
0.210
0.020
The results show that the arterial road segments exhibit the same behaviour as the data
synthesized in Tables 3 and 4. In the case of arterial roads the IHSDM model is shown to overpredict collisions/year by no less than 26 percent (2008 model with no calibration). When
expressed as collisions/mvkm, the model over predicts by at least 5 percent. Similar to Tables 3
and 4 the MPB and MAD for the collisions per year are higher than the collisions per million
vehicle kilometres with the collisions /mvkm displaying closer matches to the actual collision
rates. Again, very weak R2 values indicate poor model fit.
The data in Figure 2 illustrate the comparison between model output and actual observed
collision data for arterial roads only.
Proceedings of the 20th Canadian Multidisciplinary Road Safety Conference,
Niagara Falls, Ontario, June 6-9, 2010
e
Compte-rendu de la 20 Conférence canadienne multidisciplinaire sur la sécurité routière,
10
Arterial Roads: 2008 (No Calibration)
4.0
Actual
3.0
2.0
y = 0.0783x + 1.334
R² = 0.010
1.0
0.0
0.0
1.0
2.0
3.0
4.0
Model
Figure 2: 2008 Model – Arterial Roads (collisions/year)
Collector Roads
The data in Tables 6 and 7 summarize the average, MPB, MAD, and R2 values for the 51
collector roads that were tested. The ADTs for these roads ranged from under 1,000 vehicles
per day to over 10,000 vehicles per day.
Table 6 - Collector Roads (collisions/year)
IHSDM Model Results
Avg. of all
collectors
MPB
MAD
R2
Actual
2008
No Calibration
2008
(Level 1)
2008
(Level 2)
2009
(beta)
1.184
1.708
2.210
2.191
2.041
-
0.476
0.866
0.254
0.978
1.238
0.254
0.960
1.223
0.255
0.857
1.060
0.238
Proceedings of the 20th Canadian Multidisciplinary Road Safety Conference,
Niagara Falls, Ontario, June 6-9, 2010
e
Compte-rendu de la 20 Conférence canadienne multidisciplinaire sur la sécurité routière,
11
Table 7 - Collector Roads (collisions/mvkm)
IHSDM Model Results
Avg. of all
collectors
MPB
MAD
R2
Actual
2008
No Calibration
2008
(Level 1)
2008
(Level 2)
2009
(beta)
0.520
0.633
0.819
0.812
0.747
-
0.113
0.316
0.001
0.299
0.441
0.001
0.292
0.436
0.001
0.227
0.385
0.001
The data in Tables 6 and 7 exhibit the same behaviour as shown in the previous tables, with the
2008 un-calibrated model performing better than the calibrated and 2009 models. In terms of
collisions per year, the model gave a 44% difference between the estimated and actual crash
rates. Collisions /mvkm were over-estimated by at least 22%. In terms of MPB and MAD the
collisions/mvkm again showed the better results. All R2 values were very weak (0.001 to 0.255)
indicating that the model adds little value in estimating collisions.
Figure 3 illustrates the comparison between model output and actual observed collision data for
collector roads only.
Collector Roads ‐ 2008 (No Calibration)
6.0
5.0
Actual
4.0
y = 0.3456x + 0.5942
R² = 0.2544
3.0
2.0
1.0
0.0
0.0
1.0
2.0
3.0
4.0
5.0
6.0
Model
Figure 2: 2008 Model - Collector Roads (collisions/year)
Proceedings of the 20th Canadian Multidisciplinary Road Safety Conference,
Niagara Falls, Ontario, June 6-9, 2010
e
Compte-rendu de la 20 Conférence canadienne multidisciplinaire sur la sécurité routière,
12
Traffic Volume Sub-groups
In previous sections it was noted that five sub-groups of road segments were delineated for
varying levels traffic volumes. Modelling collision frequencies for roads in different volume bins
showed no improvement over previous results. Again, model estimates yielded very poor MPB,
MAD and R2 values.
It should be noted that in the data a trend was observed between each traffic volume sub-group
in that the maximum differences between the predicted and the actual collision rates would grow
larger as the ADT increased. The lowest absolute maximum difference between predicted
collisions/year and that observed was 1.016 in the under 1,000 ADT group and the largest
absolute maximum difference noted was 5.766 in the over 10,000 ADT group. This was only
taken in the average number of collisions per year as the ADT values were normalized when the
collision rates per million vehicle kilometres were used.
CONCLUSIONS
The collision prediction models embedded in IHSDM for two-lane rural roads were found to yield
very poor results for road segments in New Brunswick. Predicted collision frequencies were over
estimated by 38 to 78 percent (for all test segments combined) depending on the level of
calibration employed. R2 values that describe model fit ranged from 0.001 to 0.254 indicating
little added value by the model framework. Level 1 and level 2 calibration techniques typically
worsen the results over the base model. The beta version of the 2009 collision prediction model
produced poorer results than the 2008 model across-the-board. Application of the model to more
homogenized sub-groups of road segments (road class and volume bins) did not improve the
predictability of the model.
One of the main issues in the calibration of the model was the effort required to collect the
necessary data and combining the data into the appropriate sub-groups. The amount of time
and resources required to perform a full calibration for the entire province would be great. A
secondary issue with the model calibration is the accuracy of the collected data. The data may
not be up to date at the time of study and changes are generally made very year. The calibration
method was also seen as unsuccessful as the calibration factors are applied across all test
sections as one value. With a model that over-predicts and under-predicts fairly equally, very
little effect on the final results will occur.
Proceedings of the 20th Canadian Multidisciplinary Road Safety Conference,
Niagara Falls, Ontario, June 6-9, 2010
e
Compte-rendu de la 20 Conférence canadienne multidisciplinaire sur la sécurité routière,
13
REFERENCES
[1] SAITO, M., Chuo, K., Evaluation of the applicability of the Interactive Highway Safety Design
Model to safety audits of two-lane rural highways, Utah Department of Transportation
Report US-2-08, March 2008.
[2] FEDERAL HIGHWAY ADMINISTRATION, IHSDM website, www.ihsdm.org, 2009
[3] HARWOOD, D.W., Council, F.M., Hauer, E., Hughes, W.E., and Vogt, A., Prediction of the
expected safety performance of rural two-lane highways. FHWA Report FHWA-RD-99-207,
2000.
[4] TRANSPORTATION
RESEARCH
BOARD.
www.highwaysafetymanual.org, 2010.
Highway
safety
manual
website,
[5] DONNELL, E.T., Gross, F., Stodart, B.P., Opiela, K.S., Appraisal of the interactive highway
safety design model’s crash prediction and design consistency modules: case studies from
Pennsylvania, Journal of Transportation Engineering, Vol. 135, No. 2, pp. 62-72, February
2009.
[6] YE, H., Accident prediction for New Brunswick’s two-lane rural highways, Master’s Thesis,
University of New Brunswick, Fredericton, New Brunswick, September 2005.
[7] Oh, J., Lyon, C., Washington, S., Persaud, B., and Bared, J. Validation of FHWA crash
models for rural intersections. Transportation Research Record, Vol. 1840, pp. 41-49, 2003.
Proceedings of the 20th Canadian Multidisciplinary Road Safety Conference,
Niagara Falls, Ontario, June 6-9, 2010
e
Compte-rendu de la 20 Conférence canadienne multidisciplinaire sur la sécurité routière,
14
Was this manual useful for you? yes no
Thank you for your participation!

* Your assessment is very important for improving the work of artificial intelligence, which forms the content of this project

Download PDF

advertising