Survival Analysis

Survival analysis is the branch of statistics that studies the time until something happens—death, machine failure, relapse after treatment, unemployment, the end of a marriage.

10 frameworks, from 1662 to today · Statistics · Mathematics

still a working view ended · each leads to its entry below

Overview

Survival analysis is the branch of statistics that studies the time until something happens—death, machine failure, relapse after treatment, unemployment, the end of a marriage1. What makes it a distinct field rather than a corner of ordinary regression is a stubborn practical problem: the event of interest is often not observed for everyone in the data. A patient moves away, a study ends before the last subject fails, a component is still running when testing stops. These observations are censored: we know the event had not yet occurred by some time, but not when it will1. Survival analysis is, at heart, the body of methods built to extract honest answers from data like this. Its history is a story of borrowing tools from actuarial science, probability theory, and engineering, and gradually forging them into a unified statistical discipline.

Life Tables and the Actuarial Tradition

The oldest strand is the life table, a tabulation of how many people of each age die within a year, and hence of life expectancy. John Graunt's Natural and Political Observations (1662) on London's bills of mortality is an early attempt to reason about death rates from incomplete records. Edmond Halley—better known for his comet—constructed a proper life table in 1693 from Breslau parish data, and the mathematician Abraham de Moivre later gave the mathematics a cleaner form. By the nineteenth century, life insurance had turned this into a commercial necessity: actuaries needed to price policies and hold adequate reserves, and the life table became the working instrument of the profession.

The actuarial tradition contributed two lasting ideas. The first is the hazard rate (also called the force of mortality): the instantaneous risk of dying at a given age, given survival to that age1. The second is the practice of estimating survival from grouped, censored data—people who lapse policies or move away are "withdrawn alive," contributing information up to the point they leave. This is censoring handled by bookkeeping, long before anyone formalized it probabilistically.

Parametric Distributions and the Engineering Strand

The early twentieth century brought a probabilistic turn. Actuaries and biologists began fitting smooth curves to mortality. The Gompertz law (from Benjamin Gompertz, 1825) had already described mortality rising exponentially with age, and the Makeham modification added a constant background risk. In reliability engineering, a parallel need arose: predicting the failure of manufactured components. Here the exponential distribution—with its memoryless property, meaning a used component is no more likely to fail than a new one—became the baseline model, and the Weibull distribution, introduced by Waloddi Weibull in 1939 and popularized in the 1950s, offered a flexible shape for wear-out and infant-mortality patterns.

The unifying insight was that a survival distribution and its hazard function are two views of the same object. Once that link was clear, the field had a common language: model the hazard, and you model the time. But parametric models demand that you commit to a specific shape, and real data often refuse.

Nonparametric Estimation: Kaplan and Meier

The decisive break came in 1958, when Edward Kaplan and Paul Meier published the product-limit estimator—now universally called the Kaplan–Meier estimator. The problem it solved was estimating a survival curve from censored data without assuming any distributional form. The idea is disarmingly simple: at each observed event time, compute the fraction of those still at risk who survive past that moment, then multiply these fractions together across event times2. Censored observations drop out of the "at risk" pool at the moment they are censored, contributing what they can and no more.

The Kaplan–Meier curve became the standard descriptive tool of medical research, and its companion, the log-rank test (developed in the 1960s and 1970s), gave a way to compare survival between groups. Around the same time, the Nelson–Aalen estimator provided a nonparametric estimate of the cumulative hazard directly. These methods made survival analysis usable by clinicians and engineers who had no taste for distributional assumptions.

Proportional Hazards: Cox's Regression Model

The next turning point addressed a gap: Kaplan–Meier compares groups, but real studies have many covariates—age, stage of disease, treatment dose—that must be adjusted for simultaneously. In 1972, David Cox introduced the proportional hazards model. Its hazard for an individual is a baseline hazard (left completely unspecified) multiplied by an exponential function of the covariates. The covariates shift the hazard up or down by a constant factor—the hazard ratio—at every point in time3.

The model's power lies in a trick called partial likelihood, which lets you estimate the covariate effects without ever specifying the baseline hazard. This made Cox regression the workhorse of medical statistics, and it remains so. The proportional hazards assumption—that hazard ratios stay constant over time—is a real constraint, and checking it became a standard part of applied practice; when it fails, alternatives such as time-varying effects or accelerated failure time models come into play.

Counting Processes and Martingales

Through the 1970s and 1980s, probability theorists rebuilt the field on a rigorous foundation. The key move was to represent each subject's event history as a counting process—a running tally that jumps from 0 to 1 when the event occurs—and to treat the hazard as an intensity governing that process. Richard Gill, Per Kragh Andersen, Niels Keiding, Odd Aalen, and others developed this framework, drawing on the martingale theory of stochastic processes.

The payoff was more than elegance. Martingale methods yielded large-sample theory—consistency and asymptotic normality—for the Kaplan–Meier estimator, the log-rank test, and Cox's partial likelihood, all within one coherent scheme. They also generalized survival analysis to settings the earlier tools could not touch: recurrent events, time-varying covariates, and multiple event types. This is the theoretical backbone on which most modern textbooks rest.

Competing Risks and Frailty

Two complications forced the field to grow. Competing risks arise when a subject can experience one of several mutually exclusive events—death from heart disease versus death from cancer—where one event precludes the others. The naive approach of treating other causes as censoring answers the wrong question4. The field distinguishes the cause-specific hazard (the rate of a given cause among those still at risk) from the cumulative incidence function (the actual probability of experiencing that cause by a given time), and the latter requires accounting for the competing events. These ideas trace to actuarial work and were formalized statistically from the 1990s onward, notably by Jason Fine and Robert Gray's 1999 model for the subdistribution hazard.

Frailty models address a different problem: unobserved heterogeneity. Some individuals are simply more fragile than others, and if you ignore this, you misestimate how hazard changes with age—the frail die early, leaving a hardier survivors, which can make mortality appear to level off. Frailty models, developed from the 1970s and 1980s (with contributions from Vaupel, Manton, and Stallard, and later Andersen and others), introduce a random effect—a latent multiplier on the hazard—shared within a group or individual. They connect survival analysis to random-effects and multilevel modeling, and in multivariate settings they become shared frailty models for clustered data such as family or hospital groups.

Bayesian Survival Analysis

Bayesian methods entered the field by treating the baseline hazard and the regression coefficients as random quantities with prior distributions. Early work in the 1970s and 1980s used parametric priors, but the real expansion came with computational advances. The 1990s saw Bayesian nonparametric approaches—notably the gamma process prior for the cumulative hazard, developed by Kalbfleisch and others, and later the Dirichlet process and its extensions—that let the baseline hazard be flexible while still yielding full posterior inference. Markov chain Monte Carlo made these models practical5. Bayesian survival analysis is especially valued for small samples, for incorporating prior clinical knowledge, and for hierarchical structures such as multi-center trials, where it dovetails naturally with frailty ideas.

Causal Inference and Machine Learning

The most recent era has pushed survival analysis toward questions of cause and prediction. Causal survival analysis asks what would happen under a treatment or intervention, not merely what is associated with survival6. This requires adjusting for confounders that may themselves change over time, and it draws on the broader causal-inference toolkit—inverse probability weighting, g-computation, and targeted learning—adapted to censored outcomes. James Robins' work on g-methods from the 1980s and 1990s laid much of the groundwork; the survival-specific formalization grew substantially from the 2000s onward, with contributions from van der Laan, Hernán, and others.

Machine learning for time-to-event data is the newest strand. Classical survival models impose strong structure—linear covariate effects, proportional hazards. Machine learning methods relax these: survival trees and random survival forests (Ishwaran and colleagues, 2008) partition patients by risk without a linearity assumption; gradient boosting and neural-network approaches extend this further. A persistent challenge is that standard machine-learning tools assume fully observed outcomes, so censoring must be handled explicitly—through adapted loss functions, inverse-probability-of-censoring weighting, or specialized architectures. The deep-learning era has produced models that predict individualized survival curves from high-dimensional inputs such as images and genomic data, though how to validate and trust them remains an active question.

The Present

Today survival analysis is a mature but still-moving field. The Kaplan–Meier curve and Cox model remain the everyday instruments of clinical trials and reliability engineering, now joined by competing-risks methods, frailty and multilevel models, and Bayesian and machine-learning approaches. The counting-process framework provides the theoretical unity underneath. Open questions cluster around causal claims from observational data with time-varying confounding, the calibration and interpretability of flexible predictive models, and how to handle the complex, high-dimensional, and often messy data that modern studies generate. The field's oldest problem—what to do when you don't see the event—has never gone away; it has simply acquired more sophisticated answers.

References

  1. Survival analysis (introduction, Definitions of common terms in survival analysis, Hazard function and cumulative hazard function), Wikipedia
  2. Kaplan–Meier estimator, Wikipedia
  3. Proportional hazards model (Background), Wikipedia
  4. Relative survival (Competing risks survival analysis), Wikipedia
  5. Bayesian statistics, Wikipedia
  6. Causal inference, Wikipedia

The frameworks of Survival Analysis

Each framework is a way of working the field has used: when it held, and the question it answers.

  1. 1662-Present

    Life table

    How many people of each age die within a year?

    Fly to its tower in Nousfarer

  2. 1825-Present

    Parametric Survival Distributions

    What smooth curve describes time until failure?

    Fly to its tower in Nousfarer

  3. 1958-Present

    Nonparametric Survival Estimation

    How can survival be estimated without assuming a distribution?

    Fly to its tower in Nousfarer

  4. 1972-Present

    Proportional hazards model

    How do covariates shift the hazard of an event?

    Fly to its tower in Nousfarer

  5. 1975-Present

    Counting Process and Martingale Theory

    What rigorous theory underlies survival estimators?

    Fly to its tower in Nousfarer

  6. 1979-Present

    Frailty and Multilevel Survival Models

    How does unobserved individual fragility affect survival?

    Fly to its tower in Nousfarer

  7. 1980-Present

    Bayesian Survival Analysis

    How should prior knowledge update survival estimates?

    Fly to its tower in Nousfarer

  8. 1990-Present

    Causal Survival Analysis

    What is the causal effect of a treatment on survival?

    Fly to its tower in Nousfarer

  9. 1990-Present

    Competing Risks Analysis

    How likely is each of several mutually exclusive events?

    Fly to its tower in Nousfarer

  10. 2008-Present

How the ideas connect

Each arrow runs from an earlier idea to the later one it shaped. The list below says what passed between them.

Life tableParametric Survival DistributionsNonparametric Survival EstimationProportional hazards modelCounting Process and Martingale TheoryFrailty and Multilevel Survival ModelsBayesian Survival AnalysisCausal Survival AnalysisCompeting Risks AnalysisMachine Learning for Time-to-Event Da…

Beyond the field

Where Survival Analysis's ideas came from in other fields, and what they led to there: the strongest 12 of 50 links.

A city in Nousfarer: a red-carpeted stair climbing terraces to a crowned tower on an acropolis, more towers round it, forest and snowy mountains beyond.
A city in Nousfarer, the game played in the browser.

Survival Analysis is a planet in Nousfarer

Its city has a tower for each of these 10 frameworks. Land, walk in, and climb one concept at a time.

Land on Survival Analysis

Related fields

Fields its ideas are linked to

More in Statistics

See where Survival Analysis sits on the Map of Human Knowledge