Independent Censoring for Cause-Specific Cox Models

Why the Censoring Assumption Matters

A cause-specific Cox model compares individuals who are still at risk for a particular event at each event time. Loss to follow-up can change who remains in those risk sets. The key question is not whether censoring occurs, or even whether its frequency differs by predictor. It is whether, after conditioning on the predictors used in the analysis, censoring still selects people according to their future event experience.

This distinction matters because censoring related to a measured predictor does not automatically invalidate a Cox model containing that predictor. The concern is dependence that remains after conditioning on the modeled information. If that dependence changes the failure experience of those left in the risk sets, the fitted coefficient may not estimate the intended conditional association.

Model and Conditional Independent Censoring

Definition: Let \(T\) be time to the first event, \(J\) its cause, \(C\) time to loss to follow-up or other right-censoring, and \(X\) the vector of modeled baseline covariates. The observed time is \(Y=\min(T,C)\). The cause-specific event indicator is \(\Delta_k=I(T\leq C,\ J=k)\). Conditional independent censoring means \(C\) is independent of \((T,J)\) given \(X\), written \(C\perp (T,J)\mid X\). In a cause-specific Cox model for cause \(k\), the event hazard is \(\lambda_k(t\mid X)=\lambda_{k0}(t)\exp(\beta_k^\top X)\).

The conditional independence statement is about the censoring time and the event process among people with the same modeled covariates. It does not require censoring to have the same distribution for every covariate pattern. The censoring survival function \(P(C\geq t\mid X)\) may vary with \(X\), provided the event experience is independent of censoring within levels of \(X\).

For example, a treatment group could have more administrative loss to follow-up because its participants entered the study later. If entry calendar time is adequately represented in \(X\), and within that information loss to follow-up carries no further information about future event times or causes, conditional independent censoring can still be plausible. By contrast, censoring because participants become too ill to continue, when that illness also predicts the event and is not captured by \(X\), is a warning of informative censoring.

The model also requires a correctly specified cause-specific hazard relationship for the covariates of interest and adequate support: within covariate patterns relevant to the coefficient, people must remain under observation long enough to contribute to the event-time comparisons. Independent censoring cannot correct a misspecified hazard model or a lack of follow-up support.

Worked Example: Predictor-Related Censoring That Is Not Automatically a Violation

Suppose a hypothetical study models a cause-specific hazard using binary treatment \(X\) and baseline severity \(Z\). The stipulated model is \(\lambda_k(t\mid X,Z)=\lambda_{k0}(t)\exp(0.70X+0.50Z)\). Loss to follow-up is more common for \(X=1\), but investigators judge that, conditional on \(X\) and \(Z\), the timing of loss is independent of future event time and cause.

For a one-unit treatment contrast at the same severity, the fitted hazard ratio is \(\exp(0.70)=2.014\). This describes the modeled cause-specific hazard contrast, not a causal effect. The different censoring frequency by treatment does not by itself invalidate this contrast: censoring may depend on \(X\), while remaining independent of \((T,J)\) conditional on \((X,Z)\).

The assumption would be questionable if, for example, loss among treated participants were especially likely for people with an unrecorded worsening condition that also predicts the event. Then conditioning on \(X\) and \(Z\) would not make censoring independent of future event experience. The numerical hazard ratio is an interpretation of the stated model, not evidence that the assumption holds.

Why Risk-Set Selection Can Bias Effects

The cause-specific Cox partial likelihood uses the covariates of the person who experiences cause \(k\) at an event time relative to the covariates of everyone still at risk at that time. If censoring preferentially removes people with unusually high or low future cause-specific hazards, the observed risk set no longer represents the appropriate conditional comparison. The resulting coefficient can be biased, even when the hazard model would have been correct in the full cohort.

A short justification shows where the assumption enters. With no tied event times, the Cox partial likelihood contribution at an observed cause-\(k\) event time \(t_i\) is based on the conditional probability that person \(i\) is the one to fail among the observed risk set \(R(t_i)\). Under the cause-specific proportional hazards model and conditionally independent censoring, this probability has the familiar form below. Censoring determines who is observed in \(R(t_i)\); conditional independence is what allows that risk-set comparison to represent the modeled event hazards rather than a further selection on future failure.

P(i\text{ has the cause-}k\text{ event at }t_i\mid R(t_i),\text{ one cause-}k\text{ event}) = \frac{\exp(\beta_k^\top X_i)} {\sum_{j\in R(t_i)}\exp(\beta_k^\top X_j)}.

The partial likelihood multiplies these contributions over cause-\(k\) event times. When censoring is informative after conditioning on \(X\), the observed membership of the risk set can depend on unobserved future event propensity. The displayed probability then need not hold for that selected risk set, so the usual partial-likelihood fit need not target the intended \(\beta_k\).

Worked Example: Selective Censoring Can Reverse a Risk-Set Comparison

Consider a hypothetical one-interval illustration with 100 people in each of two exposure groups, \(X=1\) and \(X=0\). In each group, 50 people have an unmeasured high-risk characteristic \(U=1\), and 50 have \(U=0\). There is no true exposure effect: in the next interval the cause-specific event probability is 0.20 for \(U=1\) and 0.05 for \(U=0\), regardless of \(X\).

Before that interval, suppose 40 high-risk people in the \(X=1\) group are censored, while nobody in the other three group-by-\(U\) cells is censored. The observed \(X=1\) risk set has 10 high-risk and 50 low-risk people, for 60 total. Its expected number of events is \(10(0.20)+50(0.05)=4.5\), giving an expected interval event proportion of \(4.5/60=0.075\).

The \(X=0\) risk set still has 50 people of each risk type, for 100 total. Its expected event count is \(50(0.20)+50(0.05)=12.5\), so its expected interval event proportion is \(12.5/100=0.125\). The ratio of these illustrative interval proportions is \(0.075/0.125=0.600\), despite the stipulated absence of an exposure effect within either \(U\) group.

This arithmetic is not a Cox coefficient calculation. It demonstrates the selection mechanism: censoring has changed the risk-set mix of an unmeasured event predictor differently by exposure. Given only \(X\), censoring is informative because \(U\) predicts the event and the censoring pattern selects on \(U\). In real continuous-time data the direction and size of Cox bias depend on the timing and pattern of selection; they cannot be read directly from this one-interval ratio.

What to Check in an Analysis

Conditional independent censoring cannot generally be established from observed event and censoring times alone. Diagnostics can reveal implausible patterns and identify covariates that need substantive attention, but a nonsignificant test or similar censoring percentages is not proof of independence.

Describe censoring by modeled covariates. Compare the frequency and timing of loss to follow-up across treatment, severity, site, and other relevant predictors. Differences are not automatically violations, but they show that an unconditional censoring assumption is unsuitable and clarify what must be conditioned on.
Investigate reasons for loss. Administrative end of study may be more plausibly independent than withdrawal prompted by worsening health. Use knowledge of the follow-up process, not only a statistical test.
Check measured predictors of both loss and the event. If a measured factor predicts censoring and the cause-specific hazard, consider whether it belongs in the event model or whether conditioning on the current covariates is otherwise defensible. Adding a variable is not a universal fix: its functional form and role must be handled appropriately.
Check follow-up support. If nearly everyone in a covariate group is censored before event times of interest, the data provide little information about that group's later hazard. Conditional independence does not restore information that was never observed.
Consider unmeasured selection. If loss is plausibly driven by unrecorded prognosis, ordinary Cox fitting alone cannot verify or remove the resulting bias. A sensitivity analysis or a design and analysis strategy tailored to the censoring process may be needed.

Competing events require a separate distinction. For the cause-specific hazard of cause \(k\), a competing event is observed and removes a person from the cause-\(k\) risk set from that time onward. It is not loss to follow-up, and treating it as a removal for this hazard does not require assuming independent latent failure times for different causes. The conditional independent-censoring assumption here concerns censoring that prevents subsequent event status from being observed.

Comparisons and Limits

Situation	Implication for cause-specific Cox analysis
Censoring varies by modeled \(X\), but \(C\perp(T,J)\mid X\)	Different censoring rates by \(X\) are compatible with the assumption; the covariate-specific event model and adequate support are still required.
Censoring depends on an unmodeled factor that predicts future events	Conditioning on modeled \(X\) may not suffice; selective risk-set membership can bias fitted effects.
Follow-up ends before relevant event times in a covariate group	Late hazards for that group may be weakly identified or unsupported, even if censoring is conditionally independent.

The ordinary cause-specific Cox approach relies on conditional independent censoring and the specified event-hazard model. A model for the censoring process or a sensitivity analysis can be considered when the assumption is doubtful, but no statistical adjustment can recover information about an unmeasured predictor without additional assumptions. Methods for the subdistribution hazard use a different risk-set construction and have their own censoring requirements; they are not a direct remedy that can be substituted without changing the estimand. The choice between cause-specific and subdistribution modeling should follow the scientific question, rather than a search for a model that avoids assumptions.

Common Mistakes

Wrong: “Censoring differs by treatment, so the Cox coefficient is necessarily biased.” Correction: Different censoring distributions across treatment groups can be acceptable if censoring is independent of future event experience conditional on the modeled covariates.
Wrong: “The censoring percentages are similar, so the assumption is verified.” Correction: Similar overall percentages can conceal selection by prognosis within groups. The assumption is substantive and conditional; observed summaries cannot prove it.
Wrong: “A competing event is just unobserved censoring.” Correction: A competing event is observed and defines removal from the cause-specific risk set. Loss to follow-up prevents later event status from being observed.
Wrong: “Including every variable associated with censoring guarantees valid effects.” Correction: Variables must adequately capture relevant dependence and be modeled appropriately; unmeasured predictors and inadequate follow-up support remain limitations.

Key takeaway: For cause-specific Cox regression, censoring may depend on modeled predictors. The crucial requirement is that censoring be independent of future event time and cause after conditioning on those predictors, with sufficient follow-up support. If censoring still selects on prognosis, risk-set comparisons and fitted effects may be biased.

Check Your Understanding

Answer each question, then compare your reasoning with the solutions.

If censoring is more frequent for one treatment group but is independent of \((T,J)\) conditional on treatment and modeled severity, is different censoring frequency alone a violation?
In the worked risk-set example, what were the expected interval event proportions for \(X=1\) and \(X=0\), and what caused their difference?
Why does a competing event not require the same interpretation as loss to follow-up in a cause-specific hazard analysis?
Name one diagnostic that can raise concern about conditional independent censoring and explain why it cannot prove the assumption.

Solutions. (1) No. Predictor-related censoring is compatible with conditional independent censoring when the stated conditional independence holds. (2) They were \(4.5/60=0.075\) for \(X=1\) and \(12.5/100=0.125\) for \(X=0\); selective removal of high-risk \(X=1\) people changed the risk-set composition. (3) A competing event is observed and ends cause-\(k\) risk-set membership, whereas loss to follow-up prevents later event status from being observed. (4) For example, compare loss timing by severity and investigate whether losses follow clinical deterioration. Such patterns can reveal plausibility concerns, but observed data cannot rule out unmeasured prognostic factors that also drive censoring.