Is a censored time useless?

No; it tells us that the event was not observed before the recorded bound.

Can two identical durations mean different things?

Yes, if one is an event and the other is censored.

Is survival time always calendar time?

No; an appropriate exposure scale such as operating hours may be used.

AI Glossary / Statistics and probability / Intermediate

Survival time

Definition

Survival time is the elapsed time from a defined starting point to a specified event. In recorded survival data, the term may also refer to observed follow-up ending at either the event or censoring. A complete record must distinguish an observed event time from a duration for which the event has not yet been observed.

What it means

The word survival does not restrict the event to death. A survival time can describe time from activation to cancellation, installation to failure, or submission to completion. Its meaning depends on both endpoints.
If observation ends before the event, the true event time remains unknown. For example, an account still active after eleven days provides evidence that its cancellation time exceeds eleven days, not that cancellation occurred then. This distinction must remain visible in the data and analysis.

How it works

Choose a time origin that matches the question. It may be enrollment, installation, first use, or another clearly specified milestone. Using different origins for different groups can make durations incomparable and distort an apparent performance difference.
Define the event operationally. A cancellation request, effective closure, and inactivity threshold may occur on different dates. Record the chosen definition consistently and explain how ambiguous cases are resolved.
Select the time scale and measurement resolution. Calendar days, operating hours, age, and accumulated usage can represent different processes. Specify whether durations include the starting day and how fractional units or tied recorded times are handled.
For standard right-censored data, let T be the true event time and C the censoring time, both measured from the same origin. The observed follow-up is Y = min(T,C), accompanied by an indicator that says whether the event was observed.
Do not infer the event indicator from duration alone. Two records can both show eight days, while one is an event and the other is merely the last observation. They contribute different information to survival estimation.
When the event is discovered only during periodic inspections, its time may be interval-censored rather than exactly observed. If a device worked at day seven and had failed by day ten, the failure lies in that interval; assigning day ten as exact changes the information.
Delayed entry requires both entry time and exit time relative to the analysis origin. A unit first observed at age five is not automatically at risk under observation during ages zero through five. Appropriate risk-set construction respects that entry.
Repeated events need a clear choice between time since initial entry, time since the previous event, and other scales. Multiple intervals from one unit are related, so treating them as independent first-event times is generally inappropriate.
Keep dates, duration calculations, event definitions, and reasons for censoring traceable. Validate impossible negative durations, conflicting timestamps, and events recorded after observation ended. Good survival-time measurement is a prerequisite for sophisticated modeling, not a detail the model can repair automatically.

Three practical examples

Customer retention: measure elapsed time from subscription activation to effective cancellation.
Reliability: record operating hours from installation until the first defined failure.
AI operations: measure time from request acceptance to successful completion, separately recording requests whose observation ends unfinished.

Think of it this way

Think of starting a stopwatch for each unit at the same kind of milestone. Some stopwatches stop because the event happens; others stop because the observer leaves. Both show elapsed time, but the reason they stopped determines what the reading tells you.

Worked example

Consider an original three-account example using elapsed calendar days, calculated as end date minus activation date without counting the start day twice.
Account A activates on January 2 and cancels on January 10. Its observed duration is eight days and its event indicator is one. Account B activates on January 20 and is still active at the January 31 cutoff. Its observed duration is eleven days and its event indicator is zero; its cancellation time exceeds eleven days.
Account C activates on January 5 and is last confirmed active on January 19, after which tracking is lost. It contributes fourteen observed days with an event indicator of zero, assuming January 19 is the valid censoring date.
The arithmetic average of recorded durations is (8 + 11 + 14)/3 = 11 days. This is an average follow-up duration, not an estimate obtained by observing all three cancellation times. Account C's loss of tracking also requires investigation because censoring related to cancellation risk can undermine standard methods.

Illustrative case study

Fictional case: Elm Accounts combines cancellation records from two systems. One measures duration from registration and the other from paid activation, creating an artificial difference between customer groups. Analysts recover the dates, choose an origin consistent with the study question, and recalculate durations. They retain separate flags for cancellations and observation cutoffs. The cleaned dataset supports a meaningful retention analysis because the times now describe the same interval across groups.

Common misconceptions and mistakes

Do not treat every recorded follow-up duration as a complete event time. Do not average censored durations and call the result mean lifetime. A zero event indicator does not mean the event can never happen. Do not mix calendar age with operating time without explanation. Excluding units that fail before a later qualifying milestone can introduce selection problems rather than merely cleaning the data.

Business use

Use carefully defined survival times to make retention, maintenance, and completion analyses comparable. Record event status and observation reasons alongside durations, and preserve enough timing information to audit the calculation. Business teams should agree on the start and end definitions before interpreting differences or commissioning predictive models.

Check your understanding

What is Account A's event time?

Answer: Eight days under the stated elapsed-day convention.

Does Account B cancel on day eleven?

Answer: No; it is still active at observation cutoff.

Why is eleven days not the observed mean cancellation time?

Answer: Two of the three cancellation times are unknown.

Related terms

Time origin
Censoring
Risk set
Survival analysis
follow-up
event indicator
delayed entry
interval censoring

Learn next

Practice constructing time-and-status records from dates, then study how censoring and delayed entry affect risk sets. Compare calendar-time and exposure-time analyses for the same operational process.

Censoring

Sources

Harrell, Glossary of Statistical Terms: https://hbiostat.org/glossary/ Therneau, survival package vignette: https://cran.r-project.org/web/packages/survival/vignettes/survival.pdf The three-account date example is original.

Previous: Survival function Next: Survivorship Bias