Analyzing determinants of marriage survival by random survival forests

In social studies, the application of ensemble approaches such as random survival forests to define the most influential factors is increasing. As high precision statistical machine learning approaches, random survival forests are non-parametric and non-linear in nature and over the past decades have received considerable interests in analyzing time-to-event data. In this article, two algorithms of random survival forest with log rank split rule (RSF1) and random survival forest with log-rank score split rule (RSF2) in analyzing marriage survival are compared. In a cross sectional study, the information of 788 divorce applicants who had referred to randomly selected divorce offices from all provinces in Iran during 2017 and 2018 was gathered by a structured questionnaire. The effect of some demographic factors of selected divorce applicants and their partners along with their marriage and divorce information on their marriage survival were analyzed using R-language packages and finally RSF1 due to its lower prediction error estimate was selected. According to variable importance (VIMP) and minimal depth measures of RSF1, covariates of age, number of children ever born (CEB) and the number of years after marriage problems of couple starts were chosen as important predictors on marriage survival.