Definition
A frailty model is a survival (time-to-event) model with a random effect — the "frailty" — that multiplies the hazard to capture unobserved heterogeneity or within-group correlation. Each individual or group i has a latent frailty ωi acting on its hazard, λij(t)=ωiλ0(t)exij′β (or additively on the log-hazard), so that "frailer" subjects fail earlier than their covariates alone would predict. Therneau-Grambsch-Pankratz (2003) show these models can be fit efficiently as penalized Cox/parametric regressions (Therneau-Grambsch-Pankratz 2003).
Key Ideas
- Frailty = survival random effect. A continuous latent variable describing excess risk for a category (individual, family, centre) over and above measured covariates — the survival-analysis analogue of a mixed-model random effect, motivating away unobserved heterogeneity (Aalen).
- Shared frailty for correlated/recurrent events. A common ωi shared by observations in a group induces within-group dependence — the standard device for recurrent events (repeated seizures, hospitalizations) and clustered survival (family members, multi-centre trials).
- Gamma shared frailty ↔ penalized likelihood. The paper's key result: solutions for the gamma shared-frailty model can be obtained exactly via penalized estimation, and Gaussian (log-normal) frailty models are closely linked to penalized models — so frailty fitting reduces to adding a penalty to the ordinary Cox partial likelihood.
- Laplace approximation for other frailties. For frailty distributions without an exact penalized form, a Laplace approximation to the marginal likelihood recovers a penalized problem, extending the approach broadly.
- Cheap computation. Because standard Cox/parametric software already solves penalized regressions, frailty models inherit that machinery and fit rapidly — implemented in the
coxph/survival routines (S-Plus, R).
- Bayesian/MCMC route. The same model has a fully Bayesian formulation (Clayton 1991): a hierarchical graphical model with a gamma-process prior on the integrated baseline hazard, a gamma frailty, and a vague hyperprior on its variance, fit by Gibbs sampling (stochastic substitution). The Poisson/counting-process likelihood factorization makes the full conditionals tractable — an early (1990s) import of MCMC into survival analysis, complementary to the penalized-likelihood route.
Why It Matters
- Unobserved heterogeneity distorts survival estimates. Ignoring frailty biases hazard-ratio estimates and can make hazards look artificially "decreasing" as frail subjects are selected out over time; frailty models correct for it.
- The standard tool for correlated event times. Shared frailty is the workhorse for recurrent-event and clustered-survival data, the survival-analysis counterpart to random-intercept mixed models.
- Unifies random effects and penalization. Casting frailty as penalized likelihood links survival modelling to the broader penalized-regression family (the same penalty-↔-Gaussian-prior duality behind penalized splines and mixed models), and makes the models estimable with existing, fast solvers.
Open Questions
- Choice of frailty distribution (gamma vs. log-normal vs. positive-stable) and its effect on inference; the distribution is weakly identified from data.
- Selecting the penalty/variance-component and quantifying its uncertainty (marginal likelihood vs. REML-type approaches).
- Multivariate and nested/hierarchical frailties, and time-varying frailty.
Related