Frailty Models

A multiplicative random effect acting on the baseline hazard, representing unobserved risk factors. In a shared frailty model, individuals within the same cluster (e.g., a clinical trial site) share the same frailty value.

• Gamma Frailty: The most common distribution, chosen for mathematical convenience and its ability to produce a closed-form likelihood.
• Log-Normal Frailty: Used when the random effect is modeled on the log scale, often in a hierarchical survival framework.
• Interpretation: A frailty > 1 indicates an individual or cluster is more 'frail' and experiences the event sooner than predicted by observed covariates alone.

Frailty models are categorized by the scope of the clustering effect they aim to capture.

• Shared Frailty: Assumes individuals within a pre-defined group (e.g., family, hospital, litter) share a common frailty value, inducing a positive correlation between their event times. This is the standard model for multicenter trials.
• Individual (Unshared) Frailty: Assigns a unique random effect to each subject to account for overdispersion or heterogeneity not explained by fixed covariates, effectively capturing the 'unexplained' variation in a population.

The interpretation of covariate effects differs fundamentally from the standard Cox model.

• Conditional (Subject-Specific) Effects: The hazard ratio represents the effect of a covariate for a given individual or cluster, holding the frailty constant. These effects are typically larger in magnitude.
• Marginal (Population-Averaged) Effects: The effect of a covariate averaged over the distribution of the frailty. The marginal hazard ratio is attenuated compared to the conditional hazard ratio.
• Key Insight: A frailty model estimates the conditional effect, which is crucial for individual-level prognosis.

Since frailties are unobserved latent variables, standard maximum likelihood is not directly applicable. Estimation relies on integrating out the random effects.

• EM Algorithm: Treats the frailties as 'missing data' and iteratively estimates the baseline hazard, regression coefficients, and frailty variance.
• Penalized Partial Likelihood: The random effects are treated as a penalty term that shrinks extreme cluster effects toward zero, stabilizing the optimization.
• Bayesian MCMC: Hierarchical Bayes methods using Markov Chain Monte Carlo are common for complex, multi-level frailty structures.

A significant frailty variance indicates unobserved heterogeneity that would otherwise bias results.

• Likelihood Ratio Test: A formal test for the null hypothesis that the frailty variance is zero. A significant p-value rejects the standard Cox model in favor of the frailty model.
• Kendall's Tau: Measures the correlation between event times within a cluster. A value significantly greater than zero suggests shared frailty is present.
• Consequence of Ignoring Frailty: Estimated hazard ratios are biased toward the null, and standard errors are underestimated, inflating the Type I error rate.

An extension for multivariate survival data where the frailties themselves are correlated, often used in twin studies or genetic epidemiology.

• Additive Genetic Model: Decomposes the frailty into a sum of genetic and environmental components to estimate heritability of time-to-event traits.
• Bivariate Frailty: Models the joint survival of pairs (e.g., twins, paired organs) by assuming a correlation structure between their individual frailties.
• Application: Distinguishing the genetic predisposition to disease onset from shared environmental risk factors.

The semiparametric foundation upon which frailty models are built. The Cox model assumes that all relevant heterogeneity is captured by observed covariates, estimating hazard ratios without specifying the baseline hazard. When this assumption fails—due to unmeasured center effects or genetic predispositions—shared frailty or individual frailty extensions become necessary.

• Models the hazard as ( h(t|X) = h_0(t) \exp(X\beta) )
• Frailty models extend this by adding a multiplicative random effect ( Z )
• The proportional hazards assumption still applies conditionally on the frailty term

Frailty terms are random effects introduced into survival models to account for correlation within clusters. Unlike fixed effects, which estimate a separate baseline hazard per group, random effects assume the frailties are drawn from a distribution—typically gamma or log-normal.

• Gamma frailty: mathematically convenient, closed-form Laplace transform
• Log-normal frailty: allows flexible correlation structures and is natural for hierarchical models
• The variance of the frailty distribution quantifies the degree of unobserved heterogeneity

The most common frailty specification for multicenter clinical trials, where patients within the same hospital or study site share a common unobserved risk factor. The shared frailty induces within-cluster dependence—patients at a high-frailty center experience events sooner than those at a low-frailty center, even with identical covariates.

• Assumes conditional independence given the frailty
• Used to estimate center effects in drug trials
• Extends naturally to recurrent events where episodes cluster within a patient

When patients face multiple mutually exclusive event types, frailty models can be extended to cause-specific hazards with correlated random effects. This is critical in oncology, where a treatment may reduce cancer mortality but increase cardiovascular death risk.

• Shared frailty across competing risks captures latent susceptibility
• The Fine-Gray subdistribution hazard model can incorporate frailty for direct cumulative incidence modeling
• Identifiability requires careful consideration of the censoring mechanism

A natural extension of frailty thinking: instead of a static random intercept, the latent variable becomes a time-varying stochastic process linking repeated biomarker measurements to the hazard. The shared random effect structure corrects for measurement error and informative dropout.

• Typically uses a linear mixed model for the longitudinal submodel
• The expected value of the latent process enters the survival submodel
• Bayesian estimation via MCMC is common for these complex likelihoods

When a subject can experience the same event multiple times—such as asthma exacerbations or tumor recurrences—a subject-specific frailty accounts for the dependence between gap times. The Andersen-Gill model treats events as independent conditional on covariates, but adding a frailty term captures the reality that some patients are simply more prone to events.

• Prentice-Williams-Peterson models stratify by event rank
• Frailty variance estimates the degree of patient-level heterogeneity
• Essential for accurately estimating treatment effects in chronic disease trials