Traffic Congestion and Safety: Mixed Effects on Total and Fatal Crashes

Abstract

This paper examines the effects of traffic congestion on total crashes, fatal or serious injury (FSI) crashes, and fatal-only crashes in peak periods using a zone-level safety analysis in Greater Melbourne. Bayesian mixed-effect negative binomial models are employed to investigate the relationship between a congestion index and the frequency of total and FSI crashes. In addition, Bayesian mixed-effect binary logistic models are adopted to explore the association between the congestion index and the likelihood of having fatal crashes in Statistical Area Level 2 (SA2) zones. Modelling results indicate that traffic congestion tends to increase total crashes in both the AM and PM peak periods and FSI crashes in the AM peak period. In contrast, traffic congestion tends to decrease the likelihood of having fatal crashes at both the AM and PM peaks. These findings suggest that many policies to reduce traffic congestion may also enhance road safety by lowering the overall number of crashes. However, it is crucial to incorporate careful speed management within these policies to reduce the risk of fatal crashes effectively.

1. Introduction

Traffic congestion and road crashes continue to pose substantial challenges to the sustainability of contemporary transport systems, leading to significant economic burdens. The economic impact of traffic congestion alone in Australia in 2016 was approximated to exceed $19 billion, with nearly 30% of these costs attributed to Melbourne [1]. Road crashes contribute to over 1100 fatalities and numerous hospitalisations annually, with an estimated yearly cost of $18 billion [2]. Together, the combined cost of traffic congestion and road crashes in Australia is projected to be around 2% of the country’s GDP.

Road safety and traffic congestion are not isolated issues but are deeply interconnected [3,4]. Many studies have explored the association between road safety and traffic congestion across various spatial scales, including cities, zones, and road segments. A review revealed that most studies reported either a positive or a U-shaped relationship between congestion and crashes [5]. However, the review also highlighted mixed results, suggesting that variations in the study location and the types of crash outcomes considered may explain the inconsistencies. While some studies indicate that higher traffic volumes increase crash risk through heightened exposure [5,6], others argue that congestion may reduce the severity of crashes due to the lower speeds typically observed in such conditions [7]. This underscores the importance of considering crash severity when evaluating the impact of traffic congestion on road safety.

At the road level, for example, using annual average daily traffic (AADT) as a proxy for traffic congestion, Kononov, Bailey, and Allery [6] investigated the impacts of traffic congestion on crashes on multilane freeways in California, Colorado, and Texas. They indicated that traffic congestion was positively associated with both total and severe crashes. When AADT increased to certain critical points, the slope of the crash prediction model’s coefficients became steeper, which implied that crash rates increased faster than traffic congestion. Hughes, Kaffine, and Kaffine [7] analysed the changes in crashes during the COVID-19 period. The results indicated that the reduction in vehicle miles travelled and congestion led to a decrease in total crashes. However, the rate of severe crashes increased as travel speed increased. Concerning the level of service (LOS) of roads, Dart and Mann [8] used a traffic volume ratio, which is calculated by dividing peak hourly traffic volume by service flow rate (maximum hourly flow at a certain LOS). This variable had a positive correlation with the total number of crashes.

Several other studies have, however, shown inconclusive results. Wang et al. [9] employed a congestion index that is calculated by subtracting free-flow travel time from actual travel time and then dividing it by the free-flow travel time. An analysis of crashes on the M25 motorway in London using this variable showed no significant impact on severe crashes and minor injury crashes. Using the average total delay, Quddus et al. [10] found no relationships between levels of traffic congestion and the crash severity level on the M25 motorway in London. Extending the research of Wang, Quddus, and Ison [9] by adding the M25 motorway’s surrounding roads, Wang et al. [11] found that an increased congestion level was related to more severe crashes, and the impact was smaller on minor injury crashes.

Some studies have used high-resolution flow data to explore the relationship between congestion and crashes [12]. Zheng [13] found that the likelihood of crashes might increase by a factor of six in a congested flow compared to a free-flow condition. This study used the traffic and crash data from a section of the I-880 freeway in California. Different states of traffic flow (free flow, transition, and congestion) were defined based on a wavelet transformation of a 1 min traffic data window of speed and occupancy (a proxy of density). Using the real-time traffic conditions at the crash events, Sun et al. [14] investigated the traffic flows and road design factors that might affect the total crashes, non-congested-flow crashes, and congested-flow crashes on urban expressways in Shanghai, China. A congestion index was used to capture the congestion conditions. The results showed that congestion might increase the risk of crashes on urban expressways. Stipancic et al. [15] matched GPS travel data and crash data with all links in the Quebec City road network to find the relationship between congestion (e.g., average speed and variation of speed) and crashes (frequency and severity). The results showed that during the afternoon (PM) peak period, the congestion index had a positive correlation with crash frequency. Dias et al. [16] found that when the congestion index and traffic density increased on eight radial routes of the metropolitan expressway network of Tokyo, the crash risk might be increased. High-resolution traffic data from several Tokyo’s expressways were also used by Hossain and Muromachi [17] to calculate a congestion index. They found that the difference in the level of congestion between downstream and upstream increases crash risk at ramp areas. In addition, a positive relationship between congestion (measured by time delay) and the number of crashes was found in ten cities in Latin America by utilising available big traffic data in 2019 [18].

Non-linear relationships between congestion and crashes at the road level have also been reported. For example, by analysing hourly traffic volume data on the highway system in France, Martin [19] observed a U-shaped correlation between traffic volume and the total number of crashes. Traffic volume under 400 veh/h was associated with peak crash rates, while the lowest crash rates occurred when the volume was between 1000 and 1500 veh/h. Similarly, Potts et al. [20] found a U-shaped relationship between traffic density and crash rates on freeway segments in several US metropolitan areas. This relationship was expected to occur in both recurrent and non-recurrent congestions.

Similar to the result for hourly traffic volume, U-shape relationships between volume-to-capacity ratios (v/c) and total crashes have been observed [21]. In other words, when v/c ratios were at their lowest, the crash rate values were highest. When the v/c ratio rose, the crash rates started to fall, then reversed to a rising trend as the increase in the v/c ratio passed a certain value. The relationship between v/c and crash rates for weekdays and weekend days, and for multivehicle, rear-end, and property-damage-only crashes also followed U-shaped models. However, crashes related to single-vehicle, fixed-object, turnover, and injury and fatality crashes were negatively correlated with the v/c ratio. In another study [22], a U-shaped relationship between the v/c ratio and serious injury crashes was observed. This study suggested that higher speeds, speed differentials, and driving conditions at night might explain the increase in crashes at a low v/c level. Moreover, the majority of crashes at lower v/c ratios were of the head-on and out-of-control types, while most of the crashes that happened at higher v/c ratios were of the rear-end and side-swipe varieties.

In contrast, using the natural log of the v/c ratio on two-lane highway segments in the US, Ivan et al. [23] found a negative relationship between the v/c ratio and the single-vehicle crash rate. Shefer [24] suggested a bell-shaped relationship between the volume-to-capacity ratio and fatal crashes only on the highway. The author reported that at the beginning, when traffic volume started to rise, high speed was still maintained, and hence, fatal crashes increased. The traffic density continued to increase until it led to speed reduction; as a result, the number of fatal crashes decreased. Similarly, Lord et al. [25] identified bell-shaped correlations linking both traffic density and v/c ratio to the total number of crashes on two highways in Canada.

Using peak hours and traffic density as measures for congestion, Shefer and Rietveld [26] found a reduction in fatalities in the morning (AM) peak period (7–9 am) and a bell-shaped relationship between traffic density and crash fatalities. However, this study did not find any significant relationship between the evening peak hours and crash fatalities. In a study looking at the I-880 freeway, Xu et al. [27] indicated that due to the low driving speeds, traffic congestion might be associated with less severe crashes. Pasidis [28] found evidence of a negative correlation between congestion (measured by the increase in travel time) and the likelihood of crashes by using 2012–2014 highway traffic and accident data in England.

Inconsistent findings have also been reported at the zone or city level. For example, Wang et al. [29] reported that total crashes in planning zones could increase with higher traffic density in Kunshan City. Stempfel et al. [30] used average travel speed as a proxy for congestion. Their study explored the relationship between average speeds and rear-end crashes on weekdays in Zurich. This relationship was considered at the whole-network level as well as at the link level. The results showed a negative relationship between average travel speed and the number of crashes at the network level. The study suggested that reducing congestion might help to improve road safety. Interestingly, the paper found a non-linear relationship, with the highest crash likelihood on links with medium speeds during the afternoon peak. In another study, traffic congestion was measured by the mean of extra travel time in proportion to the free-flow travel time, aggregated at the city level [31]. This study found a U-shaped relationship between traffic congestion and fatalities per capita in 129 European cities between 2008 and 2017.

In contrast, Hadayeghi et al. [32] found that v/c was negatively correlated with both the total number of crashes and the number of severe crashes in traffic zones in the city of Toronto, Ontario, Canada. In their study, the v/c for each zone was calculated by averaging the v/c values for all road sections in that zone. Noland and Quddus [33] used employment density as an indicator of congestion but found no effects on crash casualties in London’s districts. Xiao et al. [34] indicated a negative correlation between the congestion index (measured by the ratio of free speed to congested speed) and crash frequency, focusing on local government areas and traffic-density-based zones in inner Melbourne.

In summary, existing research has yielded mixed and conflicting findings regarding the association between road safety and traffic congestion, a variability that may be partly due to the challenges related to measuring congestion, selecting appropriate spatial units, and considering differing safety outcomes across various study locations. While a substantial body of research has focused on this relationship at the road-segment level [14,15,23,25], fewer studies have examined it at the broader zone level. Studies that have done so often report either no significant effects or counterintuitive results [32,33,34]. Moreover, there is a lack of evidence regarding the effects of traffic congestion on crash severity at the zone level. Given that traffic congestion can influence mode shifts and traffic diversion patterns, examining this relationship at the zone level is essential. A deeper understanding of this dynamic is critical for developing effective policies that address both traffic congestion and road safety, as interventions targeting one issue can have significant implications for the other. This research, therefore, aims to explore traffic congestion’s effects on three crash types (total crashes, fatal or serious injury (FSI), and fatal-only crashes) using a zonal analysis of Greater Melbourne. The three crash types were selected to capture the effects of congestion on crash severity.

2. Materials and Methods

This paper used data from Greater Melbourne, Victoria, Australia. The Statistical Area Level 2 (SA2), an intermediate statistical area with an average size of about 500 square kilometres, was selected as the zonal unit, providing a balanced aggregation level of zonal data. Traffic crash data, including crashes by severity (i.e., total, FSI, and fatal only) and time period (i.e., AM peak 7–9 am and PM peak 4–6 pm), in the span of five years (July 2015 to June 2020) were obtained through the CrashStats system. The analysis employed traffic congestion modelling data, which was obtained from Infrastructure Australia [1]. This dataset provided information about traffic volume and travel speed on the road network during the peak periods (AM and PM). Considering traffic, land use, and demographic variables is crucial to zone-level analysis, as previous research has shown significant safety effects of these variables [35,36]. Accordingly, the Australian Bureau of Statistics’ census data were employed to extract variables associated with demographics, land use, and journey-to-work data. In addition, road infrastructure variables were obtained from Data.Vic. A spatial analysis was conducted in the R programming environment [37] to compute data for each SA2 zone. Figure 1a,b depict the spatial variations of total and fatal crashes in Greater Melbourne, respectively.

This research measures traffic congestion levels through a congestion index, which is computed as the average free-flow speed (derived from speed limits) divided by the average congested speed (derived from estimated speeds for AM and PM peaks). As such, a higher congestion index indicates a higher traffic congestion level. A zone’s average speed was computed as the total distance travelled divided by total travel time, taking into account traffic volumes, speeds, and travel distances across road links within a zone. It should be noted that the average free-flow speed was computed with speed limits, while the average congested speed was computed using estimated speeds in the peak periods. Figure 2a,b demonstrate the spatial variations of the congestion index in the AM and PM peak periods across Greater Melbourne, respectively. As expected, the inner and middle zones tend to have higher levels of congestion.

Due to missing data, a few SA2 zones (e.g., airports) were excluded. Consequently, the cross-sectional analysis was based on a total of 300 SA2 zones in Greater Melbourne.

Table 1. Summary of main variables (n = 300).

Variables	Mean	Std. Dev.	Min	Max
No. of total crashes in AM peak period	19.063	14.481	1	116
No. of FSI crashes in AM peak period	6.39	5.277	0	43
No. of fatal crashes only in AM peak period	0.157	0.382	0	2
No. of total crashes in PM peak period	27.46	19.31	0	156
No. of FSI crashes in PM peak period	9.053	6.995	0	63
Number of fatal crashes only in PM peak period	0.257	0.495	0	2
Congestion index in AM peak period	1.782	0.465	1	4.569
Congestion index in PM peak period	1.614	0.338	1	3.528
Vehicle kilometres travelled—VKT (/day)	306,269.1	254,261.8	11,454.4	1,537,299
Population density (10,000 persons/square km)	0.211	0.178	0.0004	1.576
No. of intersections (in 100)	5.200	3.320	0.34	19.43
Active transport mode share	0.049	0.074	0.005	0.429
Proportion of industrial land use	0.063	0.123	0	0.664

presents a summary of variables considered in the final analysis. It can be observed that the mean of FSI crashes and total crashes in the AM and PM peak periods exceeds the variance, indicating overdispersion in the crash frequency data. Therefore, Bayesian mixed-effect negative binomial regression models were adopted to study the relationship between total/FSI crashes and traffic congestion, measured by the congestion index. Several traffic, land use, and demographic factors were controlled in these models. To further account for potential heterogeneity, especially considering the limited number of variables, the mixed-effect modelling approach was adopted.

It can also be seen that the numbers of fatal crashes in SA2 zones in the AM and PM peak periods were relatively low, ranging between zero and two and averaging 0.157 and 0.257, respectively. Accordingly, Bayesian mixed-effect binary logistic regression models were employed to examine the link between the congestion index and the likelihood of having fatal crashes or not in a zone.

The estimation of the Bayesian mixed-effect negative binomial and logistic regression models were performed in the R programming environment [37,38]. The variance inflation factors (VIFs) were computed for each model, confirming no evidence of multicollinearity. The selection of variables in the final analysis was performed using the Widely Applicable Information Criterion (WAIC), a commonly used model fit statistic for the Bayesian approach. A smaller WAIC value indicates a better model fit. To establish the direct association between traffic congestion and crashes, separate models for the two peak periods were estimated. Model estimations were conducted using Markov Chain Monte Carlo (MCMC) simulation based on 20,000 samples, excluding 20,000 burn-in samples. A convergence diagnostic showed that all models converged. The overall research framework is depicted in Figure 3.

3. Results

Table 2. Bayesian mixed-effect negative binomial regression models for total crashes in the AM and PM peak periods.

	AM Peak			PM Peak
Variables	Estimate	95% Bayesian CI		Estimate	95% Bayesian CI
Intercept	−5.586	−6.593	−4.594	−4.443	−5.349	−3.565
standard deviation	0.053	0.002	0.143	0.054	0.002	0.149
Log of VKT	0.586	0.506	0.667	0.538	0.469	0.611
Population density	1.452	1.042	1.866	1.347	0.972	1.724
No. of intersections	0.036	0.017	0.056	0.054	0.037	0.071
Active transport mode share	1.529	0.619	2.467	0.839	0.033	1.669
Proportion of industrial land use	0.621	0.219	1.032	0.900	0.522	1.277
Congestion index	0.292	0.165	0.42	0.187	0.034	0.341
Dispersion parameter	8.786	6.643	11.468	9.770	7.511	12.928
WAIC	1972.7			2149.9

presents the results of the Bayesian mixed-effect negative binomial models for total crashes in the peak periods. With the Bayesian approach, an estimate is significant if the 95% Bayesian confidence interval (CI) does not contain zero. Accordingly, all estimates were significant in the AM and PM models. The standard deviations of the random intercepts and dispersion parameters were significant, indicating the appropriateness of using mixed effect and negative binomial approaches, respectively. The effects of controlling variables were expected. For example, it was found that total crashes tend to increase with greater population density and VKT, a greater active transport mode share and proportion of industrial land use, and a greater number of intersections.

Table 2 highlights that a greater congestion index is correlated with more total crashes in the AM (a positive coefficient of 0.292, 95% Bayesian CI: 0.165 to 0.42) and the PM peak periods (a positive coefficient of 0.187, 95% Bayesian CI: 0.034 to 0.341). According to the incidence rate ratio (i.e., the exponent of the coefficient), if the congestion index increased by one, total crashes would increase 1.339-fold in the AM peak period and 1.205-fold in the PM peak period.

Table 3. Bayesian mixed-effect negative binomial regression models for FSI crashes in the AM and PM peak periods.

	AM Peak			PM Peak
Variables	Estimate	95% Bayesian CI		Estimate	95% Bayesian CI
Intercept	−6.347	−7.746	−4.970	−5.260	−6.457	−4.039
standard deviation	0.069	0.003	0.191	0.060	0.002	0.167
Log of VKT	0.576	0.465	0.688	0.538	0.442	0.632
Population density	1.383	0.835	1.944	1.278	0.800	1.762
No. of intersections	0.037	0.011	0.063	0.061	0.039	0.083
Active transport mode share	1.785	0.572	3.026	0.656	−0.421	1.733
Proportion of industrial land use	0.554	0.011	1.106	0.879	0.403	1.359
Congestion index	0.182	0.008	0.353	0.002	−0.201	0.209
Dispersion parameter	6.445	4.529	9.221	7.791	5.570	10.941
WAIC	1500.3			1638.2

presents the results of the Bayesian mixed effect negative binomial models for FSI crashes in the peak periods. Similar to the models for total crashes, the standard deviations of the random intercepts and dispersion parameters were significant. Controlling variables (population, VKT, number of intersections, and industrial land use) had similar effects, except for the active transport mode share in the PM model, which was not significant. For example, a one-unit increase in population density would result in 3.99-fold and 3.59-fold increases in FSI crashes in the AM and PM peak periods, respectively. The results showed that a greater congestion index was correlated with more FSI crashes in the AM peak period, evidenced by a positive coefficient of 0.182 (95% Bayesian CI: 0.008 to 0.353). More specifically, if the congestion index increased by one, FSI crashes in the AM peak would increase 1.2-fold. However, there was no evidence of an association between the congestion index and FSI crashes in the PM peak period.

The results of the Bayesian mixed-effect binary logistic models for fatal crashes versus no fatal crashes are summarised in

Table 4. Bayesian mixed effect binary logistic regression models for the occurrence or non-occurrence of fatal crashes in the AM and PM peak periods.

	AM Peak			PM Peak
Variables	Estimate	95% Bayesian CI		Estimate	95% Bayesian CI
Intercept	−13.629	−21.967	−6.02	−1.685	−7.947	4.359
standard deviation	0.496	0.018	1.332	0.556	0.025	1.481
Log of VKT	1.131	0.508	1.806	0.104	−0.37	0.596
Population density	1.570	−2.301	5.698	2.690	−0.105	5.798
No. of intersections	−0.001	−0.137	0.129	0.224	0.106	0.354
Active transport mode share	−7.468	−21.224	2.669	−2.576	−10.852	4.459
Proportion of industrial land use	1.796	−1.021	4.613	−0.833	−4.04	2.03
Congestion index	−1.554	−2.895	−0.353	−1.663	−3.092	−0.372
WAIC	232.2			300.0

. The standard deviations of the random intercepts in both models were significant. Among controlling variables, the results showed that the likelihood of having fatal crashes in an SA2 zone tends to increase with increasing VKT in the AM model and a higher number of intersections in the PM model. Interestingly, in both the AM and PM models, the congestion index was negatively correlated with the likelihood of having fatal crashes, with negative coefficients of −1.554 (95% Bayesian CI: −2.895 to −0.353) and −1.663 (95% Bayesian CI: −3.092 to −0.372). In other words, if the congestion index increased by one, the odds of having fatal crashes in the AM and PM peaks would be reduced by factors of 0.211 and 0.190, respectively.

4. Discussion

This paper has examined the impacts of traffic congestion on total crashes, fatal or serious injury (FSI) crashes, and fatal-only crashes in peak periods using a zone-level safety analysis in Greater Melbourne. More specifically, Bayesian mixed-effect negative binomial models were estimated to study the relationship between a congestion index and the numbers of total and FSI crashes, while Bayesian mixed-effect binary logistic models were developed to examine the relationship between the congestion index and the likelihood of having fatal crashes.

Regarding total crashes, traffic congestion tends to increase the number of total crashes in both the AM and PM peak periods. A plausible reason is that traffic congestion is correlated with more traffic activity and, consequently, more traffic conflicts, resulting in more total crashes [39]. Crashes may also be associated with an increase in risky driving behaviour when drivers get out of congested areas [40]. This positive relationship is in alignment with several previous studies at the zonal level, such as Wang, Liu, and Xu [29] and Stempfel, Guler, Menéndez, and Brucks [30]. In contrast, a recent macroscopic analysis [34] found that congestion tends to decrease total crashes. However, it is noted that their study focused on inner suburbs where traffic tends to be more congested and employed different zonal units.

Regarding FSI crashes, traffic congestion is positively correlated with the number of FSI crashes during the AM peak period. This effect, however, was not observed during the PM peak period. It should be noted that the AM peak period (7–9 AM) includes the morning school drop-off hours (8–9.30 AM), while the PM peak period (4–6 PM) is outside of the afternoon school pick-up hours (2.30–4 PM). A high level of congestion during the AM peak may by associated with school drop-off activities within school zones. In addition, many vulnerable road users (e.g., pedestrians and children) can often be present in the school zones during these hours. Collisions between vehicles and vulnerable road users are more likely than vehicle-to-vehicle crashes to result in severe injuries. This highlights the need for targeted safety measures during peak congestion hours in school zones.

Most importantly, substantial evidence showed that fatal crashes had a negative correlation with congestion. In other words, increased congestion levels were associated with a lower risk of fatal crashes in the AM and PM peak periods. Speed is often a major contributing factor in fatal crashes. Decreased speeds due to increasing congestion levels could reduce the impacts in the event of crashes, resulting in fewer fatal crashes [24,27]. More specifically, at lower speeds, the kinetic energy involved in a traffic crash is significantly lower. Previous research focusing on road links in Canada also reported that lower congestion levels were associated with fatalities [15]. Similarly, another study showed a negative link between traffic fatalities and congestion measured by traffic volume in the US counties [41].

As anticipated, higher exposures (i.e., VKT and population density) are associated with more total, FSI, and fatal crashes across both peak periods. The number of intersections is also a significant contributing factor, where an increase in the number of intersections tends to be associated with more total and FSI crashes. The presence of conflict points at intersections, particularly unsignalised intersections, can lead to an increased crash risk. Active transport mode share and proportion of industrial land use tend to be positively correlated with total and FSI crashes. These are logical findings since active transport mode share directly reflects the number of vulnerable road users, while industrial land use is associated with a high number of heavy vehicles and traffic movements. These results were consistent with previous studies [32,35,42].

5. Conclusions

The findings of this study underscore the complex relationship between traffic congestion and traffic safety, offering new insights into this dynamic, particularly from a zone-level perspective during peak hours and with a focus on crash types by severity. The results suggest that many policies aimed at reducing traffic congestion may also enhance road safety by lowering the overall number of crashes. However, it is crucial to incorporate careful speed management within these policies to effectively reduce the incidence of fatal crashes. Therefore, policymakers should consider potential unintended impacts on road safety when designing interventions to mitigate traffic congestion. Integrated approaches combining traffic congestion and safety management (e.g., real-time multi-objective traffic signal optimisation) are essential to ensure that reducing congestion does not inadvertently increase fatal crashes.

While this study examined crashes by severity, future work should explore the impact of traffic congestion on other types of crashes, such as rear-end and intersection collisions, as well as on different road user groups, including pedestrians and cyclists. Additionally, as highlighted by Retallack and Ostendorf [5], incorporating temporal data could help capture the non-linear effects of congestion on crash occurrences. Methodological improvements, such as the application of advanced techniques to address spatial autocorrelation and heterogeneity and spatial effects, would further enhance the robustness of future studies. In addition, future research should investigate the potential bidirectional relationship between traffic congestion and safety.