MerQuR+Statistics Glossary

Statistics Glossary → Frailty model

Frailty modelconcept

An approach in survival analysis that models the unobserved risk shared by observations in the same cluster (family, hospital, recurrent events within a person) through a random effect that multiplies the hazard function.

The Cox model assumes that observations are independent. Yet members of the same family share genetic and environmental factors, and patients treated in the same hospital share the quality of care. In the shared frailty model, the hazard of an individual in cluster i is h(t) = uᵢ · h₀(t) · exp(Xβ), where uᵢ is the cluster's unobserved 'frailty'. It is usually assumed to come from a gamma distribution with mean 1 and variance θ: clusters with uᵢ > 1 carry above-average risk. θ = 0 means the clusters do not differ; the larger θ, the stronger the within-cluster dependence.

The coefficients of a frailty model describe within-cluster (conditional) effects, whereas the cluster-robust marginal Cox model corrects for dependence only in the standard errors and yields population-averaged (marginal) effects. The two sets of coefficients do not measure the same thing, and when frailty is present conditional effects are usually larger than marginal ones. The choice of frailty distribution (gamma, log-normal) can affect the results, and with few or small clusters the estimate of θ is imprecise. Frailty also explains the selection effect whereby, because of unmeasured risk factors, high-risk individuals experience the event first and leave the sample over time.

An everyday example

A study follows the time to heart disease among siblings from 50 families. In some families almost all the siblings fall ill at a young age, in others none do; this points to an unmeasured shared family risk. The frailty model gives each family a risk multiplier: someone from a family with a frailty of 1.8 faces, at every moment, about 1.8 times the hazard of a person with the same characteristics from an average family.

In MerQur

It is chosen with the 'Model (Cluster-robust marginal Cox / + shared gamma frailty)' parameter in Cluster-Robust (Marginal) Cox; θ is estimated from the marginal likelihood, and β and the frailties by the EM algorithm.

Analyses where this concept appears

Cluster-Robust Cox (Marginal and Frailty Models)

Related concepts

Cluster-robust (clustered) standard errorsFixed and random effects