Estimating Marginal Effects on Survival Outcomes for Clinical Interpretations and Decision Making

The Cox regression model is the most widely used method in survival analysis, but the hazard ratio has limited clinical interpretability. Survival or cumulative incidence curves are often used to illustrate treatment effects, yet there is no consensus on the most appropriate methods for generating these curves, particularly in the presence of confounding and competing risks. We reviewed original research articles published in a general medical journal in 2025, identified limitations in common practices, and describe a marginal effects framework to generate adjusted cumulative incidence curves that are coherent with the analysis model, anchored in absolute risk estimates, and adaptable to competing risks and clustered data. The framework targets the population-average difference in cumulative incidence between exposure groups, under the assumptions of no unmeasured confounding, positivity, and non-informative censoring. As a demonstration, we applied the framework to real-world data from Yek et al. (2025) and showed that it produced clinically interpretable estimates of hospital-acquired antimicrobial resistance risk that a single hazard ratio could not convey. To facilitate implementation, we describe each analytic step in detail, discuss available software, and share an R program with core routines written in C for computational efficiency.