Stochastically changing covariates may influence survival. They may
be observed, unobserved or partly observed. We review the properties of hazard
models explicitly representing the effects of unobserved, and partially
observed, stochastic covariates. Such models will increase in importance as new
longitudinal population studies, and longitudinal surveys of high dimensional
failure processes in humans, become available--many are now in progress. It is
shown that marginal survival distributions and likelihoods generated in
analytically closed form make such parametrically detailed models
computationally tractable. Several ways of defining the marginal distribution
of the data for constructing a likelihood function are considered. The most
complete models can handle both continuously and discretely evolving
covariates. Parameters can be estimated from multiple data sets to
retrospectively and prospectively evaluate covariate trajectories. Such methods
will both extract more information from a longitudinal study and use it in a
parametric structure that is logically consistent with the behavior of the
underlying processes of substantive interest.
Publié le : 1997-02-14
Classification:
Stochastic hazards,
random covariates,
Cameron-Martin,
martingales,
conditional Gaussian processes,
Wiener processes,
Kalman filters,
quadratic hazard functions
@article{1029963259,
author = {Yashin, Anatoli I. and Manton, Kenneth G.},
title = {Effects of unobserved and partially observed covariate processes
on system failure: a review of models and estimation strategies},
journal = {Statist. Sci.},
volume = {12},
number = {1},
year = {1997},
pages = { 20-34},
language = {en},
url = {http://dml.mathdoc.fr/item/1029963259}
}
Yashin, Anatoli I.; Manton, Kenneth G. Effects of unobserved and partially observed covariate processes
on system failure: a review of models and estimation strategies. Statist. Sci., Tome 12 (1997) no. 1, pp. 20-34. http://gdmltest.u-ga.fr/item/1029963259/