Frailty model
A frailty model represents unobserved heterogeneity in time-to-event data through a latent random effect on the hazard.
In a common multiplicative specification, h(t|Z=z) = z h0(t), with a positive, dimensionless multiplier Z and baseline hazard h0. Setting E[Z]=1 fixes the scale, not a measured property of a drive fleet. [1]
The population hazard is h0(t) E[Z|T>t], not generally h0(t) E[Z]. Higher-frailty units tend to fail sooner, changing the survivor mixture even when each unit's multiplier is fixed. Conditional and population hazard trends can therefore differ. [1]
Individual frailty describes variation between units. Shared frailty assigns a common latent effect to a specified group and can model dependence between its event times. The conventional shared model assumes conditional independence given that effect and included covariates. State the grouping and distribution assumptions. [2]
A frailty term models random effects; a cluster-robust variance adjustment alone does not fit a latent frailty distribution. These are different modeling approaches. [3]
An overly restrictive baseline hazard can be mistaken for frailty. Arbitrarily splitting one device's unchanged observation history into daily intervals adds no survival information; intervals from one event history are not new independent devices. This equivalence assumes uninformative splitting, not changing covariates or new events. [1]
Relations
- related
- Hazard rate, Survival function
Limitations
- A latent effect does not identify a physical failure mechanism or establish causation. A falling fleet hazard alone cannot identify the frailty distribution.
Citations
- Gutierrez (2002): Parametric frailty and shared frailty survival models, sections 2-4
- Therneau, Grambsch and Pankratz: Penalized Survival Models and Frailty, Technical Report 66 (2000), sections 1-2
- R survival 3.8-6: frailty random effects and the distinction from cluster variance
Sources & caveats
- Methodological definition, not a fitted storage model. The linked analysis uses two additive linear hazards to illustrate general survivor selection, not a fixed-multiplier frailty model.
- The Mayo source is Technical Report 66 from June 2000, not the later 2003 journal version.