Decoding pedestrian severity at crosswalks using hybrid clustering and random parameter models Authors: Swastika Barua, Michael Starewich, Tausif Islam Chowdhury, Amir Rafe, Subasish Das Published: 2026 ABSTRACT AbstractPedestrian crashes at crosswalks represent a critical safety concern due to their disproportionate contribution to severe injuries and fatalities relative to their overall frequency. This study analyzes pedestrian-involved crosswalk crashes reported in the Texas Crash Records Information System (CRIS) from 2017 to 2022 to examine context-specific determinants of injury severity. A two-stage analytical framework was employed. First, Cluster Correspondence Analysis (CCA) identified three distinct crash environments: (1) intersection crashes associated with turn-phase right-of-way violations, (2) low-speed yield-phase property-damage-only collisions at intersections, and (3) distraction-related driveway departure crashes at non-intersections. Each cluster exhibited unique combinations of movement patterns, roadway characteristics, and behavioral attributes. Second, cluster-specific Random Parameter Logit with Heterogeneity in Means (RPLHM) and Multinomial Logit (MNL) models were estimated to evaluate injury severity determinants within each context. Results indicate that weather, lighting conditions, roadway functional type, posted speed limits, and inattentive driving significantly influence severity outcomes, with effects varying across clusters. Several parameters, including daylight conditions and undivided roadway configurations, demonstrated substantial unobserved heterogeneity, suggesting context-dependent risk shifts influenced by roadway environment and demographic factors. These findings highlight the limitations of uniform countermeasures and support cluster-specific, context-sensitive interventions to mitigate pedestrian injury severity at crosswalks.Keywords: Cluster correspondence analysis, Random parameter logit, CrosswalkSubject terms: Engineering, Mathematics and computing, Risk factors Introduction Crosswalk crashes involving pedestrian, pedacyclists and motor vehicles remain a significant public safety issue across the U.S., with crosswalks often serving as the focal point of such incidents. In 2021, a total of 7,388 pedestrians were killed in traffic crashes nationwide, reflecting a 12.5% increase from the previous year and marking the highest count since 19811. That same year, an estimated 60,577 pedestrians were injured, equating to one fatality every 71 min and one injury every 9 minutes1. Although pedestrians represent a small portion of total traffic participants, their vulnerability is disproportionately high. This is especially apparent in states like Texas, where pedestrian fatalities increased by 22% between 2019 and 20232. While pedestrian-related crashes account for only 1% of all reported traffic incidents in Texas, they contribute to 19% of roadway deaths, underscoring the severity of outcomes when such events occur2. These alarming trends highlight the urgent need to examine crosswalk-related crash dynamics and inform effective safety interventions. A growing body of research has emphasized that pedestrian injury severity at crosswalks is shaped by an interconnected relationship among geometric, behavioral, and environmental factors. Using mixed logit modeling on Florida crash data, Haleem et al.3 found that higher truck volumes, older pedestrians, dark lighting conditions, and increased vehicle speeds significantly elevated the probability of severe injuries at both signalized and unsignalized intersections. Similarly, Pfortmueller et al.4 analyzed 347 pedestrian crashes and reported that incidents at zebra crosswalks were associated with lower injury severity compared to non-zebra sites, while collisions involving large vehicles such as trucks showed markedly higher fatality odd. In the European context, Olszewski et al.5 examined fatal pedestrian crashes at unsignalized zebra crossings in Poland and found that darkness, midblock crossing location, and rising speed limits were critical predictors of fatal outcomes, particularly for older pedestrians. Complementing these findings, Fu et al.6 introduced a behavioral framework based on distance-velocity diagrams to evaluate driver compliance at non-signalized crosswalks, showing that incomplete stops and delayed yielding significantly increased pedestrian conflict risk. Despite these advances, much of the existing research remains centered on geometric and operational improvements, often overlooking the multifaceted interactions among driver behavior, roadway conditions, environmental factors, and temporal variations that contribute to crash severity under real-world conditions. To address these gaps, this study developed a multi-tiered modeling framework to capture the complex and heterogeneous nature of pedestrian injury severity at crosswalks. This study analyzes crash data from the Texas Crash Records Information System (CRIS) covering the years 2017 to 2022, focusing on urban intersection and crosswalk-related incidents across the state. The dataset includes detailed information on crash circumstances, environmental factors, road and area types, vehicle and person characteristics, and injury outcomes. Using this comprehensive data, the research applies a two-stage methodology. First, Cluster Correspondence Analysis (CCA) is used to group crashes into three distinct clusters based on shared patterns among categorical variables such as roadway configuration, lighting, driver behavior, and area context. Next, random parameter logit (RPL) models including specifications with heterogeneity in means are developed separately for each cluster. Literature review Description and statistics of crosswalk-related crashes Pedestrian crashes at crosswalks remain a pressing global safety issue, contributing significantly to traffic-related fatalities and injuries. In U.S. cities, older pedestrians are at a particularly high risk when crossing at intersections without signals or stop signs, where driver expectations and yielding behavior can be inconsistent7. Additionally, crash typing analyses have shown that fatal pedestrian incidents often occur in low-light conditions and involve vehicles traveling straight or turning at high speeds8. European data reveal sharp disparities in pedestrian safety across countries, with Poland and Lithuania reporting the highest pedestrian death rates at crosswalks, while the UK and Germany reported the lowest9. In developing countries like Ghana and Bangladesh, road crossing remains hazardous due to a combination of limited pedestrian infrastructure, poor lighting, and lack of enforcement. Pedestrian non-compliance such as ignoring midblock crossings or informal crossing behavior further increases crash likelihood in these contexts10,11. Crosswalk-related injuries disproportionately affect vulnerable populations, such as seniors7,12 and residents in low-income or minority neighborhoods13, further underscoring the importance of crosswalk design and accessibility. In Switzerland, a study of 347 adult pedestrian crashes found that more than half of it occurred on zebra crosswalks, though crashes off-crosswalks tended to be more severe, with higher injury severity scores and vehicle speeds4. Factors associated with crosswalk crashes Numerous factors influence the likelihood and severity of crosswalk-related crashes. Road geometry, vehicle speed, lighting conditions, and pedestrian demographics (especially age) consistently appear as significant contributors. For instance, higher speeds and lack of street lighting at unsignalized crosswalks significantly increase fatality risk3. Vehicle type also plays a role. Larger vehicles, such as trucks and SUVs, are more likely to cause severe injuries upon impact due to their mass and driver blind spots4, and pedestrian crashes are more likely at crossings without traffic signals or stop signs7. Behavioral elements such as distraction (e.g., using mobile phones, wearing headphones), risky crossing maneuvers (e.g., running, crossing obliquely), and lack of caution further elevate crash risk11,14. Pedestrians not using designated crosswalks also face higher injury severity, as evidenced by both urban U.S. and European studies. Meanwhile, factors such as infrastructure type (e.g., refuge islands, Danish offsets) and signal timing also shape interaction dynamics and crash potential15,16. Driver behavior is also context sensitive. Yielding tendencies are affected by signal presence, pedestrian volume, and crosswalk design. Rectangular Rapid Flashing Beacons (RRFB) activation, in particular, improves driver compliance, though its effectiveness can vary depending on roadway context and pedestrian presence17,18. Methodological advancements and key findings in crosswalk crash analysis Recent studies have adopted a range of innovative methodologies to better understand crosswalk-related crashes. Machine learning approaches such as SHAP, XGBoost and Random Parameters Discrete Outcome Models17, and Hidden Markov Models18 have improved accuracy in identifying crash predictors, yielding insights into critical risk patterns. Geographic Information System (GIS) tools and LiDAR sensors have enabled precise measurement of driver compliance and pedestrian behavior at uncontrolled crosswalks15. Video analytics and naturalistic data have facilitated the real-time extraction of vehicle-pedestrian interactions19. Association Rule Mining has also been applied to uncover complex crash patterns at nighttime and intersections8,12. Several studies also explore the efficacy of countermeasures such as RRFBs, pedestrian islands, and warning systems like SeeMe, showing mixed but generally positive results in improving driver yield rates and reducing crash severity17,20. Cross-sectional and case-control analyses further highlight the nuanced interplay of road user behavior, infrastructure, and environmental conditions7, suggesting the need for multi-layered interventions in both developed and developing urban environments. Recent studies on crash datasets involving advanced vehicle technologies further reinforce the necessity of combined clustering and modeling structures. Previous research by Tamakloe et al.21 demonstrated that pattern recognition within crash clusters involving vehicles with advanced driving technologies reveals distinct behavioral signatures that pooled models systematically obscure. Similarly, prior work integrating latent class analysis with multinomial logit modeling confirmed that heterogeneity in crash patterns produces fundamentally different severity determinants across subgroups, and that ignoring this structure leads to mis-specified inference22. Ren et al.23 further established, through association rule mining of autonomous vehicle rear-end crashes augmented with built environment data, that crash severity outcomes are governed by context-specific combinations of risk factors that operate differently across crash typologies. Collectively, these studies establish that a segmentation-first, modeling-second framework is not only methodologically justified but necessary when the underlying crash population is structurally heterogeneous, as is the case with crosswalk crashes spanning intersection, non-intersection, and driveway access contexts. Table 1 summarizes recent crosswalk safety studies, highlighting diverse methodologies, data sources, and key factors influencing pedestrian crash risk and injury severity. Table 1Summary of Recent Studies on Crosswalk Safety: Methodologies, Data Sources, and Key Findings.AuthorLocation; DataMethodologyFindingsPfortmueller et al.4Switzerland; Emergency Dept. Patient DataRetrospective cross-sectional analysisNon-zebra crosswalk crashes resulted in significantly higher injury severity (ISS = 14.4 vs. 10.5). Zebra crosswalks were associated with 39% lower odds of severe injury. Truck involvement increased injury risk significantly.Das et al.8U.S.; FARS crash data (2014–2018)Association Rule MiningFatal crashes outside crosswalks in dark conditions were more likely. Key factors: lighting condition, vehicle movement type, and pedestrian age ≥ 45. Lift scores used to identify high-risk patterns.Koepsell et al.7Washington & Clifornia; Pedestrian case-controlMatched case-control studyOlder pedestrians (65+) had 2.1 times higher collision risk at marked crosswalks without traffic signals or stop signs. Markings without controls may create a false sense of safety.Høye and Laureshyn20Norway; Field Study in TrondheimBefore-after observational studySeeMe warning system improved motorist yielding by up to 39% at some locations, but effects varied. Effectiveness reduced with high false alarm rates and poor sign placement.Goswamy et al.17Florida; 2016–2020 crash dataXGBoost and Random Parameter Discrete Outcome Models (RPDOM)RRFBs significantly reduced injury severity in nighttime pedestrian crashes. XGBoost had higher prediction accuracy (97%) than RPDOM (73.8%).Guan et al.15Reno, Nevada; LIDAR trajectory dataGIS-based classification using multimodal trajectory analysisRefuge islands and Danish Offset crosswalks improved driver yielding rates up to 78%. Curb extensions alone were less effective.Kong et al.16U.S.; Near-crash pedestrian datasetSHAP-based interpretable machine learningHigh pedestrian volume associated with left-turn near crashes. Slow vehicle speeds (< 10 mph) linked to critical near-crash events. Signal cycle length interacts with pedestrian volume to influence crash likelihood.Bungum et al.14U.S.; College campus crosswalk observationsObservational pedestrian behavior studyOnly 13.5% of pedestrians displayed full caution. Distracted pedestrians were significantly less likely to follow safety behaviors like checking traffic before crossing.Olszewski et al.5Poland; National crash data (2007–2012)Binary logit model with interaction termsFatality risk increased by 37% for every 10 km/h speed increase. Nighttime, undivided roads, and midblock crossings were key risk enhancers.Haleem et al.3Florida, U.S.; Crash dataMixed logit modelHigh truck percentage, older pedestrians, and poor lighting significantly increased crash severity at both signalized and unsignalized intersections. Standard crosswalks reduced severe injury odds by 1.36%. Summary of Recent Studies on Crosswalk Safety: Methodologies, Data Sources, and Key Findings. Research gap and contribution Despite a growing body of research on crosswalk safety, several critical methodological and conceptual gaps persist that limit the applicability of existing findings for targeted intervention. First, the majority of prior studies rely on single-model frameworks, predominantly binary or ordered logit/probit models, estimated on pooled crosswalk datasets without accounting for the structural heterogeneity embedded within such data. Pooling fundamentally different crash scenarios, such as intersection turn-phase conflicts, low-speed yield-phase interactions, and distraction-induced driveway departures, into a single model produces averaged parameter estimates that may be statistically significant yet behaviorally misleading. When distinct crash mechanisms are treated as a homogeneous population, the resulting countermeasures are necessarily generic, overlooking the context-specific risk factors that drive severe outcomes in each crash type. Second, while prior studies employing machine learning approaches such as XGBoost and SHAP-based methods have improved predictive accuracy, such frameworks do not provide the probabilistic behavioral interpretation required for causal inference or policy design. Previous research applying association rule mining and hidden Markov models similarly identifies co-occurrence patterns without quantifying how specific factors shift injury severity probabilities under varying crash conditions. Third, and most critically, no previous study has explicitly recognized and modeled the coexistence of multiple latent crash environments within crosswalk datasets. Crosswalk crashes are not a single phenomenon; previous research indicates that such crashes span intersection conflicts, non-intersection driveway events, and low-speed property-damage interactions, each governed by distinct combinations of driver behavior, roadway geometry, and environmental conditions. Applying a uniform modeling framework across these typologically distinct scenarios suppresses context-dependent heterogeneity, both observed and unobserved, and risks producing biased estimates of key safety determinants such as lighting, speed, and inattentiveness. This study directly addresses these gaps through a two-stage analytical framework. First CCA is applied to objectively segment crosswalk crashes into behaviorally coherent typologies based on categorical crash attributes, without imposing a priori groupings. Second, cluster-specific Random Parameter Logit models with Heterogeneity in Means (RPLHM) are estimated within each segment, capturing both observed and unobserved sources of parameter variation that uniform models cannot accommodate. This approach moves beyond descriptive crash typing toward a rigorous, heterogeneity-aware modeling strategy that provides the cluster-specific behavioral insights necessary for precise, context-sensitive safety countermeasure design. Data preparation This study was designed to systematically analyze crosswalk-related crash outcomes using a multi-stage, data-driven approach as shown in Fig. 1. The process began with the 2017–2024 Texas crash dataset, which includes 12,816 crosswalk crash observations, which was prepared by filtering cases using the Texas CRIS variable Traffic_Cntl_ID, retaining only records where the traffic control device at the crash location was coded as ‘Crosswalk’ (n = 12,816). This criterion identifies crashes occurring at locations where a marked crosswalk facility was recorded as the applicable traffic control device, encompassing all units involved including drivers, pedestrians, and pedalcyclists, in those crash events, irrespective of the specific road user type. After initial variable selection using a Random Forest method, cross-correlation analysis (e.g. Cramér’s V) was conducted to examine relationships between key predictors. The optimal number of clusters was identified with the elbow method. CCA was then applied, segmenting the dataset into three distinct crash clusters based on shared patterns in crash attributes. For modeling injury outcomes, three categories were defined: fatal and severe injury (KA), minor and possible injury (BC), and no injury (O). Each cluster was analyzed with a sequence of increasingly flexible discrete choice models starting with the multinomial logit (MNL) model, followed by the RPL model, and finally, the RPLHM to capture both observed and unobserved sources of risk and variation. The study also computed marginal effects to quantify how each predictor influenced the probability of different injury outcomes. Flowchart of study design. Methodology Cluster corresponding analysis The CCA method combines dimension reduction with cluster analysis for categorical data, improving cluster convergence compared to earlier methods24,25. In this approach, a binary super indicator matrix Z is utilized, where the matrix dimensions are , where n representing the number of observations and Q denoting the total number of categories across all variables. Each observation encodes a single category as one, with the remaining elements set to zero. The matrix Z is structured as , where is an indicator matrix for the j -th categorical variable, and q indicates the total categorical variables. Similar to the loading matrices in continuous data analysis, the B matrix captures category quantifications across all categorical variables. The equation defines the quantifications for each j-th categorical variable, where has dimensions . Additionally, Y, an matrix, where reduced space coordinates of observations (rows of Z) are represented. , , indicates cluster membership, and G, , denotes the matrix of cluster centroids24. Cross-tabulation and contingency analysis To establish the relationship between clusters and categorical variables, a cross-tabulation of the indicator matrix (Z) and membership matrix (Z_K) is constructed. This results in a K × p_j contingency matrix, denoted as 24. Applying CCA to this contingency matrix (F) determines scaling values that maximize variance between clusters and categories. This process ensures that clusters effectively distinguish observations by optimizing variance across categorical variables, while also enhancing the distribution of categories within each variable. Steps in the CCA algorithm The CCA algorithm unfolds through several steps. Cluster Initialization: The cluster allocation matrix is initialized by randomly assigning objects to clusters. Simultaneously, category quantifications are estimated. Cluster Initialization: The cluster allocation matrix is initialized by randomly assigning objects to clusters. Simultaneously, category quantifications are estimated. Coordinate Computation: The coordinates for observations are determined by computing the averaged centered scores using the previously obtained category quantifications. This is represented by the equation: . Coordinate Computation: The coordinates for observations are determined by computing the averaged centered scores using the previously obtained category quantifications. This is represented by the equation: . Cluster Update: The K-means clustering algorithm is applied to the Y matrix, updating the cluster allocation matrix . Cluster Update: The K-means clustering algorithm is applied to the Y matrix, updating the cluster allocation matrix . Iterative Refinement: The process is repeated, returning to Step 2, where the updated matrix is used for the next iteration. Iterative Refinement: The process is repeated, returning to Step 2, where the updated matrix is used for the next iteration. Convergence: The algorithm continues iterating until stabilizes, and consequently, Y and G remain unchanged. Once this stability is achieved, the optimal cluster centroid matrix G and category quantification matrix B are finalized24. Convergence: The algorithm continues iterating until stabilizes, and consequently, Y and G remain unchanged. Once this stability is achieved, the optimal cluster centroid matrix G and category quantification matrix B are finalized24. Biplot generation and scaling adjustments These matrices, G and B derived from the clustering process are utilized to generate a biplot that visually represents clusters and category relationships. Due to dimensional disparities, standard Correspondence Analysis (CA) normalizations may fail to maintain an even distribution of row and column points. To resolve this, a scaling adjustment is recommended, ensuring that the average squared deviation from the origin remains uniform across clusters and categories. This is achieved using a scaling constant (γ), calculated as . The new scaled measures, and , maintain the same average squared deviation from the origin, ensuring consistency for biplot presentations in the analysis24. Random parameter modeling framework Injury severity outcomes were classified as KA, BC, and O, with the modeling process beginning with an MNL framework and annual model estimations to assess year-to-year transferability. Backward elimination at a 10% significance level was used to retain statistically meaningful variables, and a constant was included to capture base utility. Recognizing the MNL model’s assumption of homogeneity and independence, a RPL model was developed to allow selected parameters to vary randomly, with each variable tested for randomness in both mean and standard deviation. Model comparison relied on metrics such as the likelihood ratio test, AIC, AIC/N, and McFadden’s Pseudo R-squared. Finally, the RPLHM model was employed to assess how observable characteristics influence both the mean and variability of random parameters, providing a more comprehensive view of heterogeneity in injury severity outcomes. Although injury severity outcomes carry an inherent ordinal structure, this study employs multinomial rather than ordered response models for several reasons. Ordered logit and probit specifications impose the proportional odds assumption, requiring each covariate to exert a monotonically consistent effect across all severity transitions simultaneously26,27. This assumption is frequently violated in crash severity contexts, where environmental, behavioral, and geometric factors often operate differently across KA, BC, and O thresholds. The MNL and RPL frameworks relax this restriction, permitting variable effects to vary freely across severity outcomes and accommodating the non-proportional influence patterns documented in the crosswalk crash literature. Regarding the choice of CCA followed by cluster-specific RPLHM over a direct Latent Class Logit (LCL) approach, while LCL simultaneously estimates class membership and outcome probabilities, class assignments remain probabilistic and are defined solely by statistical fit criteria, yielding limited behavioral interpretability. The CCA-first framework employed here produces geometrically distinct crash typologies, each characterized by coherent combinations of roadway geometry, driver behavior, and environmental conditions that are interpretable as real-world crash scenarios prior to model estimation. This two-stage structure facilitates the development of cluster-specific, context-sensitive countermeasures in a manner that purely probabilistic latent class assignments cannot readily support. Crashes involving pedestrian crosswalks often include unmeasured or omitted variables such as pedestrian intent, driver yielding behavior, and the presence or clarity of signage or markings. These unrecorded influences introduce unobserved heterogeneity into the data, potentially distorting model results if not properly addressed28,29. To capture this latent variation, this study utilizes a random parameters logit model with heterogeneity in both means and variances. This modeling strategy allows for more flexible estimation of injury severity probabilities across crash events that differ by location, behavior, and environment. This approach has gained prominence in crash modeling research as it enables parameter variation across individuals and contexts, yielding a more detailed depiction of behavioral differences30,31. The methodology begins with the following utility function: Here, denotes the likelihood of crash j to result in injury severity level i. The vector represents the observable explanatory variables associated with crash j at severity level i, is the vector of parameters to be estimated, and is the unobserved random error term. The basic MNL model evaluates the probability of each severity level outcome using the following expression30: This equation expresses the probability that crash j results in injury severity level i, among a set of three possible discrete outcomes. To incorporate unobserved heterogeneity in the parameter estimates, the model takes the form30,32: In this formulation, is the probability density function of , defined by a vector of distributional parameters , including its mean and variance. All other terms retain their earlier definitions. To allow both the mean and standard deviation of random parameters to vary across observations, the model adopts the following specification33,34: In this expression, is the mean of the random parameter across observations. is a vector of variables explaining variation in the mean of the random parameters, and is a vector of estimable parameters corresponding to . is a vector of variables affecting the heterogeneity in the standard deviation of the random parameters, while represents associated estimable parameters. is the scale term, and is a normally distributed random term with zero mean and unit variance. This formulation captures variability in both the central tendency and dispersion of parameters across observations30,35. If multiple parameters are identified as random, the model accounts for potential correlation among them. This is represented by the correlated random parameters specification36: In this equation, is the overall mean vector of the random parameters, is the vector of factors that influence this mean, and is a set of estimable parameters. represents the Cholesky matrix used to express the covariance between the random parameters, while is a vector of zero-mean, unit-variance random variables. Marginal effects The influence of predictor variables on the probability of different pedestrian injury severity outcomes is assessed through the use of marginal effects. These values capture the shift in the likelihood of a given severity level when a binary variable transitions from 0 to 1, revealing how much more (or less) probable a certain injury outcome becomes due to the presence of a specific condition or characteristic, holding other factors constant. It is calculated as follows30: To ensure consistent and reliable estimates, all models were fitted using simulated maximum likelihood estimation with 1,000 Halton draws37,38. The random parameters were modeled using a normal distribution, a choice supported by prior literature indicating better empirical performance than alternative forms39,40. Marginal effects were computed from the model offering the best overall fit determined by the likelihood ratio test, both for each annual subset and for the full dataset. Code availability The custom code and analytical scripts central to the conclusions of this study are publicly available in the following GitHub repository: https://github.com/AI-in-Transportation-Lab/crosswalk-nlogit-code. The repository comprises R scripts implementing the CCA via the clustrd package and NLOGIT command syntax files used to estimate the MNL, RPL and RPLHM models for each crash cluster. Sufficient documentation is provided to enable readers to access, understand, and reproduce the complete analysis pipeline. Results and discussions Variable selection In this Random Forest analysis41 as shown in Fig. 2, a total of 21 variables were considered. The variables and their abbreviations are as follows: weather (Weather), lighting condition (Light), road type (RdType), road alignment (RdAlign), harmful event (HarmEv), intersection relation (IntRel), first harmful event (FHE), other contributing factors (OthrFac), road class (RdClass), population group (PopGrp), crash speed (CrSpeed), crash hour (Cr_Hr), contributing factor (ConFac), vehicle type (VehType), person type (PrType), person age (PrAge), driver age (DrAge), surface condition (SurCon), roalrelated (RdRel), road part (RdPart), unit type (Unit), and occupant position (OccPos). Variable importance plot utilizing Random Forest. The Random Forest variable importance plot reveals that occupant position, person type, and unit type are the most significant predictors in the model, each receiving the highest importance scores by a large margin. This result indicates that the specific position of the occupant, the type of person involved, and the type of unit (such as a vehicle, pedestrian, or cyclist) are central to understanding crash outcomes in this dataset. contributing factor, person age, and driver age also emerge as relevant predictors, highlighting the additional impact of behavioral and demographic characteristics on crash patterns. Other variables such as population group, road type, first harmful event, and other contributing factors play a moderate role, while variables including weather, crash hour, road alignment, surface condition, road part, and road relation contribute less to the model’s classification performance. The cross-correlation and Cramér’s V42 analyses as shown in Fig. 3 were conducted to examine the strength and nature of associations among the variables in the dataset. The ranked cross-correlation results indicate that several variable pairs, such as road class and road part, as well as unit type and person type, exhibit strong relationships, suggesting that these variables often co-occur or have closely linked categories. Similarly, the Cramér’s V heatmap confirms strong associations among combinations like unit type and person type, person type and occupant position, and unit type and occupant position. These findings highlight which variables share significant overlap or interaction in the context of crash data. Based on the insights gained from these correlation analyses, the following variables were selected for further CCA: weather, light, road type, road alignment, harmful event, intersection relation, first harmful event, other contributing factors, road class, population group, crash speed, crash hour, contributing factor, vehicle type, person type, person age, driver age, and injury severity. (a) Ranked cross-correlation plot and (b) Cramer’s V correlation. The elbow method43 was used to determine the optimal number of clusters for the analysis as shown in Fig. 4. In the model criterion plot, a sharp increase is observed as the number of clusters moves from two to three, followed by a noticeable plateau, which forms a clear “elbow” at three clusters. This pattern indicates that three clusters capture most of the meaningful structure in the data without adding unnecessary complexity. Choosing more than three clusters’ results in only minor improvements in model fit, confirming that three is the most suitable solution for this analysis. (a) Elbow method for optimal number of clusters (b) cluster Bi-plot. The biplot from the CCA displays the distribution of cases and variable categories across two dimensions. In this solution, three clusters were identified with sizes of 7,676 (59.9%), 2,644 (20.6%), and 2,496 (19.5%) cases, respectively as shown in Table 2. The cluster centroids are distinct in the two-dimensional space: Cluster 1 is centered at approximately (− 0.0034, − 0.0032), Cluster 2 at (0.0119, − 0.0009), and Cluster 3 at (− 0.0021, 0.0107). The within-cluster sum of squares for each group (0.0772, 0.0456, and 0.0571) reflect the compactness and separation of the clusters. Additionally, the model explains a substantial portion of the variation in the data, with the ratio of between-cluster to total sum of squares reaching 82.42%. The biplot further highlights how clusters are distributed and how key variable categories group within each cluster, providing a clear visual interpretation of the main groupings and distinguishing characteristics within the dataset. Table 2Centroid, Size, and Dimensions of Clusters.ClusterSize and PercentageDimension 1Dimension 2Within cluster sum of squares by clusterCluster 17676 (59.9%)− 0.0034− 0.00320.0772Cluster 22644 (20.6%)0.0119− 0.00090.0456Cluster 32496 (19.5%)− 0.00210.01070.0571 Centroid, Size, and Dimensions of Clusters. Findings of cluster correspondence analysis Cluster 1: Intersection pedestrian/cyclist-vehicle collisions due to turn-phase right-of-way violations In this cluster, drivers execute turning maneuvers (right and left) into marked crosswalks or dedicated bike lanes, entering the path of pedestrians and cyclists who lawfully occupy these spaces and failing to detect or yield to them. Figure 5 represents the contributing factors for this crash pattern. Most crashes occur in urban settings at both signalized and stop-controlled intersections during daylight hours when drivers fail to yield while turning at speeds of 10 to 20 mph. Obstructed sight lines from parked vehicles or signage, combined with driver distraction or misjudgment of a cyclist’s speed, often lead to these right-of-way violations. Despite the relatively low turning speeds, the vulnerability of unprotected road users means a high proportion of collisions result in moderate to severe injuries. Countermeasures such as enhanced yield markings, leading pedestrian intervals, dedicated bicycle turn lanes and improved sight triangle clearance can help reduce the frequency and severity of these incidents44–48. Cluster 1: Intersection Pedestrian/Cyclist-Vehicle Collisions Due to Turn-Phase Right-of-Way Violations. Cluster 2: Intersection pedestrian/cyclist -vehicle yield-phase property-damage-only collisions Cluster 2 represents low-speed, yield-phase interactions at signalized or marked intersections in which drivers begin to decelerate to yield to pedestrians or cyclists occupying crosswalks or bike lanes but misjudge gap acceptance or the road user’s path (see Fig. 6). The misjudgment often occurs due to limited sight distance, ambiguous signal phasing, or momentary inattention, and make mirror or bumper contact that results exclusively in vehicle and equipment damage. These collisions almost invariably occur at speeds below 15 mph, where the energy involved is insufficient to injure the vulnerable user but still causes property damage. Studies report that at unsignalized priority intersections, the vast majority of Type II bicycle-motor vehicle crashes, where motorists have the right of way yet fail to yield are classified as property-damage-only events, underscoring the prevalence of low-severity, yield-phase incidents47. property-damage-only pedestrian and bicyclist collisions often outnumber injury crashes at intersections, and mapping these incidents can guide targeted improvements to yield markings, signal timing, and driver awareness treatments49. Cluster 2: Intersection Pedestrian/Cyclist -Vehicle Yield-Phase Property-Damage-Only Collisions. Cluster 3: Driver distraction-induced non-intersection driveway departure collisions with through vehicles Cluster 3 represented in Fig. 7, is defined by non-intersection near crosswalk collisions in which an inattentive driver departs from a residential or commercial driveway or other access point, directly into the path of a through-travelling vehicle. These crashes most often occur on two-lane urban arterials or residential collectors, where the departing vehicle accelerates at 10–15 mph across one or more through lanes while the oncoming vehicle travels at 30–45 mph. Drivers engaged in secondary tasks (e.g., cell-phone dialing, infotainment adjustments, passenger interactions) experience visual and cognitive distraction, substantially increasing the likelihood of failing to detect or yield to approaching traffic50. Field studies show that crashes occurring “close to home” within one mile of a driver’s residence are disproportionately linked to lapses of attention rather than violations, highlighting the prevalence of attention-related driveway departures51. Although many impacts produce only property damage for the departing vehicle, occupants of the through vehicle face significant side-impact forces that can result in moderate to severe injury. Proven countermeasures include driveway consolidation, advance yield markings with “no-entry” symbols, roadside vegetation management to improve sightlines, and in-vehicle forward-collision warnings based on connected-vehicle data streams. Cluster 3: Driver Distraction–Induced Non-Intersection Driveway Departure Collisions with Through Vehicles. Cluster-based random parameter logit model The cluster-based descriptive statistics in Table 3 reveal several distinct patterns across crash, roadway, and demographic characteristics. In Cluster 1, crashes are most commonly associated with clear weather, daylight, city streets, and intersection-related locations, with a high prevalence of pedestrian involvement (mean = 0.748) and passenger car involvement (mean = 0.221). Notably, this cluster shows a higher rate of crashes at intersections (mean = 0.063) and intersection-related locations (mean = 0.768), as well as a greater proportion of crashes involving vehicles going straight or turning left. In contrast, Cluster 2 exhibits a substantial increase in the proportion of crashes involving motor vehicles (mean = 0.988), intersection crashes (mean = 0.213), and driveway access (mean = 0.144), along with a sharp decrease in pedestrian-involved crashes (mean = 0.005). The majority of cases in Cluster 2 also involve the occupant position being front left (mean = 0.951), highlighting a strong shift toward vehicle occupant scenarios. Cluster 3, on the other hand, is characterized by a sharp increase in non-intersection-related crashes (mean = 0.964) and a substantial rise in crashes occurring on non-trafficways (mean = 0.863) and under lower posted speed limits (mean = 0.786 for speed limits less than 25 mph). This cluster also displays a higher proportion of distracted driving (mean = 0.363) and fixed object involvement (mean = 0.031), while the prevalence of city street crashes and intersection-related crashes drops notably. Age group patterns shift across clusters as well, with younger drivers and persons under 15 years old more prominent in Cluster 2, and persons and drivers aged 55 to 64 years old slightly more common in Cluster 3. Table 3Cluster-Based Descriptive Statistics.VariablesCluster 1Cluster 2Cluster 3MeanSDMeanSDMeanSD Weather Condition Clear weather (1 if crash occurred in clear weather, 0 otherwise)0.7970.400.7840.410.8040.40Cloudy weather (1 if crash occurred in cloudy weather, 0 otherwise)0.1350.340.1410.350.1180.32Rainy weather (1 if crash occurred in rainy weather, 0 otherwise)0.0570.230.0690.250.0720.26 Lighting Condition Dark, lighted condition (1 if crash occurred in dark, lighted, 0 otherwise)0.2380.430.1420.350.1370.34Dark, not lighted condition (1 if crash occurred in dark, not lighted, 0 otherwise)0.0530.220.0160.130.0160.13daylight (1 if crash occurred in daylight, 0 otherwise)0.6720.470.8150.390.8210.38 Roadway Characteristics Two-lane, two-way (1 if a crash occurred in two-lane, two-way roads, 0 otherwise)0.3830.490.3730.480.3090.46Four or more lanes, divided (1 if crash occurred in four or more lane, divided roads, 0 otherwise)0.3190.470.3110.460.0770.27Four or more lanes, undivided (1 if crash occurred in four or more lanes, undivided roads, 0 otherwise)0.1630.370.1940.400.0380.19 Crash Characteristics Fixed object (1 if crash involved a fixed object, 0 otherwise)0.0200.14––0.0310.17Motor vehicles (1 if crash involved motor vehicles, 0 otherwise)0.0030.050.9880.110.0620.24Passenger car (1 if crash involved passenger cars, 0 otherwise)0.2210.420.0020.050.1050.31Pedestrian (1 if crash involved pedestrian, 0 otherwise)0.7480.430.0050.070.7060.46Driveway access (1 if crash occurred at or involved driveway access, 0 otherwise)0.0360.190.1440.350.0250.16Intersection (1 if crash occurred at intersection, 0 otherwise)0.0630.240.2130.410.0040.06Intersection related (1 if crash occurred at intersection related place, 0 otherwise)0.7680.420.2530.430.0070.08Non-intersection related (1 if crash occurred at non-intersection related place, 0 otherwise)0.1320.340.3900.490.9640.19Going straight (1 if crash occurred when vehicle going straight, 0 otherwise)0.4640.500.0010.030.6270.48Turning left (1 if crash occurred when vehicle turning left, 0 otherwise)0.3300.47--0.2020.40Turning right (1 if crash occurred when vehicle turning right, 0 otherwise)0.1960.400.0010.030.0500.22Distraction (1 if driver was distracted, 0 otherwise)0.1650.370.0710.260.3630.48Leaving driveway (1 if crash occurred when vehicle was leaving driveway, 0 otherwise)0.0250.160.1260.330.1270.33Vehicle movement (1 if crash occurred when one vehicle going slow and another stops, 0 otherwise)––0.3010.46–– Area Types City street (1 if crash occurred in city street, 0 otherwise)0.7690.420.7130.450.1170.32Farm to market (1 if the crash occurred on farm-to-market roads, and 0 otherwise)0.0500.220.0590.230.0020.04Non-trafficway (1 if the crash occurred on a non-trafficway, and 0 otherwise)0.0010.030.0340.180.8630.34US and state highways (1 if the crash occurred on a US and state highway, and 0 otherwise)0.1230.330.1090.310.0080.09Population group 10,000 to 24,000 (1 if the crash occurred in an area with a population between 10,000 and 24,000, and 0 otherwise)0.1330.340.1270.330.1920.39Population group 25,000 to 49,000 (1 if the crash occurred in an area with a population between 25,000 and 49,000, and 0 otherwise)0.0490.220.0890.280.1030.30Population group greater than 250,000 (1 if the crash occurred in an area with a population greater than 250,000, and 0 otherwise)0.6090.490.4480.500.3750.48Population group 50,000 to 99,000 (1 if the crash occurred in an area with a population between 25,000 and 49,000, and 0 otherwise)0.0930.290.1040.310.1060.31 Posted Speed Limit PSL 30 to 40 mph (1 if the crash occurred on a roadway with a PSL between 30 and 40 mph, and 0 otherwise)0.7660.420.6510.480.1430.35PSL 45 to 65 mph (1 if the crash occurred on a roadway with a PSL between 45 and 65 mph, and 0 otherwise)0.1290.340.1710.380.0660.25PSL greater than 65 mph (1 if the crash occurred on a roadway with a PSL greater than 65 mph, and 0 otherwise)--0.0060.080.0020.04PSL less than 25 mph (1 if the crash occurred on a roadway with a PSL less than 25 mph, and 0 otherwise)0.1050.310.1710.380.7860.41 Driver Demographic and Behavior Driver inattentiveness (1 if driver inattentiveness was a contributing factor in the crash, and 0 otherwise)0.0460.210.0720.260.1480.36Failed to yield (1 if failure to yield was a contributing factor in the crash, and 0 otherwise)0.2270.420.0110.100.1510.36No contributing factor (1 if failure to yield was a contributing factor in the crash, and 0 otherwise)0.4790.500.4780.500.4660.50Front left (1 if the occupant position was front left, and 0 otherwise)0.4730.500.9510.220.5220.50Person aged less than 15 years old (1 if the person involved in the crash was under 15 years old, and 0 otherwise)0.2060.400.2620.440.1510.36Person aged 25 to 54 years old (1 if the person involved in the crash was between 25 and 54 years old, and 0 otherwise)0.3940.490.4930.500.3450.48Person aged 55 to 64 years old (1 if the person involved in the crash was between 55 and 64 years old, and 0 otherwise)0.1130.320.0950.290.1530.36Driver aged less than 15 years old (1 if the driver involved in the crash was under 15 years old, and 0 otherwise)0.1940.400.2690.440.1400.35Driver aged 25 to 54 years old (1 if the driver involved in the crash was between 25 and 54 years old, and 0 otherwise)0.3870.490.4720.500.3250.47Driver aged 55 to 64 years old (1 if the driver involved in the crash was between 55 and 64 years old, and 0 otherwise)0.1110.310.0930.290.1490.36 Cluster-Based Descriptive Statistics. The modeling process for each cluster followed a systematic refinement strategy, starting with a standard MNL framework and progressing toward increasingly flexible specifications to capture unobserved heterogeneity. For Clusters 1 and 3, the RPLHM provided a statistically superior fit, as evidenced by the Likelihood Ratio Test and AIC metrics, and successfully identified several parameters with significant unobserved heterogeneity. However, for Cluster 2, the MNL specification was retained as the final model. Extensive testing was conducted to evaluate potential random parameters for Cluster 2; however, every candidate variable failed to yield a significant standard deviation for the random parameter distribution at the 10% level. Because the best-performing models for Cluster 2 were unable to produce significant random terms, the MNL model was determined to be the most statistically appropriate representation for this specific crash environment. Description of explanatory variables in random parameter logit models Weather condition The model estimation results in Table 4 show that weather plays a clear role in crosswalk crash patterns. In Cluster 1, the negative coefficient for cloudy weather (− 0.506) means that crosswalk crashes are less likely to happen during cloudy conditions in this group, compared to clear weather. This finding is consistent with results from Park and Bae52, who found that cloudy weather was associated with reduced pedestrian injury severity. Their study suggests that drivers may become more attentive or cautious in response to overcast or dim conditions, which could lead to fewer crashes in environments where crosswalk visibility or activity is already high. However, in Cluster 2, the positive coefficient for cloudy weather (1.085) suggests that crosswalk crashes are more likely to occur when the weather is cloudy. This may reflect regional or behavioral differences where cloudy conditions do not trigger the same level of driver caution, potentially due to overfamiliarity with such weather or less perceptible changes in roadway visibility. For Cluster 1, crashes are also more likely to take place during rainy weather (0.332), indicating that rain increases the risk of crosswalk crashes for this group. This aligns with findings by Zhai et al.53, who concluded that rainy weather increases pedestrian crash severity by reducing visibility, extending stopping distances, and impairing driver perception and control. The presence of rain likely exacerbates the complexity of crosswalk navigation for both pedestrians and drivers, particularly in environments with limited traffic control or poor surface conditions. There is no clear connection between weather and crosswalk crashes in Cluster 3, since no coefficients are reported for cloudy or rainy weather. These findings suggest that, for some clusters, bad weather, especially rain, can increase the risk of crosswalk crashes. Table 4Model Estimation of Cluster-based Best Models.VariablesCluster 1 RPLHMCluster 2 MNLCluster 3 RPLHMMeant-statMeant-statMeant-statConstant0.2533.623.71211.332.1997.99 Weather Condition Cloudy weather (1 if crash occurred in cloudy weather, 0 otherwise) [KA]− 0.506− 3.551.0852.2––Rainy weather (1 if crash occurred in rainy weather, 0 otherwise) [BC]0.3322.58–––– Lighting Condition Dark, lighted condition (1 if crash occurred in dark, lighted, 0 otherwise) [O]0.4736.99––0.4202.83Daylight (1 if crash occurred in daylight, 0 otherwise) [KA]− 1.458− 18.46− 1.572− 3.54− 17.373− 2.08 Standard deviation of the variable daylight ––––6.6972.38 Roadway Characteristics Two-lane, two-way (1 if crash occurred in two-lane, two-way roads, 0 otherwise) [BC]− 0.36− 2.79––––Standard deviation of the variable two-lane, two-way roads1.3184.41––––Four or more lanes, undivided (1 if crash occurred in four or more lanes, undivided roads, 0 otherwise) [KA]− 6.904− 2.32–––– Standard deviation of the variable four or more lanes undivided roads 6.2192.92–––– Crash Characteristics Fixed object (1 if crash involved a fixed object, 0 otherwise) [O]1.9097.4––1.2892.84Motor vehicles (1 if crash involved motor vehicles, 0 otherwise) [BC]3.7253.582.0066.02− 1.138− 4.19Passenger car (1 if crash involved passenger cars, 0 otherwise) [KA]− 1.040− 6.6––––Pedestrian (1 if crash involved pedestrian, 0 otherwise) [O]− 0.123− 1.74––− 1.127− 8.61Non-intersection related (1 if crash occurred at non-intersection related place, 0 otherwise) [O]− 0.187− 2.20.3903.23Going straight (1 if crash occurred when vehicle going straight, 0 otherwise) [KA]− 0.248− 3.154.9183.83− 2.362− 5.07Turning left (1 if crash occurred when vehicle turning left, 0 otherwise) [BC]0.76910.78––0.4423.08Turning right (1 if crash occurred when vehicle turning right, 0 otherwise) [O]− 0.280− 3.79––Distraction (1 if driver was distracted, 0 otherwise) [O]− 0.357− 4.41––− 0.426− 3.64Leaving driveway (1 if crash occurred when vehicle was leaving driveway, 0 otherwise) [KA]− 1.501− 2.85––––Vehicle movement (1 if crash occurred when one vehicle going slow and another stopped, 0 otherwise) [BC]––0.2972.43–– Area Types City street (1 if crash occurred in city street, 0 otherwise) [O]––––− 0.758− 2.57Non-trafficway (1 if the crash occurred on a non-trafficway, and 0 otherwise) [BC]––− 0.834− 1.771.5846.09Population group 10,000 to 24,000 (1 if the crash occurred in an area with a population between 10,000 and 24,000, and 0 otherwise) [KA]− 0.484− 3.35–––– Posted Speed Limit PSL 45 to 65 mph (1 if the crash occurred on a roadway with a PSL between 45 and 65 mph, and 0 otherwise) [BC]––0.3282.37–– Driver Demographic and Behavior Driver inattentiveness (1 if driver inattentiveness was a contributing factor in the crash, and 0 otherwise) [O]3.12313.730.5532.163.68713.72Failed to yield (1 if failure to yield was a contributing factor in the crash, and 0 otherwise) [BC]− 4.959− 16.91––− 5.793− 8.13Person aged less than 15 years old (1 if the person involved in the crash was under 15 years old, and 0 otherwise) [KA]− 0.246− 2.13––––Person aged 55 to 64 years old (1 if the person involved in the crash was between 55 and 64 years old, and 0 otherwise) [BC]0.1551.66––––Driver aged less than 15 years old (1 if the driver involved in the crash was under 15 years old, and 0 otherwise) [BC]− 0.400− 3.39––––Heterogeneity in mean, Parameter: VariableEffect of the presence of no contributing factor on the mean of the random parameter two-lane, two-way roads0.8735.19––––Effect of person’s aged 25 to 54 years old on the mean of the random parameter daylight––––− 2.371− 1.66Effect of pedestrians and pedacyclist presence on the mean of the random parameter daylight––––10.8172.34 Statistics Number of observations767626442496K261016Log likelihood at convergence− 5756.917− 1192.821− 1347.371Restricted log likelihood-8432.948-2904.73-2742.136McFadden Pseudo R-squared0.320.590.51AIC11565.802405.602726.70AIC/N1.5070.9101.092 Model Estimation of Cluster-based Best Models. The marginal effects for weather condition variables in Table 5 show cluster-specific influences on injury severity in crosswalk-related crashes. In Cluster 1, cloudy weather slightly decreases the probability of severe injury (KA: − 0.0037) and increases the likelihood of non-injury (O: 0.0026), while rainy weather has a minor negative effect on severe injury (KA: − 0.0007) and a small positive effect on possible injury (BC: 0.0031). In Cluster 2, cloudy weather is linked to a small increase in severe injury (KA: 0.0085) and a decrease in non-injury outcomes (O: − 0.0072), with minimal effect on possible injury (BC: − 0.0013). Marginal effects for these weather variables are not available in Cluster 3, indicating limited relevance in that cluster’s crash settings. These findings support prioritizing visibility-enhancing and weather-responsive countermeasures at high-risk crosswalk locations under adverse weather conditions. Table 5Marginal Effect Results of Cluster-based Bestt Models.VariableInjury SeverityCluster 1 RPLHMCluster 2 MNLCluster 3 RPLHMConstantKA− 0.0133− 0.0247− 0.0143BC− 0.0338− 0.4539− 0.3134O0.04710.47860.3277 Weather Condition Cloudy weather (1 if crash occurred in cloudy weather, 0 otherwise)KA− 0.00370.0085–BC0.0011− 0.0013–O0.0026− 0.0072–Rainy weather (1 if crash occurred in rainy weather, 0 otherwise)KA− 0.0007–− BC0.0031––O− 0.0024–– Lighting Condition Dark, lighted condition (1 if crash occurred in dark, lighted, 0 otherwise)KA− 0.0095–− 0.0012BC− 0.0133–− 0.0081O0.0228–0.0093Daylight (1 if crash occurred in daylight, 0 otherwise) [KA]KA− 0.0518− 0.01230.0167BC0.01560.0018− 0.0100O0.03620.0105− 0.0067 Roadway Characteristics Two-lane, two-way (1 if crash occurred in two-lane, two-way roads, 0 otherwise)KA− 0.0020––BC0.0103––O− 0.0083––Four or more lanes, undivided (1 if crash occurred in four or more lanes, undivided roads, 0 otherwise)KA0.0056––BC− 0.0023––O− 0.0033–– Crash Characteristics Fixed object (1 if crash involved a fixed object, 0 otherwise)KA− 0.0010–− 0.0001BC− 0.0034–− 0.0026O0.0044–0.0027Motor vehicles (1 if crash involved motor vehicles, 0 otherwise)KA− 0.0001− 0.00240.0001BC0.00050.2477− 0.0076O− 0.0004− 0.24530.0075Passenger car (1 if crash involved passenger cars, 0 otherwise)KA− 0.0072––BC0.0023––O0.0049––Pedestrian (1 if crash involved pedestrian, 0 otherwise)KA0.0057–0.0053BC0.0115–0.1156O− 0.0172–− 0.1209Non-intersection related (1 if crash occurred at non-intersection related place, 0 otherwise)KA0.0014− 0.0026–BC0.0034− 0.0477–O− 0.00480.0503–Going straight (1 if crash occurred when vehicle going straight, 0 otherwise)KA− 0.00830.0385− 0.0116BC0.0026− 0.00580.0057O0.0057− 0.03270.0059Turning left (1 if crash occurred when vehicle turning left, 0 otherwise)KA− 0.0063–− 0.0010BC0.0366–0.0116O− 0.0303–− 0.0106Turning right (1 if crash occurred when vehicle turning right, 0 otherwise)O− 0.0103––BC0.0072––KA0.0031––Distraction ( 1 if driver was distracted, 0 otherwise)KA0.0024–0.0008BC0.0072–0.0198O− 0.0096–− 0.0206Leaving driveway (1 if crash occurred when vehicle was leaving driveway, 0 otherwise)KA− 0.0008––BC0.0002––O0.0006––Vehicle movement (1 if crash occurred when one vehicle going slow and another stopped, 0 otherwise)KA–− 0.0003–BC–0.0366–O–− 0.0363– Area Types City street (1 if crash occurred in city street, 0 otherwise)KA––0.0007BC––0.0120O––− 0.0127Non-trafficway (1 if the crash occurred on a non-trafficway, and 0 otherwise)KA––− 0.0083BC––0.2062O––− 0.1979Population group 10,000 to 24,000 (1 if the crash occurred in an area with a population between 10,000 and 24,000, and 0 otherwise)KA− 0.0034––BC0.0010––O0.0024–– Posted Speed Limit PSL 45 to 65 mph (1 if the crash occurred on a roadway with a PSL between 45 and 65 mph, and 0 otherwise)KA–− 0.0004–BC–0.0405–O–− 0.0401– Driver Demographic and Behavior Driver inattentiveness (1 if driver inattentiveness was a contributing factor in the crash, and 0 otherwise)KA− 0.0016− 0.0037− 0.0004BC− 0.0085− 0.0676− 0.0236O0.01010.07130.0240Failed to yield (1 if failure to yield was a contributing factor in the crash, and 0 otherwise)KA0.0016–0.0001BC− 0.0106–− 0.0046O0.0090–0.0045Person aged less than 15 years old (1 if the person involved in the crash was under 15 years old, and 0 otherwise)KA− 0.0030––BC0.0010––O0.0020––Person aged 55 to 64 years old (1 if the person involved in the crash was between 55 and 64 years old, and 0 otherwise)KA− 0.0005––BC0.0029––O− 0.0024––Driver aged less than 15 years old (1 if the driver involved in the crash was under 15 years old, and 0 otherwise)KA− 0.0045––BC0.0012––O0.0033–– Marginal Effect Results of Cluster-based Bestt Models. Lighting condition Lighting conditions have a clear effect on crosswalk crash patterns as shown in Table 4. In Cluster 1, crosswalk crashes are more likely to happen at night when there is artificial lighting (coefficient = 0.473), suggesting that night-time crossings, even with streetlights, carry a higher risk of severe injury. One possible explanation is that artificial lighting, while helpful, may not fully eliminate visibility limitations, particularly when pedestrians wear dark clothing or cross unexpectedly. Drivers may also misjudge distances or struggle with glare and contrast, increasing the likelihood of severe outcomes. In contrast, crashes during daylight are much less likely in Cluster 1 (coefficient = − 1.458), meaning that daytime provides a safer environment for pedestrians at crosswalks. This same pattern is seen in Cluster 2, where crosswalk crashes are less likely during daylight hours (coefficient = − 1.572). In Cluster 3, crosswalk crashes are again more likely to occur at night with artificial lighting (coefficient = 0.420) and much less likely during daylight (coefficient = − 17.37). The high negative value for daylight in Cluster 3, along with some variation, points to very few severe crosswalk crashes during the day in this group. The results reinforce the notion that night-time conditions, even when illuminated, as associated with greater injury severity, whereas daylight consistently corresponds to safer pedestrian outcomes across clusters. Similar findings were reported by Haleem et al.3, who observed that artificial lighting does not fully mitigate crash severity at night, and by Siddiqui et al.54, who identified daylight as a protective factor for pedestrian safety at crossings. The marginal effects for lighting condition variables in Table 5 show varying impacts on injury severity across the three clusters of crosswalk-related crashes. In Cluster 1, dark but lighted conditions are associated with a reduction in both severe injury (KA: − 0.0095) and possible injury (BC: − 0.0133), while increasing the likelihood of non-injury outcomes (O: 0.0228). Daylight conditions in Cluster 1 show a larger decrease in severe injury probability (KA: − 0.0518) and a slight increase in possible injury (BC: 0.0156) and non-injury (O: 0.0362). In Cluster 2, daylight is linked to a small reduction in severe injury (KA: − 0.0123), with minor positive or negative effects on other outcomes (BC: 0.0018, O: 0.0105) (see Fig. 8). In Cluster 3, dark but lighted conditions have a minimal effect, slightly reducing severe (KA: − 0.0012) and possible injury (BC: − 0.0081), and modestly increasing non-injury (O: 0.0093). Daylight in Cluster 3 is associated with a slight increase in severe injury (KA: 0.0167), a decrease in possible injury (BC: − 0.0100), and a small decrease in non-injury (O: − 0.0067). These findings highlight the importance of prioritizing lighting upgrades and visibility-enhancing treatments at crosswalk locations, particularly in dark or low-visibility environments associated with elevated injury risks. Marginal effect plot for the variable “Daylight”. Roadway characteristics The roadway characteristics in Table 4 show how the type of road influences crosswalk crash patterns. In Cluster 1, crosswalk crashes are less likely to happen on “two lane, two way” roads (coefficient = − 0.36), meaning these roads see fewer severe crosswalk crashes in this group. This may reflect lower vehicle speeds, fewer conflict points, or better visibility between drivers and pedestrians, all of which contribute to a safer crossing environment. Kashani and Mohaymany55 found similar patterns on rural two-lane, two-way roads, noting that injury severity tended to decrease on these road types, particularly where speeds were moderate and roadside hazards were limited. There is some variation in this pattern, but overall, injury severity at crosswalks is lower on these roads. Similarly, crashes on “four or more lane, undivided” roads are also much less likely to be part of Cluster 1 (coefficient = − 6.904), indicating that wide undivided roads are not common sites for severe crosswalk crashes in this group. There is also some variation here, but the general trend remains. This finding may suggest either lower pedestrian exposure on these corridors within the dataset or successful implementation of safety features such as midblock refuge islands or enhanced signage. Pour-Rouholamin and Zhou56 previously reported elevated pedestrian injury risk on similar roadway types, emphasizing the need for targeted design interventions. The absence of severe crashes here could indicate areas where such interventions have mitigated risk. No such patterns were found in Clusters 2 and 3, as there are no coefficients reported for these road types. These results suggest that, for Cluster 1, severe crosswalk crashes are less common on both two-lane, two-way roads and on wider undivided roads, pointing to the importance of considering road design and type when looking to improve crosswalk safety. The marginal effects for roadway characteristics in Table 5 illustrate how road configuration influences injury severity in crosswalk-related crashes across clusters. In Cluster 1, crashes on two-lane, two-way roads are associated with a slight decrease in severe injury (KA: − 0.0020) and non-injury outcomes (O: − 0.0083), and a small increase in possible injury (BC: 0.0103). For crashes on four or more lane, undivided roads within Cluster 1, there is a minor increase in severe injury probability (KA: 0.0056), with corresponding small decreases in both possible injury (BC: − 0.0023) and non-injury (O: − 0.0033). These results suggest that, within intersection and crosswalk settings, two-lane, two-way roads are linked to marginally lower severe injury risk, while wider undivided roads show a slightly elevated severe injury likelihood. No significant marginal effects are estimated for these roadway variables in Clusters 2 and 3, indicating their impact is most notable in Cluster 1’s crosswalk and intersection-related crash environments. These findings support prioritizing speed-management and pedestrian refuge treatments on multilane undivided roadways with elevated crosswalk-related injury risk. Crash characteristics The crash characteristics in Table 4 show how certain types of crashes are more or less common in each cluster, and how these may affect injury severity, especially near crosswalks. In Cluster 1, crashes are more likely to involve hitting a fixed object (coefficient = 1.909) or a motor vehicle (3.725), which both suggest a higher chance of serious injury. Fixed-object crashes are often high-impact events and typically occur when a driver loses control, often at higher speeds or due to evasive maneuvers. Yamamoto and Shankar57 found that such collisions are strongly associated with severe injuries due to the absence of energy-absorbing structures and the abrupt deceleration involved. Similarly, collisions involving another motor vehicle may result in higher injury severity due to increased impact energy and the likelihood of secondary pedestrian contact. However, crashes are less likely to involve a passenger car (− 1.040) or a pedestrian (− 0.123), and less likely to occur when vehicles are going straight (− 0.248) or turning right (− 0.280). These patterns may reflect the reduced complexity and lower conflict potential during predictable movements such as straight-through travel, or the fact that right turns typically occur at slower speeds. Crashes in this cluster are more likely to happen during a left turn (0.769). Left-turn conflicts are well-documented as high-risk maneuvers for pedestrians, especially at crosswalks, due to driver blind spots, opposing traffic pressure, and poor yielding behavior. Distraction is less common here (− 0.357), potentially indicating that these crashes may result more from situational complexity than from inattentiveness. Crashes are also less likely to happen when leaving a driveway (− 1.501) or at non-intersection locations (− 0.187). Crashes involving driveways usually occur at lower speeds, often transpiring in residential or commercial areas. As shown by Agran et al.58, driveway incidents, while especially dangerous for children, tend to produce less severe injuries due to lower vehicle momentum and heightened driver awareness in those zones. For Cluster 2, motor vehicle crashes are more likely (2.006), and the chance of a crash happening while a vehicle is going straight is much higher (4.918). This could indicate a high volume, mid-block or intersection dominated environment, where pedestrian visibility or signal timing plays a critical role. The high coefficient for straight movement may reflect situations where pedestrians are struck by vehicles in motion along primary through lanes, possibly due to misjudged gaps or failure to yield. Crashes at non-intersection locations (0.390) and where one vehicle is moving slowly while another is stopped (0.297) are also more common. These crash dynamics may involve illegal maneuvers or sudden stops that lead to pedestrian conflict with bypassing or lane changing drivers. In Cluster 3, crashes are more likely to involve hitting a fixed object (1.289) or turning left (0.442), but less likely to involve a motor vehicle (− 1.138) or a pedestrian (− 1.127). This may lead to isolated or single vehicle events that still result in serious consequences, perhaps due to loss of control near pedestrian zones or failure to properly navigate left turn situations. Crashes in this group are also less likely when a vehicle is going straight (− 2.362) or if the driver was distracted (− 0.426). These results may imply that distraction plays a less prominent role in the most severe crashes in this cluster or distraction is inconsistently reported. However, previous research by Khan and Habib59 showed that distracted driving does elevate injury severity in pedestrian crashes, suggesting that even where distraction is infrequent, its presence warrants continued concern. These patterns suggest that, in the context of crosswalks, severe injuries are more common when crashes involve multiple motor vehicles, fixed objects, or left turns. Crashes during straight-through movements or involving distracted drivers are less common in these severe clusters. Because left turns and fixed object collisions are dangerous for pedestrians, special care should be taken to improve safety at crosswalks where vehicles may turn or strike roadside objects. Measures such as left turn phasing, improved fixed object setbacks, and increased visibility at complex intersections could reduce the incidence or severity of these high-risk crash types. The marginal effects for crash characteristics in Table 5 show several cluster-based differences in their association with injury severity in crosswalk-related crashes. Fixed object involvement has consistently small effects across clusters. Motor vehicle involvement produces a strong effect only in Cluster 2, greatly increasing the probability of possible injury (BC: 0.2477) and reducing non-injury (O: − 0.2453), while its effect is negligible in other clusters (see Fig. 9). Passenger car involvement mainly lowers severe injury slightly in Cluster 1 (KA: − 0.0072) but is not significant in other clusters. Pedestrian involvement raises the likelihood of severe injury in both Cluster 1 (KA: 0.0057) and Cluster 3 (KA: 0.0053), and strongly increases possible injury in Cluster 3 (BC: 0.1156) while reducing non-injury. Non-intersection-related crashes decrease severe and possible injury in Cluster 2 (KA: − 0.0026, BC: − 0.0477), but the effects are minor in other clusters. Going straight increases severe injury in Cluster 2 (KA: 0.0385) but lowers it in Clusters 1 and 3, while turning left consistently reduces severe injury and increases possible injury (see Fig. 10). Distraction has a stronger effect on increasing possible injury in Cluster 3 (BC: 0.0198), while leaving driveway and slow vehicle movement show only minor or cluster-specific impacts. These findings support prioritizing targeted countermeasures such as pedestrian-focused intersection design, turn-conflict mitigation, driveway access management, and distracted-driving enforcement at high-risk crosswalk locations. Marginal effect plot for the variable “Motor Vehicle”. Marginal effect plot for the variable “Going Straight”. Area types The area type variables in Table 4 reveal that the risk of severe crosswalk crashes varies by location. In Cluster 3, crashes are less likely to occur on city streets (coefficient = − 0.758), suggesting that severe crosswalk crashes in this group are not common in typical urban street settings. This finding aligns with previous research showing that pedestrian injury severity is often lower in dense urban environments like New York City, where slower speeds, higher traffic control density, and greater pedestrian visibility contribute to safer outcomes60,61. Instead, Cluster 3 shows a higher likelihood of crashes happening in non-trafficway areas (coefficient = 1.584), such as parking lots, driveways, or private properties, where traditional traffic controls and crosswalk markings might be absent. These environments can be particularly hazardous due to limited signage, irregular pedestrian paths, and reduced driver vigilance. In Cluster 2, severe crosswalk crashes are less likely to occur in non-trafficway locations (coefficient = − 0.834), indicating that these settings are not a significant source of risk for this cluster and possibly reflecting higher regulation or reduced pedestrian volume in such areas. For Cluster 1, the analysis shows that crashes are less likely to occur in areas with a population between 10,000 and 24,000 (coefficient = − 0.484), meaning that medium-sized communities are not prominent locations for severe crosswalk crashes in this group. This could be due to lower traffic volumes, shorter trip distances, and fewer high-speed corridors in such areas, all of which reduce crash risk and severity. These patterns highlight that severe crosswalk crashes do not always follow the most expected patterns of location and population. The marginal effects for area type variables in Table 5 indicate notable differences by cluster. In Cluster 3, crashes occurring on city streets have a minimal effect on severe injury (KA: 0.0007) and a modest increase in possible injury (BC: 0.0120), while reducing the probability of non-injury (O: − 0.0127). For non-trafficway crashes, Cluster 3 shows a small decrease in severe injury (KA: − 0.0083), a significant increase in possible injury (BC: 0.2062), and a notable decrease in non-injury (O: − 0.1979), suggesting a shift toward higher injury severity in non-trafficway environments. The population group 10,000 to 24,000 variable, which is only estimated in Cluster 1, is associated with a slight reduction in severe injury (KA: − 0.0034), a very small increase in possible injury (BC: 0.0010), and a minor increase in non-injury (O: 0.0024). These results suggest that non-trafficway environments in Cluster 3 are especially associated with higher injury severity, while city street and moderate-sized population areas have only minor effects on injury severity outcomes. These findings support prioritizing pedestrian safety improvements and access-management strategies in non-trafficway and urban crosswalk environments with elevated injury risks. Posted speed limit The posted speed limit and driver behavior variables provide important insights into factors that increase the risk of severe crosswalk crashes as shown in Table 4. In Cluster 2, crashes are more likely to occur on roadways with a posted speed limit between 45 and 65 mph (coefficient = 0.328), indicating that higher-speed roads pose a greater danger for crosswalk users in this group. This finding is supported by previous research demonstrating that pedestrian injury severity increases significantly with higher posted speed limits, due to the greater impact forces and reduced reaction times available to both drivers and pedestrians62. The marginal effects for posted speed limits between 45 and 65 mph are only significant in Cluster 2 as shown in Table 5, where such crashes slightly reduce the probability of severe injury (KA: − 0.0004), increase possible injury (BC: 0.0405), and decrease non-injury outcomes (O: − 0.0401). These findings support prioritizing speed-management and pedestrian protection measures at higher-speed crosswalk corridors with elevated injury risk potential. Driver behavior For driver behavior, inattentiveness as shown in Table 4 stands out as a strong positive factor across all clusters, with high coefficients (3.123 in Cluster 1, 0.553 in Cluster 2, and 3.687 in Cluster 3), showing that crashes where drivers are not paying attention are much more likely to lead to severe outcomes at crosswalks. This is consistent with findings from Damsere-Derry et al.63, who demonstrated that driver inattention significantly increases the likelihood of severe pedestrian injuries, especially in environments with complex pedestrian-vehicle interactions. Failure to yield is strongly negative in Cluster 1 (− 4.959) and Cluster 3 (− 5.793), suggesting that this specific behavior is less likely in these severe crash groups, possibly because other risky actions, like speeding or inattentiveness, are more dominant. The involvement of very young persons (under 15 years old) and very young drivers is less likely in severe crash clusters (with coefficients of − 0.246 and − 0.400, respectively, in Cluster 1), which aligns with the findings of Abey64, showing negative associations between young pedestrian involvement and injury severity. Meanwhile, older driver involvement (aged 55–64) is slightly more likely in Cluster 1 (0.155), supporting Okafor et al.65, who found that older drivers may have higher crash severity risks due to slower reaction times and age-related cognitive factors. These findings highlight that high-speed environments and inattentive driving greatly increase the risk of severe crosswalk crashes, emphasizing the critical need for targeted interventions such as enhanced driver education focused on attention and awareness, implementation of traffic calming measures to reduce vehicle speeds near crosswalks, and improved enforcement of traffic laws to mitigate risk behaviors. The marginal effects for driver demographic and behavior variables in Table 5 show distinct cluster-based patterns. Driver inattentiveness consistently decreases the likelihood of severe and possible injury across all clusters, with the strongest reduction for possible injury in Cluster 2 (BC: − 0.0676), while it increases the probability of non-injury, especially in Cluster 2 (O: 0.0713) (see Fig. 11). Failed to yield is associated with a slight increase in severe injury (KA: 0.0016) and non-injury (O: 0.0090), but a decrease in possible injury (BC: − 0.0106) in Cluster 1, with only very small effects in Cluster 3. For the age variables, both person and driver aged less than 15 years old show minor effects in Cluster 1, with a slight reduction in severe injury and a slight increase in non-injury. Similarly, being a person aged 55 to 64 years old yields small, mostly neutral effects. These findings support prioritizing driver awareness, yielding compliance, and targeted pedestrian safety interventions at crosswalk locations with elevated behavioral risk factors. Marginal effect plot for the variable “Driver Inattentiveness”. Random parameters and heterogeneity in means The estimation results illustrated in Table 4 show that several variables, such as daylight, two-lane two-way roads, and four or more lane undivided roads, are modeled as random parameters to account for unobserved heterogeneity in crosswalk-related crashes. In Cluster 1, both two lane two way roads (mean: − 0.36, SD: 1.318) and four or more lane undivided roads (mean: − 6.904, SD: 6.219) have large standard deviations, indicating that while their average effects reduce injury risk, there remains a substantial probability, 13.7% and 19.7%, respectively that their effects could be positive in specific situations. Similarly, daylight shows a strongly negative mean effect on severe injury, particularly in Cluster 3 (mean: − 17.373, SD: 6.697), but the high variability suggests a notable probability (13.1%) that daylight could increase risk in certain crash contexts, especially when combined with factors like age group or pedestrian presence. The analysis of unobserved heterogeneity across latent clusters revealed statistically significant heterogeneity in means for Clusters 1 and 3, with no such effects identified in Cluster 2. In Cluster 1, the presence of “no contributing factor” was found to significantly influence the mean of the random parameter associated with “two-lane, two-way roads.” Specifically, this variable produced a positive shift in the mean value of the parameter (coefficient = 0.873, t-stat = 5.19), indicating that, in the absence of identifiable contributing factors, the effect of two-lane, two-way roads on injury severity outcomes becomes more pronounced and variable. In Cluster 3, heterogeneity in the mean of the random parameter “daylight” was influenced by two key variables. First, involvement of individuals aged 25 to 54 years old led to a downward shift in the mean of the daylight parameter (coefficient = − 2.371, t-stat = − 1.66), suggesting reduced severity impacts under daylight conditions for this age group. Conversely, the presencce of pedestrians or pedacyclists in the crash significantly increased the mean of the daylight parameter (coefficient = 10.817, t-stat = 2.34), implying that daylight conditions may exacerbate injury severity when vulnerable road users are involved. These findings confirm the presence of unobserved heterogeneity in key roadway and environmental variables and illustrate how behavioral and contextual factors can systematically influence the distributional properties of random parameters across different crash environments. Modeling these effects is critical to improving predictive performance and ensuring that interventions are targeted to the varying risk profiles captured within each latent cluster. Policy implication The Safe System Approach (SSA) provides a structured framework for translating crosswalk crash severity findings into targeted countermeasures by addressing unsafe conditions across five tiers: Safe Road Users, Safe Speeds, Safe Roads, Safe Vehicles, and Post-Crash Care. The findings from the three cluster-specific models reveal distinct risk profiles intersection turn-phase conflicts, low-speed yield-phase interactions, and distraction-related driveway departure crashes each requiring context-sensitive interventions rather than uniform countermeasures and can be explained through the lens of SSA. First, lighting conditions emerged as the most consequential severity determinant across all three clusters. Daylight conditions produced the largest protective marginal effect on KA probability in Cluster 1 (-0.0518), and a strong negative coefficient in Cluster 2 (-1.572), while the Cluster 3 daylight random parameter mean of -17.373 confirmed a broadly protective effect despite substantial unobserved heterogeneity (SD = 6.697). Dark and unlighted environments correspondingly elevated injury risk. Agencies should prioritize installing or upgrading pedestrian-scale street lighting at crosswalk locations, particularly on undivided roadways and non-intersection driveway access points. High-intensity retroreflective crosswalk markings, in-pavement flashing warning lights, and RRFBs should be deployed at locations where nighttime pedestrian exposure is documented, protecting Safe Road Users. Second, driver inattentiveness was a significant contributing factor across Clusters 1 and 3, with large positive coefficients on O injury crashes (C1: +3.123; C3: +3.687), indicating that while inattentive-driver crashes frequently result in lower-severity outcomes, the behavioral pattern represents a pervasive and widespread risk at crosswalk locations. Targeted public awareness campaigns emphasizing distraction-free driving near crosswalk-controlled locations, combined with enforcement operations focused on distracted driving at high-pedestrian-volume intersections, can address this risk. Digital variable message signs placed in advance of crosswalk locations on higher-speed corridors can reinforce pedestrian presence and prompt driver alertness enhancing the elements of Safe Roads. Third, failure to yield to pedestrians, particularly during left and right turn movements at intersections produced the largest negative marginal effect on BC probability in Cluster 1 (-4.959), indicating that yield-phase violations systematically escalate crash outcomes. Leading Pedestrian Intervals (LPIs), which give pedestrians a head starts before vehicles receive a green signal, directly address this conflict type by increasing pedestrian conspicuity during the turn phase. Geometric interventions such as protected intersection designs, corner refuge islands, and channelized right-turn lanes with yield lines can further reduce turn-phase conflicts. At locations with high failure-to-yield rates, automated pedestrian detection systems with extended crossing phases should be considered. Fourth, roadway types specifically two-lane undivided (SD = 1.318) and four-or-more-lane undivided configurations (SD = 6.219) exhibited substantial unobserved heterogeneity in Cluster 1, indicating that the severity effect of these road types varies meaningfully across crashes in ways not fully captured by observed variables. This heterogeneity underscores the need for site-specific assessments rather than class-level policies. Raised crosswalks and pedestrian refuge islands on multilane undivided arterials can reduce vehicle speeds ensuring Safe Speeds and provide crossing relief, while road diets converting four-lane undivided roads to three-lane configurations with dedicated turn lanes have demonstrated pedestrian safety benefits on comparable facility types. Fifth, driveway departure crashes in Cluster 3 driven by driver distraction and occurring predominantly at non-intersection locations were further characterized by a non-trafficway area marginal effect of + 0.2062 on BC probability and − 0.1979 on O probability, indicating that crashes in private access environments consistently produce moderate rather than property-damage-only outcomes. High-visibility crosswalk markings, advance yield lines set back from driveways, and wayfinding signage alerting drivers to pedestrian crossings ahead of driveway approaches are practical near-term measures. Where land use generates high pedestrian volumes near driveway access points, consolidating driveways or restricting turn movements can eliminate the conflict class entirely. Table 6 summarizes these findings by mapping cluster-specific risk factors, and their associated model estimates directly to SSA tiers66, where Tier 1 refers to removing severe conflicts, Tier 2 refers to reducing vehicle speeds, Tier 3 indicates managing conflicts in time and Tier 4 indicates increasing attentiveness and awareness. Table 6Policy Level Implications Based on Safe System Approach.Risk FactorClusterSSA TierCountermeasuresDark, Not LightedC1, C3Tier 1: Remove severe conflictsInstall pedestrian-scale street lighting at crosswalk locations; deploy in-pavement flashing warning lights and RRFBs at unlighted crossings on arterials and driveway access points.Failure to Yield – ROW to PedestrianC1Tier 1: Remove severe conflictsImplement Leading Pedestrian Intervals (LPIs) at signalized intersections; install protected intersection geometry with corner refuge islands and channelized turn lanes with yield lines.Left/Right Turn MovementC1Tier 3: Manage conflicts in timeRestrict permissive left turns at high-pedestrian-volume intersections; implement exclusive pedestrian phases; use pedestrian hybrid beacons at unsignalized crossings on turn-heavy corridors.Four-or-more Lane Undivided RoadsC1, C3Tier 2: Reduce vehicle speedInstall raised crosswalks and pedestrian refuge islands on multilane undivided arterials; evaluate road diet conversions from four-lane undivided to three-lane configurations with dedicated turn lanes.Two-Lane, Two-Way RoadsC1Tier 2: Reduce vehicle speedDeploy advance pedestrian warning signs with flashers; install high-visibility ladder-style crosswalk markings; consider traffic calming on high-pedestrian-volume two-lane segments.Driver InattentivenessC1, C3Tier 4: Increase attentiveness and awarenessImplement targeted distracted driving enforcement campaigns near crosswalk-controlled locations; deploy variable message signs reinforcing pedestrian presence on high-pedestrian corridors.Driveway DepartureC3Tier 3: Manage conflicts in timeInstall advance yield lines and high-visibility crosswalk markings at driveway access points; consolidate or restrict driveways at locations with documented pedestrian activity; add wayfinding signage alerting drivers to pedestrian crossings ahead.Pedestrian/Pedalcyclist InvolvementC3Tier 4: Increase attentiveness and awarenessConduct outreach targeting vulnerable road user visibility and crossing behavior; install pedestrian countdown signals and accessible pedestrian signals at high-volume crossings; deploy sharrows and shared-use path designations near driveway access corridors.Intersection-Related CrashesC2Tier 1: Remove severe conflictsInstall pedestrian hybrid beacons or hawk signals at uncontrolled midblock and intersection crossings; evaluate signal timing optimization to reduce pedestrian exposure during high-conflict phases.Daylight ConditionsC1, C2, C3Tier 4: Increase attentiveness and awarenessUpgrade crosswalk retroreflectivity and marking standards to extend daylight-equivalent visibility into dusk and dawn periods; use high-contrast thermoplastic markings and retroreflective delineators at all crosswalk locations. Policy Level Implications Based on Safe System Approach. Conclusions Crosswalk-related crashes represent one of the most consequential yet preventable threats to vulnerable road users in urban environments. Pedestrians and cyclists, by virtue of their physical exposure and limited protection, bear a disproportionate share of severe and fatal outcomes when conflicts with motor vehicles occur at or near crosswalk locations. Understanding the precise conditions under which these outcomes arise is therefore not merely an academic exercise but a prerequisite for designing interventions that are both targeted and effective. This study contributes to that understanding by demonstrating that crosswalk crash severity is not governed by a single uniform set of risk factors but rather by distinct combinations of roadway geometry, driver behavior, and environmental conditions that operate differently across crash typologies. The cluster-specific analytical framework employed here moves beyond the limitations of pooled modeling approaches, revealing that crashes at intersections involving turn-phase right-of-way violations, low-speed yield-phase interactions, and distraction-induced driveway departures each constitute fundamentally different crash environments with distinct injury mechanisms. Recognizing this structural heterogeneity is essential for translating research findings into countermeasures that address the specific conditions driving severe outcomes rather than the averaged tendencies of a mixed crash population. The empirical findings provide a detailed and portrait of how severe crosswalk injuries arise across different crash contexts. Fatal and severe injury outcomes emerge most prominently in environments characterized by left-turn vehicle maneuvers at intersections, where drivers navigating competing traffic demands frequently fail to detect or yield to pedestrians lawfully occupying the crossing path. Multilane undivided roadway configurations amplify this risk by exposing pedestrians to multiple lanes of conflicting traffic without the benefit of physical separation or refuge opportunities. Adverse lighting conditions, particularly dark and unlighted environments, consistently elevated injury severity across all three crash clusters, reflecting the fundamental vulnerability of pedestrians when driver detection distances are shortened and reaction time margins are compressed. Driver inattentiveness further compounds the probability of severe outcomes by reducing the temporal window available for evasive response, a pattern observed most strongly in non-intersection driveway departure crashes where distraction-related contributing factors were disproportionately represented. These findings illustrate that severe crosswalk crashes are not random occurrences but are systematically concentrated in environments where physical design deficiencies, behavioral risk factors, and adverse lighting conditions converge to minimize the safety margin available to vulnerable road users. The random parameter specifications reveal an equally important dimension of crosswalk crash severity that aggregate models obscure: the effects of key risk factors vary substantially across individual crash events in ways that observed variables alone cannot fully explain. Daylight exposure, undivided road configuration, and driver inattentiveness each exhibit significant unobserved heterogeneity, indicating that their influence on injury severity is moderated by unmeasured contextual factors such as pedestrian crossing behavior, vehicle approach speed, sight distance conditions, and driver attentiveness at the moment of conflict. This heterogeneity carries direct implications for intervention design, as it cautions against the assumption that a countermeasure demonstrating effectiveness in one crash context will deliver equivalent safety benefits in another. The cluster-specific marginal effects analysis reinforces this point by demonstrating that the magnitude and direction of key predictors shift meaningfully across the three crash typologies, providing practitioners with actionable evidence on where particular treatments are most likely to yield measurable safety returns. Locations characterized by high left-turn volumes warrant protected intersection geometry and leading pedestrian intervals, multilane undivided arterials require raised crosswalks and pedestrian refuge islands, and non-intersection driveway corridors demand advance yield markings, sightline management, and targeted distracted driving enforcement to address the dominant risk profile of each crash environment. Several limitations temper the interpretation of these findings and define productive directions for future inquiry. Reliance on police-reported crash records introduces a recognized source of incompleteness, as distraction-related contributing factors are frequently underreported at the scene and property-damage-only events are often excluded from official records, potentially understating the true prevalence and influence of inattentiveness-related crashes. Pooling records across the 2017–2022 study period, while necessary to sustain adequate cluster and outcome sample sizes, precludes formal evaluation of temporal stability in cluster solutions and model parameter transferability across individual years. Crash mechanisms are not static; they evolve in response to changes in roadway design standards, enforcement intensity, vehicle safety technologies, and travel behavior patterns, dynamics that were particularly pronounced during and following the COVID-19 pandemic. Future research should address these constraints through year-specific or rolling-window cluster estimation, with consistency evaluated using the adjusted Rand index or normalized mutual information. Parameter transferability should be examined through likelihood-ratio tests, out-of-sample predictive validation, and hierarchical modeling frameworks allowing coefficients to vary temporally. Extension to multi-jurisdictional datasets would clarify whether the identified severity mechanisms are specific to Texas or broadly generalizable. Integration of real-time behavioral data, including video-derived pedestrian trajectories, connected vehicle telemetry, and pedestrian exposure counts, would further enrich future specifications and sharpen the precision with which practitioners can identify and prioritize interventions most capable of reducing crosswalk crash Acknowledgements The authors acknowledge the integration of Texas crash data (2017–2022) from the Texas Crash Records Information System (CRIS)67, along with publicly available transport infrastructure data, which supported the analysis in this study. Author contributions CRediT Authorship Contribution Statement- Swastika Barua: Conceptualization, Methodology, Formal analysis, Writing – original draft, Writing – review & editing; Subasish Das: Data curation, Supervision, Writing – review & editing; Michael Starewich: Formal analysis, Visualization; Tausif Islam Chowdhury: Formal analysis, Writing – original draft, Writing – review & editing; Amir Rafe: Writing – review & editing. All authors reviewed and approved the final version of the manuscript. Funding This study has not received any funding. Data availability The crash datasets used and analyzed during the current study were obtained from the Texas Department of Transportation (TxDOT) Crash Records Information System (CRIS). These data are subject to licensing restrictions and are therefore not publicly available but may be available from the corresponding author upon reasonable request and with permission. Declarations Competing interests The authors declared no potential conflicts of interest with respect to the research, authorship, and/or publication of this article. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement The crash datasets used and analyzed during the current study were obtained from the Texas Department of Transportation (TxDOT) Crash Records Information System (CRIS). These data are subject to licensing restrictions and are therefore not publicly available but may be available from the corresponding author upon reasonable request and with permission.