Abstract IA04: Validation of a simplified Rosner-Colditz breast cancer incidence model in the California Teachers' Study

Abstract

Abstract Purpose: To validate a simplified breast cancer incidence model using baseline risk factors in an independent dataset Methods: We restricted the study population to comparable age ranges at baseline (age 47-79) (Nurses' Health Study (NHS), 1994, n=64,627; California Teachers' Study (CTS), 1995, n=31,386) We fit simplified Rosner-Colditz (RC) log incidence models using baseline risk factors and estimated both a 14-year risk model (1994-2008, 3597 cases) and a 4-year risk model (1994-1998, 1616 cases) based on NHS data. Both the 14-year and 4-year risk models were compared with the Gail model over the same time periods in the CTS population (14-year model, 1995-2009, 1786 cases; 4-year risk model, 1995-1999, 543 cases). We assessed performance using measures of discrimination based on AUC and calibration based on Poisson regression. Correlated AUC methods (Rosner and Glynn, 2009) were used to compare AUC's of competing models. Calibration was assessed by using relative risks from the RC and Gail models and absolute incidence rates from SEER. Results: Variables considered in the RC models were age; age at menopause (by type of menopause), menopausal status, age at 1st birth age at menarche, nulliparity, birth index, benign breast disease, duration of HRT use among current users by type of HRT, weight at age 18, change in weight from age 18 to baseline, separately by menopausal status and HRT use, height, alcohol consumption and family Hx of breast cancer. Age-adjusted AUC estimates in the NHS population were: (14-year risk model, RC model: 0.606 ± 0.005, Gail model: 0.563 ± 0.005, pdiff<0.001; 4-year risk model, RC model: 0.611 ± 0.007, Gail model: 0.566 ± 0.007, pdiff<0.001). Age-adjusted AUC estimates in the validation (CTS) population were: (14-year risk model, RC model: 0.580 ± 0.007, Gail model: 0.549 ± 0.007, pdiff<0.001; 4-year risk model, RC model: 0.605 ± 0.012, Gail model: 0.572 ± 0.012, pdiff=0.025). Calibration of the 14-year risk model indicated an estimated E/O ratio of 1.10 (95% CI = 1.05, 1.15) for RC; 1.08 (95% CI = 1.05-1.13) for Gail. Calibration of the 4-year risk model indicated an estimated E/O ratio of 1.16 (95% CI = 1.07-1.26) for RC; 1.15 (95% CI = 1.07-1.25) for Gail. Calibration results were similar using Poisson regression. Conclusion: The simplified RC model based on baseline risk factors is practical to use in a clinical setting and has a significantly higher AUC than the Gail model when validated in an external sample. AUC is better for short-term (4-year) vs. long-term risk prediction. Calibration is slightly off using both models and indicates that expected risks are slightly higher than observed risks for both short-term and long-term models. Citation Format: Bernard A. Rosner. Validation of a simplified Rosner-Colditz breast cancer incidence model in the California Teachers' Study. [abstract]. In: Proceedings of the AACR Special Conference: Improving Cancer Risk Prediction for Prevention and Early Detection; Nov 16-19, 2016; Orlando, FL. Philadelphia (PA): AACR; Cancer Epidemiol Biomarkers Prev 2017;26(5 Suppl):Abstract nr IA04.