Optimal three-group splits based on a survival outcome

In clinical research it is often desirable to discretize a continuous or ordered covariate. In this paper, we investigate the use of various running ordered logrank tests for finding optimal 3-group splits based on a survival outcome and a single covariate. We first present a successful application of using the modified ordered logrank test (MOL) to find three prognostic groups on a myeloma dataset. We then evaluate through simulations the performance of the running ordered logrank tests and a hierarchical method based on recursive partitioning in different scenarios: (1) when the true underlying distribution has three-groups, (2) when there is a linear relationship between covariate and outcome, and (3) when there is no association between covariate and outcome. We conclude that the MOL is the most robust among all versions of the running ordered logrank tests if the underlying distribution truly has three-groups, although further research could help define when the MOL is the statistic of choice more generally for finding optimal 3-group splits.