Long-term effects of driver feedback on harsh braking behavior: insights from survival models

This study investigates the long-term effects of smartphone-based driver feedback on harsh braking behaviour using a large-scale naturalistic driving dataset. Data were collected through a 21-month experiment involving 31 car drivers and 24,904 trips, during which drivers were exposed to different feedback mechanisms across multiple phases. The analysis compares behaviour across baseline, feedback and post-feedback periods to assess both immediate and sustained effects of feedback interventions. Descriptive statistics and non-parametric tests indicate a significant reduction in harsh braking events during the feedback phase, confirming the short-term effectiveness of feedback in improving driving behaviour. However, a noticeable relapse is observed following the removal of feedback, suggesting challenges in maintaining behavioral improvements over time. To further explore the persistence of these effects, survival analysis techniques were applied, including Kaplan–Meier estimators, Cox Proportional Hazards models with frailty, Weibull Accelerated Failure Time (AFT) models with clustered heterogeneity and Random Survival Forests (RSF). The results consistently demonstrate a gradual decline in improved behaviour during the post-feedback phase, indicating that relapse occurs progressively rather than immediately, with survival probabilities declining to approximately 40.3% by the 150-trip mark. Key predictors of relapse include driver age, vehicle engine capacity and trip duration, with older drivers exhibiting more stable behaviour over time. In terms of model performance, the Weibull AFT model provided the best balance between interpretability and predictive accuracy, while the RSF model effectively captured non-linear relationships. Overall, the findings highlight the importance of continuous or reinforced feedback mechanisms to sustain safer driving behaviour and provide a methodological framework for assessing long-term behavioral dynamics using survival analysis.