IV. Literature Review of Critical Studies

Unlike many literature reviews for safety research efforts, the goal of this task was not to review a large number of studies to summarize findings on RLC program effectiveness. Because the overall Phase I goal was to produce a scientifically sound experimental plan that could overcome as many threats to validity as possible, the literature review was aimed at a shorter list of international studies judged by the study team and the oversight panel to be critical studies. To this end, all possible studies of relevance were first identified on the basis of Internet searches such as Transportation Research Information Services (TRIS), and information from parallel and recent reviews and meta-analyses conducted for FHWA, the National Cooperative Highway Research Program (NCHRP), and the Insurance Institute for Highway Safety. The final choice of critical studies included studies from both the United States and other countries with a longer history of RLC program implementation, studies that appeared to be best in terms of scientific rigor, and studies often cited by other researchers or in political discussions of RLC effectiveness. The team scanned a number of study sources and reports and ultimately defined a listing of 17 critical studies.

A study team member then reviewed each of these studies in detail. The goal was to not only extract information on measured RLC program effectiveness, but also identify problems or issues that we would attempt to overcome in this new evaluation design. To accomplish this, listings of study strengths and weaknesses were developed for each study reviewed.

In the sections that follow, a general summary of the literature findings is presented first, followed by an itemization of the lessons learned from this exercise.

Summary of Findings

The studies reviewed varied widely, including the following areas:

Accident types (all, right-angle, those caused by red-light running).
Accident severities (all, injury plus fatal, weighted).
Area of study (treated intersections, treated approaches, jurisdiction-wide).
Use and designation of comparison sites.
Treatment type (cameras only, cameras plus warning signs, red-light-running and speed cameras).
Sample sizes, ranging from 3 to 78 camera-equipped intersections.
Countries (several from Australia and the United Kingdom, but few from the United States).
Study methodology (simple before-and-after, before-and-after with comparison group, chi-squared tests, statistical modeling).

It is not surprising that estimates of the safety effect of cameras vary considerably. A summary of the more relevant study findings is provided in table 1, including a synopsis of the main difficulties.

From table 1, one could conclude that the bulk of the results support a conclusion that red-light cameras reduce right angle crashes and could increase rear end crashes; however, as the last column shows, most studies are tainted by methodological difficulties that raise questions about any conclusions from them. One difficulty, failure to account for regression to the mean, can exaggerate the positive effects, while another, ignoring possible spillover effects at intersections without RLC, will lead to an underestimation of RLC benefits, even more so if sites with these effects are used as a comparison group. ("Spillover effect" is the expected effect of RLCs on intersections other than the ones actually treated, resulting from jurisdiction-wide publicity and the general public's lack of knowledge of where RLCs are installed.) Almost all studies had one or the other of these flaws and many had both, in addition to other flaws.

Table 1. Summary of findings from past studies.

Reference	City	Camera sites	Comparison/ reference group	Crash type studied and estimated effects(negative indicates reduction)		Comment
Hillier, et al. (1993)⁽⁸⁾	Sydney, Australia	Installed at 16 intersections	16 signalized intersections	Right-angle and left-turn opposed	-50%	RTM^* possible; spillover may have affected comparison sites; results confounded by adjustment to signal timing in middle of study period
Hillier, et al. (1993)⁽⁸⁾	Sydney, Australia	Installed at 16 intersections	16 signalized intersections	Rear end	+25% to 60%
South, et al. (1988)⁽⁹⁾	Melbourne, Australia	Installed at 46 intersections	50 signalized intersections	No significant results. Looked at right angle, right-angle (turn), right against thru, rear end, rear end (turn), other, all crashes, number of casualties, no significant results		RTM^* possible, no accounting for changes in traffic volumes; comparison sites possibly affected by spillover and other treatments
Andreassen (1995)⁽¹⁰⁾	Victoria, Australia			No significant results		Lack of an effect could be that the sites studied tended to have few red-light-running related accidents; comparison sites may have been affected by spillover
Kent, et al. (1995)⁽¹¹⁾	Melbourne, Australia	3 intersection approaches at different intersections	Noncamera approaches	No significant relationship between the frequency of crashes at RLC and non-RLC sites and differences in red-light-running behavior		Cross-sectional design is problematic; likely spillover effects to the noncamera approaches at the same intersections
Mann, et al. (1994)⁽¹²⁾	Adelaide, Australia	Installed at 13 intersections	14 signalized intersections	Reductions at the camera sites were not statistically different from the reductions at the comparison sites		RTM^*and spillover to comparison sites are issues not addressed
London Accident Analysis Unit (1997)⁽¹³⁾	London, U.K.	RLC at 12 intersections and 21 speed cameras	Citywide effects examined	No significant results		The results are confounded because two programs are evaluated
Hooke, et al. (1996)⁽¹⁴⁾	Various cities in England and Wales	Installed at 78 intersections		All injury	-18%	A simple before-and-after comparison not controlling for effects of other factors, RTM^* and traffic volume changes; therefore there is limited confidence in the results.
Ng, et al. (1997)⁽¹⁵⁾	Singapore	Installed at 42 intersections	42 signalized intersections	All	-7%	RTM^*and spillover effects at comparison sites are issues
Ng, et al. (1997)⁽¹⁵⁾	Singapore	Installed at 42 intersections	42 signalized intersections	Right angle	-8%	RTM^*and spillover effects at comparison sites are issues
Retting and Kyrychenko (2001)⁽¹⁶⁾	Oxnard, CA	Installed at 11 intersections	Unsignalized intersections in Oxnard and signalized intersections in 3 similarly sized cities	All	-7%	Looked at citywide effects, not just at RLC sites29 months of before-and-after data used
				All injury	-29%
				Right angle	-32%
				Right-angle injury	-69%
				Rear end	+3% (nonsignificant)
SafeLight, Charlotte⁽¹⁷⁾	Charlotte, NC	Installed at 17 intersections	no comparison group	Angle-all approaches	-37%	Probable RTM in site selection
				Angle-camera approaches	-60%
				All-camera approaches	-19%
				Rear end-camera approaches	+4%
				All	< -1%
Maryland House of Delegates (2001)⁽¹⁸⁾	Howard County, MD	Installed at 25 intersections		Rear end	-32%	Probable RTM in site selection
				Right angle	-42%
				Other	-22%
Fleck and Smith (1998)⁽¹⁹⁾	San Francisco, CA	Installed at 6 intersections	Citywide effects examined	Citywide injury collisions caused by red-light violators; unclear how these were defined	- 9%	Question on definition of RLC crashes; did not examine specific effects at treated sites
Vinzant and Tatro (1999)⁽²⁰⁾	Mesa, AZ	6 intersections with RLC only, 6 intersections with RLC plus photo speed enforcement	6 signalized intersections	Total crash rates-crashes per million entering vehicles at each intersection		It is unclear if the assignment of treatment/no treatment to the four quadrants was random
				Combined-treatment quadrant	- 15.9%
				Photo-radar quadrant	- 7.5%
				RLC quadrant	- 9.7%
				Control quadrant	- 10.7%
Fox (1996)⁽²¹⁾	Glasgow, Scotland	Installed at 8 intersections and 3 "pelican" crossings	Area wide effects on injury crashes examined	Crossing carelessly	- 54.0%	RTM effects likelybecause the decreases in non-RLR crashes are greater than the RLR decreases at times, it is difficult to say what citywide effect the cameras have.
				Unsafe right turn	- 29.0%
				Failure to keep distance	+ 8.0%
				Other	- 29.0%
				All per month	- 32.0%
Winn (1995)⁽²²⁾	Glasgow, Scotland	6 locations on 1 approach	Various	Injury crashes related to RLR violations	- 62.0%	Probable RTM effects

^* RTM = Regression to the mean, also called "bias by selection."

A similar assessment of the literature was made independently in a recent meta-analysis, in which the review for the Insurance Institute for Highway Safety included most of the same studies cited in table 1 and some others.⁽²³⁾ That work found, expectedly, that largest safety benefits were reported by studies that did not control for regression to the mean and that small effects tend to be found where the possibility of spillover was ignored. The one study that measured both spillover and specific effects, while ensuring that regression to the mean was not a factor, was an evaluation of the Oxnard, California program by the Insurance Institute for Highway Safety.⁽¹⁶⁾ That study found a significant reduction in injury crashes overall but did not separate the specific effects at treatment sites from citywide effects. (It is understood that a follow up study is doing this.)

While it is difficult to make definitive conclusions from studies that generally fail the tests on the validity of the methodology, the results did provide some level of comfort for a decision to conduct a definitive large-scale study of U.S. installations. It was important, however, that the planned study capitalize on lessons learned from the strengths and weaknesses of the previous evaluations, many of which were conducted in an era when knowledge of potential pitfalls in evaluation studies and methods of avoiding or correcting them was not widespread. These lessons are reviewed next.

Lessons Learned and Issues Raised by Literature Search

From the literature review, a number of lessons were learned that were useful in designing a definitive U.S. study. Following is an itemization:

Number of treatment sites: This was limited in many studies, making for low significance of many results. A definitive study will have to pay careful attention to sample sizes.
RLC "spillover effects" in same city: Crashes could be affected at control/comparison sites, making it necessary to have such sites in other similar cities. This will also make it difficult to determine the effect at treated location versus all other locations in the same city, perhaps requiring a study design without a strong reliance on the use of comparison sites.
Differences in accident investigation and reporting practice between jurisdictions: This will make intracity comparisons or amalgamating data difficult and may require a separate analysis of the more severe crashes that are less likely to be affected by these differences if amalgamation is necessary to achieve large enough sample sizes.
Defining "red-light-running crashes": The lack of precise definition in past studies, the lack of clarity between angle and turning crashes on police forms, and the lack of information on "legal" right-turn-on-red crashes could cloud the definition of the outcome variable.
RLC effects on rear end crashes: There is clearly a need to consider not only this crash type in the analysis, but also to account for the tradeoff in crash severity between right angle and rear end types. To do so requires the use of the economic cost of crashes as an outcome variable.
Exposure changes between before-and-after periods: Exposure is the major determinant of intersection crashes. Therefore, it is important to account for any changes between the before and after period, particularly if these changes are triggered by the measure. All studies reviewed have failed to do this accounting, conveniently assuming that RLCs will not change exposure. It is also important to use proper methods for accounting for exposure changes because the conventional method of normalizing with crash rates (per unit of traffic volume) is a dubious one, given the non-linear relationship between crashes and exposure found in many cases.
Regression to the mean effects: In almost all studies reviewed, RLCs were installed at intersections with a high incidence of crashes, particularly those likely to be affected by RLC; this can lead to significant regression to the mean, particularly in the positive effects on RLC-targeted crashes, which must be accounted for.
Yellow interval improvements (and other intersection improvements) made at time of installation of RLCs: This makes it difficult to determine what caused the measured effect. It especially important to separate the effects of these other measures from that of RLC because some studies have shown that these other treatments can be just as effective as RLC. Thus the study should not resolve the difficulty of the confounding effects of other measures by avoiding affected sites, but should intentionally seek to include some of these sites if possible.
Disaggregate effects by signalization variables: There is little knowledge on the effect of variables such as cycle length and yellow and all-red interval combinations. Such knowledge will be useful in planning RLC programs or in explaining differential effects across sites and jurisdictions. The ability to seek this knowledge will be affected by the size of the sample and the variation in these factors.
Effect of signage: Warning the driver at a specific location may or may not change the effect, and may or may not limit the possible treatment effect on other intersections. Signing and enforcing one approach has been reported to have a benefit for all other approaches in some studies, but some studies have found otherwise. The issue will require further investigation, which will require that appropriate data be collected.
Public education level: It is desirable that this be specified and measured and its effects evaluated, in the light of other enforcement research that has shown the importance of combining enforcement with public information (PI) programs.
Type of ticketing: This affects true enforcement level and driver perception of "cost" of a ticket. The level of the fine and whether there are driver points for a violation can affect outcome, and the requirement to ticket the driver (rather than owner) can mean that only low percentage of offenders (i.e., 25 percent in San Francisco) are ticketed. This issue requires resolution, which will require that appropriate data be collected.
Definition of Red-Light violation: This could affect ticketing and public perception. Short or long "grace period" after signal turns red could have different effects. It would be desirable to isolate these differential effects, if possible.
Camera rotation: Determining if one can have a greater effect with the same number of cameras rotated to more sites is important. However, the optimum amount of rotation is undefined, so it might be desirable to develop this knowledge.
Relationship between changes in violations and changes in crashes: This will depend on many factors including the grace period chosen, driver versus owner, etc. To date, there is no knowledge on such a relationship. Establishing a link would be useful in that it would considerably simplify the task of evaluating RLC installations. This, however, will likely require a prospective study.

These "lessons learned" were then incorporated into the experimental designs for both the crash-frequency-based study and the economic analysis study covered in later sections of this report.

Previous | Table of Contents | Next

FHWA-HRT-05-048

TFHRC Home | FHWA Home | Feedback