Multiple Reasons for Censoring

Why Distinguish Censoring Causes?

A participant may stop contributing event-time information because follow-up ends at a database freeze, because the participant withdraws, or because contact is lost. These exits all censor the event time, but they need not arise from the same process. A database freeze may be largely determined by enrollment date and a fixed closing date; loss to follow-up may be related to health, access to care, or prior symptoms. Combining the reasons without checking their predictors can hide a mechanism that needs different covariates or a different model.

The purpose of separating causes is not to treat each censoring reason as an event of scientific interest. It is to describe and model how observation ends, so that assumptions and adjustments for censoring are credible. An event of interest, including a competing event, is an outcome—not a censoring reason.

For example, \(J=1\) could denote administrative closure, \(J=2\) withdrawal, and \(J=3\) loss to follow-up. The labels must be operationally defined before analysis: if several explanations are recorded, specify a consistent rule for assigning the primary reason. If a reason is unknown, retain an explicit unknown category or describe how missing cause labels are handled rather than silently assigning them to a known cause.

Cause-Specific Censoring Processes

Write \(H(t)\) for the measured history available just before time \(t\), including relevant baseline covariates and time-varying information. For each censoring cause \(k\), define its cause-specific censoring hazard as the instantaneous rate of censoring for reason \(k\), among participants who are event-free, uncensored, and under observation just before \(t\):

The total censoring hazard is the sum of these mutually exclusive cause-specific hazards:

Events remove a person from the risk set for later censoring, but they are not included in this sum. In a cause-specific censoring model, an observed event is therefore a competing exit from the censoring risk set: it ends the opportunity to be censored later. This risk-set convention does not recode the event as a censoring cause.

A model might use separate cause-specific regressions, such as a Cox model for each censoring cause, or a joint model with a cause-by-predictor interaction. In either case, the outcome is the censoring reason at its time of occurrence, and the risk set contains participants still event-free and observed. The choice should reflect the data structure and the intended use of the fitted probabilities.

Combining Reasons Into an All-Cause Censoring Probability

When causes are mutually exclusive, the chance of censoring in a short interval is the sum of the cause-specific chances, to first order in continuous time. Along a specified covariate and history path \(h(u)\), the net probability of avoiding all censoring causes through time \(t\), under a regime that prevents event exits, is

This expression is conditional on the specified path and is a net censoring survival quantity because the hazards condition on remaining event-free; it is not generally the actual probability of avoiding censoring in the competing-risks process. If \(H(u)\) is stochastic and only the starting history is conditioned on, one cannot generally substitute the hazard along an unspecified future history and claim the same exponential formula. The conditional probability for a starting history must account for the distribution of future histories. In practice, models for time-varying histories use the observed history as it evolves, with assumptions about how that history is measured and modeled.

In discrete time, let \(q_k(t)\) be the conditional probability of censoring for cause \(k\) in interval \(t\), among those still event-free and observed at the interval’s start. If causes are mutually exclusive within the interval, the conditional probability of no censoring in that interval is \(1-\sum_k q_k(t)\). Multiplying these conditional probabilities across intervals gives the corresponding net censoring survival probability under a regime that prevents event exits, along the history path; it does not generally give the actual probability of avoiding censoring when events can remove participants from later risk sets. This is different from multiplying separate cause-specific “survival” probabilities, which would incorrectly treat simultaneous possibilities as independent.

Do Different Causes Have Different Predictors?

A useful comparison asks whether a predictor has different associations with the cause-specific censoring hazards. For instance, a calendar-time variable may predict administrative closure strongly, while prior symptoms may predict withdrawal or loss to follow-up. A joint model can express this by allowing cause-specific baseline hazards and cause-by-predictor interactions. The null hypothesis that a predictor has the same log-hazard coefficient for two causes is a direct contrast between coefficients.

Do not conclude that effects differ merely because one cause-specific coefficient is statistically significant and another is not. The relevant test is the difference itself, with its standard error accounting for covariance between estimates. Cause-specific models use overlapping risk sets, so treating their coefficient estimates as independent without justification can produce an incorrect standard error.

A global comparison can test all cause-by-predictor interaction terms jointly, for example with a likelihood-ratio test comparing a model with cause-specific predictor effects against one constraining the effects to be equal. A descriptive table of censoring reasons by covariate may help identify patterns, but comparing only the characteristics of censored participants can be misleading: it ignores how many people with each characteristic were at risk and for how long.

Worked Example: Choosing a Censoring Model

Consider a hypothetical multi-site study with two recorded censoring reasons: a scheduled database closure and participant withdrawal. Database closure is determined by enrollment date and the study’s fixed calendar end; withdrawal is potentially associated with recent symptoms and treatment burden. All participants who experience the event of interest remain classified as having an event, not as censored.

A single pooled censoring model would be reasonable only if it represents the total censoring probability adequately given its predictors and functional form. If the causes have different predictors, a pooled model that omits cause distinctions can average unlike relationships. Here, a defensible starting point is a cause-specific model that includes calendar time or enrollment date for closure and symptom history and treatment burden for withdrawal. The models should use risk sets that include only participants still event-free and under observation, and event types should be recorded separately.

Before using the fitted models, check that the reason labels are reliable, that each cause has enough events to estimate its model, and that predicted censoring probabilities are not near zero in covariate histories that matter for the target analysis. Compare observed and predicted censoring patterns over time and across important covariate strata. If closure is nearly deterministic given calendar time, a flexible calendar-time term may be needed; if withdrawal depends on a time-varying symptom measure, that measure must be recorded before withdrawal and included appropriately. These checks do not prove that censoring is ignorable, but they can reveal an observation model that fails to represent known features of the data.

Once each cause-specific hazard is estimated, combine them to obtain the net all-cause censoring survival probability under a regime that prevents event exits, along the relevant history. For a hypothetical participant with predicted one-year censoring probabilities of \(0.05\) for withdrawal and \(0.02\) for closure in one interval, the no-censoring probability for that interval is \(1-0.05-0.02=0.93\), not \((1-0.05)(1-0.02)=0.931\). Small differences can accumulate over many intervals.

Assumptions, Diagnostics, and Limitations

Separating reasons improves transparency, but it does not by itself correct informative censoring. For an inverse-probability censoring adjustment to support the target event analysis, the censoring process must be conditionally independent of the relevant unobserved event experience given the modeled history, and the required censoring probabilities must be positive. A well-fitting model for recorded censoring reasons cannot compensate for unmeasured predictors of dropout or for a history variable that was not captured before censoring.

• Check cause definitions. Ambiguous or inconsistently recorded reasons can make cause-specific comparisons meaningless. Report how unknown and multiple reasons were handled.
• Check temporal ordering. Covariates used to predict censoring must be available before the censoring time. Information recorded only after withdrawal cannot be used as though it were a prior predictor.
• Check model fit by cause. Compare predicted and observed censoring patterns over time and within covariate groups. A good fit for total censoring can conceal poor fit for a particular cause.
• Check sparse causes and overlap. Few observations for a cause can produce unstable coefficients. Near-zero predicted probability of remaining uncensored can produce extreme weights and indicate a positivity problem.
• Check calendar structure. Administrative censoring may be tied to calendar date or staggered entry rather than participant characteristics. Ignoring the design can make the model’s implied follow-up probabilities wrong.

A pooled model is simpler and may be adequate when causes have similar predictor relationships or when the target only requires a sufficiently accurate model for total censoring. Separate cause-specific models are more interpretable when the mechanisms differ, but they require more parameters and can be unstable for rare causes. A joint model with cause-by-covariate interactions offers a middle ground: it shares structure while allowing selected effects to differ. The choice is a modeling decision to be checked against the study design and diagnostics, not a rule that every reason must always receive a fully separate model.

The most important pitfall is to treat every exit from follow-up as censoring. If an event of interest or a competing event is coded as a censoring cause, the model estimates the wrong observation process and may alter the estimand. The correction is to preserve event outcomes as event outcomes, use them to end the censoring-model risk set, and define censoring causes only for genuine losses of event-time observation.