Survival analysis (time-to-event analysis) comprises statistical methods for evaluating time periods until the occurrence of a defined event. In clinical research, the event is often death, disease progression, recurrence, hospitalization, or treatment discontinuation. A characteristic feature is that the event does not occur for all participants during the observation period. This results in censored data, which cannot be adequately handled using classical mean comparisons.
Survival analyses are central to oncology, cardiology, and chronic diseases, as well as medical technology (e.g., time to implant failure) and pharmacovigilance (e.g., time to onset of an adverse effect). They support benefit-risk assessment, the planning of study endpoints, and the interpretation of evidence of efficacy in regulatory dossiers. For sponsors and CROs, they are also operationally relevant because they define requirements for follow-up processes, event adjudication, and data quality.
What is measured in a survival analysis?
The target variable is a duration from a defined starting point to the event. The start can be randomization, start of treatment, surgery date, or diagnosis. It is crucial that the start and event are clearly operationalized to avoid systematic bias. Typical endpoints include overall survival, progression-free survival, disease-free survival, or time to treatment discontinuation.
A key feature is censoring: if the event is not observed by the end of the follow-up, the individual is included in the analysis with their known minimum observation time. Study withdrawal or loss to follow-up also frequently leads to censoring. The assumption that censoring is non-informative must be critically examined, as estimators may otherwise be biased. In practice, reasons for discontinuation are therefore documented and alternative assumptions are tested in sensitivity analyses, for example, through worst-case scenarios or inverse probability weighting.
Censoring, Follow-up, and Data Quality
The quality of survival data depends heavily on consistent follow-up and the clean recording of events. In multicenter studies, clear definitions in the clinical trial protocol and Standard Operating Procedures are required so that all trial sites evaluate events consistently. For time-critical endpoints, data sources should be prioritized (e.g., hospital records, registry data, death certificates) and uniform processes for query management and source data verification should be established.
Practical pitfalls include inconsistent date entries, delayed reporting, or unclear event classification. Especially for progression endpoints, standardized imaging evaluation and, if necessary, an independent review are advisable. For devices or implants, it must be defined when a “failure” occurs (e.g., revision surgery, loss of function, or radiological criterion). Additionally, it must be considered that subsequent therapies or a crossover can change the observed event profile, thereby complicating interpretation.
The choice of analysis population is also important. In the intention-to-treat population, the effect is mapped according to randomization, while per-protocol evaluations more closely reflect the treatment actually received. From a regulatory perspective, both viewpoints are often expected, with deviations in protocol compliance, rescue medication, and subsequent therapy needing transparent documentation.
Typical Methods: Kaplan-Meier, Log-Rank, and Modeling
The Kaplan-Meier method estimates the survival function as a step curve and allows for reporting median survival times or survival probabilities at specific time points. The log-rank test is frequently used to compare two groups. It tests whether the curves differ systematically without assuming a specific form of the hazard function.
To quantify the treatment effect in randomized trials, a Cox proportional hazards model is often used. The result is typically a hazard ratio with a confidence interval. Interpretation requires care: a hazard ratio describes a relative risk per time point under the assumption of proportional hazards. In the event of significant deviations (e.g., delayed onset of effect), time-dependent effects, stratified models, or evaluations via restricted mean survival time may be more appropriate. In studies with competing risks (e.g., death before progression), suitable methods must also be selected.
Regulatory Expectations and Reporting
Regulatory authorities expect transparent definitions of time points, events, and censoring rules, as well as robust sensitivity analysis. In clinical trials under EU Regulation 536/2014, data integrity, traceability, and the documentation of key analytical steps must be ensured. For medicinal product trials, the ICH E9 principles on statistical planning and the ICH E6(R3) requirements for quality management and data quality are also relevant.
In the Clinical Study Report, Kaplan-Meier curves, number at risk, event counts, censored observations, and the specification of the models should be presented transparently. Typical sensitivities concern alternative censoring rules, different analysis populations (intention-to-treat vs. per-protocol), and the handling of missing data. For interim analyses or adaptive designs, proper control of the alpha error is required. For real-world evidence analyses, additional requirements for confounding control and data source validation apply.
FAQ
Why are special methods needed for censored data?
Because the actual event time is unknown for censored observations. Survival methods utilize information up to the last known time point and avoid biases that would arise from ignoring or incorrectly imputing data.
What does a hazard ratio of less than 1 mean?
A hazard ratio of less than 1 indicates a lower risk of the event in the test group compared to the control group, assuming proportional hazards. For clinical interpretation, absolute effects and the shape of the Kaplan-Meier curves are also important.
Which metrics are helpful in addition to the hazard ratio?
Median survival times, event rates at fixed time points (e.g., 12-month rate), and restricted mean survival time are frequently reported. These measures support clinical assessment, especially when the proportional hazards assumption is questionable.
Regulatory References (Selection)
- EU Clinical Trials Regulation: Regulation (EU) No 536/2014
- ICH E6(R3): Good Clinical Practice (Quality Management, Data Integrity)
- ICH E9: Statistical Principles for Clinical Trials