Edward
Rotan Visiting Professor
Wesley O. Johnson, Ph.D.
Professor of Statistics
Semiparametric
AFT Models for Survival Data
The semiparametric proportional hazards (PH) model is
ubiquitous in survival literature.
It is flexible and easily fit using standard packages,
at least for right censored data.
However, the assumption of proportional hazard functions
may be violated and we may
seek a
proper alternative semiparametric model.
One such model is the accelerated
failure time
(AFT) model. Whereas the PH model assumes the covariates act multiplic-
atively on a
baseline hazard function, the AFT model assumes that the covariates act
multiplicatively on
the argument of the baseline survival distribution.
Early approaches to the semiparametric AFT model were
considered by Miller (1976),
Buckley and James (1978), Kalbfleisch
(1978), Koul, Susarla and Van
Ryzin (1981), and
Christensen and Johnson (1988). All these approaches are difficult or
sub-adequate in
one way
or another. As an example, Johnson and Christensen (1989) demonstrated that
the
mathematics involved for a fully Bayesian solution to even the non-censored
problem
in this
context is horrendous at best.
Recently, the analytic intractability of Bayesian
semiparametric inference for the AFT
model for
right censored data was overcome by utilizing Markov chain
(MCMC) methods. See
Doss (1994), Kuo and Mallick
(1997), Walker and Mallick (1999)
and Kottas and Gelfand (2000). Here,
we consider two alternative approaches that build
on these
works. We model the baseline survival distribution either with a mixture of Polya
tree (MPT)
processes or a mixture of Dirichlet processes (MDP),
given a standard para-
metric
family of base measures. Our models
allow for a complete Bayesian solution
using MCMC
methods and are direct nonparametric extensions of existing parametric
models.
Additionally, existing prior information for all parametric model parameters
can
be
incorporated immediately into the more general nonparametric approach. The MDP
model presented
is for interval censored data and is a straightforward generalization of
the
approach taken by Christensen and Johnson (1988), and thus provides a
numerically
tractable and
practically implementable, complete Bayesian approach
to a useful general-
ization of
the problem that they considered to be analytically intractable. The MPT model
builds on the
work of Walker and Mallick. This work is joint with
Tim Hanson at the