A comparison of spatio-temporal prediction methods of cancer incidence in the U.S

Cancer is the cause of one out of four deaths in the United States, and in 2009, researchers expected over 1.5 million new patients to be diagnosed with some form of cancer. People diagnosed with cancer, whether a common or rare type, need to undergo treatments, the amount and kind of which will depend on the severity of the cancer. So how do healthcare providers know how much funding is needed for treatment? What would better enable a pharmaceutical company to determine how much to allocate for research and development of drugs, the amount of each drug to manufacture, or the time spent to improve or reformulate those drugs? How do government planners determine which cancers need more attention than others? To answer these questions, it becomes extremely important to get accurate predictions of new cancer cases (also known as cancer incidences) that will occur in the future based on past data. Past data on cancer incidences in the U.S. is available only at certain cancer registries. These registries did not all come online at the same time, resulting in varying lengths of incidence data. Prediction into the future would require one to account for these varying lengths. Additionally, since these registries do not cover the entire United States, one needs to incorporate some spatial projection methods. In this thesis, we develop a Bayesian spatio-temporal method of predicting future cancer incidences based on past data. A conditional autoregressive prior is used for the spatial component and an autoregressive model is used for the temporal part. We use standard Bayesian Markov chain Monte Carlo techniques to develop predictions four years into the future for individual states. The method is illustrated using incidence data for some rare and common cancers