The proposed work will develop, evaluate, and apply new statistical methodology to analyze case-control data. The new methods are designed to efficiently and inexpensively combine multivariate data from numerous large strata in fitting general relative risk models. These methods appreciably curtail high computing costs in using large-scale (mainframe) computers, compared to existing methods. They significantly improve slow computing times in using smaller time-sharing minicomputers or microcomputers with limited addressable memory.
The specific aims of the proposed work are: 1) to extend the newly developed method of modified-score logistic risk analysis to more general relative risk functions; 2) to develop and test improved estimators to reduce small-strata estimation bias and speed convergence to asymptotic behavior; 3) to develop methods to efficiently combine multivariate data from numerous large and small strate, with minimal bias and computational effort; 4) to develop analysis of residuals and measures of deviance for fitting nested models and indicating model departures; 50 to develop regression diagnostics for detecting outliers and data points with undue influence on parameter estimates; 6) to develop methods for estimating multivariate attributable risk in matched or stratified case-control data; and 7) to develop methods for classification and discrimination in these study designs. Monte-Carlo sampling experiments (computer simulations) using a variety of case-control study designs will be conducted to demonstrate the performance of the new methods in comparison to alternative methods. The new methodology will be used in fitting proportional hazards models to existing data for 1884 breast cancer patients and 3432 control patients in multivariate prediction of breast cancer risk. Attributable risk of breast cancer due to single and multiple factors in the presence of all others will be estimated. Multifactorial classification of women at high and low short-term and lifetime risk will be determined. Additivity and multiplicativity of risk predictors will be investigated. A portable, control-statement operated computer program to implement the new methodology will be developed and made available to others.
