Analyzing roadway factors influencing the severity of non-motorized road users' crashes in Charlotte, North Carolina, USA

1 Introduction and literature review

Pedestrian and bicyclist fatalities represent a persistent and growing challenge in urban transportation planning. In 2022, 7,522 pedestrians and 1,084 bicyclists were killed in motor vehicle crashes on U.S. public roadways, with approximately 67,000 pedestrians and 46,195 bicyclists sustaining non-fatal injuries (IIHS-HLDI, 2025). Collectively, these vulnerable road users comprise about 20% of all traffic-related fatalities, despite representing a small fraction of total trips, roughly 1% of trips for bicyclists and less than 12% for pedestrians (NHTSA, n.d.). The economic burden of bicycle-related injuries and deaths alone exceeds $23 billion annually, encompassing healthcare costs, lost productivity, and reduced quality of life (CDC, 2023a). Bicyclist injury prevention is likewise treated as a public health priority at the national level, with emphasis on safer riding environments and visibility (CDC, 2023b).

Despite this well-documented risk, understanding precisely which factors drive crash severity rather than crash occurrence remains an active area of inquiry. Prior research has examined a wide range of contributing variables, broadly categorized into human, environmental, and infrastructure factors (Noland and Quddus, 2004; Lord and Mannering, 2010; Shah et al., 2021; Chen, 2015). Human factors include motorist and non-motorist behavior such as distraction, failure to yield, and non-compliance with traffic controls and regulatory requirements (Guo et al., 2017). Environmental factors encompass lighting, weather, and surface conditions. Infrastructure factors include roadway geometry, lane configuration, the presence of dedicated bicycle facilities, and traffic control type. Road geometry plays a particularly consequential role: intersections, curves, and high-speed arterials without pedestrian infrastructure are consistently associated with elevated crash severity (Pulugurtha et al., 2007). Research has demonstrated that divided roadways with higher vehicle speeds are disproportionately dangerous for non-motorized users, while dedicated bicycle lanes and roundabouts reduce both crash frequency and severity (Rothenberg et al., 2016; Wu et al., 2019). Motorist behavior, especially speeding, distracted driving, and failure to yield, further amplifies the severity risk, particularly on roads with higher traffic volumes (Noland and Quddus, 2004; Jacobsen and Rutter, 2012). Lighting and visibility conditions also significantly influence crash outcomes. Low-light environments are associated with higher crash severity, as reduced visibility limits both driver response time and pedestrian/cyclist detectability (Retting et al., 2003). Weather conditions such as rain, snow, and ice reduce traction and visibility, compounding these risks (Jurewicz et al., 2016). Studies also identify demographic vulnerabilities: older pedestrians and bicyclists tend to experience more severe injuries due to greater physiological fragility (Javid and Sadeghvaziri, 2023; Sadeghvaziri et al., 2023). Taken together, this body of work establishes that human, environmental, and infrastructure factors each shape crash severity. What remains comparatively underdeveloped is research that weighs these domains against one another within a single city-scale severity model, and that treats crashes lacking a documented contributing factor as an analytic category in their own right rather than discarding them. These are the two gaps the present study addresses. Methodologically, much prior severity work relies on disaggregated multi-category models whose estimates become unstable when fatal and serious-injury cells are sparse, and tends to report predictors in isolation rather than weighing behavioral against environmental contributions; more recent work has begun to formalize behavioral pathways in bicycle–motor vehicle crashes by integrating machine learning with path analysis (Lu et al., 2022), though such approaches foreground behavioral mechanisms rather than contrasting them against built-environment predictors; the present study responds to both limitations through a policy-aligned binary specification that directly contrasts the two domains.

Recent North Carolina research has advanced the modeling of vulnerable-road-user crash severity using flexible, heterogeneity-aware frameworks. Abdulrazaq and Fan (2024) applied a seasonally constrained random-parameters approach to capture temporal dynamics in pedestrian injury severity; subsequent work examined seasonal instability in the determinants of vulnerable-road-user crashes statewide (Abdulrazaq and Fan, 2025a) and developed a priority-based, multi-level heterogeneity framework to identify high-risk locations across space and time (Abdulrazaq and Fan, 2025b). These studies foreground unobserved heterogeneity and temporal or spatial instability in statewide data. The present study is complementary but methodologically and substantively distinct: it concentrates on a single, rapidly urbanizing city (Charlotte); it integrates behavioral and built-environment predictors within a parsimonious, interpretable binary framework aimed directly at municipal Vision Zero decision-making; and it treats crashes lacking a documented contributing factor as an explicit analytic category rather than excluding them. In doing so it complements the statewide, heterogeneity-focused literature with a transferable, city-scale template for prioritizing infrastructure interventions.

A critical but underexplored dimension of crash severity research concerns the role of crash documentation quality. Cases where no contributing circumstance is recorded may reflect either rare, infrastructure-driven events with few behavioral precursors, or systematic gaps in how crashes are investigated and reported. How this documentation gap relates to crash outcomes, once road design and road-user type are accounted for, remains poorly understood and motivates the present study. While national-level research provides aggregate trends, fewer studies disaggregate crash severity outcomes by crash documentation quality, road type, and localized infrastructure characteristics, the very factors most relevant for city-level safety planning (Younes et al., 2023). Charlotte, North Carolina, presents a timely opportunity to address these gaps. As one of the fastest-growing metropolitan areas in the U.S. with a population exceeding 890,000 as of 2023, Charlotte has experienced rapid urban expansion and increased demand on its multimodal transportation network (McAllister, 2024). The city has adopted Vision Zero principles, committing to eliminating traffic fatalities and serious injuries through data-driven infrastructure upgrades and equitable safety policies (City of Charlotte, n.d.a; City of Charlotte, n.d.b). These local commitments align with federal Vision Zero guidance, which emphasizes infrastructure-based countermeasures for eliminating fatal and serious injuries (FHWA, 2023).

This study addresses these gaps by examining crash severity among pedestrians and bicyclists in Charlotte from 2019 to 2023. The analysis integrates environmental variables (road configuration, lighting, traffic control) with categorized behavioral causes to identify which factors most significantly predict severe outcomes. Two binary logistic regression models are developed and validated, with findings directly informing actionable planning and infrastructure priorities for vulnerable road users in Charlotte and similar rapidly urbanizing contexts. The study’s distinct contribution is threefold. First, it provides a city-specific, infrastructure-focused severity analysis for Charlotte, complementing the statewide North Carolina literature with results tailored to one municipality’s Vision Zero program. Second, it directly contrasts behavioral fault attribution against built-environment factors within the same models, quantifying which matters more for severe outcomes. Third, it treats the “no contributing circumstance” category as an analytic signal of crash-documentation quality rather than discarding it, and shows that its apparent severity association does not survive adjustment for road design and road-user type.

2 Methodology

2.1 Study area

Charlotte, the largest city in North Carolina, is one of the fastest-growing metropolitan areas in the southeastern United States. With a population exceeding 890,000 as of 2023 (McAllister, 2024) and an expanding multimodal transportation network, the city has experienced significant increases in pedestrian and bicycle activity alongside growing traffic volumes. Charlotte has actively pursued Vision Zero strategies, aiming to eliminate traffic-related fatalities and serious injuries through infrastructure upgrades and policy reform (City of Charlotte, n.d.a; City of Charlotte, n.d.b). Despite these efforts, the city continues to face substantial safety challenges for vulnerable road users. Figure 1 illustrates the spatial distribution of pedestrian and bicycle crashes from 2019 to 2023, with the highest concentrations in the central urban core, highlighting the geographic need for targeted safety interventions.

2.2 Data sources and preparation

Crash records were sourced from the Highway Safety Information System (HSIS) (FHWA, n.d.) and included details on crash severity, timing, road geometry, lighting, traffic control type, and primary contributing factors. The dependent variable (crash severity) was derived from the original five-level severity classification: fatal injury (1), suspected serious injury (2), possible injury (3), minor injury (4), and no injury (5). Figure 2a presents the distribution of crash severity levels observed in the dataset, which reflects the typical right-skewed pattern common in crash studies, with non-severe outcomes substantially outnumbering severe ones. Figure 2b shows crash severity distributions by primary cause; crashes with no contributing circumstance documented account for a disproportionate share of severe outcomes.

Missing values in environmental variables were addressed through supplementary data collection. Crash timestamps were matched with hourly historical weather data from verified platforms to classify road surface conditions as wet or dry. Geographic coordinates were used to manually verify missing roadway characteristics, including median type, roadway relation (mid-block, intersection, driveway), and traffic control type. Crashes without a recorded primary contributing factor were excluded from the analysis. Remaining records were classified into three groups: (a) Human-related causes (e.g., inattention, failure to yield, speeding); (b) No contributing circumstance indicated; and (c) Unknown or other causes (e.g., visibility obstruction, unclassified factors). This variable served as a key predictor in both models. The data-preparation flow proceeded as follows. The dataset extracted for analysis comprised 2,046 pedestrian- and bicyclist-involved crash records, each with a recorded primary contributing factor (records lacking this field were excluded during data extraction); these 2,046 records constitute the sample for the behavioral model (Model 1). For the environmental model (Model 2), an additional 25 records were removed (21 with unclassifiable “Unknown” roadway-area codes and 4 with missing road-alignment), producing the analytic sample of 2,021. Surface condition was imputed from matched hourly weather for records with missing road-surface data, and median type was verified manually from geographic coordinates where missing. Missingness was confined mainly to road-surface condition and median/roadway-relation type; these fields were completed through the weather-matching and coordinate-based verification described above rather than dropped, so no additional records were lost to missing environmental data beyond the 25 excluded for unclassifiable roadway-area or alignment codes. Because the supplemental values were reconstructed rather than field-observed, some residual misclassification is possible, and this measurement limitation is revisited in the Limitations section. The full distribution of all model variables by severity outcome is reported in Table 1 (descriptive statistics). Annual crash counts were 437 (2019), 347 (2020), 313 (2021), 447 (2022), and 502 (2023); the lower volumes in 2020–2021 are consistent with pandemic-period reductions in travel. The severe-crash share was broadly stable across years (7–10%), with a modest elevation in 2021 (13.1%). Because the analysis pools five years, single-year anomalies are unlikely to drive the results, although findings should be generalized to non-pandemic conditions with appropriate caution.

Table 1. Descriptive statistics of model variables by crash severity (N = 2,046).

2.3 Model selection rationale

Although crash severity is inherently ordinal, ordinal logistic regression was found unsuitable due to significant violations of the proportional odds assumption (confirmed via Brant test), particularly for predictors such as traffic control type and roadway configuration. Generalized ordinal models frequently failed to converge when applied to sparse categories, and multinomial logistic regression, while converging yielded poor classification of severe outcomes (overall accuracy ~49.4%), which are the most policy-relevant. Binary logistic regression was therefore adopted, consolidating the outcome into: severe (fatal or suspected serious injury; original levels 1–2) and non-severe (possible, minor, or no injury; original levels 3–5). This framework aligns with Charlotte’s Vision Zero prioritization of fatal and serious injury elimination and maximizes model interpretability and stability. Collapsing the five KABCO levels into severe (K, A) versus non-severe (B, C, O) also reflects the small number of fatal and serious-injury cases (110 and 84, respectively): retaining five outcome categories leaves the fatal cell too sparse for stable multi-category estimation, whereas the K + A grouping corresponds to the fatal-and-serious-injury outcomes that Vision Zero explicitly targets, making the binary split both statistically and substantively appropriate. More flexible alternatives, such as random-parameters ordered logit models with heterogeneity in means and variances, can accommodate unobserved heterogeneity in severity data (Ullah et al., 2025); however, such specifications require substantially larger samples within each severity level than were available here for the rare severe-outcome class, reinforcing the choice of a parsimonious binary model.

2.4 Model specification

Two models were estimated using R’s glm() function with a binomial link function. Odds ratios were computed by exponentiating model coefficients. Significance was assessed using Wald tests; model fit was evaluated via residual deviance, AIC, and receiver operating characteristic (ROC) analysis. Calibration plots were generated using 1,000 bootstrap repetitions. The binary logistic specification assumes independent observations, a linear relationship between the predictors and the log-odds of a severe outcome, and the absence of severe multicollinearity among predictors. These assumptions were examined directly: multicollinearity was assessed with variance inflation factors (VIF) (all between 1.0 and 1.4), the appropriateness of pooling pedestrians and bicyclists was tested with a likelihood-ratio test and addressed by including road-user type as a covariate, and sensitivity to the rare severe-outcome class was evaluated with an inverse-frequency-weighted re-estimation (all reported in the Results). Continuous lane count entered the model linearly in the logit, consistent with its monotonic association with severity.

Formally, let p_*i* denote the probability that crash i results in a severe outcome (fatal or suspected serious injury). Equation 1 expresses the log-odds (logit) of a severe outcome as a linear function of the predictors:

logit ( p i ) = ln [ p i / ( 1 − p i ) ] = β 0 + β 1 x 1 i + β 2 x 2 i + … + β k x ki (1)

where β₀ is the intercept and β_k are the coefficients for the k predictors x_ki; the odds ratio for each predictor is exp(β_k). The two models differ only in the predictor set. Model 1 (behavioral) includes the reclassified primary-cause group alone (reference: human-related). This specification is given in Equation 2, where the reclassified primary-cause group is the sole predictor and human-related causes serve as the reference category.

logit ( p i ) = β 0 + β 1 ( No − contribution ) + β 2 ( Unknown / Other ) (2)

Equation 3 presents Model 2, which augments the primary-cause term with the roadway, environmental, and road-user predictors:

logit ( p i ) = β 0 + β 1 ( Primary cause ) + β 2 ( Lighting ) + β 3 ( Traffic control ) + β 4 ( Surface ) + β 5 ( Road division ) + β 6 ( Functional context ) + β 7 ( Urbanization ) + β 8 ( Alignment ) + β 9 ( Lanes ) + β 10 ( Road − user type ) (3)

2.4.1 Model 1—primary cause (behavioral) model

This model used only the reclassified primary cause variable as predictor, treating human-related causes as the reference category. It provided a baseline estimate of how behavioral fault attribution relates to severity.

2.4.2 Model 2—environmental and primary cause model

This model extended Model 1 by adding roadway and environmental variables: lighting condition, traffic control type, road surface condition, roadway configuration (median type), functional context (intersection, driveway, mid-block), urbanization level (urban/rural), road alignment, number of lanes, and road-user type (pedestrian vs. bicyclist). The reclassified primary cause variable was retained as a covariate.

3 Results

3.1 Primary cause (behavioral) model

The behavioral model was estimated on 2,046 crash records with human-related causes as the reference category.

Table 2 presents the model results. Crashes where no contributing circumstance was recorded had 2.32 times higher odds of severe injury compared to human-attributed crashes (OR = 2.32, p < 0.001), a highly significant finding. As discussed below, this elevated odds ratio is best interpreted as a crash-documentation artifact rather than a causal infrastructure effect; it attenuates to non-significance once road design and road-user type are controlled in Model 2. Crashes classified as unknown or miscellaneous showed slightly elevated severity odds (OR = 1.29), but this difference was not statistically significant (p = 0.469). The model residual deviance was 1,253.8 (AIC = 1,259.8), and the AUC was 0.60 (95% CI 0.56–0.64), only marginally above chance (0.50). Model 1 is therefore presented as a baseline comparator rather than a standalone predictive tool. Calibration was favorable, with a mean absolute error of below 0.01 (Figure 3b); because Model 1 has only three predictor levels, predicted and observed severe-crash frequencies coincide closely within each group.

Table 2. Logistic regression model results for the crash primary reason factors.

Figures 3a,b present the ROC and calibration plots for the behavioral model, respectively. The calibration plot shows predicted probabilities aligning well with observed frequencies in the lower probability range, with minimal overfitting evidenced by the close proximity of the apparent and bias-corrected curves.

A key implication of the elevated severity in the “No Contributing Circumstance” category is that these crashes may occur at locations without any traffic control infrastructure. According to the dataset, crashes occurring at locations without traffic control account for the largest share of both non-severe and severe crashes (58.9% and 64.9%, respectively), followed by intersections with stop-and-go signals (26.0 and 24.5%). In contrast, stop-sign-controlled intersections represent a smaller proportion of crashes (11.5% of non-severe and 5.3% of severe crashes).

Similarly, most non-severe crashes occur during daylight (63.1%), while severe crashes are more frequent under dark conditions, particularly at lighted locations (40.2%) and unlighted locations (18.6%). Likewise, crashes predominantly occur under dry road conditions (87.5% of both non-severe and severe crashes), whereas wet conditions account for only 12.5%.

3.2 Environmental and primary cause model

The extended model incorporated roadway and environmental predictors alongside the primary cause variable. It was estimated on 2,021 crash records, with a null deviance of 1,260.0 (df = 2,020) and residual deviance of 1,056.2 (df = 1,999), yielding an AIC of 1,100.2, a substantial improvement over the baseline model. The ROC curve (Figure 4a) shows an AUC of 0.78 (95% CI 0.74–0.82), reflecting strong discriminative ability. The calibration plot (Figure 4b) shows close agreement between predicted and observed probabilities, with a mean absolute error of 0.013. Variance inflation factors (VIF) for all predictors ranged from 1.0 to 1.4 (maximum 1.4 for lane count), indicating no problematic multicollinearity among urban classification, lane count, and road division. Road-user type was a significant covariate: bicyclists had substantially lower severity odds than pedestrians (OR = 0.24, p < 0.001), and a likelihood-ratio test confirmed that its inclusion significantly improved fit (χ² = 24.4, df = 1, p < 0.001), supporting a pooled specification adjusted for road-user type rather than fully stratified models. As a robustness check for class imbalance, an inverse-frequency-weighted logistic regression produced effects consistent in direction, magnitude, and significance with the unweighted model (e.g., daylight OR = 0.38; lane count OR = 1.14; positive median barrier OR = 2.44; road-user type OR = 0.25), indicating that the principal findings are not artifacts of the rare severe-outcome class. To further evaluate goodness of fit and out-of-sample performance, the Hosmer–Lemeshow statistic for the environmental model was χ² = 17.41 (df = 8, p = 0.026), indicating broadly adequate agreement between predicted and observed severe-crash frequencies, with only a modest departure consistent with the large sample. Stratified 10-fold cross-validation yielded a mean AUC of 0.766 (SD 0.037) for the environmental model and 0.602 (SD 0.069) for the behavioral model, closely tracking the apparent AUCs (0.78 and 0.60) and indicating minimal overfitting. Ninety-five percent confidence intervals for every odds ratio, together with coefficients and p-values for all predictors (including non-significant terms), are reported in the complete regression output in Table 3.

Table 3. Complete environmental model output: coefficients, odds ratios, 95% confidence intervals, and p-values for all predictors (n = 2,021).

Coefficients are on the log-odds scale; odds ratios are exp(β). Reference categories: primary cause = human-related; lighting = dark/dim; traffic control = no control present; surface = dry; road division = two-way, not divided; functional context = mid-block; urbanization = rural/mixed; alignment = straight, level; road user = pedestrian. Number of lanes is continuous. Model fit: null deviance 1,260.0 (df = 2,020); residual deviance 1,056.2 (df = 1,999); AIC 1,100.2; AUC 0.78 (95% CI 0.74–0.82); Hosmer–Lemeshow χ² = 17.41 (df = 8, p = 0.026); 10-fold cross-validated AUC 0.766 (SD 0.037).

Table 4 presents the predictors from the environmental model, with reference categories noted. The elevated odds for crashes with no identified human contribution seen in the behavioral model no longer reached significance once road-user type and functional context were included (OR = 1.38, p = 0.073), reinforcing its interpretation as a documentation artifact rather than a robust infrastructure signal. Daylight conditions were strongly protective, reducing severe crash odds by 66% (OR = 0.34, p < 0.001). Roadway configuration produced the largest effects: two-way divided roads with positive median barriers were associated with more than twice the odds of severe outcomes (OR = 2.52, p < 0.001), and those with unprotected medians showed roughly a 1.8-fold increase (OR = 1.77, p = 0.005). Urban locations had about 1.8 times the odds of severe crashes relative to rural or mixed-use areas (OR = 1.76, p = 0.043), and each additional travel lane increased severity odds by 18% (OR = 1.18, p = 0.001). Crashes at intersections carried more than twice the odds of severe injury relative to mid-block locations (OR = 2.21, p < 0.001). Bicyclists had substantially lower severity odds than pedestrians (OR = 0.24, p < 0.001). Among traffic controls, signalized control was associated with lower severity odds (OR = 0.57, p = 0.012), while other control types were not statistically significant. Road-alignment effects were not statistically robust: the straight-grade category, previously reported as strongly protective, did not reach significance in the revised model (OR = 0.66, p = 0.23) and rests on a small number of severe outcomes (see Discussion).

Table 4. Logistic regression results for environmental factors.

n = 2,021; AUC = 0.78 (95% CI 0.74–0.82). Reference categories: primary cause = human-related; lighting = dark/dim; traffic control = no control present; surface = dry; road division = two-way, not divided; functional context = mid-block; urbanization = rural/mixed (<70% developed); alignment = straight, level; road-user type = pedestrian. Number of lanes is continuous (odds ratio per additional lane). Additional predictors significant in the revised model but not listed above: intersection vs. mid-block (OR = 2.21, p < 0.001); signalized control vs. no control (OR = 0.57, p = 0.012); and bicyclist vs. pedestrian (OR = 0.24, p < 0.001). Variance inflation factors (VIF) for all predictors were between 1.0 and 1.4.

4 Discussion

The findings of this study reveal that infrastructure and environmental characteristics are strong, independent predictors of crash severity for pedestrians and bicyclists in Charlotte, complementing and, in some cases, outweighing the influence of behavioral fault attribution. This has meaningful implications for Vision Zero planning and active transportation policy. Rather than treating each predictor in isolation, the discussion below builds a single argument—that the physical design of the roadway, more than the attribution of behavioral fault, governs how severe these crashes become—by considering in turn the role of crash documentation, lighting, road configuration and cross-section, land-use context, and traffic control, and then drawing out the cross-cutting methodological implications.

Situating these results within prior research clarifies what is consistent and what is distinctive. The protective association of daylight echoes a robust finding across pedestrian and bicyclist severity studies, in which reduced visibility in darkness is repeatedly linked to more severe outcomes; the present estimate reinforces that pattern in a mid-sized, rapidly growing U.S. city. The strong gradient across roadway cross-section—higher severity on divided, multi-lane, and positive-barrier roadways—aligns with the broader literature associating higher-speed arterial environments with more severe vulnerable-road-user injuries, and complements recent North Carolina work by Abdulrazaq and Fan (2024, 2025a, 2025b) that emphasizes substantial heterogeneity and seasonal instability in the determinants of vulnerable-road-user injury. A notable point of difference concerns the behavioral “no contributing factor” effect, which is prominent in the behavioral model but attenuates once roadway context is controlled; this divergence from analyses that report behavioral factors as dominant suggests that part of the apparent behavioral signal in police-reported data may reflect documentation practice rather than a causal mechanism. The lower severity odds for bicyclists relative to pedestrians are likewise consistent with injury-mechanism differences reported elsewhere. Where the present findings diverge from individual prior studies, the differences plausibly stem from variation in geographic setting, in severity coding (binary versus disaggregated KABCO levels), and in the specific covariates available, underscoring the value of locally calibrated models for guiding city-level safety investment.

A behavioral pattern warranting careful interpretation is the elevated severity of crashes where no contributing circumstance was documented. While this category showed markedly higher odds in the behavioral model (OR = 2.32), the effect attenuated to non-significance in the environmental model once road design, functional context, and road-user type were controlled (OR = 1.38, p = 0.073). This pattern points primarily to a crash-documentation artifact: more severe crashes may be harder to document behaviorally, particularly when the pedestrian or cyclist is incapacitated or when fault is ambiguous post-event. A secondary, but unconfirmed, possibility is that some of these crashes occur at locations lacking traffic control infrastructure, where there are no observable violations to record. We therefore frame this category as a data-quality signal with possible co-occurring infrastructure gaps, rather than as direct evidence of infrastructure-driven severity.

Among environmental predictors, lighting emerges as the most robust protective factor. Daylight reduces severe crash odds by 66%, consistent with the broader literature on visibility and pedestrian safety (Jurewicz et al., 2016). This underscores the value of street lighting upgrades in corridors with high pedestrian and bicycle activity, particularly in Charlotte’s urban core where crashes are most concentrated (Figure 1).

Road configuration produced the largest magnitude effects in the model. Divided roads, whether with positive median barriers or unprotected medians, were associated with markedly higher severity odds (2.5 × and 1.8×, respectively). These findings likely reflect vehicle operating speeds, which tend to be higher on divided arterials, and reduced crossing opportunities for pedestrians. This is consistent with prior research linking high-speed corridors to severe non-motorized crash outcomes (Pulugurtha et al., 2007; Jacobsen and Rutter, 2012). The 18% per-lane increase in severity odds further reinforces the role of cross-sectional road design: wider, multi-lane roads are intrinsically more hostile to pedestrians and cyclists. These findings align with the study’s observation that crash concentrations are highest in Charlotte’s central urban core (Figure 1), where multi-lane arterials intersect with high pedestrian activity.

Urban environments had roughly 1.8 times the severity odds of rural or mixed areas, likely reflecting higher traffic volumes, more complex multimodal interactions, and greater pedestrian/cyclist exposure. Road-alignment effects, by contrast, should be interpreted with caution. An earlier specification suggested that straight-grade alignment was strongly protective (OR = 0.16); however, this estimate proved sensitive to model specification and rests on a small number of severe outcomes (approximately 10 severe crashes among 153 in that category). In the revised model it was not statistically significant (OR = 0.66, p = 0.23). We therefore refrain from drawing substantive conclusions from alignment categories with sparse severe-outcome counts and report them (reference category: straight-level) primarily for completeness. The plausible mechanism—clear sightlines and predictable geometry reducing conflict intensity—remains untested here and would require a larger sample to evaluate reliably.

Among traffic-control types, signalized control was associated with significantly lower severity odds (OR = 0.57, p = 0.012) once roadway geometry, behavioral factors, and road-user type were controlled, while other control types did not reach significance. The latter should not be interpreted as evidence that traffic controls are irrelevant; their protective effects may also operate through interaction mechanisms—for example, stop signs reducing speed differentials at conflict points even when violated—that are not fully captured in a main-effects-only model. Future research should explore interaction terms and compliance behaviors to better isolate the role of traffic control.

A key methodological challenge in this study was class imbalance: severe crashes constituted a small fraction of total incidents, which limits model sensitivity for these critical but rare outcomes. To assess robustness, we re-estimated the environmental model using inverse-frequency class weighting; the principal effects were stable in direction, magnitude, and significance, indicating that the findings are not artifacts of imbalance. Despite combining five years of data to increase case volume, the dataset still exhibits the typical right-skewed severity distribution common in crash studies (Figure 2a). Future work could further explore resampling methods, ensemble classifiers, or cost-sensitive approaches that prioritize severe-outcome detection without sacrificing calibration.

Several limitations should be acknowledged. First, the analysis relies on police-reported crash records from HSIS, which are subject to well-documented data constraints: under-reporting of less-severe and non-motorized crashes, inconsistencies in coded contributing factors, and the documentation gaps discussed above. The “no contributing circumstance” category is itself a manifestation of this constraint and is interpreted accordingly. Second, some environmental fields were reconstructed—road-surface condition was imputed from matched hourly weather and median type was verified manually from coordinates—which introduces possible measurement error, and the manual recoding of roadway-relation and median categories could not be independently audited. Third, potential sources of bias remain: severe outcomes are rare, so estimates for sparse categories (for example, certain alignment levels) are imprecise, and any correlation between documentation quality and severity could bias the behavioral coefficients even after adjustment. Fourth, generalizability is bounded: the findings reflect Charlotte’s specific built environment and travel patterns and may not transfer directly to rural areas or to cities with different network structures; the 2019–2023 window also spans the COVID-19 pandemic, and although pooling 5 years and the stability of effects across years mitigate this concern, the results are best generalized to non-pandemic conditions. Finally, the models estimate associations rather than causal effects, and main-effects-only specifications do not capture interactions (for example, between lighting and traffic control) that may shape severity. In addition, several factors known to influence crash severity were unavailable in the dataset and therefore omitted, including traffic volume (AADT), prevailing or posted vehicle speed, pedestrian and bicyclist characteristics (such as age, gender, and impairment), and the presence of dedicated non-motorized facilities (sidewalks, marked crossings, and bicycle lanes). Their absence may introduce omitted-variable bias: for example, the elevated severity observed on divided, multi-lane roadways likely also reflects higher operating speeds and traffic exposure that the model cannot separate from roadway geometry, so the reported geometric effects should be read as partly capturing these correlated, unmeasured conditions. Future work incorporating exposure, speed, and facility data would help disentangle these influences.

5 Conclusion

This study demonstrates that infrastructure and environmental characteristics are central drivers of crash severity for pedestrians and bicyclists in Charlotte, North Carolina. Because these results derive from observational, police-reported data, they should be understood as statistical associations rather than causal relationships: the analysis identifies roadway and environmental conditions that co-occur with more severe outcomes, not factors demonstrated to cause them. Crashes recorded without an identified human contribution appear disproportionately severe in simple comparisons, but this association is largely a crash-documentation artifact: it does not persist once road design and road-user type are accounted for. The broader implication is that behavioral fault assignment alone is an insufficient lens for understanding crash severity, and that road-design factors warrant primary attention in severity-reduction strategies.

Daylight conditions consistently reduce crash severity odds, reinforcing the importance of lighting improvements in high-activity corridors. Divided roadways and high lane counts dramatically elevate severity risk, pointing to the need for speed management, pedestrian refuge islands, and access control on multi-lane arterials. Urban environments require particular attention, as they combine high pedestrian exposure with the road types associated with greatest severity risk.

These findings offer directly actionable guidance for Charlotte’s Vision Zero program. Priority interventions should target multi-lane divided arterials in the central urban core with unprotected medians, particularly in low-light environments. Future analyses should incorporate interaction terms to better capture the conditional effects of traffic control, and address class imbalance through advanced modeling strategies. Broader replication across comparable rapidly urbanizing cities would strengthen the generalizability of these results.

Generative AI statement

The author(s) declared that Generative AI was not used in the creation of this manuscript.

Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.

Publisher’s note

All claims expressed in this article are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.

Footnotes

Edited by: S. M. Labib, Utrecht University, Netherlands

Reviewed by: Abiola Abdulrazaq, University of North Carolina at Charlotte, United States

Irfan Ullah, University of Wyoming, United States

Srinivasa Rao Gandupalli, Gandhi Institute of Technology and Management (GITAM), India