Introduction
Freight transportation is a cornerstone of the modern economy, yet its physical footprint—massive trucks navigating dense urban environments—presents significant safety risks. In the Los Angeles Greater Metropolitan Area (LAMA), freight truck crashes are notably more severe than those involving only passenger vehicles. However, the impact of these incidents is not distributed equally across the landscape. This research introduces a Deep Counterfactual Inference (DCI) model to move beyond simple correlations, uncovering the causal drivers of crash severity and addressing the critical issue of "spatial justice."
The Problem: Beyond Correlation to Spatial Justice
While existing traffic safety research effectively identifies high-risk locations, it often fails to explain the underlying causal mechanisms that make crashes more severe in specific neighborhoods. Low-income and minority populations frequently reside near major freight corridors and industrial zones, bearing a disproportionate burden of traffic risk.
Traditional statistical and machine learning models often treat socioeconomic factors as simple correlations or "black boxes." They can identify that a crash occurred in a disadvantaged neighborhood, but they struggle to answer the counterfactual question: Would this crash have been less severe if the infrastructure—such as street lighting or road design—had been improved? This study seeks to provide a causal framework to evaluate these hypothetical interventions, offering a data-driven path toward more equitable urban planning.
Methodology: Deep Counterfactual Inference (DCI)
The core innovation of this study is the DCI model, a multi-task deep learning architecture designed to estimate counterfactuals—outcomes that would have happened under different conditions. Unlike standard predictive models that only analyze factual data (what actually occurred), the DCI model allows researchers to simulate "what-if" scenarios.
Technical Architecture
- Multi-Task Learning: The model utilizes two distinct "heads"—one for factual scenarios and one for counterfactual scenarios. These heads share a joint representation layer to capture complex, non-linear interactions between road conditions, driver behavior, and the environment.
- Calibration Layer: To process a heterogeneous mix of continuous (population density), binary (alcohol involvement), and ordinal (weather) variables, the model employs a specialized calibration layer to standardize inputs.
- Causal Metrics: The researchers utilize Individual Treatment Effect (ITE) and Average Treatment Effect (ATE) to quantify the degree to which specific factors, such as functioning traffic control devices, alter the probability of a fatality or severe injury.
- Propensity Score Matching: To ensure that hypothetical scenarios remain grounded in reality, propensity scores are used to match observed crashes with similar counterfactual cases.
Data Integration
The study analyzes a decade of data (2012–2022) from the Los Angeles Greater Metropolitan Area, integrating three primary sources:
- SWITRS: 28,626 freight truck-related crash records from the California Statewide Integrated Traffic Records System.
- American Community Survey (ACS): Socioeconomic indicators, including household income, minority percentages, and population density.
- OSMnx: Road network characteristics, such as intersection density and segment lengths, extracted from OpenStreetMap.
Key Insights and Causal Findings
The DCI model revealed several critical insights into the nature of freight safety and urban inequity:
- The Rural Severity Gap: While central Los Angeles experiences a higher density of crashes, rural and sparsely populated regions (such as parts of Ventura and San Bernardino counties) see higher severity levels. This is largely attributed to increased emergency response times in remote areas.
- Infrastructure as an Equity Lever: The model demonstrated that improving street lighting in low-income areas resulted in a significantly larger decrease in crash severity (a 0.118 level drop) compared to the same improvement in high-income areas. This suggests that historical infrastructure neglect is a primary driver of current safety disparities.
- Vulnerability of Non-Motorists: Pedestrian involvement was identified as the single greatest "treatment" factor, increasing average severity by 0.85 levels. This effect is particularly pronounced in areas with high intersection density, highlighting failures in "shared space" designs where freight routes overlap with pedestrian paths.
- Persistent Inequity: Even in scenarios where traffic control devices are fully functional, minority-dense communities continue to experience higher crash severity. This indicates that spatial injustice is deeply embedded in broader land-use patterns and road geometry beyond simple equipment maintenance.
Results and Model Performance
The DCI model outperformed several state-of-the-art causal inference methods, including Causal Forests (CF), Double Machine Learning (DML), and Generative Adversarial Networks (GANs).
- Accuracy: DCI achieved the lowest Mean Squared Error (0.882 for factual and 0.978 for counterfactual scenarios), outperforming the GAN approach (1.039/1.111).
- Causal Impact (ATE) Highlights:
- Weather: Favorable weather conditions reduced severity by an average of 0.325 levels.
- Lighting: Enhanced lighting reduced severity by 0.267 levels.
- Driver Behavior: Alcohol or drug involvement increased severity by 0.198 levels, while improper turning added 0.137.
Conclusion and Future Directions
This research provides a robust argument for "targeted" urban planning. By demonstrating that infrastructure investments in underserved neighborhoods yield a higher "safety return" than in wealthier ones, the study offers a framework for prioritizing interventions that promote spatial justice.
While the current model is limited by its reliance on static data and a lack of real-time human factor variables (such as driver fatigue), it sets the stage for future work. Integrating real-time telematics and behavioral data could eventually lead to dynamic "early warning" systems, further protecting vulnerable communities from the disproportionate risks of freight traffic.
Summary updated