Joint Shared and Individual Frailties

Why Include Two Frailty Levels?

Members of a cluster can resemble one another for more than one reason. They may share a clinic, household, or site-level environment, while each person also has persistent characteristics that affect their own recurrence rate. A model with only one frailty level cannot represent both patterns separately. A joint model gives one latent multiplier to the whole cluster and another to each individual within it.

The distinction matters because dependence has a structure. A cluster-level frailty can associate the event histories of different people in the same cluster. An individual-level frailty links repeated events for the same person, but does not directly link that person to other cluster members. The individual multiplier also combines with the cluster multiplier, so both contribute to a person's overall rate.

Definition: Let $j$ index clusters and $i$ index individuals within cluster $j$. A joint shared-and-individual frailty model for recurrent events has conditional intensity $$ \lambda_{ij}(t\mid U_j,V_{ij}) = Y_{ij}(t)\lambda_0(t)\exp\{\beta^\mathsf{T}X_{ij}(t)\}U_jV_{ij}. $$ Here $Y_{ij}(t)$ indicates that individual $i$ in cluster $j$ is under observation and at risk at time $t$; $\lambda_0(t)$ is a baseline event rate; $X_{ij}(t)$ is a covariate vector; $U_j$ is the cluster frailty; and $V_{ij}$ is the individual frailty. Both frailties are positive and have mean one. Thus, conditional on the frailties, the rate is multiplied by $U_jV_{ij}$.

Model Assumptions and Meaning

A common specification takes $U_j$ to be independent and identically distributed across clusters, and $V_{ij}$ to be independent and identically distributed across individuals. The two frailty levels are independent of one another and of the measured covariates. Conditional on the covariate histories and both frailties, individuals' event processes are independent, and the model specifies the recurrent-event intensity correctly. Censoring and entry into observation must also be noninformative conditional on the modeled history and covariates; otherwise, the observed event histories may not represent the assumed process.

For interpretable baseline and regression effects, the frailties are conventionally normalized to have mean one. Write their variances as $\theta_C=\operatorname{Var}(U_j)$ and $\theta_I=\operatorname{Var}(V_{ij})$. The conditional rate ratio for a one-unit increase in a covariate $x$, holding both frailties fixed, is $\exp(\beta_x)$. The frailties are not additional measured covariates: they represent unobserved multiplicative heterogeneity. A zero variance at either level means the corresponding frailty is degenerate at one and that level contributes no latent variation.

The product $U_jV_{ij}$ is an individual's total latent multiplier. Consequently, an individual's event rate alone does not reveal whether a high multiplier came from their cluster, their own persistent characteristics, or both. The model uses the pattern across people and repeated events—not a direct observation of the frailties—to separate the components.

Worked Example: Reading the Two Multipliers

Suppose a hypothetical cluster has $U_j=1.5$. For two individuals at a time when each is at risk, assume their baseline rate times covariate multiplier is $0.4$ events per person-year. Let their individual frailties be $V_{1j}=0.8$ and $V_{2j}=1.2$.

For the first individual, the conditional intensity is $$ 0.4(1.5)(0.8)=0.48 $$ events per person-year. For the second, it is $$ 0.4(1.5)(1.2)=0.72 $$ events per person-year. The cluster multiplier raises both rates by the same factor, while the individual multipliers make the rates differ within the cluster. These are conditional rates for the hypothetical frailties, not marginal rates averaged over a population.

How the Two Levels Create Different Dependence

To see the dependence algebraically, consider event counts over specified observation intervals. Conditional on the frailties, suppose the count $N_{ij}$ is Poisson with mean $U_jV_{ij}H_{ij}$, where $H_{ij}$ is the integrated baseline rate times the covariate multiplier and at-risk indicator over that interval. This Poisson-count calculation is a useful special case of the intensity model. Assume the mean-one frailties are independent, with variances $\theta_C$ and $\theta_I$.

For two different individuals $i$ and $k$ in the same cluster, conditional independence gives zero conditional covariance. Their marginal covariance comes from the shared $U_j$. For two disjoint intervals for the same individual, both $U_j$ and $V_{ij}$ are shared, so both contribute. The short derivation below makes these distinctions explicit.

Compute the mean.
Since $E(U_j)=E(V_{ij})=1$, iterated expectation gives $E(N_{ij})=E\{U_jV_{ij}H_{ij}\}=H_{ij}$.

Compare different individuals.
For distinct individuals in a cluster, conditional independence implies $E(N_{ij}N_{kj}\mid U_j,V_{ij},V_{kj})=U_j^2V_{ij}V_{kj}H_{ij}H_{kj}$. Taking expectations and subtracting the product of means gives $\operatorname{Cov}(N_{ij},N_{kj})=\theta_C H_{ij}H_{kj}$.

Compare intervals for one individual.
For conditionally independent counts in disjoint intervals with integrated exposures $H_a,H_b$, their marginal covariance is $H_aH_b\{\operatorname{Var}(U_jV_{ij})\}$. Independence and the mean-one normalizations give $\operatorname{Var}(U_jV_{ij})=(1+\theta_C)(1+\theta_I)-1=\theta_C+\theta_I+\theta_C\theta_I$.

The same product variance determines the extra-Poisson component of a single interval count. The law of total variance gives $$ \operatorname{Var}(N_{ij})=H_{ij}+H_{ij}^{\,2}\{\theta_C+\theta_I+\theta_C\theta_I\}. $$ The first term is the conditional Poisson variation. The second is variation in the latent rate. For different individuals, the corresponding covariance contains only $\theta_C$, because their individual frailties are independent. This contrast is the central source of information for separating the two variance components.

Worked Example: Calculating the Dependence Contributions

Suppose a hypothetical model has $\theta_C=0.30$ and $\theta_I=0.50$. Two different individuals in one cluster each have integrated exposure $H=2$. Under the Poisson-count assumptions above, each has expected count $E(N)=2$, and the variance is $$ \operatorname{Var}(N)=2+2^2(0.30+0.50+0.30\times0.50) =2+4(0.95)=5.80. $$

The covariance between their counts is $$ \operatorname{Cov}(N_{1j},N_{2j})=(0.30)(2)(2)=1.20. $$ Only the shared frailty contributes to this between-person covariance. Now consider two disjoint intervals for one individual, each with integrated exposure $H_a=H_b=1$. Their covariance is $$ (1)(1)(0.30+0.50+0.15)=0.95. $$ Here $0.15=\theta_C\theta_I$. The interaction term arises because the individual rate contains the product of the two frailties; the two variance contributions are not simply additive in the total multiplier's variance.

Likelihood and Information for Identifying the Components

In likelihood estimation, the frailties are latent and must be integrated out. For a cluster's observed event histories, let $L_{ij}(u,v)$ denote the conditional likelihood contribution for individual $i$ when $U_j=u$ and $V_{ij}=v$. Conditional independence gives the cluster likelihood $$ L_j=\int_0^\infty f_U(u)\prod_i\left\{\int_0^\infty L_{ij}(u,v)f_V(v)\,dv\right\}du. $$ This expression shows the hierarchical structure: each individual's likelihood is first averaged over that individual's frailty, conditional on the shared cluster value; the result is then averaged over the cluster frailty. The distributions and their variance parameters must be specified to evaluate this likelihood. Gamma and lognormal choices are examples; the choice affects the model and its fitted likelihood.

The data need informative replication at both levels. Repeated events or separated event intervals within individuals provide information about persistent individual-level heterogeneity. Multiple individuals within clusters provide comparisons that can reveal shared variation. In the idealized count setting, if the covariance between different cluster members is divided by their exposures, it identifies $\theta_C$. The corresponding within-person covariance divided by the two interval exposures identifies $\theta_C+\theta_I+\theta_C\theta_I$. If these quantities are denoted $c$ and $q$, respectively, then $$ \theta_I=\frac{q-c}{1+c}, $$ provided $c$ and $q$ arise from the stated model. This moment argument illustrates the information structure; fitted likelihood estimates generally use the full event times, censoring, covariates, and chosen frailty distributions rather than raw covariance alone.

Worked Example: Separating Variances From Hypothetical Moments

Suppose hypothetical population moments are available for two individuals in the same cluster with exposures $H_1=1.5$ and $H_2=2$. Their count covariance is $0.90$. The normalized between-person covariance is $$ c=\frac{0.90}{(1.5)(2)}=0.30, $$ so the model implies $\theta_C=0.30$.

For one individual, suppose counts in two disjoint intervals with exposures $H_a=0.8$ and $H_b=1.2$ have covariance $0.84$. The normalized within-person covariance is $$ q=\frac{0.84}{(0.8)(1.2)}=\frac{0.84}{0.96}=0.875. $$ Since $q=\theta_C+\theta_I+\theta_C\theta_I$, solve for the individual variance: $$ \theta_I=\frac{q-\theta_C}{1+\theta_C} =\frac{0.875-0.30}{1.30}=0.4423. $$ These calculations recover model variance components from exact theoretical moments. Real sample covariances are noisy, and this calculation alone does not supply standard errors, adjust for covariates, or establish that the assumed frailty distributions and event-process model are suitable.

Diagnostics, Alternatives, and Limitations

Warning signs include variance estimates near zero when the data appear to show dependence at that level, unstable estimates when cluster sizes or repeated events are sparse, and strong sensitivity to the frailty distribution. The moment relationships also offer a basic coherence check: under the stated model, normalized within-person covariance $q$ should not be below normalized between-person covariance $c$, because $q-c=(1+c)\theta_I\geq0$. A clear contradiction in well-estimated summaries may indicate sampling fluctuation, misspecified exposure or covariate structure, dependence not represented by this model, or a mismatch between the assumed process and the data. It is not by itself a formal model test.

A shared-frailty-only model sets $\theta_I=0$; it attributes persistent dependence to cluster membership and cannot represent extra individual persistence beyond that shared level. An individual-frailty-only model sets $\theta_C=0$; it can link recurrences for one person but cannot induce dependence between different people in a cluster. The joint model is more flexible, but that flexibility requires enough information to distinguish both components. Frailty models represent latent association under their assumptions; they do not establish that a particular unmeasured cluster or individual cause exists.

Sparse clusters with one observed individual provide no direct between-person comparison. Individuals with few events or little follow-up provide limited repeated-event information. If cluster membership, follow-up, or event observation is informative in ways not captured by the model, the apparent allocation of heterogeneity across levels can be misleading. Moreover, misspecified baseline rates, time-varying covariate effects, or event-history effects can be mistaken for frailty variation. Interpret the fitted components in light of the design and model assumptions rather than as automatic decompositions of all observed variability.

Key takeaway: The shared frailty links different individuals in the same cluster; the individual frailty links repeated events for one person. Both enter each person's rate multiplicatively, so within-person dependence reflects both variance components and their product, while between-person covariance reflects the shared component.

Check Your Understanding

Answer each question from the model and assumptions given above. Solutions follow.

Which frailty component directly induces covariance between counts for two different people in the same cluster?
If $\theta_C=0$ and $\theta_I>0$, what dependence remains between separate event intervals for one person?
For $\theta_C=0.2$, $\theta_I=0.4$, and $H=1.5$, calculate the count variance in the Poisson special case.
Why is a cluster containing only one observed individual uninformative for a direct comparison of within-cluster and between-person dependence?

Solutions. (1) The cluster frailty $U_j$ is shared across people and gives covariance $\theta_C H_{ij}H_{kj}$. (2) Individual-level dependence remains: the disjoint-interval covariance is $H_aH_b\theta_I$, since the cluster multiplier is constant at one when its variance is zero. (3) The excess-rate variance is $0.2+0.4+(0.2)(0.4)=0.68$, so $\operatorname{Var}(N)=1.5+(1.5)^2(0.68)=1.5+1.53=3.03$ events squared. (4) There is no second cluster member with which to compare event histories, so the data provide no direct between-person covariance information for that cluster.