The focus of this project is development and refinement of statistical procedures for the design and analysis of cancer screening and related studies. Problems under investigation include derivation and comparison of data analysis methods, assessment of case-control studies for screening evaluation, development of models of cancer screening, and approaches to the analysis of categorical data. Properties of case-control studies in the context of screening evaluation are being considered. It was found that the odds ratio from a case control study comparing screened versus not screened individuals is subject to bias and may underestimate or overestimate the impact of screening. Extensions were derived for a stage shift cancer screening model which uses the number of cancers diagnosed by stage in screened and control groups to estimate the number of cases shifted to earlier stages by screening and the mortality impact. The model was used to investigate age-specific breast cancer mortality in the HIP study. A simple model was developed to estimate dilution in a screening trial where follow-up continues after screening has ended. This was applied to both the Overall Analysis of all randomized individuals and a Limited Analysis of restricted subgroups of cases defined at certain intervals from study entry. Criteria were developed for comparability of the restricted case subgroups used in the Limited Analysis. Data from diagnostic testing and screening can often be analyzed using techniques for missing categorical data. Simple techniques have been developed for obtaining closed-form maximum likelihood estimates and their asymptotic variances and for finding the observed information matrix when using the EM algorithm with missing categorical data. A symbolic method has been developed for determining identifiable models in experimental designs involving missing categorical data. Methods have also been developed for analyzing data from randomized screening trials to estimate the benefit of screening unaffected by lead time bias and also the average lead time. Measurements on covariates are considered in assessing the existence of a possible dependence of lead time on the benefit or other variables.
