Clinical trials in general and oncology specifically strive to follow patients for a long time. The primary endpoints tend to be mortality, overall survival (os), or disease-free progression. The extended timeline of evaluation requires specific statistical procedures to analyze time to event data. It is a useful tool in clinical research and provides invaluable information about the efficacy of an intervention. Although not unique to cancer clinical trials, the medical literature often reports Cox's proportional hazards regression model (due to impact of covariates affecting the survival rates) when performing a survival analyses. The term "survival" analysis historically was used when the primary endpoint was mortality but now we use it to measure time to event such as time in remission for example.
The accurate interpretation of survival analysis is plagued by confounding and more specifically censoring introduced by clinical trial design. For the point of this post let's define censoring and why it becomes problematic--and why you should know about it. I am going to simplify it a bit but imagine inconstancies within clinical trial findings if data is missing from certain study participants. This is important in cancer clinical trials with OS as an endpoint because when they are loss to follow-up before the end of the trial they must be censored. We don't know if they experienced the clinical endpoint because of non-occurrence of outcome event or death (if not a specific endpoint of trial), so we must adjust for this in the data. Various methods are used to analyze censored data but the more robust methods widely used in survival studies are likelihood-based approaches (survival analysis methods) that adjust for the occurrence of censoring in each observation, and allow retention of more data.
The topic of the manuscript that I am reviewing addresses the complexity of censoring and classifies the survival techniques as either nonparametric or parametric and an actual blend of the two that is beyond our interest here. Typical statistical methods are not suitable for complex censoring mechanisms, especially if we need to consider both loss to follow-up and end of study censoring. The magnitude of the bias depends primarily on the proportion of patients who are censored and on the hazard ratio between the group of patients remaining on study. Potential solutions include revising end points to include inadequate response and initial signs of clinical progression as treatment failure. In addition, all patients should be observed until progression, and sensitivity analyses performed as indicated.