AI Glossary / Statistics and probability / Intermediate
Censoring
Definition
Censoring occurs when an observation provides only partial information about a value, such as knowing that an event time exceeds a recorded follow-up time. The observation is not completely missing: its bound or interval contains useful information. Statistical methods must use that information appropriately rather than treating the recorded limit as the exact outcome.
What it means
In right censoring, an event has not been observed by a particular time, so its time is known to be later than that point under the stated event definition. Left censoring gives an upper bound, and interval censoring locates the event between two times without identifying the exact time. Censoring also occurs in measurements below detection limits or above instrument ranges. It differs from truncation, where certain units are absent from the observed sample because of a selection rule.
How it works
Record both the observed time or bound and an indicator describing what is known. For right-censored event data, a unit contributes event-free follow-up until censoring but is not counted as having the event at that time. Kaplan-Meier estimation updates survival at event times using the units still at risk, while suitable parametric models use survival probabilities for censored observations in the likelihood. The censoring mechanism matters: common methods require independence or a suitable conditional version of it. If withdrawal relates to unaccounted-for event risk, ordinary analysis can be biased and additional modeling or sensitivity analysis may be needed.
Three practical examples
- Reliability testing: a component still working when a test ends contributes a lower bound on its lifetime.
- Customer research: an account still active at the data cutoff contributes observed retention time, but its eventual churn time is not yet known.
- Laboratory measurement: a concentration reported below a detection limit provides an upper bound rather than a measured zero.
Think of it this way
Think of watching a race until the broadcast ends. You know which runners finished and how long the others had continued without finishing. The broadcast ending is not the same event as every remaining runner crossing the finish line.
Worked example
Suppose five units enter a hypothetical reliability study together. One fails at time 2, one is censored at time 3, one fails at time 4, one is censored at time 5 and one fails at time 6. With no tied times, the Kaplan-Meier estimate starts at 1. At time 2, five units are at risk and one fails, so estimated survival becomes 4/5 = 0.8. Censoring at time 3 does not lower survival; it removes that unit from later risk sets. At time 4, three units remain at risk and one fails, so survival becomes 0.8 × 2/3 = 8/15, approximately 0.5333. At time 5, another censoring again leaves the curve unchanged. At time 6, the last at-risk unit fails, giving a final estimate of zero. That final step is based on one remaining unit and is highly uncertain; it is not proof that all units in the population fail by time 6. The calculation assumes the censoring mechanism is suitable for Kaplan-Meier analysis. Treating censoring at times 3 and 5 as failures would produce a different, incorrect event history.
Illustrative case study
Fictional case: Cedar Equipment compares service lifetimes for two component designs. Many newer components are still operating at the review date. Analysts retain those observations as censored rather than deleting them or recording their current ages as completed lifetimes. They also investigate whether components removed for inspection are systematically more likely to fail. The report explains follow-up differences and uncertainty before comparing designs. This fictional case illustrates data handling and does not claim one component has superior reliability.
Common misconceptions and mistakes
Censored does not mean event-free forever, and it does not mean the event occurred at the censoring time. Deleting censored observations can waste information and distort comparisons. Replacing values below a detection limit with zero or a fixed fraction of the limit is not universally valid. Administrative study closure may support an independence assumption, but that assumption still depends on the study and time scale. A competing event is not automatically harmless censoring when estimating the probability of a particular event. The choice of estimand and competing-risk method matters.
Business use
Businesses analyzing retention, time to purchase, equipment failure or service completion should retain event definitions, observation windows and reasons for follow-up ending. Use methods compatible with the actual censoring pattern; interval observations need different treatment from exact event times. For AI prediction, avoid training on censored times as though they were complete labels. Evaluate predictions with methods that account for incomplete follow-up and inspect whether missing follow-up differs across groups. A larger dataset cannot compensate for systematically mislabeled censoring.
People also ask
Does censoring always happen at the end of a study?
No. It can occur earlier through loss of follow-up or other observation limits, and it can involve intervals or detection thresholds.
Does a censoring mark make a Kaplan-Meier curve fall?
No. It changes later risk sets; observed events cause the downward steps.
Can censoring be informative?
Yes. If its mechanism carries unaccounted-for information about event risk, standard assumptions may fail.
Check your understanding
Q: Why are only three units at risk at time 4 in the worked example?
Answer: A: One of the original five failed at time 2 and another was censored at time 3. Neither remains under event-free observation immediately before the time-4 failure.
Related terms
- Survival analysis
- Risk set
- Kaplan-Meier estimator
- right censoring
- left censoring
- interval censoring
- truncation
- informative censoring and competing risks
Learn next
draw individual follow-up timelines for the example and construct the risk set at each event time. Then compare a dataset with independent administrative censoring against one where high-risk units preferentially leave observation, explaining why the same calculation may have different validity.
More in Statistics and probability
- Categorical Naive Bayes - Categorical Naive Bayes is a classification method that models each input feature as a categorical variable and combines its class-conditional probabilities...
- Cause-effect diagram - A cause-effect diagram is a visual tool for organizing possible contributors to a clearly defined problem or outcome.
- Cell - In statistics, a cell is a particular combination of categories or factor levels, commonly represented by the intersection of a row and a column.
- Census - A census is a statistical collection that seeks information from every unit in a defined population, rather than selecting only a sample of those units.
- Centering - Centering transforms a variable by subtracting a reference value from each observation, most commonly the variable's mean.
- Centerpoint - A centerpoint is an experimental run with quantitative factors set at the middle of their selected low and high settings.
Sources
Sources: Harrell, statistical glossary, censoring definitions and informativeness: https://hbiostat.org/glossary/ ; NIST Engineering Statistics Handbook, censoring in reliability data: https://www.itl.nist.gov/div898/handbook/apr/section1/apr131.htm ; Penn State STAT 509, survival estimation and censoring assumptions: https://online.stat.psu.edu/stat509/Lesson06 .
Previous: Cellular neural network Next: Census