The Likelihood for Right-Censored Trial Survival Data

Why Censoring Changes the Likelihood

In a trial with a time-to-event endpoint, some participants experience the event while others leave follow-up without an event or reach the study’s administrative end. For the latter participants, the event time is not observed. Their follow-up still provides information: if a participant is event-free through a recorded time, the event time must exceed that time. The likelihood must represent that information without pretending to know when the event would later occur.

The key distinction is between an observed event time and a last-known event-free time. An exact event time contributes a probability density. A right-censored time contributes the probability of remaining event-free beyond that time. These are different quantities, and substituting one for the other changes the likelihood.

Definition: Let \(T\) be a participant’s event time and \(C\) the time at which observation ends. The recorded follow-up time is \(Y=\min(T,C)\), and the event indicator is \(\Delta=I(T\leq C)\), where \(I(\cdot)\) is 1 when its condition is true and 0 otherwise. Thus \(\Delta=1\) denotes an observed event at \(Y=T\); \(\Delta=0\) denotes right censoring at \(Y=C\), which tells us \(T>C\). For event-time parameter \(\theta\), write \(f(t\mid\theta)\) for the event-time density and \(S(t\mid\theta)=P(T>t\mid\theta)\) for the survival function.

This setup assumes a defined event, time origin, and time scale. For example, a trial might measure months from randomization to first hospitalization. The recorded time and the density must use the same units. This tutorial concerns right censoring: an event has either been observed at a time or is known not to have happened by a time. It does not cover interval-censored event times, for which the event is known only to have occurred within an interval.

Deriving the Participant Contribution

Suppose first that a participant has an observed event at time \(y\). For a continuous event time, the probability of an event at exactly one specified point is zero; the density describes event-time likelihood per unit of time. If the event is censored at time \(y\), the available observation is instead \(T>y\), whose probability is the survival function.

L_i(\theta)=f(y_i\mid\theta)^{\Delta_i}S(y_i\mid\theta)^{1-\Delta_i}

When \(\Delta_i=1\), the expression is \(f(y_i\mid\theta)\). When \(\Delta_i=0\), it is \(S(y_i\mid\theta)\). This compact expression is the event-time likelihood contribution for participant \(i\). The density has units inverse to the time unit, while the survival probability is dimensionless.

The justification relies on how the censoring process relates to the event time. A standard sufficient condition is that \(T\) and \(C\) are independent, conditional on any observed covariates or trial-arm information included in the analysis, and that the censoring mechanism has no parameters shared with the event-time model. This is commonly called noninformative censoring for the event-time likelihood. Under it, the observed event-time part of the likelihood has the form above; factors describing censoring do not depend on \(\theta\) and can be omitted when making inference about \(\theta\).

\begin{aligned} \Delta_i=1:\quad p(y_i,\Delta_i=1\mid\theta) &=f(y_i\mid\theta)P(C_i\geq y_i),\\ \Delta_i=0:\quad p(y_i,\Delta_i=0\mid\theta) &=g_C(y_i)S(y_i\mid\theta), \end{aligned}

Here \(g_C\) denotes the censoring-time density, where it is appropriate to describe censoring continuously. In the event case, the factor \(P(C_i\geq y_i)\) expresses that observation continued long enough for the event to be seen. In the censored case, \(g_C(y_i)\) describes when follow-up ended, while \(S(y_i\mid\theta)\) expresses that the participant remained event-free until then. If the censoring factors are free of \(\theta\), they do not change the event-time likelihood as a function of \(\theta\). The same reasoning applies when censoring is administrative or otherwise has a discrete or mixed distribution: its factor is omitted only when it does not involve the event-time parameter of interest.

Assumptions for this likelihood: The event and censoring times are recorded on a consistent time scale; the event status at each recorded time is correctly classified; and censoring is independent of the event time conditional on the observed information used in the model. The participants’ event-time observations are independent conditional on the model parameters, or their dependence is represented in the model. If censoring depends on prognosis, the conditioning information or a joint model must address that dependence.

From Participant Contributions to a Trial Likelihood

For \(n\) conditionally independent participants with a common event-time parameter \(\theta\), multiply their contributions. If trial arms have different event-time parameters, use the appropriate parameter for each participant; multiplying contributions does not imply that the arms share a parameter.

L(\theta)=\prod_{i=1}^{n} f(y_i\mid\theta)^{\Delta_i} S(y_i\mid\theta)^{1-\Delta_i}

A useful derivation follows directly by splitting the participants into those with observed events and those censored. Let \(D\) be the set of event participants and \(R\) the set of right-censored participants. Then the product can be regrouped as

L(\theta)= \left\{\prod_{i\in D} f(y_i\mid\theta)\right\} \left\{\prod_{i\in R} S(y_i\mid\theta)\right\}.

Each observed event contributes one density factor; each censored participant contributes one survival factor. This regrouping is not a different method. It makes clear exactly what information each participant supplies. If a parameter is shared across trial arms, all participants may contribute to its likelihood; if parameters are arm-specific, each factor uses its participant’s assigned arm parameter.

Worked Example: One Event and One Censored Participant

Consider a hypothetical trial participant who experiences an event at 2 months and another who is event-free when censored at 3 months. Suppose two candidate event-time parameter values, \(\theta_A\) and \(\theta_B\), give the following values from their respective event-time models: \(f(2\mid\theta_A)=0.12\) per month, \(S(3\mid\theta_A)=0.55\), \(f(2\mid\theta_B)=0.18\) per month, and \(S(3\mid\theta_B)=0.36\). The censoring is assumed noninformative for this comparison.

The event participant contributes the density at 2 months; the censored participant contributes survival beyond 3 months. Thus,

L(\theta_A)=f(2\mid\theta_A)S(3\mid\theta_A) =(0.12)(0.55)=0.066,

and

L(\theta_B)=f(2\mid\theta_B)S(3\mid\theta_B) =(0.18)(0.36)=0.0648.

Each likelihood value has units per month because there is one observed event density and one dimensionless survival probability. For these two participants, the likelihood at \(\theta_A\) is about \(0.066/0.0648=1.019\) times the likelihood at \(\theta_B\). This is a likelihood comparison, not a posterior probability or a probability that either parameter value is true.

Worked Example: Assembling a Small Trial Likelihood

Suppose a hypothetical trial has four participants in one arm. Two have events at 1 and 3 weeks, and two are censored event-free at 2 and 4 weeks. At a candidate arm parameter \(\theta\), suppose the model gives \(f(1\mid\theta)=0.10\) per week, \(S(2\mid\theta)=0.75\), \(f(3\mid\theta)=0.08\) per week, and \(S(4\mid\theta)=0.50\). Assume independent participants and noninformative censoring.

The event indicators are 1, 0, 1, 0, in the order listed. Therefore the likelihood is the product of two densities and two survival probabilities:

\begin{aligned} L(\theta) &=f(1\mid\theta)S(2\mid\theta)f(3\mid\theta)S(4\mid\theta)\\ &=(0.10)(0.75)(0.08)(0.50)\\ &=0.0030. \end{aligned}

The units are per week squared, because there are two event-time density factors. If the participants were in two trial arms with separate parameters, each factor would use the parameter for that participant’s assigned arm. The recorded times alone do not justify pooling arms under one parameter.

Checking the Likelihood and Its Assumptions

A practical check is to trace each participant from the data to exactly one likelihood factor. An observed event must use its event-time density; a right-censored record must use survival through the censoring time. The following process helps prevent coding and notation errors.

Define the endpoint and clock.
Specify what counts as the event, the time origin, and the units. Confirm that all recorded times use that scale.

Classify each record.
Set \(\Delta_i=1\) only for an observed event, and set \(\Delta_i=0\) when follow-up ends with the participant event-free.

Apply the matching factor.
Use \(f(y_i\mid\theta)\) for an event and \(S(y_i\mid\theta)\) for censoring, with the relevant arm or covariate values.

Check independence and censoring.
Assess whether censoring could depend on the participant’s unobserved event time after conditioning on the information in the model.

Potential warning signs include censoring related to worsening symptoms, withdrawal because of treatment failure, or loss to follow-up that is associated with unmeasured prognosis. In such cases, censoring may be informative: the participants who remain under observation need not represent those who leave. Simply using \(S(y_i\mid\theta)\) while ignoring the dependence can give misleading event-time inference. Possible responses include conditioning on measured predictors, modelling event and censoring jointly, or using a method designed for the particular dependence; the appropriate choice depends on the study design and estimand.

Other checks concern the data itself. Confirm that event and censoring indicators are not reversed, that censoring times are not accidentally recorded as events, and that times are nonnegative and consistently converted. If event and censoring times can tie, the protocol’s event-status convention must determine which status is recorded. For continuous event times, exact ties are usually unusual, but rounding or discrete recording can make them common. Competing events also require care: a different event type is not automatically ordinary independent censoring, because it may prevent the event of interest from occurring.

Common Mistakes and Alternatives

The most common error is to treat a censoring time as if it were the participant’s event time. A participant censored at 4 months has not necessarily had an event at 4 months; the observation says only that the participant was event-free through 4 months. The correction is to use \(S(4\mid\theta)\), not \(f(4\mid\theta)\).

A second error is to omit censored participants altogether. They contribute no event density, but they do contribute survival information. Dropping them discards the fact that their event time exceeded their observed follow-up. A third error is to interpret a density as a probability. For a continuous time, \(f(y\mid\theta)\) is a density, not the probability of an event at precisely \(y\); it is the appropriate likelihood contribution for an exact event time.

The likelihood above is the standard event-time likelihood for right-censored data under noninformative censoring. An alternative is to model the censoring process jointly with the event process. That may be necessary when censoring is informative or when parameters are shared, but it requires assumptions about the censoring mechanism as well as the event-time distribution. A simpler alternative is available only if the design or scientific question genuinely makes censoring irrelevant; discarding censored observations is not a valid substitute.

This likelihood specifies how the observed data depend on event-time parameters; it does not itself specify a prior or produce a posterior distribution. In Bayesian analysis, the likelihood is combined with a prior for the parameters of the chosen event-time model. The likelihood contribution for censoring remains a survival probability regardless of which prior is later used.

Key takeaway: Under noninformative right censoring, an observed event at time \(y\) contributes \(f(y\mid\theta)\), and an event-free participant censored at \(y\) contributes \(S(y\mid\theta)\). Multiply the appropriate contributions across participants, while checking that the censoring assumptions and time definitions fit the trial.

Check Your Understanding

Answer each question, then compare with the worked solutions below.

A participant has an observed event at 5 days. Which event-time likelihood factor applies, and what are its units if time is measured in days?
A participant is censored event-free at 6 days. Write the likelihood contribution and explain what observation it represents.
In a hypothetical two-participant dataset, one event at time 2 has density \(0.20\) per week and one participant censored at time 3 has survival probability \(0.40\). Calculate the joint event-time likelihood contribution.
Why can censoring related to worsening health invalidate the simple event-time likelihood?

Solutions. (1) The contribution is \(f(5\mid\theta)\); it has units per day because it is a density with respect to time. (2) The contribution is \(S(6\mid\theta)\), the probability that the event time exceeds 6 days. (3) Multiply the event density by the censoring survival probability: \((0.20)(0.40)=0.080\) per week. (4) If worsening health affects both the chance of leaving follow-up and the event time, censoring may depend on event risk even after conditioning on the modelled information. The censoring mechanism is then informative, so omitting it can distort inference about the event-time parameters.

The Likelihood for Right-Censored Trial Survival Data

What You'll Learn