Statistical issues in survival analyses (Part II)
January 16, 2023
In an article recently published online in The Biometrical Journal, Handorf et al seek to understand what is the best way to model survival data in the presence of nonproportional hazards, perhaps from the effect of confounding variables. The authors were particularly interested to compare this for real world data. The authors therefore went through a process of identifying several existing methods that exist in the current literature to handle such problems. They first discussed propensity score methods, which had been originally introduced by Rosenblum and Rubin in 1983. The original paper laid out the methodology in a lot of detail. Overtime, the use of their methods has either been criticized or continued to be utilized. The authors decided to restrict their interest to weighting or rather IPTW, inverse propensity score weighting. Next they discussed methods that exist under nonproportionality. One of the major methods they discussed was restricted mean survival time (RMST) since this can handle estimating survival under non-proportional hazards. After this they go to discussion of the Cox proportional hazard model where they discussed allowing the hazard to vary by natural logarithm of time and they used a piecewise constant treatment effect. They also presented parametric accelerated failure