Cautionary use of RMST regression in the presence of time-dependent covariates: insights from extensive simulation studies

The Cox proportional hazard model is commonly used for survival analysis in medical research but requires the proportional hazard assumption to be satisfied.This assumption is often violated when covariates are timedependent, such as in cancer vaccine studies where treatment effectiveness changes over time.The restricted mean survival time (RMST) regression model offers an alternative approach for non-proportional data.However, research on RMST regression performance with time-dependent covariates remains insufficient.This study examined RMST regression performance in various simulation settings involving time-dependent covariates and compared results with joint modeling analysis.We employed the RMST regression method proposed by Zhang et al. (2022), which uses a generalized linear model framework, splits covariates into time-dependent and timeindependent components, and adopts inverse probability censoring weight (IPCW) to accommodate censoring.Based on extensive numerical research and real data examples, we found several drawbacks of RMST regression.The empirical type I error rate was not appropriately controlled under the nominal level in RMST regression, while joint modeling did not suffer from this problem.However, joint models faced convergence issues with dichotomous time-dependent covariates under Weibull survival scenarios.In conclusion, RMST regression should be applied cautiously when proportional hazards assumptions are violated, particularly with time-dependent covariates present.