A parametric dynamic survival model applied to breast cancer survival times

Much current analysis of cancer registry data uses the semiparametric proportional hazards Cox model. In this paper, the time-dependent effect of various prognostic indicators on breast cancer survival times from the West Midlands Cancer Intelligence Unit are investigated. Using Bayesian methodology...

Full description

Saved in:
Bibliographic Details
Published in:Applied statistics 2002-01, Vol.51 (4), p.421-435
Main Authors: Hemming, K., Shaw, J. E. H.
Format: Article
Language:English
Subjects:
Bayesian survival
Breast cancer
Cancer
Datasets
Distribution theory
Dynamic modeling
Exact sciences and technology
Health hazards
Inference from stochastic processes
time series analysis
Markov chain Monte Carlo methods
Markovian processes
Mathematics
Medicine
Modeling
Monte Carlo simulation
Moral hazard
Multivariate analysis
Musical intervals
Neoplasia
Non-proportional hazards
Nonparametric inference
Parametric models
Piecewise exponential
Probability and statistics
Sciences and techniques of general use
Statistical discrepancies
Statistical methods
Statistical variance
Statistics
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Much current analysis of cancer registry data uses the semiparametric proportional hazards Cox model. In this paper, the time-dependent effect of various prognostic indicators on breast cancer survival times from the West Midlands Cancer Intelligence Unit are investigated. Using Bayesian methodology and Markov chain Monte Carlo estimation methods, we develop a parametric dynamic survival model which avoids the proportional hazards assumption. The model has close links to that developed by both Gamerman and Sinha and co-workers: the log-base-line hazard and covariate effects are piecewise constant functions, related between intervals by a simple stochastic evolution process. Here this evolution is assigned a parametric distribution, with a variance that is further included as a hyperparameter. To avoid problems of convergence within the Gibbs sampler, we consider using a reparameterization. It is found that, for some of the prognostic indicators considered, the estimated effects change with increasing follow-up time. In general those prognostic indicators which are thought to be representative of the most hazardous groups (late-staged tumour and oldest age group) have a declining effect.
ISSN:0035-9254
1467-9876
DOI:10.1111/1467-9876.00278