Abstract:
|
Time-to-event endpoints are very common in clinical research, both randomized clinical trials and observational studies. Both the traditional product-limit estimator of K-M curve and the partial likelihood function from Cox proportional hazard model rely on risk sets defined at each event or censoring time. The risk sets beyond the first event time consist of subjects who have not failed previously, and therefore the balance of confounders between groups is lost regardless of randomization. While valid for testing the null hypothesis of no treatment effect, quantities like the hazard ratio do not allow for a causal interpretation. Many methods have been proposed to address causal inference for survival analysis, such as (1) inverse probability of censoring weighted estimator, estimating equation, and targeted maximum likelihood estimator (classical and one-step optimization) for counterfactual average survival curve; and (2) accelerated failure time model, models based on time-transformation, and additive hazard model for effect size. In this talk we will review their underlying assumptions, strengths and limitations, focusing on practical implementation in real-world studies.
|