Semi-Parametric Modeling of Churn Confounding Competing Risks Using Time-Dependent Covariates Among Mobile Phone Subscribers in Kenya

Abstract

Mobile phone service providers are currently experiencing high churn rates. As a result, service providers are trying to develop ways to predict churn rates and uncover why subscribers’ churn occurs. However, the task of predicting churn in the mobile phone industry is complicated due to the large, sparse, and unbalanced nature of the data especially when competing risks are confounded by time-dependent covariates. This paper aimsto develop a semi-parametric model (the adjusted Cox model) by adjusting the extended Cox proportional hazards model to model competing risks confounded by time-dependent covariates and uses data from three mobile phone service providers in Mombasa and Kilifi Counties in Kenya to analyze and evaluate the validity and performance of the model. The paper establishes that the adjusted Cox model is a better model for predicting subscriber’s survival outcomes as well as for detecting the most influential covariates when competing risks are confounded with time-dependent covariates.