Attrition Modelling for Online Media Users by Cox Proportional Hazards

Abstract

In this work survival analysis was used to analyse propensity to churn for online writers of a news website in Kenya known as hivisasa.com. The study sought answers on which covariates were major determinants of writer attrition from the online platform, their statistical signi cance and magnitude. A total of one hundred and four writers who had at-least one publication for January 2019 formed part of the study sample. The sample historical data for January to July was used to determine writers who churned within the period and those that were retained. Previous literature on attrition research was reviewed and the study settled on survival methods in order to address time to event and manage censored data. Descriptive analysis was handled by tting Kaplan-Meier curves to visualize the retention curves of various categories of the covariates. Log-Rank test was then used to test the statistical signi cance of the various di erences observed. Cox proportional hazard was tted on the data including all covariates to determine the magnitude of hazard risk. Three of the covariates that is gender, number of articles published by a writer and category of articles done by the writer were signi cant in explaining writer attrition risk and magnitude. The results showed a high churn rate among female writers, writers publishing non political content on the site as well as publishing less than 148 articles for the study period. On the other hand three covariates; time spent on the platform from subscription, location of a writer and level of education were not statistically signi cant in explaining writer attrition. Even though these covariates lacked statistical signi cance Cox regression coe cients revealed that the magnitude of risk varied across them. Level of education graduate and time spent on the platform of more than 250 days reduced the chances of a writer churning 12% and 19% respectively in comparison to the reference variable holding for the e ect of other covariates. The model performance was validated by tting a ROC curve to ascertain how best the model was able to t the data. The ROC curve had an AUC of 87% which means the model had a 87% chance of predicting a churned writer as so.