Citation

Zhang Y, Li H, Ren G. Accid. Anal. Prev. 2021; 165: e106507.

Copyright

(Copyright © 2021, Elsevier Publishing)

DOI

10.1016/j.aap.2021.106507

PMID

34856506

Abstract

Numerous evaluation studies have been conducted on a variety of road safety measures. However, the issue of treatment heterogeneity, defined as the variation in treatment effects, has rarely been investigated before. This paper contributes to the literature by introducing generalized random forests (GRF) for estimation of heterogeneous treatment effects (HTEs) in road safety analysis. GRF has high functional flexibility and is able to search for complex treatment heterogeneity. We first perform a series of simulation experiments to compare GRF with three causal methods that have been used in road safety studies, i.e., outcome regression method, propensity score method, and doubly robust estimation method. The simulation results suggest that GRF is superior to these three methods in terms of model specification, especially with the existence of nonlinearity and nonadditivity. On the other hand, a large dataset is required for accurate GRF estimation. Then we conduct a case study on the UK's speed camera program. Our results indicate significant reductions in the number of road accidents at speed camera sites. And the heterogeneity in treatment effects is found to be statistically significant. We further consider the associations between the baseline accident records, traffic volume, local socio-economic characteristics, and the safety effects of speed cameras. In general, the effect of speed cameras is larger at the sites with more baseline accident records, higher traffic volume, and in more densely-populated and deprived areas. Several policy suggestions are provided based on these findings. The evaluation of HTEs likely offers more comprehensive information to local authorities and policy makers, and improves the performance of speed camera programs. Moreover, GRF can be a promising approach for revealing treatment effect heterogeneity in road safety analysis.

Language: en

Keywords

Causal machine learning; Generalized random forest; Heterogeneous treatment effects; Road safety evaluation; Speed enforcement cameras