Validating a model for predicting breast cancer and nonbreast cancer death in women aged 55 years and older

Abstract

Background

To support mammography screening decision making, we developed a competing-risk model to estimate 5-year breast cancer risk and 10-year nonbreast cancer death for women aged 55 years and older using Nurses’ Health Study data and examined model performance in the Black Women’s Health Study (BWHS). Here, we examine model performance in predicting 10-year outcomes in the BWHS, Women’s Health Initiative-Extension Study (WHI-ES), and Multiethnic Cohort (MEC) and compare model performance to existing breast cancer prediction models.

Methods

We used competing-risk regression and Royston and Altman methods for validating survival models to calculate our model’s calibration and discrimination (C index) in BWHS (n = 17 380), WHI-ES (n = 106 894), and MEC (n = 49 668). The Nurses’ Health Study development cohort (n = 48 102) regression coefficients were applied to the validation cohorts. We compared our model’s performance with breast cancer risk assessment tool (Gail) and International Breast Cancer Intervention Study (IBIS) models by computing breast cancer risk estimates and C statistics.

Results

When predicting 10-year breast cancer risk, our model’s C index was 0.569 in BWHS, 0.572 in WHI-ES, and 0.576 in MEC. The Gail model’s C statistic was 0.554 in BWHS, 0.564 in WHI-ES, and 0.551 in MEC; IBIS’s C statistic was 0.547 in BWHS, 0.552 in WHI-ES, and 0.562 in MEC. The Gail model underpredicted breast cancer risk in WHI-ES; IBIS underpredicted breast cancer risk in WHI-ES and in MEC but overpredicted breast cancer risk in BWHS. Our model calibrated well. Our model’s C index for predicting 10-year nonbreast cancer death was 0.760 in WHI-ES and 0.763 in MEC.

Conclusions

Our competing-risk model performs as well as existing breast cancer prediction models in diverse cohorts and predicts nonbreast cancer death. We are developing a website to disseminate our model.

The 2023 US Preventive Services Task Force (USPSTF) recommends biennial mammography screening for women aged 40-74 years but gives an “I” recommendation (insufficient evidence) for women aged 75 years and older because the randomized clinical trials of mammography screening excluded women aged 75 years and older (1). The USPSTF further recommends that clinicians be prepared to discuss I graded services in the context of shared decision making (2). The American Cancer Society (ACS) recommends screening women aged 45-54 years annually and women aged 55 years and older biennially who have at least a 10-year life expectancy (3); however, ACS notes that women aged 55 years and older may choose annual screening. The ACS further encourages clinicians to engage women aged 75 years and older in shared decision making because of the uncertain benefit of screening these women and the chance of harm including anxiety, complications from cancer work-up, and overdiagnosis (detection of nonlethal tumors), leading to overtreatment (4). The American College of Obstetrics and Gynecology and the American College of Physicians also recommend not screening women with less than a 10-year life expectancy at average breast cancer risk (5,6). Despite published guidelines, many women aged 55 years and older are screened annually with mammography with little understanding of the risks, and approximately 50% of women with less than a 10-year life expectancy are screened despite little chance of benefit; clinicians find it particularly challenging to discuss screening cessation with women aged younger than 75 years with less than a 10-year life expectancy (7-11).

Shared decision making is a process of communication where clinicians and patients work together to make health-care decisions (2,12). Shared decision making may increase patient engagement and satisfaction with care and lead to more equitable care (13,14). For high-quality shared decision making around mammography screening, experts and clinicians recommend considering older women’s breast cancer risk, life expectancy, and values and preferences (4,15,16). However, shared decision making around mammography screening has been challenging to implement because of a lack of validated prediction models that simultaneously estimate older women’s individualized breast cancer risk and 10-year life expectancy—information much needed to help individualize the benefits and harms of screening (16).

To inform screening decisions, we developed a competing-risk regression model that predicts 5-year breast cancer risk and 10-year nonbreast cancer death in women aged 55 years and older using Nurses’ Health Study (NHS) data (two-thirds for development, one-third for validation) (17,18). The model was developed for women aged 55 years and older because nearly all women are postmenopausal by this age and breast cancer risk factors may differ between pre- and postmenopausal women. Also, life expectancy declines with age so few women aged younger than 55 years have less than a 10-year life expectancy (19).

We previously examined model performance among women aged 55 years and older in the Black Women’s Health Study (BWHS) and found that our model performed better in predicting breast cancer when variables that only predicted nonbreast cancer death were not included and when body mass index (BMI) was included as a continuous variable (18); therefore, our final model includes these changes. When predicting 10-year nonbreast cancer death, our final model’s C index was 0.79 in the NHS validation cohort and 0.77 in BWHS, similar to the performance of models that predict overall mortality in older adults (11,12). When predicting 5-year breast cancer risk, our final model’s C index was 0.61 in the NHS validation cohort and 0.57 in BWHS, similar to the Breast Cancer Risk Assessment Tool’s (hereafter, Gail model) reported performance (20,21).

Before testing our model in clinical practice, we aimed to 1) validate our model’s performance in 2 additional diverse cohorts, the Women’s Health Initiative–Extension Study (WHI-ES) and the Multiethnic Cohort (MEC) (22,23); 2) to expand our model to predict 10-year breast cancer risk to match the 10-year time period our model uses to predict nonbreast cancer death; and 3) to formally compare our model’s performance with that of the Gail model and the International Breast Cancer Intervention Study (IBIS) model (24-26), because both are recommended in primary care and include risk estimates for women aged 75 years and older (27).

Methods

To examine our model’s performance, we included data from postmenopausal women without a history of invasive breast cancer who returned the 2004 NHS questionnaire (n = 83 330), 2009 BWHS questionnaire (n = 17 380), participated in the 2005 WHI-ES (n = 106 894), or returned MEC’s third questionnaire (completed between 2003 and 2008, n = 49 668); Figure 1 presents each cohort’s sample population. We chose these study periods because they allowed at least 10-year follow-up for nearly all participants and most women had stopped using menopausal hormone therapy, similar to current practice.

Figure 1.

Sample population. ^aThe number of total MEC participants was 215 903, and 96 990 men were excluded. ^bNHS: died before 2004; BWHS: died before 2009; WHI-ES: died before extension study 1 (2005); MEC: died before questionnaire 3 (2003-2008). ^cNHS: did not return 2004 questionnaire; BWHS: did not return 2009 questionnaire; WHI: chose not to participate in extension study; MEC: did not return questionnaire 3 (1 participant had inconsistent death and questionnaire date). BWHS = Black Women’s Health Study; MEC = Multiethnic Cohort Study; NHS = Nurses’ Health Study; WHI-ES = Women’s Health Initiative–Extension Study.

Each cohort is described in Supplementary Methods 1 (available online). In brief, NHS is a longitudinal study of 121 738 female nurses aged 30-55 years at entry in 1976; 97% were White women (28). BWHS is a longitudinal study of 59 000 self-identified Black women aged 21-69 years at entry in 1995 (29). WHI was a multicenter study that enrolled postmenopausal women aged 50-79 years to clinical trials or an observational study from 1993 to 1998 (22). Enrollment of racial and ethnic minority groups was proportionate to that of US women aged 50-79 years in 1990 (18.2%). Follow-up after March 31, 2005, involved surviving participants who provided written informed consent for follow-up through December 30, 2010 (n = 115 396, WHI-ES1) and over an open-ended subsequent period (n = 79 572, WHI-ES2); at least 80% of surviving participants provided consent on each occasion. MEC is a longitudinal study following 215 251 male and female residents of Hawaii and Los Angeles, California, aged 45-75 years at entry in 1993 and included Japanese American, Native Hawaiian, African American, Latino, and White individuals (23).

Participants in all cohorts provided lifestyle and medical history information through mailed questionnaires. For the present analysis, participants were aged 57-85 years (NHS), 55-85 years (BWHS), 56-90 years (WHI-ES), and 55-90 years (MEC) at the start of follow-up. The oldest participants at the end of follow-up were aged 95 years (NHS and BWHS) to 99 years (WHI-ES and MEC). We excluded women with a history of cancer (except nonmelanoma skin cancer) when predicting breast cancer risk because second diagnoses of cancer are not confirmed in NHS. Analyses were completed using SAS 9.4 software.

Outcomes

Detailed definitions of outcomes and risk factors are in Supplementary Methods 2 (available online). Briefly, cause of death for all cohorts was obtained through use of state-issued death certificates. Additional cause of death information on NHS, BWHS, and WHI-ES cohorts was obtained from the National Death Index, family and friends, and post office records. WHI-ES also obtained information using Medicare and Medicaid databases, cancer registries, hospital records, and large health maintenance organization databases, and causes of death were physician adjudicated. Nearly all self-reported breast cancers were confirmed in each cohort, either through medical record review or through state cancer registries.

Model validation

The competing-risk regression model in the NHS development cohort (n = 55 553), the NHS validation cohort (n = 27 777), and the BWHS has previously been described (18). In brief, of 18 breast cancer risk factors available in NHS, 9 improved model performance and were statistically significant at a P value less than .05 (17). For nonbreast cancer death, of 60 mortality risk factors considered, 21 provided the best model fit using best-subsets regression (30,31). Together, 25 variables were used in assessing risk (4 for breast cancer only, 16 for nonbreast cancer death only, and 5 for both). Supplementary Methods 3 (available online) presents our model’s risk assessment questionnaire.

Here, we examined our final competing-risk model’s performance in predicting 5-year breast cancer risk and 10-year nonbreast cancer death in WHI-ES and MEC. We also examined our model’s performance in predicting 10-year breast cancer risk. To do so, we repeated the model in the NHS development cohort with 10-years as the outcome and examined the model’s performance in the NHS validation cohort, BWHS, WHI-ES, and MEC. For these analyses, we censored participants at the end of WHI-ES1 who did not participate in WHI-ES2.

To test whether differences between risk factor hazard ratios (HRs) were statistically significant between the NHS development cohort and the other cohorts, we used the normal approximation z-test, which takes into account within-cohort standard error. To examine our model’s discrimination in predicting 5- and 10-year breast cancer risk and 10-year nonbreast cancer death, we calculated our model’s C index (C for concordance) in the validation cohorts using risk factor regression coefficients from our competing-risk model in the NHS development cohort. Concordance is based on the idea that for any pair of observed survival times for the i^th and j^th participant at t_i and t_j, where t_{i <} t_j, if the predicted survival times from the model also show the same order or predicted survival times, then we have concordance. The ratio of concordant pairs to the total number of eligible pairs of survival times estimates the C index. We used Kremer’s SAS software macro based on work by Harrell et al., Pencina et al., and Wolbers et al. to determine the C index (the ratio of all concordant pairs to the total number of evaluable pairs); Supplementary Methods 4 (available online) for additional details on how participants experiencing a competing event were treated in calculating the C index (32-35). Risk factors not assessed in a cohort were not considered in estimating participant risk. Participants missing information on assessed risk factors were excluded.

We examined our model’s calibration in predicting 5- and 10-year breast cancer risk and 10-year nonbreast cancer death in the validation cohorts based on Royston and Altman’s methods for validating models using survival analysis (36). Specifically, we compared the expected risk of each outcome (eg, 10-year breast cancer risk) estimated by the cumulative incidence function from our competing-risk model with the observed risk for each outcome estimated using the nonparametric estimation of cumulative incidence function within 5 risk groups (risk quintiles from the NHS development cohort).

We repeated these methods to examine model performance by age (55-74 years, 75 years and older). Supplementary Methods 4 (available online) includes additional details on the validation methods used.

Comparing models

We used Gail model tool’s SAS Macro(v4) (37) and IBIS(v8) (38) to also project absolute breast cancer risk at 5 and 10 years. When predicting 10-year risk in WHI-ES, we excluded women who did not participate in WHI-ES2. For the Gail model, we excluded participants missing risk factor information; IBIS allows missing risk factor information. We used Rosner and Glynn’s (39) methods to determine model C statistics and standard errors in each cohort. To assess calibration, we compared the expected number of breast cancers among participants based on the models’ calculated absolute risk with the observed number in each cohort. We calculated 95% confidence intervals (CIs) of expected-to-observed ratios using the Poisson variance for the logarithm of the observed number of breast cancers. To allow more direct comparison of our model to these models, in a sensitivity analysis only, we performed a logistic regression using our model’s predictors, applying the NHS development cohort logistic regression coefficients to each validation cohort to calculate our model’s C statistic. We repeated these analyses stratifying by age (55-74 years, 75 years and older).

Examples

Using our model’s regression coefficients and baseline survival from the NHS development cohort, we calculated absolute risk estimates for 5- and 10-year breast cancer risk, 10-year breast cancer death, and 10-year nonbreast cancer death for hypothetical women aged 57 years (the youngest women in NHS), 65 years, 75 years, and 80 years, with different breast cancer and nonbreast cancer death risk factors. We also presented absolute risk estimates for these women using baseline survival and the regression coefficients from the validation cohorts to show variation.

Results

Table 1 presents the raw and age-standardized prevalence of our model’s risk factors and participant race and ethnicity and education for each cohort. Supplementary Table 1 (available online) compares sociodemographic characteristics between participants missing risk factor information in each cohort with participants with complete data. Supplementary Table 2 (available online) presents the prevalence of the Gail model’s and IBIS’s risk factors within each cohort. On average, BWHS participants were younger at the start of follow-up, more likely to be obese and to have diabetes and hypertension, and less likely to have cancer or to use menopausal hormone therapy than participants in other cohorts. NHS participants were less likely to be nulliparous than other cohort participants. MEC and BWHS participants were more likely to abstain from alcohol compared with NHS and WHI-ES participants. WHI-ES participants were less likely to report functional impairments than NHS participants; BWHS and MEC did not assess functional impairment. Ten-year breast cancer incidence rates ranged from 3.5 (NHS) to 4.2 (WHI-ES, age-standardized) per 1000 person-years. Ten-year nonbreast cancer death incidence rates ranged from 15.4 (WHI-ES, age-standardized) to 22.5 (BWHS, age-standardized) per 1000 person-years, whereas 10-year breast cancer death incidence rates ranged from 0.2 to 0.3 across cohorts.

Table 1.

Baseline characteristics for factors in our competing risk regression model in each cohort^a

Factors in our final model	NHS (n = 83 330)	BWHS Crude (n = 17 380)	BWHS Age-standardized (n = 17 380)	WHI-ES Crude (n = 106 894)	WHI-ES Age-standardized (n = 106 894)	MEC Crude (n = 49 668)	MEC Age-standardized (n = 49 668)
Age, %
55-59 y	7.0	36.5	—	4.1	—	14.8	—
60-64 y	21.4	28.3	—	18.7	—	19.7	—
65-69 y	22.4	17.1	—	24.4	—	18.6	—
70-74 y	20.9	9.9	—	22.9	—	17.5	—
75-79 y	18.3	5.8	—	18.8	—	15.8	—
80 and older y	10.1	2.4	—	11.1	—	13.5	—
Mean (SD), y	70.1 (7.0)	63.7 (6.7)	—	71.0 (6.9)	—	69.6 (8.3)	—
Highest self-reported body mass index in past 10 years, %
<20 kg/m2	2.7	0.5	0.5	1.2	1.2	4.8	4.6
20-22.4 kg/m2	11.0	3.0	3.0	7.0	7.0	14.2	14.0
22.5-24.9 kg/m2	19.2	8.5	9.2	15.2	15.2	20.5	20.4
25-29.9 kg/m2	36.4	31.6	33.8	35.4	35.2	33.8	34.1
30-34.9 kg/m2	19.3	28.3	28.0	22.9	22.8	16.2	16.4
35-39.9 kg/m2	7.2	14.5	13.5	10.6	10.6	6.3	6.3
≥40 kg/m2	4.1	12.6	10.7	7.3	7.4	4.1	4.0
Unknown	0.1	1.0	1.3	0.5	0.5	0.1	0.1
Mean (SD), kg/m2	28.1 (5.7)	32.1 (7.0)	31.6 (6.5)	29.9 (7.3)	30.0 (7.3)	27.4 (6.0)	27.4 (5.9)
Average alcohol use per day (highest average use in past 10 years), g/d, %b
None	37.0	50.4	54.2	39.4	39.3	66.1	66.5
1-4.9	22.5	26.0	23.9	23.8	24.0	15.7	15.6
5-14.9	17.4	13.7	12.6	19.0	19.1	9.6	9.4
≥15	13.2	9.9	9.3	15.9	15.8	7.8	7.7
Unknown	9.9	—	—	1.9	1.8	0.8	0.8
Cigarette use, %
Never	44.7	53.4	50.5	51.4	51.2	58.2	58.0
Current	7.8	10.1	8.0	4.1	4.3	5.5	5.4
Past	47.3	36.4	41.6	44.5	44.5	34.9	35.2
Unknown	0.2	—	—	—	—	1.4	1.4
Limited from walking several blocks, %c,d
Not at all	61.7	—	—	69.0	70.0	—	—
A little or a lot	32.9	—	—	30.9	30.0	—	—
Unknown	5.4	—	—	0.1	0.1	—	—
Limited in bathing or dressing oneself, %c,d
Not at all	88.1	—	—	94.0	94.1	—	—
A little or a lot	6.6	—	—	6.0	5.8	—	—
Unknown	5.3	—	—	0.1	0.1	—	—
Usual walking pace outdoors, %d
Unable to walkc,e	2.7	—	—	0.4	0.4	—	—
Slow or average, <3 mph	72.9	62.0	65.2	74.6	74.1	—	—
Brisk or very brisk, ≥3mph	19.8	27.5	23.3	18.0	18.7	—	—
Unknown	4.6	10.6	11.6	6.9	6.9	—	—
High blood pressure, %f	60.3	68.7	75.4	54.0	53.2	58.9	60.0
Depression, %d	20.4	22.2	19.4	19.9	20.3	—	—
Hip fracture, %	2.1	0.8	1.1	1.7	1.6	0.8	0.8
Parkinson disease, %	0.6	0.2	0.3	0.6	0.6	0.5	0.5
Myocardial infarction, %	5.6	4.3	6.5	3.5	3.3	4.3	4.4
Congestive heart failure, %d	3.3	3.2	4.6	2.5	2.4	—	—
Stroke or transient ischemic attack, %	7.5	4.2	5.7	4.4	4.2	5.2	5.3
Emphysema or asthma, %	18.6	18.3	18.0	14.2	14.2	15.3	15.1
Diabetes, %	12.1	24.2	28.5	10.0	9.9	17.4	17.9
Dementia, %	1.2	0.2	0.5	0.5	0.5	0.1	0.1
Kidney disease, %d	0.6	1.9	2.3	0.3	0.3	—	—
Cancer, %g	12.3	5.7	7.4	9.2	9.0	10.1	10.2
Age at menopause, %, y
Younger than 45	10.6	19.3	21.2	20.4	20.6	28.9	29.2
45-49	23.6	20.1	19.2	24.9	25.8	26.5	26.4
50-54	56.2	27.4	25.9	36.4	36.1	34.2	34.2
55 and older	8.7	8.3	10.6	12.8	12.0	9.2	9.2
Hysterectomy, age unknownh	—	22.2	21.9	—	—	—	—
Unknown	1.0	2.7	1.2	5.5	5.5	1.1	1.0
Mammogram in past 2 years, %
No	12.1	14.0	15.2	15.2	15.1	27.2	26.9
Yes	79.5	86.0	84.9	84.6	84.8	72.9	73.1
Unknowni	8.4	—	—	0.2	0.2	—	—
Number of breast biopsies, %d
0	73.2	67.0	65.7	71.8	71.8	—	—
1	23.5	21.3	21.3	17.5	17.5	—	—
≥2	3.3	11.7	13.1	10.7	10.7	—	—
Postmenopausal hormone use, %
Never	22.6	40.7	37.5	31.8	31.1	24.4	23.9
Current estrogen plus progestin user <5 years	0.4	0.7	0.2	0.3	0.3	5.7	4.6
Current estrogen plus progestin user ≥5 years	2.7	1.1	0.8	1.4	1.5	3.9	4.6
Current estrogen-alone user <5 years	0.9	1.6	0.7	0.5	0.5	2.3	1.1
Current estrogen-alone user ≥5 years	9.3	6.3	5.6	6.2	6.3	7.6	8.6
Past estrogen plus progestin user <5 years	14.9	12.8	12.5	13.1	13.4	14.7	14.6
Past estrogen plus progestin user ≥5 years	15.9	5.6	7.4	20.8	21.1	9.6	10.5
Past estrogen-alone user <5 years	7.1	12.2	12.2	6.8	6.8	11.6	11.4
Past estrogen-alone user ≥5 years	13.8	13.8	18.1	19.2	19.1	11.5	12.0
Unknown	12.4	5.3	4.8	—	—	8.8	8.6
Age at first birth, y and parity, %j,k
Nulliparous	5.3	16.0	12.7	11.5	11.9	12.7	12.1
Younger than 25, 1-2 children	14.1	32.2	35.5	15.6	16.2	19.9	19.3
Younger than 25, ≥3 children	35.1	27.8	28.3	35.5	34.9	40.8	42.6
25-29, 1-2 children	15.0	11.9	10.8	10.2	10.5	10.0	9.7
25-29, ≥3 children	20.1	2.9	4.0	11.5	11.0	7.5	7.5
30 and older, 1-2 children,	5.8	7.1	6.3	5.2	5.3	5.6	5.2
30 and older, ≥3 children	2.8	0.5	0.6	2.0	1.9	1.3	1.3
Unknown	1.7	1.6	1.9	8.4	8.2	2.3	2.3
First-degree female relatives with breast cancer and age at diagnosis, %, yl
None	82.0	80.1	79.8	86.5	86.7	82.9	82.9
1 and aged younger than 50	4.2	3.8	4.2	2.2	2.2	4.2	4.2
1 and aged 50 and older	11.6	14.2	14.0	10.1	10.0	10.4	10.4
≥2 and at least one aged younger than 50	1.1	0.6	0.7	0.5	0.5	0.9	0.9
≥2 and aged 50 and older	1.1	1.3	1.3	0.7	0.7	1.6	1.6
Outcomes
5-year breast cancers per 1000 person-years	3.6	3.5	3.8	4.2	4.2	3.7	3.8
10-year breast cancers cases per 1000 person-years	3.5	3.3	3.6	4.3	4.2	3.7	3.7
10-year breast cancer deaths per 1000 person-years	0.3	0.2	0.3	0.3	0.2	0.3	0.3
10-year nonbreast cancer deaths per 1000 person-years	22.2	11.5	22.5	15.7	15.4	19.1	19.7
Median survival, y
Nonbreast cancer death	6.2	6.0	6.1	6.0	6.1	6.2	6.1
Breast cancer death	6.3	6.3	6.2	6.8	6.9	7.1	7.0
All cause death	6.2	6.0	6.1	6.0	6.1	6.2	6.2
Sociodemographics, %
Asian, Pacific Islander	0.9	0.0	0.0	2.1	2.2	36.5	36.0
Hispanic	0.9	0.9	0.1	3.0	3.2	17.1	18.0
Native American	0.2	0.0	0.0	0.4	0.4	—	—
Non-Hispanic Black	1.7	99.1	99.9	7.3	7.5	13.1	13.1
Non-Hispanic White	96.2	—	0.0	86.0	85.5	26.4	26.0
Other	—	0.0	0.0	1.0	1.1	6.8	6.9
Education, %, ym
<12	—	2.7	4.0	3.4	3.7	12.2	12.5
12	—	16.4	19.1	16.3	16.0	25.6	26.2
13-15	—	27.8	26.2	37.1	37.0	30.5	30.4
16	91.7	19.5	17.1	11.5	11.6	15.4	14.9
≥17	8.3	33.6	33.5	30.6	31.0	15.4	15.0
Unknown	0.0	0.1	0.1	0.7	0.7	1.0	1.0

^aNurses’ Health Study (NHS) included participants who completed the 2004 questionnaire. Black Women’s Health Study (BWHS) included participants who completed the 2009 questionnaire. Women’s Health Initiative–Extension study (WHI-ES) began in 2005. Multiethnic Cohort (MEC) included participants who completed questionnaire 3 (2003-2008). — = not applicable.

^bA standard drink is any drink that contains about 14 g of pure alcohol (12 oz of beer, 5 oz of wine, or 1.5 oz of liquor).

^cBWHS does not ask participants about walking several blocks, limitations in bathing or dressing oneself, or inability to walk.

^dMEC questionnaire 3 did not assess mobility, function, walking pace, depression, congestive heart failure, kidney disease, or breast biopsy.

^eWHI did not assess inability to walk; those who reported wheelchair use were placed in the unable to walk category.

^fHealth conditions were self-reported.

^gExcluded breast and nonmelanomatous skin cancers.

^hParticipants with hysterectomy and age at menopause unknown in BWHS were placed in the aged 45-49 years category for all analyses; sensitivity analyses were performed showing no statistically significant difference when placed in any other age category.

ⁱParticipants missing data on mammography use in NHS had completed a short version of the 2004 questionnaire.

^jParticipants who were parous with an unknown number of children were categorized as having 1-2 children; <1% of participants in each cohort had an unknown number of children.

^kThe categories for age at first birth in MEC were younger than 26, 26-30, 31 years or older.

^lAge at diagnosis in WHI was classified as younger than 45 or 45 years and older. Relatives with an unknown age at diagnosis were categorized as having been diagnosed at aged 50 years and older (45 years an older in WHI); <1% of participants had a relative with unknown age at diagnosis.

^mParticipants in NHS who had a master’s or doctorate degree were placed in the ≥17 years category. All others were placed in the 16 years category, as registered nurses (22.8% of these specifically reported a bachelor’s degree).

Model validation in predicting breast cancer risk

Breast cancer risk factor hazard ratios were similar across cohorts (ie, no statistically significant differences at P < .01; Table 2; Supplementary Table 3, available online). When applying development cohort regression coefficients to the validation cohorts and predicting 10-year breast cancer risk, our model’s C index was 0.597 (95% CI = 0.576 to 0.619) in the NHS validation cohort, 0.569 (95% CI = 0.541 to 0.597) in BWHS, 0.572 (95% CI = 0.562 to 0.582) in WHI-ES, and 0.576 (95% CI = 0.560 to 0.591) in MEC. Results were similar after stratifying by age, but the C index was higher for women aged 75 years and older in BWHS, WHI-ES, and MEC (Table 2;Supplementary Table 3, available online). Results were similar when predicting 5-year breast cancer risk (Supplementary Tables 4 and 5, available online). The expected-to-observed cumulative incidence function for 5- (Supplementary Table 6, available online) and 10-year (Figure 2;Supplementary Table 7, available online) breast cancer risk was similar within each risk group for each cohort.

Table 2.

Competing risk regression model modeling 10-year cumulative risk of breast cancer in each cohort^a,b,c,d

Modeling 10-year breast cancer risk	NHS development cohort		NHS validation cohort		BWHS		WHI-ESe		MEC
	HR (n = 37 628)	P	HR (n = 18 980)	P	HR (n = 13 247)	P	HR (n = 82 634)	P	HR (n = 39 206)	P
Observed No. of breast cancer cases	1353		683		400		3128		1342
Factors in our final model
Age, per year increase	1.00	.53	0.99	.20	1.01	.26	1.00	.12	1.00	.25
Highest self-reported body mass index in past 10 years, per kg/m2 increase	1.03	<.001	1.01	.04	1.02	.02	1.01a	<.001	1.04	<.001
Average alcohol use per day, highest average use in past 10 years
None	1	—	1	—	1	—	1	—	1	—
1-4.9 g/d	1.09	.23	0.99	.94	0.88	.33	1.07	.17	0.90	.17
5-14.9 g/d	1.25	.003	1.02	.87	0.96	.81	0.98a	.70	0.96a	.66
≥15 g/d	1.35	<.001	1.29	.02	1.45	.01	1.20	<.001	1.18	.09
Age at menopause, y
Younger than 45	0.85	.16	0.51	<.001	1.02	.9	0.97	.58	1.07	.40
45-49	1	—	1	—	1	—	1	—	1	—
50-54	1.11	.12	1.02	.84	1.28	.07	1.06	.21	1.12	.11
55 and older	1.36	.002	1.17	.26	1.24	.29	1.01a	.93	1.29	.01
Mammogram in past 2 years
No	1	—	1	—	1	—	1	—	1	—
Yes	1.13	.20	1.07	.61	0.78a	.07	1.15	.02	1.21	.006
No. of breast biopsies
None	1	—	1	—	1	—	1	—	—	—
1	1.37	<.001	1.38	<.001	1.35	.01	1.26	<.001	—	—
≥2	1.48	.002	1.80	<.001	1.29	.09	1.50	<.001	—	—
Age at first birth, y, and parity
Nulliparous	1.14	.35	1.47	.04	0.97	.86	1.12	.08	1.17	.09
Younger than 25, 1-2 children	1	—	1	—	1	—	1	—	1	—
Younger than 25, ≥3 children	1.01	.94	1.16	.25	0.99	.93	0.97	051	1.00	.98
25-29, 1-2 children	1.06	.55	1.25	.12	1.29	.11	1.12	.10	0.99	.90
25-29, ≥3 children	0.97	.78	1.07	.63	1.21	.51	1.20	.007	1.03	.81
30 and older, 1-2 children	1.08	.56	1.70	.003	1.40	.07	1.15	.10	1.17	.21
30 and older, ≥3 children	1.09	.63	0.84	.56	2.13	.14	1.25	.07	1.24	.36
First-degree relatives with history of breast cancer and age at diagnosis
None	1	—	1	—	1	—	1	—	1	—
1 and aged younger than 50 years	1.64	<.001	1.35	.09	1.50	.07	1.46	<.001	1.27	.06
1 and aged 50 years and older	1.43	<.001	1.34	.008	1.29	.06	1.42	<.001	1.43	<.001
≥2 and at least one aged younger than 50 years	2.08	<.001	3.38	<.001	1.36	.6	1.51	.06	2.12	<.001
≥2 and aged 50 years and older	1.82	.005	1.62	.12	0.87	.78	2.46	<.001	1.63	.009
Postmenopausal hormone use
Never	1	—	1	—	1	–	1	–	1	–
Current estrogen and progestin user <5 years	1.70	.14	0.80	.76	2.10	.11	0.80	.53	1.26	.05
Current estrogen and progestin user ≥5 years	2.09	<.001	2.07	<.001	2.02	.05	1.72	<.001	1.62	<.001
Current estrogen-alone user <5 years	1.48	.10	1.32	.44	1.49	.25	1.00	.99	1.15	.44
Current estrogen-alone user ≥5 years	1.16	.14	1.13	.38	1.22	.34	0.93	.42	1.01	.96
Past estrogen and progestin user <5 years	0.99	.89	0.86	.23	1.04	.81	1.04	.50	1.03	.77
Past estrogen and progestin user ≥5 years	1.17	.05	1.11	.38	1.38	.12	1.10	.06	1.30	.007
Past estrogen-alone user <5 years	0.95	.69	0.85	.37	1.00	.99	0.84	.05	1.01	.91
Past estrogen-alone user ≥5 years	0.88	.20	1.01	.93	0.84	.33	1.01	.90	1.05	.67
C index (95% CI) when using risk factor regression coefficients from the NHS development cohortf	0.606 (0.591 to 0.621)		0.597 (0.576 to 0.619)		0.569 (0.541 to 0.597)		0.572 (0.562 to 0.582)		0.576 (0.560 to 0.591)

^aIndicates that the hazard ratio for between-study heterogeneity is statistically ignificantly different from NHS development cohort hazard ratio (P < .05); no P values were <.01. BWHS = Black Women’s Health Study; CI = confidence interval; MEC = Multiethnic Cohort Study; NHS = Nurses’ Health Study; WHI-ES = Women’s Health Initiative–Extension Study; HR = subdistribution hazard ratio from Fine-Gray model; — = not applicable.

^bThe model predicting 10-year risk of breast cancer has a competing risk of nonbreast cancer death.

^cWe performed a complete case analysis in which all participants with missing data were excluded.

^dWe excluded women from with a history of cancer because the NHS cohort does not confirm secondary cancers.

^eWomen in WHI-ES who did not participate in the second extension study and therefore did not have 10 years of breast cancer follow-up were censored at 5 years.

^fC index for each cohort uses the regression coefficients from the NHS development cohort.

Figure 2.

Calibration of our model for predicting 10-year breast cancer risk in each cohort by risk group.^aThese analyses included women with complete data. Expected breast cancer risk was calculated using the cumulative incidence function (CIF) from our competing risk breast cancer prediction model, and observed breast cancer risk was calculated using the nonparametric estimation of CIF within 5 risk groups (risk quintiles from the NHS development cohort). BWHS = Black Women’s Health Study; MEC = Multiethnic Cohort Study; NHS = Nurses’ Health Study development cohort; WHI-ES = Women’s Health Initiative–Extension Study.

Comparing breast cancer prediction models

Our model’s C statistic was higher than the Gail model’s or IBIS’s in all cohorts when predicting 5- or 10-year breast cancer risk except when predicting 5-year risk in WHI-ES (Table 3; receiver operating characteristics curve plots demonstrating each model’s performance in each cohort are presented in Supplementary Figures 1 and 2, available online). The Gail model overpredicted breast cancer at 5 and 10 years in the NHS validation cohort, regardless of participant age, and at 10 years in BWHS. IBIS overpredicted breast cancer risk in BWHS but underpredicted risk in MEC at 10 years and in WHI-ES at 5 and 10 years, regardless of participant age (Supplementary Table 8, available online).

Table 3.

Calibration and discrimination of the Gail model, IBIS, and our breast cancer prediction model in each cohort at 5 and 10 years^a

		Gail model			IBIS model			Our model
Follow-up time	Cohort	No.a	Expected-to-Observed ratio (95% CI) b	C statistic (95% CI)c	No.a	Expected-to-Observed ratio (95% CI) b	C statistic (95% CI)c	No.	C statistic -logistic regression (95% CI)	C index, competing risk regression (95% CI)
5 years	NHS validation cohort	23 816	1.16 (1.06 to 1.28)	0.571 (0.544 to 0.598)	24 393	1.04 (0.95 to 1.14)	0.568 (0.541 to 0.594)	18 980	0.613 (0.584 to 0.642)	0.612 (0.583 to 0.641)
	BWHS	16 059	1.00 (0.89 to 1.13)	0.534 (0.500 to 0.569)	16 390	1.06 (0.95 to 1.20)	0.541 (0.508 to 0.575)	13 247	0.574 (0.536 to 0.611)	0.573 (0.536 to 0.611)
	WHI-ES	88 037	1.01 (0.96 to 1.06)	0.577 (0.563 to 0.590)	95 441	0.81 (0.78 to 0.85)	0.557 (0.544 to 0.570)	82 634	0.567 (0.553 to 0.581)	0.567 (0.553 to 0.580)
	MEC	40 516	1.01 (0.93 to 1.08)	0.546 (0.525 to 0.567)	43 349	1.01 (0.94 to 1.08)	0.550 (0.529 to 0.570)	39 206	0.571 (0.549 to 0.592)	0.570 (0.548 to 0.591)
10 years	NHS validation cohort	23 816	1.16 (1.09 to 1.25)	0.570 (0.550 to 0.590)	24 393	1.02 (0.95 to 1.09)	0.573 (0.553 to 0.592)	18 980	0.599 (0.577 to 0.621)	0.597 (0.576 to 0.619)
	BWHS	16 059	1.09 (0.995 to 1.19)	0.554 (0.529 to 0.580)	16 390	1.19 (1.09 to 1.30)	0.547 (0.521 to 0.572)	13 247	0.570 (0.541 to 0.599)	0.569 (0.541 to 0.597)
	WHI-ESd	72 953	0.84 (0.82 to 0.87)	0.564 (0.554 to 0.574)	78 743	0.75 (0.73 to 0.78)	0.552 (0.542 to 0.562)	82 634	0.578 (0.568 to 0.588)	0.572 (0.562 to 0.582)
	MEC	40 516	0.97 (0.92 to 1.03)	0.551 (0.536 to 0.567)	43 349	0.93 (0.89 to 0.98)	0.562 (0.547 to 0.577)	39 206	0.580 (0.565 to 0.596)	0.576 (0.560 to 0.591)

^aAnalyses using the Gail model included only women with complete data for Gail model risk factors. IBIS allows risk factor data to be missing. CI = confidence interval; BWHS = Black Women’s Health Study; IBIS = International Breast Cancer Intervention Study; MEC = Multiethnic Cohort Study; NHS = Nurses’ Health Study; WHI-ES = Women’s Health Initiative–Extension Study.

^bCompared the expected number of breast cancers based on the Gail model estimates and IBIS tool estimates (calculated using the Gail SAS macro and IBIS Breast Cancer Risk Evaluation Tool, respectively) with the observed number in each cohort. Used the Poisson variance of the logarithm of the observed number of cases to determine 95% confidence intervals.

^cCalculated the C statistic or area under the receiver operating characteristic curve and its standard error to assess discrimination.

^dWomen in WHI-ES who did not participate in the second extension study were censored at 5 years in the competing risk regression model and excluded from the Gail and IBIS models.

Model validation in predicting nonbreast cancer death

Although there were some statistically significant differences between hazard ratios associated with nonbreast cancer death risk factors across cohorts, qualitatively the associations were similar (Table 4). For example, whereas the hazard ratio for a BMI of at least 40 kg/m² was statistically significantly higher in MEC for nonbreast cancer death, women with the lowest and highest BMIs were at the highest risk of nonbreast cancer death in all cohorts. The hazard ratio for diabetes was similar across cohorts despite variations in prevalence. MEC considered history of asthma but not emphysema, which may explain why this factor was not associated with nonbreast cancer death in MEC. The hazard ratio for stroke or transient ischemic attack was lower in NHS and WHI-ES, likely because MEC only assessed stroke (not transient ischemic).

Table 4.

Competing risk regression model modeling 10-year cumulative risk of nonbreast cancer death in WHI-ES and MEC^a,b,c

Modeling 10-year risk of nonbreast cancer death
Subdistribution hazard ratios (HRs) from Fine and Gray model	NHS development cohort (n = 48 102)d		WHI-ES (n = 92 720)		MEC (n = 47 973)
Observed No. of cases of nonbreast cancer death	9376		13 274		8284
Factors in our final model	HR	P	HR	P	HR	P
Age, per year increase	1.11	<.001	1.10a	<.001	1.11	<.001
Highest self-reported body mass index in past 10 years
<20 kg/m2	1.63	<.001	1.53	<.001	1.29a	<.001
20-22.4	1.15	<.001	1.12	.002	1.08	.05
22.5-24.9	1	—	1	—	1	—
25-29.9	0.91	.003	0.94	.02	1.03a	.3
30-34.9	0.88	<.001	0.89	<.001	1.15b	<.001
35-39.9	0.91	.07	0.94	.10	1.26b	<.001
≥40	1.10	.10	1.01	.73	2.00b	<.001
Average alcohol use per day, highest average use in past 10 years
None	1	—	1	—	1	—
1-4.9 g/d	0.86	<.001	0.93a	.00	0.88	<.001
5-14.9 g/d	0.91	.004	0.95	.07	1.02a	.55
≥15 g/d	1.04	.24	1.02	.48	1.08	.09
Cigarette use
Never	1	—	1	—	1	—
Current	2.46	<.001	2.64	<.001	2.54	<.001
Past	1.35	<.001	1.30	<.001	1.41	<.001
Limited from walking several blocks
Not at all	1	—	1	—	—	—
A little or a lot	1.58	<.001	1.60	<.001	—	—
Limited in bathing or dressing oneself
Not at all	1	—	1	—	—	—
A little or a lot	1.47	<.001	1.45	<.001	—	—
Walking pace
Unable to walk	1.42	<.001	1.70	<.001	—	—
Slow or average, <3 mph	1	—	1	—	—	—
Brisk or very brisk, ≥3 mph	0.69	<.001	0.82b	<.001	—	—
High blood pressure	1.15	<.001	1.20	<.001	1.31b	<.001
Depression	1.17	<.001	1.01b	.71	—	—
Hip fracture	1.31	<.001	1.13a	.02	1.29	.005
Parkinson disease	2.35	<.001	1.25b	.04	2.21	<.001
Myocardial infarction	1.27	<.001	1.26	<.001	1.80b	<.001
Congestive heart failure	1.76	<.001	2.08a	<.001	—	—
Stroke or Transient ischemic attack	1.19	<.001	1.22	<.001	1.57b	<.001
Emphysema or Asthma	1.29	<.001	1.28	<.001	1.01b	.84
Diabetes	1.36	<.001	1.35	<.001	1.46	<.001
Dementia	3.07	<.001	1.20b	.15	1.05a	.9
Kidney disease	1.28	.07	1.05	.77	—	—
Cancer	1.45	<.001	1.57b	<.001	1.55	<.001
Age at menopause, y
Younger than 45	1.01	.74	0.99	.72	1.00	.91
45-49	1	—	1	—	1	—
50-54	0.93	.004	0.98	.29	0.92	.004
55 and older	0.87	.002	0.96	.20	0.89	.006
Mammogram in past 2 years
No	1	—	1	—	1	—
Yes	0.75	<.001	0.88b	<.001	0.74	<.001
C index (95% CI) when using risk factor regression coefficients from the NHS development cohorte	0.795 (0.791 to 0.800)		0.760 (0.756 to 0.764)		0.763 (0.758 to 0.768)

^aIndicates that the hazard ratio for between-study heterogeneity is statistically significantly different from NHS development cohort hazard ratio (P < .05). CI = confidence interval; MEC = Multiethnic Cohort Study; NHS = Nurses’ Health Study; WHI-ES = Women’s Health Initiative Extension Study; — = not applicable.

^bIndicates significance at P < .001.

^cThe model predicting 10-year risk of nonbreast cancer death has a competing risk of breast cancer death.

^dWe performed a complete case analysis in which all participants with missing data were excluded.

^eC index for each cohort assessed using the regression coefficients from the NHS development cohort.

Applying NHS development cohort regression coefficients, our model’s C index when predicting 10-year nonbreast cancer death was 0.760 (95% CI = 0.756 to 0.764) in WHI-ES and 0.763 (95% CI = 0.758 to 0.768) in MEC (Table 4). The model performed worse in women aged 55-74 years and 75 and older when stratified by age (Supplementary Table 9, available online), possibly because age itself is a strong mortality predictor. Figure 3 and Supplementary Table 10 (available online) demonstrate that the expected-to-observed cumulative incidence function for 10-year nonbreast cancer death was similar except for the highest risk group where our model overestimated 10-year nonbreast cancer death by 11% in MEC and 13% in WHI-ES.

Figure 3.

Calibration of our model for predicting 10-year nonbreast cancer death risk in WHI-ES and MEC by risk group. ^aThese analyses included women with complete data. Expected nonbreast cancer death was calculated using the cumulative incidence function (CIF) from our competing risk nonbreast cancer death prediction model, and observed nonbreast cancer death was calculated using the nonparametric estimation of CIF within 5 risk groups (risk quintiles from the NHS development cohort). Figures were made using Microsoft Excel. BWHS = Black Women’s Health Study; MEC = Multiethnic Cohort Study; NHS = Nurses’ Health Study; WHI-ES = Women’s Health Initiative–Extension Study.

Examples

Table 5 shows that breast cancer and nonbreast cancer death risk estimates are similar across cohorts for hypothetical women regardless of which cohort regression coefficients and baseline survival are used in the model to calculate risk.

Table 5.

Competing-risk model estimates for breast cancer, breast cancer death, and nonbreast cancer death generated within each cohort for hypothetical postmenopausal women at 4 different ages^a,b,c

Hypothetical patients
1. Low breast cancer risk (<3.0% 5-year risk) and few comorbidities or functional impairments					BMI = 22, nonsmoker, nondrinker, average walking pace, no functional limitations, has had a mammogram in the past 2 years, aged 48 years at menopause, had 3 children, first birth at aged 22 years.
2. Low breast cancer risk (<3.0% 5-year risk) and multiple comorbidities or functional impairments					BMI = 23, former smoker, <2 drinks per week, slow walking pace, needs help getting dressed, history of cancer, has depression, history of myocardial infarction, has diabetes, has had a mammogram in the past 2 years, has no family history of breast cancer, past user of estrogen and progesterone <5 years, age 50 years at menopause, had 2 children, first birth at aged 28 years.
3. Higher breast cancer risk (≥3.0% 5-year risk) and few comorbidities or functional impairments					BMI = 38, nonsmoker, ≥7 drinks per week, average walking pace, no functional limitations, has hypertension, has had a mammogram in the past 2 years, 1 first-degree relative with breast cancer aged younger than 50 years, past user of estrogen and progesterone >5 years, aged 55 years at menopause, had 2 children, first birth at aged 30 years.
4. Higher breast cancer risk (≥3.0% 5-year risk) and multiple comorbidities or functional impairments					BMI = 35, former smoker, ≥7 drinks per week, slow walking pace, limited walking several blocks, has hypertension, diabetes, chronic obstructive pulmonary disease, history of myocardial infarction, has not had a mammogram in the past 2 years, has 1 first-degree relative with breast cancer aged 50 years and older, past user of estrogen and progesterone >5 years, aged 55 years at menopause, had no children.
	Absolute breast cancer risk, 5 years, %				Absolute breast cancer risk, 10 years, %				Absolute breast cancer death risk, 10 years, %				Absolute nonbreast cancer death risk, 10 years, %
	NHS	BWHS	WHI-ES	MEC	NHS	BWHS	WHI-ES	MEC	NHS	BWHS	WHI-ES	MEC	NHS	BWHS	WHI-ES	MEC
Aged 57 years
1	0.9	0.9	1.2	0.9	2.0	1.5	3.0	2.4	0.04	0.00	0.07	0.04	2.3	3.0	1.9	1.9
2	1.1	1.0	1.9	1.0	2.6	2.3	4.1	2.5	0.09	0.35	0.11	0.14	9.0	12.7	7.5	8.1
3	4.7	3.5	4.2	5.7	10.8	9.7	8.3	11.5	0.28	0.12	0.14	0.17	1.9	3.5	1.8	2.8
4	3.9	4.3	3.3	5.0	8.4	7.2	6.6	9.8	0.19	0.24	0.13	0.23	11.3	12.5	9.1	13.4
Aged 65 years
1	0.9	1.0	1.2	1.0	2.0	1.6	2.9	2.3	0.07	0.00	0.11	0.06	5.2	6.4	4.1	4.4
2	1.2	1.2	2.0	1.1	2.6	2.5	3.9	2.4	0.14	0.52	0.15	0.19	19.5	25.5	15.8	17.6
3	5.0	4.0	4.3	6.4	10.7	10.5	8.0	11.1	0.44	0.17	0.21	0.22	4.3	7.4	4.0	6.4
4	4.2	4.9	3.5	5.5	8.2	7.7	6.3	9.5	0.31	0.36	0.19	0.31	24.3	25.1	19.1	28.2
Aged 75 years
1	1.0	1.1	1.3	1.1	1.9	1.7	2.7	2.2	0.13	0.00	0.17	0.08	14.2	15.9	10.6	12.0
2	1.3	1.4	2.1	1.3	2.5	2.7	3.8	2.3	0.25	0.85	0.25	0.27	46.2	53.7	37.0	42.5
3	5.5	4.7	4.6	7.3	10.4	11.4	7.6	10.7	0.78	0.28	0.33	0.32	11.8	18.2	10.5	17.1
4	4.7	5.8	3.7	6.3	8.0	8.4	6.0	9.1	0.55	0.59	0.31	0.45	54.8	53.0	43.5	61.2
Aged 80 years
1	1.1	1.2	1.3	1.2	1.9	1.8	2.7	2.1	0.17	0.00	0.22	0.10	22.7	24.5	16.8	19.5
2	1.4	1.5	2.1	1.4	2.5	2.8	3.7	2.3	0.32	1.1	0.31	0.33	64.8	71.3	53.2	60.7
3	5.8	5.1	4.7	7.8	10.3	11.9	7.4	10.4	1.0	0.36	0.42	0.38	19.1	27.7	16.6	27.1
4	4.9	6.2	3.8	6.8	7.9	8.8	5.9	8.9	0.73	0.75	0.39	0.53	73.8	70.6	60.8	79.6

^aPerformed a complete case analysis in which all participants with missing data were excluded. BMI = body mass index; BWHS = Black Women’s Health Study; MEC = Multiethnic Cohort Study; NHS = Nurses’ Health Study; WHI-ES = Women’s Health Initiative Extension Study.

^bRisk estimates were generated by outputting the cumulative incidence function at the 5- or 10-year time point for each set of specified covariates in each cohort’s competing risk regression model.

^cAlthough there is no universally agreed upon threshold to define older women at high risk of breast cancer, the US Preventive Services Task Force and the American Society of Clinical Oncology consider women with ≥3% 5-year risk to be at high enough risk for clinicians to discuss the benefits and harms of breast cancer prevention medications. Therefore, we used this risk threshold in these examples (55,56).

Discussion

To inform mammography screening decisions, we examined the performance of a competing-risk model that estimates 10-year risk of breast cancer, breast cancer death, and nonbreast cancer death in diverse cohorts of women aged 55 years and older. Across cohorts, our model gave well-calibrated estimates for breast cancer risk and performed similar to or better than existing prediction models. Our model also predicted 10-year nonbreast cancer death well but overpredicted nonbreast cancer death in women at highest risk, likely because of differences in how several diseases were assessed in the cohorts and because MEC did not assess function. We anticipate that our model, which may be used among diverse postmenopausal women aged 55 years and older, will facilitate shared decision making around mammography screening.

Currently, the Gail model is the most commonly used breast cancer prediction model in primary care, likely because it is quick and available online and has been validated (40). To estimate risk, the Gail model uses population-based age- and race and ethnicity–specific breast cancer incidence rates and adjusts these rates based on the associated hazard of breast cancer from a multivariable logistic regression model that included multiple risk factors and the amount of breast cancer risk that may be explained by the model (24,25). IBIS does not consider race and ethnicity but uses age-based baseline breast cancer incidence rates from the United Kingdom. IBIS requires inputting age at cancer diagnoses and age or age of death for multiple relatives, which can be time consuming (26). We found the Gail model’s C statistic to be 0.55 in MEC and 0.53 in BWHS when predicting 5-year breast cancer risk among women aged 55 years and older and between 0.55 (BWHS) and 0.57 (NHS validation cohort) when predicting 10-year breast cancer risk and that IBIS performed slightly worse than the Gail model. In general, our model performed better than these models, but its C index (and C statistic) was still modest, likely because breast cancer was uncommon regardless of risk factors. All models were generally well calibrated, which is important because providing women with personalized risk estimates, which are often lower than their preconceived risk (41), may help inform screening decisions (42).

We anticipate that our model that considers comorbidity and functional impairment will be particularly useful for supporting shared decision making around mammography screening intervals and screening cessation. This is important because there is a dearth of shared decision-making tools for older women from diverse backgrounds (43,44) and diverse older women with varying levels of education report being interested in receiving personalized information about the benefits and harms of screening rather than being directed to have mammograms (45). Although the USPSTF and ACS recommend screening women aged 55 years and older biennially, approximately 50% of older women screened are screened annually (7,8). Providing women aged 55 years and older with model risk estimates may help some women feel comfortable choosing to be screened biennially. Also, our model may help identify women with high competing-mortality risk who should consider screening cessation. Although one study found no differences in screening among women with less than a 10-year life expectancy by race, the study found that Black women aged 65 years and older with at least a 10-year life expectancy were less likely to be screened than White women aged 65 years and older with a more than 10-year life expectancy (46). Therefore, our model may help reduce disparities by identifying older women regardless of race and ethnicity in excellent health who may benefit from screening.

Although mortality indices exist for estimating adults’ 10-year overall mortality risk, none specifically predict nonbreast cancer death (47). Existing mortality indices often include few comorbidities because they were developed from population surveys or include few functional measures because they were developed from billing codes (47). Because NHS collects rich data on comorbidity and function, our model was able to consider a wide range of mortality risk factors. The inclusion of dementia in our model may be particularly useful to support caregiver and clinician shared decision making around screening.

There is increasing interest in breast cancer prediction models that consider women’s individualized competing mortality risks (48). Hedlin et al. (49) used WHI data to develop a model that simultaneously predicts myocardial infarction; stroke; hip fracture; breast, colorectal, and lung cancer; and death, within 5-15 years (C statistic was 0.59 for predicting 10-year breast cancer risk and 0.72 for predicting 10-year death). Their model includes few breast cancer risk factors, risk factors not generally predictive of breast cancer (eg, stroke family history), risk factors that may be time consuming to obtain (eg, waist circumference), and race and ethnicity. Its performance has not been tested in an external cohort, which is necessary before implementation.

Our study has limitations. Some risk factors considered in IBIS are not available in our cohorts (eg, genetic testing, breast density, paternal breast cancer). The cohorts used slightly different questions or responses categories for assessing some risk factors. For example, MEC did not assess breast biopsies, and MEC and BWHS did not assess function. Therefore, combining cohorts into one large validation cohort would have led to a loss of precision and information. In addition, participants missing risk factor information were older, less educated, and less likely to be non-Hispanic White than those with complete data; although absolute differences were small (Supplementary Table 1, available online). Also, breast density information was not available for most women in our cohorts so we could not compare our model’s performance with models that primarily consider breast density (50). Polygenic risk scores are increasingly being tested in breast cancer prediction models, but polygenic risk scores are available for few women in our cohorts (51,52). However, there are barriers to implementing genetic testing in primary care including 1) high costs, 2) need for training, 3) ensuring equitable access, and 4) regulatory approvals (53). Also, the association of polygenic risk scores with breast cancer risk declines with age, and polygenic risk scores tend to be more accurate in individuals of European ancestry (54). If polygenic risk scores become more available, we will test their inclusion in the future.

Our model uses state-of-the-science methods for estimating competing risks directly and at the person level via the Fine and Gray method, is fully validated in diverse cohorts, and is the only model that simultaneously estimates risk of breast cancer, breast cancer death, and nonbreast cancer death. It performs slightly better than existing breast cancer prediction models and includes risk factors that may be particularly relevant to older women (eg, age at menopause). To provide older women and their clinicians with our model’s risk estimates to support shared decision making, particularly around screening frequency and cessation, we are developing a website to disseminate our model (a draft prototype may be accessed at https://bcrisk55plus.shinyapps.io/risktool/).

Data availability

Investigators who wish to use data collected in the NHS are encouraged to visit http://www.nurseshealthstudy.org/researchers. Investigators who wish to use data collected in the WHI are encouraged to visit https://www.whi.org/md/working-with-whi-data. Investigators who wish to use data collected in MEC are encouraged to visit https://www.uhcancercenter.org/for-researchers/mec-data-sharing. Investigators who wish to use data collected in BWHS are encouraged to visit https://www.bu.edu/bwhs/for-researchers/data-requests/.

Author contributions

Emily A. Wolfson, MPH (Formal analysis; Methodology; Software; Visualization; Writing—original draft), Mara Schonberg, MD, MPH (Conceptualization; Funding acquisition; Methodology; Supervision; Visualization; Writing—original draft), A. Heather Eliassen, ScD (Conceptualization; Data curation; Writing—review & editing), Kimberly Bertrand, ScD (Conceptualization; Data curation; Writing—review & editing), Yurii B. Shvetsov, PhD (Conceptualization; Data curation; Methodology; Writing—review & editing), Bernard A. Rosner, PhD (Data curation; Methodology; Writing—review & editing), Julie R. Palmer, ScD (Conceptualization; Data curation; Writing—review & editing), Andrea LaCroix, PhD (Conceptualization; Data curation; Writing—review & editing), Rowan Chlebowski, MD, PhD (Conceptualization; Data curation; Writing—review & editing), Rebecca A. Nelson, PhD (Conceptualization; Data curation; Writing—review & editing), and Long Ngo, PhD (Conceptualization; Formal analysis; Methodology; Software; Supervision; Visualization; Writing—review & editing).

Funding

This work was supported by the National Cancer Institute at the National Institutes of Health (R01CA242747; U01CA164974; R01CA228357), an NHS cohort infrastructure grant (UM1 CA186107), and an NHS program project grant (P01 CA87969). Julie R. Palmer received support from the Karin Grunebaum Cancer Research Foundation and the Susan G. Komen Foundation.

Conflicts of interest

Dr Chlebowski reported receiving personal fees from AstraZeneca, Novartis, Pfizer, Amgen, AstraZeneca, outside the submitted work. HE, who is a JNCI Associate Editor and co-author on this paper, was not involved in the editorial review or decision to publish the manuscript. The other authors have no conflicts of interest to report.