Frailty Models for Unequal Cluster Sizes

Why Cluster Size Matters

In a shared-frailty model, members of a cluster share an unobserved risk factor. Conditional on that factor, their event histories may be independent; after averaging over the factor, their outcomes are associated. Cluster size therefore affects more than the number of records in a likelihood. It also affects how much within-cluster dependence the data can reveal. A large cluster supplies many pairs of members whose outcomes can reflect the same latent risk, whereas a small cluster supplies few such pairs.

The effect is not simply “large clusters count more.” Their likelihood contributions depend on all members’ event and censoring information jointly, and their influence on estimated heterogeneity depends on the model, follow-up, and observed outcomes. The comparisons below make these distinctions explicit.

For the standard likelihood, assume the cluster sizes are fixed by design or are noninformative for the frailties and event processes, conditional on modeled covariates. Assume censoring is noninformative given the modeled information and frailty. These assumptions matter: if clusters are larger because their frailties or event processes affect cluster membership or observation, conditioning on size alone may select a nonrepresentative set of clusters.

How Size Enters an Integrated Cluster Likelihood

For member \(i\) in cluster \(j\), let \(\lambda_{0}(t)\) be the baseline hazard, \(x_{ij}\) the covariate vector, and \(\beta\) the regression coefficient vector. Conditional on \(Z_j\), the hazard is \(Z_j\lambda_{0}(t)\exp(x_{ij}^{\mathsf T}\beta)\). Let \(t_{ij}\) be the observed time, \(\delta_{ij}\) indicate an event (\(\delta_{ij}=1\)) rather than censoring (\(\delta_{ij}=0\)), and \(H_{ij}\) be the cumulative hazard through \(t_{ij}\) without the frailty multiplier. Define \(D_j=\sum_{i=1}^{m_j}\delta_{ij}\) and \(A_j=\sum_{i=1}^{m_j}H_{ij}\).

Conditional independence gives a product of member likelihood contributions at a fixed value of \(Z_j\). But the members share the same frailty, so the cluster likelihood is obtained by integrating that product over the frailty distribution. With \(Z_j\sim\operatorname{Gamma}(1/\theta,1/\theta)\), where the arguments are shape and rate, its density has mean 1 and variance \(\theta\). The resulting integrated contribution is:

Here the member-specific hazard factors account for the observed events, while \(D_j\) and \(A_j\) summarize the event count and total cumulative hazard exposure relevant to the frailty integral. Thus \(m_j\) affects the likelihood through the number and pattern of observations and through the sums \(D_j\) and \(A_j\); there is no universal multiplier equal to cluster size. In particular, integrating each member’s contribution separately would replace the shared latent \(Z_j\) with separate frailties and would describe a different model.

Why Large Clusters Can Reveal Heterogeneity

A useful way to see the connection between size and heterogeneity is to consider a common follow-up time and the binary indicator \(Y_{ij}\) that member \(i\) has had an event by that time. Let \(\Lambda\) denote each member’s cumulative hazard without frailty. Given \(Z_j\), the event probability is \(1-\exp(-Z_j\Lambda)\). The gamma frailty’s Laplace transform is \(E[\exp(-sZ_j)]=(1+\theta s)^{-1/\theta}\). Consequently, the marginal event probability is \(p=1-S\), where \(S=(1+\theta\Lambda)^{-1/\theta}\).

For two distinct members in the same cluster, the probability that both remain event-free is \(S_2=(1+2\theta\Lambda)^{-1/\theta}\). Since the covariance of two event indicators equals the covariance of their event-free indicators, their covariance is \(c=S_2-S^2\). The count of events \(N_j=\sum_iY_{ij}\) therefore has variance:

The first term is the variance that would arise from independent members with marginal event probability \(p\). The second term is the contribution from the \(m_j(m_j-1)\) ordered pairs of distinct members. Under positive shared gamma frailty, \(c\) is nonnegative: members in a cluster with high latent risk tend to have more events together, while members in a cluster with low latent risk tend to have fewer.

Unequal Sizes and the Balance of Information

For a fixed total number of members, the number of within-cluster pairs depends on how those members are grouped. Consider 16 members observed at the same exposure as in the preceding example. If they form eight clusters of size 2, there are \(8\binom{2}{2}=8\) unordered within-cluster pairs. If they form two clusters of size 8, there are \(2\binom{8}{2}=56\) such pairs. The latter design has seven times as many pair opportunities to express shared dependence.

A related descriptive summary of size imbalance is \(m_{\mathrm{eff}}=(\sum_jm_j)^2/\sum_jm_j^2\), sometimes called an effective cluster count for size weighting. It equals the number of clusters when all sizes are equal and decreases when a few clusters dominate membership. It is not a replacement for a likelihood-based information calculation, but it warns that the raw number of clusters can overstate how evenly the data are distributed across clusters.

Assumptions, Diagnostics, and Alternatives

Unequal sizes are not themselves a violation of the shared-frailty model. The central modeling assumptions remain the frailty distribution, the conditional hazard structure, and conditional independence of members given frailty and covariates. The standard likelihood also needs a defensible treatment of cluster size: if size is fixed or noninformative, one can condition on the observed sizes. If size is associated with latent risk or outcomes, an analysis that ignores that relationship can misrepresent the population distribution of frailties. The size-generating process may then need to be modeled, or conclusions should be qualified as conditional on observed cluster sizes.

Useful checks include examining the distribution of \(m_j\), the number of clusters at each size, and whether a small number of large clusters dominate total membership. Compare fitted results with and without the largest clusters as a sensitivity analysis, not as a method for choosing the preferred result. Also examine whether cluster size is associated with observed covariates, follow-up, or event burden; such patterns may signal design differences, informative size, or model misspecification. A fitted frailty variance near zero does not prove that sizes are harmless, and apparent excess event-count variability alone does not establish that a gamma frailty is correct.

Alternative approaches answer different questions. A cluster-robust variance estimator can adjust uncertainty for dependence without specifying a frailty distribution, but it does not directly estimate a latent heterogeneity variance or provide the same conditional hazard model. A model with separate member-specific frailties describes additional individual heterogeneity, rather than only the shared dependence. A fixed-effects or cluster-stratified analysis conditions on cluster-specific features but does not estimate a distribution of shared frailties in the same way. These approaches should not be treated as interchangeable corrections for unequal size.

The practical limitation is that many members in one cluster do not create many independent clusters. Large clusters can yield substantial information about within-cluster association, but estimates of the population distribution of frailties still depend on the number and representativeness of clusters. If only a few clusters are very large, the fitted heterogeneity may be sensitive to those clusters and to the frailty distributional assumption. Unequal sizes should therefore be reflected in the likelihood and the interpretation, rather than summarized by total sample size alone.

Common Pitfall