Ex-post evaluation of transport interventions with causal mediation analysis

Numerous ex-post evaluation studies have answered the questions on whether an intervention made a difference to the transport system using causal inference models, whereas the questions on why and how remain underexplored. This paper contributes to the transport domain by introducing an approach for causal mediation analysis presented in the statistics literature. The approach builds on the potential outcome framework for causal inference, and permits the use of various parametric and nonparametric models. By conducting causal mediation analysis, the total causal effect can be decomposed into direct effect and indirect (or mediation) effect that relays through an intermediate variable known as mediator. Simulations are presented to show the performance of the approach in handling typical research questions in the transport domain. The simulation results suggest that the approach performs well in estimating direct and causal mediation effects under various scenarios. The approach is applied to a case study of carsharing services in San Francisco Bay Area, exploring the direct effect of carsharing on travel behaviors, as well as the mediation effect via household car ownership as mediator. Nonlinear effects have been taken into account by adopting the generalized additive model. We find that carsharing services could increase the use of transit and non-motorized modes both directly and indirectly via influencing household car ownership. Moreover, the proportion of the total effect that is mediated is around 35% and 27.5% for transit and non-motorized trips, respectively. The approach could be used more generally in the transport domain for program evaluation.