The purpose of this project is to conduct independent research in statistical methods and computer techniques with particular emphasis on those appropriate for analyzing data from clinical, diagnostic, and prevention trials and epidemiologic studies of cancer. Many of the problems studied under this project arise from the consultative activities of the Section. Among the current projects are: Comparison of Methods for Identifying Prognostic Factors and Predicting Survival for Patients with Colorectal Cancer. In collaboration with a working group of the American Joint Committee on Cancer, we have been comparing various statistical methods, one of which being Cox proportional hazards models, a) to determine the strengths and weaknesses inherent within each analytic method; b) to identify the appropriate evaluative statistical techniques by which these methods might be compared; and c) to conclude which method or methods were most successful in predicting patient survival using a dataset on patients with colorectal cancer. Interactive Statistical Programs. The Section continues to maintain and improve a group of interactive computer programs for efficient analysis of medical data, particularly those dealing with risk factors and prognostic factors using sophisticated multiple regression techniques and survival analysis. Power Analysis. The Section continues to maintain and improve an interactive program to compute statistical power and sample size requirements for a variety of experimental designs. Analysis of Diet Survey Data: Typical Consumption and Effects of Covariates. In collaboration with the University of Maryland, a parametric statistical model was constructed to model count data provided by 24-hour recall questionnaires. The model permits one to use regression models to relate abstention and average consumption to covariates such as income, race, day of week, or season. An important feature is that the model separates within and between person variation of count data. As a result, one can estimate the distribution of typical individual consumption, either for the entire population or for selected subpopulations. Biomarker Studies. Matched case-control studies are becoming increasingly popular for studying the association of biomarkers with cancer incidence. Methods of correcting relative risk estimates for the bias introduced by measurement error in biomarkers are being investigated. Corrections for measurement error are also being investigated for validation analysis of surrogate endpoints.
