Exploring the Contribution of Road Infrastructure and Environmental Factors to Crash Severity at Intersections in Mixed Traffic Settings

Abstract

Road traffic crashes claim approximately 1.19 million lives annually worldwide, with low- and middle-income countries (LMICs) bearing a disproportionately high share of this burden. Intersections in these contexts are particularly hazardous due to mixed, non-lane-based traffic and infrastructural constraints. This study analysed 1242 police-reported intersection crashes (2021–2025) from Douala and Yaoundé, Cameroon, using binary probit and logistic regression models to identify infrastructural and environmental determinants of crash severity. Results from both models were consistent, indicating that late-night and early-morning crashes (00:00–05:59) significantly increased the probability of severe outcomes by 13.5% (p-value < 0.05), while single-lane roads raised it by 21.5% (p-value < 0.05; OR = 5.38), and two-lane roads raised the probability by 9.1% (p-value < 0.05; OR = 3.90) compared with multilane sections. Additionally, centre lines were associated with safer outcomes than physical separation (p-value < 0.05; OR = 0.30). Although model fit indices were modest (Nagelkerke R² = 0.118), typical of cross-sectional crash-severity models, the findings underscore the dominant influence of road geometry and lighting-related temporal exposure in shaping intersection crash outcomes. These insights provide a basis for targeted interventions such as road widening, improved night-time illumination, and simplified midblock designs to enhance safety in Cameroon and similar LMIC urban settings.

1. Introduction

Road traffic crashes are a major global public health and socioeconomic challenge, causing approximately 1.19 million deaths and millions of injuries annually [1]. Low- and middle-income countries (LMICs) bear a disproportionate burden, accounting for over 90% of fatalities despite having fewer vehicles. Urban intersections are among the most hazardous locations due to the convergence of vehicles, pedestrians, and cyclists, creating multiple conflict points and a high potential for severe crashes [2,3]. Understanding the determinants of crash severity at intersections is critical for improving road safety and achieving Sustainable Development Goals.

In LMICs, intersections often feature mixed traffic, where motorized and non-motorized road users share limited space with minimal separation [4,5]. This heterogeneity increases crash frequency and severity, particularly in cities with unplanned growth, informal traffic behaviour, and limited pedestrian facilities [6,7,8]. Vulnerable road users, such as pedestrians and cyclists, are disproportionately affected, and intersections in these contexts are often poorly designed or inadequately controlled, exacerbating the risk of severe collisions [9].

Crash severity is influenced by a combination of human, environmental, and infrastructure-related factors. Driver characteristics, including age, gender, and risk-taking behaviour, are well-established predictors [10,11]. Environmental conditions, such as lighting, weather, and pavement state, also play a critical role [12,13,14]. Road infrastructure elements including lane width, shoulder provision, medians, signage, and intersection geometry can mitigate or exacerbate crash severity [6,15,16,17,18]. Poorly designed intersections lacking signals, markings, or channelization have consistently been associated with higher severity outcomes.

Interactions between environmental and geometric factors further influence crash risk. For example, inadequate lighting or wet pavement may amplify the negative effects of narrow lanes or insufficient shoulders [19,20,21]. Most evidence comes from high-income countries with homogeneous traffic, limiting applicability in LMIC contexts where lane discipline is weak and traffic is heterogeneous [22,23]. Empirical studies that jointly assess infrastructure and environmental determinants in mixed-traffic LMIC intersections remain scarce [24,25], highlighting the need for context-specific analyses.

This study addresses this gap by examining the impact of road infrastructure and environmental factors on crash severity at intersections in Cameroon, a representative LMIC urban setting with mixed traffic [26]. Using police-reported crash data collected over four and a half years, binary probit and logistic regression models estimate the likelihood of severe crashes based on roadway and environmental attributes [27,28]. The findings aim to inform policymakers and urban planners on targeted interventions, such as lane redesign, surface maintenance, and signal optimization, to reduce crash severity. The remainder of the paper is organized as follows: Section 2 reviews relevant literature on infrastructure and environmental determinants; Section 3 presents the methodology and data; Section 4 discusses the results; and Section 5 concludes with policy implications and recommendations for improving intersection safety in LMIC cities.

2. Literature Review

2.1. Contribution Factors in High-Income Countries

Road intersections are widely recognized as some of the most hazardous points in road networks because they concentrate multiple conflict points and involve diverse road users, including pedestrians, cyclists, and motorcyclists. Their complex traffic dynamics and high exposure make them disproportionately prone to severe crashes [11,29,30]. In Singapore, for example, intersections account for around 35% of reported crashes, while in the U.S., intersection-related crashes are hundreds of times more likely to involve turning manoeuvres with obstructed views compared to non-intersection crashes [31]. The study of crash severity at intersections has also evolved over time. Early research before 2000 largely focused on human and behavioural attributes, such as driver age, gender, alcohol use, and crash timing [10,32]. Later work progressively included vehicle characteristics, environmental factors, and particularly road infrastructure, with growing evidence that geometric and cross-sectional elements play a decisive role in shaping crash outcomes [33].

Among these elements, lane number and lane width are central in determining vehicle operating speeds and manoeuvring complexity. Crashes on single-lane roads are generally less likely to be fatal than those on multilane facilities [34]. Yet the influence of lane width is less straightforward: while narrow lanes below 2.8 m constrain movements and raise risks, excessively wide lanes above 3.1–3.2 m may encourage speeding and unsafe overtaking [15]. Indeed, some studies report that widths exceeding 3.4 m are particularly detrimental [35], though others recommend flexibility for narrower designs below 3.6 m [36]. Broader comparative analyses confirm this complexity, with [3] showing that while adding lanes and medians can reduce severity in some contexts, left-turn, angular, and head-on collisions remain highly severe.

Closely related to lanes is the design of medians, which regulate traffic flows, reduce head-on and angular crashes, and provide refuge for pedestrians. Their safety effect, however, depends on type and width. Raised or striped medians may underperform [37], while wider medians generally improve safety, particularly on high-speed approaches [38]. Evidence suggests that each additional meter of median width reduces crashes slightly [19], yet paradoxically intersections with very wide medians sometimes show higher crash counts, reflecting the complex relationship between space allocation and traffic behaviour. Other studies have identified junction design and median form among a large set of infrastructure-related risk factors shaping crash outcomes across urban networks [39].

Shoulders and roadside conditions also influence crash severity by providing recovery space and affecting sight distance, but their effects are context dependent. Wider shoulders can enhance safety on high-speed roads, yet they may also raise operating speeds and thus severity [21]. Widening unpaved shoulders by even 1 ft has been shown to lower crashes on four-lane divided roads [37]. Still, terrain, shoulder type, and side friction are critical: in some contexts, paved shoulders have been linked to higher severity, while medians and access control mitigate risks [40].

Pavement condition is another determinant of intersection safety, as it directly affects vehicle stability, braking, and skid resistance. Adequate pavement maintenance improves safety, particularly at signalized intersections [16]. Poor surface conditions increase risks of loss of control [20], yet counterintuitively, very good pavement conditions may also raise severity by encouraging speeding [41]. Studies further show that wide pavements combined with high-speed limits tend to increase crash severity, underscoring the need for continuous monitoring and maintenance strategies [31,42].

Beyond cross-sectional features, road geometry plays a major role in shaping crash outcomes at intersections. Horizontal curvature, vertical alignment, and sight distance are consistently linked with severity. Sharp horizontal curves, inadequate super-elevation, and insufficient lateral clearance raise risks considerably [43,44]. Right-turn curves are particularly hazardous during overtaking and lane-changing manoeuvres, while sharper curves affect both directions. Evidence from urban environments confirms that alignment, profile, and cross-sectional characteristics such as margins and cross slopes significantly influence crash rates [31].

The presence and quality of facilities for vulnerable road users are critical in determining severity outcomes. Signalized intersections with marked crosswalks enhance pedestrian safety by clarifying right of way, yet their effectiveness depends on road configuration. On multilane roads without medians, marked crosswalks may actually increase pedestrian crash frequency compared to unmarked crossings [45]. Sidewalk provision is assumed to reduce pedestrian crash severity, but where sidewalks end abruptly or lack accessibility features, risks may rise, particularly for persons with mobility impairments. Despite their importance, facilities for pedestrians and cyclists remain underrepresented in infrastructure-severity research, especially in LMIC mixed-traffic contexts where VRUs dominate flows [46].

Roadside barriers and guardrails are commonly deployed to shield vehicles from roadside hazards and protect pedestrians, yet their safety effects vary. Hitting a guardrail has been shown to reduce fatal and severe injury risks by nearly half [47], but outcomes differ by barrier type. For example, strong-post W-beam guardrails may increase severity compared to low-tension cable systems, while motorcycle collisions with barriers are disproportionately severe [48]. Large-scale evidence from Japan further indicates that guardrails can be associated with higher injury severity at intersections, pointing to the context-sensitive nature of their effects [49].

Finally, environmental conditions amplify the influence of infrastructure on severity outcomes. Poor weather, low visibility, and inadequate lighting increase crash risks, particularly at intersections where judgment and coordination demands are high. Studies using advanced techniques such as machine learning confirm that natural light, road category, and multi-vehicle involvement significantly interact with infrastructure features in determining severity [12,50]. This growing evidence highlights the need to integrate environmental considerations into infrastructure-focused safety strategies.

2.2. Contribution Factors in Low- and Middle-Income Countries

Road intersections in low- and middle-income countries (LMICs) present unique safety challenges due to the co-existence of heterogeneous road users, weak enforcement, and infrastructure often not designed for mixed traffic. In dense urban environments such as Dhaka, India, and Cartagena, intersections remain critical hotspots for fatal and serious injuries [51,52]. Studies across South and Southeast Asia reveal that intersections accommodate a complex fleet of motorized vehicles, pedestrians, and informal transport modes, where right-of-way ambiguity and limited channelization increase crash likelihood [23,25]. In Thailand, motorcycle-involved crashes dominate intersection injuries, reflecting the vulnerability of powered two-wheelers under weak lane discipline and inadequate geometric control [24]. Similarly, African cities such as Nairobi and Dar es Salaam exhibit high injury severities at signalized intersections, where infrastructure deficits, poor lighting, and limited pedestrian facilities amplify risks [8,9]. These findings reinforce the fact that intersection safety in LMICs cannot be fully understood through models derived from high-income countries, as local traffic compositions and behavioural patterns differ markedly.

Among infrastructure characteristics, lane configuration, road width, and median design have emerged as critical determinants of crash severity in LMIC contexts. In India and Bangladesh, narrow lanes and inadequate separation between opposing flows exacerbate overtaking conflicts and side-impact collisions, particularly where pedestrian and non-motorized vehicle activity is high [23,53]. Conversely, excessively wide lanes in urban corridors have been shown to encourage higher speeds, contributing to severe injuries among vulnerable road [51]. Research in Colombia and Brazil also highlights the role of median design, finding that undifferentiated or poorly marked medians fail to regulate traffic streams effectively, while properly dimensioned physical medians help reduce head-on and crossing crashes [52,54]. In Iran, intersections and roundabouts with improved geometric alignment and wider medians were associated with lower crash severity, suggesting that consistent delineation and channelization yield measurable safety benefits even in resource-constrained contexts [55]. However, the performance of such measures is closely linked to driver compliance and enforcement quality, both of which remain weak in most LMIC networks.

Road surface condition and roadside design also exert a significant influence on intersection safety. Studies from Brazil and Tanzania report that poor pavement maintenance and the absence of proper shoulders or recovery zones increase the likelihood of run-off-road and loss-of-control crashes [9,54]. In contrast, overengineered or recently resurfaced pavements may indirectly encourage speeding, particularly where enforcement is limited and mixed traffic persists [53]. Findings from Kenya further indicate that roadside encroachments and informal commercial activities near intersections compromise visibility and limit pedestrian refuge options, amplifying exposure to severe collisions [8]. These outcomes suggest that the relationship between infrastructure and crash severity in LMICs is inherently nonlinear and highly sensitive to environmental and behavioural factors.

The integration of safety facilities for pedestrians and motorcyclists remains limited across most LMIC intersection studies, despite their overrepresentation in crash statistics. Evidence from Bangladesh and India shows that marked crosswalks and sidewalks are often absent or discontinuous, while signal timing rarely accommodates non-motorized users [25,53]. Similarly, Thai and Brazilian studies emphasize that vulnerable road users suffer disproportionately from inadequate delineation, poor lighting, and limited protective infrastructure [24,54]. A cross-country review by [7] highlights that in low-income settings, the lack of integrated road design standards and weak institutional capacity undermine the effectiveness of conventional safety measures. These structural and institutional barriers worsen the effects of geometric deficiencies, leading to persistently high intersection fatality rates despite considerable infrastructure improvements.

Collectively, these LMIC studies underscore that the determinants of crash severity at intersections are deeply contextual, shaped by infrastructure provision, enforcement intensity, and the behavioural adaptation of road users to mixed-traffic conditions. While studies from high-income countries have consistently shown that geometric improvements such as medians, lane separation, and surface quality reduce crash frequency and severity [3,11,56], evidence from LMICs reveals that their effectiveness depends heavily on implementation quality and behavioural compliance [24,54,55]. In environments characterized by heterogeneous traffic and limited enforcement, measures effective in HICs such as wider medians or improved surface conditions may not yield comparable safety benefits and can even exacerbate risks through higher operating speeds or informal road use [23,53]. The growing body of LMIC-focused research therefore provides crucial evidence that infrastructure–behaviour interactions differ fundamentally from those observed in homogeneous traffic systems [7,29], underscoring the need for context-sensitive design and enforcement frameworks. Building on this knowledge, the present study seeks to expand this evidence base by examining how infrastructure features such as lane configuration, shoulders, median type, and pavement condition, together with environmental factors like lighting and time of day, shape crash severity patterns at intersections in Douala and Yaoundé, Cameroon, where mixed traffic and vulnerable road users prevail.

2.3. Modelling Approaches

Recent studies on crash severity have relied on a range of statistical and econometric approaches to examine the role of infrastructure and geometric design in shaping road safety outcomes. Early applications were dominated by count models, such as Poisson and negative binomial regression, which are well-suited for modelling crash frequencies. For example, Ref. [22] used these models to show that factors such as road gradient, curvature, and lane count significantly increase crash occurrences. Similarly, [57] applied negative binomial regression to Ethiopian roads, identifying a positive correlation between road length, used as a proxy for geometric design, and accident frequency. [17] also employed Poisson and multinomial logit regression models to study Abu Dhabi intersections, finding that geometric and operational conditions, including the number of lanes as a proxy for intersection configuration, directly influenced safety outcomes. More advanced extensions of count models have been explored as well [58], for instance, adopted a multivariate Poisson log-normal model to examine crashes at urban signalized intersection approaches in Nebraska, showing that arterial roads and lane configurations (right-turn, left-turn, and through lanes) increased both crash frequency and severity, though pavement condition and friction were not considered.

While count models are effective in capturing crash frequency, they are less suited to analysing injury severity, which is inherently ordinal in nature. For this reason, a large body of research has applied discrete choice models, particularly ordinal probit and logit approaches, to model the probability of different severity levels. [3] used ordinal probit models to study crash injury severity at signalized intersections in Florida, finding that lane count, medians, and right-turn channelization reduced severe injury risks, though pavement surface friction and condition were excluded. [3] extended this line of work to unsignalized intersections in Florida, applying two ordinal probit models that incorporated driver characteristics, intersection attributes, pavement surface type, and road conditions (e.g., dry or wet). Their results highlighted the role of left and right shoulder widths and turning lane counts in shaping severity, though pavement friction was again omitted. In a different context, [29] employed a generalized ordinal logit model to study crash severity at Bangladeshi intersections, identifying undivided roads, dry pavements, and rural settings as risk factors for severe injuries. Likewise, [56] analysed rural four-leg signalized intersections using ordinal probit models, concluding that tighter curves and higher speed limits increased severity, while wider medians and protected left-turn phases mitigated risks, though pavement friction was again not included.

More recently, researchers have turned to advanced econometric approaches to account for spatial effects, heterogeneity, and unobserved correlations. [59] applied a Bayesian spatial generalized ordered logit model with conditional autoregressive priors to study freeway crashes, showing that driver type, season, traffic volume and composition, emergency medical response time, and crash type significantly affected severity. They further demonstrated that vehicle type, time of day, weather condition, vertical grade, bridges, and traffic composition influenced thresholds between medium and severe crash levels. Such approaches allow for richer representation of spatial dependence and random effects than traditional logit or probit models.

However, binary probit and logistic regression models remain widely used for crash severity analysis due to their simplicity and ability to handle binary outcomes. These models are particularly effective when the focus is on distinguishing between severe versus non-severe crashes. Probit models provide marginal effects that quantify changes in crash probabilities, while logistic regression offers odds ratios that are more intuitive for policy applications. Both approaches have been applied to intersection safety, with [30] demonstrating their usefulness in quantifying the influence of crash-related attributes and geometric factors. Because it allows estimation of the likelihood of severe crashes while maintaining a direct interpretive link between explanatory variables and outcome probabilities the binary probit model alongside the binary logistic regression model are the most effective modelling approaches for our study. Unto this, and unlike ordered or machine learning models, the probit framework provides statistically interpretable coefficients and marginal effects, facilitating the identification of how specific geometric, traffic, and environmental factors influence crash severity. Recent studies have demonstrated that while machine learning techniques can outperform traditional models in predictive accuracy, they often operate as “black boxes” with limited interpretability and high data requirements [60,61]. For instance, [60] showed that handling data imbalance remains a major challenge in crash severity prediction, while [61] highlighted the trade-off between the power to predictive and explain in various machine learning models. Given that severe crashes are relatively rare and data are often unevenly distributed in developing regions, the binary probit model offers a more robust and transparent approach for hypothesis-driven analysis under such data conditions.

Taken together, the literature shows a progression from basic count models of crash frequency to more sophisticated severity models such as ordinal probit/logit and Bayesian spatial approaches. While these advanced models capture heterogeneity and spatial correlation, their application often requires extensive data and computational resources, which limits their practicality in many LMIC contexts. By contrast, binary probit and logistic regression models remain widely applied in crash severity studies because of their relative simplicity, interpretability, and ability to handle binary outcomes. Probit models provide marginal effects that quantify how explanatory variables shift the probability of severe crashes, while logistic regression offers odds ratios that are intuitive for policy and safety applications. Given the data environment and the objective of distinguishing between severe and non-severe crashes at intersections in mixed-traffic settings, this study employs both binary probit and logistic regression models to provide robust and interpretable insights into crash severity in LMICs.

3. Method

3.1. Models

Probit models, first introduced by [62], have been widely used in crash severity research as they allow the probability of an outcome to be expressed through a standard normal cumulative distribution [11]. In this study, a binary probit model was applied to examine the determinants of crash severity at intersections in Douala and Yaoundé.

Given the limited number of fatal crashes and the resulting imbalanced distribution across severity categories, the severity outcomes were aggregated into two categories levels; severe and non-severe, following the approach of [63]. This aggregation ensured model stability and reliable coefficient estimation, which could not have been achievable using ordinal framework that requires sufficient data across all categories. Severe crashes included all cases with possible injury, non-incapacitating injury, incapacitating injury, or fatality, while non-severe crashes were limited to property-damage-only (PDO) events. Accordingly, the dependent variable y was coded as:

• y = 0; if the crash is not severe (PDO)
• y= 1; if the accident is severe (injury or fatality)

The ordered probit formulation assumes that the probability of crash n falling into severity category j is:

$$\sum _{j=1}^j{P}_n\left(j\right)=F\left({\alpha }_j-{\beta X}_n,\theta \right), j=\mathrm{1,2},\dots \dots \dots , J-1$$

(1)

$$P_n\left(j\right)=1-\sum _{j=1}^{J-1}{P}_n\left(j\right)$$

(2)

where $P_n\left(j\right)$ is the probability that crash n belongs to category j; J is the number of categories (two in this case); α_j are threshold parameters; X_n is the vector of crash and road characteristics (e.g., lane count, shoulders, pavement condition, delineation); β are the estimated coefficients; F is the cumulative distribution function, assumed to be the standard normal distribution Φ, and θ is the shape parameter that controls F.

To interpret results, marginal effects were estimated. These measure the change in the probability of a severe crash associated with a one-unit change in an explanatory variable. For the probit model, the expected value of the response variable is:

E(Y) = Φ(X^Tβ)

(3)

and the marginal effect of a covariate is:

$${\displaystyle \frac{\partial \mathrm{E}\left(\mathrm{Y}\right)}{\partial \mathrm{X}}}= \left\{{\displaystyle \frac{\mathrm{d}\mathrm{F}\left({\mathrm{X}}^{\mathrm{T}}\mathsf{\beta }\right)}{\mathrm{d}\left({\mathrm{X}}^{\mathrm{T}}\mathsf{\beta }\right)}}\right\}\mathsf{\beta }=\mathrm{f}(\left({\mathrm{X}}^{\mathrm{T}}\mathsf{\beta }\right)\times \mathsf{\beta }=\mathsf{\varphi }\left({\mathrm{X}}^{\mathrm{T}}\mathsf{\beta }\right)\times \mathsf{\beta }$$

(4)

where Φ (⋅) is the standard normal probability density function. This step allowed us to quantify the influence of road and environmental factors on the likelihood of severe crashes.

In addition, logistic regression was applied as a complementary approach, as it has been extensively used to evaluate crash risk factors [64,65,66]. In this study, the dependent variable was specified in the same way as in the probit model (1 = severe crash, 0 = non-severe crash). The logistic model estimates the log-odds of a severe crash as:

$$\log \left(P\right)=\ln \left({\displaystyle \frac{P}{1-p}}\right)= {\beta }_0+ {\beta }_1{x}_1+\dots + {\beta }_i{x}_i$$

(5)

where P is the probability of a severe crash, x_i are the explanatory variables, and β_i are the coefficients. The odds ratio (OR) is then obtained as exp(β_i). In this study, the OR were used to rank the relative importance of different road and intersection characteristics. An OR greater than 1 indicated that an increase in the variable increased the odds of a severe crash, while an OR less than 1 indicated a reduction in severity risk.

Overall, the combined use of binary probit and logistic models provides a statistically rigorous, interpretable, and robust framework for analysing crash severity in the presence of limited data, as summarised in Figure 1. These models ensure comparability with prior research while avoiding the interpretability and data volume constraints associated with more complex machine learning approaches.

3.2. Analysis with SPSS

SPSS version 27 (Statistical Package for the Social Sciences), developed by IBM, was used for model implementation. The software supports a wide range of statistical techniques, including binary probit and logistic regression, and is widely applied in transportation safety research for analysing categorical and continuous variables.

For this study, the crash dataset was imported into SPSS in compatible formats (.sav or .csv) containing the binary dependent variable Severity (0 = non-severe, 1 = severe) and predictor variables (e.g., shoulder presence, number of lanes, lane separation type, and road condition). Data cleaning involved addressing missing values through imputation or exclusion and recoding categorical variables into analysable formats.

In the binary probit analysis, SPSS produced output tables summarizing model performance and parameter estimates:

SPSS generated output tables, including:

• Model Fit: log-likelihood and pseudo-R² statistics (e.g., McFadden’s R²), with a significant chi-square test (p-value < 0.05) confirming model validity.
• Parameter Estimates: coefficients (β) and p-values, indicating the direction and significance of each predictor (e.g., a positive coefficient for Alcohol, β = 0.75, p-value = 0.01\β = 0.75, p-value = 0.01 β = 0.75, p-value = 0.01, reflects a higher likelihood of severe crashes).
• Marginal Effects: calculated manually from the estimated coefficients to interpret how changes in predictors affect the probability of a severe crash (e.g., a one-unit increase in Speed raises the probability of a severe crash by X%).

Model validity was further assessed by testing for multicollinearity using Variance Inflation Factors (VIF < 5) through the Analyse > Regression > Linear option, and by inspecting residuals to identify potential outliers or misspecification.

The binary logistic regression was implemented using similar procedures. In addition to coefficients and significance levels, odds ratios (OR = exp(β_i)) were computed to quantify the relative effect of each predictor on the likelihood of severe crashes.

4. Data

4.1. Study Area and Context

This study focuses on Douala and Yaoundé, Cameroon’s largest cities and primary economic hubs. Both cities are undergoing rapid urbanization, leading to increasing travel demand, high traffic volumes, and recurrent congestion. These conditions, coupled with mixed traffic flows dominated by cars, motorcycles, and pedestrians, contribute to a high incidence of road crashes.

Travel demand in the two cities is substantial but varies in composition. In Yaoundé, approximately 8 million daily trips are generated, with cars accounting for 48.4%, pedestrians for 35%, motorcycles for 14.2%, and buses for 2.4% [67]. In Douala, about 6.9 million daily trips are generated, of which 17% are by car, 35% by pedestrians, 44% by motorcycles, 1% by bus, and 3% by other modes [26].

Despite these estimates, comprehensive and regularly updated traffic data remain unavailable. Permanent traffic sensors are lacking, and systematic data collection at intersections or along corridors is rarely conducted.

The overall transport and mobility system is inefficient, characterized by limited public transport options, irregular and uncomfortable services, severe congestion, and rising pollution levels [67]. Road safety is a pressing concern, with crashes causing thousands of deaths and serious injuries annually in both cities [26,67]. The road network consists of arterial roads, secondary roads, and local streets, but official statistics on traffic flow, vehicle composition, and intersection characteristics are largely missing.

4.2. Data Collection and Cleaning

Four years and six months of crash data from Douala and Yaoundé were used in this study, covering the period from January 2021 to December 2022 for Douala and from January 2023 to June 2025 for Yaoundé. The data were obtained from the General Delegation for National Security (DGSN) of Cameroon, specifically from the central police headquarters of both cities. Crashes were reported to the police, who prepared official crash reports that were subsequently compiled into Excel spreadsheets by road safety experts.

A two-stage approach was adopted to address missing values. First, missing information on intersection location was completed by searching for the intersection names provided in the police reports on web-based mapping platforms such as Google Maps and Google Earth, from which geographic coordinates were extracted. Second, missing details on road characteristics, such as the presence of shoulders or delineation, were retrieved from the same sources. Road variables were then categorized for analysis: some were binary (e.g., presence or absence of shoulders), while others were classified into multiple categories (e.g., shoulder width).

After removing non-essential records, specifically crashes that occurred outside intersections or at locations that could not be clearly identified, the final dataset comprised 1242 crashes. The dataset was divided into two subsets: severe crashes and non-severe crashes.

4.3. Data Description

The dataset comprised 1242 intersection crashes, of which 173 (13.9%) were classified as severe and 1069 (86.1%) as non-severe. Severe crashes included possible, minor, incapacitating, and fatal injuries, while non-severe crashes consisted of property-damage-only (PDO) events. Overall, PDO crashes represented 86.1% of the sample, while injury and fatal crashes accounted for 13.9%.

In addition to severity, the dataset included environmental and road-related variables. For example, crash time was categorized by time of day, and street lighting conditions were coded as present or absent, with 15.3% of crashes occurring at intersections without lighting. Each record in the dataset represented a unique intersection crash, with its corresponding attributes systematically classified on a categorical or binary basis.

In SPSS, the dependent variable was defined as crash severity (1 = severe, 0 = non-severe), while the independent variables were those listed in

Table 1. Data attributes.

Road Facilities	Observable Features	Quantity
Roadsides	Paved shoulder driver side	none [1], narrow (0 to 1 m) [2], medium (1 to 2.4 m) [3], wide (>2.4 m) [4]
	Paved shoulder passenger side	none [1], narrow (0 to 1 m) [2], medium (1 to 2.4 m) [3], wide (>2.4 m) [4]
	Shoulder rumble strip	present [1], absent [2]
Midblock	Median type	center line [1], central hatching [2], physical separation [3].
	Number of lanes	number of lanes per direction of traffic: 1 × 1 [1], 2 × 2 [2], 3 × 3 [3] means one lane per direction of traffic
	Road width	narrow: 0 to 2.75 m [1], medium: between 2.75 m and 3.25 m [2], large: >3.25 m [3]
	Road condition	poor [1], medium [2], good [3]
	Delineation	poor [1], good [2]: refers to lane delimitation markings (horizontal signalization)
Facilities for VRU	Crosswalk	absent [1], present (but has to be also visible) [2]
	Sidewalk	absent [1], present (but has to be in good condition visible) [2]
Environmental conditions	Street lighting	present [1]
		absent [2]
	Time of the day	Late night (00:00–05:59) [1]
		Morning peak hours (06:00–09:59) [2]
		Midday (10:00–14:59) [3]
		Afternoon peak hours (15:00–17:59) [4]
		Evening (18:00–20:59) [5]
		Night (21:00–23:59) [6]

.

Table 2. Summary statistics.

Variables	Total (%)	Non Severe Crashes (%)	Severe Crashes (%)
Roadside
Paved shoulder driver side
none [1]	496 (39.9%)	421 (39.4%)	75 (43.4%)
narrow (0 to 1 m) [2]	42 (3.4%)	39 (3.6%)	3 (1.7%)
medium (1 to 2.4 m) [3]	150 (12.1%)	133 (12.4%)	17 (9.8%)
wide (>2.4 m) [4]	554 (44.6%)	476 (44.5%)	78 (45.1%)
Paved shoulder driver side
none [1]	494 (39.8%)	418 (39.1%)	76 (43.9%)
narrow (0 to 1 m) [2]	40 (3.2%)	37 (3.5%)	3 (1.7%)
medium (1 to 2.4 m) [3]	149 (12%)	133 (12.4%)	16 (9.2%)
wide (>2.4 m) [4]	559 (45%)	481 (45%)	78 (45.1%)
Shoulder rumble strip
present [1]	386 (31.1%)	234 (24.5%)	40 (23.1%)
absent [2]	856 (68.9%)	723 (75.5%)	133 (76.9%)
Midblock
Median type
center line [1]	321 (25.8%)	301 (28.2%)	20 (11.6%)
central hatching [2]	502 (40.4%)	423 (39.6%)	79 (45.7%)
physical separation [3]	419 (33.7%)	345 (32.3%)	74 (42.8%)
Number of lanes
1 × 1 [1]	681 (54.8%)	578 (54.1%)	103 (59.5%)
2 × 2 [2]	485 (39%)	420 (39.3%)	65 (37.6%)
3 × 3 [3]	76 (6.1%)	71 (6.6%)	5 (2.9%)
Road condition
poor [1]	61 (4.9%)	57 (5.3%)	4 (2.3%)
medium [2]	390 (31.4%)	337 (31.5%)	53 (30.6%)
good [3]	791 (63.7%)	675 (63.1%)	116 (67.1%)
Delineation
poor [1]	450 (36.2%)	388 (36.3%)	62 (35.8%)
good [2]	792 (63.8%)	681 (63.7%)	111 (64.2%)
Street lighting
present [1]	1052 (84.7%)	903 (84.5%)	149 (86.1%)
absent [2]	190 (15.3%)	166 (15.5%)	24 (13.9%)
VRU facilities
Crosswalk
absent [1]	446 (35.9%)	384 (35.9%)	62 (35.8%)
present (but has to be also visible) [2]	796 (64.1%)	685 (64.1%)	111 (64.2%)
Sidewalk
absent [1]	209 (16.8%)	184 (17.2%)	25 (14.5%)
present (but has to be in good visible condition) [2]	1033 (83.2%)	885 (82.8%)	148 (85.5%)
Environmental conditions
Street lighting
present [1]	1052 (84.7%)	903 (84.5%)	149 (86.1%)
absent [2]	190 (15.3%)	166 (15.5%)	24 (13.9%)
Time of the day
Late night/early morning (00:00–05:59) [1]	69 (5.6%)	42 (3.9%)	27 (15.6%)
Morning peak hours (06:00–09:59) [2]	205 (16.5%)	172 (16.1%)	32 (18.5%)
Midday (10:00–14:59) [3]	390 (31.4%)	353 (33%)	37 (21.4%)
Afternoon peak hours (15:00–17:59) [4]	283 (22.8%)	250 (23.4%)	33 (19.1%)
Evening (18:00–20:59) [5]	208 (16.7%)	185 (17.3%)	23 (13.3%)
Night (21:00–23:59) [6]	87 (7%)	67 (6.3%)	21 (12.1%)

summarizes the descriptive statistics of the dataset.

5. Results

5.1. Binary Probit Model Results

The initial binary probit model encountered quasi-complete separation, reflected in a failure of maximum likelihood estimation (MLE) to converge. This was due to extremely large coefficients and standard errors for Paved shoulder (driver side) and Paved shoulder (passenger side) (e.g., β = 5.641, Std. Error = 11,937.38). These variables likely contained categories where severity outcomes could be perfectly or nearly perfectly predicted. The issue was compounded by the large number of predictors (34 degrees of freedom) relative to the limited number of severe crashes (n = 173), with some categories containing very few observations (e.g., narrow shoulders: 3.4% driver side, 3.2% passenger side).

To resolve this problem, the paved shoulder variables were excluded from the model. A revised estimation achieved convergence and improved overall model fit. Subsequent analysis further eliminated four non-significant infrastructural features (delineation, street lighting, crosswalks, and sidewalks) to enhance model parsimony. The resulting specification constitutes the final binary probit model, with stable parameter estimates and improved information criteria (AIC = 321.78). The revised analysis achieved convergence and provided improved fit, as shown in

Table 3. Goodness of fit.

	Value	df	Mean Deviance
Deviance	167,660	103	1628
Scaled Deviance	167,660	103
Pearson Chi-Square	167,429	103	1626
Scaled Pearson Chi-Square	167,429	103
Log Likelihood	−147,890
Akaike’s Information Criterion (AIC)	321,780
Finite Sample Corrected AIC (AICC)	322,076
Bayesian Information Criterion (BIC)	388,398
Consistent AIC (CAIC)	401,398

df: degree of freedom.

.

Following variable exclusion, the model’s AIC improved considerably (from 481.91 to 321.78), and deviance measures near unity (≈1.63) indicate acceptable model fit and residual distribution. The refined structure avoids overparameterization while retaining key explanatory power. The omnibus likelihood ratio test (p-value < 0.05) confirms that the retained predictors collectively improve model fit relative to the intercept-only specification (

Table 4. Omnibus test results.

Likelihood Ratio Chi-Square	df	p-Value
86,152	12	<0.05

).

The test of model effects identified which predictors significantly influenced crash severity (

Table 5. Test of model effects.

	Likelihood Ratio Chi-Square	df	p-Value
(Intercept)	121.648	1	<0.05
Time interval	42.996	5	<0.05
Midblock: Centre line [1], central hatching [2], physical separation [3]	20.004	2	<0.05
Number of lanes	12.607	2	<0.05
Road condition Poor [1], medium [2], good [3]	4.029	2	0.133
Shoulder rumble strip Present [1], not present [2]	2.637	1	0.104

).

From Table 5, significant determinants of crash severity include time interval, midblock configuration, and number of lanes (p-value < 0.05).

Shoulder rumble strip and road condition show marginal effects (p-value ≈ 0.10–0.14), suggesting minor or context-dependent associations.

Detailed estimates, including coefficients, confidence intervals, Wald statistics, p-values, and marginal effects, are presented in

Table 6. Binary probit estimates and marginal effect.

Predictor	Category	β	Std. Error	95% CI		Wald Chi-Square	p-Value	Marginal Effect (%)	Interpretation
				Lower	Upper
Intercept	-	−1.258	0.2726	−1.792	−0.723	21.282	<0.05	-	Baseline probability of a severe crash is approximately 14.8%.
Time Interval	Late night/early morning (00:00–05:59) [1]	0.562	0.2168	0.137	0.987	6.716	<0.05	13.5	Late-night/early-morning hours significantly increase the probability of severe crashes by 13.5%, reflecting low visibility, fatigue, or alcohol involvement.
Time Interval	Morning peak (06:00–09:59) [2]	−0.208	0.1863	−0.573	0.158	1.241	0.265	−4.7	Morning peak slightly lowers severity risk by 4.7%, though not statistically significant—likely due to congestion reducing speeds.
Time Interval	Midday (10:00–14:59) [3]	−0.485	0.1757	−0.830	−0.141	7.622	<0.05	−10.2	Midday period reduces severe-crash probability by 10.2%, possibly from better visibility and moderate traffic flow.
Time Interval	Afternoon peak (15:00–17:59) [4]	−0.426	0.1812	−0.782	−0.071	5.536	<0.05	−9.5	Afternoon peak decreases severity by 9.5%, likely reflecting slower, congested conditions.
Time Interval	Evening (18:00–20:59) [5]	−0.473	0.1914	−0.849	−0.098	6.123	<0.05	−11.1	Early-evening crashes are 11.1% less likely to be severe than late-night ones.
Time Interval	Night (21:00–23:59) [6]	0	-	-	-	-	-	-	Reference category.
Midblock Centreline	Centre line [1]	−0.670	0.1503	−0.964	−0.375	19.835	<0.05	−3.9	Central hatching or separation reduces severe-crash probability by 3.9% compared to plain centre-line sections.
Midblock Centreline	Central hatching [2]	−0.196	0.1326	−0.456	0.064	2.192	0.139	−14.8	Provides a 14.8% lower probability of severe crashes, though effect is not statistically significant.
Midblock Centreline	Physical separation [3]	0	-	-	-	-	-	-	Reference; safest configuration.
Number of lanes	One lane [1]	0.932	0.2647	0.413	1.451	12.393	<0.05	21.5	One-lane roads increase severe-crash probability by 21.5%, largely due to head-on exposure and limited manoeuvring space.
Number of lanes	Two lanes [2]	0.748	0.2557	0.247	1.249	8.557	<0.05	9.1	Two-lane segments show a 9.1% higher probability of severe crashes compared to multilane roads.
Number of lanes	3+ lanes [3]	0	-	-	-	-	-	-	Reference; wider cross-sections mitigate severity risk.
Road Condition	Poor [1]	−0.421	0.2678	−0.946	0.104	2.471	0.116	−9.4	Poor road surfaces reduce severity probability by 9.4% (non-significant), possibly due to lower driving speeds.
Road Condition	Medium [2]	−0.183	0.1154	−0.409	0.044	2.506	0.113	−4.6	Medium-quality roads show a 4.6% reduction in severe-crash probability (non-significant).
Road Condition	Good [3]	0	-	-	-	-	-	-	Reference.
Shoulder Rumble Strip	Present [1]	−0.201	0.1235	−0.443	0.042	2.637	0.104	−16.9	Presence of rumble strips lowers severe-crash likelihood by 16.9%, reflecting their early-warning benefit (marginal significance).
Shoulder Rumble Strip	Not present [2]	0	-	-	-	-	-	-	Reference.

.

5.2. Binary Logistic Regression Results

To complement the probit analysis, a binary logistic regression model was estimated, providing the advantage of interpreting results through odds ratios (OR). The overall model fit statistics are summarized in

Table 7. Binary logistic model summary.

Step	−2 Log likelihood	Cox & Snell R2	Nagelkerke R2
1	918.452 a	0.066	0.118

^a. Estimation terminated at iteration number 5 because parameter estimates changed by less than 0.001.

.

The refined model achieved a −2 Log Likelihood of 918.452, with Cox & Snell R² = 0.066 and Nagelkerke R² = 0.118, see Table 7, indicating modest explanatory power typical of cross-sectional crash severity models. Although these pseudo-R² values were slightly lower than those of the initial model (0.068 and 0.122), such minor reductions do not signify a decline in model quality, as pseudo-R² indices in logistic regression are not direct measures of variance explained and cannot be interpreted in the same way as the R² in linear regression [28,68,69]. Rather, model evaluation should rely on likelihood-based criteria, the statistical significance of predictors, and overall theoretical coherence, which together provide a more robust basis for determining model adequacy [27,70].

Furthermore, model selection should prioritize parsimony and interpretability over marginal differences in pseudo-R², since simpler models with statistically meaningful predictors often yield more generalizable and policy-relevant insights [27,71]. The Omnibus Test of Model Coefficients (χ² = 84.277, df = 12, p-value < 0.05) confirmed that the retained predictors significantly improved model fit relative to the intercept-only model, reinforcing the refined model’s statistical validity and theoretical soundness. See

Table 8. Omnibus test model coefficients.

	Chi-Square	df	p-Value
Step	84.277	12	<0.05
Block	84.277	12	<0.05
Model	84.277	12	<0.05

.

The regression coefficients, Wald statistics, p-values, odds ratios (OR), and their confidence intervals are presented in

Table 9. Binary logistic estimates.

	β	S.E.	Wald	df	p-Value	OR	95% C.I. for OR
							Lower	Upper
Time interval [6]	-	-	42.692	5	<0.05	-	-	-
Time interval [1]	0.943	0.369	6.538	1	0.011	2.568	1.246	5.292
Time interval [2]	−0.321	0.326	0.969	1	0.325	0.725	0.383	1.375
Time interval [3]	−0.864	0.315	7.550	1	0.006	0.421	0.227	0.781
Time interval [4]	−0.734	0.322	5.188	1	0.023	0.480	0.255	0.903
Time interval [5]	−0.829	0.343	5.824	1	0.016	0.437	0.223	0.856
Shoulder rumble strip Present [1]	−0.355	0.228	2.427	1	0.119	0.701	0.448	1.096
Shoulder rumble strip Present [2]	-	-	-	-	-	-	-	-
Midblock: physical separation [3]	-	-	17.657	2	<0.05	-	-	-
Midblock: central line [1]	−1.221	0.291	17.612	1	<0001	0.295	0.167	0.522
Midblock: central hatching [2]	−0.321	0.240	1.780	1	0.182	0.726	0.453	1.162
Number of lanes [3]	-	-	10.709	2	0.005	-	-	-
Number of lanes [1]	1.683	0.521	10.437	1	0.001	5.380	1.938	14.935
Number of lanes [2]	1.361	0.504	7.289	1	0.007	3.902	1.452	10.484
Road condition: good [3]	-	-	4.598	2	0.100	-	-	-
Road condition: Poor [1]	−0.878	0.550	2.551	1	0.110	0.416	0.142	1.221
Road condition: medium [2]	−0.359	0.209	2.955	1	0.086	0.699	0.464	1.052
Constant	−2.188	0.531	17.009	1	0.000	0.112	-	-

.

5.3. Comparative Analysis of Probit and Logistic Models

Both models consistently identified; time interval, midblock centreline, and number of lanes as significant predictors of crash severity (p-value < 0.05), while shoulder rumble strip and poor road condition appeared marginally significant (p-value < 0.10). The probit model yielded marginal effects (Table 6), quantifying the change in the probability of a severe crash; for instance, single-lane roads were associated with a 21.5% higher probability of severe outcomes. In contrast, the logistic model produced odds ratios (Table 9), which are more intuitive for policy communication. For example, single-lane roads increased the odds of severity by 5.38 times relative to three-lane approaches.

Although the refined models exhibited slightly lower pseudo-R² values than the initial specifications, the differences were minimal and statistically inconsequential. More importantly, the refined models demonstrated improved theoretical coherence, reduced redundancy, and enhanced interpretability, which are key criteria for model selection [27,28,72]. The convergence of results between the probit and logistic frameworks strengthens confidence in the robustness of the findings; logistic regression provides clear, policy-relevant risk multipliers, while the probit specification offers probabilistic insights that enhance understanding of the underlying crash severity dynamics. Collectively, the complementary evidence from both approaches reinforces the validity and practical applicability of the refined model.

6. Discussions

The binary probit and logistic regression analyses provide complementary insights into the determinants of crash severity at intersections in Yaoundé and Douala. Both models consistently identified time interval, midblock configuration, and number of lanes as significant predictors (p-value < 0.05), with shoulder rumble strips and poor road conditions emerging as marginally significant (p-value < 0.10). The probit model quantified changes in the probability of severe crashes, whereas the logistic model provided odds ratios that are more intuitive for safety policy and intervention design [27,69]. This dual modelling approach thus combines the interpretive strength of marginal probabilities with the communicative clarity of odds ratios, bridging empirical prediction with policy translation, an essential step for evidence-based safety management in rapidly motorising LMIC contexts.

Time of day emerged as a major factor. The probit model showed that late-night/early-morning crashes (00:00–05:59) increased severity probability by 13.5%, while midday, afternoon, and evening periods reduced it by 10.2%, 9.5%, and 11.1%, respectively. The logistic model confirmed this, showing late-night crashes were 2.57 times more likely to be severe, while midday, afternoon, and evening crashes reduced odds by 58%, 52%, and 56%, respectively. These findings reinforce global evidence linking night-time crashes with higher severity due to reduced visibility, fatigue, or risky behaviours such as speeding and alcohol use [12,31]. Comparable studies in HICs [11,73] also report elevated night-time risk, yet LMIC settings often experience amplified effects due to weaker enforcement and limited illumination, highlighting how contextual risk multipliers shape crash severity outcomes. This pattern suggests that targeted night-time enforcement and roadway illumination programs could yield measurable reductions in severe crash probability, directly translating the estimated marginal effects into operational safety strategies.

Midblock design also significantly influenced severity. The probit model indicated that centre lines reduced severity probability by 3.9% compared to physical separation, while central hatching reduced it by 14.8% (though not statistically significant). Logistic results echoed this, with centre lines reducing severity odds by 71% (OR = 0.295, p-value < 0.05). This suggests that less complex road designs may lower severity by reducing speeds and simplifying manoeuvres, consistent with prior studies [22]. Similar relationships were found by [11] in Florida and [70] in Ghana where simplified intersection geometry and delineation improved crash outcomes. These parallels reinforce that while geometric countermeasures are universally beneficial, their effectiveness depends on implementation fidelity and driver adaptation to mixed-traffic conditions. This also contributes theoretically by supporting the geometric risk–behavioural adaptation nexus, which posits that design simplicity can modulate human error probability under mixed-traffic dynamics.

The number of lanes was the most influential geometric factor. Probit results showed single-lane and two-lane roads increased severity probability by 21.5% and 9.1%, respectively, compared to three-lane roads. Logistic regression confirmed this, showing 5.38 and 3.90 times higher odds of severe crashes on single- and two-lane roads. These results align with earlier research demonstrating that narrower facilities constrain manoeuvring space and exacerbate collision impact [11,42]. In LMIC contexts, such as Ghana [70] and Ethiopia [31], this pattern is often compounded by mixed traffic and poor delineation, further elevating crash risk. In practical terms, the 21.5% marginal increase in severity probability offers a quantifiable design threshold for prioritising lane expansion or shoulder widening at high-risk intersections. These findings suggest that intersection expansion or lane channelisation could be cost-effective interventions, particularly when guided by marginal effect estimates that quantify how each geometric constraint translates into crash severity probability.

Shoulder rumble strips and road conditions showed marginal protective effects. Rumble strips reduced severity probability by 16.9% (probit) and odds by 29.9% (OR = 0.701), while poor road conditions reduced severity probability by 9.4% (probit) and odds by 58.4% (OR = 0.416). These findings suggest rumble strips may help drivers correct deviations, while poor surfaces might encourage cautious driving. However, the weak significance signals the need for more data to confirm these trends [41]. Comparable protective effects of rumble strips have been reported in the U.S. [74], while contrasting findings in Malaysia [41] indicate that road surface quality interacts strongly with enforcement levels and driver caution. Hence, these marginally significant variables may represent latent behavioural adaptations rather than direct engineering effects. Nonetheless, their inclusion broadens the model’s explanatory scope, suggesting that even subtle tactile feedback mechanisms can induce corrective driver behaviour under uncertain roadway conditions.

By contrast, delineation, street lighting, crosswalks, and sidewalks were not statistically significant. This may reflect limited variability in the dataset (e.g., most intersections had lighting) or insufficient statistical power due to the imbalance between severe and non-severe crashes (13.9% severe). Similar findings have been noted in LMIC settings, where the effects of environmental factors are highly context-dependent [51]. In contrast, HIC studies often report stronger effects for such [11,73] suggesting that contextual moderation through enforcement, driver compliance, and infrastructure quality should be modelled explicitly in future research. This underscores that environmental variables, particularly lighting condition and time of day, operate synergistically: while lighting was not directly significant, its influence is captured indirectly through the pronounced night-time effect, reinforcing the critical role of visibility in shaping crash outcomes.

From a theoretical perspective, the convergence of probit and logit estimates supports behavioural adaptation and geometric risk theory, suggesting that crash severity results from the joint influence of infrastructure-induced exposure and driver decision-making constraints. The quantified marginal effects offer direct translation into safety design, e.g., a 21.5% increase in severity probability for single-lane approaches could justify re-channelisation or added shoulders at high-volume intersections. This study therefore advances theoretical understanding by empirically demonstrating how structural road geometry interacts with temporal exposure to produce non-linear severity effects in mixed-traffic systems.

Policy-wise, these findings reinforce the importance of integrating marginal effect outputs into prioritisation tools for intersection upgrades. Decision-makers can use the probability-based estimates to rank sites by expected severity reduction potential rather than frequency alone a shift consistent with the Safe System approach promoted in both HIC and LMIC policy frameworks. Such translation of model outputs into design priorities marks a methodological contribution, as it demonstrates how econometric evidence can directly inform spatially targeted interventions and cost-effective policy sequencing.

From a modelling standpoint, the refined model’s improved AIC (321.78 vs. 481.91) and acceptable deviance ratio (≈1.63) illustrate that simplification enhanced fit while maintaining theoretical coherence. The logistic model’s modest pseudo-R² values (Cox & Snell R² = 0.066, Nagelkerke R² = 0.118) are consistent with expectations for crash-severity data, confirming that explanatory clarity and parsimony were prioritised over overfitting. Future model improvements could incorporate hierarchical structures, exposure measures, or Bayesian updating to capture unobserved heterogeneity and strengthen predictive accuracy. In addition, testing interaction terms (e.g., time interval × number of lanes) could reveal compound risk dynamics, while adopting mixed-effects or penalised regression frameworks may enhance stability in small or imbalanced samples.

Overall, the consistency across models reinforces confidence in the key results while underscoring the need for richer datasets and advanced methods to capture the complex interplay of infrastructure, behavioural, and environmental factors. By combining robust model specification with interpretable marginal effects, this study contributes a replicable framework for LMIC crash-severity analysis that informs both academic theory and practical design guidance. In essence, it bridges the gap between quantitative inference and actionable safety design, aligning empirical modelling with real-world decision-making.

7. Conclusions

This study analysed 1242 intersection crashes in Douala and Yaoundé (2021–2025) using binary probit and logistic regression models to identify determinants of crash severity. Both models consistently highlighted time interval, midblock design, and number of lanes as significant predictors, with shoulder rumble strips and road condition showing marginal effects. The consistency between models strengthens the robustness of the findings and demonstrates that even with modest pseudo-R² values, well-specified models can yield meaningful and policy-relevant insights. Together, these models provide complementary interpretive and predictive perspectives where the probit model quantifies risk changes in probability terms and the logistic model contextualises them within policy-relevant odds ratios.

The results demonstrate that single-lane roads substantially increase severity risk (probit: 21.5% probability; logistic: OR = 5.38), while late-night and early-morning crashes (00:00–05:59) are particularly dangerous (probit: 13.5%; logistic: OR = 2.57). Centre lines were associated with safer outcomes compared to physical separation (probit: –3.9%; logistic: OR = 0.295). Rumble strips and poor road conditions also appeared to reduce severity, although these effects were only marginally significant (probit: –16.9%; logistic: OR = 0.701 and probit: –9.4%; logistic: OR = 0.416, respectively) and require further study. Environmental factors, specifically lighting condition and time of day emerged as critical contextual moderators of severity outcomes. While lighting itself was statistically insignificant, its interplay with night-time crashes indicates that inadequate illumination magnifies crash severity by reducing visibility and driver response time, reinforcing the importance of continuous lighting and targeted night-time enforcement.

Collectively, these results highlight that geometric design and temporal exposure remain the most influential determinants of crash severity, yet their impacts are modulated by environmental quality. Poor illumination, uneven surfaces, and faded markings likely exacerbate crash risks during high-vulnerability periods such as late-night travel. This aligns with prior LMIC studies [31,70,75] and indicates that environmental improvements, better lighting, resurfacing, and delineation maintenance should complement geometric and behavioural interventions. Thus, addressing temporal and geometric risks without improving the environmental context may yield only partial safety benefits.

These findings highlight the central role of road geometry and temporal factors in shaping crash outcomes in LMIC urban contexts, while also pointing to concrete strategies for intervention. From a policy perspective, translating the estimated marginal effects into design priorities offers clear pathways for action. Road widening and traffic calming on narrow roads, improved night-time visibility, and the adoption of simpler midblock designs such as centre lines emerge as practical measures to reduce crash severity. The modest yet consistent protective effects of rumble strips suggest that low-cost tactile and auditory countermeasures can meaningfully mitigate driver error. By translating statistical results into actionable guidance, this study provides policymakers in Cameroon and comparable LMICs with evidence-based directions for reducing crash severity. Moreover, strengthening data systems and advancing modelling approaches remain critical to developing more precise, context-sensitive strategies that align with global road safety goals. This synthesis of marginal effects and model outcomes thus offers both strategic and operational contributions to road safety management.

The study also has limitations. The probit model showed signs of overdispersion, and the logistic model’s modest explanatory power (Nagelkerke R² = 0.118) suggests unobserved factors such as vehicle speed, driver behaviour, and traffic volume remain influential. Additionally, environmental exposure variables such as rainfall intensity, illumination quality, and intersection visibility were not directly measured, which may have understated their influence on crash severity. Data imbalance (13.9% severe crashes) constrained model sensitivity, especially for rare categories, while quasi-complete separation required the exclusion of paved shoulder variables. Nonetheless, these limitations point to promising avenues for future model enrichment rather than undermining the interpretive strength of current findings.

From a methodological standpoint, the refined model’s parsimony over the initial model indicates that simplified structures with theoretically grounded predictors may enhance interpretability and external validity, even at the cost of marginally lower pseudo-R² values [27,71,76]. This emphasizes that model refinement should prioritise explanatory clarity and policy applicability over raw fit metrics. Such parsimony ensures that model outputs remain actionable, scalable, and transferable across urban LMIC contexts.

Future research should integrate richer datasets including driver, vehicle, and exposure variables, apply interaction terms (such as time × lanes), and adopt advanced methods such as penalized regression or machine learning to improve predictive accuracy [30,51,77]. Incorporating continuous environmental indicators (e.g., light intensity, surface friction, and weather data) could further clarify how environmental degradation amplifies or mitigates crash severity risks. In addition, it would also be valuable to integrate traffic conflict analysis in place of, or alongside, crash data for earlier detection of risk patterns, supporting a proactive approach to safety analysis and complementing traditional crash-based modelling. These extensions will deepen both theoretical understanding and practical usability of crash-severity models in resource-limited contexts.

This study contributes both empirical and methodological value: empirically, by quantifying how geometric and temporal features interact with environmental conditions to influence crash severity; and methodologically, by demonstrating that balanced probit–logit modelling provides interpretable, transferable insights even in data-constrained LMIC contexts. Ultimately, the study provides a replicable analytical framework that advances both theory and practice linking statistical modelling to targeted, evidence-based road safety design and policy implementation.