Asymmetric effects of dynamic usage behavior on duration in subscription-based online service

In this article, we develop a model to predict the lifetime duration of customers for a subscription-based online content provider. In contrast to prior research, this article employs a dynamic measure of usage and compares alternate model specifications. A proportional hazard model is used with time-varying usage covariates obtained from Web logs while controlling for initial purchase characteristics such as purchase price and billing frequency. We test and find evidence for violation of the proportionality assumption in the hazard model and conclude that a stratified hazard model provides the best model fit. The results show that customers who are billed monthly for annual subscriptions maintain their subscriptions longer than do customers billed quarterly, when compared to annual subscriptions. Furthermore, dynamic changes in usage have an asymmetric effect on duration, with increases in usage having a stronger effect than do decreases in usage. Managers can use this model to identify the customers to be targeted for both retention and upgrade.