Right Censoring Likelihoods With Omics Covariates

Why Censoring Changes the Likelihood

In an omics survival analysis, each subject may have a vector of baseline molecular measurements and a follow-up time. If the event occurs during follow-up, its time is observed. If follow-up ends first, the event time is not observed; we know only that it exceeds the time follow-up ended. This is right censoring. Treating a censored follow-up time as though it were an event time would therefore enter information that was never observed into the analysis.

The likelihood represents what the data actually say about event times. An observed event contributes an event-time density (or probability mass), while a censored observation contributes the probability that the event time exceeds the censoring time. The construction relies on assumptions about how event and censoring times relate, including whether that relationship is independent after conditioning on the recorded covariates.

Definition: For subject \(i\), let \(T_i\) be the event time, \(C_i\) the censoring time, \(X_i\) the observed baseline covariate vector (which may include omics measurements), \(Y_i=\min(T_i,C_i)\) the observed time, and \(\delta_i=I(T_i\le C_i)\) the event indicator. Thus \(\delta_i=1\) denotes an observed event, including a tie, and \(\delta_i=0\) denotes censoring before the event.

The event-time distribution conditional on covariates is described by \(F_T(t\mid x)=P(T\le t\mid X=x)\), with density \(f_T(t\mid x)\) when a density exists, and survival function \(S_T(t\mid x)=P(T>t\mid X=x)\). Define \(F_C\), \(f_C\), and \(S_C\) for censoring in the same way. A density has units inverse to time; a survival probability is dimensionless.

Likelihood Contributions and the Censoring Assumption

A central assumption for the standard conditional likelihood is independent censoring given the recorded covariates: \(T\perp C\mid X\). It means that, within subjects with the same included \(X\), knowing the censoring time provides no additional information about the event time. This does not require censoring to be independent of \(X\). For example, follow-up may vary by clinic or treatment group, provided the relevant covariates are included and the conditional independence assumption is reasonable.

If event and censoring times are continuous, ties occur with probability zero. Under conditional independence, the joint observed-data density for a subject with covariates \(x\) is

L_{\mathrm{obs}}(y,\delta\mid x) = \{f_T(y\mid x)S_C(y\mid x)\}^{\delta} \{f_C(y\mid x)S_T(y\mid x)\}^{1-\delta}.

For an event, the subject has event time \(y\) and censoring time at least \(y\), giving \(f_T(y\mid x)S_C(y\mid x)\) in the continuous case. For censoring, censoring occurs at \(y\) while the event time exceeds \(y\), giving \(f_C(y\mid x)S_T(y\mid x)\). Each product has units of inverse time: it contains one density and one dimensionless survival probability.

The strict-versus-inclusive distinction matters if either distribution has atoms. With the convention that ties count as observed events, a measure-based form of the contributions is

\begin{aligned} \delta=1:\quad &dF_T(y\mid x)\,P(C\ge y\mid x),\\ \delta=0:\quad &dF_C(y\mid x)\,P(T>y\mid x). \end{aligned}

These expressions are contributions with respect to the relevant event-time or censoring-time probability measure, not necessarily ordinary densities. An event at \(y\) requires \(C\ge y\), so its factor is inclusive. A censored observation at \(y\) requires \(T>y\), because \(T=y\) is assigned to the event category. In continuous models these distinctions have probability zero and the measure-based expressions reduce to the familiar density formulas. When censoring has an atom at \(y\), replacing \(P(C\ge y\mid x)\) with \(S_C(y\mid x)=P(C>y\mid x)\) can give the wrong event contribution.

Worked Example: Event and Censoring Contributions With Continuous Times

Suppose, hypothetically, that for one covariate profile the event time is exponential with rate \(0.20\) per month and the censoring time is exponential with rate \(0.10\) per month. The two times are conditionally independent, and both distributions are continuous. For an event observed at month 3, the full observed-data density contribution is

f_T(3)S_C(3) =(0.20e^{-0.20(3)})(e^{-0.10(3)}) =0.20e^{-0.9} \approx 0.08131\ \text{per month}.

The event-time part alone is \(f_T(3)=0.20e^{-0.6}\approx0.1098\) per month; the censoring survival factor is \(e^{-0.3}\approx0.7408\), dimensionless. Their product retains units of inverse months.

For a different subject censored at month 4, the full contribution is

f_C(4)S_T(4) =(0.10e^{-0.10(4)})(e^{-0.20(4)}) =0.10e^{-1.2} \approx 0.03012\ \text{per month}.

The event-model part of this censored subject's contribution is \(S_T(4)=e^{-0.8}\approx0.4493\), the probability of surviving beyond month 4. The censoring density supplies the remaining factor in the full observed-data density.

Separating the Event Model From the Censoring Model

For inference about the event-time distribution, the usual right-censoring likelihood retains the event-model factors and treats the censoring mechanism as a nuisance mechanism. With continuous times, the full likelihood for \(n\) independent subjects factors as

L_{\mathrm{obs}} = \underbrace{\prod_{i=1}^n f_T(Y_i\mid X_i)^{\delta_i} S_T(Y_i\mid X_i)^{1-\delta_i}}_{L_T} \underbrace{\prod_{i=1}^n S_C(Y_i\mid X_i)^{\delta_i} f_C(Y_i\mid X_i)^{1-\delta_i}}_{L_C}.

If the event-model parameters and censoring-model parameters are distinct, and the censoring mechanism is ignorable for the target event-model inference, \(L_C\) can be omitted when estimating the event-model parameters. The event-model likelihood is therefore

L_T=\prod_{i=1}^n f_T(Y_i\mid X_i)^{\delta_i} S_T(Y_i\mid X_i)^{1-\delta_i}, \qquad \ell_T=\sum_{i=1}^n \left\{\delta_i\log f_T(Y_i\mid X_i) +(1-\delta_i)\log S_T(Y_i\mid X_i)\right\}.

A censored subject contributes no event-time density at the observed time: the event is known only to occur later, so that subject contributes \(S_T(Y_i\mid X_i)\). An observed event contributes \(f_T(Y_i\mid X_i)\). Omitting the censoring factor is not the same as assuming no censoring; it relies on being able to separate the censoring mechanism from the event-time parameters. If censoring and event processes share parameters or censoring remains informative after conditioning on \(X\), this simplification need not be valid.

Worked Example: An Exponential Event Model With an Omics Covariate

Consider a hypothetical model with one standardized omics covariate \(x\) and event hazard \(\lambda(x)=\lambda_0e^{\beta x}\). For this example, take \(\lambda_0=0.10\) per month and \(\beta=\log 2\). The exponential model has \(f_T(y\mid x)=\lambda(x)e^{-\lambda(x)y}\) and \(S_T(y\mid x)=e^{-\lambda(x)y}\). Assume conditional independent censoring and continuous event and censoring times.

Subject A has \(x=0\), with an event observed at month 2. The rate is \(\lambda(0)=0.10\) per month, so the event-model contribution is

f_T(2\mid 0)=0.10e^{-0.10(2)} \approx 0.08187\ \text{per month}.

Subject B has \(x=1\), with follow-up censored at month 3 and no event observed. The rate is \(\lambda(1)=0.10e^{\log 2}=0.20\) per month. The event-model contribution is a survival probability, not a density:

S_T(3\mid 1)=e^{-0.20(3)}=e^{-0.6} \approx 0.5488.

For these two observations, the event-model likelihood at the specified parameter values is \(0.08187(0.5488)\approx0.04493\) per month. Its log is

\ell_T =\log(0.10)-0.20-0.60 \approx -3.1026.

The contribution has one observed event density and hence units of inverse months. The covariate changes the modeled event rate; it does not turn a censored observation into an observed event.

Derivation and a Simple Likelihood Calculation

The factorization follows directly from the observation rule and conditional independence. For an event at \(y\), the observed information is \(T=y\) and \(C\ge y\). Independence gives the event-time density times the censoring survival probability. For censoring at \(y\), the observed information is \(C=y\) and \(T>y\), giving the censoring density times the event-time survival probability. Dropping censoring terms leaves precisely the event-model contribution for each subject.

A small calculation shows how the event-model likelihood uses censored follow-up. Suppose three hypothetical subjects have event times 2 and 4 months, and one censoring time of 3 months. Assume independent observations and a common exponential event rate \(\lambda\), measured per month. The event contributions are \(\lambda e^{-2\lambda}\) and \(\lambda e^{-4\lambda}\); the censored contribution is \(e^{-3\lambda}\). Thus

L_T(\lambda)=\lambda^2e^{-9\lambda}, \qquad \ell_T(\lambda)=2\log\lambda-9\lambda.

Differentiating gives \(d\ell_T/d\lambda=2/\lambda-9\), so the maximum occurs at \(\hat{\lambda}=2/9\approx0.2222\) per month. It is a maximum because \(d^2\ell_T/d\lambda^2=-2/\lambda^2<0\) for \(\lambda>0\). At this estimate,

L_T(\hat{\lambda}) =\left(\frac{2}{9}\right)^2e^{-2} \approx 0.006683\ \text{per month}^2, \qquad \ell_T(\hat{\lambda})=2\log(2/9)-2\approx-5.008.

The censored subject contributes three months of survival information through \(e^{-3\lambda}\); treating month 3 as an event would incorrectly add a factor \(\lambda\) and alter the total event exposure. This calculation illustrates the likelihood construction, not a general prescription for fitting complex omics models.

Checking the Assumption and Comparing Approaches

Independent censoring conditional on \(X\) is a substantive assumption, not a consequence of right-censoring notation. It is plausible for a fixed administrative study end that is unrelated to prognosis after accounting for study entry and covariates. It may be implausible when sicker subjects are more likely to withdraw, or when follow-up intensity depends on clinical status that is not included in \(X\). Omics predictors alone do not guarantee that relevant clinical causes of both censoring and event time have been controlled.

Useful checks include tabulating censoring frequency and follow-up duration across measured clinical and omics covariates, checking whether censoring patterns differ by site or enrollment period, and investigating documented reasons for loss to follow-up. Strong patterns can flag a questionable assumption or a missing predictor of censoring. Lack of visible patterns does not prove conditional independence: dependence on unmeasured prognosis cannot generally be ruled out from observed data alone. A sensitivity analysis or an explicit censoring model may be needed when the assumption is doubtful.

Approach	What it contributes	Key qualification
Event-model likelihood with right censoring	Uses an event density for events and event survival for censored observations.	Requires an adequate event model and an appropriate independent-censoring assumption.
Full joint likelihood for event and censoring times	Models both processes and retains their observed-data factors.	Requires a specified censoring model; dependence between processes must be handled.
Analysis treating every observed time as an event	Uses all recorded times as if event times were known.	Incorrect for censored subjects because their event times are only known to exceed follow-up.

The event-model likelihood is compact and often appropriate when the censoring process is ignorable and its parameters are distinct. A full joint model can be useful when the censoring process is itself of interest or the mechanisms require joint modeling, but it adds assumptions and parameters. Neither approach repairs a misspecified event-time model. This tutorial concerns ordinary right censoring; delayed entry and interval-censored observations require different likelihood contributions.

Common Mistakes and Practical Limits

Using the observed time as an event time when \(\delta=0\). A censoring time gives the lower bound \(T>Y\), not the equality \(T=Y\). Use \(S_T(Y\mid X)\) in the event-model likelihood.
Using unconditional independence automatically. Censoring can depend on recorded covariates while still being independent of event time conditional on those covariates. Conversely, conditioning on an incomplete covariate set may leave informative censoring.
Ignoring atoms while assigning ties to events. When ties can occur, an event contribution uses \(P(C\ge y\mid x)\), not \(P(C>y\mid x)\). The censoring contribution uses \(P(T>y\mid x)\) under the stated event-on-tie convention.
Misreading the units. A density times a survival probability has units inverse time, not inverse time squared. A censored event-model contribution is a probability and is dimensionless.
Assuming a good fit proves independent censoring. Event-model fit and censoring independence are different questions; the latter may fail because of unmeasured factors.

These likelihoods also assume the event and censoring times are measured on the stated time scale and that the covariates used in the conditional model are observed as specified. If omics measurements are unavailable for some subjects or measured after follow-up begins, the analysis requires additional assumptions beyond the basic right-censoring construction.

Check Your Understanding

Try each question, then compare your reasoning with the solutions below.

For a censored subject followed for 5 months, what event-model factor belongs in the likelihood, and what does it mean?
Under continuous times, an event is observed at \(y\). Write its full observed-data density contribution and identify the units.
If event and censoring times may tie at \(y\), and ties count as events, why may \(S_C(y)\) be the wrong censoring factor for an event?
In the exponential model, what is the event-model contribution for an event at time \(y\) and a censored observation at time \(y\), conditional on \(x\)?

Solutions. (1) The factor is \(S_T(5\mid X)\), the conditional probability that the event time exceeds 5 months. (2) The contribution is \(f_T(y\mid x)S_C(y\mid x)\), with units inverse time because the density has units inverse time and the survival probability is dimensionless. (3) \(S_C(y)=P(C>y)\) excludes a censoring time exactly equal to \(y\), but the event-on-tie convention allows \(C=y\); the required factor is \(P(C\ge y)\). (4) For rate \(\lambda(x)\), the event contribution is \(\lambda(x)e^{-\lambda(x)y}\), while the censored event-model contribution is \(e^{-\lambda(x)y}\). The first is a density and the second is a survival probability.