Use your PIARC access.> Create an account> Forgot your password?
Road Safety Manual
A manual for practitioners and decision makers
on implementing safe system infrastructure!
Effective decisions regarding the use of road safety infrastructure interventions can only be made with adequate knowledge regarding its effectiveness. The key measure that is used for this assessment is the expected reduction in crashes, expressed as a crash modification factor (CMF) for the intervention. This indicates the degree to which the intervention is expected to reduce crashes. CMFs are also described in detail in Effective Safe System Interventions in Intervention Options And Selection. CMFs help policy-makers in the decision-making process, and the results from previous studies are typically used to provide a realistic estimate of the expected effect (SafetyCube, CMF Clearinghouse). However, there are many gaps in knowledge regarding the effectiveness of interventions.
An international collaboration on effective evaluation for road safety infrastructure interventions has recently been completed. The OECD/ITF (2012) document Sharing Road Safety provided the following key messages:
Some of the key recommendations from this work were that:
The ultimate measure of the success of a road safety policy or intervention is the effect that it has had on crash reduction, particularly the reduction in fatal and serious injuries. Unfortunately, difficulty arises in considering crashes on their own, as it may be necessary to wait several years after the countermeasure or package of measures has been introduced to be able to validate the changes in crash statistics. Given an indication of effectiveness is often required in a shorter timeframe, particularly to determine that nothing has gone wrong, more immediate feedback may be required. Proxy measures for safety might be useful to monitor the effectiveness of schemes (see also Identifying Data Requirements) for the link between intermediate and final outcome measures). These proxy measures are usually observational-type measurements. It is often recommended to conduct an evaluation in two stages: a short-term phase (e.g. 6 months) and a longer-term phase (e.g. 3 to 5 years).
Where practical, the treated site, route or area should be observed immediately after completion of the construction work; with regular visits made in the following days, weeks or months after completion until the team is satisfied that the scheme is operating as expected.
It is recommended that any earlier behavioural measurements made at the investigation stage be repeated later (e.g. traffic conflict counts, speed measurements), as this will help determine/support the need for any further changes, or may in fact prove the success of the intervention. This behavioural study is also required because some features of an intervention may, for instance, produce an unforeseen reaction in road users and subsequently create a potentially hazardous situation. Monitoring and analysis should highlight this problem at an early stage so that appropriate action can be taken quickly to remove this hazard.
At best, it may be possible to alleviate this hazard easily, e.g. by realigning the kerb lines to prevent a hazardous manoeuvre. At worst, it could lead to the complete withdrawal of a scheme and the need to reassess alternative schemes.
Some of the issues that may be relevant to assess include:
It would be impractical to carry out detailed behavioural studies for all minor alterations, but such studies may be particularly important for expensive schemes such as area-wide or mass action treatments.
It may be preferable to allow the scheme to operate for about two months before conducting a behavioural ‘after’ study. This should serve as a ‘settling-in’ period, during which regular users get used to a new road feature and any learning effects have largely disappeared. The evaluation should verify the impacts in terms of safety for all users. The overall effects of a project should be understood as a whole. For example, the implementation of a bike lane can be dangerous if used by motorcyclists.
The following six case study from Malaysia, Puerto Rico, Portugal, Slovakia and Italy provides a number of excellent evaluations on the safety effects of various treatment.
The case study describes a set of safety countermeasures in via Pistoiese (Firenze – Italy), that is an urban collector road classified as a high accident concentration section. A detailed safety study has been undertaken to identify the possible applicable safety countermeasures. Accident analysis, road safety inspections and driving simulation studies were performed. The intervention to be implemented combines physical and perceptual treatments. Read more (PDF, 869 kb).
The most important form of evaluation of any safety measure is determining its effect on crashes, and whether the intervention has reduced the number of crashes (particularly fatal and serious injury ones) by the expected amount. Evaluation involves the analysis of crash data by comparing the safety performance before a change was made to the safety performance afterwards. This often involves statistical analysis. To be reasonably sure that the random nature of crashes has been taken into account, it will normally be necessary to wait for several years for a valid result to be available. As mentioned earlier, more immediate feedback is often necessary and sometimes a shorter-term monitoring method may be applied.
Apart from the time-scale issue, there are other factors that complicate the (apparently straightforward) process of assessing the effectiveness of crash changes at treated sites, routes or areas. The main ones to consider are:
Each of these is discussed below.
This effect complicates evaluations at high-crash locations. Crash locations for treatment are typically chosen because they were the scene of numerous crashes. In some cases, the high crash numbers may be the result of just one particularly bad year, and this may have happened purely by chance. Crash numbers at such sites will tend to fall in the next year even if no intervention is applied. Even if a three-year period is used, crash occurrence may be due to random fluctuations, and in subsequent years these sites will experience lower numbers. This is known as regression to the mean. It is believed that the regression to the mean effect can overstate the effectiveness of a treatment by 5% to 30%. Figure 1 shows a graphical representation of regression to the mean.
Mountain, Maher and Fawaz (1998) provide a clear example of regression to the mean that can result in overestimation of intervention effectiveness. In a large study covering over 900 sites and 13 612 crashes, they compared crashes occurring during the site identification ‘before’ period, the lag period (the period between a site being selected for treatment and the actual installation of the treatment), and the ‘after’ period. The crash numbers were adjusted to take the different period lengths into account, and the data is summarised in Table 12.1. The difference between the number of crashes occurring in the lag period and the number of crashes occurring in the before period was taken as a measure of the regression to the mean effect.
|Treatment||Biased estimate of reduction due to the treatment (%)||Unbiased estimate of the treatment effect (%)||Estimated reduction due to regression to the mean (%)|
Surface treatments – wet road crashes only
Other junction improvements
Other link improvements
Street lighting schemes
Traffic management schemes
All treatment types
Source: Adapted from Mountain, Maher and Fawaz (1998).
The most robust way of accounting for both the regression to the mean effect and changes in the environment may be through the use of control sites that have been chosen in exactly the same way as the treated sites and have been identified as having similar problems, but have been left untreated. The choice as to whether a site is treated or is left untreated is based on a random allocation. This type of controlled experiment (or randomised control trial) is rare in road safety infrastructure projects as it is difficult to justify not treating a site that has been identified as high risk.
In recent years, the Empirical Bayes approach has emerged as an effective way of minimising the impact of regression to the mean. Although the Empirical Bayes approach is recognised as good practice in road safety evaluation, not all countries (especially those in LMICs) will have experience in using this technique. If not using this recommended approach, at the very least, a before-and-after analysis using comparison sites should be undertaken, with at least three (but preferably five) years of data collected for the before period. This is because the regression to the mean effect does tend to diminish if considered over longer periods of time. For example, in a study conducted in two counties in the UK, Abbess et al. (1981) calculated that regression to the mean had the following effects:
Where the Empirical Bayes approach has not been used, the above allowances could be made when calculating the reduction in crashes produced by the countermeasures.
When evaluating changes in crashes (or for most of the monitoring measures described in Observation in 12.4 Evaluating Road Safety Infrastructure Interventions), other factors not affected by the treatment might still influence that measure, and therefore will have to be taken into account. Examples include a change in the speed limit on roads that include the treated crash site; national or local road safety campaigns; or traffic management schemes that might affect traffic volume (e.g. closure of an intersection near the site, producing a marked change in traffic patterns). Changes related to external factors may be compensated for by comparing the site under study, for the same before and after periods, with comparison sites (sometimes termed ‘control’ sites) that have not been treated. In order for this data to be valid, it is important that these other sites experience exactly the same changes as the site under evaluation.
Comparison data can be collected either by matched pairs or area controls. A matched pair control site involves finding a site that is geographically close to the treated site (but not close enough to be affected by any traffic diversion), and has similar general characteristics. This is so that the control site will be subject to the same local variations that might affect safety (e.g. weather, traffic flows, safety campaigns, etc.).
In practice it may be difficult to find other sites with similar safety problems that will be left untreated purely for the sake of statistical analysis. Area comparisons comprising a large number of sites are, therefore, frequently used.
Comparison group sites should be as similar as possible to the treated sites and they should not be affected by the treatment.
There is still some controversy over whether or not the ‘crash migration’ effect exists. It has been found that crashes tend to increase at sites adjoining a successfully treated site, producing an apparent crash transfer or ‘migration’. Why this effect occurs is unclear, but one hypothesis is that drivers are ‘compensating’ for the improved safety at treated sites by being less cautious elsewhere.
Obviously, to detect such an occurrence, the crash frequencies in the surrounding area of the treated sites before and after treatment need to be compared with a suitable comparison group. In other words, the area of an evaluation study needs to be expanded to include routes that may be impacted by the project, and comparison sites must be identified that have not been affected.
However, there are no established techniques yet available to estimate such an effect for a particular site. The first reported occurrence of this feature (Boyle & Wright, 1984) found an overall crash increase in the surrounding areas of about 9%, and a later study (Persaud, 1987) of a larger number of sites estimated the increase to be 0.2 crashes per site per year.
A study by Austroads (2010) identified the effects that certain interventions have on the redistribution of traffic, and suggest that this may be a cause of migration. If traffic is reduced at the treated location, it is likely that it will be increased at a nearby location (this is likely in all situations, except the rare circumstance where trip numbers are reduced). It is suggested that this redistribution is more likely for certain types of treatments. For these interventions, evaluation of the effects should include a broader geographic area to capture locations where exposure (and therefore risk) may be increased. Interventions identified as potentially causing such an effect included:
There are several key documents that provide detailed accounts of evaluation methods for road safety infrastructure. These include:
One or more of these documents should be consulted for a full account of how to conduct a road safety evaluation.
The ‘gold standard’ in evaluation methodology is a ‘controlled experiment’ or randomised control trial (RCT). As indicated earlier, this approach is very rare in evaluation of infrastructure interventions. This is mainly due to concerns with leaving high risk sites untreated, but is also due to issues such as lack of knowledge regarding this approach. By randomly allocating sites to a treatment or control group, any biases which arise from treating sites with the worst crash history should be eliminated. External factors which come into play as the evaluation proceeds (such as unforeseen enforcement programmes in the area of the trials) are also catered for, as the external factors can be assumed to affect the treatment and control sites equally. The difference between the treatment and control sites in the after period is a true reflection of the influence of the treatment.
The currently recommended approach for evaluation of infrastructure interventions is the Empirical Bayes (or EB) method. Hauer (1997) explains this procedure in its simplest form, which is recommended reading for those wishing to understand the logic behind this approach as it applies to road safety. T
The approach uses the concept of an ‘expected’ number of crashes, or the long-term average, calculated over as long a period as is considered useful. The second concept is the ‘reference population’, or a set of similar sites or routes for which data is available (e.g. all intersections of a certain type in a network). The reference population acts like a comparison group. In the classical version of EB, the number of crashes which would be expected at a treatment site if no intervention had taken place is estimated and compared with the number of crashes that actually occurred. The comparison of the actual number of crashes with the expected number of crashes indicates the extent of the CMF.
A commonly applied method for the evaluation of infrastructure safety is the before-and-after study utilising a comparison group. Even though this approach does not fully address the issue of regression to the mean (although as indicated above, using a longer ‘before’ time period can reduce this effect), it does limit the impact of external factors. The approach compares the outcomes at the treatment sites with the outcomes at a set of comparison sites (sometimes termed ‘control’ sites), which have similar characteristics to the treatment sites in all important aspects, except that the treatment is not installed. This approach assumes that external factors act on both the treatment sites and the comparison sites in an identical way, and so can be measured and allowed for. Since the treatment sites and the comparison sites are subject to the same sets of external variables, any difference in safety outcomes must be due to the treatment.
The most basic form of evaluation (and one that has often been applied) is to simply compare crashes in the period before an intervention is installed with the crashes after (termed a simple or naïve before-and-after study – i.e. without a comparison group). This approach is not recommended, as it does not adequately account for regression to the mean or external variables.
Cross-sectional studies have also been used to try and identify the effects of safety interventions. These studies compare safety performance of sites with a particular safety feature (or features) with sites that do not have the same feature. It is assumed that the difference in safety performance is due to this feature. There are many issues with using this approach (particularly differences between sites other than the feature of interest; and differential crash rates that may have led to the installation of the feature in the first place), therefore it should not be used for this type of evaluation (however, if it is used, the limitations should be well-understood and documented).
In evaluating a particular treatment, the answers to the following questions will usually be required:
The rare and random nature of road crashes can lead to fairly large fluctuations in crash frequencies at a site from year to year, even though there has been no change in the underlying safety level. This extra variability makes the effect of the treatment more difficult to detect, but a test of statistical significance can be used to determine whether the observed change in crashes is likely to have occurred by chance or not.
The main problem with using crash data for evaluation (even assuming high recording accuracy) is distinguishing between changes due to the treatment and changes due to other sources. As explained earlier, even if the selected sites are good comparison groups that take into account the environmental influences, there are other confounding factors that need to be considered.
There are a number of points to take into account when choosing time periods used to compare crashes occurring before and after treatment; these include:
As identified throughout this chapter, there are a number of documents that provide a detailed guide to evaluating intervention effectiveness. There are also various tools to help with this task. Box 12.2 provides a summary of the Countermeasure Evaluation Tool from SafetyAnalyst.
Developed by the American Association of State Highway and Transportation Officials (AASHTO), the Countermeasure evaluation tool is part of the SafetyAnalyst suite of tools. It provides step-by-step guidance for conducting evaluation of the benefits of safety interventions using the Empirical Bayes approach. Therefore, it is able to account for the effect of regression to the mean (see above). As well as detecting the overall change in crashes (the CMF), the tool can also identify the effect on specific target crash types. For further details see Harwood et al. (2010) or http://www.safetyanalyst.org/evaltool.htm.
Source: Harwood et al. (2010).