trb 2013 annual meeting paper revised from original...
TRANSCRIPT
Downloaded from amonline.trb.orgDownloaded from amonline.trb.orgDownloaded from amonline.trb.orgDownloaded from amonline.trb.orgDownloaded from amonline.trb.orgDownloaded from amonline.trb.orgDownloaded from amonline.trb.orgDownloaded from amonline.trb.orgDownloaded from amonline.trb.orgDownloaded from amonline.trb.orgDownloaded from amonline.trb.orgDownloaded from amonline.trb.orgDownloaded from amonline.trb.orgDownloaded from amonline.trb.org
Crash Frequency Modeling for Highway Construction Zones 1
2
Ozgur Ozturk (Corresponding Author) 3 Graduate Research Assistant 4
Rutgers Intelligent Transportation Systems (RITS) Lab 5
Department of Civil and Environmental Engineering 6
Rutgers, The State University of New Jersey 7
623 Bowser Road, Piscataway, NJ 08854, USA 8
Tel: (732) 445-5496, Fax: (732) 445-0577 9
Email: [email protected] 10
11
Kaan Ozbay, Ph.D. 12 Professor & Director 13
Rutgers Intelligent Transportation Systems (RITS) Lab 14
Department of Civil and Environmental Engineering 15
Rutgers, The State University of New Jersey 16
96 Frelinghuysen Road, Piscataway, NJ 08854, USA 17
Tel: (732) 445-2792 Fax: (732) 445-0577 18
Email: [email protected] 19
20
Hong Yang, Ph.D. 21 Post-Doctoral Associate 22
Rutgers Intelligent Transportation Systems (RITS) Lab 23
Department of Civil and Environmental Engineering 24
Rutgers, The State University of New Jersey 25
623 Bowser Road, Piscataway, NJ 08854, USA 26
Tel: (732) 445 0579 ex.127 Fax: (732) 445-0577 27
Email: [email protected] 28
29
Bekir Bartin, Ph.D. 30 Research Associate 31
Rutgers Intelligent Transportation Systems (RITS) Lab 32
Department of Civil and Environmental Engineering 33
Rutgers, The State University of New Jersey 34
623 Bowser Road Piscataway, NJ 08854, USA 35
Tel: (732) 445-5496 Fax: (732) 445-0577 36
E-mail: [email protected] 37
38
39
40
41
42
Word count: 4683 Texts + 4 Figures + 6 Tables=7183 43
Abstract: 217 44
Submission Date: August 1, 2012 45
46
47
48
TRB Paper ID: 13-4555 49
Transportation Research Board’s 92nd Annual Meeting, Washington, D.C., 2013 50
TRB 2013 Annual Meeting Paper revised from original submittal.Downloaded from amonline.trb.org
Ozturk, Ozbay, Yang and Bartin 2
ABSTRACT 1 Throughout the country, work zone safety issues have received considerable attention in recent years due 2
to increasing work zone crashes along numerous highway renovation and reconstruction projects. In 3
general, previous studies were able to consider limited number of contributing factors mainly due to the 4
limitations in data availability. The main goal of this paper is to remedy this major problem related to data 5
availability and estimate improved models using data from multiple sources. Work zone project drawings, 6
crash database and straight line diagrams are used to create an integrated work zone safety database. 7
Work zone crash data is plotted in time and space to validate, locate and adjust work zone related 8
information. The negative binomial regression approach is used as the appropriate model to predict crash 9
frequency within these work zones. Traffic volume is adjusted for daytime and night time conditions in 10
terms of hourly distribution of daily traffic. The duration-based and period-based models are also 11
developed to address relationship between potential factors and to predict crash frequency on work zones 12
in terms of property damage only (PDO) and injury crashes. Compared with previous frequency models, 13
additional parameters such as number of lanes closed and speed reduction are used. These additional 14
factors identified as significant can help traffic engineers to further improve safety of work zone projects. 15
TRB 2013 Annual Meeting Paper revised from original submittal.Downloaded from amonline.trb.org
Ozturk, Ozbay, Yang and Bartin 3
INTRODUCTION 1 Work zone safety has received much attention in the recent years due to numerous highway renovation 2
projects that resulted in many work zone crashes. There were 87,606 nationwide work zone crashes 3
occurred in 2010 and 37,476 of which were an injury crashes, resulting approximately one injury work 4
zone crash at every 14 minutes (1). 5
Similar statistics are observed at the state level. New Jersey (NJ) has been called the corridor 6
state, an apparent consequence of its perceived role and proximity to the strong markets in New York 7
City, Philadelphia and the Boston-Washington Northeast Corridor. It experiences the second highest 8
travel delay and total congestion cost in the nation (2). Moreover, its aging infrastructure requires regular 9
maintenance and replacement. About half of the state-maintained roads are in deficient condition (3). 10
Furthermore, NJ is the state with the highest capital and bridge disbursement and the maintenance 11
disbursement per mile (4). Accordingly, an increase is expected in number of work zone projects and 12
work zone crashes as well. In 2010, NJ crash statistics shows this increase distinctly for work zone 13
labeled crashes. Yearly average number of work zone crashes between 2004 and 2009 is 5,395; however 14
this number increased in 2010 by 26.7 percent to 6,837 crashes equal to 7.8 percent of the nationwide 15
total number of work zone crashes (5). Thus, there is a need to study the specific factors that contribute to 16
work zone crash frequencies in NJ. 17
The majority of the work zone crash studies focused on severity analysis and the related 18
contributing factors. There are only a few studies that directly addressed work zone crash frequency 19
modeling (6-11). Negative binomial regression method is generally used to model work zone crash 20
frequency with model parameters such as annual average daily traffic (AADT), duration, length, work 21
zone type and urban indicator to find contributory factors related to work zone crashes. Some of these 22
studies reported limitations in terms of data. By using different available data sources, an enhanced model 23
can be developed for work zone crash frequency by accounting more contributing factors. 24
The main objective of this study is to develop an improved work zone crash frequency model by 25
incorporating additional important parameters such as speed reduction, the number of lanes closed, road 26
systems and number of intersections. These parameters are not generally included in the work zone crash 27
frequency models proposed by previous studies (e.g. 7, 9). Seasonally and hourly adjusted directional 28
AADT values are also used in the model. NJ crash database (2004-2010), project drawings and NJ 29
straight line diagrams are combined to create an integrated database. Crash frequency is investigated for 30
injury and property damage only (PDO) crashes for day and night time conditions separately. In the 31
duration-based model, the number of total crashes is taken into account for the work zone duration, and in 32
the period-based model three monthly crash counts are used to set models. Poisson and negative binomial 33
regressions are employed to model crash count data. 34
This paper is organized as follows. Next section provides a comprehensive review of the 35
literature related to crash frequency modeling of the work zones. The following section describes the 36
integrated database used for modeling and presents the model specification. The last section presents 37
estimation results and summarizes major findings. 38
39
LITERATURE REVIEW 40 Previous studies provide important information for work zone impact on crash rates. Juergens (12) 41
reported an increase in crash rate between 7.0 to 21.4 percent for ten work zone locations. Graham (13) 42
stated an average increase of 7.5 percent for crash rates within 79 work zone sites. Rouphail et al. (14) 43
and Hall and Lorenz (15) concluded in their study that crash rates increased with the presence of work 44
zones more than 88 percent and 26 percent respectively. Garber and Woo (16) reported that the crash 45
rates at work zones on multilane highways increased on the average by 57 percent and the crashes at work 46
zones on two-lane urban highways increased about 168 percent on average. The research by Pigman and 47
Agent (17) also showed increasing crash rates for 14 out of 19 work zones when compared to the before 48
work zone conditions. Rouphail et al. (14) and Ha and Nemeth (18) found that work zone crashes are less 49
severe than pre-construction crashes. On the contrast, Pigman and Agent (17) and Garber and Zhao (19) 50
stated that work zone crashes are more severe than non-work zone crashes. Benekohal et al. (20) reported 51
TRB 2013 Annual Meeting Paper revised from original submittal.Downloaded from amonline.trb.org
Ozturk, Ozbay, Yang and Bartin 4
the survey among truck drivers that 90 percent of them consider traveling through work zone is more 1
dangerous than non-work zone areas. Rouphail et al. (14) and Wang et al. (21) found that rear-end crashes 2
increased significantly during work zone periods. Khattak et al. (9) found that crash rates were higher 3
with the presence of work zones 23.5 percent for non-injury crashes and 17.5 percent for injury crashes. 4
Contrary to other studies described above, Jin et al. (22) observed lower crash rates during construction 5
periods in Utah. Dissanayake and Akepati (23) found that most of the work zone crashes at five states 6
(Iowa, Kansas, Missouri, Nebraska, and Wisconsin) occurred during the daytime and under clear weather 7
(no adverse) conditions. According to the descriptive studies about work zone crash rates, general 8
opinions are that work zones have an increasing effect on crash frequency when compared to pre-work 9
zone conditions. 10
Crash frequency models have been used for road safety analysis in many studies. Negative 11
binomial models (Poisson-Gamma) are generally used for modeling crash count data. In addition, there 12
are several methods to model crash frequencies such as Poisson model, zero inflated negative binomial 13
and Poisson, truncated regression, generalized additive model, Conway Maxwell Poisson model, negative 14
multinomial model etc. (24). There are only a few studies in the literature that are directly related to the 15
work zone crash frequency modeling. The following studies are focused on modeling work zone crash 16
occurrence by using different factors as model components. 17
As shown in Table 1, different methodologies and parameters were applied to develop work zone 18
crash frequency models. Pal and Sinha (6) investigated the safety effects of lane closure strategies at 19
interstate work zones in Indiana. Normal regression, negative binomial and Poisson regression models 20
were developed by using duration, traffic interaction and crash rate at normal conditions. They found that 21
normal regression predicts crash counts better when compared to negative binomial and Poisson 22
regression results. Venugopal and Tarko (7) proposed two models for crash frequency on approaches to 23
work zones and inside the work zones. They used project cost and work zone type as additional 24
parameters based on previous studies. Duration and length variables were statistically significant and the 25
dependence was almost linear. Elias and Herbsman (8) used Monte Carlo simulation to develop a work 26
zone crash frequency probability function. According to results, work zone crashes increase incrementally 27
when daily traffic volume (ADT), duration, and work zone length increase. Khattak et al. (9) developed 28
negative binomial model for crash modeling by using following parameters: ADT, length, duration, urban 29
indicator, injury indicator, work zone indicator. Separate models were constructed for both injury and 30
non-injury in the pre-work zone and during work zone periods. According to negative binomial regression 31
results, duration has significant effect on both injury and non-injury work zone crash frequencies. Qi et al. 32
(10) modeled rear-end work zone crashes by using truncated negative binomial regression. They found 33
that occurrence of rear-end crashes are more likely in work zones with flaggers, alternating one way 34
traffic, and higher AADT conditions. Srinivasan et al. (11) attempted to model the location of crashes 35
within work zones as a function of the length of work zone segments, traffic volume and weather 36
parameters. Multinomial logit model was used to construct crash probabilities per lane-mile for different 37
segments. The work zone components were investigated separately. Among various weather and traffic 38
volume scenarios, bad weather and high traffic volumes made the advance warning area relatively unsafe, 39
and the termination area was found to be the most unsafe segment during off-peak hours. 40
Work zone length, AADT, duration, project cost and work zone type are the most common 41
parameters used to set crash frequency models statistically. Based on previous studies, additional 42
parameters related to work zone safety structure such as number of lanes closed, speed reduction, road 43
system and number of intersections are identified as possible additional variable to model the relationship 44
between work zone and crash frequency. Several statistical methods are used to model work zone crash 45
frequency. The negative binomial regression is one of the most favored techniques adopted to model 46
crash frequency because of its performance on possible overdispersion in the data (24). 47
TRB 2013 Annual Meeting Paper revised from original submittal.Downloaded from amonline.trb.org
Ozturk, Ozbay, Yang and Bartin 5
TABLE 1 Work Zone Crash Frequency Modeling Literature Summary 1
Work Zone Crash
Frequency Models
Pal & Sinha
(1996)(6)
Venugupal
& Tarko
(2000)(7)
Elias &
Herbsman
(2000)(8)
Khattak
et. al
(2002)(9)
Qi et. Al
(2005)
(10)
Srinivasan
et. Al
(2008)(11)
This
Study
Duration * * * * *
*
AADT * * * * * * *
Length * * * *
* *
No. of Operating Lanes
*
Work Zone Speed Limit
*
Cost of Project
*
Lane Closure
*
*
Speed Reduction
*
Urban Indicator
* *
Road System
*
Weather *
*
Crash Rate
*
Intersection
*
*
Ramp
*
*
Daytime-Nighttime
*
PDO-Injury
*
*
Control Device
*
Type of Work
*
*
Sample Size (Site) 34 116 - 36 - 1 60
Model NB, NLR,P NB Monte Carlo NB TNB MNL NB
NB:Negative Binomial Regression, P:Poisson, NLR:Normal Linear Regression, TNB:Truncated NB, MNL:Multinomial Logit Regression
2
DATA DESCRIPTION 3 Statistical software R is used to merge and analyze the crash data for this study. Crash counts are 4
clustered by three monthly periods (seasonal) for each work zone and separated into four categories as 5
following; daytime PDO, daytime injury, nighttime PDO and nighttime injury crashes. Fatal crashes are 6
assumed as injury crashes to avoid bias. Table 2 shows the number of crashes for each set in this study. 7
8
TABLE 2 Number of Work Zone Crashes for Each Category 9
Category PDO Injury Total
Daytime Crashes 2915 862 3777
Nighttime Crashes 1192 413 1605
Total Crashes 4107 1275 5382
10
AADT is a critical parameter for crash models since it represents the exposure of traffic. Hence, 11
accuracy of the AADT values is important for modeling. Khattak et al. (9) stated that directional AADT 12
should be used for modeling to determine crash distribution more precisely. To avoid over-representation 13
of AADT within the model, directional AADT from the NJ Straight Line Diagrams is used for given 14
milepost and time post of work zones. AADT values are adjusted by using New Jersey Department of 15
Transportation (NJDOT) seasonal adjustment factors since these values are estimated yearly (5). Instead 16
of using AADT directly, a traffic volume parameter is used for daytime and nighttime conditions 17
separately. It is generated by adjusting the seasonal-directional-AADT value based upon hourly 18
TRB 2013 Annual Meeting Paper revised from original submittal.Downloaded from amonline.trb.org
Ozturk, Ozbay, Yang and Bartin 6
adjustment factors. Bourne et al. (25) reported an example of normalization issues for results that daytime 1
work zone crashes are overrepresented. Nighttime traffic is approximately estimated as a quarter of the 2
total daily traffic. In this way, a biased relationship between AADT and crash counts is avoided by using 3
reduced traffic volume for nighttime condition. Figure 1 shows hourly distribution of all work zone 4
crashes occurred in NJ between 2004-2010. 5
6 FIGURE 1 Hourly distribution of all work zone crashes in NJ, 2004-2010 (5). 7
8
Data from the NJ crash database for the period between 2004 and 2010, work zone project 9
drawings and NJ straight line diagrams are combined to create the integrated database for modeling work 10
zone crash frequencies. After separating each work zone by direction, and filtering incomplete data 11
mainly because of unclear project drawings, 60 work zones are used in frequency modeling. Work zone 12
spatial information is gathered from project drawings. Project drawings do not include time and date 13
information regarding work zones. Accordingly, spatio-temporal diagrams of work zone crashes for each 14
work zone are plotted to better determine location and duration of the work zones. Advance warning, 15
transition and termination areas are not included in project mileposts. Therefore project mileposts are 16
sometimes updated to capture all of work zone labeled crashes. A sample spatio-temporal diagram of the 17
work zone crashes along I-78 and related project drawing are shown in Figure 2 and Figure 3. 18
19
20 FIGURE 2 Temporal spatial plotting of construction zone crashes at I-78. 21
0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23
Time of Day
Nu
mb
er
of W
ork
Zo
ne
Cra
sh
es
0500
1000
1500
2000
2500
3000
608524 483
368 333437
911
1747
2485
22192095
2381
267826582724 2762
26782690
2010
1447
12171314 1276
940
TRB 2013 Annual Meeting Paper revised from original submittal.Downloaded from amonline.trb.org
Ozturk, Ozbay, Yang and Bartin 7
1 FIGURE 3 Sample project drawing (cover page) for the work zone project on I-78 (26). 2
3
Figure 4 shows a sample for the spatio-temporal distribution of the work zone crashes within 4
different work zone sections. Milepost information for work zone components is gathered from project 5
drawings. Ramp and intersection locations are also included. From the Figure 4, it can be seen that work 6
zone crashes are likely to be denser near ramps and intersections. 7
8
9 FIGURE 4 Spatial distribution of crashes within work zone on US 1 North (5). 10
11
Work zone length, milepost, number of operating lanes and lane closure information are obtained 12
from project drawings. NJDOT crash records provide seven different light conditions, however, for the 13
sake of simplicity; these are categorized into two levels as daytime and nighttime. Estimated project 14
duration is specified in terms of days. NJ Straight line diagrams provide posted speed limit for normal 15
conditions. Work zone speed limits are gathered from NJ crash records according to the distribution of 16
2007
2008
2009
2010
2011
Milepost
Date
31.0 31.4 31.8 32.2 32.6 33.0 33.4 33.8 34.2 34.6
Adv ance Warning
( 24 % )
Transition
( 0.5 % )
Activ ity Area
( 74 % )
Termination
( 1.5 % )AADT ( 33899 )
Non-WZ Crashes
WZ Crashes
Intersection
Ramp
TRB 2013 Annual Meeting Paper revised from original submittal.Downloaded from amonline.trb.org
Ozturk, Ozbay, Yang and Bartin 8
posted speed limits within work zones. Work zone speed reduction and lane drop parameters are 1
generated by estimating difference between work zone and normal conditions. 2
Road types are categorized in two levels: Interstate and state highways. Number of lanes, number 3
of ramps and intersection for each work zones are obtained from NJ straight line diagrams. The duration-4
based model is developed by using 120 work zone crash count data for entire duration. Period-based PDO 5
and injury models are developed based on 950 work zone crash count data within 3 monthly seasonal 6
periods. Since work zones represent temporary conditions on the roadway and each work zone has its 7
own construction schedule, it is not possible to aggregate crashes over a longer time period. Thus, a three 8
month period is selected as the shortest appropriate duration. 9
10
MODEL SPECIFICATION 11 There are many statistical methods employed by previous related studies to model crash frequencies 12
including Poisson model, zero inflated negative binomial and Poisson, truncated regression, generalized 13
additive model, Conway Maxwell Poisson model, and negative multinomial model etc. (24). Crash counts 14
are non-negative integers that are affected by several factors. Elias and Herbsman (8) found best fitted 15
distributions for crash counts as negative binomial and Poisson. To model such crash count data, negative 16
binomial modeling (Poisson-Gamma) is favored since it allows over dispersion caused by other variables 17
(7, 9). Accordingly, crash frequency is modeled by using negative binomial method, where the dependent 18
variable is the total work zone crash counts for a given period. 19
Let Yi represents the number of crashes at work zone i for an exact length and duration, crash 20
occurrence for work zone i is independent and probability density can be Poisson distributed (9,27); 21
22
(1) 23
24
Where, is observed number of crashes and is the expected crash frequency for the work 25
zone . represents explanatory variables such as duration, length and AADT. Mean and variance values 26
of Yi are equal to mean and variance of . Eq. (2) describes this statement, where is the estimated 27
coefficient and is the value of explanatory variables. Overdispersion is included by error term , which 28
represents a random effect due to omitted explanatory variables and unmeasured heterogeneity. 29
30
(2) 31
32
In the negative binomial model, is assumed as gamma distributed with mean 1 and 33
variance (9, 28). Natural form of overdispersion is; 34
35
(3) 36
37
If we insert instead of , dispersion rate will be; 38
39
(4) 40
41
The negative binomial model differentiates from the Poisson model by the parameter , which is 42
necessary to determine overdispersion. If is significantly different than zero, the negative binomial 43
model is appropriate to use with such a count data, else Poisson model can be used. Safety performance 44
function (SPF) for predicting the number of work zone crashes in an interval of given length and duration 45
can thus be built as follows; 46
(5) 47
48
TRB 2013 Annual Meeting Paper revised from original submittal.Downloaded from amonline.trb.org
Ozturk, Ozbay, Yang and Bartin 9
Where is the expected number of crashes within work zone; represents the work zone length 1
for the whole period of construction, is the traffic volume during the period of study Di represents 2
duration of the work zone. represents other explanatory variables and represents the coefficients for 3
model parameters. 4
Based on the crash information available from NJDOT crash database, project files and straight 5
line diagrams, following variables are selected for crash frequency modeling explained in Table 3. These 6
variables are length, light conditions, adjusted traffic volume, posted speed, speed reduction, the number 7
of operated lanes, the number of lane closures, road type, the number of ramps, and the number of 8
intersections within the work zone. Natural log values are used for duration, traffic and work zone length 9
parameters. Based on these parameters, the duration-based and period-based models are as follows: 10
11
Duration-based model; 12
(7) 13
14
Period-based model; 15
(8) 16
17
TABLE 3 Variables Considered in the Negative Binomial Model 18
Variable Symbol Type Description
Length length continuous length of the work zone (mile)
Light Conditions Night indicator daytime = 0, nighttime = 1
Traffic Volume traffic continuous adjusted ADT per lane (by direction,
seasonal factor, time factor)
WZ Speed wzspeed continuous reduced posted speed limit (mph)
Operated Lane operatedlane indicator number of operating lanes
Lane Drop lanedrop continuous number of closed lanes
Speed Reduction speedreduction continuous reduction in posted speed limit (mph)
Road System roadsystem indicator Interstate = 0, state = 1
Ramp Number ramp continuous number of ramps
Intersection intersection Continuous number of intersections
Duration duration continuous duration of the work zone (days)
19
MODEL ESTIMATION RESULTS 20 The estimated parameters shown in Table 4, Table 5 and Table 6 are used to determine the relationship 21
between independent variables and the frequency of work zone crashes. The results of the duration based 22
and period based models are presented at the 95 percent level of significance. 23
According to the duration-based work zone crash frequency modeling results shown in Table 4, 24
the project duration is found to be as the most significant parameter affecting the occurrence of crashes. 25
Similar results were found in previous studies for the duration of the work zones (7, 9). Work zone length 26
and traffic volume within the construction site are also found to be significant for the duration-based 27
model. For one percent increase in duration, traffic volume and length variables crash frequency increases 28
by 0.71 percent, 0.51 percent and 0.48 percent, respectively. Traffic parameter which represents the 29
volume per lane in conjunction with the number of operating lanes captures the effect of exposure. 30
TRB 2013 Annual Meeting Paper revised from original submittal.Downloaded from amonline.trb.org
Ozturk, Ozbay, Yang and Bartin 10
TABLE 4 Estimated Parameters of Duration-Based Crash Frequency Model (N=120) 1
Variable Coefficient Std. Err. Z P>|z| Significant
ln(length) 0.477 0.133 3.580 0.000 ***
night -0.080 0.227 -0.350 0.725
ln(traffic) 0.512 0.158 3.230 0.001 ***
wzspeed -0.023 0.014 -1.600 0.110
operatedlane 0.642 0.141 4.540 0.000 ***
lanedrop 0.158 0.129 1.230 0.220
speedreduction 0.025 0.013 2.000 0.045 *
roadsystem 0.203 0.193 1.050 0.292
ramp 0.011 0.028 0.410 0.680
intersection 0.014 0.010 1.380 0.167
ln(duration) 0.710 0.084 8.500 0.000 ***
intercept -6.731 1.097 -6.140 0.000 ***
alpha 0.190 0.030
chibar2 = 614.89 Prob>=chibar2 = 0.000
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
2
Work zone PDO crash counts for three month periods are modeled using the negative binomial 3
regression. The period-based PDO crash frequency model results are shown in Table 5. Results depict 4
that traffic volume has significant incremental effect on PDO crashes. One percent raise in traffic volume 5
causes 0.45 percent increase in total number of work zone PDO crashes. Additional operating lane causes 6
78.8 percent increase in total PDO crashes since traffic volume is estimated per lane. Similarly, lengths' 7
marginal effect of one percent increase causes 0.48 percent more PDO crashes. Different than normal 8
crash modeling parameters, lane drop and speed reduction are also found to have a significant effect on 9
the number of PDO crashes. One additional lane drop causes 28.8 percent increase in PDO crashes when 10
other parameters are kept constant. Since we used reduced volume for nighttime conditions, interpretation 11
is not biased due to the less traffic during nighttime. Considering volume adjustment, overall crash counts 12
at night are decreased by 26.9 percent when compared to daytime condition. If the work zone is on state 13
highways, 60.4 percent more PDO crashes are expected based on the modeling results. One more ramp 14
within the work zone section causes 4.3 percent more PDO crashes for the construction site. Proportional 15
increase in PDO crashes is found to be higher than injury crashes. 16
According to the period-based injury crash frequency modeling results at 95 percent level of 17
significance, length of the work zone is one of the most important variables for estimating the injury 18
crashes when compared to the PDO crashes (Table 6). One percent increase in project length causes 0.64 19
percent increase in injury crash counts. Traffic volume and number of operating lanes which represent the 20
exposure to traffic are both significant. Ullman et al. (29) stated that nighttime does not cause 21
significantly greater risk for crash severity than daytime for motorists at work zone condition. Similarly, 22
nighttime injury crash occurrence probability at work zone is found to be 31.8 percent less than daytime 23
work zone. Lane drop is another efficient factor to increase injury crash frequency at work zones. One 24
more lane drop causes 48.1 percent more injury crashes at the work zone site. Srinivasan et al. (30) 25
reported in their study that total crashes were increased by 66.9 percent for daytime and 60.1 percent for 26
night time compared to normal traffic conditions by temporary lane closure conditions at work zones. 27
28
29
30
TRB 2013 Annual Meeting Paper revised from original submittal.Downloaded from amonline.trb.org
Ozturk, Ozbay, Yang and Bartin 11
TABLE 5 Estimated Parameters of Period-Based PDO Crash Frequency Model (N=950) 1
Variable Coefficient Std. Err. z P>|z| Significant
ln(length) 0.476 0.118 4.04 0.000 ***
night -0.314 0.152 -2.07 0.039 *
ln(traffic) 0.446 0.110 4.06 0.000 ***
wzspeed -0.011 0.009 -1.22 0.223
operatedlane 0.581 0.099 5.87 0.000 ***
lanedrop 0.253 0.084 3.00 0.003 **
speedreduction 0.036 0.010 3.46 0.001 ***
roadsystem 0.473 0.138 3.42 0.001 ***
ramp 0.042 0.020 2.15 0.032 *
intersection 0.013 0.007 1.84 0.066 .
intercept -4.648 0.876 -5.30 0.000 ***
alpha 0.501 0.040
chibar2 = 896.1 Prob>=chibar2 = 0.000
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
2
TABLE 6 Estimated Parameters of Period-Based Injury Crash Frequency Model (N=950) 3
Variable Coefficient Std. Err. z P>|z| Significant
ln(length) 0.642 0.153 4.20 0.000 ***
night -0.381 0.189 -2.02 0.043 *
ln(traffic) 0.325 0.137 2.36 0.018 *
wzspeed -0.016 0.011 -1.46 0.143
operatedlane 0.590 0.126 4.69 0.000 ***
lanedrop 0.393 0.102 3.84 0.000 ***
speedreduction 0.024 0.014 1.81 0.071 .
roadsystem 0.959 0.184 5.21 0.000 ***
ramp 0.023 0.024 0.96 0.337
intersection -0.001 0.009 -0.16 0.873
intercept -4.752 1.138 -4.18 0.000 ***
alpha 0.489 0.061
chibar2 = 172.1 Prob>=chibar2 = 0.000
Signif. codes: 0 ‘***’ 0.001 ‘**’ 0.01 ‘*’ 0.05 ‘.’ 0.1 ‘ ’ 1
4
CONCLUSIONS 5 In this study, the duration-based and the period-based safety performance functions are developed to 6
identify risk factors that affect crash frequency in work zones. Negative binomial regression is found to 7
be an appropriate approach to model count data because mean and variances differ significantly for the 8
data set used in this study. Crash database, work zone project drawings and NJ straight line diagrams are 9
used to estimate parameters for frequency models. The dataset used in safety performance function 10
models consists of 60 construction work zones on NJ interstate and state highways. Project drawings and 11
TRB 2013 Annual Meeting Paper revised from original submittal.Downloaded from amonline.trb.org
Ozturk, Ozbay, Yang and Bartin 12
crash database provide satisfactory information for analyzing work zone crashes. By identifying the 1
effects of work zone length, daily traffic exposure, and other explanatory variables on the frequency 2
models, this study provides considerable information to transportation agencies. 3
According to the estimation results, “project duration”, “length of the work zone”, and “traffic 4
volume” are found to be the most important factors that increase work zone crash frequency. This result is 5
consistent with previous studies (6-11). By considering adjusted traffic volume based on hourly 6
distribution, nighttime PDO and injury crash occurrence probability is found to be 26.9 and 31.8 percent 7
lower than daytime condition, respectively. Additionally, lane closure is found to be an increasing factor 8
for number of PDO and injury work zone crashes. Speed variance between posted and work zone speed 9
limits are observed to increase PDO crashes. Number of ramps and intersections are found to be 10
important factors that increase crash frequency as well. 11
Based on findings from the modeling results, following recommendations could be useful to 12
increase safety within work zones: 13
“Project duration” could be minimized by including incentives and disincentives within 14
contracts , an approach adopted some states (31), 15
To avoid exposure resulting from “heavy volume”, traffic could be diverted to alternate 16
routes or work schedule could be switched to off-peak and weekend periods when 17
appropriate conditions exist (32), 18
Nighttime schedule for work zones can also be recommended to reduce injury and PDO 19
work zone crashes, 20
The variance between the posted and work zone speed limits should be optimized to 21
prevent increase of rear-end crash occurrence. A sharp reduction in speed limits may 22
cause more risk within work zones, and 23
Lane closure strategies should be revised to minimize the number of lane drops. If a lane 24
closure is necessary, countermeasures such as an intelligent lane merging system (33) can 25
be implemented to allow a smoother transition within the work zone to reduce crash 26
frequency. 27
28
ACKNOWLEDGEMENTS 29 The contents of this paper reflect views of the authors who are responsible for the facts and accuracy of 30
the data presented herein. This research is partially supported by NJDOT and also fellowship from 31
Turkish Government. The contents of the paper do not necessarily reflect the official views or policies of 32
the agencies. 33
34
REFERENCES 35 1. FHWA. Facts and Statistics Work Zone Injuries and Fatalities 36
www.ops.fhwa.dot.gov/wz/resources/facts_stats/injuries_fatalities.htm. Accessed May 15, 2012. 37
2. Schrank, D., T. Lomax, and B. Eisele. TTI’s 2011 Urban Mobility Report Powered by INRIX 38
Traffic Data. Texas Transportation Institute, 2011. 39
http://tti.tamu.edu/documents/mobility-report-2011.pdf Accessed May 18 2012. 40
3. FY 2013 - 2022 Statewide Capital Investment Strategy. New Jersey Department of 41
Transportation. http://www.state.nj.us/transportation/capital/cis/pdf/scis1322.pdf Accessed May 42
13 2012. 43
4. Hartgen, D. T., R.K. Karanam, M. G. Fields, T. A. Kerscher. 19th Annual Report on the 44
Performance of State Highway Systems (1984-2008). Report No. Policy Study 385. 45
Transportation Research Board of the National Academies, Washington, D.C., 2010. 46
5. Ozbay, K., H. Yang, O. Ozturk, M. Yildirimoglu, S. Demiroluk, and B. Bartin. Work Zone Safety 47
Analysis. Working report. New Jersey Department of Transportation, New Jersey. 48
TRB 2013 Annual Meeting Paper revised from original submittal.Downloaded from amonline.trb.org
Ozturk, Ozbay, Yang and Bartin 13
6. Pal, R. and K.C. Sinha. Analysis of Crash Rates at Interstate Work Zones in Indiana. 1
Transportation Research Record: Journal of the Transportation Research Board, No. 1529, 2
Transportation Research Board of the National Academies, Washington, D.C., 1996, pp. 43–53. 3
7. Venugopal, S. and A. Tarko. Safety Models for Rural Freeway Work Zones", Transportation 4
Research Record: Journal of the Transportation Research Board, No. 1715, Transportation 5
Research Board of the National Academies, Washington, D.C., 2000, pp. 1-9. 6
8. Elias, A.M. and Z.J. Herbsman. Risk Analysis Techniques for Safety Evaluation of Highway 7
Work Zones. Transportation Research Record: Journal of the Transportation Research Board, 8
No. 1715, Transportation Research Board of the National Academies, Washington, D.C., 2000, 9
pp. 10–17. 10
9. Khattak, A. J. and F. M. Council. Effects of Work Zone Presence on Injury and Non-Injury 11
Crashes. Accident Analysis and Prevention, Vol. 34, 2002, pp. 19-29. 12
10. Qi, Y., R. Srinivasan, H. Teng, and R.F. Baker. Frequency of Work Zone Accidents on 13
Construction Projects. Report No: 55657-03-15. Region 2, University Transportation Research 14
Center, City College of New York, New York, 2005. 15
11. Srinivasan, S., G. Carrick, X. Zhu, K. Heaslip, and S. Washburn. Analysis of Crashes in Freeway 16
Work Zone Queues: A Case Study. Final Report Research No: 07-UF-R-S3. Transportation 17
Research Center, University of Florida, Florida, 2008. 18
12. Juergens, W.R. Construction Zone, Detour and Temporary Connection Accidents. Business and 19
Transportation Agency, California Division of Highways, 1972. 20
13. Graham, J., R. Paulsen, and J. Glennon. Accident and Speed Studies in Construction Zones. 21
Report No. FHWA-RD-77-80. FHWA, U.S. Department of Transportation, 1977. 22
14. Rouphail, N.M., Z.S. Yang, and J. Fazio. Comparative Study of Short- and Long-Term Urban 23
Freeway Work Zones. Transportation Research Record: Journal of the Transportation Research 24
Board, No. 1163, Transportation Research Board of the National Academies, Washington D.C., 25
1988, pp. 4-14. 26
15. Hall, J.W. and V.M. Lorenz. Characteristics of Construction-Zone Accidents. Transportation 27
Research Record: Journal of the Transportation Research Board, No. 1230, Transportation 28
Research Board of the National Academies, Washington D.C., 1989, pp. 20-27. 29
16. Garber, N. J. and T. H. Woo. Accident Characteristics at Construction and Maintenance Zones in 30
Urban Areas. Research Report VTRC 90 – R12, Virginia Transportation Research Council, 31
Charlottesville, Virginia, 1990. 32
17. Pigman, J.G. and K.R. Agent. Highway Accidents in Construction and Maintenance Work Zone. 33
Transportation Research Record: Journal of the Transportation Research Board, No. 1270. 34
Transportation Research Board of the National Academies, Washington D.C., 1990, pp. 12-21. 35
18. Ha, T. and Z. Nemeth. Detailed Study of Accident Experience in Construction and Maintenance 36
Zones. Transportation Research Record: Journal of the Transportation Research Board, No. 37
1509, Transportation Research Board of the National Academies, Washington, D.C., 1995, pp. 38
38-45. 39
19. Garber, N.J. and M. Zhao. Distribution and Characteristics of Crashes at Different Work Zone 40
Locations in Virginia. Transportation Research Record: Journal of the Transportation Research 41
Board, No. 1794, Transportation Research Board of the National Academies, Washington, D.C., 42
2002, pp. 19–25. 43
20. Benekohal, R., E. Shim, and P.T.V. Resende. Truck Drivers' Concerns in Work Zones: Travel 44
Characteristics and Accident Experiences. Transportation Research Record: Journal of the 45
Transportation Research Board, No.1509, Transportation Research Board of the National 46
Academies, Washington D.C., 1995, pp. 55-64. 47
21. Wang, J., W. E. Hughes, F.M. Council, and J.E. Paniati. Investigation of Highway Work Zone 48
Crashes: What We Know and What We Don’t Know. Transportation Research Record: Journal 49
of the Transportation Research Board, No.1529, Transportation Research Board of the National 50
Academies, Washington D.C., 1996, pp. 54-64. 51
TRB 2013 Annual Meeting Paper revised from original submittal.Downloaded from amonline.trb.org
Ozturk, Ozbay, Yang and Bartin 14
22. Jin, T. G., M. Saito, and D. L. Eggett. Statistical Comparisons of the Crash Characteristics on 1
Highways between Construction Time and Non-Construction Time. Accident Analysis and 2
Prevention, Vol.40, 2008, pp. 2015-2023. 3
23. Dissanayake, S. and S.R. Akepati. Characteristics of Work Zone Crashes in the Swzdi Region: 4
Differences and Similarities. Proceedings of the 2009 Mid-Continent Transportation Research 5
Symposium, Ames, Iowa, 2009. 6
24. Lord, D. and F. Mannering. The statistical analysis of crash-frequency data: A Review and 7
Assessment of Methodological Alternatives. Transportation Research Part A, 2010, Vol. 44, pp. 8
291–305. 9
25. Bourne, J.S., T.A. Scriba, C. Eng, R.D. Lipps, , G.L. Ullman, , D.L. Markow, K.C. Markow, D. 10
Gomez, D.L. Holstein, B. Zimmerman, and R. Stargell Best Practices in Work Zone Assessment, 11
Data Collection, and Performance Evaluation. Scan Team Report (Scan 08-04), National 12
Cooperative Highway Research Program (NCHRP) Project 20-68A U.S. Domestic Scan, 2010. 13
26. State of New Jersey, Department of Transportation, Plans of Route I-78 Local and Express 14
Roadway Improvements, Contract No. 050003731, 2006. 15
27. Greene, W. Econometric Analysis, Third edition. Macmillan, New York, 1997. 16
28. Cameron, C.A. and P. K. Trivedi. Regression Analysis of Count Data. Econometric Society 17
Monograph No. 30. Cambridge University Press, Cambridge, 1998. 18
29. Ullman, G.L, M.D. Finley, J.E. Bryden, R. Srinivasan, and F.M. Council. Traffic Safety 19
Evaluation of Nighttime and Daytime Work Zones. NCHRP Report No. 627, Transportation 20
Research Board of the National Academies, Washington D.C., 2008. 21
30. Srinivasan, R., G.L. Ullman, M.D. Finley, and F.M. Council. Use of Empirical Bayesian Methods 22
to Estimate Temporal-Based Crash Modification Factors For Construction Zones. Transportation 23
Research Record: Journal of the Transportation Research Board, No. 2219, Transportation 24
Research Board of the National Academies, Washington, D.C., 2011, pp. 29-38 25
31. Zhu, Y., I. Ahmad, and L. Wang. "Estimating work zone road user cost for alternative contracting 26
methods in highway construction projects." Journal of Construction Engineering and 27
Management, Vol. 135 (7), 2009, pp. 601–608. 28
32. Ullman, G.L., V. Iragavarapu, and D. Sun. Work Zone Positive Protection Guidelines. Report No. 29
FHWA/TX-11/0-6163-1, Texas Transportation Institute, 2010. 30
33. Pei, Y. and L. Dai. Study on Intelligent Lane Merge Control System for Freeway Work Zones. 1. 31
Intelligent Transportation Systems Conference. IEEE, Seattle, WA, 2007, pp. 586-591. 32
TRB 2013 Annual Meeting Paper revised from original submittal.Downloaded from amonline.trb.org