American Statistical Association New York City Metropolitan Area Chapter Mailman School of Public Health Columbia University Department of Biostatistics Colloquium

This work aimed to update the breast cancer risk prediction model of Gail et al. (1989). We first analyzed data from the Breast Cancer Detection Demonstration Project (BCDDP) to build a multivariate breast cancer relative risk model. Standard breast cancer risk predictors such as family history and age at menarche were collected on all subjects in an age-stratified case-control study nested within the BCDDP, but a new risk predictor, mammographic density (MD), were only available for about half the cases and controls. We developed a pseudo-likelihood method to accommodate missing MD measurements in the logistic regression analysis of both standard risk predictors and MD. Our method adapted a maximum likelihood method of Scott and Wild (1997) to incorporate a novel scheme for maximizing the contribution of complete covariates to the precision of inference. We showed that this procedure was substantially more efficient than a previously proposed weighted-likelihood method. We applied this method to the analysis of BCDDP data and obtained an updated relative risk model. We showed that this updated model, when combined with another model obtained from a different source of data in the BCDDP, led to a breast cancer absolute risk prediction model that had a higher discriminatory power than the original model of Gail et al. (1989).