Conditional Independent Censoring in Joint Models

Why Censoring Requires an Assumption

In a joint model, the event-process likelihood uses each subject’s observed event or censoring time. A censored subject contributes information that the event had not occurred by the censoring time, rather than an observed event time. That contribution is interpretable only under assumptions about how censoring arose. If subjects with a high unobserved event risk tend to leave observation earlier, censoring changes which event histories are seen and can distort inference.

The central distinction is between censoring that is independent of the event time after conditioning on modeled information, and censoring that remains related to the event process. In a shared random-effects model, conditioning on random effects can remove dependence that would otherwise remain. But because those effects are latent, conditional independence alone does not always make the censoring mechanism ignorable in the marginal likelihood.

Definition: For subject $i$, let $T_i$ be the event time, $C_i$ the censoring time, $X_i$ the modeled covariates, and $b_i$ the random effects. The observed time is $Y_i=\min(T_i,C_i)$, and the event indicator is $\Delta_i=I(T_i\le C_i)$. Conditionally independent censoring means $$ T_i \mathrel{\perp\!\!\!\perp} C_i \mid X_i,b_i. $$ Equivalently, conditional on the modeled covariates and random effects, knowing the censoring time provides no additional information about the event time.

Write $f_T(t\mid X_i,b_i)$ and $S_T(t\mid X_i,b_i)$ for the event-time density and survival function. Similarly, let $f_C$ and $S_C$ denote the censoring-time density and survival function, with the same conditioning variables when needed. For continuous times, the conditional-independence assumption gives the observed-data density

L_i^{\mathrm{full}}= \left\{f_T(Y_i\mid X_i,b_i)^{\Delta_i} S_T(Y_i\mid X_i,b_i)^{1-\Delta_i}\right\} \left\{S_C(Y_i\mid X_i,b_i)^{\Delta_i} f_C(Y_i\mid X_i,b_i)^{1-\Delta_i}\right\}.

If an event is observed at $Y_i$, its density is multiplied by the probability of remaining uncensored through that time. If censoring occurs at $Y_i$, the censoring density is multiplied by the probability that the event has not yet occurred. The first brace is the event-process contribution,

L_{Ti}=f_T(Y_i\mid X_i,b_i)^{\Delta_i} S_T(Y_i\mid X_i,b_i)^{1-\Delta_i}.

This is the familiar contribution used in the event submodel: an observed event contributes a density, and a right-censored observation contributes survival through the censoring time. Whether the second brace can be omitted from a joint-model likelihood requires further care.

Deriving the Conditional Factorization

The factorization follows directly by considering the two possible observation outcomes. Assume $T_i$ and $C_i$ have continuous distributions conditional on $X_i,b_i$, so a tie has probability zero. If $\Delta_i=1$, then $T_i=Y_i$ and $C_i\ge Y_i$. Conditional independence gives the joint density $f_T(Y_i\mid X_i,b_i)S_C(Y_i\mid X_i,b_i)$. If $\Delta_i=0$, then $C_i=Y_i$ and $T_i>Y_i$, giving $f_C(Y_i\mid X_i,b_i)S_T(Y_i\mid X_i,b_i)$. Combining the cases yields the displayed full likelihood.

Thus the event-process factor is valid conditional on $X_i,b_i$ under conditional independent censoring. It does not follow that the censoring factor is always irrelevant after random effects are integrated out. In a joint model the subject likelihood typically integrates over $b_i$; a censoring contribution that depends on $b_i$ changes how the integral weights possible random effects.

L_i^{\mathrm{marginal}} = \int L_{i,\mathrm{long}}(b_i)\, L_{Ti}(b_i)\, L_{Ci}(b_i)\, p(b_i)\,db_i.

Here $L_{i,\mathrm{long}}$ denotes the longitudinal-data contribution, $p(b_i)$ the random-effects density, and $L_{Ci}$ the censoring-process factor in the full likelihood. If censoring depends only on observed covariates and its parameters are distinct from the event and longitudinal parameters, $L_{Ci}$ can often be separated from the integral and omitted for inference on those model parameters. If censoring depends on $b_i$, it generally cannot be pulled outside the integral. Omitting it can change the implied distribution of $b_i$ among subjects with the observed censoring patterns.

Key distinction: Conditional independence $T_i\mathrel{\perp\!\!\!\perp}C_i\mid X_i,b_i$ justifies factorizing the conditional event and censoring contributions. Ignoring the censoring contribution in a marginal joint likelihood additionally requires that the censoring mechanism be ignorable for the parameters and latent effects being inferred. Censoring independent of $b_i$, conditional on observed covariates, is a common sufficient condition when censoring parameters are distinct.

Worked Likelihood Calculations

Worked Example: An Administrative Censoring Time

Suppose a hypothetical subject has a fixed study end at 6 years, set by the study design independently of the subject’s event risk. The subject has no event by 6 years, so $Y_i=6$ and $\Delta_i=0$. Conditional on the subject’s covariates and random effects, suppose the event hazard is constant at $0.10$ per year.

For a constant hazard $\lambda=0.10$, the event survival function is $S_T(t)=\exp(-\lambda t)$. Thus the event-process likelihood contribution is

L_{Ti}=S_T(6)=\exp(-0.10\times 6)=\exp(-0.60)=0.5488.

The value $0.5488$ is the model probability of remaining event-free through 6 years for this subject’s specified covariates and random effects. Because the censoring time is fixed by design rather than selected according to the subject’s event risk, it is reasonable to condition on that study end and use this event contribution. The calculation does not assert that censoring is harmless in every setting; its justification depends on how the end time was determined.

Worked Example: Conditional Independence Can Fail Marginally

Consider a hypothetical population with two equally common random-effect groups, $b_i=H$ and $b_i=L$. Let $A$ indicate an event by year 5 and $D$ indicate censoring before year 5. Suppose, conditional on $b_i$, $A$ and $D$ are independent. Their conditional probabilities are $P(A=1\mid H)=0.8$, $P(D=1\mid H)=0.8$, $P(A=1\mid L)=0.2$, and $P(D=1\mid L)=0.2$.

The marginal probabilities are $P(A=1)=0.5(0.8)+0.5(0.2)=0.5$ and $P(D=1)=0.5$. But the joint probability is

P(A=1,D=1)=0.5(0.8\times0.8)+0.5(0.2\times0.2) =0.32+0.02=0.34.

If the two indicators were marginally independent, this probability would be $0.5\times0.5=0.25$, not $0.34$. Their association arises because both are related to the same random effect, even though they are independent conditional on it. Moreover, observing early censoring changes the distribution of the random effect: $P(H\mid D=1)=0.8$, since $P(H,D=1)=0.5(0.8)=0.4$ and $P(D=1)=0.5$. Therefore, when censoring depends on latent effects, its contribution can matter to the marginal likelihood.

Worked Example: Comparing Event Contributions for Two Subjects

Suppose a hypothetical model has a constant conditional event hazard of $0.08$ per year for one subject and $0.12$ per year for another, given their covariates and random effects. The first subject has an observed event at 3 years. The second is censored at 3 years. For an exponential event model, $f_T(t)=\lambda\exp(-\lambda t)$ and $S_T(t)=\exp(-\lambda t)$.

For the subject with the event, the contribution is the event density:

L_{T1}=0.08\exp(-0.08\times3) =0.08\exp(-0.24) =0.06293\ \text{per year}.

For the censored subject, the contribution is survival through 3 years:

L_{T2}=\exp(-0.12\times3) =\exp(-0.36) =0.6977.

The density and survival contributions have different units and meanings: the first is a density at an observed event time, while the second is a probability of surviving beyond the censoring time. These are conditional event-process contributions, not by themselves complete marginal likelihoods when random effects must still be integrated.

Assessing the Assumption and Choosing an Approach

Conditional independent censoring is primarily a statement about the data-generating process, not a property that can be established by a single test of the observed data. It is more credible when censoring is administrative or driven by recorded variables included in the model. It is less credible when withdrawal, loss to follow-up, or treatment discontinuation is driven by unrecorded health deterioration related to the event risk.

Useful checks combine study knowledge with examination of observed patterns. Compare censoring rates across recorded covariate groups and across observed longitudinal histories; investigate whether reasons and timing of dropout suggest worsening health; and ask whether the model includes plausible shared latent features that could drive both event risk and censoring. These checks can reveal problems, but observed data alone generally cannot verify independence conditional on an unobserved $b_i$.

The main approaches differ in what they assume and model:

Approach	What it handles	Main caution
Condition on a noninformative censoring schedule	Administrative censoring or censoring independent of modeled latent effects	Requires a defensible design or mechanism argument
Model the censoring process	Censoring related to covariates or latent effects represented in the model	Requires a credible censoring model and appropriate joint integration
Inverse-probability-of-censoring weighting	Observed-covariate-dependent censoring under specified positivity and model conditions	Weights can be unstable; unobserved drivers remain a concern
Sensitivity analysis	Uncertainty about unobserved causes of dropout	Conclusions depend on the range of departures examined

A diagnostic warning is a pattern in which dropout becomes more frequent after worsening measured marker values or within groups expected to have greater event risk. Such a pattern does not prove informative censoring, but it should prompt scrutiny of the assumed conditioning set. A particularly serious failure occurs when an unmeasured time-varying health change affects both dropout and the event, while the modeled random effects do not adequately represent that relationship.

Common Mistake and Limitations

Wrong reasoning: “Event time and censoring time are independent given the random effects, so I can always omit the censoring model.” Correction: That independence gives a conditional factorization. If censoring also depends on latent random effects, the censoring factor can alter their distribution in the marginal likelihood and may need to be included.
Wrong reasoning: “Censoring is independent because it is not significantly associated with observed covariates.” Correction: Lack of an observed association does not establish independence from latent effects or unmeasured health changes. The assumption needs substantive justification.
Wrong reasoning: “A censored subject contributes no event information.” Correction: Under the stated event-process model, a subject censored at $c$ contributes $S_T(c\mid X_i,b_i)$, the probability of remaining event-free through $c$.

No statistical adjustment can recover information that is unsupported by assumptions or by the observed data. A censoring model can reduce bias when it captures the relevant mechanism, but its results depend on that model. When plausible censoring mechanisms remain uncertain, report sensitivity analyses rather than presenting one assumption as verified.

Key takeaway: State what variables condition the censoring assumption. Conditional independence given covariates and random effects supports the event-process factor, but latent-effect-dependent censoring can still influence marginal inference and may require an explicit censoring contribution.

Check Your Understanding

Answer each question, then compare with the solutions.

A subject is censored at 4 years. What is the event-process likelihood contribution, and what does it mean?
Why can event and censoring indicators be associated marginally even if they are independent given a random effect?
State one condition under which a censoring contribution can be omitted from the marginal joint likelihood for event-model inference.
A study finds that dropout tends to follow worsening observed marker values. Does this prove informative censoring? What should be examined?

Solutions. (1) The contribution is $S_T(4\mid X_i,b_i)$, the conditional probability that the event time exceeds 4 years. (2) Both indicators can depend on the same random effect; averaging over that effect induces marginal association. (3) For example, censoring may depend only on observed covariates, with censoring parameters distinct from the event and longitudinal parameters, so its factor separates from the random-effects integral. (4) No. The pattern is a warning, not proof; investigators should examine dropout reasons, covariate and longitudinal histories, and whether plausible shared latent drivers are represented in the model.