Independent Censoring

Independent Censoring: Concept Architecture

What can be learned after follow-up stops

In time-to-event analysis, independent censoring means that, at a given analysis time and within the groups or histories used for the analysis, leaving observation does not reveal additional information about a person's future event time. It is an assumption that permits the observed patients still at risk to represent relevant patients whose event times are no longer observed. It is often called non-informative censoring, but independence must be stated relative to a particular event, estimand, and conditioning set.

Right censoring occurs when follow-up ends before the event is observed. A study may know that someone remained event-free until their last contact but cannot know their subsequent event time. Planned study closure, loss to follow-up, withdrawal, treatment change, and a competing event can all end or alter observation, but they do not automatically meet the same assumptions. Independent censoring does not mean that censored people never experience the event; it means the analysis can account for their observed event-free time without systematically misrepresenting their unobserved future.

Define event time, censoring time, and observed data

Let $T$ be the event time from a clearly defined origin and $C$ the time observation would stop in the absence of the target event. For ordinary right-censored data, the recorded duration and event indicator are:

$$Y=\min(T,C),\qquad \Delta=I(T\leq C).$$

Here $\Delta=1$ denotes an observed target event and $\Delta=0$ denotes censoring. State the convention for simultaneous event and censoring times. A subject censored at month six contributes information that the event did not occur before month six; their status at month seven is unknown. Censoring is therefore different from an observed event-free outcome at the full analysis horizon and from excluding that subject entirely.

A simple unconditional statement is $T\perp C$ within the population to which an unadjusted survival curve is applied. More often, a defensible assumption is conditional independence $T\perp C\mid X$ for measured baseline covariates $X$, treatment arm, or an appropriate observed history. These are not equivalent: if frailer patients are more likely to leave, censoring may be independent within measured risk groups yet informative for a pooled, unadjusted curve. With time-varying factors, one needs a sequential condition given the information available just before each possible censoring time, not a future variable measured after censoring.

Why the Kaplan–Meier risk set needs the assumption

The Kaplan–Meier estimator multiplies the fraction surviving at each observed event time. If $d_j$ target events occur among $n_j$ people at risk just before event time $t_j$, then:

$$\widehat{S}(t)=\prod_{t_j\leq t}\left(1-\frac{d_j}{n_j}\right).$$

Someone censored before $t_j$ leaves that risk set. For the product to estimate survival in the intended population, the remaining risk set must give a valid representation of those at risk at that time, within the strata or model being used. Random administrative end of study can be plausible for some estimands, but delayed entry, calendar-time changes, prognosis-related withdrawal, and other design features still need examination. Random treatment allocation does not make later loss to follow-up independent.

An unadjusted Kaplan–Meier curve does not generally recover marginal population survival if censoring depends on a baseline prognostic factor, even if censoring and event time are independent conditional on that factor. One option is to estimate valid within-stratum survival curves and average them using the target population's baseline stratum distribution. A correctly specified regression or inverse-probability-of-censoring weighted approach may be appropriate when stratification is insufficient. Each remedy has additional model and positivity requirements.

An illustrative risk-mix example

Suppose a cohort has 100 high-risk and 100 low-risk people, with all target events scheduled at one year for simplicity. In the full cohort, 40 high-risk and 10 low-risk people would have the event by then, so true one-year event-free survival is $(200-50)/200=75\%$. Before that event time, 80 high-risk people are censored, selected independently of their future outcome within the high-risk group. The 20 high-risk people still followed therefore have eight events in this idealized example; all 100 low-risk people are observed and have ten events.

If all events occur together at one year, the pooled risk set contains 120 people and 18 events, giving an unadjusted Kaplan–Meier survival of $1-18/120=85\%$. It overstates survival for the original 200-person cohort because high-risk people were disproportionately censored. Averaging the two valid group estimates with the original half-and-half population mix gives $0.5(0.60)+0.5(0.90)=0.75$. These counts are deliberately constructed to illustrate the mechanism; real studies involve multiple event and censoring times and sampling uncertainty.

The example does not imply that merely including a risk-group label always repairs bias. All variables that jointly predict censoring and future events need appropriate measurement and modeling, with sufficient observed follow-up in each relevant group. If nobody in a subgroup remains under observation to a target time, its survival at that time is not identified from the observed data without further assumptions.

Check plausible censoring mechanisms

Study closure can produce administrative censoring when event-free participants reach a predefined end date. Its independence is more plausible if enrollment and calendar time do not carry unmodeled prognostic information for the time scale under analysis. Loss to follow-up following worsening symptoms, toxic effects, inability to attend visits, or imminent death can be informative. Censoring when a patient switches treatment may also select patients by clinical trajectory; whether it is appropriate depends on whether the estimand asks about assigned treatment regardless of switching or a hypothetical no-switch strategy.

In oncology, progression-free survival analyses can be biased when people starting another therapy are censored because their treatment decision reflects impending progression. Conversely, an observed death may itself be part of a composite event definition rather than a censoring event. In competing-risks analyses, censoring a person at a competing event and applying $1-\widehat{S}$ for a cause-specific event does not generally give that cause's real-world cumulative incidence. Define whether the target is a cause-specific hazard, a hypothetical world without competing events, or the actual cumulative incidence, and choose the estimator accordingly.

Adjustment, assumptions, and sensitivity analysis

If censoring is independent given observed predictors, inverse-probability-of-censoring weighting (IPCW) estimates the chance of remaining observed and gives more weight to otherwise similar people who remain under follow-up. In simplified notation, a contribution while under observation may be weighted by $1/G(t\mid X)$, where $G(t\mid X)=P(C\geq t\mid X)$ is the modeled probability of no censoring through $t$. The exact risk-set and event-time weight conventions depend on the estimator. Require positivity, $G(t\mid X)>0$, for histories and times being used, and avoid interpreting enormous weights as reliable recovered information.

IPCW relies on a suitable censoring model and on no unmeasured common causes of censoring and the event after conditioning. Outcome regression, multiple imputation of unobserved event times under stated assumptions, joint models, or sensitivity analyses can address related questions, but none can verify unseen outcomes directly. Display censoring by arm, time, site, and prognostic group; document reasons for withdrawal and report follow-up coverage, numbers at risk, and the period over which estimates remain supported by data. Sensitivity analyses should ask how conclusions change under plausible event patterns among those lost to follow-up, especially when few remain at risk late in follow-up.

Implications for health economic models

Biased survival curves can distort extrapolated life-years, QALYs, state occupancy, treatment duration, costs, and incremental cost-effectiveness. A small difference in early censoring patterns may be amplified by a long model horizon. Compare adjusted and unadjusted estimates when justified, check whether treatment arms have different follow-up mechanisms, and carry uncertainty in survival and censoring adjustments into the economic model. Do not treat a late Kaplan–Meier plateau with few patients still observed as evidence of a permanent cure without further clinical and modeling support.

When presenting a survival estimate, name the event, time origin, censoring rules, target population, assumption (unconditional or conditional), observed adjustment variables, censoring counts, and sensitivity analyses. Independent censoring is a bridge from partial follow-up to an event-time estimate; credibility depends on the study design and on what is known about those who leave observation.

Sources and further reading

• Satten GA, Datta S. The Kaplan–Meier estimator as an inverse-probability-of-censoring weighted average. The American Statistician (2001; open manuscript). Establishes the link between product-limit estimation and censoring weights.
• Campigotto F, Weller E. Impact of informative censoring on the Kaplan–Meier estimate of progression-free survival in phase II clinical trials. Journal of Clinical Oncology (2014). Examines the consequences of censoring associated with subsequent cancer therapy.
• Therneau TM. R survival package: Surv documentation. Primary software documentation for the observed time and event indicator used for right-censored data.
• Prince T, et al. On the estimation of inverse-probability-of-censoring weights for the evaluation of survival prediction error. PLOS ONE (2025). Examines practical estimation of censoring weights for a survival prediction metric.