Accident Analysis and Prevention 59 (2013) 260266

Contents lists available at SciVerse ScienceDirect

Accident Analysis and Prevention

journal homepage: www.elsevier.com/locate/aap

A comparison of collision-based and conict-based safety

evaluations: The case of right-turn smart channels
Emanuele Sacchi a,1 , Tarek Sayed a, , Paul deLeur b
a
b

The University of British Columbia, Department of Civil Engineering, 2002 6250 Applied Science Lane, Vancouver, B.C., Canada V6T 1Z4
Insurance Corporation of British Columbia, 530 171 W. Esplanade, North Vancouver, B.C., Canada

a r t i c l e

i n f o

Article history:
Received 2 February 2013
Received in revised form 30 May 2013
Accepted 2 June 2013
Keywords:
Smart channels
Treatment effectiveness index
Observational beforeafter studies
Full Bayes estimation
Trafc conict techniques

a b s t r a c t
This study presents the results of a collision-based full Bayes (FB) beforeafter (BA) safety evaluation of
a newly proposed design for channelized right-turn lanes. The design which is termed Smart Channels
decreases the angle of the channelized right-turn to approximately 70 . Its implementation is usually
advocated to afford drivers a better view of the trafc stream they are to merge with and to allow also
for safer pedestrian crossing. The evaluation used data for three treatment intersections and several
comparison sites in the city of Penticton, British Columbia. The evaluation utilized FB univariate and
multivariate linear intervention models with multiple regression links representing time, treatment, and
interaction effects as well as the trafc volumes effects. As well, the models were extended to incorporate
random parameters to account for the correlation between sites within comparison-treatment pairs.
The results showed that the implementation of the right-turn treatment has resulted in a considerable
reduction in the severity and frequency of collisions.
Another objective of the paper was to compare the results of the collision-based evaluation with the
results of a trafc conict-based evaluation of the same treatment intersections. The comparison showed
remarkable similarity between the overall and the location specic reductions in conicts and collisions
which provides support for using trafc conicts in BA studies. The results also provide positive empirical
evidence that can support the validity of trafc conict techniques.
2013 Elsevier Ltd. All rights reserved.

1. Introduction
A signicant volume of right-turning vehicles at intersections
can have adverse operational and safety effects. Therefore, channelized right-turn lanes are usually implemented at intersections with
high right-turn trafc volumes to reduce vehicle delay and improve
safety. Studies have shown that the provision of these lanes provides signicant safety benets (Agent et al., 1996; Harwood et al.,
2002).
To further improve the safety of channelized right-turn lanes, an
alternative right-turn design has recently been proposed (Zegeer
et al., 2002) that is more pedestrian-friendly and that supports
improved trafc operations. The new design, termed Smart
Channels, decreases the angle of the channelized right-turn to
approximately 70 . The design reduces the pedestrian crossing distance which can lead to shorter distance exposure, shorter signal
cycles, and reduced potential for pedestrians to be in conict with

Corresponding author. Tel.: +1 604 822 4379.

E-mail addresses: esacchi@mail.ubc.ca (E. Sacchi), tsayed@civil.ubc.ca (T. Sayed),
Paul.deLeur@icbc.com (P. deLeur).
1
Tel.: +1 604 822 8785.
0001-4575/$ see front matter 2013 Elsevier Ltd. All rights reserved.
http://dx.doi.org/10.1016/j.aap.2013.06.002

vehicles (Zegeer et al., 2002; The City of Ottawa, 2009). As well,

smart channels provide drivers with a better view of the trafc
stream they are to merge with (Fig. 1).
Conventional channelized right-turn lanes can require head
turns of as much as 150 . However, guidelines such as the Transportation Association of Canadas (TAC) guidelines require that
drivers should not have to look more than 120 back to check
approaching trafc (Transportation Association of Canada, 1999).
This is advocated because of the difculty of older drivers to turn
their head more than 90 to view intersecting trafc and the importance for drivers to clearly see potential conicting vehicles at the
end of an auxiliary right-turn (Staplin et al., 2001). Smart channels
can accommodate both issues by widening the visibility cone of
drivers (light-gray areas in Fig. 1).
In 2009, the Insurance Corporation of British Columbia (ICBC), in
partnership with the British Columbia Ministry of Transportation
and Infrastructure, initiated a pilot project to evaluate the safety
performance of smart channels. The evaluation of the conversion
of right-turn lanes to smart channels was carried out for 3 intersections in the city of Penticton, British Columbia. A before-and-after
(BA) analysis of the safety impact was performed using automated
detection and analysis of trafc conicts with computer vision techniques (Autey et al., 2012). The results showed that the conversion

E. Sacchi et al. / Accident Analysis and Prevention 59 (2013) 260266

261

Fig. 1. Visibility conditions: conventional versus smart channelization.

of channelized right-turn lanes to smart channels can lead to a considerable reduction for both conict frequency and severity. The
overall reduction in total conicts was estimated at approximately
51%. The total severity of all conicts, normalized to trafc volumes, was observed to decrease by approximately 41% following
the treatments (Autey et al., 2012). One main advantage which supported the use of trafc conicts for the evaluation was that trafc
conicts are much more frequent than road collisions and BA studies based on trafc conicts can be conducted over much shorter
periods. Moreover, the use of automated video-based conict analysis considerably facilitates the collection of trafc conict data
and overcomes the reliability and repeatability problems usually
associated with manual trafc conict observations.
However, although the use of collision surrogates such as trafc conicts can offer several advantages for BA studies, the link
between conict reduction and potential collision reduction still
needs to be clearly established before a wider application of the
trafc conicts technique in BA studies. In fact, establishing the
linkage between conicts and collisions is still a challenging subject in todays road safety research as witnessed by many studies
(Gettman et al., 2008; Davis et al., 2008; El-Basyouny and Sayed,
2013). As well, BA studies based on collision data rely on more
solid statistical techniques which were developed over many years
of research compared to the automatic detection and analysis of
trafc conicts. Therefore, the main objectives for this study are to:
(a) Supplement the conclusions from Autey et al.s 2012 evaluation
by conducting a BA study using collision data and state-ofthe-art full Bayes (FB) statistical technique. Collision and trafc
volume data for the same treated intersections in Penticton are
considered in addition to data for a set of control/comparison
sites;
(b) Compare the results of the FB collision-based evaluation and
the conict-based evaluation conducted earlier to further
strengthen the validity of using the trafc conict technique.
This validation should lead to a wider application of the trafc conicts technique and better understanding of the link
between road safety, driver behavior, and dynamic trafc interactions.
2. Collision-based beforeafter studies
Several methods are available to analyze collision data and to
determine the effectiveness of safety treatments. Generally, these

methods can be classied into conventional and Bayesian techniques.

Conventional methods assume that the collisions observed at a
location can be used to directly provide an estimate of the true collision frequency. Examples of conventional evaluation techniques
include the simple BA studies and simple BA studies with a control
group (Hauer, 1997; Highway Safety Manual, 2010).
However, it can be argued that the observed accident frequency
alone is not a reliable measure to estimate road safety since collision
frequency is a stochastic variable of which the observed collision
frequency is but one realization. Bayesian methods treat collision
frequency as a random variable with its own probability distribution. The probability distribution is obtained in two steps. The
rst step involves determining a prior distribution and the second
step involves using Bayes theorem to convert the prior distribution into a posterior distribution with observed data. Examples of
Bayesian evaluation techniques include the empirical Bayes (EB)
and full Bayes BA studies. The EB technique is considered the current most popular road safety evaluation approach (Hauer, 1997;
Sayed et al., 2004; Persaud and Lyon, 2007; Sayed et al., 2010).
However, it is important to recognize that despite being the
current state-of-the-practice, the EB method suffers from several
limitations. These limitations make the use of the FB technique
more appealing for conducting before and after safety studies.
Researchers have shown that the FB technique: (i) accounts for
more uncertainty in the data and model parameters; (ii) provides
more detailed inference (credible intervals and parameter distributions); (iii) adds more exibility in selecting collision distributions;
(iv) allows inference at more than one level; (v) efciently integrates the estimation of the SPF and treatment effects in a single
step, reducing the need for a large data set; and (vi) conducts multivariate analysis (Li et al., 2008; Persaud et al., 2010; El-Basyouny
and Sayed, 2009a,b, 2010, 2011). Therefore, the FB approach will
be used in this study to estimate the effectiveness of the proposed
design for channelized right-turn lanes (smart channels).
2.1. Fully-Bayesian approach
If we consider a BA study where collision data are available for a
reasonable period of time before and after the intervention and, in
addition, we also consider the availability of a comparison/control
group for the before and after period of the intervention at treatment sites, it is possible to write the foundational model for FB

262

E. Sacchi et al. / Accident Analysis and Prevention 59 (2013) 260266

analysis of collision data in the form of a Poisson-lognormal intervention (PLNI) model. It is usually assumed that Yit denote the
collision count recorded at site i (i = 1, 2, . . ., n) during year t (1,
2, . . ., tB , tB + 1, . . ., tB + tA ), where tB is the last year before treatment and tA is the number of years after treatment, independently
distributed as (El-Basyouny and Sayed, 2012a,b,c):

where the only difference in the PLNI model in Eq. (3) is the additional subscript p(i) = 1, 2, . . ., NC which denotes which treatment
group the regression coefcient belongs to (with NC equal to the
number of comparison groups). This formulation implies the same
value of the model parameter in the group of similar sites (treated
and comparison).

Yit |it Poisson (it )

(1)

2.2. Treatment effectiveness index calculation

ln (it ) = ln (it ) + i

(2)

ln (it ) = 0 + 1 Ti + 2 t + 3 [t (tB,i + 1)]Iit + 4 Ti t

+ 5 Ti [t (tB,i + 1)]Iit + 6 Ti Iit + 1 ln (V1,it ) + 2 ln (V2,it )

(3)

With the FB method, the treatment effectiveness (TE) index,

from which it is possible to estimate how much collision experience
is expected to change following a countermeasure implementation,
can be calculated as (El-Basyouny and Sayed, 2012a):
ln (TEi ) = ln (TAi ) + ln (CB ) ln (TBi ) ln (CA )

with
i Normal (0, 2 )

(4)

where it is the expected number of collisions at site i (comparison or treatment) in year t, Ti denotes the treatment indicator
(equals 1 for treatment sites, 0 for comparison sites), tB+1,i denotes
the intervention year for the ith treatment site and its matching
comparison group, Iit denotes the time indicator (equals 1 in the
after period, 0 in the before period), 1 represents the difference
in log collision count between treated and comparison sites, 2
represents a linear time trend, 3 represents the slope due to the
intervention, 4 and 5 allow for different time trends and different intervention slopes across the treated and comparison sites,
6 which is usually referred to as jump term accounts for a
possible sudden change (drop or increase) of collisions at treated
sites in the post-intervention period, V1,it , V2,it denote the annual
average daily trafc (AADT) at the major and minor approaches (for
intersections), and 1 , 2 their respective coefcients, i accounts
for random effects for latent variables across the sites and  represents the extra-Poisson variation. This is the most common form
for the univariate PLNI model.
In addition, as already mentioned, for a collision count of severity level k (k = 1, 2, . . ., K), the FB methodology allows to conduct
multivariate analysis (El-Basyouny and Sayed, 2009a). In this case,
the PLNI model remains the same except for adding the superscript
k to the coefcients. However, in order to account for the correlation among crash counts of different severity levels at site i, it is
assumed for random effects that (El-Basyouny and Sayed, 2009b):
i = (1i , 2i , . . . , Ki )NK (0, )

(5)

where is a covariance matrix (the diagonal element  kk represents the variance of Ki , whereas the off-diagonal element  jk
j

represents the covariance of i and ki ).

Since comparison sites are selected to be as similar as possible to treatment sites, this may generate a correlation in collisions
between sites within comparison-treatment pair. To account for
this correlation, the variation due to comparison-treatment pairing can be represented by allowing the model coefcients to vary
randomly from one pairs to another, such that (El-Basyouny and
Sayed, 2012a):
ln (it ) = p(i),0 + p(i),1 Ti + p(i),2 t + p(i),3 [t (tB,i + 1)]Iit
+ p(i),4 Ti t + p(i),5 Ti [t (tB,i + 1)]Iit + p(i),6 Ti Iit
+ p(i),1 ln(V1,it ) + p(i),2 ln(V2,it )

(6)

and
p(i),j N(j, j2 ),

j = 0, 1, 2, 3, 4, 5, 6

(7)

p(i),j N(j, j2 ),

j = 1, 2

(8)

(9)

where TB and TAi are the predicted collision counts for the ith
treated site averaged over appropriate years during the before and
after periods, respectively, and CB and CA the corresponding
quantities for the comparison group where the predicted collision
counts are averaged over all sites in the matching comparison group
and years. Eq. (9) is equivalent to TEi = TAi CB /TBi CA = TAi /TAi
where TA = TBi (CA /CB ) represents the predicted collisions in
the after period for the ith treated site had the countermeasures not
been implemented and the term CA /CB is used to adjust the prediction for general trends between the before and the after periods
(Park et al., 2010).
Finally, the overall index, TE, can be computed from:
1
ln (TEi )
n
n

ln (TE) =

(10)

i=1

where the overall percentage of reduction in predicted collisions

counts can be easily estimated as (1 TE) 100.
3. Evaluation of right-turn smart channels
A collision and trafc data set made available from ICBC for the
years 20072011 was used for the analysis. As stated before, the
three treated locations with smart channels, evaluated in Autey
et al.s project, were selected also for this study. The details for the
three locations are summarized in Table 1.
In addition, other intersections in the city of Penticton suitable
as comparison/control sites were identied. Non-treatment sites
were selected and dived into two categories: (i) comparison sites
with right-turn channel, i.e., nearby intersections with right-turn
lane in the before and after period (where no countermeasure has
been implemented), and (ii) control sites that were nearby intersections comparable for geographic proximity and trafc volumes
to a treatment site.
Table 2 lists the comparison and control sites for this study. To
increase the size of the comparison sites data set, two additional
intersections with conventional right-turn channelization in the
municipality of Peachland were included in the analysis. Each control site was then matched to a treatment site (matched pairing).
The main criterion to do that was geographic proximity. Therefore,
a reasonable number of sites was available to develop FB models
with random parameters among matched pairs.
The collision data included fatal-plus-injury (F + I) and propertydamage-only (PDO) collision records for all the treated and
comparison/control sites from January 2007 to December 2011 (all
kind of collisions). In this study, two selected study periods are
investigated: monthly, and 4 months. Since the countermeasures
were implemented in the early summer of 2010, it was not possible to take into account a yearly study period. Indeed, PLNI models
require at least two study periods for the post-treatment analyses.
This led to a range of before-periods of 40 months and after-periods

E. Sacchi et al. / Accident Analysis and Prevention 59 (2013) 260266

263

Table 1
Treatment sites details.
Code

Main street

Cross street

Conguration

Signalized or stop control

Penticton
T1
T2
T3

Channel Parkway
Channel Parkway
Channel Parkway

Green Ave
Warren Ave
W Duncan Ave

3-leg
3-leg
4-leg

Stop
Stop
Signal

of 16 months. The months from May to August 2010 were excluded

from the analysis.
Table 3 provides summary statistics for collision frequency
(average number of collisions at an intersection during 4 months)
by severity level at treated sites during the period preceding and
following the intervention. As well, the table shows the percentage of reduction/increase of collisions. Overall, a strong reduction
for both PDO and F + I crashes is shown among treatment sites,
except for site T1 where an increase equal to 7% in F + I collisions
was observed. These values represent essentially a simple observational BA study without taking into account any confounding factor.
Thus, no denitive conclusions can be taken on the countermeasure
effectiveness using these values.
4. Platform and standards applied for model estimation
The statistical software WinBUGS (Spiegelhalter et al., 2005)
was selected as the modeling platform to obtain the FB estimates.
Several multivariate/univariate PLNI models were implemented
using different time intervals with xed/random coefcients
among pair sites.
The main goal of the FB analysis is to obtain an estimate of the
unknown parameters introduced in Section 2.1. Therefore, prior
distributions for all parameters are assumed and then the posterior
distributions are sampled using Markov Chain Monte Carlo (MCMC)
techniques. In this study, the whole set of parameters (i.e., j and
j ) for univariate models with xed and random parameters among
pairs were assumed as non-informative with Normal distribution,
zero mean and large variance, i.e., Normal(0, 103 ), to reect the
lack of precise knowledge of the their value (prior distribution).
The variance i2 and 2 of random effects were assumed InverseGamma(0.001, 0.001). Then, the inverse of the covariance matrix,

1 , for multivariate models was assumed Wishart(I, 2), where I

is the identity matrix and 2 is the number of degrees of freedom
which correspond to a vague prior knowledge (El-Basyouny and
Sayed, 2009a,b).
The code produced draws from the posterior distribution of the
parameters, and given those draws, the MCMC technique was used
to approximate the posterior mean and standard deviation of the
parameters. Hence, the posterior summaries in this study where
computed by running two independent Markov chains for each of
the parameters in the models for 40,000 iterations. Chains were
thinned using a factor of 10 and the rst 20,000 iterations in each
chain were discarded as burn-in runs. The convergence was monitored by reaching ratios of the Monte Carlo errors relative to the
standard errors for each parameter less than about 5%, using the
BGR statistics of WinBUGS and also using visual approaches such
as observing trace plots.
5. Results
5.1. Treatment effectiveness
The rst two FB models considered were univariate PLNI (UPLNI)
functions on both monthly and 4-month basis. Only the comparison sites in Table 2 (sites with conventional right-turn channel
before and after) were taken into account. Hence, the regression
coefcients of the model did not vary from a site to another. The set
of estimated parameters is shown on the left-hand side of Table 4
as well as the primary outcomes of the models (TE) which represent an average index of treatment effectiveness across the three
treated locations.
Stated in terms of the expected change in collisions, Table 4
indicates 65% (i.e., 100 [1 0.349]) expected reduction in total

Table 2
Comparison (c) and control (C) site details.
Code
c = comparison
C = control

Main street

Cross street

Conguration

Signalized or stop control

Matching pairing

Penticton (BC)
c1
c2
c3
c4
C1
C2
C3
C4
C5
C6
C7
C8
C9
C10
C11
C12
C13
C14

Channel Parkway
Railway St
Eckhardt Ave W
Government St
Channel Parkway
Skaha Lake Rd
Skaha Lake Rd
Main St
Main St
Main St
Main St
Government St
Government St
Government St
Eckhardt Ave W
Westminster Ave
Winnipeg St
S Main St

Fairview Rd
W Eckhardt Ave
Westminster Ave W
Carmi Ave
Skaha Lake Rd
Green Ave
Kinney Ave
McDougall Ave
Warren Ave
Industrial Ave
Duncan Ave
Industrial Ave
Duncan Ave
Eckhardt Ave
Winnipeg St
Winnipeg St
Wade Ave
Green Ave

4-leg
4-leg
3-leg
4-leg
3-leg
4-leg
4-leg
4-leg
4-leg
4-leg
4-leg
4-leg
4-leg
4-leg
4-leg
4-leg
4-leg
4-leg

Signal
Signal
Stop
Stop
Signal
Signal
Signal
Signal
Signal
Signal
Signal
Signal
Signal
Stop
Signal
Signal
Signal
Signal

T2
T3
T3
T3
T1
T1
T2
T2
T2
T2
T3
T2
T3
T3
T3
T3
T3
T1

Peachland (BC)
c5
c6

Highway 97
Highway 97

Princeton Ave
Clements Cres

4-leg
3-leg

Signal
Stop

264

E. Sacchi et al. / Accident Analysis and Prevention 59 (2013) 260266

Table 3
Four-month crash frequencies by severity level for treated sites.
Treatment site

T1 Channel/Green
T2 Channel/Warren
T3 Channel/Duncan

Average F + I crash frequency

Average PDO crash frequency

Before

After

Change

Before

After

Change

1.4
1.8
2.7

1.5
0.75
0.75

+7%
58%
72%

1.9
3.5
4.5

0.5
0.5
1.75

74%
86%
61%

Table 4
Parameters, standard errors and Indexes of Effectiveness for Univariate (Total Collisions) and Multivariate Analysis.
Parameter

UPLNI 4-months

UPLNI monthly

MPLNI 4-months
PDO

0
1
2
3
4
5
6
1
2


5.832
0.786
0.028
0.085
0.034
0.07
0.900
0.625
0.079
0.458

5.468
0.682
0.008
0.024
0.002
0.033
1.341
0.452
0.075
0.458

TE mean value

0.386

0.349

TE C.I.
5%
95%

0.225
0.597

0.214
0.524

0.152
0.564

F+I

8.541 2.354
4.099 1.119
0.895 0.376
0.638 0.244
0.028 0.028*
0.025 0.020*
*
*
0.091 0.039
0.050 0.021
*
*
0.064 0.049
0.005 0.004
0.549 0.190
0.471 0.143
*
1.94 0.722
0.033 0.065
0.859 0.356
0.318 0.097
0.064 0.015
0.114 0.019

 2,2 = 0.499 0.184

 1,1 = 0.550 0.201
 1,2 =  2,1 = 0.181 0.062
0.325
0.524

1.405
0.320
0.05*
0.038*
0.027*
0.05*
0.687*
0.286
0.013
0.204

1.483
0.394
0.0005*
0.024*
0.008*
0.034*
0.542
0.198
0.014
0.193

0.219
0.982

The sign of the estimate is not signicant at the 95% condence level.

crashes after the installation of Smart Channels for the monthly

basis and 61% for the 4-month basis. The TE condence interval
(CI) for this study was set at 95%. Thus the reduction of collisions
in this case was signicant at the 95% condence level because the
ranges do not include a TE value equal or higher than 1.
In addition, a multivariate analysis on 4-month basis with comparison sites was carried out to better comprehend the collision
reduction by severity level (right-hand side of Table 4). By separating PDO from F + I collisions, the time frame on monthly basis
was not taken into account due to the high number of no-crash
periods. In this case, the overall expected change in PDO collisions was estimated higher than the F + I one (67% vs. 48%).
Also in this case, the results were signicant at the 95% level of
condence.
Overall, Table 4 shows that the model coefcients that mainly
affected the TE values were those signicantly different from zero.
In particular, for both univariate and multivariate models, these
were the intercept, 0 , the coefcient that accounts for the difference in log collision count between treated and comparison sites,
1 , the jump term, 6 , (when signicant) and the trafc volume
coefcient at the major approach, 1 .
The second study performed was a FB evaluation with
comparison-treatment pairs. This analysis was important to highlight possible different collision reductions among treated sites.
To this end, the set of comparison sites was not sufcient to create matched pair groups and, for this reason, the evaluation was
extended to nearby intersections comparable for trafc/geographic
proximity (i.e., the control site group in Table 2). With this intention in mind, the two comparison sites in Peachland were excluded
from the analysis.
The means and standard errors (SEs) of the model coefcients,
which varied randomly from one pairs to another, are reported in
Table 5. Moreover, it is worth to note how the posterior estimates
of the standard deviations,  i , were signicantly greater than zero,
especially for  1 , and this supported once again the use of random
over x parameters among pairs.

In addition to the overall estimates of j and j provided in

Table 5, there were 3 sets of p(i),j and p(i),j (corresponding to the 3
treatment-comparison pairs) both for monthly and 4-month analysis (not shown to save space). Such estimates were used to compute
the relative TE indexes and the results are reported in Table 6.
The highest reduction of collisions was estimated at Channel
Parkway & Warren Avenue. The lower reduction was calculated at
Channel Parkway & Green Avenue. In this latter case, the reduction was not signicant at the 95% level but only at the 90% with a
value equal to 0.991 (4-month basis) and 0.986 (monthly basis).
The insignicant results may be associated with the low number of matched control sites collected for this treatment site. In
fact, as shown in Table 2 only three sites were available for the
Table 5
Parameters for the Analysis with Matched Pair Sites (Total Collisions).
Parameter

0
0
1
1
2
2
3
3
4
4
5
5
6
6
1
 1
2
 2

*

Mean value (4-month)

Mean value (monthly)

Mean

SE

Mean

SE

0.898
0.185
0.839
0.346
0.035*
0.063
0.018*
0.116
0.031*
0.088
0.114*
0.211
0.867*
0.399
1.224
1.409
0.364
0.288
0.321

0.130
0.092
0.411
0.101
0.076
0.033
0.060
0.072
0.095
0.139
0.168
0.101
0.544
0.439
0.550
0.604
0.109
0.091
0.098

0.507
0.187
0.752
0.344
0.009*
0.056
0.003*
0.069
0.001*
0.062
0.048*
0.093
1.331
0.401
1.170
1.374
0.360
0.270
0.338

0.092
0.090
0.437
0.180
0.050
0.032
0.031
0.045
0.011
0.117
0.071
0.044
0.524
0.440
0.444
0.611
0.101
0.054
0.100

The sign of the estimate is not signicant at the 95% condence level.

E. Sacchi et al. / Accident Analysis and Prevention 59 (2013) 260266

265

Table 6
Indexes of effectiveness for the analysis with matched pair sites (total collisions).
Intersection

Time frame

TE

Condence interval
5%

95%

T1Channel/Green

Monthly
4 months

0.617
0.636

0.260
0.275

1.168
1.203

T2Channel/Warren

Monthly
4 months

0.309
0.349

0.133
0.148

0.555
0.635

T3Channel/Duncan

Monthly
4 months

0.414
0.449

0.215
0.228

0.673
0.751

Overall

Monthly
4 months

0.407
0.444

0.249
0.260

0.609
0.678

Table 7
Collisions versus trafc conicts reduction.
Intersection/severity

Trafc conicts (average hourly conicts)

Collisions (on 4-month basis)

T1 Channel/Green
T2 Channel/Warren
T3 Channel/Duncan
Overall

33%
57%
55%
51%

36.4%a
65.1%a
55.1%a
55.6%a

Total severity

Conict severity: 41%

PDO: 67.5%b
F + I: 47.6%b

a
b

PLNI model with matched pairs.

MPLNI model.

matched pairing analysis for site T1. Finally, the overall TE indexes
were found similar but slightly higher compared to the univariate analysis with xed coefcients (see Table 4). This difference
may be attributable to the different set of comparison groups taken
into consideration and the type of analysis used (univariate versus
matched pairs).

is consistent between the two studies. These results should offer

strong support for the validity of using trafc conicts in BA studies. As well, the results indicate that reducing trafc conicts can
help road safety agencies to achieve their ultimate goal of reducing
collisions.
6. Summary and conclusions

5.2. Comparison between collision-based and conict-based

results
As mentioned earlier, an evaluation study was conducted in
2012 to evaluate the same three treatment intersections using
automated video-based trafc conict analysis (Autey et al., 2012).
The study demonstrated the usefulness of using trafc conicts in
BA safety evaluations as the desired sample size for analysis can be
obtained in much shorter time periods. It was also demonstrated
that new developments in the use of computer vision techniques
to automate the extraction of trafc conicts from video data
can overcome the shortcomings of the traditional manual conict
observation methods. The results showed that the conversion of
channelized right-turn lanes to smart channels can lead to a considerable reduction in both conict frequency and severity. The
overall reduction in total conicts was estimated at approximately
51%. The total severity of all conicts, normalized to trafc volumes, was observed to decrease by approximately 41% following
the treatments. The results of the Autey et al.s study are summarized in Table 7 along with those from this study using the analysis
with matched pair sites (the one that relies on the largest dataset).
Since the multivariate analysis was available only on 4-month basis,
the same time period was used also for the univariate analysis with
matched pairs.
The table shows that the trafc conict reductions are very
consistent with collision reductions. However, collision reductions
are always slightly higher than conict reductions except in the
case of Channel/Duncan intersection. However, the differences in
the reductions are very small and the uncertainties related to the
estimated reductions, as reported in Table 6, have always to be considered when comparisons are made. As well, the ranking of the
three intersections according to the magnitude of the reductions

This study presented the results of a BA safety evaluation of

the conversion of channelized right-turn lanes to smart channels.
Several FB models were developed using different study periods:
(i) univariate PLNI models with xed parameters among different
treated sites, on monthly and 4-month basis; (ii) multivariate PLNI
models with xed parameters among different treated sites, on 4month basis only; and (iii) univariate PLNI models with random
coefcients among matched pairs, on both monthly and 4-month
basis. The FB methodology allowed obtaining a set of effectiveness
indexes for each treated location, severity level and time frame.
Moreover, considerably smaller data was required to perform the
analysis compared to other state-of-practice methodologies, such
as the EB method. Overall, the BA evaluation using collision records
showed a considerable improvement of safety levels after converting conventional right-turn lane into smart channels.
As well, the study compared the results of the collision-based
evaluation to the results of an earlier evaluation using automated
trafc conict techniques. The similarity of overall and the location specic reductions in conicts and collisions was remarkable
which provides strong support for using trafc conicts in BA studies. This result is signicant since currently there is little research
that clearly establishes the linkage between trafc-conict and
collision-frequency reductions. As well, the results add another
aspect of validity for recent research on modeling the relationship
between collisions and conicts (Davis et al., 2008; El-Basyouny
and Sayed, 2013).
It should be noted that the conict reductions in Autey et al.s
study were based on trafc events with associated minimum
time-to-collision (TTC) of less than 3 s. Expanding the research to
investigate the relationship between collision and conict reductions for other conict indicators (e.g., gap time, post encroachment

266

E. Sacchi et al. / Accident Analysis and Prevention 59 (2013) 260266

time) and different conict thresholds should be a worthwhile area

of future research.
Acknowledgement
The authors gratefully acknowledge the Insurance Corporation
of British Columbia (ICBC) for its nancial and resource commitment to this project.
References
Agent, K., Stamatiadis, N., Jones, S., 1996. Development of Accident Reduction Factors. Report KTC-96-13. Kentucky Transportation.
Autey, J., Sayed, T., Zaki, M., 2012. Safety evaluation of right-turn smart channels
using automated trafc conict analysis. Accident Analysis and Prevention 45,
120130.
Davis, G.A., Hourdos, J., Xiong, H., 2008. Outline of causal theory of trafc conicts
and collisions. In: Transportation Research Board 87th Annual Meeting (No. 082431).
El-Basyouny, K., Sayed, T., 2009a. Collision prediction models using multivariate poisson-lognormal regression. Accident Analysis and Prevention 41 (4),
820828.
El-Basyouny, K., Sayed, T., 2009b. Accident prediction models with random corridor
parameters. Accident Analysis and Prevention 41 (5), 11181123.
El-Basyouny, K., Sayed, T., 2010. A full Bayes approach to beforeafter safety evaluation with matched comparisons: a case study of stop-sign in-ll program.
Transportation Research Record: Journal of the Transportation Research Board
2148, 18.
El-Basyouny, K., Sayed, T., 2011. A full Bayes multivariate intervention model with
random parameters among matched pairs for beforeafter safety evaluation.
Accident Analysis and Prevention 43 (1), 8794.
El-Basyouny, K., Sayed, T., 2012a. Measuring direct and indirect treatment
effects using safety performance intervention functions. Safety Science 50 (4),
11251132.
El-Basyouny, K., Sayed, T., 2012b. Measuring safety treatment effects using full Bayes
non-linear safety performance intervention functions. Accident Analysis and
Prevention 45, 152163.
El-Basyouny, K., Sayed, T., 2012c. Linear and nonlinear safety intervention models:
novel methods applied to the evaluation of shoulder ruble strips. Transportation
Research Record: Journal of the Transportation Research Board 2280, 2837.

El-Basyouny, K., Sayed, T., 2013. Safety performance functions using trafc conicts.
Safety Science 51 (1), 160164.
Gettman, D., Pu, L., Sayed, T., Shelby, S., 2008. Surrogate Safety Assessment Model
and Validation: Final Report, Publication No. FHWA-HRT-08-05.
Hauer, E., 1997. Observational BeforeAfter Studies in Road Safety. Pergamon Press,
Elsevier Science Ltd., Oxford, United Kingdom.
Harwood, D.W., Bauer, K.M., Potts, I.B., Torbic, D.J., Richard, K.R., Kohlman Rabbani,
E.R., Hauer, E., Elefteriadou, L., 2002. Safety Effectiveness of Intersection Leftand Right-Turn Lanes. Report No. FHWA-RD-02-089. Federal Highway Administration, U.S. Department of Transportation.
Highway Safety Manual, 2010. American Association of State Transportation Ofcials (AASHTO), Washington, DC.
Li, W., Carriquiry, A.L., Pawlovich, M., Welch, T., 2008. The choice of statistical models
in road safety countermeasure effectiveness studies in Iowa. Accident Analyses
and Prevention 40, 15311542.
Park, E.S., Park, J., Lomax, T.J., 2010. A fully Bayesian multivariate approach
to beforeafter safety evaluation. Accident Analysis and Prevention 42 (4),
11181127.
Persaud, B., Lyon, C., 2007. Empirical Bayes beforeafter safety studies: lessons
learned from two decades of experience and future directions. Accident Analysis
and Prevention 39 (3), 546555.
Persaud, B., Lan, B., Lyon, C., Bhim, R., 2010. Comparison of empirical Bayes and full
Bayes approaches for beforeafter road safety evaluations. Accident Analysis
and Prevention 42 (1), 3843.
Sayed, T., deLeur, P., Sawalha, Z., 2004. Evaluating the insurance corporation of British Columbia road safety improvement program. Transportation
Research Record: Journal of the Transportation Research Board 1865,
5763.
Sayed, T., deLeur, P., Pump, J., 2010. Impact of rumble strips on collision reduction on
highways in British Columbia, Canada: comprehensive before and after safety
study. Transportation Research Record: Journal of the Transportation Research
Board 2148, 915.
Spiegelhalter, D., Thomas, A., Best, N., Lunn, D., 2005. WinBUGS User Manual. MRC
Biostatistics Unit, Cambridge http://www.mrc-cam.ac.uk/bugs
Staplin, L., Lococo, K., Byington, S., Harkey, D., 2001. Highway design handbook
for older drivers and pedestrians. Federal Highway Administration. Report No.
FHWA-RD-01-103.
The City of Ottawa, 2009. Ottawa Pedestrian Plan Draft.
Transportation Association of Canada, 1999. Geometric Design Guide for Canadian
Roads.
Zegeer, C.V., Seiderman, C., Lagerwey, P., Cynecki, M., Ronkin, M., Schneider, R.,
2002. Pedestrian facilities users guide providing safety and mobility. Federal
Highway Administration. Report No. FHWA-RD-01-102.