Premature casual carpooling in Texas: analyzing customer churn in the Metropia experiment with survival analysis and machine learning

Abstract

This study makes a timely contribution to peer-to-peer mobility research by employing a hybrid analytical framework that integrates survival analysis (Cox Proportional Hazards model) with advanced machine learning techniques (XGBoost and SHAP values) to uncover nuanced patterns of user churn in the Metropia app. The Metropia experiment, a peer-to-peer carpooling program, was conducted in Austin and El Paso, TX, with a sample of 3,227 users who first engaged with the app between January 1, 2016, and December 31, 2018. By connecting sociodemographic factors, trip characteristics, and app incentives to user retention, this study offers a data-driven view of behavior on mobility platforms. Results show that older users, females, and telecommuters are more likely to stay, while higher-income households, larger families, long commutes, and experienced drivers show higher churn. Several app features show complex, non-linear effects on disengagement. Interpreted through the lenses of transaction cost economics, network effects, and prospect theory, these findings highlight the importance of reducing perceived user costs and enhancing value through thoughtful app design. The study not only identifies critical behavioral and structural challenges facing casual carpooling but also sheds light on its premature development stage. It offers actionable insights for platform developers and planners to enhance engagement and retention, while calling on local governments to support shared mobility through strategic policy and urban design. This research ultimately advances the discourse on sustainable urban transportation and the future of the sharing economy.