Interaction Effects of Intersection Safety Countermeasures: Evidence from 1,892 Accident-Prone Intersections Nationwide

1. Traffic Safety Countermeasures and Effect Analysis Method

2. Literature Review

1) Effectiveness analysis of traffic safety countermeasures

There are many previous studies analyzing the effectiveness of countermeasures, most of which have examined the reduction in accidents following the implementation of traffic safety countermeasures. <Table 1> summarizes previous studies on each of the 13 traffic safety countermeasures covered in this study, with four studies summarized for each countermeasure.

Table 1.

Summary of related studies on the effectiveness of traffic safety countermeasures

Installing traffic lights has been reported to reduce accidents by 6% to 47% (Jami, 2011; see <Table 1>). Installing traffic sign reduced accidents by 18% to 42% (Lee et al., 2015; see <Table 1>). Installing delineation devices reduced accidents by 9% to 55% (Park et al., 2006; see <Table 1>). Installing road lighting reduced accidents by 4% to 42% (Hovey and Chowdhury, 2005; see <Table 1>). Installing jaywalking prevention facilities reduced accidents by -1% (1% increase) to 77% (Baek, 2012; see <Table 1>). Skid-resistant pavement reduced accidents by 27% to 78% (Lee et al., 2011; see <Table 1>). Installing speed hump reduced accidents by 15% to 77% (Werner, 2015; see <Table 1>). Installing automated traffic enforcement system reduced accidents by 12% to 24% (Goldenbeld et al., 2019; see <Table 1>).

In addition, signal timing improvement has been reported to reduce accidents by 7% to 45% (Stevanovic et al., 2013; see <Table 1>). Lane reconfiguration reduced accidents by 16% to 29% (Zhou et al., 2022; see <Table 1>). Channelization reduced accidents by 5% to 34% (Choi et al., 2016; see <Table 1>). Installing sidewalk & crosswalk reduced accidents by 10% to 43% (Chen et al., 2013; see <Table 1>). Intersection geometry redesign reduced accidents by 27% to 65% (Elvik, 2017; see <Table 1>).

1. Selection of Analysis Targets

In this study, analysis targets were selected to precisely compare the effects and interactions of traffic safety countermeasures.

To this end, an expert survey was conducted targeting practitioners performing technical support tasks (7,106 cases in 2023) at the Korea Road Traffic Authority. Key countermeasures repeatedly found in technical support reports were derived as a primary candidate group (26 items) and then evaluated by 50 practitioners having at least five years of work experience.

Three evaluation criteria were set: practical applicability, policy significance, and accident reduction effectiveness. Practical applicability refers to the feasibility and ease of on-site implementation, policy significance refers to the relevance to national traffic safety policy goals, and accident reduction effectiveness refers to the contribution to reducing traffic accidents.

The relative importance of each criterion was assessed using the Analytic Hierarchy Process (AHP), and the priorities of candidate countermeasures were derived using the Best–Worst Method (BWM). AHP has strengths in estimating weights of the criteria, while BWM has strengths in ensuring consistency and efficiency in comparing alternatives. By combining the two techniques, a more reliable decision-making foundation was provided.

According to the AHP analysis results, the weight of the criteria was in the order of practical applicability (0.48), policy significance (0.35), and accident reduction effectiveness (0.17), and the consistency ratio of the responses was all below 0.10. The comprehensive score calculated by combining the BWM analysis and AHP weights is shown in <Table 2>, and this study selected 13 countermeasures corresponding to the top 50% of this comprehensive score as the final analysis targets (Figure 2).

Table 2.

Results of AHP and BWM analyses

The 13 countermeasures selected in this way are representative items frequently applied in actual fields, and they were used in the following chapter as basic data for interaction effect analysis based on the application status and combined implementation type at each intersection.

2. Analytical Method

This study employed a moderated regression model to precisely examine the effects of combined implementation of traffic safety countermeasures. This model goes beyond simple before-and-after evaluation and incorporates an interaction term, allowing for statistical identification of the combined effects between countermeasures. Therefore, this method is suitable for understanding the structural impact of combined implementation. However, the before-and-after evaluation with a comparison group faces challenges in securing a realistic comparison group, and the empirical Bayes evaluation suffers from the subjectivity of prior distribution settings and computational complexity. The moderated regression model addresses these shortcomings while also offering high practical applicability, making it suitable for the analytical purposes of this study.

Moderated regression analysis based on the ordinary least squares (OLS) method is a regression analysis method that uses a moderator variable (Hayes, 2018; Baron and Kenny, 1986; Helm and Mark, 2012). A moderator variable refers to a third variable that moderates the effect of the independent variable X on the dependent variable Y, as shown in Z in <Figure 3> (Hong and Jeong, 2014; Kwak, 2023).

Figure 2.

Prioritization results for traffic safety countermeasures

Figure 3.

Moderated regression conceptual model

The moderated regression model applied in this study is expressed as in Equation (4).

In this study, 13 countermeasures were each set as independent variables, and 12 countermeasures, excluding the identical countermeasures, were entered as moderator variables, resulting in a total of 306 moderation models. To mitigate multicollinearity among variables and enhance the clarity of regression coefficient interpretation, mean centering was applied to the independent variables and moderator variables (Goldstein, 2010; Kelley et al., 2017). This is a common procedure that facilitates the interpretation of interaction terms by using the mean-subtracted value of each variable.

In addition, since heteroscedasticity was confirmed in the error term analysis, the HC3 (heteroscedasticity-consistent) robust standard error estimate was applied to ensure the reliability of the regression coefficients (White, 1980). In R, the significance of the coefficients was tested using the vcovHC (model type=“HC3”) option.

3. Analytical Data

1) Overview of the Traffic Accident-Prone Area Improvement Project

The analytical data of the present study is based on the effect analysis data of the Accident-Prone Area Improvement Project. This project, which aims to improve road geometry and safety facilities to reduce recurring accidents at specific locations, has been conducted since the “Comprehensive Traffic Safety Plan” was established in 1987. Recently, a certain amount of design and construction has been implemented annually in accordance with the “Seventh Accident-Prone Area Improvement Project Plan (2022–2026)”.

The project is implemented in collaboration with the Office for Government Policy Coordination, the Ministry of Land, Infrastructure and Transport, the Ministry of the Interior and Safety, municipal governments, the National Police Agency, and the Korea Road Traffic Authority, and is conducted through the following procedures: accident data investigation and analysis, site selection, basic improvement plan development, improvement construction, and before-and-after effect evaluation. One cycle, from accident investigation to effect evaluation, takes approximately four years. For example, when the project is implemented based on accident data from 2024, the following steps are taken in the order of accident investigation in 2024, site selection and basic design in 2025, improvement construction in 2026, and effect evaluation in 2028 (accident statistics for 2027 will be compiled in 2028).

Accident-prone areas are selected based on a certain accident frequency occurring at the same location over the past year. As shown in <Figure 4>, the spatial boundary is 30 meters from the stop line at intersections and crosswalks, 200 meters for single-lane roads in urban road segments and 400 meters in rural road segments.

Figure 4.

Spatial scope of the accident-prone area improvement program (MOCT, 2002)

According to the cumulative effect analysis from 1991 to 2024, the improvement project exhibited a certain level of safety improvement, such as a decrease in the accident frequency by approximately 28.9%, a decrease in the number of deaths by 45.3%, and a decrease in the number of injuries by 29.2% (Korea Road Traffic Authority, 2024b).

2) Compiling of analytical data

The analytical data for this study was compiled based on the “Accident-Prone Area Basic Improvement Plan and Effect Analysis” report published by the Korea Road Traffic Authority, covering the past 10 years (2015 to 2024). Of the 2,313 intersections for which the effect analyses were conducted, intersections unrelated to the 13 countermeasures analyzed in this study were excluded, leaving a total of 1,892 intersections (three-way and four-way intersections) for final analysis.

Countermeasure information was consolidated and organized into similar items, referencing the improvement-specific symbols in the report. For example, traffic signs (A00, A10, A21) and delineation devices (B91, B96) were grouped into the same category for analysis. Furthermore, the scale of countermeasures applied at each intersection was calculated by dividing the scale by the total number of access ways to the intersection, based on a drawing of the intersection in each direction. For example, in a three-way intersection where skid-resistant pavement was applied in two directions, the scale was calculated as 0.67.

The dependent variables are the accident frequency and the accident severity. The accident frequency was obtained from the appendix of the report, and the accident severity was calculated using the unit cost of damage over the past 10 years presented in the Korea Road Traffic Authority's “Estimation and Assessment of Road Traffic Accident Costs” instead of the existing equivalent property damage only (EPDO) method (deaths×12, injuries×3). The weights were recalculated based on reported accidents (Table 3). As a result, the weights were 218.6 for deaths, 29.3 for serious injuries, 2.0 for minor injuries, and 1.0 for reported injuries.

Table 3.

Traffic accident costs by year and type of damage(unit: 10,000 KRW)

The regional distribution of the analysis data constructed through this process is shown in <Figure 5>, and among the 1,892 intersections, Seoul (11.2%) was the most common, followed by Gyeonggi (10.9%) and Gyeongbuk (8.0%). Furthermore, a total of 4,242 countermeasures were applied to the intersections analyzed, with an average of 2.2 countermeasures per intersection. While the number of countermeasures applied was high for traffic lights (740 locations) and skid-resistant pavement (505 locations), the number of countermeasures applied was relatively low for signal timing improvement (72 locations) (Table 4). The average scale of application by countermeasure was high for signal timing improvement (0.769) and intersection geometry redesign (0.724), while the average scale of application for the automated enforcement system (0.283) was the lowest.

Figure 5.

Distribution of intersections selected for analysis by region

Furthermore, to precisely analyze the interaction effect according to the combined implementation, the sample size for each combination of countermeasures needs to be clearly presented. In this study, the application of all binary combinations of the 13 countermeasures was investigated, and the number of intersections for each combined implementation type was calculated separately. The sample size for each combination refers to the number of actual observations used to estimate the interaction term (X×Z) in the analysis model, and is different from the number of single countermeasures applied in <Table 4>. The sample size for each combined implementation type is presented in the <Appendix> to ensure the transparency and reproducibility of the study. In addition, when a specific combination was implemented in a very small scale (N<10), caution was taken in presenting the results in consideration of the possibility of interpretation bias.

Table 4.

Status of traffic safety countermeasures implemented at analyzed intersections

1. Analysis Overview and Model Suitability Review Results

The moderated regression analysis model was constructed using each of the 13 traffic safety countermeasures as a baseline and includes pairwise interaction terms with the remaining 12 countermeasures. The dependent variables were accident frequency and accident severity. While 156 models (13×12) are theoretically possible for each dependent variable, 153 models were analyzed, excluding combinations that are not present in the data, such as delineation devices–intersection geometry redesign and intersection geometry redesign–road lighting.

A goodness-of-fit test revealed that 282 of the 306 models (92.2%) were statistically significant (139 accident counts and 143 accident severity). Furthermore, as shown in <Table 5>, the adjusted R² of the accident frequency models was generally higher than that of the accident severity models, confirming their relatively greater explanatory power for accident frequency.

Table 5.

R2 of a regression model

In this study, a “statistically significant model” was defined as one with an F-test significance level of less than 0.05. Meanwhile, considering the variability of actual traffic accident data, the omission of various external factors, and the structure of the regression model including interaction terms, the adjusted R² of 0.1 to 0.4 is within the range generally observed in existing traffic safety model studies. Therefore, this study utilized the adjusted R² as an indicator for comparison between models rather than as an absolute interpretation criterion, and policy discussions were centered on the F-test significance of the model and the statistical test results of the interaction terms.

2. Interpretation of Model Results

Analysis of the effects of single implementation of each countermeasure and the interaction effects of combined implementations revealed differences in the magnitude and direction of the effects across countermeasures. For most countermeasures, combined implementations showed either increased or decreased effects compared to single implementations, with statistical significance varying across countermeasures.

Traffic lights showed an accident reduction effect of approximately 7% in the case of single implementation, but this effect was somewhat reduced when combined with other countermeasures. Conversely, some combinations exhibited increased reductions when combined with an automated enforcement system or lane reconfiguration. Traffic signs showed a slightly lower average effect when implemented in combination, compared to single implementation, but their effect was rather enhanced when combined with delineation devices and sidewalks & crosswalks.

Overall, combined implementation of delineation devices, skid-resistant pavement, and lane reconfiguration tended to increase the effects of these countermeasures. Specifically, delineation devices demonstrated a greater effect in both accident frequency and accident severity compared to single implementation, with statistically significant increases observed in multiple combinations. Conversely, some measures, such as road lighting and speed hump, exhibited reduced effects in certain combinations.

Intersection geometry redesign demonstrated a high single implementation effect of approximately 20% or more on both accident frequency and accident severity, but its combined implementations reduced its effect in some combinations. However, combined implementation with traffic signs demonstrated a slightly increased accident frequency-reducing effect in some cases.

Meanwhile, some combinations exhibited asymmetric interaction effects. For example, when increasing the number of traffic lights was the primary strategy, installing an automated enforcement system enhanced the accident frequency-reducing effect. However, when an automated enforcement system was the primary strategy, the effects of increasing traffic lights were reduced. This inconsistent trend may be attributed to statistical and environmental factors, such as sample heterogeneity, multicollinearity between variables, and differences in implementation environments. Furthermore, even when the interaction term is statistically significant, the actual effect size may be minimal; therefore, interpretation requires considering both the effect size (%) and significance.

In summary, rather than categorically interpreting the direction of the effect of a specific combination, a cautious approach is required, taking into account data limitations and statistical variability. Further research conducted by controlling traffic environment characteristics is necessary for more robust conclusions. The final interaction analysis results are presented in <Table 6>.

Table 6.

Results of a moderated regression analysis by traffic safety countermeasures

The statistical test results for the interaction effect are presented in <Table 6>, but it is difficult to intuitively compare the relative magnitude and direction of the effects of each combination of countermeasures using the table format alone. Therefore, the interaction effect between accident frequency and accident severity was visualized for statistically significant combinations (p<0.05). <Figure 6> presents the interaction coefficients for the accident frequency-reducing effect and the accident severity-reducing effect, sorted by effect magnitude. The two figures clearly show whether a specific combination of countermeasures exhibits a synergistic effect (positive effect) or a reduced effect (negative effect) when implemented together, and complementarily visualize the interaction structure between countermeasures that would otherwise be difficult to grasp in a text-based interpretation.

Figure 6.

Significant interaction effects (p<0.05): accident frequency (left) and accident severity (right)

Meanwhile, the accident-prone area improvement project data used in this study compares the average accidents over the three years before and one year after the improvement; therefore, the presented accident reduction effect (%) is limited as being a relative indicator. Thus, for policy interpretation, supplementation is necessary to confirm the actual annual accident reduction rate. Furthermore, while it is internationally common to utilize active evaluation indicators that consider traffic volume, such as Annual Average Daily Traffic (AADT) and Million Vehicle Kilometers (MVK), the data used in this study did not include AADT and MVK, making them inapplicable. Future research should incorporate traffic environment variables such as traffic volume and geometry to conduct a more substantive analysis of the effects, including absolute reductions and traffic volume-adjusted indicators.

3. Results of Residual Diagnostics

In this study, the normality, homoscedasticity, and presence of extreme values in residuals, which are fundamental assumptions of regression analysis, were diagnosed for a total of 306 moderated regression models.

Normality was analyzed using Kline's (1998) skewness and kurtosis measures, commonly used in regression and structural equation models. All models met the criteria (skewness of 3 or less and kurtosis of 8 or less), demonstrating overall normality.

As mentioned in Chapter 3, homoscedasticity was tested by applying robust standard errors (Robust SE) during model building to address heteroscedasticity issues. The Breusch–Pagan test was then used to verify the criteria for all models, ensuring the stability of regression coefficient estimation and the reliability of interpretation.

Extreme values (standardized residuals exceeding 3) were within the standard range in all models, with the average extreme value rate being approximately 1.3%. This met the recommended criterion (less than 5% of the total sample), indicating no significant impact on model interpretation.

These results suggest that the moderated regression models constructed in this study generally meet statistical assumptions, and that interpretation of model estimates and policy discussions are reliable.

1. Policy Recommendations

Through the analytical results of this study, the effects of each traffic safety countermeasure applied alone were compared with the interaction effects when applied in combination with other countermeasures. This comparison showed whether positive or negative changes occurred in each combination of countermeasures, and effective implementation strategies for reducing accident frequency and severity were discussed.

For binary interactions of countermeasures, the difference between the single implementation effect and interaction effects of each combination was calculated and presented as a heat map, as shown in <Figure 7>. The change in effect in <Figure 7> indicates the magnitude of the interaction, which was calculated as "interaction effect – single implementation effect." In the heat map, positive interaction magnitudes are colored red (increased accident reduction effect), while negative ones are colored blue (decreased accident reduction effect).

Figure 7.

Heatmap of interaction effects between traffic safety countermeasures: interaction magnitude of accident frequency effects (left) and interaction magnitude of accident severity effects (right)

As shown in <Figure 7>, the magnitude of interaction varied depending on the combination of countermeasures. <Table 7> summarizes the top five and bottom five combinations of countermeasures by interaction effect magnitude, along with the resulting change in effect. These results confirm that some combinations of traffic safety countermeasures are relatively effective in reducing accident frequency and severity, while others are less effective.

Table 7.

Top/Bottom 5 traffic safety countermeasure combinations by interaction effect size

Based on the above analysis, three policy recommendations are proposed in this article for developing effective traffic safety improvement strategies.

First, a strategy that maximizes complementary effects should be adopted.

When signal timing improvement is applied as a primary countermeasure, the effect may be enhanced when combined with delineation devices or skid-resistant pavement. Furthermore, when an automated enforcement system is applied as a primary countermeasure, significant synergy effects have been observed when combined with traffic signs, sidewalks & crosswalks, and jaywalking prevention facilities. Therefore, it is advisable to give policy priorities to these combinations and apply them simultaneously.

Second, a non-combined strategy is needed to avoid mutual interference.

When a speed hump is used as a primary countermeasure, its effect may be diminished when combined with additional traffic signals. Furthermore, when road lighting is used as a primary countermeasure, its effect may also be diminished when used in combination with speed humps. Therefore, in these cases, combined implementations should be avoided, and a strategy that combines each countermeasure with other countermeasures is necessary.

Third, a customized combination strategy is needed according to safety improvement objectives.

The results in <Figure 7> demonstrate the effects of reducing accident frequency and accident severity. Therefore, goal-oriented combinations of countermeasures may be considered based on the accident characteristics at each location. For example, combining signal timing improvement with delineation devices or skid-resistant pavement can be expected to significantly reduce accident frequency. Similarly, combining automated enforcement system with jaywalking prevention facilities or sidewalks & crosswalks can be expected to significantly reduce accident severity.

Meanwhile, when discussing these policy implications, the statistical significance of the effects should be considered in addition to the magnitude of the regression coefficients. Statistically insignificant reduction effects cannot be ruled out because they may be the result of within-sample variation. Therefore, it is more appropriate to understand them as a reference-level tendency rather than as a basis for policy. Complementarily applying these perspectives can further enhance the reliability of policy recommendations and the consistency of interpretation based on the model results.

2. Conclusions

This study analyzed the interaction effects among 13 countermeasures, considering the characteristics of various countermeasures being applied in combination to traffic safety projects in Korea. Specifically, the accident frequency-reducing effect and the accident severity-reducing effect that occur when each countermeasure is combined with others were quantified.

To this end, data from accident-prone area improvement projects over the past 10 years were used to collect data on the scale of countermeasure application and accident data before and after improvement at 1,892 intersections. Moderated regression analysis models were used to identify single implementation effects and interaction effects.

The main results are as follows. First, 282 of the 306 models were statistically significant, confirming the presence of interaction effects by combined implementation. Second, the explanatory power of the accident frequency model was higher than that of the accident severity model, confirming that the influence of the combination of countermeasures was relatively stronger on accident frequency. Third, there were numerous cases where the combined implementation enhanced or weakened the effects compared to single implementations. On average, delineation devices, automated enforcement systems, signal timing improvement, and sidewalks & crosswalks strengthened both the accident frequency-reducing effects and the accident severity-reducing effect, whereas traffic lights, traffic signs, jaywalking prevention facilities, speed humps, and intersection geometry redesign tended to weaken the effects.

Furthermore, the accident reduction effect derived from this study was lower than in previous studies. This is interpreted as a result of the precise examination of the scale of countermeasures applied at each intersection access way. This allowed for a comparison of the relative effect of the 13 countermeasures frequently applied to traffic safety projects in Korea. Furthermore, the analysis newly identified the interaction effects of combined applications, which were difficult to identify in previous studies.

However, since this study focused on interrupted flow sections centered on intersections, the results may be influenced by the physical and traffic characteristics of the target intersections (e.g., urban/rural ratio, intersection shape, countermeasure application patterns). Therefore, direct generalization of these results to uninterrupted flow sections or other diverse traffic environments is limited. Future studies may need to be conducted on expanded analyses that incorporate diverse regions and traffic flow characteristics.

Furthermore, the negative interaction effects observed in some combinations may be due to complex factors, including design conflicts between countermeasures, increased driver cognitive load, and desensitization due to repetition and overlapping. Considering these factors, the combined application of countermeasures should be carefully carried out, taking into account not only the simple summation of countermeasures but also the implementation environment, driving characteristics, and driver responses. Future research should more precisely propose policy and design improvements, such as design methods that minimize negative interactions and facility layout principles that reduce driver cognitive load.

This study analyzed the single implementation effects of various traffic safety countermeasures and the interaction effects of their combined implementation, suggesting improvement strategies. However, the present study has the following limitations.

First, the analysis method relies on a simple before-and-after comparison, making it difficult to completely eliminate the problem of regression to the mean. Future research should address this limitation through sophisticated analyses utilizing multi-year time-series data.

Second, traffic accidents are complexly influenced by various external factors, including weather conditions, driver behavior, time of day, land use characteristics, and road congestion. However, data limitations prevented this from being reflected in the regression model of this study. Future research should employ multi-faceted analyses that consider traffic environment variables such as driver characteristics and spatial factors to precisely identify the causes of differences in the effect among countermeasures.

Third, due to limitations in accident data and facility application history, important traffic environment variables such as road geometry, traffic volume (AADT), and traffic flow type (interrupted/uninterrupted) were not included in the model. These variables may alter the magnitude and direction of the effects of countermeasures. Therefore, future research should systematically verify the combined implementation effects through interaction analyses of traffic environment variables or through subdivided models based on traffic characteristics.

Finally, this study focused on an exploratory analysis of the combined implementation effects of countermeasures, suggesting combination strategies based on the magnitude and statistical significance of interaction effects. Future use of artificial intelligence-based predictive models or nonlinear models to quantify the differences in the effect across countermeasure combinations will increase their practical applicability, such as in packaged road safety project planning.