STATISTICAL INNOVATIONS IN RISK MODELING

Brigham And Women'S Hospital, Boston MA

Investigators

Linked publications & trials

Abstract

The overarching goal of this project is to further epidemiologic knowledge of breast, colon, and ovarian cancers, by exploring innovative risk modeling approaches. Another goal is to further the understanding of the study of combined outcomes, such as total mortality. There are three aims subsumed under these goals. Our first aim is to construct log- incidence models of colon cancer and ovarian cancer, analogous to the work that has been done with breast cancer. Compared to conventional logistic regression models, these non-linear allows more efficient testing of complex time-dependent etiologic hypotheses They also permit the straightforward calculation of cumulative incidence, in addition to commonly-used relative risk measures. A second aim is to examine how well the above log-incidence models perform at clinical risk prediction at the individual level. Can these models accurately classify individual women with regard to disease status better than conventional logistic regression models, and can they perform well in an absolute sense? This issue of discriminatory accuracy is growing in importance; epidemiologic risk models are increasingly used by clinicians to provide patients with estimates of their personnel risk of disease, so that these individuals may make informed decisions may make informed decisions about prevention options. We will compare the discriminatory accuracy of the log- incidence models to the accuracy of conventional logistic regression models, and will compare results and conclusions obtained from ROC curve analyses to those obtained from traditional goodness-of-fit analyses which assess the difference between observed and expected values of (average) risk within specific subgroups of the sample population. It is likely that goodness-of-fit analyses are examination of combined endpoints, a common practice in epidemiology. The specific outcomes which comprise the combined outcome (e.g., mortality from different causes, in the combined outcome of total mortality) may evince different risk factor profiles. Statistical models which assume constancy of relative risk for a given risk factor are likely inappropriate when a combined outcome is being evaluated, though such models are often used. We will develop and expand the methodology of polychotomous logistic regression, which will permit the more accurate quantification of relative risk and benefits of exposures with respective to overall risk of adverse health outcome.