Comparing Machine Learning Methods for Survival Analysis: An Application to Tenure at the Ministry of Foreign Affairs

Data on ministerial tenure often covers a vast amount of information about individual ministers. This type of high-dimensional data provides an ideal ground to compare theories of cabinet change with machine learning algorithms that select the most informative features in data in order to improve prediction. This paper uses new data covering the tenure of Foreign Ministers during the long 19th Century to compare the results and predictions of a Cox model of ministerial tenure with the output of two machine learning methods applied to survival analysis: Random Survival Forests (Ishwaran et al. 2008) and Component-wise Gradient Boosting (Hofner et al. 2014). The paper makes two contributions to the literature: it tests the prediction power of theories of ministerial change, and compares two machine learning methods applied to survival analysis, one of which is not subject to the limitations of the proportional hazard assumption.