A patient enters a five-year study.
At the end of five years, the event being studied has not happened.
What is the patient’s event time?
We do not know.
But we know something important: it is longer than five years.
That is censoring: not a blank, not a zero, and not a completed observation.
Quick Read
Censoring occurs when the value of interest is only partly known because observation ends, begins too late, or captures only an interval containing the event. In time-to-event research, right censoring means the event is known to occur later than the last observed time; left censoring means it occurred before a known point; interval censoring means it occurred somewhere between two observations.
One-sentence answer: censoring is lossy because the process contains a more precise event time than the observer was able to record, yet the surviving partial information still constrains what the missing value can be.
Censoring Is Not Missing Data in the Ordinary Sense
If an event time is completely missing, we may know nothing about it.
If it is right-censored at 60 months, we know the event time exceeds 60 months.
The observation is incomplete but informative.
This is why replacing a censored value with 60 months is wrong: 60 is the last known event-free time, not necessarily the event time.
Right Censoring: We Stop Watching First
A study ends before a participant experiences the event. A machine is still working when the reliability test finishes. A subscriber remains active when the observation window closes.
In each case, the true duration extends beyond the last observed point.
The exact endpoint is lost, but the lower bound survives.
Left Censoring: The Event Began Before We Could See It
Sometimes observation begins after the relevant process has already crossed a threshold.
A pollutant is already detectable at the first measurement. A disease marker is already above a threshold when monitoring starts. The exact crossing time lies somewhere earlier.
We know an upper bound, not the exact event time.
Interval Censoring: We Know the Event Happened Between Visits
A patient tests negative in January and positive in April.
The transition happened somewhere in that interval.
If we record April as though it were the exact event date, we manufacture precision the observation schedule never supplied.
Sampling and censoring meet here: the event exists continuously, while the observer checks intermittently.
Survival Analysis Was Built for This Kind of Partial Knowledge
Ordinary averages can behave badly when many event times are censored because the incomplete cases cannot simply be treated as completed durations.
Survival-analysis methods explicitly incorporate censored observations. The Kaplan–Meier estimator, for example, updates estimated survival at observed event times while accounting for how many individuals remain at risk.
The important idea is not the formula alone. It is the epistemic discipline: use every constraint the observation gives you without inventing the endpoint it did not give you.
Informative Censoring Is the Dangerous Case
Many standard methods rely on assumptions about censoring.
If people leave a study for reasons related to their unobserved event risk, censoring can become informative. A participant whose condition worsens may drop out. A failing machine may be removed from service before formal failure. A struggling student may stop attending before the final assessment.
Now the missing endpoint is entangled with why observation stopped.
The censoring mechanism becomes part of the model.
Censoring Is Not Truncation
With censoring, the case is present but incompletely observed.
With truncation, some cases never enter the observed dataset at all because they fall outside an inclusion window or detection rule.
This distinction matters enormously. Censoring leaves a visible partial record. Truncation can remove the existence of the case from the analyst’s world.
Reliability Engineering Lives With Censoring
Test one hundred components for 10,000 hours.
Some fail. Others are still operating when the test ends.
The survivors are not 10,000-hour failures. They are components known to have survived at least that long.
Throwing them away wastes information. Treating them as failures corrupts information. Censoring-aware analysis preserves the partial truth.
Education Has Censored Learning Trajectories
A student leaves a tuition programme in June.
The final examination is in October.
The tutor knows the student’s trajectory until June, not the October outcome.
Filling the missing result with the June prediction would confuse forecast with observation. Ignoring the student entirely would discard months of evidence.
The honest representation keeps the trajectory and marks where observation ended.
Archives Can Be Censored by Time
A diary ends in 1943.
Did the writer stop believing what they wrote? Leave the country? Lose the notebook? Die? Continue elsewhere?
The surviving record tells us the history up to its boundary. The boundary itself does not identify the cause.
Historical reasoning often works with censored lives and institutions even when the statistical term is not used.
When Censoring Is Harmless Enough
Censoring is not automatically a fatal flaw.
If the observation process and analytical assumptions are appropriate, censored cases can contribute valid information. Modern survival methods exist precisely because incomplete follow-up is common and manageable.
The danger is pretending partial observation is complete observation.
A Practical Censoring Audit
- Boundary: where did observation begin or end?
- Type: is the case right-, left- or interval-censored?
- Constraint: what lower, upper or interval bound survives?
- Mechanism: is censoring plausibly independent of the unobserved event after conditioning on relevant information?
- Method: does the analysis explicitly support censored observations?
- Confusion: has a last-seen time been mistaken for an event time?
- Provenance: can downstream readers tell which observations were complete and which were censored?
Continue Through eduKateSG
Continue with How Lossy Works | Imputation and How Lossy Works | Sampling. Sampling creates gaps between observations; censoring marks where an endpoint remains only partly known; imputation asks whether and how a missing value should be reconstructed.
Final Thought: Incomplete Does Not Mean Empty
A censored observation is a disciplined kind of ignorance.
It says: I cannot tell you the exact answer, but I can tell you where the answer cannot be.
Good reasoning protects that boundary instead of filling it with confidence.