Posterior Frailty Prediction for Observed Clusters

Why Predict Frailty for an Observed Cluster?

A shared frailty represents unobserved differences in event propensity between clusters. After observing a cluster’s recurrent-event history, its frailty is no longer described only by the population distribution: the number and timing of events, together with time at risk, provide information about its latent risk level. Posterior frailty prediction combines that cluster’s history with the frailty distribution assumed by the model.

The key distinction is between predicting with a frailty distribution and substituting a single frailty value. A posterior mean is useful for describing the cluster’s average conditional intensity, but a future event probability generally requires averaging over the entire posterior distribution. The latter accounts for uncertainty about the cluster’s frailty.

Model and Posterior Update

Consider clusters indexed by \(j\), with members indexed by \(i\). Conditional on a shared cluster frailty \(u_j\), member \(i\)'s recurrent-event process has intensity \(u_j\lambda_0(t)\exp(x_{ij}(t)^\top\beta)\) while that member is at risk. Here \(\lambda_0(t)\) is the baseline intensity, \(x_{ij}(t)\) is a vector of observed covariates, and \(u_j>0\) multiplies the intensity for every member of cluster \(j\). A value above one indicates higher conditional intensity than the frailty-free level; a value below one indicates lower intensity.

For each cluster, let \(D_j\) be the total number of observed events over its members and let \(A_j\) be the accumulated frailty-free hazard exposure over the observed at-risk periods:

Thus \(A_j\) incorporates observed follow-up and covariate values, but not the unknown multiplier \(u_j\). An individual’s time under observation contributes exposure while that individual is at risk, including periods ending in right censoring. A period after a terminal event does not contribute if the person is no longer at risk.

Assume the frailties are independent across clusters with a gamma prior, parameterized by shape \(a\) and rate \(b\):

Under this shape-rate convention, the prior mean is \(a/b\) and the prior variance is \(a/b^2\). A common mean-one parameterization sets \(a=b=1/\theta\), giving prior variance \(\theta\).

Why the Posterior Has This Form

Conditional on \(u_j\), the recurrent-event likelihood has a factor \(u_j\) for each event and an exponential survival factor for the cumulative intensity. Terms involving event times, covariates, and the baseline intensity do not depend on \(u_j\), so the part of the likelihood that does depend on it is

This expression assumes the conditional recurrent-event intensity model is appropriate and that the censoring mechanism contributes no additional information about \(u_j\) after conditioning on the modeled history. For example, censoring may determine when exposure ends, but its own likelihood must not depend on the frailty in an unmodeled way.

Multiplying the likelihood by the gamma prior gives the posterior kernel. The powers of \(u_j\) add, as do the exponential rates:

This is the kernel of a gamma density with shape \(a+D_j\) and rate \(b+A_j\), establishing the update. The event count increases the posterior shape; accumulated exposure increases its rate. Both affect the posterior mean, but in different ways.

From Posterior Frailty to Cluster-Specific Risk

At a future time \(t\), for a member with specified covariates who remains at risk, the posterior mean conditional intensity is the frailty-free intensity multiplied by the posterior mean frailty:

This describes average instantaneous intensity under the posterior. For prediction over a future interval, let \(B_j\) be the future frailty-free integrated intensity across the person or members of interest. Conditional on \(u_j\), the probability of no event in that interval is \(\exp(-u_jB_j)\). Averaging this probability over the posterior gives

The last equality follows from the gamma Laplace transform: if \(U\sim\operatorname{Gamma}(r,s)\) in the shape-rate convention, then integrating its density gives \(E(e^{-BU})=(s/(s+B))^r\). This is a posterior predictive event probability for the specified future interval and risk set, under the model.

Assumptions, Checks, and Limitations

The gamma update is exact for the specified model when the conditional likelihood has the stated \(u_j^{D_j}\exp(-u_jA_j)\) dependence and the prior is gamma. Its use therefore relies on several substantive conditions:

• The frailty acts multiplicatively on the modeled intensity and is shared by the members whose histories are pooled.
• The event-count and accumulated-exposure likelihood correctly represent the recurrent-event process, including how risk sets change over time.
• The prior distribution and its parameters are appropriate. In fitted models, these parameters and the baseline and regression parameters are usually estimated; treating them as fixed omits their estimation uncertainty from this posterior calculation.
• Censoring is noninformative for frailty given the modeled information, or its dependence is explicitly represented. If high-frailty members are more likely to be censored in an unmodeled way, observed exposure alone does not repair the resulting bias.

A practical check is to reconstruct \(D_j\) and \(A_j\) from the records and verify that each event contributes one event factor and each at-risk segment contributes its integrated intensity. Unexpected posterior estimates can result from omitted exposure, double-counted event time, an incorrect risk set, or a mismatch between the assumed and actual recurrent-event process. A cluster with few events and little exposure will generally retain substantial posterior uncertainty, so its point estimate should not be treated as a precise measurement.

This approach differs from using only the posterior mean: the mean is sufficient for the posterior average intensity, while a nonlinear prediction such as the probability of at least one event requires integrating the relevant probability over the posterior. It also differs from a population-average prediction, which averages across the frailty distribution rather than conditioning on one observed cluster’s history. A gamma prior yields the convenient closed-form update above; with a non-gamma prior, the same likelihood contribution can still be combined with that prior, but a closed-form posterior or predictive risk may not be available.

These are model-based predictions, not causal effects. They are conditional on the chosen covariate values, future risk interval, baseline and regression parameters, and assumptions about the event and censoring processes. They also do not automatically account for uncertainty in estimated model parameters. Consequently, a single cluster’s posterior estimate should not be interpreted as proof of a stable, intrinsic property independent of the model or its observed history.

Common Mistake