Shared Frailty Induces Within-Cluster Dependence

Why Shared Frailty Creates Dependence

Two people in the same cluster may have related event times even when their event processes are independent after conditioning on a shared latent risk. For example, a cluster-level factor that raises both members’ event hazards can make early events in one member informative about the likely risk of the other. The dependence arises when that common factor is unobserved and integrated out; it is not the same as conditional dependence given the factor.

The derivation below concerns one event time for each of two cluster members. It uses a shared frailty and assumes conditional independence of their event times given that frailty. This gives a precise way to describe how a latent common risk induces dependence in the marginal event-time distribution.

The mean-one convention separates the overall hazard scale from the frailty variation: \(\theta=0\) corresponds to a nonrandom frailty equal to one, while larger \(\theta\) means greater variation in shared risk. A common covariate model has \(\lambda_i(t)=\lambda_0(t)\exp(x_i^\mathsf{T}\beta)\), so \(A_i(t)=\Lambda_0(t)\exp(x_i^\mathsf{T}\beta)\). The dependence derivation itself requires only the two cumulative hazards and the shared-frailty assumptions above.

From Conditional Independence to Marginal Dependence

Let \(L_U(s)=E\{\exp(-sU)\}\) be the Laplace transform of the frailty distribution. Conditional independence gives a product of conditional survival probabilities. Integrating that product over the shared \(U\) yields the joint marginal survival:

By contrast, each marginal survival is \(P(T_i>t_i)=L_U\{A_i(t_i)\}\). If the two event times were marginally independent, the joint survival would equal the product of those marginal survivals. In general, \(L_U(A_1+A_2)\ne L_U(A_1)L_U(A_2)\). The same random \(U\) appears in both conditional probabilities, so averaging their product is not the same as multiplying their separate averages.

To see the sign of that covariance more generally, define \(f(U)=\exp(-A_1U)\) and \(g(U)=\exp(-A_2U)\), for fixed nonnegative cumulative hazards \(A_1,A_2\). Both functions decrease as \(U\) increases. Let \(U'\) be an independent copy of \(U\). Expanding the product and using independence of \(U\) and \(U'\) gives

The inequality holds because the two differences always have the same sign: if \(U>U'\), both are nonpositive, and if \(U<U'\), both are nonnegative. Thus the event-free indicators have nonnegative covariance. It is strictly positive when the frailty is nondegenerate and both cumulative hazards are positive. If \(\theta=0\), \(U=1\) with probability one and this source of marginal dependence disappears.

Gamma Frailty and Horizon-Specific Dependence

For a gamma frailty with mean one and variance \(\theta>0\), use the shape-rate parameterization \(U\sim\operatorname{Gamma}(1/\theta,1/\theta)\). Its Laplace transform is \(L_U(s)=(1+\theta s)^{-1/\theta}\). Consequently, for horizons \(t_1,t_2\), the joint survival and the covariance of event-free indicators \(I_i=I(T_i>t_i)\) are

Here \(A_i\) denotes \(A_i(t_i)\) in the second line. A horizon-specific correlation is this covariance divided by the product of the standard deviations of the two indicators:

This correlation describes whether two members are event-free at their respective specified horizons. It is not automatically the correlation between their event times. Both the covariance and the correlation depend on the horizons and cumulative hazards; the correlation can also behave differently from the covariance because its denominator changes.

Dependence Between Event Times

Joint survival fully describes the bivariate event-time distribution in this continuous-time setting, but it can be useful to calculate a familiar event-time correlation as well. Consider the special case of a constant baseline hazard \(\lambda\), shared gamma frailty \(U\), and conditionally independent event times. Given \(U\), each \(T_i\) has an exponential distribution with rate \(\lambda U\). Therefore \(E(T_i\mid U)=1/(\lambda U)\), and conditional independence gives \(E(T_1T_2\mid U)=1/(\lambda^2U^2)\).

For the gamma parameterization above, inverse moments exist only when their order is below the shape \(1/\theta\). In particular, \(E(U^{-1})\) is finite for \(\theta<1\), and \(E(U^{-2})\) is finite for \(\theta<1/2\). Under the latter condition, \(E(U^{-1})=1/(1-\theta)\) and \(E(U^{-2})=1/\{(1-\theta)(1-2\theta)\}\). The event-time covariance is therefore

Assumptions, Comparisons, and Failure Modes

The calculations rely on several substantive assumptions. The frailty is shared by the two members, multiplies their hazards, and has the stated distribution and mean. Event times are independent conditional on that frailty and any specified covariates. The event-time example adds a constant conditional baseline hazard; its covariance and correlation formulas also require the indicated finite inverse moments. If these assumptions do not describe the data-generating process, the formulas need not describe the observed dependence.

Shared frailty is one explanation for within-cluster association. A model with no shared latent factor but with conditional dependence specified directly is a different approach; it need not produce the same joint survival structure. Similarly, conditional independence under shared frailty is not marginal independence. Cluster-robust uncertainty calculations address a different question: they can adjust uncertainty for clustering but do not, by themselves, specify this joint event-time distribution.

Dependence summaries also have limitations. A positive covariance of event-free indicators at a selected horizon does not fully characterize the shape or strength of all dependence between event times. Conversely, a single event-time correlation can hide horizon-specific patterns and may not exist when the marginal event-time variance is infinite. In gamma models, the inverse-moment conditions above are essential, not technical decoration.

• Common mistake: multiplying marginal survivals. The reasoning “the members are conditionally independent, so their marginal survivals multiply” omits the shared random \(U\). The correction is to multiply conditional survivals first, then average: \(E\{\exp[-U(A_1+A_2)]\}=L_U(A_1+A_2)\).
• Common mistake: treating a horizon correlation as an event-time correlation. The correlation of \(I(T_i>t_i)\) concerns binary status at chosen horizons. The correlation of \(T_1,T_2\) concerns the event times themselves and requires finite second moments.
• Diagnostic warning: interpreting the fitted frailty as proof of a mechanism. Shared frailty can represent unmeasured common risk, but an observed association alone does not establish its cause. Different unmeasured factors or direct dependencies can also produce clustered outcomes.
• Numerical warning: rounding too early. Compute the joint survival, marginal survivals, covariance, and variances at full precision, then round the final result. Subtracting separately rounded probabilities can change the reported last digit of a small covariance.