CASE-COHORT STUDY OF MAMMOGRAPHY AND BREAST CANCER

点击翻译按钮获取中文摘要

Statistical techniques that reduce costly data collection would be of great value to cancer epidemiologists. Such methods include reduction of duration of follow-up and reduction of data collection on relatively uninformative subjects. Surrogate outcomes for cancer mortality could substantially reduce the duration of follow-up and the number of participants in large intervention studies. Prentice has proposed criteria for assessing the validity of surrogate outcomes for mortality. He has also proposed a case-cohort design which concentrates data collection on influential subjects and reduces data collection on those without the outcome of interest. This design allows for a proportional hazards analysis of time to failure outcomes. Nonetheless, the methods have not found widespread acceptance. The conditions for determining the validity of the surrogate outcomes are stringent and difficult to test. The case-cohort design is seldom used due to a computationally difficult variance estimator and lack of software to fit the model. This proposal demonstrates application of both methods to existing data concerning the efficacy of mammography screening in reducing breast cancer mortality. The case-cohort design will be shown to be an efficient method for evaluating cancer screening using observational data. It also allows testing the validity of using advanced breast cancer as a surrogate outcome in measuring screening efficacy. The population consists of 94,656 women aged 40 and older who were enrolled in a large regional HMO in the period 1982-88. There were 1,144 incident breast cancer cases including 126 who subsequently died of breast cancer. Follow-up will be extended to include an additional 100 deaths. The case-cohort analysis will include mortality and advanced disease as outcomes. Advanced disease is defined by different combinations of tumor stage and size. The validity of advanced disease as a surrogate for mortality will be tested by the statistical model. Software will be developed allowing interactive model fitting. The study results and software will be disseminated to epidemiologists and applied statisticians.