Posterior Survival Probabilities Under Heavy Censoring

Why Heavy Censoring Matters for Survival Estimates

A survival estimate at an early follow-up time may be supported by many participants, while an estimate at a later time may depend on only a handful. This is common in clinical trials: participants enter at different times, follow-up ends, or observation is otherwise censored before the endpoint of interest. A late survival estimate can still be calculated under a specified model, but its uncertainty should reflect how little information remains.

Here the focus is survival at a fixed time represented by several discrete intervals. The model uses an event hazard for each interval and a beta prior for each hazard. Its main lesson is not that heavy censoring makes estimation impossible in every case, but that the posterior for late survival can be wide, sensitive to prior assumptions, or unsupported by any late risk-set observations at all.

Model and Assumptions

Suppose \(r_j\) participants are observed to be at risk at the start of interval \(j\), \(d_j\) have an event during that interval, and \(r_j-d_j\) are known to complete it event-free. The likelihood contribution for the interval hazard is proportional to \(h_j^{d_j}(1-h_j)^{r_j-d_j}\). In this tutorial, censoring is assumed to occur at interval boundaries, so each person counted as event-free is known to have completed the interval without an event. If a participant is censored partway through an interval, that participant cannot automatically be counted as a non-event for the whole interval; the observation model must account for the partial follow-up.

Assume independent priors \(h_j\sim\operatorname{Beta}(a_j,b_j)\), where both shape parameters are positive. The interval outcomes are conditionally independent given their hazards, and censoring is non-informative for the event process under the model: censoring does not select participants according to an unaccounted-for risk of the event. These assumptions matter. A beta update does not repair informative dropout, an incorrect interval definition, or a hazard model that poorly represents the trial.

Updating the Hazards and Combining Survival

The beta distribution is conjugate to the binomial likelihood. Given \(d_j\) events and \(r_j-d_j\) observed event-free intervals, the posterior is

Because \(q_j=1-h_j\), its posterior distribution is \(\operatorname{Beta}(b_j+r_j-d_j,\ a_j+d_j)\). If the prior distributions and likelihood factor across intervals, the posterior distributions of the hazards (and therefore the \(q_j\)) are independent across intervals. This independence makes the mean and second moment of \(S_J\) straightforward to calculate.

For \(q_j\sim\operatorname{Beta}(A_j,B_j)\), the beta moments are \(E(q_j)=A_j/(A_j+B_j)\) and \(E(q_j^2)=A_j(A_j+1)/[(A_j+B_j)(A_j+B_j+1)]\). Consequently,

These are posterior summaries under the stated model. The posterior mean is not a guarantee that the true survival probability equals that value, and it does not describe the uncertainty by itself. A credible interval or a posterior probability such as \(P(S_J>c\mid\text{data})\) is also useful when the clinical decision depends on a threshold \(c\).

Why a Small Risk Set Produces More Uncertainty

The late interval’s posterior is informed by its own events and event-free observations, not by the original enrollment count alone. With a uniform prior, the six interval-3 observations in the example produce \(h_3\mid\text{data}\sim\operatorname{Beta}(2,6)\). The posterior mean hazard is \(2/(2+6)=0.25\), and the posterior mean interval survival is 0.75. That point summary should not obscure that the data are sparse.

For comparison, suppose the same interval data—one event and five event-free observations—were analyzed using a \(\operatorname{Beta}(2,2)\) prior for the hazard. The posterior would be \(\operatorname{Beta}(3,7)\), giving mean hazard \(3/10=0.30\) and mean interval survival \(7/10=0.70\). The data are unchanged; the prior contributes relatively more when the risk set is small. This is a sensitivity check, not a reason to select a prior after seeing which answer is preferred.

The uncertainty can also be seen in the variance. Under the uniform prior, \(q_3\mid\text{data}\sim\operatorname{Beta}(6,2)\), so

A late risk set with more observations would generally give more likelihood information, although the exact posterior uncertainty also depends on the observed event pattern and prior. Heavy censoring therefore calls for a careful account of who remains observed and how much the posterior depends on assumptions.

When the Late Hazard Has No Data

The most severe case occurs when nobody enters a later interval. Then \(r_j=0\), and the likelihood contribution for that interval is constant with respect to \(h_j\): it contains no event or event-free observations. The posterior for that hazard is its prior. A numerical late survival estimate can still be produced by combining it with earlier intervals, but the data have not updated the late hazard.

Diagnostics, Alternatives, and Limitations

Before interpreting a posterior survival probability under heavy censoring, check the data summaries and the model specification. Useful checks include:

• Report the number at risk at the start of each interval, alongside the events and completed event-free intervals. A large original sample does not imply a large late risk set.
• Verify that each participant’s interval contribution is justified. In particular, do not count someone censored partway through an interval as known to have survived the entire interval.
• Check whether censoring could depend on prognosis after accounting for the model. If participants at higher event risk are more likely to leave observation, a non-informative censoring assumption may fail.
• Assess whether the interval definitions and a constant hazard within each interval are defensible for the question. Different interval choices can alter the model and the inferred survival.
• Examine posterior uncertainty and sensitivity to plausible priors, especially in intervals with few observations or none.

A non-Bayesian Kaplan–Meier estimate is a standard alternative for right-censored survival data and summarizes the observed event-time pattern without specifying beta priors for discrete interval hazards. It has its own sampling uncertainty and can also become imprecise at late times when few participants remain at risk. A parametric continuous-time model, such as an exponential model, instead imposes a form on event times and can use the timing of events and censoring. These approaches answer related questions under different assumptions; they do not make sparse late follow-up informative by itself.

The discrete-time beta-hazard calculation is transparent and gives a direct posterior distribution under its assumptions. Its limitations include dependence on interval choices, the assumption of a common hazard within each interval, and possible sensitivity to priors when risk sets are small. In particular, the posterior mean product is not the same as multiplying posterior hazard means: survival uses \(1-h_j\), and uncertainty in the product matters. Likewise, multiplying the endpoints of separate interval credible intervals does not generally produce a credible interval for the product. To obtain a credible interval or \(P(S_J>c\mid\text{data})\), use the posterior distribution of the product, for example by drawing independent \(q_j\) values from their posterior distributions, multiplying them for each draw, and summarizing those product draws.