Survival Analysis

Survival analysis is a collection of statistical methods designed to analyze time to event data, where the outcome of interest is the time elapsed until a specific event occurs, such as death, disease recurrence, treatment failure, or hospital discharge. The defining feature of survival analysis is its ability to handle censored observations, meaning participants who have not experienced the event by the end of the study or who are lost to follow up. Standard statistical techniques that ignore censoring produce biased estimates, making survival analysis the standard approach when participants are followed for different lengths of time and not all of them experience the event during the observation period.

In clinical and epidemiological research, many outcomes are inherently time dependent. Simply comparing the proportion of participants who experienced an event at the end of a study ignores when events occurred and discards information from participants with incomplete follow up. Survival analysis keeps the information in time to event data by using every participant's contribution up to the point of censoring or event occurrence. This is critical for designing and analyzing clinical trials in oncology, cardiology, and infectious disease, where event rates accumulate over time and treatments may delay rather than prevent events. Survival analysis also supports health economic modeling by providing the time based transition probabilities needed for cost effectiveness evaluation.

The core tools of survival analysis include the Kaplan Meier estimator, the log rank test, and the Cox proportional hazards model. The Kaplan Meier method produces a step function estimating the survival probability over time, with steps occurring at each observed event time. The log rank test compares survival curves between groups, testing the null hypothesis that the survival distributions are identical. The Cox model is a semiparametric regression approach that estimates hazard ratios for one or more predictors without specifying the baseline hazard function, though it assumes that the ratio of hazards between groups remains constant over time (the proportional hazards assumption). When this assumption is violated, researchers may use stratified Cox models, time varying coefficients (interactions between a covariate and time), or models that do not assume proportional hazards, such as accelerated failure time models, in which covariates accelerate or decelerate the time to event rather than multiplying the hazard. Parametric models such as the exponential and Weibull models specify the shape of the hazard over time; they also assume proportional hazards, so on their own they do not resolve a violation.

Survival analysis should be used whenever the research question concerns the time until an event occurs, especially when some observations are censored. This includes randomized trials with mortality or relapse endpoints, cohort studies following participants for incident disease, and reliability studies in engineering. Use the Kaplan Meier method for descriptive visualization and univariable comparison, the log rank test for formal group comparisons, and the Cox model for multivariable adjustment. Parametric survival models are appropriate when the underlying survival distribution can be reasonably assumed or when extrapolation beyond the observed data is required, as in health economic modeling.

A frequent error is assuming that censoring is noninformative when it is informative, meaning participants who are censored have different event risks than those who remain under observation. Standard survival methods, including the Kaplan Meier estimator, the log rank test and the Cox model, assume that censoring is unrelated to the risk of the event, and violation of this assumption can bias the results. Another mistake is ignoring the proportional hazards assumption in Cox regression. If the treatment effect changes over time, the reported hazard ratio is an average that may misrepresent the treatment effect at clinically relevant time points. Researchers also sometimes confuse median survival time with mean survival time or misinterpret Kaplan Meier curves by reading probabilities at specific time points without accounting for the confidence intervals, which widen as the number at risk decreases.

A phase III oncology trial randomizes 600 patients with advanced lung cancer to a new immunotherapy (n = 300) or standard chemotherapy (n = 300) and follows them for overall survival over three years. At the analysis cutoff, 180 deaths have occurred in the chemotherapy arm and 140 in the immunotherapy arm, with the other 280 patients (120 in the chemotherapy arm and 160 in the immunotherapy arm) censored due to ongoing follow up. The Kaplan Meier curves show a median overall survival of 14.2 months for chemotherapy and 18.6 months for immunotherapy. The log rank test yields p = 0.003. A Cox proportional hazards model adjusted for age, performance status, and tumor stage estimates a hazard ratio of 0.72 (95% CI: 0.58 to 0.89), indicating a 28% reduction in the hazard of death with immunotherapy. A Schoenfeld residual test shows no evidence that the proportional hazards assumption is violated, although a nonsignificant test cannot prove that it holds.