Loading field map
Survival analysis is the branch of statistics concerned with the time until a well-defined event occurs. The event might be death, equipment failure, disease recurrence, recovery, or any transition of interest. The defining feature of the data is that for some subjects, the event has not yet happened by the end of the observation period; these observations are called censored. The central challenge of the field is to make valid inferences about the timing of events despite this incomplete information.
To understand why survival analysis exists as a distinct discipline, consider a simple question: what is the average time to recovery for patients on a new treatment? If the study ends after one year, some patients will have recovered, but others will still be recovering. Simply averaging the recovery times of only those who recovered ignores the information that the others survived at least a year without recovering. Worse, it systematically underestimates the true average recovery time. The patients who have not yet recovered are not missing at random; they are, by definition, the ones with longer recovery times.
Censoring is not a nuisance to be discarded but information to be used. A patient who has not recovered after one year tells us that their recovery time is greater than one year. Survival analysis provides the mathematical framework to incorporate such partial information. The most common form is right-censoring, where observation ends before the event occurs. Left-censoring occurs when the event has already happened before observation begins, and interval-censoring occurs when the event is known to have happened within a window of time but the exact time is unknown. Each form requires different handling, but all share the same principle: the likelihood of the data must account for the fact that censored observations contribute only the knowledge that the event time lies in a certain range.
Two related functions form the backbone of the field. The survival function, denoted S(t), gives the probability that the event has not yet occurred by time t. It starts at 1 at time zero and decreases toward 0 as t increases. The hazard function, denoted h(t), describes the instantaneous rate of the event occurring at time t, given that it has not occurred before t. It is a rate, not a probability, and can take any non-negative value. A constant hazard means the event is equally likely to occur at any moment, regardless of how much time has passed—a property of the exponential distribution. An increasing hazard describes aging or wear-out processes; a decreasing hazard describes early-failure or healing processes.
The two functions are mathematically equivalent: given either one, the other can be derived. The survival function is more intuitive for describing the overall experience of a population, while the hazard function is more useful for understanding the mechanism of risk at each moment. Much of the field's development has been about estimating these functions from censored data and comparing them across groups.
The simplest and most widely used tool in survival analysis is the Kaplan–Meier estimator, developed in the 1950s. It estimates the survival function from data with right-censoring without assuming any particular mathematical form for the distribution of event times. The idea is to divide time into intervals between observed events. At each event time, the estimated probability of surviving past that moment is the proportion of those still at risk who did not experience the event. The overall survival estimate is the product of these conditional probabilities across all event times.
This estimator is nonparametric because it makes no assumptions about the shape of the survival curve. It is the survival-analysis analogue of the empirical cumulative distribution function, adapted to handle censoring. Its limitations are that it only describes a single group and cannot easily incorporate continuous predictors. For comparing two groups, the log-rank test provides a nonparametric hypothesis test of whether their survival curves differ, though it does not estimate the magnitude of the difference.
The most influential development in the field is the Cox proportional hazards model, introduced by David Cox in 1972. It addresses the question that dominates applied survival analysis: how do multiple covariates—age, treatment, genetic markers, environmental exposures—affect the risk of the event? The model assumes that the hazard for an individual is a baseline hazard multiplied by a factor that depends on the covariates. Specifically, the hazard at time t for an individual with covariates x is h(t) = h₀(t) × exp(βx), where h₀(t) is an unspecified baseline hazard common to all individuals, and β is a vector of coefficients to be estimated.
The key innovation is that the baseline hazard is left completely unspecified. This makes the model semiparametric: the effect of covariates is modeled parametrically through the exponential term, but the shape of the baseline hazard is free. Cox showed that the coefficients β can be estimated without estimating the baseline hazard at all, using a partial likelihood that compares the covariate values of the individual who experienced the event at each event time with those still at risk. This approach is remarkably efficient and robust, and it has become the default method for analyzing survival data in medicine, epidemiology, and engineering.
The proportional hazards assumption—that the ratio of hazards for two individuals is constant over time—is the model's central assumption. When it fails, the model can misrepresent the data. Extensions allow for time-varying covariates, stratification by groups with different baseline hazards, and tests for non-proportionality. Despite these complications, the Cox model's combination of flexibility and interpretability has made it the workhorse of the field.
Before the Cox model, and continuing alongside it, is the parametric approach. Here, the event time distribution is assumed to follow a specific mathematical form, such as exponential, Weibull, log-normal, or gamma. Each distribution imposes a particular shape on the hazard function: the exponential has constant hazard, the Weibull allows monotonic increasing or decreasing hazard, and the log-normal allows a hazard that rises then falls.
Parametric models have several advantages. They are more efficient than semiparametric models when the distributional assumption is correct, meaning they require less data to achieve the same precision. They allow direct estimation of quantities like the median survival time or the probability of surviving past a given time, which are awkward to obtain from the Cox model. They also permit prediction of survival beyond the observed follow-up period, though such extrapolation is only as reliable as the distributional assumption. Their disadvantage is the risk of misspecification: if the true distribution differs from the assumed one, estimates can be badly biased. In practice, parametric models are often used when the mechanism of failure is well understood, such as in reliability engineering, or when the data are sparse and the Cox model's flexibility is a liability.
A substantial extension of the basic framework arises when there are multiple possible events, and the occurrence of one prevents the others. For example, a patient might die from the disease of interest or from an unrelated cause. These are competing risks. The naive approach of treating deaths from other causes as censored observations is incorrect, because it implicitly assumes that the patient would eventually experience the event of interest, which is not true. The correct analysis requires estimating the cause-specific hazard for each event type, and the cumulative incidence function, which gives the probability of each event occurring by time t.
Multi-state models generalize this further by allowing individuals to move through several states over time—for example, from healthy to diseased to dead, with possible recovery and relapse. These models describe the entire trajectory of a process rather than a single event. They are built from transition hazards between states and are particularly important in chronic disease epidemiology and health economics, where the cost and quality of life depend on the sequence of states a patient occupies.
Another important extension accounts for unobserved heterogeneity among individuals. Even after adjusting for measured covariates, some individuals may be inherently more or less prone to the event. This unobserved variation is modeled as a frailty—a random multiplicative factor on the hazard. Frailty models are the survival-analysis analogue of random-effects models in other branches of statistics. They are essential when analyzing clustered data, such as patients within hospitals or components within systems, where individuals in the same cluster share unmeasured influences. Ignoring frailty when it exists leads to underestimation of the hazard's decline over time, because the individuals most prone to the event tend to experience it earlier, leaving a population that is increasingly composed of robust survivors.
Contemporary survival analysis is characterized by several converging developments. Machine learning methods, including random survival forests and deep learning approaches, have been adapted to handle censored data and high-dimensional predictors, though they often sacrifice the interpretability that makes the Cox model so useful in clinical and regulatory settings. The field has also embraced more flexible modeling of the baseline hazard and time-varying effects, blurring the boundary between parametric and semiparametric approaches.
A major practical concern is the analysis of large-scale administrative and electronic health record data, where follow-up is irregular, events are recorded with error, and the proportional hazards assumption is frequently violated. Methods for handling interval censoring, measurement error in event times, and dependent censoring—where the censoring time is correlated with the event time—remain active research areas. The field has also developed sophisticated tools for sample size calculation and interim analysis in clinical trials, where survival endpoints are common.
Throughout its history, survival analysis has been driven by applied problems, particularly in medicine and engineering, and its methods have in turn shaped how those fields design studies and interpret results. The core intellectual contribution remains the same: a principled way to learn from incomplete data about the timing of events, and a set of tools that respect the fundamental uncertainty introduced by censoring.