AR

WT

a

ARRA

KPIRST

1

1

oVacar

mrsOf

0h

Accident Analysis and Prevention 62 (2014) 238– 247

Contents lists available at ScienceDirect

Accident Analysis and Prevention

j ourna l h om epage: www.elsev ier .com/ locate /aap

n evaluation of Winnipeg’s photo enforcement safety program:esults of time series analyses and an intersection camera experiment

ard Vanlaar ∗, Robyn Robertson, Kyla Marcouxraffic Injury Research Foundation, 171 Nepean Street, Suite 200, Ottawa, Ontario, Canada K2P 0B4

r t i c l e i n f o

rticle history:eceived 11 December 2012eceived in revised form 11 July 2013ccepted 25 September 2013

eywords:hoto enforcementntersection cameraed-light runningpeedingime series analysis

a b s t r a c t

The objective of this study was to evaluate the impact of Winnipeg’s photo enforcement safety programon speeding, i.e., “speed on green”, and red-light running behavior at intersections as well as on crashesresulting from these behaviors. ARIMA time series analyses regarding crashes related to red-light running(right-angle crashes and rear-end crashes) and crashes related to speeding (injury crashes and propertydamage only crashes) occurring at intersections were conducted using monthly crash counts from 1994to 2008. A quasi-experimental intersection camera experiment was also conducted using roadside dataon speeding and red-light running behavior at intersections. These data were analyzed using logisticregression analysis. The time series analyses showed that for crashes related to red-light running, therehad been a 46% decrease in right-angle crashes at camera intersections, but that there had also been aninitial 42% increase in rear-end crashes. For crashes related to speeding, analyses revealed that the instal-lation of cameras was not associated with increases or decreases in crashes. Results of the intersectioncamera experiment show that there were significantly fewer red light running violations at intersectionsafter installation of cameras and that photo enforcement had a protective effect on speeding behavior

at intersections. However, the data also suggest photo enforcement may be less effective in preventingserious speeding violations at intersections. Overall, Winnipeg’s photo enforcement safety program hada positive net effect on traffic safety. Results from both the ARIMA time series and the quasi-experimentaldesign corroborate one another. However, the protective effect of photo enforcement is not equally pro-nounced across different conditions so further monitoring is required to improve the delivery of thismeasure. Results from this study as well as limitations are discussed.

. Introduction

.1. Background

Speeding and red light running are among the leading causesf road crashes in Canada and the United States (Goldenbeld andan Schagen, 2005; McGee and Eccles, 2003; Tay, 2000). Drivingbove the speed limit has been shown to increase one’s risk ofrash involvement, injury and death. Likewise, red light runninglso increases the risk of crashing, injury, and death for obviouseasons (Kloeden et al., 2001).

The consequences of speeding and red light running vary inagnitude. In Quebec, red light running has been shown to be

esponsible for more than one quarter of all traffic injuries at inter-

ections with traffic lights (Brault et al., 2007). According to anntario study (Ministry of Transportation Ontario, 1998), 42% of

atal crashes and 29% of injury crashes involved disobeying traffic

∗ Corresponding author. Tel.: +1 613 238 5235; fax: +1 613 238 5292.E-mail address: wardv@tirf.ca (W. Vanlaar).

001-4575/$ – see front matter © 2013 Elsevier Ltd. All rights reserved.ttp://dx.doi.org/10.1016/j.aap.2013.09.023

© 2013 Elsevier Ltd. All rights reserved.

signals. Therefore, approximately 61 fatal crashes and 4800 injurycrashes occurred in Ontario each year as a result of drivers runningred lights. The crashes that result from red light running also varyin severity. Red light running has generally resulted in right-anglecrashes which have a higher injury and fatality rate than most othertypes of crashes, including rear-end crashes (Helai et al., 2008).

Intersections can be even more hazardous when drivers arespeeding. Generally, as speed increases, so does the risk of beinginvolved in a crash as well as the severity of that crash (Evans, 2006;Elvik, 2005; Hess, 2004). In fact, the risk of being involved in a crashincreases proportionately to the increase in speed. Increasing theaverage driving speed by as little as 1% raises the risk of fatality by4–12% (Evans, 2004); driving 10 km/h above the speed limit morethan doubles the risk of being involved in a crash (Kloeden et al.,2001), while driving 20 km/h above the limit increases this risk upto six times.

Photo enforcement devices such as speed cameras and/or red

light cameras are increasingly being used in conjunction with tra-ditional police traffic enforcement techniques. In general, photoenforcement has been shown to bring about significant behavioralchanges in motorists that have resulted in reduced disregard for

sis an

tttsns2tme2

mot2a2Aghabp

miacadbcec

c2ododbPettaMnigTntms

1

eaWa

W. Vanlaar et al. / Accident Analy

raffic signals and designated speed limits (Blakey, 2003). However,here is considerable variation in the literature regarding the effec-iveness of photo enforcement programs in general. While mosttudies have found an overall reduction in speeding, red light run-ing, and associated crashes, some studies have not found anyignificant improvement (Andreassen, 1995; Burkey and Obeng,004) or found results that suggest photo enforcement is effec-ive only at some locations or under certain conditions and that

ore research is needed to better understand the impact of photonforcement and how this measure can best be employed (Erke,009; Garber et al., 2007; Kent et al., 1997).

Additionally, while some studies have found that photo enforce-ent works to reduce traffic violations at the camera sites alone,

ther studies suggest there is a spill-over or halo effect, i.e., a reduc-ion in surrounding non-camera intersections as well (Chen et al.,000; Hess, 2004; Ministry of Transportation Ontario, 1995; Rettingnd Kyrychenko, 2002; Retting et al., 1999; Shin and Washington,007), while some find no spillover effects; for example, in Phoenix,rizona (Shin and Washington, 2007). Such spillover effects sug-est a more generalized change in driver behavior. Further, itas been noted that spillover effects “are a key advantage ofutomated speed enforcement that are not generally achievedy traditional police speed enforcement” (Retting et al., 2008,. 444).

Many studies have investigated the impact of photo enforce-ent devices at intersections on red-light running and found

mprovements in the overall safety of intersections (see McGeend Eccles, 2003). However, some researchers have voiced con-erns about undesirable side effects of red-light cameras such as

possible increase in rear-end crashes. For example, a study con-ucted in 2005 examining the effects of red-light cameras using aefore-after research design found a 25% decrease in right-anglerashes, but also found a 15% increase in rear-end crashes (Councilt al., 2005). Unfortunately, if and how the effects of these deviceshange over time has not been studied.

Furthermore, while studies have looked at the effect of speedameras in general (cf. Wilson et al., 2010; Pilkington and Kinra,005), few evaluations of the impact of photo enforcement devicesn speeding behavior at intersections specifically have been con-ucted. The majority of studies on the effects of photo enforcementn speeding have focused on the use of mobile speed cameraevices. When fixed cameras have been examined, the effects ofoth mobile and fixed cameras are often examined together (e.g.,ilkington and Kinra, 2005). It should also be noted that few photonforcement programs have utilized photo enforcement cameraso detect “speed-on green” which is a type of photo enforcementhat captures vehicles as they speed through intersections on greennd amber lights. In Canada, only two jurisdictions, Alberta andanitoba have used speed cameras in this way (CCMTA, 2010) and

o evaluation has been conducted on the use of this technologyn Canada. In fact, the City of Winnipeg was one of the first pro-rams in North America to use the speed on green technology.hus, there is little information available regarding the effective-ess of these devices on speeding at intersections in particular. Forhis reason, there is a need to evaluate the use of photo enforce-

ent to detect speeding at intersections. This is the focus of thistudy.

.2. The Winnipeg photo enforcement safety program

The City of Winnipeg photo enforcement safety program was

stablished in 2003 to augment conventional traffic enforcements a potential solution to enhancing traffic safety. The goal of theinnipeg photo enforcement safety program is to reduce crashes

nd injuries by reducing red-light running and excessive speeding.

d Prevention 62 (2014) 238– 247 239

With respect to both speeding and red light runningoccurring at intersections the Winnipeg photo enforcementprogram utilizes a system that was designed by GatsometerBV. This technology can detect both speeding and red-lightoffences.

To detect red-light running at intersections, the City of Win-nipeg uses the “violation on entrance approach”, meaning theautomated photo enforcement system is activated only once thetraffic signal has turned from amber to red. At this point, any vehi-cle that passes over the magnetic sensors will trigger the camerato photograph the violating vehicle as it passes through the inter-section. Thus, only when the signal turns red, the sensors becomeactive (in essence, this means vehicles that entered the intersec-tion when the light was still amber but exit it when the light hasalready turned red are not in violation). This is different from thestricter “violation on exit” approach, where a violation is logged ifthe signal turns red upon exiting the intersection, even when thesignal was not red when entering it (in essence, this means vehi-cles that entered the intersection when the light was still amberbut exit it when the light has already turned red are indeed inviolation).

In Winnipeg, when a vehicle is detected passing over the acti-vated sensors two photographs are taken. The first photographtaken is of the vehicle outside the intersection (at the stop line)and shows that the signal is red. The second photograph shows thesame vehicle in the intersection and must show that the light is stillred. Any vehicle that is waiting to turn left or caught in the inter-section due to traffic backlog would not be photographed. Notethat these pictures are reviewed manually by photo enforcementstaff as part of a validation process to avoid issuing tickets for falsepositives.

To detect speeding at intersections, these same sensors detectthe presence of vehicles and calculate their speed using time anddistance. If the speed of the vehicle exceeds the predeterminedspeed threshold, the camera will be triggered to photograph theviolating vehicle as it passes through the intersection. The system’slevel of accuracy in measuring a driver’s speed is accounted forby using tolerances (accuracy below 100 km/h is ±2 km; accuracyabove 100 km/h is ±2%). While such tolerances are used, the Citycommunicates to the public that speeding is enforced at the postedspeed limits.

1.3. Objectives

The objective of this study was to evaluate the photo enforce-ment safety program in Winnipeg, Manitoba and determine theimpact of the program on crashes and violations related to speed-ing and red-light running. The Traffic Injury Research Foundation(TIRF) was contracted by the City of Winnipeg to evaluate the Win-nipeg Photo Enforcement Safety Program of the Traffic Safety Unitof the Winnipeg Police Service (see Vanlaar et al., 2011 for the fullreport about this evaluation).

This paper presents the results of time series analyses regardingcrashes related to red-light running (right-angle crashes and rear-end crashes) and crashes related to speeding (injury crashes andproperty damage only crashes) occurring at intersections. Suchanalyses allowing for the examination of trends over time havenot yet been widely applied to the study of photo enforcement asmost evaluations have used a before/after research design ratherthan time series.

This paper also presents the results of a quasi-experimental

intersection camera experiment using roadside data on speedingand red-light running behavior to assess the impact of the pro-gram on speeding and red-light running violations occurring atintersections.

2 sis an

2

2

2

imfrvtdocgrn(uaat

oorc

Wffitrc(mm

spottbibSois

2

ninctssetf

w

40 W. Vanlaar et al. / Accident Analy

. Methods

.1. Time series analyses

.1.1. Crash typesTime series analyses were adopted for this study to provide

nsight into trends regarding collisions before and after the imple-entation of the photo enforcement program in Winnipeg. Many

actors that affect road safety change with time such as weather,oad quality, vehicle design, traffic flow, the number of registeredehicles, etc. Such changes typically occur gradually which makeshe task of accurately measuring the true effect of an interventionifficult. Time series analyses can help to ensure that the effectf a program actually occurred after the intervention, i.e., thatrashes were not already declining for other reasons prior to pro-ram implementation. Time series analysis can also help to rule outegression to the mean – or the tendency of extremes to sponta-eously move closer to the mean – as a rival plausible explanationPalys, 2003). Perhaps most importantly, time series analysis wassed in this study to enable us to measure changing behavior at

more refined level compared to using a more crude before/afternalysis where all the before data and the after data are lumpedogether.

Two sets of ARIMA time series analyses have been conducted,ne set of analyses regarding crashes related to red-light runningccurring at intersections and one set of analyses regarding crasheselated to speeding occurring at intersections. These analyses wereonducted using Stata release 11 (StataCorp, 2009a).

In terms of types of crashes, an approach comparable to Shin andashington (2007) was used. In particular, crashes in the analyses

or red-light running were grouped into categories of crash con-guration: right-angle crashes and rear-end crashes – these arewo types of crashes that are typically associated with red-lightunning. Crashes in the analyses for speeding were grouped intoategories of severity: injury crashes and property damage onlyPDO) crashes. Fatal crashes have not been analyzed because many

onthly counts – if not all – for this type of crashes are zero so noeaningful conclusions from these data can be drawn.Crashes that could not have resulted from red light running or

peeding at intersections were not included in the analyses. Exam-les of such crashes include crashes where the driver was reversingr parked. In addition, with regard to analyses related to speeding,o increase the likelihood that the crashes under scrutiny in this sec-ion about speeding could logically have been the result of speedingehavior, only crashes during non-peak periods have been included

n these analyses (i.e., all crashes during the weekend and crashesetween 9:00 am and 4:00 pm and between 7:00 pm and 6:00 am).electing crashes according to this criterion can be justified basedn the knowledge that many drivers are stuck in traffic jams dur-ng peak hours and would therefore not be able to freely choose thepeed at which they want to drive.

.1.2. DataMonthly counts of crashes that happened in the city of Win-

ipeg between January 1994 and December 2008 have been usedn these analyses. Crash data were obtained from the City of Win-ipeg. Frequency distributions of crashes in the data provided wereompared to those reported in the City of Winnipeg’s Annual Reporto ensure accuracy of the received data. In order to make the analy-es robust, data from as early as possible, i.e., 1994 were used. In theame vein, monthly counts were used rather than daily counts tonsure the counts were sufficiently robust and normality assump-

ions hold (daily counts at the city level were too low and volatileor reliable analysis).

The cameras used in the Winnipeg photo enforcement programere installed at four distinct periods in time. The first 12 cameras

d Prevention 62 (2014) 238– 247

were installed in January 2003; the second set of 12 cameras wasinstalled in August of 2003; the third set in July/August 2004; and,finally, the last set of 12 cameras was installed in July/August 2005.In total, 48 camera intersections are included in this evaluation.Four dummy variables have been created to indicate when eachset of cameras was installed. Each observation is assigned a value0 or 1 for this dummy variable; 0 for all observations (i.e., crashes)before installation of a particular set of cameras and 1 for allobservations (i.e., crashes) after installation of this particular set ofcameras.

2.1.3. DesignOne set of time series analyses were conducted for crashes

related to red-light running occurring at intersections, notablyright-angle crashes and rear-end crashes, and one set of analyseswere conducted regarding crashes related to speeding occurring atintersections, notably injury crashes and PDO crashes.

In addition to both the analyses related to red-light running andto speeding that happened at intersections with photo enforcement(i.e., where cameras are used) further analyses were conductedwith crashes that happened at intersections in Winnipeg as a whole,excluding the camera intersections. These latter analyses have beenconducted to investigate possible spill-over effects, i.e., whetherthe impact of the photo enforcement program on red-light run-ning and speeding extended beyond intersections where camerasare used.

Finally, the models for both sets of crashes (i.e., those at cameraintersections and those in Winnipeg excluding camera intersec-tions) are also analyzed using data coming from a comparisongroup (see Section 2.1.4 for more information about the comparisongroup).

2.1.4. Comparison groupComparison group data come from data regarding comparable

crashes at comparable times in the province of New Brunswick.Data were cross-referenced and checked for accuracy. The sameselection criteria for crashes are applied as those used with thedata from Winnipeg (e.g., excluding crashes where a driver wasreversing or parked).

Note that during the monitoring period (January 1994 throughDecember 2008) New Brunswick does not have photo enforcementand, as such, is suitable as a comparison group (CCMTA, 2010).Also, New Brunswick was one of the few suitable jurisdictionsthat did not change the way crash configuration (e.g., right-angleversus rear-end) was captured during the monitoring period. Otherjurisdictions did change this, which disrupted the time series andrendered them unsuitable for use as a comparison group. In addi-tion, in both Winnipeg and the province of New Brunswick, theminimum reporting threshold for property damage only collisionswas $1000 in total collision damages during the monitoring period(Manitoba, 2012; New Brunswick, 2012). Furthermore, as Cana-dian jurisdictions, both New Brunswick and Winnipeg are subjectto comparable macro-economic developments, which is anotherimportant consideration in support of New Brunswick as a com-parison group.

The total number of crashes in Winnipeg from 1994 through2008 was 455,497 while the total number for New Brunswick was282,057. The distribution of crashes according to injury severity inboth jurisdictions is comparable with the large majority being prop-erty damage only (Winnipeg: 73%; New Brunswick: 62%), followedby injury crashes (Winnipeg: 26%; New Brunswick: 37%) and fatalcrashes (Winnipeg: 0.2%; New Brunswick: 1%). Finally, by virtue of

selecting crashes in New Brunswick based on crash configuration,the large majority of selected crashes took place on urban roads.

It warrants mentioning that extensive efforts were madeto obtain monthly counts of such variables as unemployment,

sis an

pm

2

dd

S

D

woε

warts2Ata

otNma(fp

2

2

cts

swittgigwb

2

ceuasN&tto

W. Vanlaar et al. / Accident Analy

opulation, traffic, etc. in Winnipeg and in New Brunswick for theonitoring period, but these data were not available.

.1.5. AnalysesARIMA time series modeling was used in Stata to analyze the

ata (see StataCorp, 2009a). In general terms, this model can beescribed as follows:

tructural equation : yt = xt ̌ + �t

isturbance, ARMA(1, 1) : �t = p�(t−1) + �ε(t−1) + εt

here p is the first order autocorrelation parameter, � is the first-rder moving-average parameter, and εt ∼ N(0, �2), meaning thatt is a white-noise disturbance (StataCorp, 2009c).

Pre-intervention series of raw crash data have been investigatedith special attention given to the overall pattern, outliers and vari-

nce of the data. Due to outliers and non-stationary variance, theaw crash numbers from all series have been transformed usinghe natural log transformation to mitigate their impact. Also, robusttandard errors have been used when modeling the data (StataCorp,009a). Selection of the final model was based on a comparison ofIC and BIC values of potential models, along with ARMA terms

hat were significant as well as within the bounds of stationaritynd invertibility (see McCleary and Hay, 1980; Yaffee, 2000).

Four dummy variables representing the installation of four setsf cameras along with data from a comparison group were simul-aneously introduced in the final model. Data from the province ofew Brunswick are used as a comparison group in the time seriesodels (see Section 2.1.4). Effects of the intervention dummy vari-

bles are described using adjusted monthly percentage changescoefficients of the log transformed data in the final model are trans-ormed using the number ‘e’). Effects were considered significant if-values were smaller than 0.05 (i.e., 5% level).

.2. Intersection camera experiment

.2.1. Research designA quasi-experimental evaluation design with intervention and

omparison groups was used to determine whether the implemen-ation of the intersection safety cameras led to fewer drivers whopeed and/or run red lights.

For each intervention site (four in total: 50 km/h in winter andummer and 60 km/h in winter and summer), two comparison sitesere included in the design (12 locations in total). These compar-

son sites are sites in the city of Winnipeg that are comparable tohe intervention sites, but at which no photo enforcement was usedhroughout the entire duration of the evaluation. Note that the pro-ram began long before the evaluation, so any site that could bencluded in the evaluation had already been exposed to the pro-ram. This is true for sites that were chosen as intervention sites asell as for those that were chosen as comparison sites; this has to

e borne in mind when interpreting the results from this study.

.2.2. DataRoadside data on speeding and red-light running behavior were

ollected to examine the effect of the intersection safety cam-ras on both speeding and red-light running. Data were collectednobtrusively with the same technology at the intervention sitesnd comparison sites, more precisely, using wireless pucks (Sen-ys Wireless Vehicle Detection System manufactured by Sensysetwork, Inc.) that are mounted in the road by Tri-Star Traffic

Distributing Inc. Devices (i.e., pucks) were tested periodicallyhroughout the implementation of the study to ensure proper func-ioning during data collection. Note that several evaluation studiesf these wireless devices have been conducted focusing on stop

d Prevention 62 (2014) 238– 247 241

bar detection as well as speed accuracy (see Cheung et al., 2005;Cheung and Varaiya, 2007; Day et al., 2010; Haoudi et al., 2008;Margulici et al., 2006) and these studies have reported high levelsof accuracy, comparable to traditional magnetics loops. The man-ufacturer reports counting errors of 1% compared to “ground truthas established by the traffic videos” (Sensys Networks Inc., 2007: p.5) and that “the essential ability of the [Sensys] wireless sensors tocapture traffic patterns is at least close to, if not better, than that ofloops” (Sensys Networks Inc., 2007: p. 5).

Sites that are comparable to already-existing enforcement sitesof the program, but at which photo enforcement had not yet beendeployed were chosen for this component of the evaluation. Infra-structure and construction were key factors in determining thefeasibility of installing cameras at selected locations.

More specifically, data were collected at two intervention siteswhere the maximum speed limit is 50 km/h, one location in thewinter and one in the summer, and at two intervention sites witha speed limit of 60 km/h, also one in the winter and one in thesummer. Data were also collected at two comparison sites per inter-vention site. Data were collected during the winter of 2009–2010(December 2009 through April 2010) and during the summer of2010 (May 2010 through September 2010). Note that due to unan-ticipated events, no data were available for the 50 km/h site in thesummer (the installation of traffic signals at a nearby intersectionled to trenching at the intervention location and, as a consequence,one traffic lane at the intervention location was closed during theexperiment). Thus, analyses were not conducted for this locationwith missing data.

After collecting three weeks of pre-data, i.e., data collectedbefore photo enforcement was implemented, the cameras wereinstalled and tested during a period of two weeks and then threeweeks of post-data were collected, i.e., data collected after photoenforcement had been implemented.

During weeks 4 and 5, signs were also erected at approachesto the intervention locations to indicate that photo enforcementwas taking place at these sites. The evaluation was unobtrusiveand evaluated actual photo enforcement. As such, if someone wasfound speeding or ran a red light at the intervention sites, theywould have been issued a ticket. As mentioned previously, the Cityof Winnipeg communicated to the public that the posted speedlimits were enforced, even though the enforcement cameras useda tolerance to account for measurement errors (see Section 1.2).

2.2.3. AnalysisFor the purposes of this study, speeding and red-light running

data as measured by the pucks, both before and after the installa-tion of the enforcement cameras, both at the intervention sites andcomparison sites were used to conduct the analyses.

Data were analyzed in Stata release 11 (StataCorp, 2009a). Withregard to speeding, analyses were first performed to determine ifthe installation of the cameras had an impact on average speed bymeasuring the change in average speed over time at each locationbefore the installation of cameras (weeks 1–3) and after the instal-lation (weeks 6–8). As mentioned previously, the cameras wereinstalled and tested during weeks 4 and 5. Differences in averagespeed were formally tested using a t-test.

Logistic regression analysis was also used to determine theimpact of the photo enforcement program. In particular, logisticregression analysis is suitable to determine the impact of a set ofindependent variables on a binary dependent variable, in our case‘speeding versus not speeding’ or ‘running a red light versus not

running a red light’. The formal model is expressed as follows (seeStataCorp, 2009b):

xj is defined as the (row) vector of independent variables, aug-mented by 1, and b is the corresponding estimated parameter

2 sis and Prevention 62 (2014) 238– 247

(T

P

ocdcpvllaodf

twstot

ltwnp

3

3

oino

fffitd

cwiice

TP

0

1

2

3

4

1995m1 2000m1 2005m1 2010m1t2

Intervention group Comparison group

Fig. 1. Log-transformed number of right angle crashes in the intervention group(Winnipeg, 48 intersections) and the comparison group (New Brunswick).

Table 2Percentage change in rear-end crashes at 48 camera intersections.

Red-light running: rear end crashes

Installation of cameras Coefficient ω (s.e.) p-Value % change(100 × (eω − 1))

First 12 cameras 0.353 (0.123) 0.004 42.33%Second 12 cameras −0.209 (0.121) 0.085 −18.86%

42 W. Vanlaar et al. / Accident Analy

column) vector. Ij = xjb is the predicted index of the jth observation.he predicted probability of a positive outcome is:

j = exp(Ij)1 + exp(Ij)

Logistic regression analysis was used to investigate the impactf the cameras on speeding violations for the intervention sitesompared to the comparison locations. Speeding violations wereefined as driving at least 1 km/h over the speed limit, as this isongruent with the City’s messaging to the public about enforcingosted speed limits. These results were calculated using all obser-ations as well as only those observations that took place duringow density traffic times, more precisely when traffic count wasower than 900 vehicles per hour per lane – this corresponds tollowing an average of 4 s headway for each vehicle and limits thebservations used in the analyses to those that took place whenrivers were free to choose what speed to drive at, rather than beingorced to go with the flow.

While it was important to look at speeding behavior in relationo messaging from the City about enforcing posted speed limits,e also looked at serious speeding violations separately. Serious

peeding violations were defined as driving at least 13 km/h overhe speed limit. Again, these analyses were conducted both for allbservations as well as those observations taking place during lowraffic density times.

Finally, with regard to analyses related to red-light running,ogistic regression was used to investigate the difference betweenhe number of red light running violations of drivers at intersectionsith photo enforcement compared to the number of red light run-ing violations of drivers at two comparison intersections withouthoto enforcement.

. Results

.1. Time series analyses related to red-light running

Analyses of crash data show (see Table 1) that the installationf the first set of cameras was associated with a non-significantncrease in right angle crashes of 12.75%, followed by a highly sig-ificant decrease of 46.10% (p = 0.003); a non-significant decreasef 10.68%; and a non-significant increase of 10.96%.

Fig. 1 contains the experimental group with right angle crashesrom Winnipeg and the comparison group with right angle crashesrom New Brunswick. The vertical line indicates the time when therst set of cameras was installed in Winnipeg, which coincides withhe initial 12.75% decrease in right angle crashes. The significantrop of 46.10% in right-angle crashes is clearly visible in this figure.

In an effort to gauge whether there were spill-over effects in theity of Winnipeg the analyses of right-angle crashes were replicatedith data from all signalized intersections in Winnipeg, exclud-

ng the 48 camera intersections. These analyses showed that otherntersections without cameras in Winnipeg did not experience aomparable significant decrease in right-angle crashes, nor did theyxperience an increase.

able 1ercentage change in right-angle crashes at 48 camera intersections.

Red-light running: right angle crashes

Installation of cameras Coefficient ω (s.e.) p-Value % change(100 ×(eω − 1))

First 12 cameras 0.120 (0.108) 0.267 12.75%Second 12 cameras −0.618 (0.211) 0.003 −46.10%Third 12 cameras −0.113 (0.285) 0.691 −10.68%Fourth 12 cameras 0.104 (0.197) 0.598 10.96%

Third 12 cameras −0.135 (0.091) 0.139 14.45%Fourth 12 cameras −0.029 (0.080) 0.718 −2.86%

Regarding rear-end crashes, the analyses do suggest the instal-lation of cameras was associated with an initial significant 42.33%increase in crashes (p = 0.004). This was followed by a non-significant 18.86% decrease (p = 0.085). The effects associated withthe installation of the third (14.45%; p = 0.139) and fourth set ofcameras (−2.86%; p = 0.718) were not significant (see Table 2).

Analyses of rear-end crashes at other intersections in Winnipegwhere no cameras were installed were also conducted. There wasa 25.36% increase which approached significance at the 5% level(p = 0.051) in rear-end crashes. The installation of following sets ofcameras was not associated with significant effects.

3.2. Time series analyses related to speeding

Time series analyses of injury crashes revealed that there wereno significant effects. Only the installation of the fourth set of cam-eras was associated with an effect that approached significance.Specifically, the installation of the last set of cameras was associ-ated with a 23.51% decrease (p = 0.053) in injury crashes at cameraintersections (see Table 3). The analyses of speeding related injurycrashes were replicated with data from Winnipeg at intersections,

excluding the 48 camera intersections. No significant effects werefound.

Table 3Percentage change in injury crashes at 48 camera intersections.

Speeding: injury crashes

Installation of cameras Coefficient ω (s.e.) p-Value % change(100 × (eω − 1))

First 12 cameras 0.044 (0.131) 0.735 4.50%Second 12 cameras 0.060 (0.159) 0.707 6.18%Third 12 cameras 0.215 (0.151) 0.887 23.98%Fourth 12 cameras −0.268 (0.138) 0.053 −23.51%

W. Vanlaar et al. / Accident Analysis an

Table 4Percentage change in PDO crashes at 48 camera intersections.

Speeding: property damage only crashes

Installation of cameras Coefficient ω (s.e.) p-Value % change(100 × (eω − 1))

First 12 cameras 0.256 (0.113) 0.024 29.18%

sifn1ec0

wy(ae(

3

3

itt(a

3

oTiy

TP

TP

Second 12 cameras −0.437 (0.119) 0.001 −35.40%Third 12 cameras 0.183 (0.079) 0.021 20.08%Fourth 12 cameras −0.145 (0.087) 0.096 −13.50%

Regarding PDO crashes, Table 4 shows the installation of the firstet of cameras was associated with a significant 29.18% increasen PDO crashes at camera intersections (p = 0.024), but this wasollowed by a significant 35.40% decrease (p = 0.001), another sig-ificant 20.08% increase (p = 0.021), and another non-significant3.50% decrease (p = 0.096). When only considering significantffects, the net effect is virtually zero: a 0.20% increase in PDOrashes at camera intersections (ω = 0.256 − 0.437 + 0.183 yields a.20% increase using the formula (100 × (eω − 1))).

Analyses of PDO crashes at other intersections in Winnipeghere no cameras were installed were also conducted. These anal-

ses showed that there was a non-significant increase of 17.94%p = 0.057) followed by a non-significant 13.50% decrease (p = 0.059)ssociated with the use of the first set and second set of cam-ras respectively. The net effect is nominal: a 2.02% increaseω = 0.165 − 0.145).

.3. Intersection camera experiment

.3.1. Average speedAn examination of average speeds revealed little differences

n average speed before the installation of the cameras comparedo after the installation of the cameras. This was the case at bothhe intervention and comparison locations for all three conditions50 km/h winter; 60 km/h winter and summer; note that no datare available for the 50 km/h summer condition).

.3.2. Speeding violationsRegarding the number of speeding violations (at least 1 km/h

ver the speed limit) at the 50 km/h locations in the winter (seeable 5), decreases in speeding violations were apparent at thentervention location and the comparison locations but the anal-ses showed that the decrease at the intervention location was

able 5ercentage change in speeding violations (1 km/h over speed limit) at 50 km/h winter cam

Speeding violations: 50 km/h winter condition

Odds ratio (s.e.)

All trafficExperimental (main effect) 0.43 (0.008)

Comparison (interaction effect) 1.89 (0.042)

Low density trafficExperimental (main effect) 0.43 (0.008)

Comparison (interaction effect) 1.99 (0.045)

able 6ercentage change in serious speeding violations (13 km/h over speed limit) at 50 km/h w

Serious speeding violations: 50 km/h winter condition

Odds ratio (s.e.)

All trafficExperimental (main effect) 0.46 (0.027)

Comparison (interaction effect) 1.37 (0.086)

Low density trafficExperimental (main effect) 0.46 (0.027)

Comparison (interaction effect) 1.44 (0.092)

d Prevention 62 (2014) 238– 247 243

significantly greater: the decrease in speeding violations at theintervention location was 57% while the decrease was only 19%at the comparison locations. When looking at low traffic densityobservations only, the difference is more pronounced: a decreaseof 57% at the intervention location and only 14% at the comparisonlocations.

When looking at serious speeding violations (13 km/h overspeed limit) at the 50 km/h locations in the winter (see Table 6),a decrease of 54% in serious violations was found at the interven-tion location. The reduction at comparison sites in serious speedingviolations was 37%. The results are somewhat more pronouncedfor serious speeding violations when looking at low traffic densityobservations only: a decrease of 54% at the intervention locationversus only 34% at the comparison locations.

Data from the 60 km/h location in the winter show that therewas a more modest, yet still significant decrease of 12% in speed-ing violations (at least 1 km/h over speed limit) at the interventionsite (see Table 7). At the comparison sites, however, there was asignificant increase in speeding violations of 13%. When lookingat low traffic density data only, an increase in speeding violationswas apparent at the intervention and comparison locations butthe increase was significantly smaller at the intervention location.The increase at the intervention site was 40% while it was 57% atthe comparison locations. A possible explanation of this effect isprovided in Section 4.2.

When analyzing serious speeding violations (13 km/h overspeed limit) at the 60 km/h locations in the winter, a different pic-ture emerges. The increase in serious speeding violations is 83% atthe intervention location while the increase was significantly lowerat 13% at the comparison locations. Comparable conclusions can bedrawn when looking at low traffic density data only (an increase of114% at the intervention site versus an increase of 15% at the com-parison site). These differences are significant (see Table 8). Theseresults are also discussed in Section 4.2.

Regarding the 60 km/h locations in the summer, there was asignificant 22% decrease in speeding violations (at least 1 km/h overspeed limit) at the intervention site and a significant 13% increaseat the comparison sites. Results for low traffic density observationsare a significant 27% decrease at the intervention location and a

significant 23% increase at the comparison locations (see Table 9).Of interest, the general increase in speeding violations apparent atthe 60 km/h location in the winter condition is no longer apparentin the summer condition.

era sites.

p-Value % change

<0.001 −57% ((0.43 − 1) × 100)<0.001 −19% (((0.43 × 1.89) − 1) × 100)

<0.001 −57% ((0.43–1) × 100)<0.001 −14% (((0.43 × 1.99) − 1) × 100)

inter camera sites.

p-Value % change

<0.001 −54% ((0.46 − 1) × 100)<0.001 −37% (((0.46 × 1.37) − 1) × 100)

<0.001 −54% ((0.46 − 1) × 100)<0.001 −34% (((0.46 × 1.44) − 1) × 100)

244 W. Vanlaar et al. / Accident Analysis and Prevention 62 (2014) 238– 247

Table 7Percentage change in speeding violations (1 km/h over speed limit) at 60 km/h winter camera sites.

Speeding violations: 60 km/h winter condition

Odds ratio (s.e.) p-Value % change

All trafficExperimental (main effect) 0.88 (0.007) <0.001 −12% ((0.88 − 1) × 100)Comparison (interaction effect) 1.28 (0.015) <0.001 +13% (((0.88 × 1.28) − 1) × 100)

Low density trafficExperimental (main effect) 1.40 (0.014) <0.001 +40% ((1.40 − 1) × 100)Comparison (interaction effect) 1.12 (0.017) <0.001 +57% (((1.40 × 1.12) − 1) × 100)

Table 8Percentage change in serious speeding violations (13 km/h over speed limit) at 60 km/h winter camera sites.

Serious speeding violations: 60 km/h winter condition

Odds ratio (s.e.) p-Value % change

All trafficExperimental (main effect) 1.83 (0.043) <0.001 +83% ((1.83 − 1) × 100)Comparison (interaction effect) 0.62 (0.021) <0.001 +13% (((1.83 × 0.62) − 1) × 100)

Low density traffic

lwtctitss

3

t5Mv

TP

TP

Experimental (main effect) 2.14 (0.063)

Comparison (interaction effect) 0.54 (0.025)

The results for serious speeding violations (13 km/h over speedimit) at the 60 km/h locations in the summer are as follows. There

as a non-significant 3% decrease in serious speeding violations athe 60 km/h intervention site. More importantly, the effect at theomparison locations is significantly different from the interven-ion location. At these comparison locations there was a significantncrease in serious speeding violations of 18%. When using lowraffic density data only the results are a decrease approachingignificance of 7% (p = 0.056) at the intervention location and aignificant 32% increase at the comparison locations (see Table 10).

.3.3. Red-light runningThe results regarding the number of red light violations show

hat there was a modest, yet significant difference between the0 km/h intervention site and its comparison sites in the winter.ore precisely, there was a significant 26% decrease in red light

iolations after installation of photo enforcement cameras at the

able 9ercentage change in speeding violations (1 km/h over speed limit) at 60 km/h summer c

Speeding violations: 60 km/h summer condition

Odds ratio (s.e.)

All trafficExperimental (main effect) 0.78 (0.012)

Comparison (interaction effect) 1.45 (0.023)

Low density trafficExperimental (main effect) 0.73 (0.017)

Comparison (interaction effect) 1.68 (0.041)

able 10ercentage change in serious speeding violations (13 km/h over speed limit) at 60 km/h s

Serious speeding violations: 60 km/h summer condition

Odds ratio (s.e.)

All trafficExperimental (main effect) 0.97 (0.023)

Comparison (interaction effect) 1.22 (0.033)

Low density trafficExperimental (main effect) 0.93 (0.036)

Comparison (interaction effect) 1.42 (0.060)

<0.001 +114% ((2.14 − 1) × 100)<0.001 +15% (((2.14 × 0.54) − 1) × 100)

intervention location and this decrease was significantly lower(21%) at the comparison locations (see Table 11).

At the 60 km/h locations during the winter the decrease in redlight violations was 44% at the intervention site versus a decreaseof only 3% at the comparison locations (see Table 11).

The results for the 60 km/h locations in the summer show thatthere was a significant 40% decrease in red light violations at theintervention site and a significant 20% increase at the comparisonsites (see Table 11).

4. Discussion

The objective of this study was to evaluate the impact of

Winnipeg’s photo enforcement safety program on crashes and vio-lations related to red-light running and speeding. Overall, it appearsthe photo enforcement had a positive net effect on traffic safety inthe city of Winnipeg.

amera sites.

p-Value % change

<0.001 −22% ((0.78 − 1) × 100)<0.001 +13% (((0.78 × 1.45) − 1) × 100)

<0.001 −27% ((0.73 − 1) × 100)<0.001 +23% (((0.73 × 1.68) − 1) × 100)

ummer camera sites.

p-Value % change

0.205 −3% ((0.97 − 1) × 100)<0.001 +18% (((0.97 × 1.22) − 1) × 100)

0.056 −7% ((0.93 − 1) × 100)<0.001 +32% (((0.93 × 1.42) − 1) × 100)

W. Vanlaar et al. / Accident Analysis and Prevention 62 (2014) 238– 247 245

Table 11Percentage change in red-light running violations at 50 km/h site, winter condition and 60 km/h site, winter and summer condition.

Odds ratio (s.e.) p-Value % change

Red-light running violations: 50 km/h winter conditionExperimental (main effect) 0.74 (0.010) <0.001 −26% ((0.74 − 1) × 100)Comparison (interaction effect) 1.07 (0.018) <0.001 −21% (((0.74 × 1.07) − 1) × 100)

Red-light running violations: 60 km/h winter conditionExperimental (main effect) 0.56 (0.013) <0.001 −44% ((0.56 − 1) × 100)Comparison (interaction effect) 1.73 (0.073) <0.001 −3% (((0.56 × 1.73) − 1) × 100)

tcilpgetpptlsr“w

4

niewiattaif

Wtraa

ociWittepbfi

r

Red-light running violations: 60 km/h summer conditionExperimental (main effect) 0.60 (0.041)Comparison (interaction effect) 2.00 (0.151)

Of importance, the time series analyses allowed for an examina-ion of the effect of intersection cameras over time. Such analysesontrol for long term trends in the number of collisions, changesn traffic volume, and changes in other risk factors influencing col-ision occurrence over time. Furthermore, the findings from bothortions of the study (i.e., time series and quasi-experiment) sug-est a consistency in findings across the two methods used tovaluate the effects of the photo enforcement cameras. Never-heless, it should be noted that the Winnipeg photo enforcementrogram may not be directly comparable to photo enforcementrograms in other jurisdictions. For example, not every jurisdic-ion communicates the same messages related to enforcing speedimits and tolerances to the public. Also, not all jurisdictions neces-arily use the less stringent “violation on entry” approach to defineed-light running violations, but may rather use the more stringentviolation on exit” approach. Thus, results should be interpretedith caution.

.1. Time series analyses

Findings of the time series analyses related to red-light run-ing revealed general reductions in right-angle crashes (46%) and

ncreases in rear-end crashes (42%). The initial 42% increase in rear-nd crashes was followed by a 19% decrease, although this decreaseas not significant at the 5%-level. While this decrease is not signif-

cant, it does suggests it is possible drivers adjusted their behaviorfter they initially used their brakes more abruptly for fear of beingicketed when approaching camera intersections. This could leado an initial increase in crashes followed by a decrease. More datare needed to confirm whether the initial increase in Winnipeg wasndeed followed by such a decrease or not. Our results suggest thaturther monitoring of this possible effect is warranted.

Regardless of whether the increase in rear-end crashes ininnipeg was truly followed by a decrease or not, it warrants men-

ioning that right-angle crashes have a higher injury and fatalityate than rear-end crashes (Helai et al., 2008), so there is generally

net benefit in terms of lives saved and serious injuries preventeds well as a crash-cost benefit (Council et al., 2005).

The analyses also suggest there may have been a spill-over effectn rear-end crashes at other intersections in Winnipeg where noameras were installed. A notable effect that was found was a 25%ncrease associated with the installation of the first set of cameras.

hile strictly speaking this effect was not significant (p = 0.051), its sufficiently close to the threshold to warrant further investiga-ion. One possible explanation, albeit at the level of speculation, ishat this increase at other intersections in Winnipeg without cam-ras is the result of drivers incorrectly assuming that cameras areresent at these intersections and, as a consequence, they used their

rakes abruptly when approaching an intersection for fear of beingned. No data were available to test this hypothesis.

Regarding the results from the time series analyses of speedingelated crashes, there were no significant effects with respect to

<0.001 −40% ((0.60 − 1) × 100)<0.001 +20% (((0.60 × 2.00) − 1) × 100)

injury crashes. Of interest, the installation of the last set of cameraswas associated with a decrease in injury crashes of 23.51% that wasalmost significant (p = 0.053). Regarding PDO crashes, there wereno increases or decreases in PDO crashes, not at the camera inter-sections (when combining the significant effects, the net result is a0.2% increase) and not at other intersections in Winnipeg. Furthermonitoring is recommended.

It was not possible to analyze the effects of photo enforcementon fatalities as there were very few fatalities over the study timeperiod, with many monthly counts for fatalities being zero. Otherlimitations of the time series analyses include the lack of monthlycounts of such variables as unemployment, population, traffic, etc.in Winnipeg and in New Brunswick for the monitoring period. Tofurther bolster our results the use of other analysis techniques canbe considered as recommended by Hamed et al. (1999). Also, nodetailed or reliable information regarding intersection design andpossible crash causes was available to refine the analyses. Finally,relevant data from another city could not be obtained for use as acomparison group though efforts were made to choose the mostsuitable jurisdiction for comparison. Thus, provincial data wereused. This was not ideal and can be considered a limitation of thisstudy.

4.2. Intersection camera experiment

The results of the intersection camera experiment show photoenforcement does have a protective effect on speeding behavior.However, it appears that average speed was not sufficiently sen-sitive to measure the impact of photo enforcement in Winnipeg.This can perhaps be explained by the fact that the evaluation tookplace in an area where a photo enforcement program had beenconducted for several years prior to the start of this evaluation.Furthermore, public awareness efforts began two months prior tothe beginning of the program and the campaigns were continu-ously maintained post-implementation. In fact, a public opinionpoll conducted in the Winnipeg Census Metropolitan Area in Mayof 2009 found that about 95% (95%-CI: 93.0–96.4%) of respondentsconfirmed they knew about the program (Vanlaar et al., 2011). Itis possible that spill-over effects led to low average speeds in thebefore period prior to the beginning of the experiment. Indeed, theaverage speed at all sites included in the evaluation in the beforeperiod was well under the maximum speed limit.

Of importance, if there was spill-over in Winnipeg, then it is fairto assume it has equally affected all sites included in the evaluation,regardless of whether they were used as an experimental locationor comparison location. The expectation, then, is that any effect ofthe impact of photo enforcement would be underestimated. Results

of this evaluation, if any, are therefore conservative. In the extreme,this would mean no more safety gains can be made in Winnipeg byinstalling more photo enforcement cameras at intersections and nomeaningful conclusions could be drawn from such an evaluation.

2 sis an

Hr

lossditsptkst

lwadtSoo(td

6tcsiwildcbtala

pl

Ttsaoomaso

5

mse

46 W. Vanlaar et al. / Accident Analy

owever, when looking at another evaluation measure – violationsather than average speed – this does not hold true.

Generally speaking, there were either decreases in speeding vio-ations at the experimental site compared to smaller decreasesr increases at the comparison sites or there were increases inpeeding violations at the experimental site that were significantlymaller than the increases at the comparison sites. However, theata also suggest that photo enforcement may be more effective

n preventing speeding violations in general (at least 1 km/h overhe speed limit) but is perhaps less effective in preventing seriouspeeding violations (at least 13 km/h over the speed limit). Oneossible explanation to consider is that serious speeding viola-ions are more commonly committed by high-risk drivers. It isnown that many traffic safety measures are less effective withuch high-risk drivers because they are less amenable to changingheir behavior (Robertson et al., 2010; Simpson et al., 2004).

The fact that this pattern of increased serious speeding vio-ations was primarily apparent at one 60 km/h experimental site

hile less pronounced at the other 60 km/h site and not at allt the 50 km/h experimental site, could perhaps be the result ofifferences in the intersection design between those experimen-al locations that have contributed to these different patterns.uch differences need further investigation, preferably by meansf a detailed and controlled experiment in an effort to enhanceur understanding of how best to implement photo enforcementGarber et al., 2007; Washington and Shin, 2005). A limitation ofhis study is that data regarding such differences in intersectionesign were not available.

Small increases in speeding violations were apparent at all three0 km/h winter locations (1 experimental plus 2 comparison loca-ions) when drivers could freely choose their speed. Note that dataollection for the 60 km/h sites started in the winter but, strictlypeaking, ended in the spring so it is possible that everywheren Winnipeg people started driving faster as a result of improved

eather conditions. This may partially explain why there is anncrease in speeding violations at all three locations, especially inight of the fact that this particular analysis only looks at thoserivers who could freely choose their speed. If changing weatheronditions was the cause of this increase it would indeed be evidentoth at the experimental locations as well as the comparison loca-ion; and, in fact, a significant difference between the experimentalnd comparison locations was found confirming the increase in vio-ations was significantly greater at the comparison locations thatre not protected by photo enforcement.

Regarding red light running violations, analyses showed aositive impact of photo enforcement with significantly fewer vio-

ations after installation of cameras across all conditions studied.The intersection cameras study design however, was limited.

he program was in place long before the study was conducted, thushe random assignment of sites to the intervention and compari-on groups was not possible. Further, since the use of cameras waslready widespread prior to the study, it is possible that the numberf red-light running and speeding violations may have already beenn the decline prior to the collection of data, and that both experi-ental and control locations included in this study may have been

ffected by spill-over effects. While it may be fair to assume thatpill-over effects may have had a comparable impact on both typesf locations, it was not possible to account for this potential bias.

. Conclusions

In sum, the time series analyses revealed that the photo enforce-ent program seems to have had a positive net effect on traffic

afety levels in the city of Winnipeg. Further, the intersection cam-ra experiment results suggest photo enforcement has a protective

d Prevention 62 (2014) 238– 247

effect leading to fewer speeding violations and red light runningviolations. Thus, both methods used to investigate the impact ofWinnipeg’s photo enforcement on safety at intersections revealedimprovements in traffic safety. These latter findings also help vali-date the former because fewer violations should logically result infewer crashes. Perhaps most importantly, by virtue of using timeseries analyses in this study, it became possible to demonstrate thatinitial increases in rear-end crashes can be followed by decreases insuch crashes. Many other studies use simple before/after designs,more recently often augmented with a Bayesian component. Whilesuch studies are important, they are limited in that they cannotillustrate the effects of an intervention over time. Indeed, our find-ings suggest that decreases in the most severe crash types (i.e.,right-angle crashes) may remain consistent, but increases in lesssevere crash types (i.e., rear-end crashes) over time may shrink andeventually turn into decreases. While the decrease in our study wasnot significant, our results do warrant further investigation of sucha possible effect using time series analysis. It is recommended toupdate the analyses when more crash data become available sothese effects can be further monitored and investigated.

Acknowledgements

The following individuals reviewed the times series analyses ofcrashes. We are grateful for their insightful feedback that helped usto further improve the quality of our work.

• Scott Masten, Research Manager II, California Department ofMotor Vehicles, Research and Development Branch

• Robert Hagge, Research Manager II, California Department ofMotor Vehicles, Research and Development Branch

• Prof. Tom Brijs, Program Leader Traffic Safety, TransportationResearch Institute (IMOB), Hasselt University

References

Andreassen, D., 1995. A Long Term Study of Red Light Cameras and Crashes. RoadResearch Board, Victoria, Australia.

Blakey, L.T., 2003. Red-light cameras: effective enforcement measures for intersec-tion safety. Institute of Transportation Engineers 73 (3), 34–43.

Brault, M., Auger, A., Montégiani, M., 2007. La sécurité aux intersections: Analysedes comportements des conducteurs au feu rouge. In: Présentation à la XVIIeConférence Canadienne Multidisciplinaire sur la Sécurité Routière à Montréal,3–6 juin 2007.

Burkey, M., Obeng, K., 2004. A Detailed Investigation of Crash Risk Reduction Result-ing from Red Light Cameras in Small Urban Areas. North Carolina Agriculturaland Technical State University, Greensboro, NC.

CCMTA, 2010. Speed and Intersection Safety Management (SISM). Annual Monitor-ing Report 2009. Canadian Council of Motor Transport Administrators, Ottawa.

Chen, G., Wilson, J., Meckle, W., Cooper, P., 2000. Evaluation of photo radar programin British Columbia. Accident Analysis and Prevention 32 (4), 517–526.

Cheung, S.Y., Coleri, S., Dundar, S., Ganesh, Tan, C.W., Varaiya, P., 2005. Traffic Mea-surement and Vehicle Classification with Single Magnetic Sensors. Journal of theTransportation Research Board, No. 1917. Transportation Research Board of theNational Academies, Washington, DC, pp. 173–181.

Cheung, S.Y., Varaiya, P., 2007. Traffic Surveillance by Wireless Sensor Networks:Final Report. UCB-ITS-PRR-2007-4. University of California, Berkeley.

Council, F.M., Persaud, B., Eccles, K., Lyon, C., Griffith, M.S., 2005. Safety Evaluationof Red-Light Cameras – Executive Summary Federal Highway Administration.FHWA-HRT-05-048.

Day, C.M., Premachandra, H., Brennan, T.M., Sturdevant, J.R., Bullock, D.M., 2010.Operational Evaluation of Wireless Magnetometer Vehicle Detectors at Sig-nalized Intersection. Journal of the Transportation Research Board, No. 2192.Transportation Research Board of the National Academies, Washington, DC, pp.11–23.

Elvik, R., 2005. Speed and road safety: synthesis evidence from evaluation studies.Transportation Research Record 1908, 59–69.

Erke, A., 2009. Red light for red light cameras? A meta-analysis of the effects of redlight cameras on crashes. Accident Analysis and Prevention 41 (5), 897–905.

Evans, L., 2004. Traffic Safety. Science Serving Society, Bloomfield Hills, MI.Evans, L., 2006. Traffic Safety, 2nd ed. Science Serving Society, Bloomfield Hills, MI.Garber, N.J., Miller, J.S., Abel, R.E., Eslambolchi, S., Korukonda, S.K., 2007. The Impact

of Red Light Cameras (Photo-Red Enforcement) on Crashes in Virginia. VirginiaTransportation Research Council, Charlottesville, VA.

sis an

G

H

H

H

H

K

K

M

M

M

M

M

M

N

P

P

Wilson, C., Willis, C., Hendrikz, J.K., Le Brocque, R., Bellamy, N., 2010. Speed cam-

W. Vanlaar et al. / Accident Analy

oldenbeld, C., Van Schagen, I., 2005. The effects of speed enforcement with mobileradar on speed and crashes: an evaluation study on rural roads in the Dutchprovince Friesland. Accident Analysis and Prevention 37, 1135–1144.

amed, M.M., Al-Eideh, B.M., Al-Sharif, M.M., 1999. Traffic accidents under the effectof the Gulf crisis. Safety Science 33, 59–68.

aoudi, A., Kavaler, R., Varaiya, P., 2008. Wireless magnetic sensors for traffic surveil-lance. Transportation Research Part C 16, 294–306.

elai, H., Chor, C.H., Haque, M.M., 2008. Severity of driver injury and vehicle dam-age in traffic crashes at intersections: a Bayesian hierarchical analysis. AccidentAnalysis and Prevention 40, 45–54.

ess, S., 2004. Analysis of the effects of speed limit enforcement cameras: differen-tiation by road type and catchment area. Transportation Research Record 1865,28–34.

ent, S., Corben, B., Fildes, B., Dyte, D., 1997. Red-light-running behaviour at red-light camera and control intersections. Monash University Accident ResearchCentre – Report #73.

loeden, C.N., Ponte, G., McLean, A.J., 2001. Travelling Speed and the Rate of CrashInvolvement on Rural Roads. Australian Transport Safety Bureau, Civic Square,ACT (CR 204).

anitoba, 2012. The Highway Traffic Act, C.C.S.M. c. H60. http://web2.gov.mb.ca/laws/statutes/archive/2009-06-10/h060e.php (accessed 03.04.12).

argulici, J.D., Yand, S., Tan, C.W., 2006. Evaluation of Wireless Traffic Sensors bySensys Networks, Inc. California Department of Transportation, Sacramento, CA.

cCleary, R., Hay, R., 1980. Applied Time Series Analysis for the Social Sciences. SagePublications, Beverly Hills, CA.

cGee, H.W., Eccles, K.A., 2003. Impact of red light camera enforcement on crashexperience. A synthesis of highway practice. In: NCHRP Synthesis 310.

inistry of Transportation Ontario, 1995. Photo Radar Safety Evaluation PreliminaryFour Month Speed Results. Safety Research Office. Ministry of TransportationOntario, Downsview, Ontario.

inistry of Transportation Ontario, 1998. Ontario Road Safety Annual Report 1997.Ministry of Transportation Ontario, Toronto.

ew Brunswick, 2012. Motor Vehicle Act. http://laws.gnb.ca/en/ShowPdf/cs/M-17.pdf (accessed 03.04.12).

alys, T., 2003. Research Decisions: Quantitative and Qualitative Perspectives, 3rded. Nelson, Scarborough, Ontario.

ilkington, P., Kinra, S., 2005. Effectiveness of speed cameras in preventing road traf-fic collisions and related casualties: systematic review. British Medical Journal,http://dx.doi.org/10.1136/bmj.38324.646574.AE, BMJ online, BMJ.com.

d Prevention 62 (2014) 238– 247 247

Retting, R.A., Farmer, C.M., McCartt, A.T., 2008. Evaluation of automated speedenforcement in Montgomery County, Maryland. Traffic Injury Prevention 9 (2),440–445.

Retting, R.A., Kyrychenko, S.Y., 2002. Reductions in injury crashes associated withred light camera enforcement in Oxnard, California. American Journal of PublicHealth 92, 1822–1825.

Retting, R.A., Williams, A.F., Farmer, C.M., Feldman, A.F., 1999. Evaluation of redlight camera enforcement in Fairfax, Virginia. Institute of Transport Economics69, 30–34.

Robertson, R., Marcoux, K., Wood, K., Vanlaar, W., 2010. Ten Percenters: Final Report.I-95 Corridor Coalition, Maine/Florida.

Sensys Networks, Inc., 2007. Evaluation of the Sensys Wireless Vehicle DetectionSystem by the California Center for Innovative Transportation. Sensys Networks,Inc., Berkeley, CA.

Shin, K., Washington, S., 2007. The impact of red light cameras on safety in Arizona.Accident Analysis and Prevention 39, 1212–1221.

Simpson, H., Beirness, D., Robertson, R., Mayhew, D., Hedlund, J., 2004. Hard coredrinking drivers. Traffic Injury Prevention 5 (3), 261–269.

StataCorp, 2009a. Stata Statistical Software. Release 11. StataCorp LP, College Station,TX.

StataCorp, 2009b. Stata Base Reference Manual. vol. 2: I-P. Release 11. StataCorp LP,College Station, TX.

StataCorp, 2009c. Stata Time-Series Reference Manual. Release 11. StataCorp LP,College Station, TX.

Tay, R.,2000. Do speed cameras improve road safety? In: Proceedings of the Sec-ond International Conference on Traffic and Transportation Studies. AmericanSociety of Civil Engineers, Reston.

Vanlaar, W., Robertson, R., Marcoux, K., July 2011. Evaluation of the Photo Enforce-ment Safety Program of the City of Winnipeg: Final Report. Traffic InjuryResearch Foundation, Ottawa, Canada.

Washington, S., Shin, K., 2005. The Impact of Red Light Cameras (AutomatedEnforcement) on Safety in Arizona. Arizona Department of Transportation,Phoenix.

eras for the prevention of road traffic injuries and deaths. Cochrane Database ofSystematic Reviews (11), Art. No. CD004607 (review).

Yaffee, R., 2000. An Introduction to Time Series Analysis and Forecasting with Appli-cations of SAS and SPSS. Academic Press Inc., New York, NY.