1 Fundamentals
Survival analysis studies the time until a specified event occurs. The event may be death in a medical study, failure of a mechanical component, job exit in labor research, or another outcome of interest. A defining feature of the field is its treatment of incomplete follow-up, since many subjects do not experience the event before observation ends.
1.1 Time-to-event data
Time-to-event data record both the duration of observation and the occurrence status of the endpoint. Unlike ordinary continuous outcomes, these data are indexed by time and often include individuals whose exact event time is not observed. The starting point may be birth, enrollment, device installation, diagnosis, or another origin defined by the study.
1.2 Event types
The endpoint in survival analysis can vary widely depending on the research question. Some studies focus on a single first occurrence, while others allow repeated events over time. The choice of event definition shapes the statistical model and interpretation.
1.2.1 Single-event outcomes
Single-event outcomes involve one terminal event per subject, such as death, disease onset, or equipment failure. Once the event happens, the subject leaves the risk set for that endpoint. This setting is the most common and is often used for estimating survival probabilities and comparing groups.
1.2.2 Recurrent events
Recurrent events are outcomes that may happen multiple times for the same individual, such as hospital readmissions, seizures, or system outages. These data require methods that account for dependence among repeated episodes within a subject. Analysts may study event counts, event gaps, or the rate of recurrence over time.
1.3 Censoring and truncation
Incomplete observation is central to survival analysis. Censoring means that the exact event time is only partially known, while truncation refers to observations being absent from the dataset because of the way they enter or are selected from the study. Both issues must be handled carefully to avoid bias.
1.3.1 Right censoring
Right censoring occurs when the event has not happened by the end of follow-up, or when observation stops before the event is seen. The only known information is that the event time exceeds the last observed time. This is the most frequent form of censoring in practice.
1.3.2 Left censoring
Left censoring arises when the event has already occurred before the time it is first detected, but the exact time is unknown. This situation appears in studies where the event is observed only at periodic checks. The recorded time gives an upper bound rather than an exact value.
1.3.3 Interval censoring
Interval censoring means the event is known to have happened between two observation times, but not exactly when. It is common in screening studies and other designs with scheduled assessments. Specialized methods are used because standard tools for exact event times do not directly apply.
1.3.4 Left truncation and right truncation
Left truncation occurs when subjects must survive or remain event-free until a certain entry time to be included in the study. Right truncation happens when only subjects whose event times fall before a cutoff are observed. Both forms affect the sampled population and require adjustment in analysis.
1.4 Survival and hazard concepts
Survival analysis uses functions that describe how event occurrence changes over time. These functions provide complementary views: one emphasizes the chance of remaining event-free, another the instantaneous event rate, and another the accumulated risk over time.
1.4.1 Survival function
The survival function gives the probability that an individual has not yet experienced the event by a given time. It is a nonincreasing function ranging from 1 at baseline toward 0 as time passes. In many applications, it offers the most intuitive summary of persistence or durability.
1.4.2 Hazard function
The hazard function describes the instantaneous rate at which events occur among those still under observation. It is not a probability in the usual sense, but a rate conditional on survival up to that moment. Hazards are especially useful for modeling covariate effects and changes in risk over time.
1.4.3 Cumulative hazard function
The cumulative hazard function aggregates hazard over time. It helps connect the survival and hazard functions mathematically and is often used in estimation and model checking. Large cumulative hazard values correspond to lower survival probabilities.
1.5 Life tables and event-time summaries
Life tables summarize survival experience within time intervals rather than at exact event times. They are useful when data are grouped or observation is coarse. Such summaries may include numbers at risk, numbers of events, and estimated survival proportions across successive intervals.
2 Descriptive methods
Descriptive methods summarize the observed timing of events without imposing a strong parametric form. They are often the first stage of analysis, providing curves and tables that display how survival evolves and where groups appear to diverge.
2.1 Kaplan–Meier estimation
The Kaplan–Meier estimator is the standard nonparametric method for estimating the survival function from censored data. It produces a stepwise curve that changes only at observed event times. Because it uses all available follow-up information, it is widely applied in clinical and reliability studies.
2.1.1 Construction of the estimator
The estimator is built from successive conditional survival probabilities at each event time. At each failure time, the proportion surviving that interval is multiplied into the running estimate. Censored observations reduce the number at risk later on but do not create a drop in the curve at the censoring time.
2.1.2 Confidence intervals
Confidence intervals around Kaplan–Meier estimates quantify uncertainty due to sampling variation. They are typically wider when few subjects remain under observation or when event times are sparse. Several transformations are used to keep interval estimates within valid probability bounds.
2.2 Nelson–Aalen estimation
The Nelson–Aalen estimator provides a nonparametric estimate of the cumulative hazard function. It is constructed by summing the observed event counts divided by the number at risk at each event time. This estimator is especially useful when analysts want a hazard-based summary rather than a survival curve.
2.3 Median survival and quantiles
Median survival is the time by which the estimated survival probability falls to one-half. Other quantiles, such as the 25th or 75th percentile, give additional landmarks for comparing distributions. These summaries are easy to interpret but may be unavailable if the curve does not drop far enough during follow-up.
2.4 Plotting survival curves
Survival curves are commonly displayed as step functions, often with marks showing censored observations. Multiple curves may be plotted together to compare groups, treatment arms, or device types. Good graphical practice includes clear labels for time scale, event definition, and number at risk.
3 Inferential methods
Inferential methods aim to compare groups or estimate the effect of covariates on event timing. These approaches move beyond description to formal statistical testing or model-based prediction.
3.1 Group comparison
Group comparison asks whether two or more survival experiences differ more than would be expected by chance. Such comparisons are common in trials, observational studies, and reliability testing. The choice of test depends on whether differences are expected early, late, or throughout follow-up.
3.1.1 Log-rank test
The log-rank test compares survival curves across groups under the assumption that differences are proportional over time. It gives equal weight to event times across the follow-up period. This makes it a standard choice when the main interest is an overall difference in hazards.
3.1.2 Wilcoxon-type tests
Wilcoxon-type tests place greater weight on early events, which can make them more sensitive when curves separate near the beginning of follow-up. They are sometimes preferred when early failure is especially important. Their emphasis differs from the log-rank test, so the two tests may lead to different conclusions.
3.2 Parametric survival models
Parametric models assume a specific distribution for event times. By imposing structure, they can provide smooth estimates, extrapolate beyond observed follow-up, and yield direct summaries such as expected survival time. Their accuracy depends on whether the chosen distribution fits the data well.
3.2.1 Exponential model
The exponential model assumes a constant hazard over time. It is mathematically simple and often used as a baseline reference. Because the hazard does not change, the model is suitable only when event risk is approximately stable across the study period.
3.2.2 Weibull model
The Weibull model allows the hazard to increase or decrease with time, making it more flexible than the exponential form. It is widely used because of its tractable properties and broad applicability. Depending on parameter values, it can represent early failures, aging processes, or decreasing risk.
3.2.3 Log-normal and log-logistic models
Log-normal and log-logistic models describe event times whose logarithms follow familiar continuous distributions. They can produce nonmonotonic hazard shapes, including hazards that rise and later fall. These models are useful when risk does not follow a simple increasing or decreasing pattern.
3.3 Semi-parametric models
Semi-parametric models combine an unspecified baseline hazard with covariate effects expressed through regression. This approach offers flexibility while still allowing inference about predictors. It is among the most influential frameworks in survival analysis.
3.3.1 Cox proportional hazards model
The Cox proportional hazards model relates covariates to hazard through multiplicative effects. It does not require a parametric form for the baseline hazard, which makes it broadly adaptable. The model is especially valued for estimating relative risk and adjusting for multiple predictors simultaneously.
3.3.2 Partial likelihood
Partial likelihood is the estimation method that underlies the Cox model. It uses the ordering of event times and the composition of the risk set rather than specifying the full survival distribution. This feature allows regression coefficients to be estimated without modeling the baseline hazard directly.
3.3.3 Stratified Cox models
Stratified Cox models permit separate baseline hazards across strata while estimating common covariate effects. They are useful when groups differ in baseline risk but the proportional effect of a predictor can still be assumed within strata. Stratification can reduce bias when one variable strongly affects the baseline hazard.
3.4 Accelerated failure time models
Accelerated failure time models describe covariate effects as speeding up or slowing down the event process. Rather than acting on the hazard directly, predictors stretch or compress the time scale. These models can be more natural when the main question concerns how long an event is delayed or hastened.
4 Model assumptions and diagnostics
Statistical models for survival data rely on assumptions that should be examined rather than accepted automatically. Diagnostics help determine whether the chosen model fits the observed pattern of event times and whether estimates are stable.
4.1 Proportional hazards assumption
The proportional hazards assumption states that hazard ratios between groups remain constant over time. This condition is central to the Cox model and to many hazard-based comparisons. When it fails, effect estimates may still be useful, but their interpretation becomes time dependent.
4.2 Goodness-of-fit assessment
Goodness-of-fit assessment evaluates whether a model reproduces the structure seen in the data. Analysts may compare fitted and observed survival curves, inspect residual patterns, or check whether predicted risks align with observed events. Adequate fit supports credible inference and prediction.
4.3 Residuals and influence diagnostics
Residuals in survival analysis measure departures between observed outcomes and model expectations. Influence diagnostics identify subjects or clusters that have a disproportionate effect on fitted results. Together, these tools help detect outliers, misspecification, and unstable estimates.
4.3.1 Martingale residuals
Martingale residuals are often used to examine functional form and the adequacy of covariate effects. They can reveal whether a continuous predictor is better represented through a transformation or nonlinear term. Because their distribution is asymmetric, interpretation requires care.
4.3.2 Schoenfeld residuals
Schoenfeld residuals are closely associated with checking proportional hazards. They are examined over time to see whether covariate effects remain stable or drift during follow-up. Systematic trends suggest that the effect may vary with time.
4.3.3 Deviance residuals
Deviance residuals are transformed residuals that are easier to interpret than martingale residuals in some settings. They help identify observations with unexpectedly early or late events. Large values may indicate unusual cases or model mismatch.
4.4 Handling non-proportional hazards
When proportional hazards do not hold, analysts may use time-varying effects, stratification, alternative models, or piecewise analyses. The most suitable method depends on the scientific question and the nature of the departure from proportionality. In some studies, reporting time-specific effects is more informative than forcing a single summary estimate.
5 Extensions
Extensions broaden survival analysis to more complex event structures and dependence patterns. These methods address situations where simple first-event models are not enough.
5.1 Competing risks
Competing risks occur when multiple different event types can prevent the event of primary interest from occurring. For example, one cause of failure may preclude another from ever being observed. Standard survival methods can overstate the incidence of the target event if competing events are ignored.
5.1.1 Cause-specific hazards
Cause-specific hazards model the instantaneous rate of each event type separately among those still event-free. They are useful for studying etiologic associations and understanding how covariates influence particular causes. Interpretation must remain tied to the chosen event category.
5.1.2 Cumulative incidence functions
Cumulative incidence functions estimate the actual probability of each event type over time in the presence of competing risks. Unlike ordinary survival curves, they account for the fact that one event can eliminate the chance of another. They are often preferred when prediction of real-world incidence is the goal.
5.2 Multi-state models
Multi-state models describe movement through a sequence of states over time, such as healthy, ill, recovered, and dead. They generalize survival analysis by allowing transitions among multiple intermediate conditions. These models are useful when the path of events matters as much as the final endpoint.
5.3 Recurrent event analysis
Recurrent event analysis studies repeated occurrences of the same event within individuals. Methods may focus on the number of events, the timing between events, or the intensity of recurrence. Such analyses are common in medical follow-up and reliability monitoring.
5.4 Frailty models
Frailty models introduce random effects to capture unobserved heterogeneity among subjects or clusters. The term frailty refers to latent susceptibility that makes some individuals experience events earlier than others. These models are valuable when repeated observations or clustered data show extra dependence beyond measured covariates.
5.5 Time-dependent covariates
Time-dependent covariates are predictors whose values change during follow-up, such as treatment status, biomarker levels, or age. Incorporating them allows the model to reflect evolving risk factors more realistically. Correct handling is important to avoid biased estimates and incorrect timing.
6 Applications
Survival analysis is applied wherever timing matters and incomplete follow-up is common. Its methods help quantify duration, compare risk, and forecast future events in many fields.
6.1 Medicine and clinical trials
In medicine, survival analysis is used to study patient survival, disease progression, treatment response, and time to relapse. Clinical trials often rely on survival curves and hazard models to compare therapies. These methods are particularly useful when follow-up times differ among participants.
6.2 Reliability engineering
Reliability engineering uses survival analysis to study the lifespan of machines, components, and systems. Analysts may estimate failure distributions, plan maintenance, and compare designs. The same methods help assess warranty periods and operational durability.
6.3 Public health and epidemiology
Public health and epidemiology use survival methods to examine onset of disease, duration of risk, and time to adverse outcomes. They support studies of screening, prevention, and population-level incidence. The ability to handle censoring makes them well suited to long-term observational data.
6.4 Economics and labor studies
In economics, survival analysis can model unemployment duration, time to job change, firm exit, or loan default. It is useful for examining how policies, market conditions, or personal characteristics affect timing. The approach is often called duration analysis in this context.
6.5 Social science research
Social science research applies survival methods to events such as marriage, divorce, school completion, and migration. These models help describe how individual histories and contextual factors influence timing. They are also used when event occurrence is observed only over survey intervals.
7 Practical considerations
Practical work in survival analysis requires careful study design, transparent handling of incomplete observations, and appropriate software. Good practice also includes checking assumptions and reporting results in a way that is both accurate and interpretable.
7.1 Study design and follow-up
Design choices determine what counts as time zero, how events are defined, and how long subjects are observed. Clear rules for entry, follow-up, and endpoint assessment reduce ambiguity. Adequate follow-up is important for estimating later event times and for comparing groups fairly.
7.2 Missing data and informative censoring
Missing data can complicate survival analysis when covariates are incomplete or when censoring is related to the underlying event process. Informative censoring occurs if the reason for leaving the study is associated with the chance of event occurrence. Analysts may need specialized methods or sensitivity analyses to address this problem.
7.3 Sample size and power
Sample size and power calculations for time-to-event studies depend on expected event rates, follow-up length, censoring, and effect size. Because not all subjects will experience the event, the number of observed events often matters more than the total sample alone. Proper planning helps ensure that comparisons have enough precision.
7.4 Software implementations
Survival analysis is supported by most major statistical platforms. Software differs in syntax and advanced capabilities, but core functions for estimation, regression, diagnostics, and plotting are widely available. Choice of tool often depends on the data structure and the analyst’s workflow.
7.4.1 R packages
R offers extensive support through packages for Kaplan–Meier estimation, Cox regression, competing risks, and multi-state models. Its open ecosystem makes it especially popular in research and teaching. Many packages also provide visualization and diagnostic tools.
7.4.2 Python tools
Python provides survival-related functionality through statistical and machine-learning libraries. These tools are commonly used for integration with data pipelines and predictive workflows. Their ecosystem is smaller than that of specialized statistical software, but it continues to expand.
7.4.3 SAS and Stata support
SAS and Stata include long-established procedures for survival estimation, regression, and model checking. They are widely used in clinical research and applied social science. Both platforms offer standardized output that is convenient for reporting and reproducible analysis.