You are on page 1of 12

See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/335454223

Transportation Safety—Crash Prediction and Assessment

Conference Paper · August 2019


DOI: 10.1061/9780784482575.016

CITATIONS READS
0 57

3 authors:

Rijesh Karmacharya Sunanda Dissanayake


Kansas State University Kansas State University
3 PUBLICATIONS   0 CITATIONS    137 PUBLICATIONS   791 CITATIONS   

SEE PROFILE SEE PROFILE

Imalka Matarage
Kansas State University
10 PUBLICATIONS   2 CITATIONS   

SEE PROFILE

Some of the authors of this publication are also working on these related projects:

Road Safety View project

Calibrating the HSM Safety Prediction Tools for Kansas Freeways and Ramps View project

All content following this page was uploaded by Imalka Matarage on 18 September 2019.

The user has requested enhancement of the downloaded file.


International Conference on Transportation and Development 2019 154

Calibrating the Highway Safety Manual Crash Prediction Models for Urban and Suburban
Intersections in Kansas
Rijesh Karmacharya1; Sunanda Dissanayake, Ph.D., P.E.2; and Imalka C. Matarage3
1
Dept. of Civil Engineering, Kansas State Univ., 2118 Fiedler Hall, Manhattan, KS 66506. E-
mail: rijeshkarmacharya@ksu.edu
2
Dept. of Civil Engineering, Kansas State Univ., 2118 Fiedler Hall, Manhattan, KS 66506. E-
Downloaded from ascelibrary.org by Kansas State University Libraries on 08/30/19. Copyright ASCE. For personal use only; all rights reserved.

mail: sunanda@ksu.edu
3
Dept. of Civil Engineering, Kansas State Univ., 2118 Fiedler Hall, Manhattan, KS 66506. E-
mail: imalkam@ksu.edu

ABSTRACT
The crash prediction models provided in the Highway Safety Manual predict crashes using
traffic and geometric data for various roadway facilities, which are incorporated through safety
performance functions and crash modification factors. In this study, calibration factors were
estimated for four types of intersections in urban and suburban locations in Kansas: 3-legged
signalized intersections (3SG), 4-legged signalized intersections (4SG), 3- legged unsignalized
intersections with stop control on the minor approach (3ST), and 4-legged unsignalized
intersections with stop control on the minor approach (4ST). The calibration factors were
estimated as 0.64 for 3SG, 0.51 for 3ST, 1.17 for 4SG, and 0.61 for 4ST. The calibration factors
show that the HSM methodology underpredicted crashes for 4SG, and overpredicted crashes for
the other three intersection types. The reliability of the calibration factors was checked with the
help of cumulative residual plots and coefficient of variation. Furthermore, calibration functions
were developed, and their reliability was examined. The results showed that the calibration
functions have better reliability as compared to calibration factors for all intersection types.
Keywords: calibration, urban intersections, calibration factor, calibration functions

INTRODUCTION
In 2010, the American Association of State and Highway Transportation Officials
(AASHTO) published the Highway Safety Manual (HSM), which contains methodologies to
predict crashes on various types of facilities in a roadway network. In addition, the HSM also
provides guidelines to develop state-specific Safety Performance Functions (SPFs) using local
data to increase the accuracy of the predictive models (AASHTO 2010).
Among the various facility types covered by the HSM, the predictive methodology for
intersections at urban and suburban locations are covered in Part C, Chapter 12. The calibration
methodology is presented in Part C, Appendix A of the HSM. The HSM methodology allows
users to predict crashes using the Safety Performance Functions (SPFs), which accounts for the
traffic characteristics, and the Crash Modification Factors (CMFs) which account for the
geometric characteristics of a given facility type. Accordingly, the objectives of this study were,
1. To estimate calibration factors for 3- and 4-legged intersections at urban and suburban
locations in Kansas,
2. To assess the reliability of the estimated calibration factors, and
3. To develop calibration functions to increase the accuracy of crash prediction, and check
the reliability of the developed calibration functions.

© ASCE

International Conference on Transportation and Development 2019


International Conference on Transportation and Development 2019 155

LITERATURE REVIEW
The HSM provides different crash prediction models to predict crashes for various facility
types. The crash prediction is carried out using regression models for SPFs and the SPFs are
multiplied by CMFs when the sites deviate from the base conditions (AASHTO 2010). SPFs are
mathematical equations, which relate the number of crashes to the traffic and geometric
characteristics (Srinivasan et al. 2013). A summary of studies relevant to this research are
Downloaded from ascelibrary.org by Kansas State University Libraries on 08/30/19. Copyright ASCE. For personal use only; all rights reserved.

presented in this section.

Studies relevant to calibration of HSM SPFs


Shin, Lee, and Dadvar (2014) estimated Local Calibration Factors (LCFs) for 18 facility
types in Maryland at urban and rural locations. The observed crashes were obtained from the
Maryland State Police database for a three-year period (2008-2010), and the predicted crashes
were estimated following the prediction methodology provided in the HSM. For urban and
suburban intersections in Maryland, the LCFs ranged from 0.1562 to 0.4782, showing that the
HSM methodology overpredicted crashes in Maryland.
The calibration of the HSM predictive methodology in Missouri estimated calibration factors
as 1.06 for 3ST, 3.03 for 3SG, 1.30 for 4ST, and 4.91 for 4SG (Sun et al, 2013). For the study
period of 2009 to 2011, the calibration factors show that the HSM methodology underpredicted
the number of crashes at urban intersections in Missouri. The calibration factors estimated for
Ohio for the same study period of 2009 to 2011, and for the same facility types, show
underprediction of crashes as well. For the intersections at urban and suburban locations, the
estimated calibration factors ranged from 1.34 to 3.71 (Troyer, Bradbury, and Juliano, 2015).
The calibration factors for urban intersections in North Carolina also underpredicted the number
of crashes (Srinivasan, and Carter, 2011). Srinivasan et al. (2011) estimated calibration factors
for various facility types in Florida for the period of 2005 to 2009. For signalized intersections at
urban locations, calibration factors were developed for individual years, with all the values
greater than 1.00, showing that the HSM methodology underpredicted the number of crashes. In
case of unsignalized intersections, calibration factors were not developed due to the lack of
sufficient data.
The calibration factors for urban intersections in Oregon estimated calibration factors ranging
from 0.35 to 1.10. For the study period of 2004 to 2006, the estimated calibration factors show
the HSM methodology underpredicted as well as overpredicted crashes at urban and suburban
intersections in Oregon (Dixon et al., 2012).
The calibration factors for different states show that the HSM crash prediction models
underpredicts as well as overpredicts crashes for the same facility type due to the difference in
the traffic and geometric conditions. The calibration factors developed for one jurisdiction might
not be a good representative for a different jurisdiction.

Selection of appropriate sample size


The HSM recommendation for sample size for calibration is 30 to 50 sites, which have
collectively experienced at least 100 crashes per year (AASHTO, 2010). However, studies have
shown that a larger sample size results in a higher level of accuracy of the estimated calibration
factor.
Shirazi et al. (2016) suggested a range of sample size with relation to the coefficient of
variance (CV) of crash data. The sample size ranged from 30 to 1,500 for three levels of

© ASCE

International Conference on Transportation and Development 2019


International Conference on Transportation and Development 2019 156

confidence, 70%, 80%, and 90%, for a range of CV. The results show that the HSM
recommendation might not be sufficient to achieve high confidence levels. Shin et al. (2014)
adopted a sample size based on 90% level of confidence for determining sample size in
Maryland. Initially, 30 sites were selected, and the sample size was increased for the 90%
confidence level. A larger sample size would increase the confidence level but would not be
feasible given the time constraints. For the estimated calibration factors for highway segments in
Washington, the quality of the calibration factors were assessed. The results showed that the
Downloaded from ascelibrary.org by Kansas State University Libraries on 08/30/19. Copyright ASCE. For personal use only; all rights reserved.

HSM recommended 30-50 sample size did not produce satisfactory results for all the facility
types (Banihashemi, 2012).
These studies show that the HSM recommendation might not be able to provide a reliable
calibration factor, with larger sample size giving higher reliability. Bahar and Hauer (2014)
provide five alternatives to estimate sample size to achieve a desired level of accuracy. The
variance or standard deviation of the estimated calibration factor, and agency specific data are
the factors based on which an appropriate sample size can be determined.

Minor street Average Annual Daily Traffic data estimation models


The Average Annual Daily Traffic (AADT) for the minor approach is a required data
element for the prediction and calibration methodologies. Several studies have adopted various
methods for estimating minor street AADT whenever the AADT data is not available. Among
these various methods, multiple linear regression modeling is the most common method.
Shin et al. (2014) developed multiple regression models to estimate minor street AADT in
Maryland. For the selection of the best available model, R-squared, adjusted R-squared, leaps
and bounds, Akaike Information Criterion, and Bayesian Information Criterion were used. The
models with R-squared values greater than 0.50 were deemed a good fit. In Oregon, Dixon et al.
(2012) developed multiple linear regression models with log 10 transformation of the continuous
variables to negate the unequal variances. The R-squared value for the two final models were
greater than 0.60, which were acceptable. Similarly, Mohamad, Sinha, Kuczek, and Scholer
(1998) developed AADT prediction models for county roads in Indiana, with log 10
transformation of the continuous variables. Cross validation was done for model validation, with
errors of 16.78% on average.
These studies provide a wide range of variables for developing regression models for the
estimation of minor street AADT. In addition, different methods for model validation and
selection of variables are also used.

Calibration functions, and assessment


Srinivasan, Carter, and Bauer (2013) provided a decision-making procedure for calibrating
SPFs or developing jurisdiction specific SPFs. Using the methods provided for assessing the
calibration factors, SPFs can be developed as the last step if reliability of the calibration factors
is not satisfactory. Srinivasan, and Bauer (2013) have provided a guide for developing
jurisdiction-specific SPFs, which shows that this process is data extensive and requires a
significant amount of time. Developing calibration functions could help increase the accuracy of
crash prediction, as compared to calibration factors.
Hauer (2016) provides a detailed methodology in developing calibration functions using
Microsoft Excel Solver add-in. Srinivasan et al. (2016) developed calibration functions for
predicting crashes in Arizona using three methods, Ordinary Least Squares, Poisson Regression,
and Negative Binomial Regression. The results showed that the calibration functions had good

© ASCE

International Conference on Transportation and Development 2019


International Conference on Transportation and Development 2019 157

reliability. To identify the models that resulted in a higher accuracy in crash prediction, Claros,
Sun, and Edara (2018) compared calibration factors, calibration functions, and jurisdiction-
specific models. To analyze the calibration functions, CURE plots, overdispersion parameter,
and log-likelihood were used. Comparing the calibration functions with calibration factors by
ranges, the calibration functions had a lower accuracy.
The quality of the developed calibration functions and estimated calibration factors can be
assessed using cumulative residual (CURE) plots and coefficient of variation (CV). If the percent
Downloaded from ascelibrary.org by Kansas State University Libraries on 08/30/19. Copyright ASCE. For personal use only; all rights reserved.

deviation of the residuals is less than 5%, or if the CV is less than 0.15, the calibration factors
and calibration functions are reliable (Lyon, Persaud, and Gross, 2016). For assessing the
calibration factors and interpreting the results, several studies have used CURE plots (Troyer et.
al, 2015; Claros et al., 2018; Persaud and Lyon,inc., and Felsburg Holt & Ullevig, 2009; Smith et
al., 2017).
In summary, CURE plots and CV values can be used for the assessment of the estimated
calibration factors and developed calibration functions. If the calibration functions have better
reliability in crash prediction, then developing calibration functions can be an alternative to
developing SPFs.

Table 1. Data elements required for HSM predictive methodology


Data Elements 3ST 3SG 4ST 4SG
AADT for major leg    
AADT for minor leg    
Number of lanes crossed by a
X  X 
pedestrian on any approach
Pedestrian volume X  X 
Presence of left-turn lanes    
Type of left-turn phasing X  X 
Presence of right-turn lanes    
Use of right-turn-on-red signal X  X 
Use of red-light cameras X  X 
Presence of bus stops within
X X X 
1,000 ft.
Presence of schools within
X X X 
1,000 ft.
Presence of alcohol sales
X X X 
establishment within 1,000 ft.
 = required or desired data element
X = not required for the identified facility type

DATA COLLECTION AND SAMPLING


The study period for this study was considered as a three-year period, 2013 to 2015 for 3SG,
3ST, and 4SG, and 2014 to 2016 for 4ST, based on the availability of the most recent data at the
time of the study. The crash data were obtained from the Kansas Crash & Analysis Reporting
System (KCARS) database. AADT data were obtained from the Federal Highway
Administration (FHWA) website (https://www.fhwa.dot.gov/policyinformation/hpms

© ASCE

International Conference on Transportation and Development 2019


International Conference on Transportation and Development 2019 158

/shapefiles.cfm) which was extracted using ArcGIS (ESRI, 2012). Other data elements were
obtained using Google Earth and an online application, the KML Circle Generator
(http://kml4earth.appspot.com/circlegen.html). The data elements required for the HSM
predictive methodology are given in Table 1.
The number of sites for individual facility types were selected for 90% level of confidence.
The sample sites were selected randomly from urban cities in Kansas, with the HSM definition
of urban areas being areas with population 5,000 or more.
Downloaded from ascelibrary.org by Kansas State University Libraries on 08/30/19. Copyright ASCE. For personal use only; all rights reserved.

METHODOLOGY
The calibration procedure is given in Part C, Appendix A of the HSM first edition, and the
predictive methodology of urban and suburban intersections is given in Part C, Chapter 12 of the
HSM (AASHTO, 2010). The calibration factor is calculated using equation 1.

Calibration Factor  Ci  
 Observed Crashes Eq. 1
 Predicted Crashes
The crashes are predicted using the following equations:
Npredicted int  Ci   Nbi  Npedi  Nbikei  Eq. 2
Nbi  Nspf int   CMF1i CMF2i  CMF6i  Eq. 3
Where,
Npredicted int = predicted average crash frequency of an intersection for the selected year;
Nbi= predicted average crash frequency of an intersection (excluding vehicle-pedestrian and
vehicle-bicycle collisions);
Nspf int = predicted total average crash frequency of intersection-related crashes for base
conditions (excluding vehicle-destrian and vehicle-bicycle collisions);
Npedi = predicted average crash frequency of vehicle-pedestrian collisions;
Nbikei = predicted average crash frequency of vehicle-bicycle collisions;
CMF1i…CMF6i = crash modification factors for intersections; and
Ci = calibration factor.
The SPFs are used to predict crashes for four types of collisions; multiple-vehicle collisions,
single-vehicle crashes, vehicle-pedestrian collisions, and vehicle-bicycle collisions. In addition to
the SPFs, nine CMFs are used as multiplicative factors to SPFs to address the changes from the
base conditions. The SPFs for the four collision types to predict crashes are given by the
following equations:
 Multiple-vehicle collisions and single-vehicle crashes

Nbimv  exp a m  bm  ln  AADTmaj   cm  ln  AADTmin   Eq. 4


Nbisv  exp a s  bs  ln  AADTmaj   cs  ln  AADTmin   Eq. 5
Where,
Nbimv = predicted number of multiple-vehicle collisions,
Nbisv = predicted number of single-vehicle crashes,
AADTmaj = average daily traffic volume (vehicles/day) for major approach (both directions
of travel combined),
AADTmin = average daily traffic volume (vehicles/day) for minor approach (both directions
of travel combined),
am, bm, cm = regression coefficients for multiple-vehicle collisions,

© ASCE

International Conference on Transportation and Development 2019


International Conference on Transportation and Development 2019 159

as, bs, cs = regression coefficients for single-vehicle crashes.


 Vehicle-pedestrian collisions
Npedi  Nbi  f pedi Eq. 6
Where,
Npedi = predicted number of vehicle-pedestrian crashes, Nbi is calculated using Equation 3.
fpedi = pedestrian crash adjustment factor.
 Vehicle-bicycle collisions
Downloaded from ascelibrary.org by Kansas State University Libraries on 08/30/19. Copyright ASCE. For personal use only; all rights reserved.

Nbikei  Nbi  f bikei Eq. 7


Where,
Nbikei = predicted number of vehicle-bicycle collisions, Nbi is calculated using Equation 3.
fbikei = bicycle crash adjustment factor.

Minor AADT estimation models for 3ST


Multiple linear regression models were used to develop estimation models for AADT data
for the minor approaches for 3SG and 3ST facility type for which the actual AADT data were
not available. The general form of a multiple regression equation is given by equation 8.
Y  βo  β1X1  β2 X2    βn Xn  ε Eq. 8
Where,
Y = dependent variable,
X1…Xn = independent variables,
βo…βn = parameter estimates, and
ε = error.
For different combinations of independent variables, multiple linear regression models were
developed using data of intersections having available minor street AADT data. The regression
models were developed using SAS 9.4 (SAS Institute Inc.). The dependent variable for the
regression model was the AADT for the minor approach. The independent variables considered
for the regression models were the AADT for the major approach, the area of the city and
county, the population, per capita income, median age, and population per household of the city.
In addition to the demographic data, other independent variables were the number of left and
right turn lanes in the major and minor approach, the type of left turn signal, the number of
through lanes, and the speed and functional class of the major and minor approach. Natural log
and log10 transformation of the continuous variables were tested in developing the regression
models. The final model for 3SG is given by equation 9 and final model for 3ST is given by
equation 10.
Ln  minorAADT   1.40  0.70 ln  majorAADT   0.26 LtMinor  0.52 RTmajor –
0.41 LtSigMaj  0.33 LtSigMin – 0.03 SLmajor  0.02 SLminor Eq. 9
 0.57 FcMajAr  0.90 FcMajCl
log10  minorAADT    1.73  0.55 log10  majorAADT   0.96 log10  PcInc  – 0.86
log10  MedAge   0.96 log10  PplHH   0.07 LTmajor  0.15 Eq. 10
ThLanes  0.14 Mi_fc_ar  0.26 Mi _ fc _ cl
Where,
minorAADT = AADT of the minor approach (veh/day),
majorAADT = AADT of the major approach (veh/day),

© ASCE

International Conference on Transportation and Development 2019


International Conference on Transportation and Development 2019 160

PcInc = Per capita income of the city (in dollars),


MedAge = Median Age of the city (years),
PplHH = Number of people in a household,
LTmajor = Number of left turn lanes in the major approach,
ThLanes = Number of through lanes in the major approach,
RTmajor = number of right turn lanes in the major approach,
LTminor = number of left turn lanes in the minor approach,
Downloaded from ascelibrary.org by Kansas State University Libraries on 08/30/19. Copyright ASCE. For personal use only; all rights reserved.

LtSigMaj = Type of left turn signal in the major approach,


LtSigMin = Type of left turn signal in the minor approach,
SLmajor = speed limit of the major approach in mph,
SLminor = speed limit of the minor approach in mph.
Mi_fc_ar = Functional class of minor approach, 1 if arterial, else 0,
Mi_fc_cl = Functional class of minor approach, 1 if collector, else 0,
FcMajAr = If the major approach is an arterial, value = 1, else 0, and
FcMajCl = If the major approach is a collector, value = 1, else 0.
Forward, stepwise, and backward selection methods were used at 5% level of significance for
variable selection. For model selection, the criteria used were R-squared, Akaike Information
Criterion (AIC), Bayesian Information Criterion (BIC), and Mallow's C p. The R-squared value
for the regression model in equation 9 was 0.5314, and for equation 10 was 0.3218. Data from
the year 2012 were used for model validation by assessing the mean error between the predicted
and the actual minor AADT. The mean error for model for 3SG was 43.95% and error for model
for 3ST was 59.74%.

Calibration Functions
The general form of a calibration function is given by equation 11.
N  a * NHSM_pred b Eq. 11
Where,
N = calibrated predicted crashes using the calibration function,
NHSM_pred = predicted number of crashes using the HSM methodology, and
a, b = regression constants.
The calibration function was developed using Negative Binomial (NB) regression, for which
the Log-Likelihood (LL) function is given by equation 12. Using Microsoft Excel Solver add-in,
calibration functions were developed with the desired result of the LL as a maximum possible
value for the given data.
1 1 1
LL NB  ln(X  ) – 1/ k *ln ( )  Xln  y  – (  X)ln ( y) Eq. 12
k k k
Where,
k = over-dispersion parameter,
X = observed crashes, and
y = fitted predicted crashes.

Evaluation of Calibration Factors


For the evaluation of the estimated calibration factors and calibration functions, Cumulative
Residual (CURE) plots were developed and analyzed (Troyer, Bradbury, and Juliano, 2015;
Lyon, Persaud, and Gross, 2016). The assessment of the developed calibration factors is based on

© ASCE

International Conference on Transportation and Development 2019


International Conference on Transportation and Development 2019 161

two criteria (Lyon, Persaud, and Gross, 2016),


 If the fitted SPF is within five percent of threshold of 2 Standard Deviations, then the
SPF is satisfactorily calibrated to the specific jurisdiction data.
 If the first criterion is not met, the calibration factors are still acceptable if the Coefficient
of Variation is less than 0.15.
CURE plots were developed for calibration factors and calibration functions when
considering fatal-and-injury only crashes, and all crashes. The calibration factors and functions
Downloaded from ascelibrary.org by Kansas State University Libraries on 08/30/19. Copyright ASCE. For personal use only; all rights reserved.

were assessed for the reliability as well.

RESULTS
A calibration factor less than 1.00 indicates that the HSM methodology overpredicts the
number of crashes and value greater than 1.00 indicates underprediction of crashes. Considering
only FI crashes at 4SG intersections, the calibration factor was estimated as 2.00, which means
the number of predicted crashes was half the number of actual crashes. However, the number of
predicted FI crashes at 3SG was almost twice the actual number of crashes, for which the
calibration factor was estimated to be 0.52.
The calibration factors for urban and suburban intersections in Kansas estimated using the
HSM predictive methodology and calibration procedure are presented in Table 2.

Table 2. Estimated calibration factors for urban and suburban intersections in Kansas
Observed Predicted Calibration
Facility Type Sample Size
Crashes crashes Factor
3ST (FI crashes) 234 95 234.58 0.40
3ST (all crashes) 234 321 625.07 0.51
3SG (FI crashes) 89 89 170.56 0.52
3SG (all crashes) 89 310 481.36 0.64
4ST (FI crashes) 167 153 211.73 0.73
4ST (all crashes) 167 352 577.74 0.61
4SG (FI crashes) 198 956 475.88 2.00
4SG (all crashes) 198 1644 1400.49 1.17

The calibration factors for 3ST, 3SG and 4ST show that the HSM methodology
overpredicted the number of crashes in Kansas. However, crashes at 4SG intersections were
underpredicted by the HSM methodology. Using the same methodology for the same
jurisdiction, crashes were underpredicted as well as overpredicted for different facility types.
For 3ST and 3SG intersections, regression models were used to estimate minor street AADT
data for some intersections. The calibration factors for these two intersection types based on sites
with actual data and sites with estimated data are given in Table 3.
The calibration factors in Table 3 shows that the factors for set of intersections with
estimated minor street AADT data were smaller than the calibration factors for intersections with
actual data. The calibration factors indicated that the overprediction of crashes was more for set
of intersections with estimated minor street AADT data, which could be a result of the errors of
the regression models.
The estimated calibration factors were assessed for their reliability and calibration functions
were developed for all the intersection types. The reliability of the calibration functions was also
checked. The results are given in Table 4.

© ASCE

International Conference on Transportation and Development 2019


International Conference on Transportation and Development 2019 162

Table 3. Calibration factors for set of intersections with actual and estimated minor street
AADT data
Facility Sample Observed Predicted Calibration
Description
Type Size Crashes crashes Factor
Intersections with actual data
128 57 115.60 0.49
(FI)
Intersections with estimated
Downloaded from ascelibrary.org by Kansas State University Libraries on 08/30/19. Copyright ASCE. For personal use only; all rights reserved.

106 38 118.98 0.32


data (FI)
3ST
Intersections with actual data
128 180 319.93 0.56
(all crashes)
Intersections with estimated
106 141 305.13 0.46
data (all crashes)
Intersections with actual data
68 75 133.11 0.56
(FI)
Intersections with estimated
21 14 37.45 0.37
data (FI)
3SG
Intersections with actual data
68 256 376.12 0.68
(all crashes)
Intersections with estimated
data (all crashes) 21 54 105.24 0.51

Table 4: Calibration functions, and assessment of calibration factors and calibration


functions
Calibration Functions (developed using NB
Facility Type Calibration Factor
Regression)
(crashes)
C.F. CV CURE deviation a b k CV CURE deviation
3ST (FI) 0.40 0.15 61 % 0.24 0.47 0.00 0.15 0%
3ST (all) 0.51 0.08 93 % 0.51 0.46 0.00 0.08 0%
3SG (FI) 0.52 0.16 3% 0.52 0.97 0.00 0.16 3%
3SG (all) 0.64 0.11 15 % 0.77 0.76 0.11 0.11 3%
4ST (FI) 0.73 0.12 17 % 0.60 0.77 0.00 0.12 13 %
4ST (all) 0.61 0.08 75 % 0.66 0.58 0.00 0.08 0%
4SG (FI) 2.00 0.08 15 % 2.00 0.93 0.54 0.08 4%
4SG (all) 1.17 0.06 80% 0.94 1.22 0.16 0.06 5%

The percent CURE deviation for calibration factors in Table 4 shows that the estimated
factors were not reliable, apart from calibration factor for 3SG intersections considering FI
crashes only. For the developed calibration functions, the CURE deviation shows that the
functions were reliable as the deviation is within 5% threshold. However, for 4ST intersections
considering FI crashes only, the calibration functions were reliable even though the CURE
deviation was not within acceptable limits since the CV value was less than 0.15. Figure 1 shows
a CURE plot for calibration factor and calibration function for 4SG intersections for all crashes.
The CURE plots in Figure 1(a) shows major portion lied outside the threshold. For the same
facility type, CURE plot for calibration functions showed that the majority of the cumulative

© ASCE

International Conference on Transportation and Development 2019


International Conference on Transportation and Development 2019 163

residuals lied inside the two standard deviations threshold.


Downloaded from ascelibrary.org by Kansas State University Libraries on 08/30/19. Copyright ASCE. For personal use only; all rights reserved.

(a) (b)
Figure 1. CURE plots for 4SG intersections (all crashes) for (a) calibration factors, and (b)
calibration functions
CONCLUSIONS
The estimated calibration factors show that the HSM methodology overpredicted the number
of crashes for 3ST, 3SG and 4ST intersections, and underpredicted the number of crashes for
4SG intersections. One of the reasons for this deviation could be the differences in the traffic
patterns and characteristics of Kansas as compared to the states whose data were used to develop
the prediction models in the HSM. The property-damage-only crash reporting threshold in
different states could also be one of the reasons for this situation. To increase the accuracy of
crash prediction, Kansas-specific SPFs can be developed using recent data from Kansas.
However, developing SPFs requires an extensive database and is time consuming. As the results
from the assessment of calibration factors and calibration functions using goodness-of-fit tests
showed that the calibration functions are reliable, calibration functions may be used instead of
developing new SPFs. In addition, calibration factors and calibration functions can also be
developed for AADT ranges, which could increase the accuracy of crash prediction. The
estimated calibration factors and developed calibration functions are recommended to be
recalibrated every two to three years, as recommended by the HSM (AASHTO, 2010). The
recalibration would help update the change in traffic and geometric characteristics, thus
increasing the crash prediction accuracy and the reliability.

REFERENCES
AASHTO (2010). Highway Safety Manual. American Association of State Highway
Transportation Officials, Washington DC.
ArcGIS shapefiles. Retrieved from
https://www.fhwa.dot.gov/policyinformation/hpms/shapefiles.cfm. Accessed July 2017.
Bahar, G., & E. Hauer (2014). User’s guide to develop highway safety manual safety
performance function calibration factors. National Cooperative Highway Research Program,
Transportation Research Board, Washington, DC.
Banihashemi, M. (2012). Sensitivity analysis of data set sizes for Highway Safety Manual
calibration factors. Transportation Research Record: Journal of the Transportation Research
Board, 2279, 75-81.

© ASCE

International Conference on Transportation and Development 2019


International Conference on Transportation and Development 2019 164

Claros, B., Sun, C., & Edara, P. (2018). HSM calibration factor, calibration function, or
jurisdiction-specific safety model–A comparative analysis. Journal of Transportation Safety
& Security, 1-20.
Dixon, K., Monsere, C., Xie, F., & Gladhill, K. (2012). Calibrating the Highway Safety Manual
Predictive Methods for Oregon Highways. Publication FHWA-OR-RD-12-07. FHWA, US
Department of Transportation.
ESRI. (2012). ArcGIS Desktop: Release 10.1: Redlands, CA: Environmental Systems Research
Downloaded from ascelibrary.org by Kansas State University Libraries on 08/30/19. Copyright ASCE. For personal use only; all rights reserved.

Institute.
Google Earth. http://www.google.com/earth. Accessed May 2018.
Hauer, E. (2016). The art of regression modeling in road safety. New York: Springer.
Lyon, C., B. Persaud, & F. Gross (2016). The Calibrator- An SPF Calibration and Assessment
Tool User Guide. Report No. FHWA-SA-17-016, FHWA, U.S. Department of
Transportation.
Mohamad, D., Sinha, K., Kuczek, T., & Scholer, C. (1998). Annual average daily traffic
prediction model for county roads. Transportation Research Record: Journal of the
Transportation Research Board, 1617, 69-77.
Persaud and Lyon, Inc., & Felsburg Holt & Ullevig (2009). Safety performance functions for
intersections. Report No. CDOT-2009-10. Colorado Department of Transportation.
SAS Institute Inc., SAS 9.4, Cary, NC. SAS Institute Inc: 2013.
Shin, H., Lee, Y. J., & Dadvar, S. (2014). The development of local calibration factors for
implementing the Highway Safety Manual in Maryland. Report No. MD-14-SP209B4J.
Maryland State Highway Administration.
Shirazi, M., Lord, D., & Geedipally, S. R. (2016). Sample-size guidelines for recalibrating crash
prediction models: recommendations for the Highway Safety Manual. Accident Analysis &
Prevention, 93, 160-168.
Srinivasan, R., & Bauer, K. (2013). Safety performance function development guide: Developing
jurisdiction-specific SPFs. Report no. FHWA-SA-14-005. FHWA, U.S. Department of
Transportation.
Srinivasan, R., & Carter, D. (2011). Development of safety performance functions for North
Carolina. Report No. FHWA/NC/2010-09. North Carolina Department of Transportation.
Srinivasan, R., Carter, D., & Bauer, K. (2013). Safety performance function decision guide: SPF
Calibration vs SPF development. Report No. FHWA-SA-14-004. FHWA, U.S. Department
of Transportation.
Srinivasan, R., Colety, M., Bahar, G., Crowther, B., & Farmen, M. (2016). Estimation of
calibration functions for predicting crashes on rural two-lane roads in Arizona.
Transportation Research Record: Journal of the Transportation Research Board, 2583, 17-
24.
Srinivasan, S., Haas, P., Dhakar, N. S., Hormel, R., Torbic, D., & Harwood, D. (2011).
Development and calibration of highway safety manual equations for Florida conditions.
Report No. TRC-FDOT-82013-2011. Florida Department of Transportation.
Sun, C., Brown, H., Edara, P. K., Claros, B., & Nam, K. (2013). Calibration of the highway
safety manual for Missouri. Report No. cmr14-007. Missouri Department of Transportation.
Troyer, D., Bradbury, K., & Juliano, C. (2015). Strength of the Variable: Calculating and
Evaluating Safety Performance Function Calibration Factors for the State of Ohio.
Transportation Research Record: Journal of the Transportation Research Board, 2515, 86-
93.

© ASCE

View publication stats International Conference on Transportation and Development 2019

You might also like