Causality in Travel Mode Choice Modeling: A Novel Methodology that Combines Causal Discovery and Structural Equation Modeling

Causal discovery identifies causal relationships between variables in a dataset. This study investigates the potential of causal discovery in extracting causal connections from transportation behavioral data. To do so, four causal discovery algorithms are tested: Peter-Clark (PC), Fast Causal Inference (FCI), Fast Greedy Equivalence Search (FGES), and Linear Non-Gaussian Acyclic Models (LiNGAM). Their performances are compared to determine the most appropriate algorithm for travel choice modeling. Next, we propose a novel methodology to combine causal discovery with structural equation modeling (SEM) to model travel mode choice. This modeling approach can overcome some of the limitations of SEM, by combining both the strengths of causal discovery and SEM. The results show that LiNGAM best captures causality in transportation behavior modeling, among the four algorithms tested since the LiNGAM-based SEM achieved the lowest values of Chi-square, Root Mean Square Error of approximation (RMSEA), along with greater than 0.95 Comparative Fit Index (CFI), Goodness-of-Fit Index (GFI), and Adjusted Goodness-of-Fit index (AGFI), Normed Fit Index (NFI), and Tucker-Lewis Index (TLI). The modeling results provide insights in causal relations leading to choosing private vehicles, public transit, or walking as a travel mode. The analyses are conducted on data from the 2017 National Household Travel Survey in the New York Metropolitan area.