Objective: We examined whether a breast cancer risk prediction model other than the Gail et al. model performs better at discriminating between women who will and who will not develop the disease.

Methods: We applied the two published versions of the Rosner and Colditz log-incidence model of breast cancer, developed on data from the Nurses' Health Study, to the estimation of 5-year risk for the period 1992 to 1997 in the same cohort. The first version contained reproductive factors only, and the second version contained a more extensive list of risk factors.

Results: Both versions of the model fit well. The ratio of expected to observed numbers of cases (E/O) in the first version was 1.00 (95% confidence interval [CI] 0.93-1.07); for the extended version the E/O was 1.01 (95% CI 0.94-1.09). The age-adjusted concordance statistic was 0.57 for the first model version and 0.63 for the extended version.

Conclusion: The discriminatory accuracy of the two versions was modest, although the addition of the variables in the extended version meaningfully increased the discriminatory accuracy of risk prediction over that found with the more parsimonious model.