Censoring Contributions in Survival Power Priors

Why censoring changes the likelihood contribution

Historical survival data often include people whose event time was not observed. A person may experience the event during follow-up, or may leave follow-up event-free, reach the administrative end of the study, or otherwise be observed only up to a time when the event has not yet occurred. That second record is not an observed event time. It tells us that the event time is later than the last observed time, and this information contributes to the likelihood.

In a survival power prior, the historical likelihood is raised to a discount power. The power discounts the information represented by the historical data; it does not change what an event or a censored record means. Thus the first task is to construct the correct likelihood from both types of records. Only then should the discount be applied.

Definition: For historical subject \(i\), let \(T_i\) be the event time, \(C_i\) the right-censoring time, \(Y_i=\min(T_i,C_i)\) the observed time, and \(\delta_i=I(T_i\le C_i)\) the event indicator. Thus \(\delta_i=1\) means the event was observed at \(Y_i\), while \(\delta_i=0\) means the event time is known only to exceed \(Y_i\). Let \(f(t\mid\theta)\) and \(S(t\mid\theta)=P(T_i>t\mid\theta)\) denote the event-time density and survival function under parameter \(\theta\).

The likelihood below assumes that censoring is independent of the event time, conditional on the model information, and that the censoring mechanism has no parameter shared with the event-time model being estimated. This includes censoring times fixed by the study design. Under these conditions, the censoring part of the joint likelihood can be omitted when making inference about \(\theta\). If censoring depends on the event time in a way not accounted for, or its mechanism shares unknown parameters with the event-time model, this simplification need not be valid.

Event and right-censoring contributions

For an event record, \(Y_i\) is an observed event time. Its contribution to the event-time likelihood is the density \(f(Y_i\mid\theta)\). For a right-censored record, the event time is not observed; the information is \(T_i>Y_i\). The corresponding contribution is the probability of surviving beyond that time, \(S(Y_i\mid\theta)\).

The censored contribution follows directly by summing over the unobserved event times that agree with the record. If censoring occurs at time \(y\), every possible event time greater than \(y\) is compatible with the observation. The total likelihood contribution is therefore the integral of the event-time density over that range:

\int_y^\infty f(t\mid\theta)\,dt = P(T_i>y\mid\theta) = S(y\mid\theta).

Consequently, for independent historical records the event-time likelihood is a product of density terms for events and survival terms for right-censored observations. The censoring assumption is what permits this event-time likelihood to be used without also modeling the censoring distribution.

Theorem: Under independent, noninformative right censoring, with a censoring mechanism that can be omitted for inference about \(\theta\), the historical event-time likelihood is

L_H(\theta) = \prod_{i=1}^{n_H} f(Y_i\mid\theta)^{\delta_i} S(Y_i\mid\theta)^{1-\delta_i}.

Here \(n_H\) is the number of historical subjects. The exponent \(\delta_i\) selects the density for an event, and \(1-\delta_i\) selects the survival probability for a censored record. The expression assumes one of those two record types for each subject. It does not treat a censoring time as if it were an observed event time.

Applying the power-prior discount

For a fixed discount \(a\), with \(0\le a\le1\), the historical contribution to the prior is proportional to \(L_H(\theta)^a\), multiplied by the baseline prior density for \(\theta\). For \(a=1\), the historical likelihood is used at full power; for \(a=0\), it contributes the constant 1. Intermediate values discount all historical likelihood contributions through the same power.

L_H(\theta)^a = \prod_{i=1}^{n_H} f(Y_i\mid\theta)^{a\delta_i} S(Y_i\mid\theta)^{a(1-\delta_i)}.

This identity follows from the product rule for powers: \((\prod_i L_i)^a=\prod_i L_i^a\). It makes clear that an event term becomes \(f(Y_i\mid\theta)^a\), while a censored term becomes \(S(Y_i\mid\theta)^a\). In a normalized power prior, the conditional prior for \(\theta\) also includes its normalizing factor; the displayed expression gives the powered historical likelihood that enters that construction. If the discount itself is modeled as random, the normalization must be handled as a function of the discount, rather than dropped as though it were constant in that parameter.

Worked Example: Powering One Event and One Censored Record

Suppose event times follow an exponential distribution with rate \(\lambda=0.1\) per year, so \(f(t\mid\lambda)=\lambda e^{-\lambda t}\) and \(S(t\mid\lambda)=e^{-\lambda t}\). In a hypothetical historical data set, one person has an observed event at 2.5 years and another is right-censored at 6.4 years. Assume independent, noninformative censoring and use fixed discount \(a=0.6\). Calculate their likelihood contributions and the powered joint contribution at this stated rate.

The event record contributes a density, and the censored record contributes a survival probability:

L_{\mathrm{event}}=0.1e^{-0.1(2.5)} =0.1e^{-0.25} \approx 0.07788, \qquad L_{\mathrm{censored}}=e^{-0.1(6.4)} =e^{-0.64} \approx 0.5273.

Applying the discount separately gives

L_{\mathrm{event}}^{0.6} =(0.1e^{-0.25})^{0.6} =0.1^{0.6}e^{-0.15} \approx 0.2162, \qquad L_{\mathrm{censored}}^{0.6} =(e^{-0.64})^{0.6} =e^{-0.384} \approx 0.6811.

Their product is approximately \(0.2162(0.6811)=0.1473\). Directly powering the joint likelihood gives the same result:

\left[(0.1e^{-0.25})(e^{-0.64})\right]^{0.6} =(0.1e^{-0.89})^{0.6} \approx 0.1473.

These are likelihood values at a stated rate, not probabilities that the rate takes a particular value. In particular, the event term is a density, with units determined by the time scale. The arithmetic is consistent up to rounding.

Exponential model and a conjugate update

The exponential model is useful for showing how censoring affects the information about a survival rate. Its density and survival function are \(f(t\mid\lambda)=\lambda e^{-\lambda t}\) and \(S(t\mid\lambda)=e^{-\lambda t}\), for \(\lambda>0\). For \(D=\sum_i\delta_i\) observed events and total observed time \(R=\sum_iY_i\), multiplying the event and censored terms yields

L_H(\lambda) =\lambda^D e^{-\lambda R}, \qquad L_H(\lambda)^a =\lambda^{aD}e^{-a\lambda R}.

Every subject contributes observed time at risk through \(R\), whether or not an event occurred. Only observed events contribute factors of \(\lambda\). This distinction is central to the update.

Worked Example: Gamma Updating With Censored Historical Follow-Up

In a hypothetical historical sample, two subjects experience events at 2 and 5 years, while two others are right-censored at 3 and 7 years. Assume exponential event times, independent noninformative censoring, and a fixed discount \(a=0.5\). Let the baseline prior be \(\lambda\sim\operatorname{Gamma}(2,4)\), using shape and rate parameters. Find the conditional posterior distribution given this historical contribution.

There are \(D=2\) events, and the total observed time is \(R=2+5+3+7=17\) years. Under the exponential model, the historical likelihood is \(\lambda^2e^{-17\lambda}\), and its powered form is

L_H(\lambda)^{0.5} =(\lambda^2e^{-17\lambda})^{0.5} =\lambda e^{-8.5\lambda}.

A Gamma prior with shape 2 and rate 4 has density proportional to \(\lambda^{2-1}e^{-4\lambda}\). Multiplying by the powered likelihood gives a density proportional to

\lambda^{2-1}e^{-4\lambda}\, \lambda e^{-8.5\lambda} =\lambda^{2}e^{-12.5\lambda}.

This is the kernel of a Gamma distribution with shape \(3\) and rate \(12.5\). Thus the conditional posterior is \(\operatorname{Gamma}(3,12.5)\), with posterior mean \(3/12.5=0.2400\) events per year. Censored subjects contribute to the total follow-up time \(R\), but they do not increase the event count \(D\).

Why censoring assumptions and diagnostics matter

The likelihood construction is only as credible as its assumptions and the recorded follow-up information. Before borrowing historical survival data, check that event indicators, time origins, units, and censoring definitions are consistent with the event-time model. A mislabeled event indicator can turn survival contributions into density contributions; a different time scale changes numerical rates; and follow-up cut short by a process related to unobserved event risk may be informative censoring.

The independent-censoring assumption is not established merely because the data contain censoring. Investigators should consider why follow-up ended and whether those reasons plausibly carry information about the event time after accounting for the model. Administrative end-of-study censoring is often more defensible than censoring caused by worsening health or loss to follow-up associated with prognosis. If informative censoring is plausible, an event-time likelihood that omits the censoring mechanism may be inadequate; a joint model or an explicit model for the censoring process may be needed.

A further check is to compare the historical data definition with the intended endpoint and follow-up window. The survival likelihood assumes the recorded time is measured from a meaningful and comparable origin, and that the status at that time is correctly classified. It also assumes the event-time model is suitable. For example, the exponential model imposes a constant event hazard; a good algebraic update does not make that assumption appropriate for every clinical endpoint.

With a fixed \(a\), the powered likelihood discounts the recorded historical likelihood without repairing a biased or incompatible source. It cannot correct informative censoring, differences in event definitions, or a misspecified event-time model. In practice, the discount and the adequacy of the model are separate concerns: reducing historical influence does not guarantee that the remaining information is valid.

Common mistake and alternatives

A frequent error is to count a censored record as an event at its last observed time. That substitutes \(f(Y_i\mid\theta)\) for \(S(Y_i\mid\theta)\), adding an event that was never observed and changing the likelihood’s dependence on \(\theta\). The opposite error—dropping every censored record—throws away the information that the person remained event-free through \(Y_i\). The correct contribution, under the stated censoring assumptions, is \(S(Y_i\mid\theta)\), and it too is raised to the historical discount power.

The main alternative to omitting the censoring mechanism is to model it jointly, particularly when censoring may be informative or its parameters are related to the event-time parameters. A complete-case analysis that excludes censored records is generally not equivalent to the correct survival likelihood and can discard useful follow-up information. Treating censoring as an event is not a valid alternative: it misrepresents the observed data. The simple powered likelihood is attractive when its assumptions are credible, but it does not remove the need to examine data provenance, censoring, and model fit.

Classify each record.
Use the event indicator to distinguish an observed event from a right-censored time; do not infer event status from the time alone.

Write the unpowered likelihood.
Use \(f(Y_i\mid\theta)\) for an event and \(S(Y_i\mid\theta)\) for a censored record, under independent noninformative censoring.

Apply the historical discount.
Raise the full historical likelihood to \(a\), equivalently raising each subject’s correct contribution to \(a\).

Check the assumptions and arithmetic.
Confirm compatible event definitions and time scales, assess the censoring mechanism, and verify the powered product against the powered joint likelihood.

Key takeaway: Under independent, noninformative right censoring, an observed event contributes its event-time density and a right-censored record contributes its survival probability. A survival power prior discounts the product of these correctly classified contributions; it does not turn censoring into an event.

Check Your Understanding

Answer each question, then compare with the solutions below.

For an event at time \(y\), what is the event-time likelihood contribution? For a right-censored record at time \(y\), what is the contribution?
A hypothetical exponential-model data set has one event at 4 years and one censoring time at 6 years. Write the historical likelihood in terms of \(\lambda\), then write its powered form for \(a=0.5\).
For the data in question 2 and prior \(\lambda\sim\operatorname{Gamma}(1,2)\), what is the conditional posterior distribution?
Why is treating a right-censored time as an event incorrect? State one condition under which the simple event-time likelihood may also be inadequate.

Solutions. (1) An event contributes \(f(y\mid\theta)\); a right-censored record contributes \(S(y\mid\theta)\), under the stated censoring assumption. (2) The event term is \(\lambda e^{-4\lambda}\), and the censored term is \(e^{-6\lambda}\). Hence \(L_H(\lambda)=\lambda e^{-10\lambda}\), and \(L_H(\lambda)^{0.5}=\lambda^{0.5}e^{-5\lambda}\). (3) The prior kernel is \(\lambda^{1-1}e^{-2\lambda}\). Multiplying by the powered likelihood gives \(\lambda^{0.5}e^{-7\lambda}\), so the posterior is \(\operatorname{Gamma}(1.5,7)\), with shape-rate parameterization. (4) The event was not observed; the record establishes only that the event time exceeds the censoring time. Treating it as an event incorrectly inserts a density factor. The simple likelihood may be inadequate if censoring is informative or shares unknown parameters with the event-time model.