Non-parametric analysis of transient data: a pseudo-competing event approach

Transient data analysis, which evaluates the impact of ordinal time-dependent covariates on survival, poses unique challenges. In a motivating study investigating increasing mixed chimerism (IMC) as a biomarker for disease relapse for post-transplant leukemia patients, existing approaches prove inadequate for a fallacious abrupt decline to zero in IMC2 stage, coupled with inflated type I error control in risk comparisons. To address these limitations, we propose a novel non-parametric approach to enhancing estimation and hypothesis testing for transient data. By conceptualizing state transitions as pseudo-competing events, we reformulate the analysis as a competing events problem, which enables enlarged risk sets of later transient states, facilitating robust analysis and intuitive interpretation. Moreover, the estimation of survival probabilities is calibrated to mitigate systematic bias from the pseudo-competing transition risks. A non-parametric bootstrap approach is introduced for uncertainty quantification and statistical testing. Simulation studies demonstrate robust estimation, outperforming competing approaches. An application to the motivating data improves statistical testing, highlighting the method's broad applicability to transient data.