We developed models that predict individualized probabilities of developing breast, endometrial, and ovarian cancers among US white women aged 50+ y. We chose these three cancers because they share several risk factors, presumably reflecting a common hormonal etiology, and because management decisions may depend jointly on these risks.
To our knowledge, there is no other absolute risk model for endometrial cancer, despite the fact that it is the fourth most common cancer in women 
and its absolute risks are quite high, particularly in obese women (). Knowledge of endometrial cancer risk might inform decision-making about diagnostic workup, clinical management, surgical interventions, and the use of agents, such as tamoxifen or unopposed estrogen that increase endometrial cancer risk. Such a model might also aid in designing intervention trials to prevent endometrial cancer and in identifying women with elevated risk who might benefit from such interventions. The endometrial cancer model may also be useful in assessing the burden of that cancer in the general population, which may increase, as more than a third of all US women now have a BMI of 30 kg/m2
or higher 
. In combination with data on trends in the prevalence of obesity, one could use the model to investigate the extent to which the increasing prevalence of obesity accounts for the significant 2% per year increase in endometrial cancer incidence seen among white women 2006–2010 
Unlike most other models for breast cancer, our model includes factors that are potentially modifiable for the individual or in populations over time, such as alcohol consumption, BMI, and use of MHT. This model had slightly better discriminatory accuracy (AUC
0.58) in the validation data than the widely used BCRAT (“Gail model”) (AUC
0.56), which predicts breast cancer risk based on reproductive factors, number of breast biopsies, and atypical hyperplasia. However, despite this increase in AUC, the discriminatory accuracy of the breast cancer absolute risk model is still modest, and limits its clinical applicability, particularly for screening. Several breast cancer risk prediction models include non-modifiable risk factors such as family history (e.g., 
), mammographic density 
, and reproductive and medical factors such as age at menarche and number of breast biopsies 
. A few models for US women include potentially modifiable risk factors, such as BMI and alcohol use 
. Our model also incorporates use, and duration of use, of combined estrogen and progestin MHT and other types of MHT. Use of estrogen and progestin MHT had the second largest RR in our model, and had the same RR (RR
1.40) for a one category increase in duration as benign breast disease. Petracci et al. 
illustrate public health and counseling applications of a breast cancer model with modifiable risk factors in Italian women, and similar calculations could be performed for the US with the model we developed. These calculations could also aid in understanding the impact of increases in obesity on US breast cancer incidence.
For ovarian cancer, a model based on the NHS includes age at menopause, age at menarche, OC use, and tubal ligation 
To build the models, we combined data from two large prospective studies from well-characterized populations. We used another large cohort, the NHS cohort, to independently validate our models. The RR estimates agreed well with those in the NHS cohort for all three models. The discriminatory power, as assessed by the AUC, was 0.58 and 0.59 for the breast and ovarian cancer models, respectively. While these values indicate modest discriminatory ability, they are similar to those reported for other cancer risk models for these cancers 
. The AUC value for endometrial cancer was 0.67, which is larger than that seen for most models of cancer incidence.
The breast and ovarian model were well calibrated in the NHS cohort; however, because women in the NHS cohort were censored at the age of diagnosis of an in situ breast cancer, the breast model may have underestimated slightly. The endometrial model significantly over-predicted the number of endometrial cancers. This reflects the fact that the NHS cohort has considerably lower endometrial cancer rates () than those seen in SEER (). Further studies of the calibration of this model in additional cohorts would be desirable to assess its applicability to the general US population.
Well-calibrated risk models, even those with modest discriminatory accuracy, have several public health applications. These include designing cancer prevention trials, assessing the absolute burden of disease in the population and in subgroups, and gauging the potential absolute reductions in risk from preventive strategies. Using risk models to select individuals for screening or other interventions usually requires high discriminatory accuracy 
. Well-calibrated risk models with modest discriminatory accuracy can also aid in individual decision-making. Such models can provide realistic information on level of risk that is useful in making decisions, such as whether or not to have a mammogram 
. Such models are also useful in decisions on whether or not to take an intervention that has both beneficial and harmful health effects 
There are several potential limitations to our models. The exact number of breast biopsies, an important predictor in BCRAT, was not available in our cohorts, and the models were restricted to women aged 50+ y. Our RR models were built using white, non-Hispanic women and may not generalize to other races or ethnicities. We adjusted the age-specific endometrial and ovarian cancer incidence rates for the prevalence of hysterectomy and oophorectomy among US women estimated from population-based surveys, but some residual error may exist. Another limitation is that MHT use was probably more prevalent during the period of study of PLCO and NIH-AARP, from 1993 to 2005, than it is currently. While such changes in the prevalence of risk factors would not affect model calibration, they could reduce the variation of risk in the population and hence reduce the discriminatory accuracy of the models. Among the strengths of our models are large study sample size, nearly complete end point ascertainment, and information on most of the important risk factors.
Our models are not intended to predict the probability of the three cancers among women known to be at much higher than average risk, e.g., women with a mutation in BRCA1 or BRCA2 or with hereditary non-polyposis colorectal cancer (HNPCC). Each model is applicable to women without a prior diagnosis of that particular cancer, and thus in principle the breast cancer model can be applied to predict breast cancer risk for women with a prior diagnosis of any other cancer, including endometrial cancer. However, in applying the risk estimates, one needs to consider that a woman's risk may be altered by treatment for a previous cancer.
In conclusion, we developed and assessed models that project the probabilities of developing breast, endometrial, or ovarian cancer among white, non-Hispanic women aged 50+ y. These models might improve the ability to identify potential participants for research studies and assist in clinical decision-making related to the risks of these cancers.