Seminar

Ram Tiwari, Ph.D.

"Prediction of U.S. Mortality Counts Using Semiparametric Bayesian Techniques"

Accurate prediction of cancer mortality figures for the current and upcoming year are extremely essential for public health planning and evaluation. Due to delay in reporting cause-specific mortality for the US, there is a 3-year lag between the latest year for which such figures are available and the current year. Prior to 2004, the American cancer Society (ACS) used to predict cancer mortality counts by first fitting a time series model with quadratic trend and autoregressive error to the past data and then projecting this model into the future. Beginning 2004 the ACS has begun to implement a new methodology, developed by the NCI, in its annual publication Cancer facts & Figures, 2004. The new method, known as the state-space method (SSM), uses a quadratic trend with random time-varying coefficients to model the mortality counts, quickly adjusts to sudden changes in the observed trend, and, as a result, generally provides predictions of mortality counts that are closer to their observed values compared to the corresponding predictions obtained from the previous method.

In this talk we present Bayesian versions of the SSM. In particular, we present two models for short-term prediction of the number of deaths that arise from common cancers in the United States. The first is a local linear model, in which the slope of the segment joining the number of deaths for two consecutive time periods is assumed to be random with a nonparametric distribution, which has a Dirichlet process prior. For slightly, longer prediction periods, we present a local quadratic model. The proposed methods are used to obtain the predictive distributions of the future number of deaths through Markov chain Monte Carlo techniques. We illustrate our methods by runs on data from selected cancer sites and provide guidelines on how to choose prior parameters that balance model flexibility with degree of smoothness in the prediction process. The Bayesian models are compared with both the SSM and the previous ACS method.

Seminar Date:
June 5, 2007