Marginal and Conditional Independent Censoring

Why the Level of Independence Matters

Right censoring hides an event time when follow-up ends before the event is observed. The key question is not simply whether censoring occurs, but whether people who remain under observation represent those who have been censored with respect to their event experience. That question can have different answers before and after accounting for observed covariates.

For example, a baseline measure of disease severity might predict both earlier events and earlier withdrawal. Within severity groups, withdrawal could be unrelated to event time, while the groups with more withdrawal also have different event risks. A pooled analysis can then have distorted risk sets even though censoring is independent within each group.

Marginal independence is a stronger claim than conditional independence in the sense relevant to analysis: it asserts independence without conditioning on \(X\). Conditional independence does not generally imply marginal independence. If \(X\) predicts both event risk and censoring, mixing over \(X\) can create an association between \(T\) and \(C\), even when they are independent within every value of \(X\).

The independence assumption is not established just because censoring is administrative, or because censoring times are recorded accurately. It is an assumption about the relation between censoring and the event process. Covariates used to make that assumption plausible should be observed before the relevant follow-up and should capture shared predictors of event risk and censoring. Conditioning on a variable that is itself affected by prior follow-up requires a more careful analysis than the baseline-covariate setting considered here.

What Conditioning Changes

A marginal survival probability at time \(t\) is \(S(t)=P(T>t)\). If \(X\) has covariate values \(x\), then this target is a population average of the covariate-specific survival probabilities \(S(t\mid x)=P(T>t\mid X=x)\). When the censoring distribution also differs across \(x\), people still under observation at time \(t\) need not have the population distribution of \(X\).

Here \(G(t\mid x)=P(C\ge t\mid X=x)\), and the sums are written for discrete covariates. Under \(T\perp C\mid X\), the probability of surviving past \(t\) among those with \(X=x\) and still observed at \(t\) is \(S(t\mid x)\). But the weights \(P(X=x\mid C\ge t)\) among those still observed can differ from the target population weights \(P(X=x)\). Thus the displayed probability need not equal \(S(t)=\sum_x P(X=x)S(t\mid x)\).

This calculation is a selection argument, not a claim that the Kaplan–Meier estimator is merely the proportion of observed people who survive to \(t\). Kaplan–Meier uses the sequence of event times and risk sets to account for censoring. The argument explains why, when censoring depends on covariates that predict events, pooled risk sets may not support the intended marginal estimate. Marginal independence is a standard sufficient assumption for an unadjusted Kaplan–Meier analysis; conditional independence alone does not generally justify pooling over \(X\).

When a Pooled Analysis Can Be Misleading

Valid Conditional Analyses and Their Requirements

If censoring is independent only conditional on \(X\), an analysis can use the covariate information rather than treating the pooled sample as if censoring were marginally independent. One direct approach is to estimate survival within covariate groups and average those estimates over the covariate distribution of the target population. For continuous or numerous covariates, a regression model can estimate conditional survival, followed by averaging predictions over that target population. Either approach must match the estimand: averaging over the study sample targets that sample’s covariate distribution, while a different target population requires its own weights or covariate distribution.

Conditional independence is not enough by itself. For a conditional analysis to support the intended estimate, the following conditions also matter:

• Relevant covariates are observed. After conditioning on \(X\), there should be no remaining common predictors of censoring and event time that induce dependence.
• Covariates are appropriately measured. The values used for adjustment must represent information available at the relevant point in follow-up. Misclassification or poorly timed measurement can leave important dependence uncontrolled.
• Positivity holds. For covariate values in the target population, people must have a positive probability of remaining uncensored over times at which survival is to be estimated. In notation, \(G(t\mid x)>0\) over the relevant time range for target values \(x\).
• The conditional estimation procedure is adequate. Stratification must leave enough observations in relevant groups; a regression approach must represent the conditional survival relationships well enough for the target estimate.
• The target population is specified. Conditional survival probabilities must be averaged using the covariate distribution appropriate to the question.

These requirements explain why simply adding covariates to a model does not guarantee valid inference. A covariate must address relevant dependence, and the analysis must have enough follow-up information in each target subgroup. Conditional independence is an assumption about the data-generating process; a fitted model cannot prove it.

Choosing an Analysis and Checking Its Weak Points

The main alternatives address different assumptions and targets. An unadjusted Kaplan–Meier analysis is appropriate under marginal independent censoring for the population survival curve. A stratified or regression-based analysis can use conditional independence given observed covariates, with standardization when a marginal population survival estimate is desired. Inverse-probability-of-censoring weighting is another approach that uses a model for remaining uncensored to reweight observed follow-up; it also depends on adequate covariate measurement, a correctly specified censoring model, and positivity. None of these approaches repairs unmeasured predictors of both censoring and event time merely by changing the estimator.

Useful diagnostics include comparing censoring frequency and follow-up across important baseline predictors, checking whether covariate groups retain adequate follow-up over the time range of interest, and assessing whether conclusions are sensitive to reasonable alternative covariate specifications. A relationship between \(X\) and censoring does not by itself prove that censoring is informative after conditioning; it signals why a pooled analysis deserves scrutiny. Conversely, similar censoring rates across recorded groups do not rule out dependence through unmeasured factors.

A frequent mistake is to reason, “censoring is independent within each group, so pooled Kaplan–Meier is valid.” The correction is that conditional independence does not generally imply marginal independence: different censoring rates can alter the covariate mix of pooled risk sets. Either justify the marginal assumption for the pooled analysis or use an analysis that conditions on the relevant covariates and targets the desired population.

Finally, censoring must be distinguished from an observed competing event. In a competing-risks setting, a competing event is an observed outcome of a different type, not automatically a censoring time. Treating it as censoring invokes additional assumptions and changes the quantity being estimated; the independence discussion here applies to genuine right censoring of the event-time information.