|
Abstract:
|
The logrank test (i.e., score test from a Cox proportional hazards (PH) model), is commonly used for survival analysis in randomized clinical trials. When the PH assumption is violated, the estimated hazard ratio is hard to interpret and the power of the logrank test can be suboptimal. Statistical strategies to guard against the potential adverse effects of non-PH include (i) use of a stratified logrank test based on pre-specified stratification factor(s), (ii) use of weighted logrank tests (e.g., “MaxCombo”) and (iii) a treatment comparison of restricted mean survival times. We propose a novel approach motivated by the observation that overall non-PH is often due to the overall population being a mixture of risk-based homogeneous subpopulations (“strata”), with PH evident within most risk strata. Our approach entails using a pre-specified algorithm to separate patients into risk-based strata and combining estimated stratum-level effects for overall inference. We describe the application of our 5-step stratified testing and amalgamation routine (5-STAR) using a real dataset. Simulations show that 5-STAR is a promising alternative to the aforementioned methods for tackling non-PH.
|
