Can a survival curve rise?

A valid survival function cannot increase with time.

Is survival always about death?

No; it applies to defined event times.

Is one minus survival always a specific competing-event probability?

No; the event definition and competing-risk analysis matter.

AI Glossary / Statistics and probability / Intermediate

Survival function

Definition

A survival function gives the probability that an event time exceeds a specified time. For an event time T, the usual definition is S(t) = P(T > t). It describes how the probability of remaining event-free changes over time, whether the event is failure, departure, purchase, or another clearly defined outcome.

What it means

A survival curve answers a horizon-specific question: what proportion of the relevant population is expected to remain free of the event beyond this duration? It does not predict an exact event date for each individual.
The name comes from survival studies, but the mathematical object applies more broadly. For time until a customer makes a first purchase, for example, survival means still not having purchased. The business interpretation therefore depends on the event definition, not on whether a high curve sounds desirable.

How it works

Define the event time and its origin. Time since registration, time since activation, and calendar date answer different questions. State the units and whether the event is the first occurrence, a repeat occurrence, or a transition between states.
The survival function is the complement of the cumulative distribution function: S(t) = 1 - F(t), where F(t) = P(T <= t). Thus S(t) cannot increase as t grows and must remain between zero and one.
For continuous event times, the distinction between strictly beyond t and at least t does not change probabilities. For discrete times or point masses, endpoint conventions matter. State the definition rather than switching inequalities unnoticed.
A survival function may come from a stipulated probability distribution or be estimated from observations. Kaplan-Meier estimation handles suitable right-censored data through event-specific risk sets. Parametric or regression models add assumptions that can support conditional predictions.
Distinguish the curve from a hazard. For a continuous distribution with density f, the hazard is f(t)/S(t) where S(t) is positive. It describes the instantaneous event rate among those still event-free, rather than the cumulative probability of having an event.
Conditional survival beyond t+u given survival beyond t is S(t+u)/S(t), provided S(t) is positive. This relationship helps update a forecast for a unit that has already remained event-free for some time.
Summaries of the curve include a median event time, commonly defined as the first time the curve falls to or below one half, and restricted mean survival time, the area under the curve up to a specified horizon. The unrestricted mean requires an integrable tail and adequate assumptions or observation of that tail.
Censoring and competing events affect estimation. An apparent flat tail after few units remain under observation is not strong evidence that risk has disappeared. For a particular event in the presence of competing events, distinguish its cumulative incidence from one minus a survival curve obtained by censoring the alternatives.
When presenting an estimated curve, include uncertainty and numbers at risk. A smooth fitted line can hide limited follow-up or strong extrapolation beyond the observed data.

Three practical examples

Retention planning: estimate the chance that an account remains active beyond six months.
Reliability: describe the probability that a device operates beyond a specified age without its first failure.
AI workflow evaluation: estimate time until a task finishes, interpreting survival as still awaiting completion.

Think of it this way

Think of a timeline with a question at every point: who has not yet experienced the event? The survival function connects those probabilities across the timeline. It is a population-level description, not a schedule assigning each individual their event date.

Worked example

Use an original hypothetical exponential event-time model with rate 0.10 per month. Its survival function is S(t) = exp(-0.10t) for nonnegative t. This model assumes a constant hazard; that is a stipulated simplification.
At ten months, S(10) = exp(-1), approximately 0.3679. The probability of an event by ten months is 1 - 0.3679 = 0.6321. These quantities are complements, not alternative estimates of the same outcome.
The median solves exp(-0.10m) = 0.5, giving m = log(2)/0.10, approximately 6.931 months. The mean is 1/0.10 = 10 months, so the mean and median differ.
For a unit already event-free at five months, the probability of remaining event-free for another ten months is S(15)/S(5) = exp(-1.5)/exp(-0.5) = exp(-1), approximately 0.3679. This equality with S(10) reflects the exponential model's memoryless property. It is not true for every survival distribution.

Illustrative case study

Fictional case: Cedar Queue estimates time until an AI-assisted request is completed. A manager sees a high survival curve and assumes it indicates good performance. Analysts explain that the event is completion, so a high curve means more requests remain unfinished. They relabel the display, add completion probabilities at useful horizons, and show how many requests remain under observation. The team also checks whether withdrawn requests are appropriately handled rather than interpreting every disappearance as successful completion.

Common misconceptions and mistakes

Do not confuse survival probability with cumulative event probability. A hazard is not a probability at a single instant. A high survival curve is not always a favorable outcome. The median may be unestimable within observed follow-up if the curve never falls to or below one half. A model's extrapolated tail is not directly observed evidence. The exponential memoryless property is a special assumption.

Business use

Use survival functions to communicate time-based reliability, retention, and completion expectations at horizons that matter for decisions. Keep the event label visible and explain whether the curve is empirical, fitted, or extrapolated. Combine probabilities with consequences and costs when choosing operational thresholds.

Check your understanding

What is the ten-month event-free probability in the example?

Answer: Approximately 36.79%.

What is the probability of the event by ten months?

Answer: Approximately 63.21%.

Why does the conditional ten-month probability stay the same?

Answer: Because the stipulated exponential distribution is memoryless.

Related terms

Learn next

Compare survival, cumulative incidence, and hazard plots for the same event process. Then study how nonconstant hazards and censoring change estimation and interpretation.

Sources

Harrell, Glossary of Statistical Terms: https://hbiostat.org/glossary/ SciPy, Exponential distribution: https://docs.scipy.org/doc/scipy/reference/generated/scipy.stats.expon.html Therneau, survival package vignette: https://cran.r-project.org/web/packages/survival/vignettes/survival.pdf The exponential example is original.

Previous: Survival analysis Next: Survival time