From Canopy to Collision: A Hybrid Predictive Framework for Identifying Risk Factors in Tree-Involved Traffic Crashes

Technical Summary: From Canopy to Collision: A Hybrid Predictive Framework for Identifying Risk Factors in Tree-Involved Traffic Crashes

Problem Statement
Tree-involved crashes represent a critical subset of run-off-road (ROR) collisions, characterized by disproportionately severe injury and fatality outcomes compared to other crash types. National data indicates that collisions with trees account for approximately one-quarter of all fatalities involving fixed objects. The high severity stems from the biomechanical incompatibility between rigid tree structures and vehicle energy-absorbing systems; trees act as narrow, non-deformable objects that concentrate collision forces, often bypassing structural load paths and causing severe intrusion into the occupant compartment. Despite their severity, tree-involved crashes have received limited attention in crash severity research compared to broader fixed-object literature, and existing studies often rely on traditional statistical methods that struggle to capture complex, non-linear relationships and interaction effects among the multifactorial causes (vehicle, driver, roadway, and environmental factors).

Methodology
To address these gaps, the authors developed a hybrid analytical framework integrating machine learning (ML) with traditional statistical modeling to analyze the Crash Report Sampling System (CRSS) database spanning 2020–2023. The study utilized a sample of 3,778 tree-involved crashes, filtering for injury outcomes (excluding Property Damage Only) and aggregating injury severity into a binary classification: Fatal/Incapacitating (KA) versus Non-Incapacitating/Possible (BC).

The modeling process followed a four-step workflow:

1. CatBoost Classification: A gradient-boosting decision tree algorithm (CatBoost) was employed to identify key factors associated with binary crash injury severity. CatBoost was selected for its superior handling of categorical variables and ability to prevent target leakage through ordered boosting and ordered target statistics.
2. SHAP Interpretability: The SHapley Additive exPlanations (SHAP) tool was applied to the CatBoost model to quantify and visualize the marginal effects of top influential factors. This step provided feature importance rankings and directional effects (e.g., whether a factor increases or decreases severity risk).
3. Binary Logistic Regression: Guided by SHAP insights, a binary logistic regression model was developed to statistically estimate factor effects and validate the ML-derived importance measures. Multicollinearity was assessed using Cramér's V, and highly correlated variables were removed to ensure model stability.
4. Interaction Analysis: SHAP interaction plots were generated to examine the combined effects of key contributing factors, revealing non-linear relationships and synergistic risk mechanisms.

Key Contributions

• Hybrid Framework: The study advances crash severity analysis by integrating the predictive power and pattern recognition of ML (CatBoost) with the inferential transparency and effect-size quantification of logistic regression.
• Interaction Discovery: Unlike previous studies that often focus on main effects, this framework explicitly identifies and quantifies complex interactions between variables (e.g., lighting conditions × vehicle age, speeding × road surface).
• Comprehensive Variable Set: The analysis incorporates a broad spectrum of 26 independent variables, including driver behavior, vehicle attributes, roadway geometry, and environmental conditions, specifically tailored to the context of tree-involved collisions.

Key Results
The analysis identified several dominant predictors of crash severity:

• Restraint Use: The single most influential predictor. Unrestrained occupants were nearly three times more likely to experience fatal or incapacitating injuries (Odds Ratio [OR] = 2.90) compared to those wearing seatbelts. SHAP values confirmed this, showing strong positive contributions to severity for non-use.
• Vehicle Age: Older vehicles significantly increased severity risk (OR = 2.27 for older vs. newer), reflecting reduced crashworthiness and outdated safety features.
• Driver Impairment and Speeding: Alcohol/drug impairment (OR = 1.71) and speeding violations (OR = 1.50) were substantial predictors, linked to reduced control capabilities and increased impact forces.
• Airbag Deployment: Counterintuitively, airbag deployment was associated with higher severity outcomes. The authors interpret this not as a causal factor but as a proxy for high-energy impacts; airbags deploy only when crash sensors detect sufficient deceleration, indicating the crash was already severe.
• Environmental Factors: Dark, unlighted conditions and wet road surfaces were associated with increased severity, though the effects were often moderated by other variables.

Interaction Effects
The SHAP interaction plots revealed critical compounding risks:

• Speeding × Lighting: The severity risk of speeding is dramatically amplified in dark, unlighted conditions. While speeding in daylight showed moderate risk, speeding in dark-not-lighted environments generated the highest positive SHAP values, indicating a catastrophic convergence of reduced sight distance and excessive speed.
• Vehicle Age × Restraint Use: The safety deficiencies of older vehicles are most catastrophic when combined with restraint non-use. Unrestrained occupants in older vehicles exhibited the highest severity outcomes, while newer vehicles provided a baseline of protection that was significantly diminished without seatbelt use.
• Lighting × Vehicle Age: In dark conditions, drivers of newer vehicles showed slightly higher severity risks than those in older vehicles, potentially indicating risk compensation behaviors (e.g., driving faster) due to a false sense of security provided by advanced vehicle technology.

Significance and Claims
The paper claims that its findings provide critical insights for developing targeted, safe-system-based interventions. By validating ML-derived patterns with statistical rigor, the study confirms that human behavioral factors (restraint non-use, impairment, speeding) are the primary drivers of tree-involved crash severity. The authors argue that interventions must move beyond single-factor approaches to address the compounding nature of these risks.

Specific recommendations derived from the results include:

• Enforcement: High-visibility nighttime seatbelt enforcement and sobriety checkpoints, particularly along tree-lined corridors.
• Infrastructure: Implementation of shoulder widening, rumble strips, and selective removal of high-risk trees or installation of energy-absorbing barriers.
• Vehicle Policy: Incentives for vehicle retirement and safety upgrades to address the elevated risk posed by older vehicle fleets.
• Education: Public messaging that highlights the biomechanical severity of tree impacts and the specific dangers of combining speeding with low-visibility conditions.

The study concludes that while limitations exist regarding data granularity (e.g., lack of specific tree attributes) and potential reporting biases, the hybrid framework successfully elucidates the complex mechanisms underlying tree-involved crashes, offering a robust basis for problem-specific safety countermeasures.