Abstract

Popular hazard models (like the Cox proportional hazard models) are based on the tacit assumption that effects of covariates on the event of interest are constant over time. In real-life situations, however, it is common to encounter situations where covariate-effects may change at specific time-points or, generally, behave dynamically.

The present work presents dynamic hazard models (as extension of the standard hazard models) in order to accommodate covariates whose effects change over time. The formulation makes the Bayesian approach a more appropriate method of inference as it allows flexibility to incorporate past information and advanced relationships among parameters through the prior distribution. In so doing, however, the Bayesian approach is plagued by correlations among the time-varying effect parameters. To handle these correlations more efficiently, we apply a Sequential Monte Carlo (SMC) method commonly known as Particle Filter.

The issues are illustrated by re-analyzing Carter et al. (1983) data on survival times of 90 gastric cancer patients equally divided in two treatment groups (see also Gamerman, 1991). Preliminary results replicate the findings of Gamerman (1991) but the present results exhibit smaller variations though these variations are dependent on the choice of the variance at the initial stage. 

References

Gamerman, D. (1991), Dynamic Bayesian Models for Survival Data. Applied Statistics 40, 63-79

Carter, W . H., Wampler, G . L., and Stablein, D . M. (1983), Regression Analysis of Survival Data in Cancer Chemotherapy. New York: Dekker.