Red light camera interventions for reducing traffic violations and traffic crashes: A systematic review

Abstract Background Road traffic crashes are a major and increasing cause of injury and death around the world. In 2015, there were almost 6.3 million motor vehicle traffic crashes in the United States. Of these, approximately 1.7 million (27%) involved some form of injury and 32,166 (0.5%) resulted in one or more fatalities (National Highway Traffic Safety Administration, 2016, Traffic Safety Facts 2013: A Compilation of Motor Vehicle Crash Data from the Fatality Analysis Reporting System and the General Estimates System). The most common cause of urban crashes appears to be drivers running red lights or ignoring other traffic controls and injuries occur in 39% of all of these types of crashes (Insurance Institute for Highway Safety, IIHS, 2018, Red light running). While many drivers obey traffic signals, the possibility for violations exists due to issues such as driver distraction, aggressive driving behaviors, or a deliberate decision to ignore the traffic signal. One researcher suggests that eliminating traffic violations could reduce road injury crashes by up to 40% (Zaal, 1994, Traffic law enforcement: A review of the literature). Red light cameras (RLCs) are an enforcement mechanism that permit police to remotely enforce traffic signals; they may serve as a deterrent to drivers who intentionally engage in red light running (RLR). The one previous systematic review of RLCs found that they were effective in reducing total casualty crashes but also found that evidence on the effectiveness of cameras on red light violations, total crashes, or specific types of casualty crashes was inconclusive. However, this review searched only a small number of electronic databases and was limited to a handful of studies published in 2002 or earlier. Objectives This report updates and expands upon the previous Cochrane systematic review of RLCs. The aim of this review is to systematically review and synthesize the available evidence on the effectiveness of RLCs on the incidence of red light violations and the incidence and severity of various types of traffic crashes. Search Methods This study uses a four‐part search strategy that involves: (a) searching 27 online electronic bibliographic databases for published and unpublished evaluations of RLCs; (b) searching the websites of 46 international institutes and research agencies focusing on transportation issues for reports and other gray literature; (c) searching the reference lists of published studies to identify additional published and unpublished works; and (d) conducting a keyword search using Google and Google Scholar to search for additional gray literature. Selection Criteria The criteria for inclusion were determined before the search process began. To be eligible, studies must have assessed the impact of RLCs on red light violations and/or traffic crashes. Studies must have employed a quantitative research design that involved randomized controlled trials, quasi‐random controlled trials, a controlled before‐after design, or a controlled interrupted time series. Research that incorporated additional interventions, such as speed cameras or enhanced police enforcement, were excluded, although normal routine traffic enforcement in the nonintervention control condition was not excluded. Both published and unpublished reports were included. Studies were eligible regardless of the country in which they were conducted or the date of publication. Qualitative, observational, or descriptive studies that did not include formal comparisons of treatment and control groups were excluded from this research. Data Collection and Analysis Initial searches produced a total of 5,708 references after duplicates were removed. After title and abstract screening, a total of 121 references remained. Full‐text review of these works identified 28 primary studies meeting the inclusion criteria, in addition to the 10 studies identified in the prior Cochrane review. Because several of the primary studies reported on multiple independent study areas, this report evaluates 41 separate analyses. At least two review authors independently assessed all records for eligibility, assessed methodological risk of bias, and extracted data from the full‐text reports; disagreements were resolved by discussion with a third review author. To facilitate comparisons between studies, a standardized summary measure based on relative effects, rather than differences in effects, was defined for each outcome. Summary measures were calculated for all studies when possible. When at least three studies reported the same outcome, the results were pooled in a meta‐analysis. Pooled meta‐analyses were carried out when at least three studies reported the same outcome; otherwise, the results of individual studies were described in a narrative. Heterogeneity among effect estimates was assessed using χ 2 tests at a 5% level of significance and quantified using the I 2 statistic. EMMIE framework data were coded using the EPPIE Reviewer database. Results The results of this systematic review suggest that RLCs are associated with a statistically significant reduction in crash outcomes, although this varies by type of crash, and suggest a reduction in red light violations. RLCs are associated with a a 20% decrease in total injury crashes, a 24% decrease in right angle crashes and a 29% decrease in right angle injury crashes. Conversely, however, RLCs are also associated with a statistically significant increase in rear end crashes of 19%. There was also some evidence that RLCs were associated with a large reduction in crashes due to red light violations. There is no evidence to suggest that study heterogeneity is consistently explained by either country or risk of bias, nor did the presence or absence of warning signs appear to impact the effectiveness of RLCs. Studies accounting for regression to the mean tend to report more moderate decreases for right angle crashes resulting in injury than studies not accounting for regression to the mean. Studies with better control for confounders reported a nonsignificant decrease in right angle crashes, compared with a significant decrease for all studies. Authors' Conclusions The evidence suggests that RLCs may be effective in reducing red light violations and are likely to be effective in reducing some types of traffic crashes, although they also appear linked to an increase in rear end crashes. Several implications for policymakers and practitioners have emerged from this research. The costs and benefits of RLCs must be considered when implementing RLC programs. The potential benefits of a reduction in traffic violations and in some types of injury crashes must be weighed against the increased risk of other crash types. The economic implications of operating an RLC program also must be considered, including the costs of installation and operation as well as the economic impact of RLC effects.

accounting for regression to the mean tend to report more moderate decreases for right angle crashes resulting in injury than studies not accounting for regression to the mean. Studies with better control for confounders reported a nonsignificant decrease in right angle crashes, compared with a significant decrease for all studies.
Authors' Conclusions: The evidence suggests that RLCs may be effective in reducing red light violations and are likely to be effective in reducing some types of traffic crashes, although they also appear linked to an increase in rear end crashes. Several implications for policymakers and practitioners have emerged from this research.
The costs and benefits of RLCs must be considered when implementing RLC programs. The potential benefits of a reduction in traffic violations and in some types of injury crashes must be weighed against the increased risk of other crash types. The economic implications of operating an RLC program also must be considered, including the costs of installation and operation as well as the economic impact of RLC effects.
1 | PLAIN LANGUAGE SUMMARY 1.1 | Red light cameras (RLCs) reduce injuries but may have no effect on total crashes RLCs photograph violators at traffic signals. They can reduce red light running (RLR), total injury crashes, and right angle crashes. However, they may also increase the risk of rear end crashes. The impact of RLCs on other types of crashes, including total crashes overall, is unclear.

| What is this review about?
Road traffic crashes are a major cause of injury and death around the world. Many crashes occur because drivers run red lights. RLCs

| What studies are included?
Included studies measure RLC effectiveness by comparing intersections with cameras to those without them. Studies that examined area-wide programs, in which RLCs operated at some but not all signalized intersections in the community were also included.
Before-after studies were only included when they had a distinct control group and collected data for treatment and control conditions both before and after RLCs were put into operation. Studies involving additional interventions, such as speed cameras or enhanced police enforcement, were excluded.
This review summarizes 38 studies that contain 41 eligible evaluations of the effects of RLCs on RLR and/or traffic crashes. The studies come from four countries, with the majority carried out in the United States or Australia. Five of the 38 studies were assessed as having a low risk-of-bias and eight were assessed as having a moderate risk-of-bias. any given time, thus ensuring orderly movement of traffic, reducing delays for waiting vehicles, and reducing the frequency of vehicular crashes (Federal Highway Administration [FHA], 2004a). The Federal Highway Administration's Manual on Uniform Traffic Control Devices (FHA, 2012) specifically identifies situations where traffic conditions require the installation of traffic signals; generally, these relate to situations where conflicting traffic movements that create crash potential could exist (Bochner & Walden, 2010).

| The use of RLCs as an intervention
Motorists run red lights for a variety of reasons. However, survey research suggests that many drivers consider RLR to frequently be an intentional act that has few legal consequences (FHA, 2004b). There are a number of engineering countermeasures that focus on engineering design as a way to reduce RLR. One increasingly popular method of enforcing compliance with traffic signals is through the use of RLCs.
RLCs are a fully automated photo detection system that includes three key elements: cameras, sensors or triggers, and a computer.
The cameras may take still or video images, or both; modern systems generally use digital cameras but some older systems may use 35-mm cameras. They may be located on one arm of an intersection where a RLR problem has been identified or be placed on all four corners of an intersection, so that vehicles coming from any direction may be photographed from multiple angles. Cameras are activated if the vehicle is moving over the triggers at a predetermined speed; if the vehicle has stopped on an induction loop or activates only the first of the two triggers, the computer will not signal the cameras. Most systems take at least two photographs and also superimpose the date and time of the violation, the location of the intersection, the speed at which the vehicle was traveling and the amount of time that elapsed between the light turning red and the vehicle entering the intersection (FHA, 2004b).
After the RLCs capture images of vehicles as they violate a red traffic signal and the evidence is reviewed, penalty tickets are sent to the address where the violating vehicle is registered. RLCs thus have the potential to reduce traffic law offenses by increasing offenders' perceptions of the risks of being caught and being brought to justice if they run a red light.
RLCs permit police to remotely enforce traffic signals. Unlike traditional manual enforcement methods, which are resource intensive and high risk, RLCs operate continuously and without human intervention, freeing up officers to engage in other activities. They do not lead to potentially dangerous high-speed pursuits and they provide a physical record of all violations (Bochner & Walden, 2010). Their mechanical nature also reduces the possibility of accusations of human bias, discrimination, or selective enforcement (Aeron-Thomas & Hess, 2005).
Studies have shown that drivers who intentionally engage in RLR appear to be most likely to be influenced by countermeasures of this type (FHA, 2004b).
The use of RLCs, however, remains somewhat controversial, particularly in the United States. Some police departments in the United States have had difficulty sustaining the financial viability of RLC programs, there have been a number of legal challenges to the use of RLCs, and their effectiveness in reducing RLR and vehicular crashes has been questioned (Langland-Orban, Pracht & Large, 2008;IIHS, 2018). In the United Kingdom, RLCs are generally more accepted as bringing about positive road safety benefits with a rapid growth in their numbers since initial use began in the 1990s (see e.g., Hooke, Knox & Portas, 1996).

| Prior reviews
One previous Cochrane systematic review of the effect of RLCs on the incidence of red light violations as well as the incidence and severity of road crashes and casualties has been conducted, examining research published in or before 2002 (Aeron- Thomas & Hess, 2005). Although no randomized controlled studies were located, the review did identify a number of CBA studies. The study concluded that RLCs were effective in reducing the total number of casualty crashes but also found that evidence regarding the effect of RLCs on red light violations, total collisions, or specific types of casualty crashes was inconclusive. The review concluded that larger and better-controlled studies were needed.

| The use of EMMIE within systematic reviews
This review was conducted in support of the efforts of the What Works Centre for Crime Reduction, which is hosted by the UK College of Policing. The What Works Centre emphasizes the development of an evidence-based approach to policing by coordinating collaborations among academics and practitioners and creating a program to foster systematic reviews of research into policing and crime reduction practices. The Centre focuses on providing "robust and comprehensive evidence that will guide decision-making on public spending" (College of Policing, 2016).
The results of this review will be incorporated in an online toolkit devised by researchers at the University College of London (UCL) Jill Dando Institute of Security and Crime Science and hosted by the What Works Centre. The toolkit uses the EMMIE framework, which assesses interventions based on five key dimensions: effect, mechanism, moderators, implementation, and economic cost (Johnson, Tilley & Bowers, 2015).

| Contribution of this review
This systematic review has expanded and updated the previous Cochrane systematic review (Aeron-Thomas & Hess, 2005), which only searched a small number of electronic databases and only included a small number of studies published in 2002 or earlier. Since this study was conducted, the use of RLCs has expanded considerably (IIHS, 2020). RLC technology has also continued to improve; for example, new radar technology has been developed that improves the images obtained from the camera and also allows the system to enforce other traffic violations in addition to RLR (Mitchell, 2012).
This updated review involves broader and more extensive searches and incorporates more recent research from as many countries as possible, as well as carrying out more detailed and extensive metaanalyses and examining economic data when available. Additionally, the review has been expanded to include information from the EMMIE framework. The results of this review have the potential to inform the police and influence government policies and procedures intended to increase traffic safety.

| OBJECTIVES
The main objective of this review was to assess the impact of RLCs on the incidence of red light violations and the incidence and severity of traffic crashes by locating and examining all the major empirical studies on the effect of RLCs on traffic patterns. The update has been expanded by including information under the EMMIE framework (see above) on mechanisms, moderators, implementation, and economic costs of RLC interventions. The description of each study includes the setting (e.g., nature of roads), theoretical basis for the intervention, characteristics, and outcomes (including traffic law violations).
Where sufficient numbers of well-designed controlled evaluations were identified, estimates of the effect of interventions on the defined primary outcome (number of red light violations) and secondary outcomes (e.g., road traffic crashes) are included to assess the effectiveness of interventions. In addition to examining the impact of RLCs on road traffic crashes overall, the effect on different types of traffic crashes, such as rear end and right angle crashes, was evaluated separately. This study has also investigated potential moderators of intervention effects, and summarized the different aspects of implementation of traffic enforcement devices and their respective costs. Although every effort was made to comply with the original protocol, some changes were deemed to be essential. These primarily occurred because the original protocol was created by the researchers at FIU, before the FIU/LSHTM collaboration developed. After this collaboration began, some modifications to the protocol were necessary to conform to the requirements of the grant that the LSHTM researchers received from the UK College of Policing. Following is a list of deviations from the protocol: 1. The original FIU protocol included a requirement that studies must collect data for a minimum of one year to be eligible for inclusion in the review. This requirement was removed at the request of the LSHTM researchers. However, the majority of the included studies still meet the original requirement.
2. The list of keywords for the agency and gray literature searches was expanded from the original list in the protocol, as additional relevant keywords were identified. The LSHTM search protocol for electronic databases was also added.
3. The original screening and review process was revised to include the incorporation of studies identified by LSHTM and the National Police Library.
4. The use of the EPPI Reviewer 4 software and the EMMIE framework data extraction were incorporated to meet the requirements of the LSHTM agreement with the College of Policing.

| Types of studies
To be eligible for inclusion, a study must have measured the effectiveness of RLCs by comparing intersections that received the treatment (the treatment condition) with intersections that did not (the control condition). We identified studies were eligible for inclusion if they involved one of the following research designs, as they were defined in the original review (based on the Cochrane Effective Practice and Organisation of Care group): 1. Experimental design/RCT: This category included studies that used random assignment to assign intersections to the treatment and control groups.
2. Quasi-random design/quasi-RCT: This category included studies that allocated the treatment and control conditions using quasirandom processes, rather than truly randomizing treatment allocation.
3. CBA design: This quasi-experimental category included studies in which data were collected for both treatment and control conditions before and after the treatment was initiated.

| Types of interventions
Eligible studies must have tested the effect of RLCs on traffic red light violations or crashes. An RLC is considered to be a digital or film still and/or video camera that is used to detect red light violators and identify them so that they may be charged with their violations.
Studies that examined RLCs as part of a larger traffic enforcement initiative, specifically those that examined the joint effect of red light and speed cameras, were excluded.
Studies were included when the interventions included cameras at intersections or junctions that were designed to detect red light violators.
Studies examining area-wide programs where RLCs operated at some but not all signalized intersections in the community were also included.

| Types of outcome measures
Eligible studies had to have measured and reported data on at least one of the following outcome measures: • Red light violations, based on the number of vehicles passing through a junction after entering on a red light. Vehicles that enter a junction on a yellow/amber light but do not clear the intersection before the light changes to red are not considered to be violators.
• Number, severity, and type of road traffic crashes. This may include the number of total crashes, the number of crashes resulting in injury, the number of property damage-only (PDO) crashes, and the number of specific types of crashes (e.g., rear end crashes; right angle crashes).
Data on economic outcomes, including the costs of providing the intervention and the income it generated, and process outcomes (e.g., implementation data) were also collected when available.

| Types of settings
Studies were eligible regardless of the country in which they were conducted or the form in which they were published. When studies were not published in English, efforts were made to obtain translations. There was only one possibly-eligible study for which this was not possible (Giaever & Tveit, 1998).
No date restrictions were placed on this study. However, RLCs have only been used for traffic enforcement since the 1960s (Retting, Ferguson, & Hakkert, 2003)

| Electronic searches
For all agency website and gray literature searches, the following keywords were used: • Red AND light AND camera(s) • Red light AND camera(s) • Red AND light AND violation(s)

• Red light AND violation(s)
• Traffic AND camera(s) • Traffic AND violation(s) • Traffic AND light(s) These were adapted when necessary to meet specific requirements of individual search engines or to conform to international terminology variations and spelling conventions. Search terms were intentionally general in nature to ensure that searches cast the broadest possible net and that relevant background material was also identified.
All electronic database searches were conducted using the full search strategy as outlined in Appendix C and only superficially adapted for each database. Additional specialized searching was conducted by the College of Policing's National Police Library.

| Screening and review process
All studies identified through the LSHTM and National Police Library search process were first exported to the EndNote bibliographic database for de-duplication. Once duplicate records had been removed, records were combined with FIU search results in a spreadsheet to identify and remove further duplications. Once duplicate records were removed, each study was screened to determine if it met basic inclusion criteria, specifically: 1. The study dealt with the use of RLCs to reduce/prevent traffic light violations and/or traffic crashes.
2. The study included both a treatment and a comparison/control group.
3. The study reported results on at least one of the following outcome measures: incidence of red light violations, incidence of road traffic crashes, and severity of road traffic crashes.
4. Extraneous variables were controlled by at least one of the following methods: randomization, matching, or pre-test and posttest measures of violations and/or crashes.
At least two review authors independently examined the titles, abstracts, and keywords of electronic records for eligibility according to the inclusion criteria above. Results of this initial screening were cross-referenced between the authors and full-texts were then obtained for all potentially relevant reports of studies. The publication status of the study (unpublished vs. published) did not affect study eligibility.
The full-text reports of potentially eligible studies were independently assessed for final inclusion in the review by two review authors using screening codes in EPPI Reviewer 4. Any disagreements were resolved by discussion with a third review author. Reference lists of all eligible trials were inspected for further eligible studies.

| Data extraction and management
All studies were managed using the EPPI Reviewer 4 software. For the EMMIE framework data extraction, data were coded independently by two review authors in EPPI Reviewer, using a standardized data coding set (see Appendix D: EPPI Reviewer standardized data coding set for EMMIE framework). The remaining coding of studies (including study characteristics, risk of bias, measurement of effect) was conducted using Microsoft Excel.

| Details of study coding categories
All eligible studies were coded on a variety of criteria, including details related to the source of the study (publication source, title, authors, etc.), study characteristics (methodological type, dates of data collection, etc.), sample characteristics (size, location, etc.) study methods and procedures (selection process, characteristics of treatment and control areas, associated publicity campaigns, etc.), descriptions of the independent and dependent variables (construct, operationalization, etc.), effect size data (if any), adjustment for bias (regression to the mean [RTM], spillover/diffusion, etc.), and study conclusions.
Every eligible study was coded by two review authors, using a standardized data extraction instrument (see Appendix E). All disagreements were identified and resolved by discussion with a third review author.

| Descriptive analyses
The review includes all studies meeting the inclusion criteria. Descriptive statistics extracted from each study included: •

| Statistical analyses
To facilitate comparisons of studies, a standardized summary measure was defined for each outcome. Summary measures were based on relative effects, rather than differences in effects, where the outcome after intervention was divided by the outcome before intervention, as an expression of the proportional change in outcome.
Summary measures were calculated for all studies where possible (i.e., where required information was reported or adequate data were available for the calculation).
Rate ratios were estimated by dividing the count of the outcome post-and preintervention in the intervention area by the corresponding ratio in the control area. For example, the estimated rate ratio for road traffic collisions was: / / collisions after collisions before in intervention area collisions after collisions before in control area .
Assuming that traffic volume remains the same on the roads postintervention in the control and intervention areas, this rate ratio estimates the change in the collision rate in intervention areas compared to that in control areas. For outcomes expressed as counts or rates the intervention effect was estimated using rate ratios with a 95% confidence interval (CI), calculated assuming a Poisson

| Data synthesis
Results were pooled in a meta-analysis where three or more studies reported the same outcome (Cooper, 2003). The logarithm of the rate ratio and its SE (calculated assuming a Poisson distribution for the number of collisions in each area and time period) were pooled. If there were too few studies for a meta-analysis, the results of individual studies were described in a narrative.
Heterogeneity (variability) among the effect estimates was assessed using a χ 2 test at a 5% significance level and quantified using the I 2 statistic, the percentage of between-study variability that is due to true differences between studies (heterogeneity) rather than due than to sampling error. An I 2 value greater than 50% was considered to reflect substantial heterogeneity. Substantial heterogeneity would mean that the results of different studies vary substantially more than would be ex-

Spillover (diffusion of benefits)
Spillover occurs when the treatment has an effect outside the targeted area or population. In the case of traffic enforcement, spillover occurs when a safety measure such as an RLC that is placed at one intersection affects driver behavior at other intersections that do not have RLCs. This may occur because RLC programs frequently involve not only the placement of cameras but also the posting of warning signs and widespread publicity campaigns. As a result, driver behavior may be affected throughout the area, rather than just at those intersections that have cameras. To reduce spillover effects, control and comparison sites should be located outside the area affected by RLC program publicity.

RTM
RTM refers to a statistical phenomenon that appears when making repeated measurements of the same variable. In general, observations that produce extremely high or low values tend to be followed by values that are closer to the mean (see e.g., Barnett, van der Pols, & Dobson, 2005

Risk of bias assessment
The expanded risk of bias analysis was based on six dimensions that focused on the design of the study, the analysis of the data, and the contents of the study report. These six dimensions, which conform to whether it was True, False, or Unclear and these were used to assess each study on whether it presented a high, low, or unclear risk of bias across the six domains.
Risk of bias assessment was performed independently by three review authors (E. G. C., S. K., and C. P.). For the studies identified in the previous review, the same three review authors independently assessed the risk of bias of the included studies. Any discrepancies were resolved by deferment to further review authors (R. S. and P. E.). All disagreements were resolved by consensus.

| RESULTS
This section presents the results of the systematic review and meta-analysis of studies examining RLCs. It is organized around the EMMIE framework which includes measures of effect as well as discussions of the mechanisms through which RLCs are believed to work, the moderating factors that may influence the activation of these mechanisms, various elements that may affect the successful implementation of RLC programs, and the economic costs and benefits associated with the use of RLCs.

| Results of the search
The search strategy produced a total of 5,708 records after duplicates were removed. Title and abstract screening resulted in the exclusion of 5,587 records, leaving a total of 121 references that were potentially eligible for inclusion in this study. Full-text review of these works identified 28 primary studies that met the inclusion criteria, in addition to the 10 studies that had been identified in the prior Cochrane review. Eight of these newly identified studies (and one study from the previous review) had associated publications, which were subsumed in the primary studies (Cunningham & Hummer, 2010;Fitzsimmons, 2007aFitzsimmons, , 2007bFitzsimmons, , 2007cGarber, Miller, Abel, Eslambolchi, & Korukonda, 2007;Hallmark, Orellana, McDonald, Fitzsimmons, & Matulac, 2010;Miller, Khandelwal, & Garber, 2006;Sayed & de Leur, 2007;Shin & Washington, 2007;Retting & Kyrychenko, 2002 from the previous review). Two of the newly identified studies reported on more than one independent study area: Fitzsimmons (2007a,b,c) and  The search process is diagrammed in Figure 1, which also shows the number of records excluded, with a summary of reasons. Further details of some of the excluded studies are also available in Appendix F.

| Description of included studies
A total of 38 studies were included in this analysis. Of these, 37 were CBA studies with a distinct control group (27 newly identified studies plus 10 previously identified in the prior Cochrane review). One of the newly identified studies employed a controlled ITS design.
No RCTs were found. See Appendix F for information on the studies included in and excluded from this review. The characteristics of the studies in the previously published Cochrane systematic review are not duplicated in this report (for copyright reasons).
A summary of outcome measures covered by the current and previous reviews is provided in Table 1.  & Obeng, 2004;Ko, Geedipally, & Walden, 2013;Miller et al., 2006;Sayed & de Leur, 2007) and 28 received a high risk of bias across more than one domain (18 newly identified studies plus 10 previously identified). Table 2 provides a list of the included studies with their risk of bias assessments. Studies were stratified across each of these domains for additional sensitivity analyses.

Total crashes
Twenty-four studies reported 27 estimates of effect on the total number of all types of crashes (including PDO). Twenty of these studies (accounting for 23 estimates of effect) reported these estimates with CIs or SEs.
The effects of RLCs on the total number of crashes at intersections were highly heterogeneous (I 2 = 90.4%, Q = 228.4, and df = 22, p < .001). The direction of the estimated effects was also inconsistent.  Sayed & de Leur, 2007). Pulugurtha and Otturu (2014) reported an EB analysis that found the RLC program was not effective in reducing crashes at the majority of selected signalized intersections. This study found some evidence that the RLCs were more effective at reducing total crashes at intersections with less than 40,000 vehicles entering per day, but reported more uncertain and marginal effects at intersections with greater than 40,000 vehicles per day. Using EB analysis to estimate the effect on total crashes, Richardson (2003) reported a reduction of 0.65 crashes per site per year as a result of RLC installation which was equated to an 11.7% reduction in total crashes per year. A before and after study conducted on a single RLC camera site (Ross & Sperley, 2011) reported an increase in total crashes after camera installation, around 77% increase in average number of total crashes per month. Sayed and de Leur (2007) conducted an EB analysis, reporting reductions in total crashes at 20 of the 25 locations after RLC implementation and a reduction in total crashes of 11.1%. These results are aligned with the meta-analysis which found the effects on total crashes to be heterogeneous, and inconsistent in the estimated direction of effect. Some of these additional studies highlighted variation within crash and intersection types impacting the effect of RLCs.

Total crashes stratified by country
As shown in Figure 3, the subgroup analysis by country indicates uncertain effects of RLCs on total crashes in both Australia (2% decrease; 95% CI; 6% decrease 7% increase) and the United States (1% decrease; 95% CI; 11% decrease-9% increase). In both cases, the CIs overlap the reference line and a further test of group differences indicates no evidence of difference in the subgroup effects (Q = 0.10, p = .76). In Australia, the effect estimates across studies are consistent (I 2 = 0%, p < .001) while in the United States, there was much greater heterogeneity (I 2 = 86.3%, p < .001).

Total injury crashes
Seventeen studies examined injury crashes, providing 18 estimates of effect. Fifteen of these studies (16 estimates) reported CIs. Figure 4 shows that the estimates of effect on total injury crashes were also highly heterogeneous (I 2 = 93.1%, p < .001). The overall pooled estimate of effect suggests that RLCs reduced total injury crashes by 20% with a 95% CI (32-5% decrease). In Llau et al. (2015), the number of injury crashes specifically included possible injuries; this distinction was not made for other studies.
Two newly identified studies (Kull, 2014;Sayed & de Leur, 2007) reported results from which SEs could not be obtained, but reported reductions similar to studies in the meta-analysis. Kull (2014) reported a reduction in the rate of all injury crashes of 0.48 per intersection per year, while Sayed and de Leur (2007) reported reductions in injury crashes of 6.1% after RLC implementation.
Several other studies used different criteria to measure injuries and therefore are not included in the pooled analysis. Burkey and Obeng (2004)

Total injury crashes stratified by country
As seen in Figure 5, a meta-analysis of studies from the United States reported a decrease in injury crashes of 19%, although the effect was uncertain (95% CI; 40% decrease-9% increase) and there was evidence of significant heterogeneity (I 2 = 94.3%, p < 0.001). A meta-analysis of studies from Australia reported a decrease in injury crashes of 15% (95% CI; 23-6% decrease) with no evidence of heterogeneity (I 2 = 4.8%,

Total PDO crashes
Seven studies examined PDO crashes, six of which included CIs or SEs. These are shown in Figure 6. Overall, there was a nonsignificant increase of 5% in PDO crashes (95% CI; 8% decrease-20% increase).
It was not possible to obtain SEs from one study (Sayed & de Leur, 2007), and this reported reductions in PDO crashes of 14.3% after RLC implementation.

Total crashes from RLR
Six studies looked specifically at RLR crashes (producing seven estimates of effect). RLR crashes were identified as those caused directly by a driver running a red light or failing to yield during a turn on red, or any crash where a red light violation ticket was issued. The pooled estimates showed a 47% overall reduction in the total number of RLR crashes, although this was nonsignificant (95% CI; 62% decrease-2% increase; I 2 = 98.4%; p < .001). ' Council Bluffs study (2007a, 2007b, 2007c  This study was excluded from final statistical analyses for turning, same roadway crashes where SE was incalculable. It was included for total injury crashes, right angle injury crashes, and rear end injury crashes where SE could be extracted.

Figure 7 identifies Fitzsimmons
F I G U R E 2 Effects of red light cameras on total crashes. CI, confidence interval; ES, effect size F I G U R E 3 Effects of red light cameras on total crashes-stratified by country. CI, confidence interval; ES, effect size F I G U R E 4 Effects of red light cameras on total injury crashes. CI, confidence interval; ES, effect size the extremely large reductions in RLR crashes found in this study. As a sensitivity analysis, the meta-analysis was repeated excluding this study; the direction of the overall estimate of effect was unchanged but the estimated reduction in crashes dropped by over half to 18%, which was now significant (95% CI; 25-11% decrease). Removing this study also reduced heterogeneity (I 2 = 14.9%; p = .320). When City of Lubbock (2008) was also removed from the pooled analysis, RLR crashes were significantly reduced by 19% (95% CI; 25-11% decrease) and heterogeneity remained low (I 2 = 17.5%, p = .304).
Two further studies reported estimates of effect of RLCs on road traffic crashes from RLR. However, these limited "total" crashes to only specific types, rather than all crashes due to RLR. Therefore, they were not considered comparable and were excluded from the pooled analysis. Andreassen (1995) reported a 7% increase in total RLR crashes, which were defined as the sum of pedestrian, right angle, turning (same roadway) and rear end crashes. Cunningham and Hummer (2010) similarly reported a 5% increase in RLR crashes, which were defined as the sum of turning (both same roadway and different roadways), right angle and rear end crashes.

Right angle injury crashes
Seven studies reported eight estimates of effect; one of these studies (Kull, 2014) reported effect estimates without CIs.
The pooled analysis of studies with reported CIs, which is shown in Figure 10, estimates a significant overall reduction of 29% in right angle injury crashes (95% CI: 42-14% decrease), with moderate evidence of heterogeneity (I 2 = 59.1%, p = .023).

Right angle injury crashes stratified by country
There were not enough studies to estimate the pooled effect of RLCs on right angle crashes in the USA. In Australia, a pooled estimate indicated a significant 43% decrease (95% CI: 53-32% decrease) with no evidence of heterogeneity (I 2 = 0%, p = .706), as shown in Figure 11.

Total turning, same roadway crashes
Five studies reported turning, same roadway crashes as an outcome with seven estimates of effect. Two of these studies reported effect estimates without CIs (Andreassen, 1995;Richardson, 2003).
Three studies were found where SE/CIs could not be obtained.
Andreassen (1995)  0-22% increase), with no evidence of heterogeneity (I 2 = 0%, p = .622). This is shown in Figure 14. The test of group differences indicates no evidence against the null hypothesis that the subgroup effects are the same (Q = 1.52, p = .22).

Total rear end injury crashes
Eight studies (nine estimates of effect) examined total rear end crashes that resulted in injury; of these seven studies reported eight estimates of effect with CI/SE. The pooled analyses of these eight estimates, which is shown in Figure 15, suggested a nonsignificant decrease in rear end injury crashes of 1% (95% CI: 20% decrease-24% increase) and moderate heterogeneity (I 2 = 55.5%, p = .028).
F I G U R E 1 0 Effects of red light cameras on right angle injury crashes. CI, confidence interval; ES, effect size COHN ET AL.

| 19 of 52
Using annual rear end injury crash rates, Kull (2014) found an increase of 17% following RLC installation.

Rear end injury crashes stratified by country
There were not enough studies from the United States to estimate a pooled effect on rear end injury crashes. In Australia, a meta-analysis found a nonsignificant 1% reduction in rear end injury crashes (95% CI: 27% decrease-36% increase) with no evidence of heterogeneity (I 2 = 39%, p = .178). This is shown in Figure 16.

Total rear end crashes from RLR
Three studies estimated the effect of RLCs on total rear end crashes that were specifically identified as resulting from a red light violation (four estimates of effect). As seen in Figure 17, a pooled estimate indicated a nonsignificant 6% increase in rear end crashes from RLR (95% CI: 34% decrease-69% increase) with evidence of heterogeneity (I 2 = 82.5%, p < .001).

| Summary of effects
This systematic review shows that RLCs resulted in a significant reduction of of 24% in right angle crashes (95% CI: 35%-10% decrease), of 29% in right angle injury crashes (95% CI: 42%-14% decrease), and of 20% in total injury crashes (95% CI: 32%-5% decrease; see Table 3). A large reduction in red light violations was suggested; however, only two of three studies reported CIs preventing appropriate meta-analysis. RLCs are also associated with a significant increase in rear end crashes of 19% (95% CI: 9%-31% increase).
There was no evidence to suggest that study heterogeneity was consistently explained by country, although results were more consistent across outcomes for studies conducted in Australia. Additional subgroup analyses found some evidence that when stratified by risk of bias according to control for confounding, studies with a low risk of bias reported a nonsignificant reduction in right angle crashes.

| Mechanisms
Mechanisms focus on how and why an intervention works.

| Moderators
Moderators are "variables that may explain outcomes across different studies" (Johnson et al., 2015, p.462). Possible moderators could include factors such as road type, intersection geometry, traffic flow, number of lanes, speed limits, weather conditions, and country or area level effects.
In the meta-analysis, studies were stratified by country (United States and Australia) to examine country-level effects, with further subgroup analyses conducted on risk of bias, RTM, and the use of warning signs.
However, there was insufficient data to allow further analysis of other potential moderators in this manner. Some of the individual studies did note, discuss, and in some cases control for a number of possible moderating factors. However, there was only limited evidence regarding their effect on RLC program outcomes.

| Time of day and day of week
Time of day and day of week have been suggested as possible moderators for RLC effects. Retting, Williams, and Greene (1998); Porter and Berry (1999); and Kamyab, McDonald, and Stribiak (2000) all found that the incidence of red light violations is greater on weekdays than weekends. They also found that red light violations varied by time of day: Retting et al. (1998) and Porter and Berry (1999) indicated that the average number of red light runners was greater during the day than at night while Kamyab et al. (2000) reported a higher incidence of red light violations specifically between 3:00 pm and 5:00 pm. Fitzsimmons (2007a)  Conversely, Arup Transportation Planning (1992) found that prior to the installation of RLCs, the number of violations per hour was lower during peak periods (defined as 7:30 am to 9:30 am and 4:00 pm to 6:00 pm), regardless of traffic flows. However, they also stated that during these peak periods, signal coordination was designed to favor major traffic flows, reducing delays and decreasing the potential for violations.
It should be noted that some of the studies in the meta-analysis did not collect data around the clock. Chin (1989), for example, only collected data on weekdays and only during specific periods of the day that fell between 8:00 am and 7:00 pm.

| Signal timing
Another possible moderator is signal timing. Retting and Greene (1995) found that longer change intervals, which may be created by increasing the length of the yellow signal interval, may reduce red light violations and traffic crashes. Bonneson and Son (2003) reported that RLR rates are higher when the duration of the yellow signal interval is shorter than suggested by traffic engineering equations and Bonneson and Zimmerman (2004) found that after increasing the yellow interval duration, there was a decrease in the frequency of red light-related crashes. Retting, Ferguson, and Farmer (2008) found that longer yellow signal lengths were associated with a 36% decrease in red light violations.

| Speed limit variations
Variations in the speed limit on the roads used in RLC studies may also serve as a moderator variable. Some studies in the meta-analysis did control for this, such as Miller et al. (2006) and Fitzsimmons (2007a, 2007b, 2007c and A UK study conducted on RLR in Birmingham (Lawson, 1991) reported some "average" characteristics associated with increased visibility through a stretch of road on the intersection exit, or the presence of multiple lanes. Lawson also noted a great variation in crash patterns between intersections and within approaches of intersections, and suggested crash history at individual sites, rather than general intersection characteristics, is the best indicator of where to place RLCs.

| Implementation
Implementation involves the context in which RLC programs are put into practice. Variations in implementation may influence program effectiveness, affecting the level of deterrence. The structure and organization of RLC programs vary greatly by jurisdiction and a number of these variations may affect how driver behavior is influenced by RLCs, although most were not tested directly by the included studies.

| Public awareness and program publicity
Public knowledge of the implementation of a photo enforcement program is considered essential for program success. The use of publicity campaigns to enhance public awareness, as well as posted warning signs (discussed below), are believed to increase the general deterrent effect of the cameras and to create spillover from intersections with cameras to the wider area (see e.g., Ross & Sperley, 2011;Shin & Washington, 2007). Retting and Kyrychenko (2002) stated that the primary mechanism for preventing red light violations is driver awareness and stressed the importance of publicizing RLC enforcement programs.
In the United States, the Federal Highway Administration Many of the included studies specifically mentioned the use of public awareness programs and media publicity regarding the operation of RLCs. Some jurisdictions mailed written notices to local residents to alert them to the new photo enforcement program. In many cases, the programs began publicity campaigns months before the cameras were installed and some even provided specific information as to which intersections were selected for RLC monitoring. To further increase awareness, and possibly enhance specific deterrence, many jurisdictions implemented a warning period prior to beginning formal RLC enforcement. During this period, violators were issued warnings rather than being ticketed. Warning periods, when used, generally lasted 30 days or less (see, e.g., Fitzsimmons, 2007aFitzsimmons, , 2007bFitzsimmons, , 2007cHallmark et al., 2010;Llau et al., 2015;Porter, Johnson, & Bland, 2013;Retting & Kyrychenko, 2002;Retting et al., 1999a;Ross & Sperley, 2011).
While none of the included studies specifically examined the impact of publicity programs on camera effectiveness, many emphasized the importance of these programs. As Ross and Sperley (2011, p.9) pointed out, "To maximize the impact of red light camera enforcement, drivers must be aware of the enforcement…"

| Warning signs
A number of the included studies reported that the jurisdictions under study posted signs to inform drivers that RLCs were in use (see Figure 19 for a complete list). Along with publicity campaigns, warning signs are designed to increase driver awareness of the automated enforcement programs and to enhance their deterrent effect.
Warning signs may be posted either at or near the specific intersections at which cameras are installed or at major entrance points to the city (or both). In some cases, the use and placement of signage may be affected by legislative requirements; in the United States, this varies by state. In North Carolina, for example, state statutes require warning signs to be posted on all approaches of any intersection at which an RLC is installed, regardless of which approaches are actually monitored by RLCs (Cunningham & Hummer, 2010). Conversely, in Oregon, cities operating RLCs are required to post signs on all major routes into the city; Ross and Sperley (2011) reported that Salem, Oregon also posted warning signs on each approach where a camera was operating to help increase driver awareness of photo enforcement. In California, warning signs are required but local governments are given the option of placing signs either at each monitored intersection or at all major city or county entrances (see e.g., California State Auditor, 2002;Retting & Kyrychenko, 2002;Retting et al., 1999a). However, using the latter option does mean that there is a risk that some major entrances may be overlooked. California State Auditor (2002) reported that in 1999 approximately 700 citations for red light violations were dismissed in Sacramento after a traffic court ruling found the city had failed to fully comply with the law when installing warning signs. Following this ruling, the city placed warning signs at each RLC-monitored intersection.
The location of warning signs may affect their utility. California State Auditor (2002) , 2005) reported that the use of warning signs at intersections was associated with a smaller benefit than warning signs at both intersections and city limits.
A meta-analysis was conducted to explore the effect of warning signs on total crashes ( Figure 18) and total injury crashes ( Figure 19) in the primary studies. There was no significant difference in the effect of RLCs on total road traffic crashes or total injury crashes between the studies that mentioned the use of warning signs and those that did not. However, for both crash types, studies reporting the use of warning signs do show a slightly greater trend toward a reduction in crashes than studies that did not mention the use of warning signs (albeit insignificant).

| Obstacles to implementation
Very few of the included studies provided any detailed discussion of the types of implementation issues and obstacles that RLC programs face. However, several key obstacles that were mentioned included problems with contract vendors, issues involved in operating a legally-compliant RLC program, and public attitudes and concerns.

Vendor-related concerns
A vendor is the agency that supplies an RLC system for a community, and also operates the system in jurisdictions where the police do not do so. Sharpnack (2009) Auditor, 2002;Fitzsimmons, 2007aFitzsimmons, , 2007bFitzsimmons, , 2007cHallmark et al., 2010). is that the data may become accessible not only to traffic enforcement agencies but also to other governmental agencies and officers and to non-governmental entities and may even become public record. This could result in a serious violation of personal privacy.

Public concerns about privacy
Another concern is possible misuse of the photographs taken by the RLCs. While California law prohibits the use of RLC photographs for any purpose other than the prosecution of the motorist for the traffic violation, California State Auditor (2002) reported that nearly all of the local California governments studied had or would use these photos for other purposes as well, including the investigation of unrelated criminal activities and as evidence in court.

Legal challenges
A number of studies reported issues relating to the legalities involved in operating an RLC program, particularly relating to issuing citations. | 27 of 52 violation regardless of who was actually driving the vehicle at the time, stating, "Guilt is presumed over innocence." Additionally, because of the delay between the actual violation and the receipt of the ticket, which is sent by mail, the owner of the vehicle may have difficulty remembering the event, or even who was actually driving the vehicle on that day. This may make it extremely difficult, if not impossible, to make any reasonable defense to the citation.
In Chicago, a class-action lawsuit recently was brought against the city alleging that the city violated due process in the method used to notify drivers of violations and of late fees for failure to pay fines on time.
This is part of a larger scandal in the city alleging corruption and mismanagement of the program, including bribery, inconsistent enforcement, and the use of unfair criteria to issue tickets (Kidwell, 2016).

Public concerns about safety versus revenue
Another issue that has been raised is that of the purpose of RLC programs and the belief that generating revenue for the local government may take priority over improved public safety (see, e.g., Hobeika & Yaungyai, 2006). Simons (2016) states that while they may be intended as safety programs, "increasingly, the public sees them as money-making scams that can actually make roads less safe." Essentially, many view RLCs as "cash cows" and money machines. Cunningham and Hummer, (2010; as discussed in Cunningham & Hummer, 2004) argues that this is a legitimate concern because if conflicts arise between revenue and safety, a system that is revenue-focused will not be able to maximize safety.
Because the vendors are for-profit businesses, they have an incentive to increase ticket revenue, regardless of public safety implications. Similarly, local and state governments may rely on the revenue generated by these programs. In 2016, a California state senator proposed a bill to reduce the base fine for rolling turn violations in the state.
The bill resulted from growing concerns that after installing RLCs in several California cities, these violations, which cause only 0.5% of all traffic crashes, were accounting for nearly all red light citations. These tickets generated huge sums for the cities; the amount of money cleared by Rancho Cordova from red light infractions increased by over 100% between 2012 and 2016 (from approximately $73,000 to $742,000) because of the installation of four RLCs. The high fine and associated fees and other penalties, which were primarily intended to apply to much more dangerous violations, have caused severe hardship to lower-income drivers and the bill was not only approved by the state Senate but was supported by the state ACLU, the National Motorists Association, and the California Association of Highway Patrolmen. However, the bill died in the Assembly appropriations committee "after it was determined that it would cause a significant loss of revenue at the state and local levels" (Flynn, 2016). Programs that appear to be revenue-driven rather than a mechanism to improve road safety may generate public opposition, making implementation and continued operation more difficult.

| Other implementation factors
Both the US Department of Transportation and the UK Department for Transport provide guidelines that address the implementation of traffic enforcement camera programs. The Department for Transport (2007) published a report on best practices for implementing RLCs (as well as speed cameras and combined red light and speed cameras, known as traffic safety cameras) in the United Kingdom; the report includes guidance on deployment, visibility, and signage. It is based in part on an influential study (Hooke et al., 1996)  • Signage: Appropriate signage must be used when deploying RLCs.
As noted above, in the United States, the placement of warning signs may be mandated by state legislation, but both the US and UK guidelines recommend warning signs be posted even if they are not required by law or ordinance. Signs should be visible at all times and should not be obscured in any way; the FHA (2005) recommends regularly monitoring signs to ensure they are cleared of any vegetation growth that may interfere with visibility.
• Publicity: Both the US and UK guidelines emphasize the importance of publicizing the RLC program. The campaign should not only include information about the program itself but also educate the public about the dangers of RLR and emphasize the RLC program's goal of enhancing road safety. Publicity needs to be ongoing to counteract habituation and to maintain drivers' awareness of the risk of detection.
• System Assessment/Monitoring: Ongoing monitoring of RLC enforcement program efforts is essential to measure the effectiveness of the program. Data collection should include crash and RLR data; the Department of Transport (2007) recommends that other information, such as public opinion regarding RLCs, also be collected.
The FHA (2005) recommends collecting data at both camera locations and control sites without photo enforcement to measure the effects of the RLC program absent other external factors. Many of the studies did not include any economic information;

| Economics
those that did varied widely in the amount of detail provided. Only one report (South et al., 1988) considered monetized cost and crashes, but the brief analysis that was conducted excluded revenue from fines. None of the studies conducted a full cost-benefit analysis that included both fiscal viability and societal benefits (including crash costs), and there was insufficient comparable information available to permit anything other than a qualitative synthesis of economic information from the included studies.

| Implementation and operational costs
Only a small number of studies provided information on implementation costs, which can include factors such as the capital cost of the RLC equipment, the costs involved in setting up an RLC program, and operational or running costs.
According to Fitzsimmons (2007aFitzsimmons ( , 2007bFitzsimmons ( , 2007c and Hallmark et al., (2010), a typical RLC system can cost US $50,000 or more, depending on both the type of intersection and the number of cameras to be installed. Washington and Shin (2005) also estimated the cost of a 35-mm wet film camera system to be around US $50,000 to $60,000, including the camera and housing, pole, and loop detectors, as well as installation, and pointed out that digital systems are significantly more expensive, costing up to US $100,000.
Operational costs include the ongoing costs involved in running the RLC program. One of these is the fee charged by the RLC vendor. This can be either a flat monthly fee or a cost per citation issued. Washington and Shin (2005) stated that monthly fees tend to be approximately US $5,000 per camera system and Fitzsimmons (2007aFitzsimmons ( , 2007bFitzsimmons ( , 2007c and Hallmark et al., (2010) reported similar monthly fees per intersection. Fitzsimmons (2007aFitzsimmons ( , 2007bFitzsimmons ( , 2007c and Hallmark et al., (2010) also noted that in Iowa, multiple possible payment structures were implemented.
While a flat monthly operating fee was one option, cities could elect to pay the vendor a fee per citation. Depending on the vendor, the fee was either fixed or decreased as the number of citations per day increased.
Shin and Washington, (2007; as discussed in Washington & Shin, 2005) suggested that RLC programs could be operated as revenueneutral programs, so that the operating costs are equal to the fines generated from the program. However, they also pointed out the difficulty in locating published estimates of installation and operating costs.

| Fiscal viability
The studies that included economic information focused either on the program's fiscal viability, which involved comparing implementation and operation costs to revenue generated, or the safety and social benefits of the program, which focused on cost changes resulting from changes in the number of traffic crashes. None of the studies conducted a complete analysis that included both operational costs and societal benefits (including crash costs).
Of the studies that examined fiscal viability, three reported negative economic outcomes (Andreassen, 1995;City of Lubbock, 2009;Sharpnack, 2009 which did not even cover program operating costs. As Andreassen (1995) did not report any significant reduction in crashes from the RLC program, it does not appear that the program was fiscally viable. Sharpnack (2009) found that between 2003  It is interesting to note that RLC programs are often expected to be fiscally viable. This is a unique requirement, as other types of crime prevention programs generally are not held to this standard.
Obviously, developing a revenue-neutral program is desirable given the budgetary constraints faced by police departments today. However, it is not clear why a program needs to break even or actually produce any positive economic outcomes to be considered worthwhile if it produces a significant increase in road safety. The measure of success for a crime prevention program is usually its ability to prevent crime, not to pay for itself.

| Crash safety benefits
Five studies (Garber, 2007; South et al., 1988) examined the economic effect of changes in the number of traffic crashes that resulted from the use of RLCs, separate from the installation and operational costs of the programs. Ross and Sperley (2011) examined the overall cost of crashes at one intersection in Salem, Oregon before and after an RLC was installed and found that the average monthly cost of crashes increased by about 70% during the post-installation period, rising from US $16,296 to US $27,738. They noted that during the period prior to camera installation there was a higher percentage of injury crashes and that most of the crashes occurring after the cameras were installed were rear end crashes, which tend to produce less severe injuries. However, despite this, the overall increase in the average number of crashes per month after the cameras were installed resulted in an overall higher monthly cost.

| Other economic factors
Early cost-benefit analysis studies have identified significant benefits for RLCs. In the United Kingdom, Hooke et al. (1996) found that the average fixed cost per site of RLCs was just over £9,200 and average annual costs were over £5,600 per site. The study compared the costs of installing and operating RLCS (including capital costs, maintenance, and prosecution costs) to benefits resulting from crash reduction and fixed penalty revenues and found that RLCs generate significant net benefits to the police and the community as a whole.
Seven of the 10 areas examined achieved a net positive return within 1 year of camera installation and all but one had done so after 5 years. Overall, the researchers reported that the return on investment was twelve times the initial investment after five years.

| Deterrence and spillover
As noted in the Results section above, a key mechanism by which RLCs may reduce RLR is deterrence, which focuses on the threat of punishment to discourage offending.

| Deterrence theory
According to Gibbs (1975), deterrence is "the omission or curtailment of a crime from the fear of legal punishment" (p. 39). Deterrence theory posits that fear of punishment encourages potential wrongdoers to comply with the law. Empirical research examining the ef-

fectiveness of traffic lights and traffic regulations on traffic violations
suggests that deterrence may result when violation is associated with penalty and the potential for penalty (Cramton, 1969). This suggests that other mechanisms, such as RLCs, may also affect driver behavior and deter drivers from running red lights. Porter and Berry (1999) conducted a survey of drivers and found that respondents did not believe that the police would catch most red light runners, so they felt there were few consequences for this behavior. However, surveys of communities in which RLC enforcement has been implemented show different results. For example, a survey conducted by Retting and Williams (2000) found that in cities with RLCs, 61% of respondents considered it likely that someone would get a ticket for running a red light in their city, compared with only 46% of respondents in cities without camera enforcement. Additionally, respondents in cities with cameras were more than twice as likely to know someone who had received a traffic ticket for running a red light as those in cities without cameras. RLCs not only increase the risk of punishment, and thus the negative consequences of RLR, but also increase the perception of risk among the general population. This suggests that the mechanism by which RLCs may reduce RLR is deterrence.
It is important to note that deterrence theory assumes that the target behavior is committed intentionally: drivers may deliberately run red lights or may intentionally try to "beat the light" by speeding up when the light turns yellow. In the latter case, even though drivers may not have actually intended to run the red light, they were aware that their attempt to beat the yellow light created the potential to do so. Deterrence assumes that drivers' behavior can be influenced by RLCs; however, some drivers may be unable to stop or clear the junction safely at the end of the green signal phase (during the yellow signal phase), this is referred to as the dilemma zone. This has implications for safety, as driver behavior may be more unpredictable; one driver may choose to stop and become at risk for a rear end collision, while a driver who continues may be at risk of crossing on the red light and prosecution and at greater risk of a collision at the intersection and potential prosecution (Maxwell & Wood, 2006).
Enforcement mechanisms that attempt to deter behaviors such as RLR are unlikely to affect situations in which the behavior is committed inadvertently or unintentionally, such as when a driver was unable to see the signal, was inattentive and did not realize that the traffic light was red, or was unable to stop. Therefore, it is likely that RLCs will be more effective in reducing RLR among drivers who deliberately engage in this behavior.

| Types of deterrence
There are two main types of deterrence. Specific deterrence focuses primarily on punishing apprehended offenders, with the assumption that they will be deterred from reoffending out of a desire to avoid future punishment. General deterrence focuses on the population at large and assumes that the threat of punishment will deter people from initial law violations. The greater the perception of risk of punishment, the greater the likelihood that general deterrence will be effective. To be effective, traffic enforcement policies need to do both, so that a sanction not only impacts the individual who is being punished but also others who do not directly experience the sanction (Bates, Soole, & Watson, 2012).
As a specific deterrent, RLCs focus on reducing repeat offending by punishing those individuals who offend. Because RLCs are fully automated and operate remotely, without requiring human intervention, drivers who run red lights are more likely to be detected and punished. One purpose of this punishment is to deter these drivers from running red lights again in the future.
General deterrence is achieved through increasing the risk of apprehension and punishment. This may include both the actual/ objective risk of being detected and the perceived/subjective risk, which reflects drivers' belief about the likelihood that they will be detected in a violation. Overall, perceived risk appears to be most likely to influence driving behavior (Zaal, 1994). RLCs are designed to create what Belin, Tillgren, Vedung, Cameron, and Tingvall (2010) calls a "feeling of continued surveillance" that suggests a greater risk of apprehension and punishment for running red lights.
In many cases, the primary aim of enforcement strategies such as RLC programs is general rather than specific deterrencethe focus is on preventing red light violations rather than on offender detection (see, e.g., MacLean, 1985;Zaal, 1994). South et al. (1988) suggests that increased enforcement, obtained through the use of RLCs, may deter potential offenders and Retting and Kyrychenko (2002) also found that the effects of RLCs were not limited to only the specific intersections with cameras, providing further support for a general deterrent effect. Retting et al. (1999a, p.173) suggests that "The presence of cameras may promote a general readiness to stop at red lights."

| Spillover: The halo effect
According to Hu et al. (2011, p.277), "A high likelihood of apprehension helps convince motorists to comply with traffic laws." RLCs are designed to increase the actual risk of apprehension as well as increasing the perception of increased risk, to create a general deterrent effect. However, their impact may not be limited to only those intersections with cameras. RLCs may also create spillover, so that they have a more general effect on driver behavior at signalized intersections, reducing red light violations not only at those locations where cameras are placed but also at surrounding non-camera intersections. Kyrychenko (2002, p. 1825) suggests that the potential for positive spillover is a key element of RLC programs because "the goal of highly conspicuous traffic enforcement is to produce generalized changes in driver behavior with respect to traffic safety laws." Attitudinal surveys suggest that RLCs do create an overall perception of increased risk. As discussed above, Retting and Williams (2000) found that in cities with RLC enforcement programs, significantly more respondents believed it was likely that someone who runs a red light would get a ticket. Arup Transportation Planning (1992) surveyed drivers in Brisbane, Australia before and after the implementation of an RLC program. He found that prior to the introduction of RLCs, most drivers did not consider it likely that they would be caught disobeying a red traffic light (only 38% of drivers COHN ET AL. | 31 of 52 rated their chances of being caught as "likely" or "very likely"); this increased to 52% after the RLC program was implemented. He also stated that a survey in Adelaide, Australia after the implementation of RLCs found that 85% of people surveyed believed that RLCs "will make it more likely for drivers to get caught running a red light" and 72% believed RLCs "would change their driving habits" (p.27). South et al. (1988) has pointed out that for RLC enforcement to have an effect, motorists must be aware that cameras are in operation. He discussed a study of RLCs in Victoria, Australia in which RLCs were installed and monitored prior to media publicity of the program. During the period when the public was unaware of the program, approximately 300 violators per week were photographed by the RLCs; this decreased to about 20 violators per week after the media began publicizing the RLC program.
To increase the potential for spillover, and therefore increase the effectiveness of the enforcement program, Ross and Sperley (2011) suggests the use of jurisdictional boundary signs warning drivers about the use of photo enforcement mechanisms, rather than posting signs at the specific traffic intersections where cameras are installed.
Shin and Washington (2007)  The researchers found significantly greater spillover effects in Scottsdale; they suggested that the drivers in Phoenix knew which intersections and approaches had cameras and modified their behavior accordingly, while drivers in Scottsdale were not as certain of where the cameras were placed and therefore were more likely to be deterred from traffic violations citywide.
In general, empirical tests of the spillover effect of RLCs have yielded inconsistent results. Retting and Kyrychenko (2002) and Retting et al. (1999a) found red light violations in Fairfax and Oxnard decreased at both camera and non-camera intersections; they argued that in both cities the decline at non-camera sites was due to spillover (see e.g., Retting, Williams, Farmer, & Feldman, 1999b). Similarly, Kull (2014) found a decline in right angle crash rates in Chicago at intersections with and without RLCs (36% and 27%, respectively), and concluded that the decrease in crashes at the non-camera sites was due to spillover. AECOM Canada, Ltd. (2014) reported relatively large spillover effects as well, suggesting that the reduction in collisions at non-RLC intersections was due to changes in driver behavior that may have resulted from the widespread publicity of the RLC program throughout the jurisdiction as well as the public's lack of knowledge as to which specific intersections were equipped with cameras. Chin (1989) found that the installation of RLCs at intersections in Singapore not only reduced RLR on those approaches covered by the cameras but on other approaches as well.
Conversely, while Hobeika and Yaungyai (2006) found a spillover effect of RLCs on PDO crashes in Fairfax, they also concluded that there was no spillover effect for injury crashes. Garber (2007) found that there were no significant changes in various types of traffic crashes at non-camera sites in Virginia that were used to test for spillover.

| Comparisons with the previous Cochrane review
The original systematic review (Aeron- Thomas & Hess, 2005) included only 10 studies which were all published during 2002 or earlier and which came from only three countries (the United States, Australia, and Singapore). This updated review has expanded the search to a more comprehensive list of databases and websites and has increased the number of included studies to 38. Additionally, the research located by this updated search tends to be of higher quality.
The updated review includes four studies that have been classified as high quality, while the original review had none; conversely, 80% of the previously identified studies were of low quality, compared to 64% of the newly identified studies.
The increased number of studies included in the updated review, combined with the addition of a more extensive meta-analysis, has increased the precision of the effect estimates. The findings of this updated review support those of the original review in some areas, but not in all. As reported in the original review, the updated review found that RLCs were effective in reducing total injury crashes.
However, while the original review did not find any significant effect of RLCs on various types of crashes, the updated review did report differential effects of RLCs on specific crash types. RLCs were found to be associated with a reduction in right angle crashes and right angle injury crashes and with an increase in rear end crashes.
Additionally, the updated review has incorporated the EMMIE coding system. Although the original review discussed effect sizes, the other dimensions of the EMMIE scheme were not addressed. In contrast, the updated review includes discussions of the underlying mechanisms, potential moderators, and implementation factors that may influence RLC effectiveness, and economic costs and benefits of RLC programs. EMMIE coding additionally found that while spillover (or diffusion of benefits) is reported in a number of studies, the magnitude of this effect is not established and factors that trigger the underlying mechanism of general deterrence (e.g., warning signs and publicity campaigns) have not been featured widely when devising measures of effectiveness of RLC programs. Practitioners must undertake careful implementation of schemes to ensure legal compliance. Full cost-benefit analysis including capital, maintenance, operational costs, and revenue alongside societal costs and benefits (including costs or savings associated with increases or decreases in crashes) are lacking in the current literature.

| Comparisons with prior meta-analyses
Two previous meta-analyses were conducted on the effects of RLCs on traffic crashes (Erke, 2009;Høye, 2013; note that Høye was an update and replication of Erke). A comparison of the results of this study with the prior meta-analyses was inconsistent.
When including all studies, regardless of quality level, both this study and Erke (2009) Erke (2009) reported no effect of RLCs on this type of crash.
All three studies found that RLCs were associated with a decrease in right angle crashes (nonsignificant in Høye, 2013); similarly, right angle injury crashes were found to decrease in all three studies, although this result was nonsignificant in Erke (2009). In all three meta-analyses, rear end crashes increased with the use of RLCs (nonsignificant in Erke, 2009), as did rear end injury crashes (significant in Erke, 2009 only).
Høye (2013) suggested some evidence that warning signs improve total crash outcomes when used with RLCs. However, this study did not find any consistent pattern or significant differences between those studies that reported the use of warning signs and those that did not.
Both Erke (2009) and Høye (2013) found that those studies that controlled for RTM reported less favorable effects of RLCs; this study found some evidence of this when examining right angle and right angle injury crashes. Erke (2009, p.903) suggests that failing to control for RTM "can lead to an overestimation of the effects of and Singapore). Aside from the addition of two studies conducted in Canada (AECOM Canada, Ltd., 2014;Sayed & de Leur, 2007), the updated review was still limited to research conducted in these countries.
Additionally, the lack of methodological rigor, particularly the failure of many studies to account for both spillover and regression to the mean, and of some to control for additional confounders affects the quality of the evidence in those studies.

ACKNOWLEDGMENTS
We would like to thank the following individuals: Andrew Hutchings,

CONFLICT OF INTERESTS
The authors declare that there are no conflict of interests.

PRELIMINARY TIMEFRAME
Approximate date for submission of the systematic review: December 31, 2018.

PLANS FOR UPDATING THE REVIEW
This review will be updated on a 5-year basis. This will require identifying and coding any new studies and rerunning the analyses.
• Transport Research and Innovation Portal (TRIP)  Jul 1977-Jul 1979 and 1 year of after data (Jul 1979-Jul 1980. Interventions 25 RLCs at typical 4 way intersections (3-4 lanes on each approach and 50 km/hr speed limit) 46 non-camera control sites (with same traffic and environmental conditions as treated sites), some located close to treated sites so could not account for spillover. 100 reference group sites (similar characteristics to treated sites) used to develop collision prediction model Outcomes Total crashes, severe crashes (fatal and injury crashes), property damage-only crashes, right angle head on crashes, rear end crashes.

Notes
Analysis carried out on 25 sites, one site appeared to be an outlier and analysis was also conducted on 24 sites, extracted only data for the evaluation of all 25 sites.  Golub et al. (2002) No independent control group, used non-monitored/unenforced approaches of RLC intersection as comparison.
Guerin (2010) Speed cameras and RLCs activated at the same timecannot separate effect of RLCs from speed cameras.

Hebert-Martinez and
Porter (2006) Focus on identifying and predicting characteristics of red light runner drivers after implementation of RLC program. Outcome measures not of interest. Kent et al. (1995) Ordinal logistic modelnot a CBA, no before data Loftis et al. (2011) No non-RLC control intersection used. RLC intersections not selected at random. Study used simulation approach to model collisions with and without RLCs installed Lum and Wong (2003a) No independent control group-nonmonitored approaches of RLC intersection used as control. No examination of collision, only red light running. Study examines impact of RLCs on after-red timesdoes RLC effect mean after-red times on camera vs non-camera approaches Lum and Wong (2003b) Same data as Lum and Wong (2003a), some of the same results reported. No independent control group-nonmonitored approaches of RLC intersection used as control Lum and Wong (2003c) Same data as Lum and Wong (2003a), some of the same results reported. No independent control group-nonmonitored approaches of RLC intersection used as control McCartt and Hu (2014) No before data. Only data during warning period at which time cameras and signs had been installed and public announcement made McFadden and McGee (1999) Report is a synthesis of other RLC studies. No new data provided MVA Consultancy 1995 Some non-camera approaches of RLC intersections used as controls (only 19 hr after monitoring) Obeng and Burkey (2008) Focus is on offsetting driver behavior -RLC intervention not the primary intervention of interest Radalj (2001) No control before after crash data. Excluded from previous review "no base data provided on control" Shah (2010) Not a controlled before after study. Mentions a report of a controlled before after study that was acquired but only reports percentage decreases and none of the data is included Shah (2014) Not a controlled before after study. Mentions a report of a control before after study that was acquired but only reports percentage decreases and none of the data is included Smith et al. (2000) Report is a synthesis of other RLC studies. No new data provided Sun et al. (2012) No before after comparison. Data collected using video-based system set up by researchers. (1) Selection and matching of intervention and control areas The characteristics of the study and control sites were the same/similar There were no changes in the control sites during the study period The control sites were not adjacent to the intervention sites It is unlikely that the control group received the intervention (2) Blinding of data collection and analysis The outcome data were obtained from routine reporting systems (not originally collected for study) The analysis was conducted blind to intervention and control groups (3) Lengths of data collection time period pre-and postintervention Length of before period is at least 1 year Length of after period is at least 1 year (4) Reporting of results The main findings of the study are clearly described The authors report uncertainty due to random variability (confidence intervals) COHN ET AL.

| 47 of 52
Appropriate statistical tests were used to assess the main outcomes reported (p-Values)

(5) Control of confounders
The authors describe potential confounders The distributions of confounders in intervention and control sites were assessed and similar Do the authors discuss the effect of confounders on the results? (6) Any other potential sources of bias Did the study control for potential bias due to regression to the mean? Did the study report, or control for "spill-over" effects (e.g., use control sites located away from red light camera sites and associated publicity)? Were any other sources of bias addressed?
APPENDIX H: SUBGROUP ANALYSES F I G U R E H 1 Effects of red light cameras on right angle crashes resulting in injury-stratified by whether or not studies account for regression to the mean (Q = 7.78, p = .01). CI, confidence interval; ES, effect size F I G U R E H 2 Effects of red light cameras on right angle crashes resulting in injury-stratified by whether or not studies were peer reviewed prior to publication (Q = 11.50, p = .001). CI, confidence interval; ES, effect size F I G U R E H 3 Effects of red light cameras on rear end crashes-stratified by risk of bias according to control for confounders (Q = 17.11, p < .001) F I G U R E H 5 Effects of red light cameras on total crashes-stratified by risk of bias according to control for confounders (Q = 0.98, p = .01). CI, confidence interval; ES, effect size