A model for cancer trends by histologic type incorporating uncertain diagnoses

When the time trends of various cancers are studied by histologic type, a serious problem arises in how to deal with the category of Not Otherwise Specified (NOS) type because this can account for a high percentage of the data. In the Connecticut Tumor Registry (CTR) data, the NOS type for lung cancer is usually treated as a separate type in considering the time trends. For malignant melanoma in CTR, more than 50% of the cases are the NOS cases which prevents researchers from studying the time trends by histologic type. Most studies do not take into account the fact that NOS cases are a mixture of known histologic types. The time trends will depend on how to allocate the NOS cases back to the appropriate known histologic types. The well known concept of missing at random (MAR) (Rubin, 1976) is one of the assumptions that may be used for incorporating the NOS type. In this dissertation, our goal is to investigate the time trends of the three temporal effects, i.e. age, period and cohort, for male lung cancer and malignant melanoma in Connecticut. We will generalize the previous work on the age-period-cohort (APC) model by employing a technique for handling missing data, and creating a unified model for cancer time trend studies. This new model depends on a pre-specified set of allocation probability distributions. Identifiability and sensitivity of the time trends are discussed in general. The EM algorithm and the Markov Chain Monte Carlo (MCMC) sampling techniques are developed in the proposed model to make inferences on the time trends. Connecticut Tumor Registry Data will be used in our study to demonstrate the proposed methods.