More powerful logrank permutation tests for two-sample survival data

Abstract

Weighted logrank tests are a popular tool for analysing right-censored survival data from two independent samples. Each of these tests is optimal against a certain hazard alternative, for example, the classical logrank test for proportional hazards. But which weight function should be used in practical applications? We address this question by a flexible combination idea leading to a testing procedure with broader power. Besides the test's asymptotic exactness and consistency, its power behaviour under local alternatives is derived. All theoretical properties can be transferred to a permutation version of the test, which is even finitely exact under exchangeability and showed a better finite sample performance in our simulation study. The procedure is illustrated in a real data example.