📊 statistics

Estimating Nonlinear and Heterogeneous Determinants of Crash Severity Using Causal Machine Learning

Using causal machine learning on over 500,000 UK crash records, this study estimates that roads with speed limits of 50 mph or higher increase the likelihood of fatal or serious accidents by approximately 6–7 percentage points, with these effects being most pronounced during daytime under dry and favorable conditions.

Original authors: Nakachew Assefa Kebede, Biruhi Tesefaye Abeje, Hassan M. Al-Ahmadi

Published 2026-06-24

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Nakachew Assefa Kebede, Biruhi Tesefaye Abeje, Hassan M. Al-Ahmadi

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Picture: Why Do Some Crashes Hurt More Than Others?

Imagine you are trying to figure out why some car accidents result in serious injuries or death, while others only cause a little bit of damage. For a long time, researchers have looked at data to find patterns. But there's a problem: correlation is not causation.

Just because two things happen at the same time (like driving fast and getting hurt) doesn't mean one caused the other. Maybe the fast drivers were also driving in bad weather, or maybe they were tired. To really know if speed is the culprit, you need to compare two identical crashes: one happening at high speed and one at low speed, with everything else (weather, time of day, number of cars) being exactly the same.

Since we can't time-travel to create a perfect "identical twin" crash for every accident, this study uses a clever trick called Causal Machine Learning to simulate that comparison using real-world data from Great Britain.

The Ingredients: What They Studied

The Data: They looked at over 500,000 car crashes reported to the police in Great Britain between 2020 and 2024.
The "Treatment" (The Variable): They split these crashes into two groups:
1. High-Speed: Crashes on roads with a speed limit of 50 mph or more.
2. Low-Speed: Crashes on roads with a speed limit below 50 mph.
The Outcome: Did the crash result in a severe injury or death (Severe), or was it just a minor scratch (Non-Severe)?
The Confounders (The "Noise"): To make a fair comparison, they had to account for other factors like the time of day, the weather, the road surface (wet or dry), and how many cars were involved.

The Method: The "Fair Match" Game

The researchers used a sophisticated digital tool to play a game of "matchmaker."

The Propensity Score (The Matchmaker): Imagine you have a huge pile of high-speed crash reports and a huge pile of low-speed crash reports. The tool looks at the details of a specific high-speed crash (e.g., "It was Tuesday, sunny, dry road, 2 cars") and finds a low-speed crash that looks almost exactly the same.
Trimming the Edges: Sometimes, a high-speed crash happens in a situation where no low-speed crash ever happens (like a very specific type of highway exit). The researchers threw these "unmatchable" cases out of the study to ensure they were only comparing apples to apples.
The Weighing Scale (Overlap Weighting): They used a special mathematical scale to ensure the groups were perfectly balanced. This is like putting weights on a scale so that the "High-Speed" side and the "Low-Speed" side have the exact same average weather, time of day, and traffic conditions.

Once the groups were perfectly balanced, they could finally ask: "If we took a low-speed crash and magically made it happen on a high-speed road (keeping everything else the same), how much worse would the outcome be?"

The Results: The "Speed Penalty"

The study found a clear, causal answer:

The Main Finding: Being on a high-speed road (50 mph+) increases the chance of a crash becoming severe or fatal by about 6 to 7 percentage points compared to a similar crash on a slower road.
- Analogy: Imagine you have 100 identical minor fender-benders. If you move 6 or 7 of them to a high-speed road, those specific 6 or 7 would likely turn into serious injuries or deaths, whereas they would have been minor on a slow road.
The "It Depends" Factor (Heterogeneity): The danger of speed isn't the same everywhere. The study found that speed is most dangerous when:
- It is daylight.
- The road is dry.
- The weather is clear.
- Why? The researchers suggest that when conditions are perfect, drivers tend to feel confident and drive faster, which amplifies the damage. When it's raining, dark, or icy, drivers naturally slow down and drive more carefully, which actually reduces the extra danger that high speed usually adds.

The Tools: The "Super-Computers"

To get these answers, the researchers didn't use simple math. They used advanced Causal Machine Learning models (specifically called DR-Learner and Causal Forest).

Analogy: Think of traditional statistics as a simple ruler. It can measure length, but it struggles with complex, twisting shapes. These new AI tools are like a 3D scanner that can map out the complex, twisting relationships between speed, weather, and injury to find the true cause-and-effect hidden inside the data.

The Bottom Line

This paper proves that speed limits aren't just numbers; they are a direct cause of injury severity.

If you are on a road with a 50 mph limit, the risk of a crash turning into a tragedy is significantly higher than on a slower road, even if the weather and time of day are the same. However, this risk is highest when the weather is nice and the road is dry, because that's when drivers are most likely to push the speed limit.

The study concludes that to save lives, we need to manage speed not just in bad weather, but especially on clear, dry days when drivers feel safe enough to speed up.

Technical Summary: Estimating Nonlinear and Heterogeneous Determinants of Crash Severity Using Causal Machine Learning

Problem Statement
Traffic collisions remain a critical global safety issue, with the United Kingdom experiencing thousands of serious and fatal accidents annually despite overall safety improvements. Traditional research into crash severity has relied heavily on statistical association models (e.g., logistic regression, ordered probit) and predictive machine learning. However, these approaches face two primary limitations:

Causal Gap: Correlation does not imply causation. Existing models often fail to answer the policy-relevant question of what would happen to injury severity if a specific environmental factor (such as speed limit) were modified.
Structural Limitations: Crash data exhibits complex nonlinearities and unobserved heterogeneity. Traditional parametric models often require pre-specified functional forms and interaction terms, which may not capture the nuanced ways in which speed interacts with lighting, weather, and road surface conditions.

This study addresses these gaps by investigating the causal effect of high-speed road environments (defined as speed limits $\ge$ 50 mph) on crash severity in Great Britain, utilizing causal machine learning to estimate both average and heterogeneous treatment effects.

Methodology
The study employs a rigorous causal inference framework applied to 503,475 police-reported collisions from the STATS19 database (2020–2024). The analysis targets the binary outcome of "severe crash" (fatal or serious) versus "non-severe" (slight).

Data and Variables:
- Treatment: Exposure to a high-speed road environment (posted speed limit $\ge$ 50 mph).
- Outcome: Binary indicator of crash severity.
- Covariates: Nine pre-crash covariates were retained to ensure conditional exchangeability: collision year, day of week, number of vehicles, pedestrian crossing facilities, lighting conditions, weather conditions, road surface conditions, special site conditions, and carriageway hazards. Post-crash variables and potential leakage factors were excluded.
Causal Design:
- Propensity Score Estimation: An XGBoost classifier was trained to estimate the probability of exposure to high-speed conditions.
- Support Trimming: Observations with propensity scores outside the common support region ( $0.05 \le e(X) \le 0.95$ ) were trimmed, removing 23.17% of the training data to ensure comparability.
- Overlap Weighting: To target the population where treated and untreated units are most comparable, overlap weighting was applied. This resulted in a final effective sample size of approximately 229,318.
- Balance Diagnostics: Post-weighting balance was verified using weighted standardized mean differences (SMD), with the largest absolute SMD being 0.031 (well below the 0.10 threshold).
Estimators: Two causal machine learning estimators were implemented to estimate the Average Treatment Effect (ATE) and Conditional Average Treatment Effect (CATE):
1. DR-Learner: A doubly robust estimator combining outcome modeling and treatment modeling.
2. CausalForestDML: A generalized random forest approach specifically designed to detect heterogeneity in treatment effects.
Interpretability: SHAP (SHapley Additive exPlanations) values were derived from a surrogate XGBoost model fitted to the CausalForestDML estimates to identify the drivers of treatment effect heterogeneity.

Key Results

Average Treatment Effect (ATE): Both estimators confirmed a statistically significant positive causal effect of high-speed environments on crash severity.
- DR-Learner: Estimated an overlap-weighted ATE of 0.0627 (95% CI: 0.0616–0.0639).
- CausalForestDML: Estimated an ATE of 0.0682 (95% CI: 0.0677–0.0686).
- Interpretation: Collisions in high-speed environments are approximately 6 to 7 percentage points more likely to result in a fatal or serious injury compared to similar collisions in lower-speed environments, after adjusting for covariates. This causal estimate is lower than the unadjusted risk difference (8.81%), indicating that observed pre-crash characteristics explain part of the raw disparity.
Heterogeneity: The effect of speed is not uniform. The CATE distribution shows substantial variation, with the median effect at 0.0716.
- Contextual Drivers: The severity penalty is most pronounced during daylight, on dry roads, and under fine weather conditions.
- Mitigating Factors: The effect diminishes in darkness, on wet/damp surfaces, and under frost/ice.
- SHAP Analysis: Lighting conditions and road surface conditions were identified as the primary drivers of heterogeneity, suggesting that drivers may exercise greater caution (reducing speed) in poor visibility or traction conditions, thereby attenuating the marginal severity penalty of high-speed limits.

Contributions and Significance
The paper claims three primary contributions:

Methodological Advancement: It demonstrates the application of causal machine learning (specifically DR-Learner and CausalForestDML) to crash severity analysis, moving beyond predictive modeling to provide evidence on causal mechanisms.
Policy-Relevant Evidence: By quantifying the causal impact of speed environments, the study provides a basis for targeted speed management policies. The findings suggest that interventions should not be limited to adverse conditions (e.g., fog or ice) but must also address "normal" operating conditions (daylight, dry roads) where the marginal severity penalty is highest.
Contextual Understanding: The research highlights that speed is a context-dependent determinant of severity. The interaction between speed and environmental factors (visibility and traction) is critical, challenging the notion of a single global speed penalty.

Limitations and Scope
The authors acknowledge the inherent limitations of observational data, including the inability to capture unobserved confounders such as driver impairment, detailed vehicle safety features, or emergency response times. Consequently, the causal interpretation relies on the assumptions of conditional exchangeability and overlap. The results are interpreted as effects for the population represented by the observed data and the specific modeling design, rather than as universal laws. The study focuses on the collision level rather than individual casualty trajectories.