Causal inference in transportation safety studies: Comparison of potential outcomes and causal diagrams

TL;DR

This study compares two causal inference frameworks, Potential Outcomes and Causal Diagrams, in transportation safety, estimating the effect of pavement marking retroreflectivity on nighttime crashes using real data.

Contribution

It evaluates the applicability and differences of these frameworks in a transportation safety context, highlighting their sensitivity to assumptions and method choices.

Findings

Increased retroreflectivity reduces nighttime crash probability.

Causal effect estimates vary significantly with the method used.

Assumption violations impact the reliability of causal estimates.

Abstract

The research questions that motivate transportation safety studies are causal in nature. Safety researchers typically use observational data to answer such questions, but often without appropriate causal inference methodology. The field of causal inference presents several modeling frameworks for probing empirical data to assess causal relations. This paper focuses on exploring the applicability of two such modeling frameworks---Causal Diagrams and Potential Outcomes---for a specific transportation safety problem. The causal effects of pavement marking retroreflectivity on safety of a road segment were estimated. More specifically, the results based on three different implementations of these frameworks on a real data set were compared: Inverse Propensity Score Weighting with regression adjustment and Propensity Score Matching with regression adjustment versus Causal Bayesian Network. The effect of increased pavement marking retroreflectivity was generally found to reduce the probability of target nighttime crashes. However, we found that the magnitude of the causal effects estimated are sensitive to the method used and to the assumptions being violated.