Censored Data

Censored data are observations in survival analysis for which the exact time to the event of interest is not known because the event has not yet occurred by the end of the study period, the participant was lost to follow up before the event, or the participant experienced a competing event that prevents the event of interest from being observed (cause specific analyses treat such participants as censored, whereas competing risks methods treat the competing event as a separate outcome). Censoring is distinct from missing data because it provides partial information: we know the event had not occurred up to a certain time, even though we do not know if or when it occurred afterward.

Censoring is a defining feature of survival analysis and must be handled correctly to avoid biased estimates of event rates and survival probabilities. In clinical trials, censoring is common because trials have fixed end dates, participants withdraw, or, when the endpoint is specific to the disease, participants die from causes unrelated to the disease under study. Ignoring censored observations by excluding them from the analysis wastes valuable information and can introduce bias. Treating censored observations as though the event occurred at the censoring time is equally problematic. Proper statistical methods for censored data, such as the Kaplan Meier estimator and the Cox proportional hazards model, are essential tools in clinical and epidemiological research.

The most common type is right censoring, where observation ends before the event occurs. This can happen because the study ends (administrative censoring), the participant withdraws (voluntary censoring), or the participant is lost to follow up. In each case, we know that the participant was event free up to the last observation time but do not know what happened afterward.

Left censoring occurs when the event happened before observation began, so the exact event time is unknown but is known to be earlier than a certain date. Interval censoring occurs when the event happened between two observation times, but the exact date is unknown. For example, if a patient tests negative for a biomarker at month 6 and positive at month 12, the conversion event is interval censored between months 6 and 12.

The Kaplan Meier method is the standard nonparametric approach for estimating survival curves from censored data. It calculates the probability of surviving past each observed event time, using the number of participants at risk just before each event. Censored observations reduce the number at risk but do not contribute events, preserving their partial information in the analysis.

The Cox proportional hazards model extends survival analysis to the multivariable setting (several predictors at once), allowing researchers to examine the association between predictor variables and the time to event while properly accounting for censoring. The model produces hazard ratios, which quantify the relative rate of the event occurring in one group compared to another.

A critical assumption in most standard survival analyses is that censoring is non informative, meaning that censored individuals have the same future risk of the event as those who remain under observation. If censoring is related to the outcome (for example, if sicker patients are more likely to be lost to follow up), this assumption is violated and the results may be biased.

Survival analysis methods for censored data should be used whenever the outcome is the time until an event occurs and some participants have not experienced the event by the end of observation. This is the standard situation in clinical trials with time to event endpoints, prospective cohort studies, and epidemiological follow up studies. Standard methods (Kaplan Meier, Cox regression) are appropriate when censoring is non informative. When informative censoring is suspected, sensitivity analyses or methods such as inverse probability of censoring weighting should be considered.

A common error is excluding censored observations from the analysis, which not only reduces the sample size but also biases the results by overestimating the event rate (since only participants who experienced the event are retained). Another mistake is treating censored observations as events or as non events rather than using appropriate survival analysis methods that account for their partial information.

Researchers sometimes also fail to assess the non informative censoring assumption, which is critical to the validity of standard survival analysis methods but cannot be tested directly from the observed data. If participants who are censored differ systematically from those who are not, the survival estimates may be biased. Comparing baseline characteristics of censored and uncensored participants can provide some evidence on this point. Finally, ignoring competing risks (events that prevent the event of interest from occurring, such as death from an unrelated cause in a study of cancer recurrence) can overstate the probability of the event of interest (its cumulative incidence), for example when one minus a Kaplan Meier estimate that treats competing events as censored is read as that probability. Competing risks methods, such as the Fine and Gray model, should be considered when appropriate.

A randomized trial evaluating a new chemotherapy regimen for advanced colon cancer follows 400 patients for up to five years. The primary endpoint is death from colon cancer. By the end of the study, 210 patients have died of colon cancer (events), 95 are still alive at the study's end (administratively censored), 45 were lost to follow up at various time points (also right censored, at their last contact), and 50 died from causes unrelated to cancer (competing events, censored at their date of death in the Kaplan Meier and Cox analyses). The Kaplan Meier estimate of five year cancer specific survival is 38% (95% CI: 33% to 43%). Because deaths from other causes compete with death from colon cancer, one minus this figure (62%) overstates the proportion of patients who died of colon cancer within five years; a cumulative incidence function from a competing risks analysis estimates that proportion directly. The Cox regression model, adjusting for age, stage, and performance status, shows that the new regimen reduces the hazard of death from colon cancer by 28% compared to the standard regimen (HR = 0.72, 95% CI: 0.55 to 0.94, p = 0.017).