As public awareness of breast cancer incidence rates and risk factors has grown, women increasingly are seeking information from clinicians on their individual breast cancer risk. It is believed that providing such information will aid women in making decisions regarding "lifestyle" factors such as alcohol consumption and postmenopausal hormone use, as well as making decisions about practices such as screening frequency and chemoprevention. The discriminatory accuracy of models of disease risk is rarely examined, yet is key to the use of epidemiologic risk equations for clinical or individual risk prediction. We propose to evaluate the discriminatory accuracy of various models of breast cancer risk in the Nurses' Health Study. We will construct various logistic regression models of breast cancer risk in this cohort, each model containing different combinations and categorizations of established breast cancer risk factors, and we will assess the ability of each model to accurately discriminate between women who did, and who did not, develop breast cancer by the end of the follow-up period. We will compare, for a given set of risk factors, the conventional logistic regression model to Rosner's nonlinear log-incidence model of breast tissue aging in terms of discriminatory ability. Finally, we will compare results and conclusions obtained from receiver operating characteristic curve and accuracy curve analyses to those obtained from traditional goodness-of-fit analyses. A detailed study of how well statistical models of breast cancer risk predict individual outcome has not yet been carried out, and is important for several reasons. First, if epidemiologic risk models are to be used to provide individuals with personal risk estimates, it is important to assess the accuracy of such prediction. There are obvious costs to providing women with estimates of risk that are little better than random guesses. Second, through our proposed examination of discriminatory accuracy, we will emphasize the distinction between estimating average risk within a group and predicting individual outcome. We will elucidate conditions under which a model of breast cancer risk "fits" the data well, in terms of accurately predicting number of cases within strata, but performs poorly at discrimination. Such clarification may suggest ways to improve communication regarding breast cancer risk. Third, the examination of the predictive accuracy of breast cancer risk models will help to quantitatively illustrate several important arguments made by Geoffrey Rose regarding ability to predict the future of individuals in a setting where we have good ability to estimate relative risk and average group risk. We will carry out the proposed work on the Nurses' Health Study cohort, which has a wealth of longitudinal data on breast cancer risk factors and excellent follow-up for breast cancer diagnoses. We will benefit from the knowledge gained over the past twenty years about breast cancer risk factors in this cohort.
If you are accessing this page during weekend or evening hours, the database may currently be offline for maintenance and should operational within a few hours. Otherwise, we have been notified of this error and will be addressing it immediately.
Please contact us
if this error persists.
We apologize for the inconvenience.
- The DCCPS Team.