Statistical Issues in Early Detection of Diseases
There is a growing interest in public health programs targeted at diagnosing chronic diseases earlier-especially breast, cervical, colorectal, ovarian, and prostate cancers. The goal is to diagnose these diseases earlier relative to the stage at diagnosis under usual care with the expectation that mortality will be reduced.In collaboration with Dr. Marvin Zelen, we have developed stochastic models using the natural history of disease to address issues that arise in screening programs. These models have helped to guide public health programs in the choice of examination schedules. The main features of such schedules are the initial age to begin examinations, the number of examinations, and the intervals between them. We have introduced two new ideas, threshold method and schedule sensitivity, that have been used to generate and compare screening examination schedules.We have further expanded the stochastic models by developing a series of probability equations to predict the mortality reduction associated with screening programs. Predicting mortality is a function of the characteristics of the case-finding process and the detection modality. The model makes two basic assumptions: first, that the disease is described by a progressive disease model and second, that potential benefit of early detection arises from a stage shift.The model can be used to predict the mortality benefit of early detection programs. More importantly, however, it can be used to assess proposed schedules for early detection programs which will enable policy makers to compare costs and mortality benefits. We have applied theoretical results mainly to the area of breast cancer screening, but will also apply results to other disease sites such as colon, lung, and prostate cancers.