1 Concept and definitions

Time-to-event is the study of durations until a defined event occurs. In applications ranging from clinical follow-up to reliability assessment and service analytics, it provides a structured way to quantify “waiting time” and to compare how quickly different populations experience outcomes.

1.1 Event of interest

The event of interest is the specific occurrence whose timing is recorded. It can be the first time an outcome appears (for example, first failure), the occurrence of a specific transition (such as recovery), or the moment a predefined criterion is met. Careful event definition is essential because the statistical conclusions depend directly on what is counted as an event.

1.2 Time origin (start point)

The time origin is the reference point from which waiting time is measured. It might be study enrollment, device activation, contract start, or the moment a monitoring process begins. Choosing the origin affects the distribution of observed times and can introduce systematic differences if groups have different effective start points.

1.3 Censoring and incomplete observation

Censoring occurs when the event is not observed within the observation window or when follow-up ends for reasons unrelated to the event. For instance, a person may leave a study before experiencing the outcome, or monitoring may stop while the individual remains event-free. Many time-to-event methods incorporate censoring so that partial information still contributes to estimation rather than discarding incomplete records.

1.4 Types of time-to-event data (exact vs interval)

Time-to-event observations can be recorded at different levels of precision. “Exact” event times occur when the timing is known precisely. “Interval” or grouped data arise when the event is known to have happened between two assessments, leading to interval-censored likelihoods. The data structure determines which estimators and software options are appropriate.

2 Core measures

Time-to-event analysis uses several related quantities that describe survival, failure behavior, and event dynamics. These measures are often interconnected: for example, the cumulative distribution of event times determines survival, and hazard relates to instantaneous failure propensity.

2.1 Survival function

The survival function, commonly denoted \(S(t)\), is the probability that the event has not occurred by time \(t\). It summarizes how long individuals or units tend to remain event-free up to different horizons. Plotting \(S(t)\) over time provides an intuitive view of “remaining” risk.

2.2 Cumulative distribution function

The cumulative distribution function, \(F(t)\), gives the probability that the event has occurred by time \(t\). Since \(F(t)=1-S(t)\), either representation communicates the same information, though they may be presented differently depending on whether the focus is on survival or on failure accumulation.

2.3 Hazard rate

The hazard rate \(h(t)\) describes the event risk at time \(t\), given that the unit has survived up to that time. It is often interpreted as an instantaneous rate, capturing how the propensity to experience the event changes over time. In many settings, hazard is the target for modeling with covariates.

2.4 Median time-to-event and quantiles

A median time-to-event is the time by which 50% of the population has experienced the event, when defined. More generally, quantiles identify time thresholds such as the 25th or 75th percentile. Quantile-based summaries can be preferable when survival curves cross or when stakeholders want specific waiting-time benchmarks.

2.5 Risk sets and event counting

Risk sets are the collection of units “at risk” just before each event time. In discrete event-time accounting, statistical procedures often rely on counting the number of events and the number at risk within ordered time points. This bookkeeping underlies many nonparametric estimators and many likelihood-based methods.

3 Modeling approaches

Modeling approaches translate the time-to-event framework into estimators that handle censoring and optionally include covariates. Methods range from minimal assumptions to flexible semiparametric and parametric forms.

3.1 Nonparametric methods

Nonparametric techniques estimate survival-related functions without specifying a full parametric distribution for event times. They are commonly used for descriptive comparisons and baseline estimation.

3.1.1 Kaplan–Meier estimator

The Kaplan–Meier estimator provides a nonparametric estimate of the survival function. It uses observed event times and adjusts for censoring through risk-set denominators. The resulting step function allows visualization of survival probabilities over time.

3.1.2 Nelson–Aalen estimator

The Nelson–Aalen estimator estimates the cumulative hazard function. Rather than directly producing survival probabilities, it accumulates hazard contributions over event times. Under standard relationships, it can be used to reconstruct survival estimates or to support hazard-based interpretation.

3.2 Parametric models

Parametric models assume a specific functional form for how event times behave. They can yield smooth hazard and survival curves and often support direct extrapolation beyond observed ranges, though such extrapolation depends on the correctness of model assumptions.

3.2.1 Exponential model

The exponential model assumes a constant hazard rate over time. Its simplicity makes it a useful baseline, but it often fits poorly when risk changes substantially with time. When appropriate, it yields straightforward closed-form survival and hazard relationships.

3.2.2 Weibull model

The Weibull model allows hazard to vary with time, including increasing or decreasing patterns depending on the parameterization. It is widely used because it captures many real-world shapes while remaining mathematically tractable.

3.2.3 Log-normal and log-logistic models

Log-normal and log-logistic models treat the logarithm of survival time as following a normal or logistic-type distribution, respectively. They are often chosen when hazard rises to a peak and later declines, or when survival tails behave in a way that constant or monotone hazards cannot represent.

3.3 Semi-parametric models

Semi-parametric methods offer a balance between flexibility and interpretability, typically modeling part of the structure while leaving other components unspecified.

3.3.1 Cox proportional hazards model

The Cox proportional hazards model relates covariates to the hazard function multiplicatively through a regression term, without requiring specification of the baseline hazard’s shape. The central idea is that hazard ratios between individuals are constant over time under the proportional hazards condition.

3.4 Accelerated failure time perspective

The accelerated failure time (AFT) approach models how covariates change the timing of events rather than the hazard directly. In AFT models, covariates act to accelerate or decelerate the event process, often expressed as multiplicative effects on log survival time. This perspective can be intuitive when stakeholders think in terms of shortening or lengthening expected waiting times.

4 Assumptions and diagnostics

Time-to-event conclusions rely on conditions that, if violated, can bias estimates or undermine comparisons. Diagnostics help assess whether the chosen model is compatible with the observed data patterns.

4.1 Proportional hazards assumption

For Cox-type models, the proportional hazards assumption requires that hazard ratios for covariate effects remain constant over time. Deviations can appear when curves suggest different convergence patterns or when residual-based checks indicate systematic changes in effect strength.

4.2 Checking functional form in parametric models

Parametric models depend on the assumed distribution family. Functional-form checks include comparing predicted survival curves to empirical patterns, evaluating tail behavior, and assessing whether the fitted distribution reproduces the observed hazard shape. Misfit is especially visible when event timing patterns show features inconsistent with the chosen family.

4.3 Goodness-of-fit and residual concepts

Goodness-of-fit tools quantify how well a model reproduces observed event timing under censoring. Residual concepts adapted to survival contexts help detect systematic departures, such as underestimating hazards in certain time windows. While many diagnostics are available, selection should match the model type and the censoring mechanism.

4.4 Handling competing risks (overview-level)

Competing risks arise when multiple event types prevent each other, so that observing one outcome changes the probability of others. In an overview sense, competing-risk methods distinguish event-specific failure probabilities from overall survival and use appropriate formulations to avoid treating competing outcomes as merely censored. This distinction matters when different endpoint types occur with substantial frequency.

5 Estimation and inference

Inference quantifies uncertainty around survival estimates, effect sizes, and comparisons across groups. Because censored data complicate likelihood behavior, standard error estimates and test statistics are designed for survival settings.

5.1 Confidence intervals for survival quantities

Confidence intervals can be constructed for survival probabilities, medians, and other derived quantities. Common strategies include transforming estimates into approximately normal forms or using simulation-based approaches. For tail probabilities where data are sparse, intervals can widen substantially, reflecting higher uncertainty.

5.2 Hypothesis testing

Hypothesis testing evaluates whether observed differences could reasonably arise under a null assumption. Tests must align with the estimand and the model structure implied by the data.

5.2.1 Log-rank test

The log-rank test compares survival curves between groups using event-time orderings. It effectively gives more weight to times where many participants are still at risk. When the hazard ratio is roughly constant, this test is often a powerful choice.

5.2.2 Alternative weighted tests

Weighted variants place different emphasis on earlier or later event times. If group differences concentrate early or emerge mainly in the tail, alternative weighting schemes can improve sensitivity relative to the standard log-rank test.

5.3 Covariate effects and interpretation

In regression models, covariate effects are expressed via hazard ratios (Cox) or time-acceleration factors (AFT). Interpretation depends on the modeling scale: a positive association with hazard implies shorter event times on average, but the exact translation requires attention to the assumed hazard or time model.

5.4 Stratification and time-varying effects

When baseline risk differs across strata, stratification allows different baseline hazards while estimating covariate effects within strata. Time-varying effects extend models by allowing covariate influence to change as time progresses, helping address potential non-proportionality or evolving risk patterns.

6 Practical workflows

Practical use of time-to-event analysis involves careful planning before any modeling step. The quality of results depends on event definition, censoring rules, and transparent reporting.

6.1 Defining the timeline and event rules

A workflow typically starts by specifying the timeline (what “time zero” is), the event criteria (what counts as an event), and the observation window (when follow-up starts and ends). This step benefits from clear documentation and consistent application across records.

6.2 Data preparation and censoring rules

Data preparation translates raw records into a survival dataset: each unit receives an observed time and an event indicator. Analysts define censoring when follow-up ends without event, when events are missing, or when measurements stop, ensuring that censoring is recorded according to the study’s rules rather than convenience.

6.3 Model selection strategy

Model selection balances interpretability, flexibility, and assumptions. A common strategy begins with nonparametric summaries to inspect survival curves, then chooses parametric or semi-parametric models based on whether the hazard shape appears compatible and whether covariate effects are plausibly proportional or time-varying.

6.4 Reporting results and visualizations

Reporting often includes survival plots, estimated medians or quantiles, effect estimates with confidence intervals, and a description of censoring. Visualizations can include Kaplan–Meier curves by group and, when relevant, hazard-related summaries. Clear language helps readers understand what “event-free” means at each time point.

6.5 Common pitfalls (bias, missingness, mis-specified events)

Pitfalls include mis-specified event definitions, incorrect censoring (such as censoring individuals in a way that depends on unobserved future event times), and missing covariate handling that introduces bias. Another common issue is selecting a model without checking its core assumptions, leading to overconfident or misleading inference.

Extensions broaden the framework to address more complex study designs and event structures. Related concepts can clarify how to interpret competing outcomes or repeated failures, and how to incorporate evolving information.

7.1 Left truncation and delayed entry

Left truncation occurs when units enter the risk set only after they have already survived for some time, such as when inclusion depends on having passed an initial eligibility window. Delayed entry adjusts the risk set so that only eligible time contributes to estimation, preventing systematic overestimation or underestimation of early risk.

7.2 Recurrent events (repeated occurrences)

Recurrent events involve multiple event occurrences for the same unit over time. Methods can model event rates across episodes or incorporate dependence between repeated events. These approaches differ from “first event only” analyses, since timing and frequency both matter.

7.3 Time-dependent covariates

Time-dependent covariates change during follow-up, such as lab values or exposure levels. Models must update covariate values as time advances, and care is needed to ensure that covariate measurement timing aligns properly with risk-set membership and event observation.

7.4 Competing endpoints vs competing risks (distinctions)

Competing endpoints can refer more loosely to multiple outcomes of interest recorded in a study. Competing risks is a more specific statistical situation where one endpoint precludes another for the unit, affecting the probability structure. Distinguishing these notions is important: methods for competing risks treat alternative outcomes as informative rather than as independent or non-interfering endpoints.