Enhancing Safety Measures at Stop-Controlled Intersections: A Study on LED Backlit Signs and Drivers’ Behavior in Montréal, Québec

Abstract

This study evaluates the safety impacts of upgrading traditional STOP signs to light-emitting diode (LED)-illuminated backlit STOP signs at urban intersections, aiming to address visibility and conspicuity concerns that affect driver behavior and intersection safety. STOP signs are critical for regulating traffic flow and minimizing conflicts, yet their effectiveness can diminish under low-visibility conditions. To assess the effectiveness of LED-enhanced signage, a before–after study was conducted using surrogate safety measures. Key performance indicators included vehicle speeds, driver compliance rates, and vehicle-to-vehicle interactions, recorded both prior to and following LED implementation. A multinomial logistic regression model was used to analyze driver behaviors, and a calibrated microscopic simulation model, optimized using a genetic algorithm (GA), was applied to estimate traffic conflict frequencies. Video data were processed to extract driver trajectories and reactions under varying signage conditions. Results showed LED STOP signs improved compliance rates from 60% to 85%, reduced average vehicle speeds by 25%, and increased post-encroachment times. Conflict analysis revealed significant reductions in vehicle-to-vehicle and pedestrian conflicts, particularly at night. These findings highlight the effectiveness of LED signage in enhancing intersection safety and offer important implications for urban traffic management and the adoption of advanced traffic control technologies.

1. Introduction

STOP signs and other traffic control devices such as traffic lights are among the fundamental components of urban traffic management in North America, which are installed and used under the well-established guidelines and warrants provided by federal, state, and provincial governments. In Canada, these guidelines are detailed in the Manual on Uniform Traffic Control Devices for Canada (MUTCDC) [1] and Tome V: TCDs [2] maintained by the Transportation Association of Canada (TAC) and Ministry of Transportation, Quebec (MTQ). These warrants provide criteria for determining the appropriate stop control devices at intersections considering factors such as vehicle volumes, speeds, delays, visibility, and safety concerns.

Intersections are among the most critical components of road networks and are often the most susceptible points for crashes, representing approximately 45% of all reported accidents [3]. Although night-time driving accounts for only about one-quarter of total vehicle travel, collisions during this period represent nearly 55% of all accidents, highlighting a considerable safety risk [4]. Among different crossing types with varying control measures, those regulated by STOP signs (SCIs) exhibit the highest crash frequencies, frequently resulting in severe or fatal injuries [4]. In Canada, 1739 individuals lost their lives at SCI during 2008–2017 [5].

Enhanced signage such as LED illuminated STOP signs emerged as a promising safety countermeasure to improve visibility and conspicuity at high-risk intersections [6]. Based on the hypotheses, a new sign with even improved conspicuity features may reduce accidents. The Strategic Highway Safety Plans (SHSP) recommend that the effectiveness of any new roadway safety intervention or device be systematically evaluated [7]. Research examining the safety performance of recently implemented regulatory signs, without depending solely on crash records, remains limited. Instead, some studies emphasize surrogate safety indicators or other directly observable traffic behaviors. This methodology enables earlier assessment of safety impacts and benefits from computational modeling to replicate real-world traffic scenarios, given that complete crash data is scarce due to the infrequency of such incidents. Constraints in the accuracy and availability of historical datasets have further motivated the use of surrogate indicators [8,9]. Additionally, the extended time required to collect adequate accident data for before–after evaluations poses another notable challenge. A sufficient duration of study is required to observe meaningful changes in safety outcomes [10].

This study seeks to address the safety impacts of LED STOP signs at urban SCI in Montréal, Québec, applying a before–after observational approach. In addition, we also aim to review driver compliance, vehicle speeds, and conflict rates before and after installing LED-enhanced signage by leveraging a comprehensive experimental design. Statistical analyses including multinomial logistic regression and generalized linear models (GLMs) are employed to assess the performance of LED signs while accounting for environmental, traffic, and intersection geometry factors. The present study contributes to the growing body of evidence supporting LED STOP signs as an effective countermeasure for high-risk intersections. The results assist transportation agencies and policymakers in refining guidelines for enhanced signage and advancing road safety strategies. While prior research has demonstrated that LED signage improves visibility and compliance, our study advances this knowledge by employing a hybrid methodology that combines field data with calibrated microsimulation models using surrogate safety indicators (time to collision (TTC), post-encroachment time (PET). We also compare multiple treatments, including backlit LED stop (BLS) signage, which has not been evaluated using such integrated tools in an urban setting.

The remainder of this paper is structured as follows: Section 2 reviews existing literature on stop-controlled intersections and surrogate safety measures. Section 3 describes the study methodology, including the field experiment, simulation modeling, and statistical analyses. Section 4 presents the results related to speed, compliance, and conflict frequency. Section 5 discusses key findings and concludes with recommendations for future research and policy implementation.

2. Literature Review

SCIs are among the critical components of transportation infrastructure to regulate traffic flow and enhance safety [3,5]. Despite their widespread implementation, these intersections are often related to safety challenges due to issues such as driver non-compliance and limited visibility. Such shortcomings highlight the necessity for effective traffic management measures to mitigate potential risks.

Traffic signs, which are regulated in the US by the Federal Highway Administration (FHWA) and are widely adopted internationally, are among the fundamental instruments for traffic control and safety enhancement. These signs help reduce accident frequency and severity by enabling drivers to detect, interpret, and act appropriately based on their properties and placement [11]. Their effectiveness underscores the need for further investigation to optimize their design and implementation, especially in the context of SCI.

2.1. Safety Studies on SCIs

Studies on safety measures at SCIs have evolved significantly due to the high proportion of accidents created by non-compliance, distraction, and inadequate visibility. The studies conducted to enhance safety at SCIs emphasize innovative signage and driver behavior analysis.

Intersection crashes are frequently attributed to driver errors and visibility challenges. Navarro et al. indicated that issues such as obstructed sight lines, poor lighting, and limited visibility significantly contribute to collision risks at SCIs. It was observed that upgrading intersections from minor-approach-only stops (MAS) to all-way stops (AWS) enhanced safety outcomes by lowering vehicle speeds and improving driver compliance [12]. In addition, Tucker et al. argued that STOP signs often become obscured by factors such as overgrown vegetation, poor placement, and adverse weather, leading to non-compliance and higher crash risks [13].

The advent of LED-enhanced signage has provided new opportunities to assess the aforementioned challenges. Multiple studies have examined how flashing LED signs influence driver behavior, reporting higher compliance levels and a reduction in roll-through maneuvers [14,15,16,17]. The use of actuated LED warning systems has been evaluated both on STOP signs and on speed alert devices [18]. Further, Shaaban and Wall claimed that intersections equipped with LED-enhanced signs experienced fewer rolling and more complete stops, highlighting the significance of dynamic visual cues [19]. Furthermore, Gates et al. reported that LED signs improve conspicuity and compliance, especially in low-light conditions, resulting in increasing their efficiency in reducing accidents [9]. An assessment of overhead and pole-mounted flashing red beacons at 100 multi-way SCIs in North and South Carolina showed reductions of (13.3 ± 4.6%) in angle collisions and (10.2 ± 4.9%) in injury-related crashes [20]. In a subsequent study conducted in North Carolina, the installation of overhead flashing beacons at rural intersections was associated with an average crash reduction of 12% (±6%) [21].

Other innovations include LED-backlit signs powered by renewable energy. In another study, Weidemann et al. investigated such systems at rural intersections with poor visibility created by challenging geometric constraints. These systems actively detected approaching vehicles and warned drivers with LED blinker signs, resulting in improving safety and compliance [16]. Patella et al. asserted that illuminated crosswalks equipped with LED systems reduced vehicle speeds even in the absence of pedestrians, indicating the potential of active lighting systems to enhance safety [22].

Material improvements were explored as a cost-effective solution. Such enhancements are expected to enhance safety and reduce accidents [23]. In addition, Gates et al. emphasized the utility of retroreflective and fluorescent sheeting to improve sign conspicuity during day and night [9]. Further, Fiolić et al. highlighted the role of high-contrast materials in reducing driver cognitive load, enabling quicker decision-making, and enhancing safety outcomes at critical junctions [24]. Furthermore, Portera et al. studied LED-enhanced road markings, indicating that such systems help drivers maintain safer trajectories and better lateral control, especially at night [25].

Nighttime conditions pose unique challenges for SCIs, with limited ambient lighting affecting drivers’ ability to perceive signs and react appropriately. In another study, Schnell et al. emphasized the significance of maintaining optimal luminance levels for nighttime sign legibility, recommending standards which ensure visibility even for older drivers [11]. Rista and Fitzpatrick compared LED-embedded signs, pedestrian hybrid beacons, and rectangular rapid flashing beacons, indicating that LED systems significantly enhance nighttime driver yielding rates [26].

In another investigation [27], SCIs were compared with uncontrolled approaches using video observations from 100 intersections encompassing 130,000 traffic events. The findings indicated that STOP signs substantially lowered approach speeds and reduced speed variability, although no statistically significant reduction in vehicle–pedestrian interactions was observed. The study suggested considering alternative interventions, as STOP signs alone may offer limited benefits for pedestrian protection. Two surrogate safety indicators, TTC and PET, were employed, along with the analysis of multiple covariates. Vehicle-vehicle interactions showed improvement in TTC and PET. However, there were a few significant variables which limited conclusions for pedestrian safety [27].

Recent studies have continued to explore the role of lighting and signage at SCIs using advanced experimental and analytical tools. Kunnah and Hassan [28] used a high-fidelity driving simulator to study the impact of varying street lighting conditions (e.g., full, partial, and no lighting) on surrogate safety measures like TTC and PET. They observed a 37% increase in TTC and a 15% reduction in vehicle speeds under improved lighting, underscoring the critical role of illumination in nighttime intersection safety.

Similarly, Mimi et al. [29] analyzed crash data involving older pedestrians in Texas and found that inadequate lighting was a key factor contributing to nighttime crash severity. Their findings support the need for better lighting and visibility solutions, especially for vulnerable users.

In a complementary line of work, Bhattacharya et al. [30] proposed data-driven models for optimizing MH-based streetlight installations in developing countries. Their use of regression and neural networks to predict energy efficiency and luminance provides an evidence-based framework for smart infrastructure design that could inform urban SCI retrofitting strategies. Cunningham et al. [31] examined the impact of retroreflective and LED-enhanced wrong-way signage at Kansas highway ramps. Their findings confirmed a significant reduction in incident rates, supporting the broader use of LED-based interventions for improving conspicuity and deterring hazardous maneuvers.

In contrast to earlier LED STOP sign studies that focused on individual compliance metrics, this research applies a combined empirical-simulation framework to assess BLS performance under various traffic and lighting conditions. The inclusion of surrogate safety indicators, GA calibration, and cross-comparison with overhead beacon systems adds methodological depth and practical relevance to current literature.

2.2. Surrogate Measures of Safety

The limitations of crash-based studies such as the rare occurrence of accidents and the need for extended observation periods have led to alternative methods for reviewing intersections safety [8,9]. For instance, ref. [32] demonstrated the use of the traffic conflict technique at unsignalized intersections through a simulation-based model (TSC-Sim), where TTC was employed as the critical indicator for conflicts. Their study established how conflict-based methods can overcome limitations of crash data and provide standardized measures of frequency and severity for intersection safety assessment. Surrogate measures of safety (SMoS) emerged as a reliable alternative to traditional crash-based analysis for evaluating the safety of SCIs. These measures enable researchers to assess potential safety issues without relying on crash data, which often require extensive observation periods and large datasets. SMoS provide actionable insights into intersection performance under varying conditions by focusing on near-miss events, driver behavior, and other safety-critical metrics.

PET and TTC are among the most widely used SMoS, which quantify the temporal or spatial proximity of conflicting road users. Navarro et al. applied PET to evaluate the impact of converting MAS into AWS intersections, indicating that AWS implementations significantly increased PET, representing safer interactions.

Recent reviews offer a broader perspective on surrogate safety measures (SSMs) beyond their application in STOP-controlled intersections. For instance, Sharma et al. (2024) provide a comprehensive overview of SSMs including TTC, PET, deceleration-based measures, and emerging sensor-driven indicators such as connected vehicle data and machine learning approaches [33,34]. Venthuruthiyil and Chunchu (2024) investigate alternative SSMs like Anticipated Collision Time, highlighting the importance of adapting thresholds to context [35]. Additionally, recent work by researchers such as Singh and Dass (2025) emphasizes the lack of standardized thresholds and promotes AI-driven frameworks for dynamic risk assessment in mixed traffic environments [36].

Furthermore, several studies demonstrate the established use of PTV VISSIM coupled with Surrogate Safety Assessment Model (SSAM) for evaluating traffic conflicts in safety analysis. Fan et al. (2013) developed a two-stage calibration and validation process using VISSIM and SSAM to estimate freeway merge conflicts, significantly improving consistency between simulated and field-measured conflicts [37]. Hasanpour and Persaud (2022) applied a similar framework to assess the safety and operational impacts of Leading Pedestrian Intervals at signalized intersections in Toronto, using microsimulation and SSAM to estimate vehicle-to-pedestrian conflicts under various traffic and geometric scenarios [38]. More recently, studies have used VISSIM in mixed-traffic scenarios with AVs and cruising behavior under varying weather conditions, incorporating SSAM for safety evaluation in urban and rural settings [39,40].

Conflict frequency and severity are among the essential components of SMoS, as well. Researchers can examine the potential risk at intersections without waiting for actual crashes to occur by investigating near-miss events. For example, vehicle-vehicle and vehicle-pedestrian conflicts can be classified into safe, mild, or dangerous interactions based on metrics such as PET or TTC. In addition, Navarro et al. reported a significant reduction in high-risk interactions after AWS implementation, with the proportion of safe interactions increasing across all the intersection approaches [12].

Vehicle speed and speed profile are considered also as a relevant traffic SMoS. Speed reductions are directly linked to lower crash severity and frequency. Some studies found that enhanced traffic control measures can be used to moderate speeds. For example, a study by Monsere et al. argued that a dynamic advanced curve warning system reduced mean vehicle speeds by 2–3 mph, as well as altering speed distributions appropriately [41]. Furthermore, Navarro et al. observed significant decreases in vehicle speeds on major approaches after STOP sign installations, with median speeds dropping by 60% [12]. In another study, Fiolić et al. highlighted the role of optimized road markings and signage visibility in increasing driver speeds during nighttime conditions, which results in improving cognitive load and situational awareness [24]. According to Arnold and Lantz, employing flashing LED STOP signs and optical speed bars is related to speed reductions. They reported significant decreases in hazardous road segments, ranging from 1 to 10 mph depending on the context [17]. These results emphasize the role of advanced warning systems and well-designed signage in achieving effective speed management.

Driver compliance, especially in yielding behaviors and adherence to traffic controls, is regarded as a cornerstone of roadway safety. Compliance rates, defined as the proportion of drivers who perform appropriately, play a critical role in vehicle-pedestrian interactions. Fitzpatrick and Park asserted that pedestrian hybrid beacons (PHBs) maintained high yielding rates (up to 97%) even at higher speeds, indicating their effectiveness during nighttime conditions [42]. In addition, Gates et al. found that fluorescent and LED-enhanced STOP signs improved driver compliance at intersections by reducing the number of non-yielding vehicles [9]. Further, Navarro et al. reported an increase in yielding rates from 45.7 to 76.7% after implementing AWS controls [12].

In summary, SMoS such as PET, speed, yielding rates, and conflict analysis provide a robust framework for examining safety at SCIs. These measures enable a nuanced understanding of the interplay between driver behavior, intersection design, and traffic control measures, paving the way for data-driven improvements in road safety. By leveraging the above-mentioned instruments, this study aims to provide new insights into the effectiveness of LED-backlit signs in enhancing safety at SCIs based on the data collected at a study site in Montréal, Québec.

Previous research has also investigated driver behavior and compliance at intersections using advanced modeling approaches. For example, Al-Sabban and Al-Ahmadi examined conversion decisions from roundabouts to signalized intersections through a simulation-based approach in Jeddah and Al-Madinah, highlighting the operational and safety trade-offs in intersection control selection [43]. Similarly, Alghamdi developed deep learning models to analyze red-light crossing violations, using variance-based sensitivity analysis to identify key behavioral and contextual predictors of non-compliance [44]. While these studies address different intersection types and modeling methods, they underline the broader significance of data-driven approaches in improving traffic safety. In contrast, the present study focuses on stop-controlled intersections and evaluates the effectiveness of LED-enhanced and backlit STOP signs in reducing blow-through and roll-through violations.

3. Proposed Method

The proposed method for addressing the safety performance of LED-enhanced STOP signs at urban intersections contained three main steps including field study, a traffic model simulation, and a statistical study (Figure 1).

3.1. Signage Treatment Definitions

To ensure clarity, the following definitions are consistently used throughout this study:

-

Standard STOP Sign: A conventional reflective stop sign with no lighting enhancements.

-

LED STOP Sign: A standard STOP sign fitted with flashing red LEDs around its border.

-

Backlit STOP Sign: A stop sign with a uniformly illuminated face powered by internal LEDs, providing full-surface lighting (not flashing).

-

Overhead Beacon: A red flashing beacon mounted above or adjacent to an intersection (typically on a pole or mast), independent of the STOP sign face.

3.2. Field Study

The effectiveness of active signage systems was studied through a carefully designed field experiment at SCIs. Several SCIs with comparable geometric layouts and traffic characteristics were selected for this study. These intersections were located along an urban corridor in Montréal (Figure 2), ensuring uniformity in traffic exposure and environmental conditions. The experimental setup allowed the same traffic stream to encounter different signage types within a short temporal and spatial range, resulting in reducing variability in the data collection process.

For the experiment, two signage types were employed: standard STOP signs fitted with LEDs and a novel BLS featuring a uniformly illuminated display. The standard signage was present at all of the study intersections. During the experimental phase, the intersections at the 18th and 25th Ave. were upgraded to BLS signage, resulting in enabling comparative analysis. To observe potential shifts in driver behavior, data were collected before and after BLS installation, with the intervention eliminated after two months to restore the original signage (Figure 3).

Advanced radar systems and video-based trajectory analysis were combined to collect the data. Then, three high-definition radars were deployed alongside 189 h of footage over ten days, recording data on vehicle volume, speed, trajectories, and compliance with signage [45]. Supplementary information including weather conditions and intersection layouts was integrated into the analysis. Speed measurements were taken at 30 m markers upstream of the stop-line, while braking points and compliance were reviewed utilizing image processing techniques. As discussed before, the trajectory data were annotated to identify driver responses including braking initiation and crossing behaviors [45,46]. Compliance and reaction patterns were categorized according to the Institute of Transportation Engineers (ITE) guidelines [17]. The dataset was used to develop microsimulation models and perform statistical analyses, which helped evaluate the effectiveness of active signage systems.

3.3. Traffic Model Simulation

Conflict analysis is employed as a robust alternative to discuss the limitations of using crash data in road safety assessment due to their infrequency [45]. A microsimulation model was developed utilizing PTV VISSIM and calibrated with the SSAM to estimate conflicts including rear-end, crossing, and lane-change incidents (Figure 4). Adjustments to the car-following model incorporated safe-to-stop braking distance constraints, and lane-changing behavior was restricted to reflect observed field conditions, ensuring accurate representation of vehicle queuing and turning behavior.

The VISSIM standard model fails to replicate real-world variability in driver behavior at STOP signs, where drivers may perform full stop and roll- or blow-through. To overcome this obstacle, two dummy links were included on approaches to allow deviations from mandatory stops. These dummy links were calibrated based on statistical and qualitative analyses of field data, resulting in creating behavior profiles as follows.

- Full stop: Drivers fully stop before proceeding.

- Roll-through: Drivers reduce speed without stopping.

- Blow-through: Drivers proceed with minimal deceleration.

The application of the aforementioned dummy links was determined by the treatment type and traffic conditions at different times of the day. Traffic volumes for four-time intervals, which were sourced from field observations and municipal data, were incorporated into the model (Figure 5).

Calibrating the microsimulation model ensured that simulated outputs closely matched real-world conditions. Given the focus on conflicts, standard calibration employing vehicle speeds or densities was complemented by turning movement counts (TMC) as the objective function. A GA was used to optimize eight key car-following parameters as follows:

-

Max deceleration of the subject vehicle;

-

Accepted deceleration of the subject vehicle;

-

Deceleration reduction distance;

-

Max look ahead distance;

-

Standstill distance;

-

Min clearance;

-

Additive part of safety distance;

-

Multiplicative part of safety distance.

The selection of these parameters was informed by prior global sensitivity analysis research on VISSIM/Wiedemann driver-behavior parameters by Sayed and colleagues [32]. Figure 6 shows the calibration process and GA workflow. The algorithm iteratively selected, mutated, and recombined parameters based on their fitness scores utilizing an initial population of nine chromosomes. Termination occurred after four generations without improvement or after 15 iterations, which one occurs first (Figure 7a). The final calibrated parameters were extracted, and the simulated turning counts were validated against observed data, indicating an appropriate representation of the observed counts (Figure 7b).

The mean squared error (MSE) was applied as a goodness-of-fit metric to evaluate the validity of the calibration results.

$$MSE={\sum }_{i=1}^n{\left(C_i^E-C_i^M\right)}^2$$

(1)

where $C^E$ is the observed TMC for turn iii, $C^M$ represents the simulated TMC for turn iii, and n is the total number of TMCs.

The MSE ensured that higher penalties were assigned to larger discrepancies, and errors of opposite signs did not offset one another.

Validation was conducted employing trajectory data from field experiments and conflict outputs from SSAM. Conflict thresholds were defined as follows.

-

TTC: ≤1.5 s;

-

PET: ≤5 s;

-

Angles for rear-end and crossing conflicts: 30–80 degrees.

A two-sample Mann–Whitney U-Test was used to compare observed and simulated conflict distributions. The test provided a robust statistical evaluation as the conflict data were discrete and non-normally distributed (Figure 7c).

$$U_i=n_1{n}_2+{\displaystyle \frac{n_1\left(n_1+1\right)}{2}}-R_i\hspace{1em}\hspace{1em}Z=(U_l-{\displaystyle \frac{n_1{n}_2}{2}})/\sqrt{{\displaystyle \frac{n_1{n}_2(n_1+n_2+1)}{12}}}$$

(2)

where $n_1$ and $n_2$ are the sample sizes, and $R_i$ is the sum of ranks for group iii.

The calculated Z-value of 0.454 fell within the acceptance region [−1.96, 1.96], with a p-value of 0.65. No significant difference was reported between the observed and simulated conflicts, resulting in validating the model.

While SSAM provides a practical means to analyze conflict potential in micro simulated networks, it has known limitations in urban environments, including simplified car-following assumptions and lack of pedestrian modeling. These limitations are acknowledged and discussed further in Section 5.

3.4. Statistical Study

3.4.1. Speed Analysis

The vehicle speed was analyzed to determine the effects of active signage on drivers’ speed profiles as they approached SCIs. ANOVA was conducted to compare mean speeds among intersections featuring standard signs, LED STOP signs, and backlit LED STOP signs.

3.4.2. Driver Compliance Analysis

Driver compliance at stop lines was analyzed using previously defined behavioral categories: full stop, roll-through, and blow-through [30,31]. The experimental scenarios were as follows.

-

Same drivers: Observing repeated behaviors of consistent drivers.

-

Same direction: Aggregating data for vehicles traveling in one direction.

-

Opposing directions: Comparing behavior across multiple approaches at a single intersection.

Multinomial logistic regression models were developed utilizing SPSS v28.01.1 to evaluate the effects of signage type and environmental factors (e.g., lighting and opposing traffic) on compliance levels (

Table 1. Variable description and characteristics.

Variable Category	Variable Name	Variable Type	Number of Outcomes	Variable (Description)
Dependent	Driver Compliance	Nominal-Category	3	FS (Full stop)
				RT (Roll-through)
				BT (Blow-through)
Independent	Treatment type	Nominal-Category	4	Stop (Standard stop)
				LED stop
				BLS
				Overhead beacon
Independent	Pavement surface condition	Boolean	2	0 (dry)
				1 (wet)
Independent	Conflict potential	Boolean	2	0 (no opposing traffic/ped)
				1 (opposing traffic/ped)
Independent	Natural ambient light	Boolean	2	0 (prior to civic dawn)
				1 (after civic dawn)
Independent	Maneuver	Boolean	2	0 (through)
				1 (turn)
Independent	LED stop	Boolean (dummy)	2	0 (standard stop)
				1 (LED stop)
Independent	BLS *	Boolean (dummy)	2	0 (standard stop)
				1 (backlit stop)
Independent	Overhead Beacon	Boolean (dummy)	2	0 (standard stop)
				1 (stop with overhead beacon)

* Treatment performance on the study.

).

The full stop category was set as the reference for logistic regression models, while blow-through and roll-through were modeled as relative categories. For instance, the probability of a blow-through (category A) relative to a full stop (category C) was modeled as follows.

$$\ln \left({\displaystyle \frac{p\left(A\right)}{p\left(C\right)}}\right)=a_1+{\beta }_1{x}_{i1}+{\beta }_2{x}_{i2}+\dots +{\beta }_m{x}_{im}$$

(3)

The likelihood ratio chi-squared test and McFadden’s Pseudo-R² were applied to examine model performance (Equation (5)).

$$McFadden={\displaystyle \frac{{Deviance}_{null}-{Deviance}_{full}}{{Deviance}_{null}}}$$

(4)

3.4.3. Conflict Analysis

The SMoS were employed to investigate traffic conflicts given the limited availability of collision data over the study period. Metrics such as TTC (TTC ≤ 1.5 s), PET (PET ≤ 5 s), and conflict angles (30–80 degrees) were extracted from vehicle trajectories using the SSAM. A significant relationship was reported between signage type and conflict occurrence (p < 0.05), with active signage being related to fewer conflicts. A Poisson regression model was utilized to estimate conflict occurrences. However, a negative binomial (NB) regression model was applied due to overdispersion (variance exceeding the mean). The NB model accommodated data heterogeneity and demonstrated superior fit, evidenced by lower Akaike Information Criterion (AIC) values compared with the Poisson model.

Goodness-of-fit evaluations, including Pearson’s Chi-squared and Deviance tests, supported the reliability of the statistical models. The McFadden Pseudo-R² and AIC validated the model’s predictive performance, ensuring the reliability in estimating the safety outcomes.

The generalized model for conflict frequency at intersection is expressed as follows.

$${\lambda }_i=\mathit{EXP}\left({\beta }_0+{\beta }_1\mathit{ln}\left(F_{1i}\right)+{\beta }_2\mathit{ln}\left(F_{2i}\right)+{\beta }_3{X}_{3i}+\dots +{\beta }_n{X}_{ni}\right)$$

(5)

4. Results

This section presents the outcomes of empirical and simulation-based evaluations of different STOP sign treatments. The analysis is structured into three parts, vehicle speed, driver compliance, and conflict frequency, offering a comprehensive perspective on the effectiveness of LED and BLS compared to standard STOP signs and overhead beacon systems.

4.1. Speed Analysis

To assess how different STOP sign treatments influence driver behavior, a one-way analysis of variance (ANOVA) was conducted using a dataset of 2000 vehicle observations. The results demonstrated that both LED and BLS signs significantly reduced vehicle approach speeds compared to traditional STOP signs. Specifically, LED signs reduced mean approach speeds by 5.3% (±2.0%), while BLS signs resulted in a 3.5% (±1.8%) decrease. The difference between LED and BLS was not statistically significant, indicating comparable effectiveness in moderating vehicle speed. Conversely, the presence of an overhead beacon did not produce a significant reduction, suggesting that its impact on speed control in urban environments is limited.

The validity of these comparisons was supported by Levene’s test, which confirmed the homogeneity of variance assumption. As illustrated in Figure 8, the boxplots of vehicle speeds show lower median and upper-quartile values for LED and BLS treatments, along with reduced variability and fewer outliers. These findings suggest that enhanced signage fosters more consistent driver behavior. Additionally, data from pedestrian crossings indicated that drivers approaching intersections with BLS signage exhibited significantly lower speeds than those approaching standard STOP signs, with the difference confirmed by an independent t-test (t₍₁₅₅₀₎ = 10.9, p = 0.066). The results collectively support the role of active signage in improving speed control and enhancing safety margins, particularly in conditions of low visibility.

4.2. Compliance Analysis

The effectiveness of signage treatments in promoting driver compliance was analyzed using multinomial logistic regression models. Driver behavior was categorized into full stops, roll-throughs, and blow-throughs, in alignment with ITE guidelines. The statistical models evaluated three experimental scenarios and controlled environmental factors such as pavement conditions, lighting, maneuver types, and opposing traffic. The reference category in all models was full stop, enabling direct comparison of the likelihood of non-compliant behavior.

The analysis revealed that LED and BLS signs significantly improved compliance compared to traditional STOP signs. For intersections with LED signs, the odds of blow-through versus full-stop behavior were 0.349, while the odds of roll-through versus full-stop were 0.501. For BLS signs, these odds were even lower: 0.096 for blow-through and 0.172 for roll-through. These results indicate that active signage, particularly BLS systems, is associated with a lower likelihood of non-compliant behavior. The statistical models were validated by likelihood ratio chi-squared tests and McFadden’s Pseudo-R² values, which confirmed the explanatory power of the full models compared to the null.

Scenario-specific findings offered additional insight. Scenario 2, which aggregated data by travel direction, showed a drop in classification accuracy from 73.7% to 63.7%, emphasizing the importance of filtered sample selection in Scenario 1. Scenario 3, which compared opposing directions at a single intersection, did not exhibit statistically significant differences in compliance, and variables such as ambient light and turning maneuvers were excluded due to their limited influence. Figure 9 further supports these conclusions by showing earlier braking points for drivers exposed to LED and BLS signs, reinforcing the observed increase in compliance.

Table 2. Compliance odds ratio in scenario 2 and delta (δ) with scenario 1.

Compliance Comparison	Sign Type	Odds Ratio	δ
Full Stop vs. Roll-Through	STOP	1.56	+4.6%
	LED	2.52	+8.2%
	BLS	2.88	+5.1%
	Overhead Beacon	2.20	+2.6%
Full Stop vs. Blow-Through	STOP	10.49	+2.9%
	LED	18.14	+11.1%
	BLS	23.73	+14.2%
	Overhead Beacon	15.34 †	+20.2%
Full Stop Rates (Opposing Traffic = No)	STOP	40.96%	+1.01%
	LED	54.36%	−2.17%
	BLS	53.48%	−3.86%
	Overhead Beacon	51.88%	+3.43%
Full Stop Rates (Opposing Traffic = Yes)	STOP	16.66%	−2.10%
	LED	14.51%	+0.28%
	BLS	18.52%	+2.61%
	Overhead Beacon	13.93%	−4.51%
Total Full Stop Rate	STOP	57.62%	−1.10%
	LED	68.87%	−1.89%
	BLS	72.00%	−1.25%
	Overhead Beacon	65.81%	−1.09%
Roll-Through Rate	STOP	36.89%	+1.04%
	LED	27.34%	+1.56%
	BLS	24.97%	+0.87%
	Overhead Beacon	29.90%	+0.28%
Blow-Through Rate	STOP	5.49%	+0.06%
	LED	3.80%	+0.33%
	BLS	3.03%	+0.38%
	Overhead Beacon	4.29% †	+0.81%

Note: Odds ratios < 1 imply greater likelihood of non-compliance relative to full stop. BLS signs achieved the lowest blow-through rates. † Indicates this value was rejected in the final MNL regression model due to insignificance. “δ” indicates the percentage difference in compliance metrics or odds ratios relative to the reference values in Scenario 1.

summarizes the odds ratios across signage types, clearly demonstrating the superior performance of BLS and LED treatments. To clarify the comparative effect across different modeling scenarios, Table 2 includes a column labeled “δ” (delta), which represents the change in odds ratios or compliance proportions relative to the values obtained in Scenario 1, the most controlled scenario based on filtered driver samples. A positive delta indicates an increase in the respective outcome compared to Scenario 1, while a negative value denotes a reduction. This helps capture how the results vary under broader traffic conditions in Scenarios 2 and 3.

To assess the classification performance of the multinomial logistic regression (MNL) model, confusion matrices were computed for Scenarios 1 and 2. In Scenario 1, the model correctly classified full-stop behavior in 96.3% of cases, while roll-through and blow-through behaviors were correctly identified in 37.8% and 28.7% of instances, respectively. The overall classification accuracy was 73.7%, corresponding to a 26.3% misclassification rate. In Scenario 2, accuracy dropped to 63.7%, with 98.2% accuracy for full stops, but only 2.0% and 0.0% for roll-through and blow-through, respectively, indicating reduced performance in broader, less filtered datasets. Blow-through events were extremely rare in Scenario 2, which resulted in a classification accuracy of 0.0% for this behavior. This reflects the well-known limitation of multinomial logit models in handling sparse categories, and while overall accuracy remained acceptable, future work could explore rare-event or penalized logistic regression approaches to improve estimation of infrequent behaviors.

Table 3. Confusion Matrix and Classification Accuracy of MNL Model for Scenarios 1 and 2.

Scenario	Behavior	Correctly Classified (%)	Misclassification Rate (%)
1	Full Stop	96.3	—
	Roll-Through	37.8	—
	Blow-Through	28.7	—
	Overall	73.7	26.3
2	Full Stop	98.2	—
	Roll-Through	2.0	—
	Blow-Through	0.0	—
	Overall	63.7	36.3

provides a summary of these results. While the predictive performance for rare behaviors was limited, the models still revealed statistically significant directional effects of signage on driver compliance, which remains the primary objective of the analysis.

4.3. Conflict Frequency Analysis

In this study, two surrogate safety indicators were used to identify and evaluate traffic conflicts: TTC and PET. TTC represents the time remaining before two road users would collide if they continued at their current speed and trajectory. PET refers to the time gap between when one road user exits a potential conflict point and another enters it. These metrics were extracted from trajectory data using the SSAM, a widely validated tool for conflict analysis. Based on established thresholds in the literature and prior studies using SSAM, conflict events were defined using TTC ≤ 1.5 s for vehicle-to-vehicle interactions and PET ≤ 5 s for vehicle-to-pedestrian interactions. These values reflect conservative cutoffs that capture only high-risk encounters likely to result in actual collisions under minor variation in conditions.

Given the limitations of crash-based evaluations for low-frequency events, this study used SSMs to estimate traffic conflict frequency. Time-to-collision (TTC ≤ 1.5 s) and post-encroachment time (PET ≤ 5 s) were selected as critical safety indicators, in line with established thresholds. Conflict events were extracted using a calibrated microsimulation model developed in PTV VISSIM and analyzed through the SSAM. The model was previously validated using a Mann–Whitney U-test, which showed no significant difference between observed and simulated conflict distributions, confirming its reliability.

To examine how compliance and traffic volume influence conflict frequency, three statistical models were developed: a baseline or null model using only observed conflict counts; Model A, incorporating driver compliance ratios derived from multinomial logistic regression; and Model B, which additionally included total intersection traffic volume as a covariate.

Table 4. Regression analysis of critical conflicts in relation to traffic volume and driver compliance.

	Null Model (Observed Conflict Data)	Model A (Compliance Odds Ratio)	Model B (Traffic Volume as Covariate)
Parameter	$\begin{array}{l}E\left\{{Conf}_i\right\}\\ =\mathit{EXP}({\beta }_0+{\beta }_1\mathrm{D}\mathrm{u}\mathrm{m}\\ +{{\beta }_2}_{{\displaystyle \frac{full stop}{nonfull-stop}}}+{{\beta }_3}_{{\displaystyle \frac{full stop}{roll-through}}}\\ +{{\beta }_4}_{{\displaystyle \frac{full stop}{Blow-through}}})\end{array}$	$\begin{array}{l}E\left\{{Conf}_i\right\}=\mathit{EXP}({\beta }_0+{\beta }_1\mathrm{D}\mathrm{u}\mathrm{m}+{{\beta }_2}_{{\displaystyle \frac{full stop}{nonfull-stop}}}\\ +{{\beta }_3}_{{\displaystyle \frac{full stop}{roll-through}}}+{{\beta }_4}_{{\displaystyle \frac{full stop}{Blow-through}}}).{\left(f sum\right)}^{{\beta }_5}\end{array}$	$\begin{array}{l}E\left\{{\mathit{Conf}}_i\right\}=\mathit{EXP}({\beta }_0+{\beta }_1\mathrm{D}\mathrm{u}\mathrm{m}\\ +{{\beta }_2}_{{\displaystyle \frac{full stop}{nonfull-stop}}}\\ +{{\beta }_3}_{{\displaystyle \frac{full stop}{roll-through}}}).{\left(f sum\right)}^{{\beta }_4}\end{array}$
Intercept	3.610 (p < 0.001) ***	−3.186 (p = 0.191)	−2.881 (p = 0.234)
Variable 1	0.294 (p = 0.155)	0.430 (p = 0.030) *	0.414 (p = 0.038) *
Variable 2	−0.302 (p = 0.204)	−0.406 (p = 0.608).	−0.448 (p = 0.036) *
Variable 3	−0.558 (p = 0.060)	−0.866 (p = 0.003) **	−0.745 (p = 0.040) **
Variable 4	0.012 (p = 0.651)	0.022 (p = 0.351)
Variable 5	-	0.927 (p = 0.005) **	0.8892 (p = 0.072) **
Degree of Freedom	30.0000	29.0000	30.0000
alpha (α)	0.2480	0.1935	0.1985
phi	4.0326	5.1680	5.0377
Deviance	37.230	37.488	37.457
Pearson	32.7117	35.6722	36.3877
Over Dispersion	1.0904	1.2301	1.2129
Log Likelihood	−141.6963	−138.0624	−138.4248
AIC	295.3926	290.1247	288.8496

Notes: p < 0.05 = * (statistically significant); p < 0.01 = ** (highly significant); p < 0.001 = *** (very highly significant).

summarizes the coefficients, p-values, and fit statistics for each model.

Five explanatory variables were included in these models. Variable 1 is a dummy indicator for evening time intervals, taking the value of 1 during evening hours and 0 otherwise. Variable 2 represents the ratio of full-stop to non-full-stop driver behaviors. Variable 3 measures the ratio of full-stop to roll-through behaviors, while Variable 4 captures the ratio of full-stop to blow-through behaviors. Variable 5 denotes the total volume of vehicles entering the intersection. These variables capture both temporal and behavioral influences on conflict occurrence.

The results indicate that Model B provided the best fit, with the lowest Akaike Information Criterion (AIC = 288.85) and acceptable overdispersion (1.21). Variables 1, 3, and 5 were statistically significant (p < 0.05), confirming that evening traffic, higher rates of non-compliant behavior, and increased volume were positively associated with conflict frequency. Model A, which did not include traffic volume, also performed better than the null model but less accurately than Model B. These findings reinforce the importance of combining behavioral and operational data to predict safety outcomes at intersections.

Figure 10 further illustrates the relationship between traffic volume and critical conflicts. Subfigure (a) shows the direct correlation between hourly volume and conflict counts, while subfigure (b) presents the regression surface generated from Model B, demonstrating how changes in compliance ratios and traffic levels jointly influence the probability of a conflict.

Building on these model results,

Table 5. Expected frequency of conflict with and without treatment (average conflict derived from the simulation).

Condition	Sign Type	Vehicle Conflicts	Pedestrian Conflicts
Peak Hours	Regular STOP	51.49 (sim: 50.33)	47.55 (sim: 45.67)
	BLS	27.14 (sim: 37.10)	34.35 (sim: 35.00)
	Reduction	47.3%	27.7%
Nighttime	Regular STOP	36.71 (sim: 30.10)	37.64 (sim: 25.00)
	BLS	12.69 (sim: 14.20)	16.70 (sim: 12.00)
	Reduction	65.4%	55.6%

compares the expected frequency of conflicts at intersections with and without BLS signage. For standard STOP signs, peak hour simulations predicted 51.5 conflicts on average, while intersections equipped with BLS signs saw a reduction to 27.1. During nighttime, BLS intersections averaged 12.7 conflicts compared to 36.7 without treatment. Pedestrian conflicts followed a similar pattern, with reductions of 27.8% during peak hours and 55.6% at night. These findings indicate that BLS signage is particularly effective in low-light conditions, reducing both vehicle and pedestrian-related conflicts by nearly half.

In summary, the conflict frequency analysis confirms the significant impact of active signage treatments on safety outcomes. The integration of compliance behavior and traffic volume into the predictive model greatly enhances its accuracy and interpretability. These results demonstrate that LED and BLS STOP signs not only influence driver behavior but also translate into measurable reductions in safety-critical events.

4.4. Study Limitations

While the results demonstrate promising safety impacts of LED and BLS signage, several limitations should be acknowledged. First, the use of surrogate safety indicators such as PET and TTC, while practical, may not fully capture long-term crash trends and may vary with context. Second, although the microsimulation model was calibrated using field data, it may not perfectly reflect driver behavior under different environmental or cultural conditions. Third, the study was limited to one urban corridor in Montréal, which may affect the generalizability of the findings to other regions or roadway types. Lastly, the observation period was relatively short, and longer-term behavioral adaptations were not captured.

In addition, the findings should be interpreted with consideration of certain contextual constraints. The analysis was conducted in an urban environment in Montréal, under MUTCDC regulations, which may limit the generalizability of the results to other jurisdictions with different traffic control standards. Driver behavior and compliance may also differ in rural contexts, where traffic volumes, enforcement practices, and roadway characteristics vary considerably. In addition, seasonal variations in lighting conditions in Québec could have influenced driver responses to illuminated signage. These contextual factors should be taken into account when extrapolating the results to other settings, and future research is encouraged to replicate this work in diverse geographic and regulatory environments to enhance external validity. In addition, future research should conduct a comprehensive global sensitivity analysis of the modeling assumptions. Techniques such as Sobol’s first- and total-order indices or Morris screening could be applied to quantify how TTC/PET thresholds and GA-calibrated driver behavior parameters influence conflict frequency outcomes. This would provide greater transparency on the robustness of surrogate safety estimates and the calibration framework. Moreover, future research should include explicit multicollinearity diagnostics, such as variance inflation factors, and residual analyses (e.g., Pearson and deviance residual checks) to complement the count-model specification tests. Although this study found that Negative Binomial models outperformed Poisson alternatives, supporting the stability of coefficient estimates, additional diagnostics would further validate the robustness of the results. Incorporating penalized or rare-event regression approaches could also improve estimation for infrequent behaviors such as blow-through violations.

5. Conclusions

This study introduces a novel methodological framework that combines high-resolution field data, surrogate safety modeling, and microsimulation to evaluate a new class of uniformly backlit STOP signs. These contributions fill a critical gap in the literature by offering deeper insights into the behavioral and conflict-level impacts of active signage in dense urban environments.

The present study seeks to assess the safety performance of LED-enhanced and BLS at SCIs through field experiments, traffic microsimulation models, and conflict-based statistical analyses. To this end, a robust before–after experimental design was employed where similar geometry, environmental conditions, and traffic flow ensured reliable results.

The results demonstrate that BLS and LED-enhanced signage significantly outperform traditional STOP signs by reducing blow-through and roll-through behaviors, while improving driver compliance rates and lowering vehicle speeds. Evaluating SMoS such as PET and TTC provided valuable insight into conflict frequency and severity. Notably, compliance odds ratios were inversely correlated with conflicts, indicating that higher compliance effectively reduces crash frequency. BLS signs reduced conflicts by 65.5 and 46.8% at night and daylight, while pedestrian crossing conflicts decreased by 55.6 and 27.8%, respectively.

A calibrated microsimulation model using GA optimization and validation through observed traffic movement counts reinforced the above-mentioned results. A significant relationship was reported between the observed and simulated conflicts, confirming the reliability of SMoS as alternatives to crash-based studies. These results underscore the limitations of traditional volume-based models since conflict inclusion provided more accurate accident predictions, especially under conditions of varying traffic flow and compliance levels.

Based on the findings from surrogate safety indicators and simulation results, BLS and LED STOP signs appear to offer measurable safety benefits under specific conditions, particularly in low-light environments where visibility is limited. These results suggest that such devices can improve driver compliance, reduce approach speeds, and minimize the occurrence of vehicle and pedestrian conflicts at stop-controlled intersections.

While the compliance models revealed statistically significant relationships, the McFadden’s Pseudo-R² values remained relatively low (0.011–0.023 in some cases), reflecting the complexity and variability of human driving behavior. Such values are common in multinomial models applied to discrete compliance outcomes. Nonetheless, we interpret these results with caution, emphasizing the directional trends over precise prediction. As such, the policy implications drawn from these findings are exploratory and intended to guide further pilot testing rather than immediate large-scale deployment.

It is important to note that SSAM’s use in urban microsimulation presents certain constraints. Although it does not fully account for pedestrian interactions or lateral dynamics, its integration with calibrated VISSIM models enables reasonable estimation of vehicle conflicts. Conflict outputs in this study were validated against field observations using non-parametric testing, and multi-period simulation was used to reflect temporal traffic variation, offering sensitivity across realistic conditions.

Although the field data collection period was limited to 10 days, this study leveraged a combination of high-resolution trajectory data, surrogate safety indicators (TTC, PET), and validated microsimulation models to ensure robustness. The inclusion of varied lighting and surface conditions, modeled across multiple time intervals, allowed for meaningful behavioral and conflict analysis. This integrated approach, supported by non-parametric validation and statistical modeling, ensures that the findings offer generalizable insights despite the limited observation window.

However, it is important to acknowledge the study’s limitations. The behavioral and conflict analyses were based primarily on surrogate safety indicators and simulation rather than comprehensive crash data. While a preliminary review of crash records on the study corridor was conducted to support the problem statement, the main evaluation relied on predictive modeling and observed driver behavior. Furthermore, the study was limited to a single urban corridor, and the generalizability of the results remains to be tested at other sites. Given the substantial cost of LED and SCI devices, broad deployment should be approached cautiously. We therefore recommend limited initial investment in additional pilot installations across diverse urban settings, followed by long-term evaluation based on crash data. A full cost-effective analysis, considering installation, maintenance, and competing safety priorities, will be essential before scaling implementation. This cautious approach will help ensure that limited safety budgets are directed toward the most effective and economically viable interventions.