Obsidian/Zotero/001_artiklid/koeffectsredlightrunning2017.md

---

category: literaturenote citekey: koeffectsredlightrunning2017 title: Effects of red light running camera systems installation and then deactivation on intersection safety authors: "Ko, Myunghoon; Geedipally, Srinivas Reddy; Walden, Troy Duane; Wunderlich, Robert Carl" year: 2017 date: 2017-00-00 2017 doi: 10.1016/j.jsr.2017.06.010 publication: Journal of Safety Research url: "https://doi.org/10.1016/j.jsr.2017.06.010" zotero_key: JPW6QJ2V zotero_storage: 327RD3CQ collections: transpordi_planeerimine folder: 001_artiklid firstAuthor: "Ko, Myunghoon"

status: converted

Journal of Safety Research

journal homepage:

Effects of red light running camera systems installation and then deactivation on intersection safety

Myunghoon Ko, ⁎ Srinivas Reddy Geedipally, Troy Duane Walden, Robert Carl Wunderlich

Texas A&M Transportation Institute, The Texas A&M University System, 3135 TAMU, College Station, TX 77843-3135, United States

article info abstract

Article history: Received 31 August 2016 Received in revised form 2 January 2017 Accepted 13 June 2017 Available online 22 June 2017

Keywords: Intersection safety Red-light running cameras Enforcement Cameras deactivation

Introduction: The primary objective of this paper is to evaluate the safety impacts of red-light running camera (RLC) system installation and then deactivation at 48 intersections in Houston, Texas. The second objective is to evaluate the spillover effect at nearby non-treated intersections in Houston after the deactivation. Methods: To accomplish study objectives, an Empirical Bayes (EB) before-after analysis was used.

Results: The results indicate statistically significant collision reductions on all red-light running (RLR) crash types (37 percent) as well as right-angle RLR crashes (47 percent) at the treated intersections after RLC activation. By way of comparison, the RLC deactivation analysis indicated that crashes increased by 20 percent for all RLR crash types and by 23 percent in right-angle RLR crashes at the formerly treated intersections. After deactivation, all severity RLR crashes increased more than expected at nearby non-treated intersections, which indicates the possibility of an adverse spillover effect. However, fatal/injury crashes associated with rear-end decreased after deactivation at both formerly treated and non-treated intersections, although those rear-end crashes account for smaller proportions when compared to all crash types/right-angle crashes.

Practical applications: Overall, removing RLC treatments results in a negative reaction to the safety benefits that the treatment provides when it is in place and actively working and to the nearby intersections where the treatment has not been implemented. This study helps define the effects that RLCs have on safety at signalized intersections after installation and deactivation.

© 2017 National Safety Council and Elsevier Ltd. All rights reserved.

1. Introduction

In 2012, there were approximately 30,000 crashes at intersections due to drivers disregarding stop signs or traffic signals in Texas (Texas Department of Transportation (TxDOT), 2012a). Disregarding a stop sign or red-light running (RLR) is one primary contributing factor for crashes at signalized intersections. With the purpose of reducing RLR violations, red light enforcement cameras (RLCs) have been generally used as an enforcement treatment to improve intersection safety. RLC systems detect a vehicle that passes a stop line and proceeds through the intersection after a traffic signal has turned red. In addition, the system takes a series of photographs and/or video images of the red light violation, as well as records the information such as date, time, and

The state of Texas is beginning to see a reduction in the number of communities using RLC systems as safety treatments. In 2010, Texas had 42 communities that operated RLC systems that monitored 389 signal controlled intersections. In 2011, there were 50 communities using RLC systems that monitored 398 signal controlled intersections. In 2012, only 39 communities operated RLC systems at 341 signalized intersections in Texas. While there is significant evidence that suggests RLC improve intersection safety by reducing crashes, fewer communities are adopting the practice of installing RLCs as a safety treatment. In addition, many more communities are removing the RLCs that were once in place as a result of mounting public pressure and formal municipal referendum elections.

On August 24, 2011, local news announced the ordinance publicly that the City of Houston deactivated their RLC systems at 66 individual intersections as a result of a municipal referendum election that abolished the use of RLCs in the city. In the later stages, systems were physically removed at the intersections in the city gradually. This action provided a valuable opportunity to assess the impact that removing RLC

E-mail addresses: M-Ko@tti.tamu.edu (M. Ko), Srinivas-G@tti.tamu.edu (S.R. Geedipally), T-Walden@tti.tamu.edu (T.D. Walden), r-wunderlich@tti.tamu.edu (R.C. Wunderlich).

time elapsed since the beginning of the red signal (Texas Department of Transportation (TxDOT), 2012b).

^⁎ Corresponding author.

Fig. 1. Illustration on the phases used in the research.

treatments had on traffic safety, most notably crash frequency and injury severity post-removal. The objectives of this study were to:

• 1. Evaluate the performance of RLC systems (after installation) using an Empirical Bayes (EB) before-after analysis (Phase I and Phase II as illustrated in Fig. 1).
• 2. Evaluate intersection safety (i.e., the changes of RLR crashes and severity) using an EB before-after analysis after RLC deactivation at the formerly treated intersections (Phase II and Phase III in Fig. 1).
• 3. Evaluate the spillover effect (i.e., the cameras influence on safety of locations where cameras have not been installed) on RLR crashes

and severity at nearby non-treated intersections in Houston after the RLCs were deactivated (Phase II and Phase III in Fig. 1)

While there have been many studies performed on the safety benefits of installing RLC systems to increase intersection safety, few have addressed the safety impact that deactivation has on intersection safety by just analyzing the RLR violations. This study analyzed the safety effects of RLC installation and then deactivation using the traffic crashes occurred at the same intersections in different periods. This study uses an EB methodology to address regression-to-mean (RTM) bias. RTM phenomenon suggests that there is a possible tendency for a fluctuating

Table 1 Summary of research studies on the RLC installation and deactivation.

Category	Study	Method	Crash/violation change percentagea (crash type)	Notes
After RLC	Erke (2009)	Meta-analysis	• +15% (all crashes) • +40% (RE) • −10% (RA)	Control for RTM and Spillover effect
	Hoye (2013)	Meta-analysis	• +6% (all crashes) • +13% (all injury) • −33% (RA) • +39% (RE)
	Ng, Wong, and Lum (1997)	CG method	• −7% (all type) • −8% (RA)	No control for RTM and Spillover effect
	Retting and Kyrychenko (2002)	Generalized linear regression model	• −7% (all type) • −32% (RA) • +3% (RE)	Control for RTM and Spillover effect, but the results are based on city-wide effects, not just at RLC sites
	Persaud, Council, Lyon, Eccles, and Griffith (2005)	EB method	• −25% (RA) • +15% (RE)	Control for RTM and Spillover effect
	Washington and Shin (2005)	EB method	• −20% (RLR angle) • −45% (RLR left turn) • +41% (RLR RE)	Control for RTM and Spillover effect
	Walden, Geedipally, Ko, Gilbert, and Perez (2011)	CG method	• −26% (all RLR type) • −19% (RLR RA) • +44% (RLR RE)	Partial control for RTM and Spillover effect
	Ko, Geedipally, and Walden (2013)	EB method	• −20% (all RLR crashes) • −24% (RLR RA) • +37% (RLR RE)	Control for RTM and Spillover effect
Spillover effect	Hoye (2013)	Meta-analysis	• −7% (RA) • +8% (RE)	Non-RLC intersections
	Persaud et al. (2005)	EB method	• −8.5% (RA) • +1.8% (RE)	Non-RLC intersections
	Washington and Shin (2005)	EB method	• −17% (RLR RA) • −40% (RLR left turn) • +45% (RLR RE)	Other approaches at RLC intersections
After RLC turning off	Walden et al. (2011)	Relative risk	• +200% (RLR violation): one year later after turning off
	Porter, Johnson, and Bland (2013)	Relative risk	• +300% (RLR violation): immedi ately after turning off • +400% (RLR violation): one year later after turning off

Note: RA stands for right-angle crash type; RE stands for rear-end crash type.

^a Negative values represent decrease in crashes/violations after the treatment installation/deactivation.

characteristic of the treatment site to return to a typical value in the period after an extraordinary value has been observed (Hauer, 1997). The EB method was used to estimate intersection safety at targeted sites based upon reference sites with similar traits and where RLC systems were not installed.

2. Background

Previous research provides varied and meaningful viewpoints in understanding the impact that RLC treatments have on intersection safety. Table 1 provides a summary list of research results regarding RLC.

2.1. Before and after RLC installation

Comparing before and after RLC installation data, in many cases reductions were observed for all crash types, right-angle, and left-turn collisions except for the studies of Erke (2009) and Hoye (2013). Ng et al. (1997) evaluated the safety benefit of 42 intersections in Singapore that were treated and non-treated. The non-treated intersections used as comparison had a similar configuration to that of the treated intersections. The study results indicated a 7% reduction in all crash types and an 8% reduction in right-angle crashes after the RLC treatments were used.

Some studies controlled for the spillover effect. Retting and Kyrychenko (2002) analyzed RLC system effects by considering 29 months of before and after crash data from 125 intersections that included 11 intersections with RLCs. Three similarly sized cities that did not implement RLCs were used for comparison. The comparison cities were located more than 100 miles from treated intersections to control the spillover effect. The results of the study indicated that all crash types within the treated city were significantly reduced by 7% and that rightangle crashes were reduced by 32%. The study revealed, though not significantly, that rear-end crashes increased by 3%.

Persaud et al. (2005), using an EB methodology, evaluated the effects of RLC treatments that occurred at 132 intersections in seven communities across the United States. For 6 individual communities that employed RLC treatments, the evidence suggested that right-angle crashes decreased between 14% and 40% while increasing by less than 1% at one. Rear-end crashes were found to have increased between 7% and 38% in all communities. When all the communities were combined, right-angle crashes decreased by 25% and rear-end crashes increased by 15%.

In the research using a meta-analysis approach, the findings indicated increases for all crash types and rear-end crashes after RLC were installed and a decrease for right-angle crashes (Erke, 2009; Hoye, 2013). According to Erke (2009), all crash types increased by 15% and rear-end crashes increased by 40%. At the same time, right-angle crashes decreased by 10%. Unfortunately all the studies detailed above considered all crashes at intersections without identifying if crashes were caused by RLR violations. This is important to the process of RLC evaluation because the primary role of the RLC system is to provide a deterrent to drivers who run red signals and to reduce crashes due to RLR violations.

Washington and Shin (2005) analyzed RLR crashes at 10 Phoenix and 14 Scottsdale intersections that were equipped with RLC systems. Based on a comparison group (CG) methodology and using Phoenix data, the researchers discovered that angle crashes decreased by 42% and left-turn crashes decreased by 10%. However, rear-end crashes were discovered to have increased by 51%. Using the Scottsdale data and an EB methodology, the authors discovered that angle crashes declined by 20% and left-turn crashes were reduced by 45% at the treated intersections. As was observed in the Phoenix study, rear-end crashes were observed to have increased by 41%. The overall conclusions suggest that RLC treatments provided a positive influence in reducing angle and left-turn RLR crashes. Consequently, RLC treatments negatively influenced rear-end RLR crashes.

Walden et al. (2011) evaluated RLC effectiveness using a CG methodology across 296 intersections in 39 Texas communities. The research findings established that all crash types decreased by 26% and that right-angle RLR crashes decreased by 19%, respectively. As was determined in other research, researchers discovered that rear-end RLR crashes increased by 44%.

Ko et al. (2013) conducted additional research using an EB methodology at 254 RLC treated intersections in Texas. The results indicated that all crash types decreased by 20% and that right-angle RLR related crashes decreased by 24%. Similarly, as was found in the CG study, rear-end RLR crashes increased by 37%. The researchers also assessed the impact of placement of RLC at intersections based on crash rates. It was discovered that if RLCs are placed at intersections with RLR crash rate of less than 1 per 10,000 vehicles then all RLR crash types increase by 13% and right-angle RLR crashes increase by 11%. When the intersections have a crash rate greater than or equal to 1 but less than 2 per 10,000 vehicles all crash types decrease by 22% and right-angle RLR crashes decreased by 14%, respectively. For intersections with the rate greater than or equal to 2, the results indicated that all crash types were significantly reduced by 29% and right-angle RLR crashes significantly decreased by 27%. These results suggest that as crash rates at signal controlled intersections increase, the more successful the RLCs will be at improving traffic safety. This supports the position of placing RLC treatments only at signal controlled intersections that exhibit higher crash rates.

2.2. Spillover effect of RLC installation

The spillover effect describes the reaction that takes place in other locations as a result of the installation or removal of a treatment from a specific location. Usually this is a similar reaction that is caused by the treatment. One example would be the placement of a RLC at a specific intersection. As a result of the placement of the RLC (over time), all crash types and right-angle crashes are reduced. At nearby intersections where no RLC treatment was placed, researchers found that they too experience reductions in all crash types including rear-end collisions and right-angle collisions (Hoye, 2013; Persaud et al., 2005; Washington & Shin, 2005). In short, the safety benefit of the treatment spilled over to the non-treated intersection.

Washington and Shin (2005) evaluated the safety benefits of nonmonitored approaches at RLC treated intersections. The researchers concluded that there were benefits in crash reductions for the unmonitored lanes but that the majority of the safety benefit was associated with the approaches monitored by the RLC system. In the Hoye (2013) and Persaud et al. (2005) studies, the researchers conducted evaluations at non-treated intersections. Their findings indicated that there were decreases in right-angle crashes by 7% and 8.5% and increases in rear-end crashes by 8% and 1.8%. Table 1 provides a summary of research that has been performed relating to spillover effect and RLC deactivation.

2.3. RLC deactivation

Porter et al. (2013) evaluated RLR behavior at signalized intersections in Virginia that were formerly treated with RLC systems. The researchers measured the frequency of traffic light cycles with at least one red light runner during observation periods. In the immediate months after the RLC were deactivated, RLR violations were three times greater than in the period when the treatments were active. One year later, at the same intersections, the RLR violations had risen to four times greater than when the RLCs were enforcing RLR signal violations.

Similar results were observed at intersections in College Station, Texas. In November 2009, the City of College Station deactivated the RLC systems within the city. Walden et al. (2011) analyzed the changes of RLR violation frequencies at the intersections before and after the RLCs were deactivated. Significant increases in the frequency of RLR violations were observed at the intersections post removal of the treatments. RLR violations increased three times higher after deactivation relative to the periods of time when RLC were operational.

The Porter and Walden studies suggest that communities who choose to remove RLC systems from treated intersections will experience increases in risk taking behavior by drivers at signalized intersections.

3. Data collection

Crash information originated from electronic copies of stored crash records maintained in the Texas Department of Transportation (TxDOT) Crash Records Information System (CRIS) database. Individual crash records were remotely accessed by interfacing with CRIS and searching the database using crash identification numbers assigned to each crash record. Researchers collected crash and traffic data at 48 intersections that were formerly treated with RLCs during Phase I, II, and III in Houston, Texas (see Fig. 1). The researchers have made sure that there were no countermeasures during the study period, other than the RLCs.

Target (or RLR) crashes included those crash events taking place inside the intersection where one vehicle disregards the red signal, plus any intersection-related rear-end crash event occurring as a consequence of heavy braking of a leading vehicle in anticipation of a yellow signal turning to red while both leading and following vehicles are traveling in the same approach direction (Ghorghi, 2016). Each of the crash reports was manually reviewed to determine the crash types and a crash associated with RLC operation. Although some crashes occurred due to signal violations or as a result of RLC influence, they were not counted toward the target crashes when they occurred under the following conditions: 1) driving under the influence, 2) adverse weather condition, such as heavy rain or icing on roadways, 3) cut-in front of traffic from a side lane, or 4) related to emergency vehicles responding in emergency circumstances. For fatal/injury RLR crashes, they were defined as RLR crashes resulting in any fatality, incapacitating/nonincapacitating injury, or possible injury.

Average Daily Traffic (ADT) data were collected from the postactivation reports submitted to TxDOT and geographic information and management system (GIMS) operated by Houston. However, researchers could not collect each year ADT data during the Phase I and II at a few treated intersections and the Phase II at nearby non-treated intersections due to a lack of information on traffic data. Researchers reviewed the changes of traffic volumes at other intersections and highway segments in Houston during the periods and then averaged the changes in traffic volume (i.e., a 3% increase). For the missing data, ADT during the periods were extrapolated with the assumption that traffic volume annually increased by 3%. Table 2 summarizes the number of crashes for all, right-angle, and rear-end crash types at the studied intersections.

Crash and traffic data at 27 nearby non-treated intersections were evaluated to identify spillover effects for the data that were collected between 2008 and August of 2014. Intersections were selected because of their close proximity to the closest RLC treated intersection. Collected crash data were processed for the purposes of identifying target crashes. Summarized information for formerly treated and non-treated intersections can be found in Table 2.

After the RLC crashes were categorized into Phases I, II, and III, there were about 594, 323, and 354 RLR all crash types per year respectively. For fatal/injury RLR crashes, there were about 164 and 157 crashes per year during the Phases II and III respectively. However, the crash severity information was not available in Phase I period. In order to measure spillover effect, all crash types had to be obtained for the non-treated intersection for Phases II and III respectively. At the non-treated intersections, there were about 58 and 76 RLR all crash types observed per year during the Phase II and Phase III. Also there were 24 and 31 fatal/injury crashes per year during the same phases.

To develop safety performance functions (SPFs), crashes, severity index, and traffic data at 66 intersections without RLC systems were collected between 2007 and 2010. In order to minimize spillover effects, reference intersections used for comparison purposes were selected based upon their distance from the treated location. Distance must have been at least 2 miles away from the closest treated intersection. All reference intersections were located in 6 jurisdictions in Texas. At the reference intersections for developing SPFs, 108 all crash types, 57

Table 2 Summary statistics of formerly treatment and non-treated intersection groups.

Intersection types	Period	Variable		Min	Max	Mean	Std. dev.	Sum
Treated	Phase I	RLR crashes per year	All RLRa	0	70	12.4	16.72	593.6
			RA RLRa	0	70	12.2	16.47	584.9
			RE RLRa	0	2	0.1	0.33	5.3
		Study period (yr)		1	4	2.4	0.61	115
	Phase II	RLR crashes per year	All RLRa (FIb )	0 (0)	39.9 (20.9)	6.7 (3.4)	7.4 (4.24)	323.1 (164.3)
			RA RLRa (FIb )	0 (0)	34.9 (20.9)	5.3 (2.8)	6.6 (3.89)	254.4 (135.4)
			RE RLRa (FIb )	0 (0)	2.2 (0.8)	0.5 (0.2)	0.54 (0.24)	23.4 (7.9)
		Study period (yr)		3.6	3.6	3.6	0	176
	Phase III	RLR crashes per year	All RLRa (FIb )	0 (0)	38.6 (17.6)	7.4 (3.3)	7.85 (3.74)	353.7 (156.7)
			RA RLRa (FIb )	0 (0)	30 (14.7)	5.6 (2.6)	6.61 (3.25)	268.3 (126.3)
			RE RLRa (FIb )	0 (0)	3.3 (1.7)	0.8 (0.3)	0.78 (0.39)	36 (14.7)
		Study period (yr)		3	3	3	0	144
Nearby non-treated	Phase II	RLR crashes per year	All RLRa (FIb )	0 (0)	6.3 (2.8)	2.2 (0.9)	1.7 (0.8)	58.2 (23.9)
			RA RLRa (FIb )	0 (0)	3.85 (2.5)	1.4 (0.6)	1.17 (0.67)	37.6 (15.7)
			RE RLRa (FIb )	0 (0)	0.7 (0.3)	0.1(0.01)	0.18 (0.06)	3 (0.3)
		Study period (yr)		3.6	3.6	3.6	0	99.3
	Phase III	RLR crashes per year	All RLRa (FIb )	0.3 (0)	7.3 (3.7)	2.8 (1.2)	1.72 (0.87)	75.7 (31.3)
			RA RLRa (FIb )	0 (0)	4 (2.7)	1.7 (0.8)	1.21 (0.65)	46.3 (20.3)
			RE RLRa (FIb )	0 (0)	1.3 (0.7)	0.5 (0.1)	0.37 (0.19)	12.7 (3.3)
		Study period (yr)		3	3	3	0	81
Reference	2007 to 2010	RLR crashes per year	All RLRa (FIb )	0 (0)	5.8 (5.5)	1.6 (1)	1.45 (1.04)	108 (64.3)
			RA RLRa (FIb )	0 (0)	5 (4.3)	0.9 (0.5)	1.12 (0.83)	57.3 (34.3)
			RE RLRa (FIb )	0 (0)	2 (1.3)	0.4 (0.2)	0.42 (0.29)	26.5 (16)
		Study period (yr)		4	4	4	0	192

Note: RA stands for right-angle crash; RE stands for rear-end crash.

^a All severity crashes including property damage only (PDO).

^b Fatal/injury crashes.

right-angle and 27 rear-end RLR crashes per year were reported. For fatal/injury RLR crashes, there were 64 all crash types, 34 right-angle, and 16 rear-end crashes per year (Table 2).

Table 3 provides a summary of statistics regarding variables collected at the intersections in reference group conditions. Other collected variables included lane width, shoulder width, yellow time, red time, channelization, and median width among others.

4. Methodology

The primary concern in a before-after study that evaluates the effectiveness of RLC systems at treated sites is the RTM bias. This bias exists due to the methods prediction concerning the expected number of target crashes from the treatment site based upon before-period crash numbers only. An EB method is useful for adjusting the RTM bias. The primary element in this method is to predict what would have been the expected frequency of target crashes in the after period for each treated site, had the treatment not been applied. An EB method is superior to other methods in that it predicts the expected number of target crashes of a site based on two pieces of information: (a) actual number of crashes at treated site during the before period and (b) the information about the safety of reference sites with similar geometric characteristics. The expected crash frequency (E[k|K]) at a treated site is a result from the combination of the predicted crash count (E[k]) based on the reference sites with similar traits and the crash history (K) of that site. Therefore, given the predicted crash frequency (E[k]) and the historic crash frequency (K),

The expected crash frequency is:

$$E[k|K] = w \cdot E[k] + (1-w) \cdot K \tag{1}$$

And its variance V[k|K] is:

$$V[k|K] = (1-w) \cdot E[k|K] \tag{2}$$

where w is a weight between 0 and 1 and is calculated as:

$$w = \frac{1}{1 + \frac{V[k]}{E[k]}} \tag{3}$$

The parameter E[k] is estimated from SPFs developed using a negative binomial regression (also known as, Poisson-gamma) model under the assumption that the covariates in the SPFs represent the main safety traits of the reference sites (Washington & Shin, 2005). The procedure for using the before-after study with the EB method is described below.

4.1. Step 1. Develop SPFs

Using crash, traffic, and geometric data from the reference sites, develop SPFs using the negative binomial regression models for each of all RLR crash types, right-angle RLR crashes, rear-end RLR crashes for all severity RLR crashes including property damage only (PDO), as well as ones for fatal/injury RLR crashes. The negative binomial regression

Table 3 Summary statistics of variables at reference intersections.

Intersection types	Variable			Min	Max	Mean	Std. dev.	Sum
Reference	RLR crashes	All RLRb (FIc )		0 (0)	23 (22)	6.5 (3.9)	5.8 (4.2)	432 (257)
		RA RLRb (FIc )		0 (0)	20 (17)	3.5 (2.1)	4.5 (3.3)	229 (137)
		RE RLRb (FIc )		0 (0)	8 (5)	1.6 (1.0)	1.7 (1.2)	106 (64)
	Major	ADTa		5,750	64,914	23,884	9,780	–
		One wayd		0	1	0.02	–	–
		LT baye		0	2	1.89	–	–
		RT bayf		0	2	0.77	–	–
		RT channelizatione		0	2	0.45	–	–
		Median presencee		0	2	1.15	–	–
		Protected LT signalf		0	1	0.30	–	–
		Yellow interval (second)		4.0	4.7	4.06	–	–
		All red (second)		1.0	2.6	1.67	–	–
		Thru lane	Dir 1	0	3	2.26	–	–
			Dir 2	1	3	2.30	–	–
		Lane width (ft)	Dir 1	9	15	11.14	–	–
			Dir 2	9	16	11.14	–	–
		Speed limit (mph)	Dir 1	30	55	39.85	–	–
			Dir 2	30	55	39.62	–	–
	Minor	ADTa		2,080	29,885	15,629	7,901	–
		One wayd		0	1	0.02	–	–
		LT baye		0	2	1.89	–	–
		RT baye		0	2	0.86	–	–
		RT channelizatione		0	2	0.59	–	–
		Median presencee		0	2	1.29	–	–
		Protected LT signalf		0	1	0.33	–	–
		Yellow interval (second)		3.0	5	4.01	–	–
		All red (second)		1.0	2.6	1.71	–	–
		Thru lane	Dir 1	0	4	2.29	–	–
			Dir 2	1	3	2.33	–	–
		Lane width (ft)	Dir 1	9	13	11.24	–	–
			Dir 2	10	20	11.33	–	–
		Speed limit (mph)	Dir 1	30	50	37.88	–	–
			Dir 2	20	50	37.73	–	–

^a ADTmaj and ADTmin are the ADT for major and minor roadways at intersections. ^b All crashes including PDO.

^c Fatal and injury crashes.

^d 0 = two ways, 1 = one way.

^e 0 = no bay, lane, or median, 1 = bay, lane(s), or median in one direction only, 2 = bay, lane(s), or median in both directions.

^f 1 = protected-only left-turn operation, 0 = permissive or protected-permissive; "–" = no data.

model is a pervasive type of model that is regularly used by transportation safety analysts for explaining and expressing the nature of traffic crashes. The negative binomial regression model is preferred over other mixed-Poisson models since the gamma distribution is a conjugate to the Poisson distribution. The Poisson distribution expresses the probability of a given number of events that occur in a fixed interval of time and/or space if the events occur within a known average rate independent of the time of the last event.

The negative binomial regression model has the following structure: the number of crashes 'Yit' for a particular i th site and time period t when conditional on its mean μit is Poisson distributed and independent over all sites and time periods.

$$Y{it}|\mu{it} \sim Po(\mu_{it})$$ $i = 1, 2, ..., t$ and $t = 1, 2, ..., t$ (4)

The mean of the Poisson is structured as:

$$\mu{it} = f(X; \beta) \exp(e{it}) \tag{5}$$

where:

f(.) is a function of the covariates (X).

β is a vector of unknown coefficients.

eit is the model error independent of all the covariates.

Although different functional forms with the collected variables were tried, the best fit functional form used for each crash type in this study is as follows:

$$E[k_z] = e^{\beta0} \cdot \left( ADT{maj} + ADT_{min} \right)^{\beta_1} \tag{6}$$

where:

E[kz] is the predicted number of crashes for specific crash type with z = All, RA, or RE (i.e., All = all RLR crash types, RA = rightangle RLR crashes, and RE = rear-end RLR crashes).

β^a is a vector of unknown coefficients (to be estimated) (a = 0,1).

4.1.1. Estimation for SPFs

Coefficients were combined with Eq. (6) to obtain the crash mean for each crash type. In general, the trend of each variable is logical and intuitive. Particularly, the estimation results suggest that RLR crashes increase with the increases in total traffic flow at the signal controlled intersections. Table 4 summarizes the estimation results of SPFs for all crash types, right-angle, and rear-end RLR collisions.

4.2. Step 2. Estimate the predicted number of crashes (^ E½k-

Apply the crash modification factor (CMF) to account for the effect of RLC systems on safety, estimate the predicted number of crashes for time period (t) and crash type (z) in the before period at each treatment site (i) using the developed SPF in Step 1 as follows:

$$\hat{E}[k_{itz}] = CMF_z \cdot e^{\beta0} \cdot \left(ADT{maj.\ it} + ADT_{min.\ it}\right)^{\beta_1} \tag{7}$$

where:

CMFz is a crash modification factor for specific crash type with z = All, RA or RE.

ADT maj. it and ADT min. it are the ADT for major and minor roadways at a treated site (i) for time period (t).

In the analyses related to the RLC installation at the treated intersections and the spillover effect at non-treated intersections, all CMFs are 1. For the RLC deactivation analysis, CMFs from Ko et al. (2013) are used (CMFAll=0.80; CMFRA=0.76; and CMFRE=1.37).

4.3. Step 3. Compute the predicted number of crashes (Pi)

Compute the sum of the annually predicted numbers of crashes by each crash type in the before period at each treatment sites using the equation:

$$P{iz} = \sum{t=1}^{t{GL}} \hat{E}[k{itz}] \tag{8}$$

where t0i⁻ denotes the before period in the year which RLC systems were installed/removed at site i.

4.4. Step 4. Predict the expected number of crashes (π) and estimate Var½π^-

Based on Eq. (1), predict the expected number of crashes at particular i th site with the equation as follows:

$$\hat{\pi}{iz} = C{iz} \cdot { w{iz} \cdot P{iz} + (1 - w{iz}) \cdot K{iz} }$$ (9)

where Kiz is the sum of the reported number of all types, RA, or RE RLR crashes during the before period at site i and the adjustment factor, Ciz, can be calculated as follows:

$$C{iz} = \frac{\sum{t=t{0i}}^{t} \hat{E}[k{itz}]}{\sum{t=1}^{t=t{0i}} \hat{E}[k_{itz}]}$$ (10)

where t0i⁺ denotes the after period in the year which RLC systems were installed/removed at site i. Also, the wiz in Eq. (9) can be calculated as:

$$w_{iz} = \frac{1}{1 + \alphaz \cdot P{iz}} \tag{11}$$

where α^z is the over-dispersion parameter of a negative binomial regression model. The expected number of after-period crashes and

Table 4 Estimates SPFs for reference intersections.

| Variable | All severitya | | | Fatal/injury (FI) | | | | |--------------------------|-----------------|----------------|------------------|-------------------|-----------------|-----------------|--| | | All | RA | RE | All | RA | RE | | | Constant (β0) | −4.783 (2.888) | −2.308 (4.017) | −11.3326 (3.387) | −5.722 (3.154) | −4.236 (4.462) | −12.195 (3.882) | | | ADTmaj+ADTmin(β1) | 0.5005 (0.2746) | 0.2060 (0.382) | 0.9848 (0.319) | 0.5401 (0.299) | 0.3401 (0.424) | 1.0183 (0.346) | | | Dispersion parameter (α) | 0.7779 (0.1783) | 1.6317 (0.385) | 0.3844 (0.213) | 0.8345 (0.212) | 1.8202 (0.4832) | 0.2406 (0.244) | | | Log-likelihood | −192.8 | −154.4 | −108.9 | −161.8 | −125.1 | −84.9 | | | AIC | 391.7 | 314.9 | 223.9 | 329.7 | 256.2 | 175.9 | | | BIC | 398.3 | 321.5 | 230.5 | 336.3 | 262.7 | 182.4 | |

Note: ADTmaj and ADTmin are the ADT for major and minor roadways at intersections; Akaike's information criterion (AIC); Bayesian information criterion (BIC). ^a All severity RLR crashes including PDO; the values in parentheses represent standard errors.

their variances for a group of sites had the treatment not been implemented at the treated site is given as:

$$\hat{\pi}z = \sum{i=1}^n \hat{\pi}_{iz} \tag{12}$$

n represents the total number of sites in the treatment group, and π^^z is the expected after-period crashes at all treated sites by crash types had there been no treatment. The estimate of variance of π^^z can be calculated from the equation as follows:

$$V\hat{a}r[\hat{\pi}{iz}] = C{iz} \cdot (1 - w{iz}) \cdot \hat{\pi}{iz} \tag{13}$$

$$V\hat{a}r[\hat{\pi}z] = \sum{i=1}^n V\hat{a}r[\hat{\pi}_{iz}] \tag{14}$$

4.5. Step 5. Compute the observed target crashes (λ)

For a treated site, the crashes in the after-period are influenced by the implementation/deactivation of the treatment. The safety effectiveness of a treatment is known by comparing the actual crashes with the treatment to the expected crashes without the treatment. The estimated crash frequency in the after-period crashes for a group of treated sites is given as:

$$\lambdaz = \sum{i=1}^n \lambda_{iz} \tag{15}$$

where λiz is the observed all type, RA, or RE RLR crash frequency during the after period at a site i. The λ^z is equal to the sum of the observed number of crashes at all treated sites by the crash types during the after study period.

4.6. Step 6. Estimate δ and θ

The 'change in the safety (δ)' and 'index of effectiveness (θ)' are defined as the difference and the ratio of what safety is with the treatment to what it would have been without the treatment respectively. These parameters give the overall safety effect of the RLC treatment and are given by:

$$\hat{\delta}{z} = \hat{\pi}{z} - \hat{\lambda}_{z} \tag{16}$$

$$\hat{\theta}{z} = \frac{\left(\frac{\hat{\lambda}{z}}{\hat{\pi}{z}}\right)}{\left(1 + \frac{V\hat{a}r(\hat{\pi}{z})}{\hat{\pi}_{z}^{2}}\right)}$$

$$(17)$$

If ^δ^z is greater than zero and ^θ^z is less than one, then the treatment has a positive safety effect. In addition, the percent decrease in the number of target crashes due to the treatment is calculated as 100ð1−^θ^zÞ percent.

4.7. Step 7. Estimate Var½ ^δ and Var½ ^θ-

The estimated variance and standard error of the estimated safetyeffectiveness are given by:

$$V\hat{a}r\left[\hat{\delta}{z}\right] = \hat{\pi}{z} + \hat{\lambda}_{z} \tag{18}$$

$$V\hat{a}r\left[\hat{\theta}{z}\right] = \frac{\hat{\theta}{z}^{2} \cdot \left[\frac{1}{\hat{\lambda}{z}} + \frac{V\hat{a}r(\hat{n}{z})}{\hat{n}{z}^{2}}\right]}{\left[1 + \frac{V\hat{a}r(\hat{n}{z})}{\hat{n}_{z}^{2}}\right]^{2}}$$ (19)

$$s.e.\left[\hat{\theta}{z}\right] = \sqrt{V\hat{a}r\left[\hat{\theta}{z}\right]} \tag{20}$$

5. Analysis results

This section reports the findings of the EB evaluation that was performed to assess the changes in safety that occurred as a result of deactivating RLC systems at formerly treated intersections in Houston, Texas. In addition, the evaluation also measures the change in fatal/injury crashes at formerly treated intersections as well as accounts for the spillover effect at non-treated intersections in Houston after RLC deactivation. Table 5 presents the results of EB analyses to the effect of the RLC enforcement systems at the 48 treated intersections and 27 nearby nontreated intersections in Houston.

5.1. Effects of RLC systems installation in Houston at treated intersections (Phase I vs. Phase II)

According to the result of RLC installation in Table 5(a), there were 1,176 RLR crashes reported during the period of time when RLC were in place and active (Phase II). The analysis revealed that if the treatment had not been installed, the expected number of RLR crashes would be approximately 1,862 crashes. Thus, there was a positive safety effect and one would expect to see approximately 686 fewer RLR crashes with the use of RLC systems. The resulting index of effectiveness is 0.63, which indicates that the estimated average safety effect of the RLC systems is to decrease all RLR crash types by 37%. This finding is statistically significant at a 95% confidence level. One may reasonably expect a decrease in all RLR crash types between 32% and 42%.

For right-angle RLR crashes, a reduction of 819 crashes is expected as an outcome of the treatment being present and active. The average safety effect of the RLC system on right-angle RLR crashes indicates that there is a significant decrease in crashes by 47%. At a 95% confidence level, a decrease of the crashes between 43% to 51% is expected.

Contrary to all crash types and right-angle RLR crashes, an increase was observed in rear-end RLR crashes after the implementation of RLCs. Overall, there were 14 additional rear-end RLR crashes occurred at the treated intersections since RLC were installed. The average safety effect of RLCs systems on rear-end RLR crashes is estimated to be an increase in crashes by 18%. However, this result is not statistically significant.

5.2. Effects of RLC systems deactivation at treated intersections (Phase II vs. Phase III)

In Phase III (deactivation) of the evaluation, 1,061 RLR crashes were reported. It was determined that if the RLC treatments had been allowed to continue, the expected number of the crashes would have been approximately 881 RLR crashes (see Table 5(b)). As a result of deactivating the RLCs, the intersections experienced greater numbers of certain types of collisions. This counter-effect on intersection safety resulted in 180 additional RLR crashes when the RLC systems were not present and active. The average safety effect of deactivating the RLC systems is estimated to increase all RLR crash types by 20%. The standard error of this estimate is 5%. At a 95% confidence level, the result is statistically significant and one may expect an increase in crashes ranging from 11% to 30%. The level of all RLR crashes returned to a level similar to that which existed prior to camera installation after the deactivation at same intersections.

Table 5 Average safety effects of treated intersections.

	(a) Effects of RLC systems installation in Houston at treated intersections (Phase I vs. Phase II)
Crash type	λ^ (observed crashes)	π^ (expected crashes)	^δ	^δ s:e:½		^θ (index of effectiveness)
				(change in safety π^−λ^)	(standard error)	Estimate	s .e.	Confidence level
All RLRa	1,176	1,862	686		49.18	0.63	0.02	N99%
RA RLRa	926	1,745	819		55.15	0.53	0.02	N99%
RE RLRa	85	71	−14		11.15	1.18	0.16	76.2%
	(b) Effects of RLC systems deactivation at treated intersections (Phase II vs. Phase III)
Crash type	λ^	π^	^δ	^δ s:e:½	^θ
					Estimate	s.e.		Confidence level
All RLRa	1,061	881	−180	41.14	1.20	0.05		N99%
RA RLRa	805	654	−151	35.40	1.23	0.06		N99%
RE RLRa	108	95	−13	12.35	1.14	0.13		68.8%
	(c) Effects of RLC systems deactivation on fatal/injury crashes at treated intersections (Phase II vs. Phase III)
Crash type	λ^	π^	^δ	^δ s:e:½	^θ
					Estimate	s.e.		Confidence level
All RLRb	470	427	−43	27.27	1.10	0.07		85.9%
RA RLRb	379	335	−44	24.33	1.13	0.18		91.2%
RE RLRb	44	57	13	7.7	0.77	0.1		92.9%
	(d) Effects of RLC systems deactivation at nearby non-treated intersections (Phase II vs. Phase III)
Crash type	λ^	π^	^δ	^δ s:e:½	^θ
					Estimate	s.e.		Confidence level
All RLRa	227	177	−50	18.81	1.28	0.12		98.3%
RA RLRa	139	108	−31	14.69	1.27	0.15		93.5%
RE RLRa	38	25	−13	6.83	1.48	0.29		90.0%
	(e) Effects of RLC deactivation on fatal/injury crashes at nearby non-treated intersections (Phase II vs. Phase III)
Crash type	λ^	π^	^δ	^δ s:e:½	^θ
					Estimate	s .e.		Confidence level
All RLRb	94	78	−16	12.04	1.20	0.16		76.9%
RA RLRb	61	48	−13	9.62	1.26	0.22		77.6%
RE RLRb	10	18	8	3.61	0.56	0.18		98.3%

Note: a negative values means crash increases; λ^ is the sum of the observed number of crashes at all treated sites during the after study period; π^ is the sum of the expected after-period crashes at all treated sites had there been no treatment; ^δ is the difference between ^ ^λ and ^π^; ^^θ is index of effectiveness; ^s:e:½ ^δ is the standard error of ^δ; RA stands for right-angle crash; RE stands for rear-end crash.

For right-angle RLR crashes, an increase of 151 additional crashes would be expected at formerly treated intersections after deactivation. The average safety effect of RLC deactivation on right-angle RLR crashes indicates that there is an increase in crashes by 23%. At a 95% confidence level, one may expect an increase in crashes ranging from 11% to 35%.

For rear-end RLR crashes, there was also an increase in crashes after the deactivation of RLCs. Overall, there were approximately 13 more rear-end RLR crashes that occurred at the formerly treated intersections (Phase III). However, the average safety effect of RLCs systems deactivation on RE RLR crashes is not statistically significant.

5.3. Effects of deactivation on fatal/injury crashes at treated intersections (Phase II vs. Phase III)

For all RLR crash types, there were 470 fatal/injury RLR crashes reported after the RLCs were deactivated (Phase III). The analysis results indicated that if the treatment had not been deactivated, the expected frequency of fatal/injury crashes would have been 427 crashes (Table 5(c)). This would mean that there would have been 43 fewer fatal/injury crashes at the intersections that had previously been treated. The safety effect of deactivation is estimated to increase fatal/injury RLR crashes by 10%. This result, however, is not statistically significant at a 95% confidence level.

For right-angle RLR crashes, there were 379 fatal/injury crashes reported after the RLC systems were deactivated. Had the RLC been in place and operational, approximately 335 fatal/injury crashes would have expected. As such, there was an increase of approximately 44 additional crashes related to RLR. The average safety effect of the RLC deactivation on right-angle fatal/injury crashes indicated an increase by 13%. This result, however, is not statistically significant.

The research results suggest that there is a reduction of fatal/injury crashes related to rear-end RLR after the deactivation of RLC treatments at previously treated intersections. Overall, there were 44 rear-end RLR crashes reported during the time in which the RLC were deactivated (Phase III). It was further determined that if RLC systems remained in place and operational there would have been approximately 57 crashes for that same time period. As a consequence of deactivating the RLC treatments from the intersection, 13 fewer fatal/injury crashes have occurred. Safety benefits of RLCs systems on fatal/injury crashes that involve rear-end RLR crashes are improved after the RLCs treatments are removed from the intersections. However, this result is not statistically significant at a 95% confidence level.

^a All severity crashes including PDO.

^b RLR crashes resulted in fatal/injury.

5.4. Effect of RLC deactivation at nearby non-treated intersections on crashes of all severity (Phase II vs. Phase III)

There were 227 all crash type RLR crashes occurred at the nearby non-treated intersections during the period of time after the RLC treatments were deactivated (Phase III). The analysis results suggest that if the treatment had not been deactivated, the expected number of RLR crashes would have been 177 during that same time period (Table 5). The results suggested that an increase of all crash type RLR crashes by 28% at 27 nearby non-treated intersections would be expected after deactivation of the RLCs at treated intersections. The range of an increase is between 5% to 51% at a 95 confidence level.

For right-angle RLR crashes, there were 139 crashes reported during the period of time immediately after deactivation (Phase III). The results of the analysis indicated that an expected 108 crashes would have occurred if the treatment had not been deactivated. Based upon the information observed, there was a difference of the frequency of right-angle RLR crash types (31 more crashes; 27% increase) at nearby non-treated intersections between before and after RLC deactivation. However, this result is not statistically significant at a 95% confidence level.

For rear-end RLR crashes, there were 38 crashes reported at nearby non-treated intersections after the RLCs were deactivated. The analysis results indicated that if the treatment had not been deactivated at treated intersections, the expected frequency of crashes would have been 25 crashes at the non-treated intersections. There were about 13 more crashes (48% increase) after the deactivation at the non-treated intersections. However, this result is not statistically significant at a 95% confidence level.

5.5. Effects of RLC deactivation on fatal/injury crashes at nearby nontreated intersections (Phase II vs. Phase III)

If RLC treatment had not been deactivated at treated intersections, the expected frequency of fatal/injury crashes at nearby non-treated intersections would have been 78 for all crash types and 48 for right-angle (Table 5(e)). Those expectations are 16 for all crash types and 13 for right-angle crashes fewer than the number of observed crashes at nearby non-treated intersections. However, both results are not statistically significant at a 95% confidence level. For rear-end fatal/injury crashes, 8 fewer crashes (44% reduction) are expected. This result is statistically significant with the expected decrease range of 8% to 80% at a 95 confidence level. The deactivation of RLR systems reduced rear-end fatal/injury crashes at nearby non-treated intersections. It is noted that rearend crashes account for smaller proportions when compared with all crash types/right-angle crashes. In this study, rear-end crashes account for 10 to 20% of all crash types at the intersections.

6. Discussions & conclusions

This study evaluated the safety performance of RLC systems at treated intersections and the potential safety effects at nearby non-treated intersections in Houston, Texas. The focus on the assessment was to evaluate safety benefits associated with RLC systems at sites where the systems had been installed, were operational, and later were deactivated. This was achieved by assessing crashes related to RLR after installation and then again after subsequent deactivation. Another purpose was to investigate safety at nearby non-treated intersections to see if the deactivation of the RLCs had any "spillover" effect.

In the analysis of before and after the installation, there are significant reductions on RLR crashes overall and right-angle RLR crashes by 37% and 47%, respectively. Those results are higher than the ones documented in the study of Ko et al. (2013), in which they concluded that there are significant decreases in all types and right-angle RLR crashes by 20% and 24% respectively from the analysis of 245 treated intersections in Texas. The reason for additional reductions at the intersections in Houston could be explained by the RLR crash records at the intersections in Phase I. The results of Ko et al. (2013) study indicated that the effectiveness of RLC systems is more likely to have significant and positive impacts at intersections with a greater RLR crash rate during the period prior to RLC installation. In their previous study, the authors noted that the average crash rate at the treated intersection was 2.35; which basically means that there were approximately 2 RLR crashes annually per 10,000 vehicles. In this study, the average crash rate at the targeted intersections was 4.17 in the before period. There are about 2 more RLR crashes per 10,000 vehicles per year at the intersections in Houston than the ones used in Ko et al. (2013). As a result, the RLC installation at the intersections with greater RLR crash rates led to more safety benefits than might otherwise be expected.

Though the RLC installation caused a positive decline in RLR violations and crashes, the deactivation resulted in an increase in crashes and a negative effect on safety. The analysis of RLC deactivation showed that RLR crashes significantly increased by 20% for all crash types and by 23% for right-angle RLR crashes. After the deactivation of the RLC systems the intersections experienced increases in all crash types and right-angle RLR crashes. According to the studies of Walden et al. (2011) and Porter et al. (2013), drivers generate more RLR violations at formerly treated intersections after RLC systems are deactivated: up to four times more than that during the RLCs operation. This study demonstrates that crashes due to RLR violations increase after RLC systems deactivation, but not to the extent that violations do.

At an intersection, a right-angle crash is a collision between vehicles on perpendicular approaches and this crash type can lead more severe injury than a rear-end crash which a collision between vehicles in the same direction. The before-after analysis associated with crash severity showed that right-angle fatal/injury RLR crashes increased since RLC systems deactivation, while rear-end RLR fatal/injury crashes decreased. RLC deactivation affects a negative safety benefit for right-angle fatal/injury crashes but positive safety benefit for rear-end fatal/injury crashes at formerly treated intersections.

In addition to the effect at the treated intersections, RLC deactivation affects RLR related crashes at nearby non-treated intersections. It is interesting to note that there were increases of RLR crashes by 28% for all crash types, 27% for right-angle, and 48% for rear-end. Those numbers are greater than that (i.e., 20% for all crash types, 23% for right-angle, and 14% for rear-end) of the treated intersections. While there are greater percentage changes at non-treated intersections than treated ones, it is important to note that the findings are assessed with the smaller crash frequencies at non-treated intersections. During the period before deactivation, there were about 2 RLR crashes per nearby nontreated intersection per year while almost 7 RLR crashes per treated intersection per year. A minor change in the crash frequency at nearby non-treated intersections can result in large percentage change. As a result, those findings reported should be interpreted with caution.

In the analysis of fatal/injury RLR crashes at nearby non-treated intersections, after deactivation of the cameras, rear-end crashes decreased by 44%. This result is totally opposite to that of all severity crashes. All severity rear-end crashes went up by 48% after deactivation at the same intersections. Rear-end crashes colliding between vehicles in the same direction can result in less severe injuries than other crash types at intersections. Similarly, rear-end fatal/injury crashes went down at treated intersections after deactivation.

This study evaluated the safety performance of installing RLC systems and also assessed RLR crashes after the RLC systems were deactivated. Researchers concluded that there is a negative effect on safety when cameras are removed at formerly treated intersections and nearby non-treated intersections as well as a positive effect related to specific crash type and severity. All crashes and right-angle RLR crashes declined after the RLC system was installed and the same types of crashes increased after the RLC system deactivation. However, rear-end crashes associated with fatal/injury decreased since RLC deactivation at both treated and nearby non-treated intersections although they accounted for small percentage of all crashes at the intersections. Overall, the degree of safety at intersections that had RLC deactivation declined and the crashes returned to similar levels that existed prior to camera installation.

Acknowledgments

The authors would like to thank Dr. Eva Shipp for providing useful comments to improve the contents of this paper.

References

• Erke, A. (2009). Red light for red-light cameras? A meta-analysis of the effects of red-light cameras on crashes. Accident Analysis and Prevention, 41, 897–905.
• Ghorghi, F. B. (2016). Behavioral, operational and safety effects of red-light cameras at signalized intersections in Alabama. (Doctoral dissertation) Auburn University.
• Hauer, E. (1997). Observational before-after studies in road safety. Oxford, United Kingdom: Pergamon Press, Elsevier Science Ltd.
• Hoye, A. (2013). Still red light for red light cameras? An update. Accident Analysis and Prevention, 55, 77–89.
• Ko, M., Geedipally, S., & Walden, T. D. (2013). Effectiveness and site selection criteria for red light camera systems. Transportation Research Record, 2327, 53–60.
• Ng, C. H., Wong, Y. D., & Lum, K. M. (1997). The impact of red light surveillance cameras on road safety in Singapore. Journal of Road and Transport Research, 2, 72–80.
• Persaud, B., Council, F. M., Lyon, C., Eccles, K., & Griffith, M. (2005). Multijurisdictional safety evaluation of red light cameras. Transportation Research Record, 1922, 29–37.
• Porter, B. E., Johnson, K. L., & Bland, J. F. (2013). Turing off the cameras: Red light running characteristics and rates after photo enforcement legislation expired. Accident Analysis and Prevention, 50, 1104–1111.
• Retting, R. A., & Kyrychenko, S. Y. (2002). Reductions in injury crashes associated with red light camera enforcement in Oxnard, California. American Journal of Public Health, 92(11), 1822–1825.
• Texas Department of Transportation (TxDOT) (2012a). Texas motor vehicle crash statistics. http://www.txdot.gov/government/enforcement/annual-summary.html (accessed June 12, 2013).
• Texas Department of Transportation (TxDOT) (2012b). Red light cameras-annual data reports. http://www.txdot.gov/government/enforcement/signage/red-light/reports. html (accessed June 12, 2013).
• Walden, T. D., Geedipally, S., Ko, M., Gilbert, R., & Perez, M. (2011). Evaluation of automated traffic enforcement systems in Texas. Texas Transportation Institute (Retrieved from) http://ftp.dot.state.tx.us/pub/txdot-info/trf/red\_light/auto\_traffic.pdf.

• Washington, S., & Shin, K. (2005). The impact of red light cameras (automated enforcement) on safety in Arizona. Report no. 550. Tucson: University of Arizona.

• Dr. Myunghoon Ko is an assistant research engineer in the Texas A&M Transportation Institute's Crash Analysis Group, the Center for Transportation Safety. He worked for the Korea Ministry of Land, Infrastructure and Transport for four and half years prior to initiating his studies for a doctoral degree. He earned Ph.D. degree in Civil Engineering of Texas A&M University in 2011. Then he has continued to work at the Texas A&M Transportation Institute. Dr. Ko has participated in Federal Highway Administration and Texas Department of Transportation research projects as well as an international project focused on traffic safety and highway geometric design. His interest areas are statistical analysis of collected crash/observed data and evaluation of safety measures, as well as sustainable highway geometric design.

• Dr. Srinivas Geedipally received his doctorate from Texas A&M University and has been with Texas A&M Transportation Institute since 2005. He is currently the Associate Research Engineer in the Center for Transportation Safety and has more than eleven years of experience in traffic safety research. He has more than forty five papers published in high-standard international journals and conferences. Dr. Geedipally has participated in numerous traffic safety research projects with state and federal governments and international sponsors. He has been a key contributor in the development of the Highway Safety Manual. He is a registered professional engineer in the state of Texas. Dr. Geedipally is a two-time recipient of the Young Researcher Award, and a Fred Burggraf award winner from Transportation Research Board.

• Dr. Troy Walden is a Research Scientist and is the Director of the Center for Alcohol and Drug Education Studies within the Center for Transportation Safety at the Texas A&M Transportation Institute (TTI). Dr. Walden has over 28 years invested in highway safety as a law enforcement officer, crash re-constructionist, traffic safety advocate, and transportation safety researcher. He has led studies on the development of safety performance functions, automated traffic enforcement systems, data driven approaches to crime and traffic safety, strategic highway safety planning and the investigation of safety issues for traffic crash analysis and alcohol enforcement.

• Mr. Robert Wunderlich is a Senior Research Engineer and the Director of the Center for Transportation Safety at the Texas A&M Transportation Institute. The Center is focused on efforts to understand and develop safer roadways and safer users through its Crash Analysis, Human Factors, Behavioral Research, Planning and Evaluation, and Teens in the Driver Seat programs. As a traffic engineer, consultant, and public works director, Mr. Wunderlich has extensive experience and private sector experience in conceiving, designing, and implementing safety improvements. He is a former International President of the Institute of Transportation Engineers and a current member of the Association of State Highway and Transportation Officials Technical Committee on Geometric Design (The Green Book Committee). He has made numerous presentations at conferences regarding traffic engineering, traffic safety and management. Mr. Wunderlich is also the Associate Director of the ATLAS Center for Advancing Transportation Leadership and Safety, a Tier 1 University Transportation Center collaboration between the University of Michigan Transportation Research Institute and the Texas A&M Transportation Institute.