Quantifying the causal effect of speed cameras on road traffic accidents\n via an approximate Bayesian doubly robust estimator

Daniel J. Graham, Cian Naik, Emma J. McCoy, Haojie LiarXiv 2017 · DOI · arXiv

Abstract

This paper quantifies the effect of speed cameras on road traffic collisions using an approximate Bayesian doubly-robust (DR) causal inference estimation method. Previous empirical work on this topic, which shows a diverse range of estimated effects, is based largely on outcome regression (OR) models using the Empirical Bayes approach or on simple before and after comparisons. Issues of causality and confounding have received little formal attention. A causal DR approach combines propensity score (PS) and OR models to give an average treatment effect (ATE) estimator that is consistent and asymptotically normal under correct specification of either of the two component models. We develop this approach within a novel approximate Bayesian framework to derive posterior predictive distributions for the ATE of speed cameras on road traffic collisions. Our results for England indicate significant reductions in the number of collisions at speed cameras sites (mean ATE = -15%). Our proposed method offers a promising approach for evaluation of transport safety interventions.

Introduction

The effectiveness of speed cameras (SCs) in reducing road traffic collisions is a subject of significant debate among policymakers and the public. While many studies suggest that cameras improve safety, empirical estimates of their impact vary wildly—ranging from negligible effects to reductions of over 50%. A primary reason for this discrepancy is the difficulty of establishing causality in the presence of "Regression to the Mean" (RTM).

This paper addresses these challenges by introducing a novel approximate Bayesian "doubly robust" (DR) estimator. By applying this method to data from England, the study provides a rigorous quantification of the causal effect of speed cameras, separating the actual safety benefit of the intervention from the statistical noise of accident hotspots.

The Problem: Selection Bias and Regression to the Mean

Evaluating speed cameras is inherently difficult because they are not placed randomly. Authorities typically install cameras at "hotspots"—sites with a history of high accident rates. This creates a selection bias: if a site is chosen because it had an unusually high number of crashes in a given year, the number of crashes is likely to decrease the following year simply by chance, regardless of whether a camera was installed. This is known as Regression to the Mean.

Previous research often relied on simple before-and-after comparisons or basic outcome regression models. These methods frequently fail to account for the fact that the "treated" sites (with cameras) are fundamentally different from "control" sites (without cameras). Without adjusting for these differences, the perceived effectiveness of cameras is often inflated.

Methodology: A Doubly Robust Causal Framework

To overcome these biases, the authors employ a Doubly Robust (DR) estimation strategy. This approach combines two distinct statistical models to estimate the Average Treatment Effect (ATE):

  1. Propensity Score (PS) Model: This model predicts the probability of a site receiving a speed camera based on its characteristics, such as historical collision rates and traffic flow.
  2. Outcome Regression (OR) Model: This model predicts the expected change in collisions based on site-specific covariates.

The "doubly robust" property means that the ATE estimator remains consistent and asymptotically normal if either the PS model or the OR model is correctly specified. This provides a "two-chances-to-get-it-right" safeguard against model misspecification.

The Bayesian Extension

While standard DR estimators are frequentist, providing only a point estimate and standard error, this paper develops an approximate Bayesian framework. Using a Bayesian Bootstrap approach (specifically the Muliere and Secchi method), the authors generate a posterior predictive distribution for the ATE. This allows for a more nuanced understanding of uncertainty, providing "credible intervals" that represent the probability that the true effect falls within a certain range.

To capture complex, non-linear relationships between variables (like traffic volume and accident frequency), the authors utilize Generalized Additive Mixed Models (GAMMs) within their component models.

Data and Analysis

The study utilizes a comprehensive dataset from eight administrative districts in England, including Greater Manchester, Merseyside, and the West Midlands. The data includes:

  • 771 Treated Sites: Locations where fixed speed cameras were installed.
  • 4,787 Control Sites: Randomly sampled points on the road network to serve as a comparison group.
  • Collision Records: Data on Personal Injury Collisions (PICs) from the police "STATS 19" database covering 1999 to 2007.
  • Covariates: Factors such as site length, 85th percentile speeds, and Annual Average Daily Flow (AADF).

Key Findings

The analysis reveals a significant causal link between speed cameras and accident reduction, though the effect is more modest than naive studies suggest.

  • The Causal Effect: The Doubly Robust model estimates a mean reduction in collisions of approximately 15% (specifically between 14.4% and 15.5%).
  • Quantifying Bias: Naive models that did not account for confounding estimated a 33.6% reduction. The drop to 15% in the causal model suggests that more than half of the improvement observed in simpler studies is actually attributable to RTM and selection bias rather than the cameras themselves.
  • Statistical Certainty: The 95% credible interval for the DR model was [-21.8%, -7.3%]. Because this interval does not include zero, there is high statistical confidence that speed cameras do indeed reduce collisions.
  • Methodological Validation: The Bayesian point estimates were nearly identical to frequentist results (-14.36% vs -14.37%), validating the robustness of the approximate Bayesian approach.

Conclusion and Policy Implications

The findings confirm that speed cameras are an effective tool for road safety, contributing to a roughly 15% decrease in personal injury collisions. However, the study also highlights the importance of rigorous causal inference. By failing to account for the "hotspot effect," previous evaluations may have overpromised the benefits of speed cameras.

The approximate Bayesian doubly robust framework presented here offers a powerful new tool for transport researchers. It provides a flexible way to evaluate safety interventions while accounting for both selection bias and the inherent uncertainty in complex traffic data. Future work could extend this model to investigate "spillover effects"—whether cameras reduce accidents globally or simply displace them to nearby roads—and apply the method to other interventions like variable speed limits or infrastructure redesigns.

Summary updated

Abstract

This paper quantifies the effect of speed cameras on road traffic collisions\nusing an approximate Bayesian doubly-robust (DR) causal inference estimation\nmethod. Previous empirical work on this topic, which shows a diverse range of\nestimated effects, is based largely on outcome regression (OR) models using the\nEmpirical Bayes approach or on simple before and after comparisons. Issues of\ncausality and confounding have received little formal attention. A causal DR\napproach combines propensity score (PS) and OR models to give an average\ntreatment effect (ATE) estimator that is consistent and asymptotically normal\nunder correct specification of either of the two component models. We develop\nthis approach within a novel approximate Bayesian framework to derive posterior\npredictive distributions for the ATE of speed cameras on road traffic\ncollisions. Our results for England indicate significant reductions in the\nnumber of collisions at speed cameras sites (mean ATE = -15%). Our proposed\nmethod offers a promising approach for evaluation of transport safety\ninterventions.\n
BibTeX
@misc{graham2017quantifying,
  author = {Daniel J. Graham and Cian Naik and Emma J. McCoy and Haojie Li},
  title = {{Quantifying the causal effect of speed cameras on road traffic accidents\textbackslash\{\}n via an approximate Bayesian doubly robust estimator}},
  year = {2017},
  eprint = {1703.05926},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1703.05926},
}
Figure