Proportion of U.S. Trends in Breast Cancer Incidence Attributable to Long-term Changes in Risk Factor Distributions

U.S. breast cancer incidence increased until about 2000, and then decreased slightly. Some of these changes might be from changes in the distribution of known breast cancer risk factors. For example, a drop of 6.7% in 2003 versus 2002 (1), has been attributed to less use of hormone replacement therapy, following the Women's Health Initiative report in July 2002 (2). Changes in the distributions of other breast cancer risk factors, including reproductive factors such as age at first live birth or the number of live births, and personal and behavioral characteristics, including alcohol consumption and body mass index (BMI), may have influenced these trends also. Quantifying the long-term impact of risk factor trends on incidence is important to understand the future breast cancer burden and may inform prevention efforts.

Few epidemiologic studies accrue and follow women long enough to capture secular or birth cohort changes in risk factor distributions. Freedman and colleagues (3) published trends by birth cohort for smoking and various reproductive variables among women in a U.S. cohort of radiation technologists. Nichols and colleagues (4) analyzed secular trends in ages at menarche and menopause, and reproductive life span using pooled data from female controls in three population-based case–control studies in Wisconsin, Massachusetts, and New Hampshire. These publications did not evaluate the association of changes in risk factor distributions with breast cancer incidence trends.

We assembled data from U.S. national surveys and analyzed trends in breast cancer risk factor distributions over nearly 30 years for U.S. women overall and in strata (<50 vs. ≥50 years old; black vs. white women). We used Poisson regression, with allowance for overdispersion, to estimate ecologic associations of BMI, age at menarche, age at first live birth, and number of births with age- and period-specific breast cancer incidence in U.S. women. Using these associations, we estimated how much breast cancer incidence would have differed from the observed incidence from 1980 to 2011 had risk factors been fixed at reference levels.

We present methods for ecologic analyses, including an attributable risk-like calculation, that are illustrated and motivated by the breast cancer application but have broad potential use.

We focused on four well-established breast cancer risk factors, BMI, age at menarche, number of live births, and age at first live birth (see for example, ref. 5).

We obtained risk factor information for women ages ≥20 years at examination and born between 1917 and 1988 from nationally representative surveys described in the following sections and in Supplementary Table S1.

National Health and Nutrition Examination Survey (NHANES) samples are designed to be nationally representative of the civilian, noninstitutionalized U.S. population. Participants are selected using a complex, stratified, multistage probability cluster sampling design. Following a home interview, subjects were examined in mobile examination centers (MECs). All information on our variables was obtained at the MECs, and we used MEC sampling weights for all NHANES surveys. To increase the sample size in overlapping age-birth cohort groups for women in the eligible age and birth-year range, we combined data from 7,040 women from the NHANES I, 1971 to 1975 (6) and 4,147 women from NHANES II, 1976 to 1980 (7) and revised the weights (8, page 282). From NHANES III (9), 1988 to 1994, we used data for 7,769 women. We combined 5 two-year NHANES surveys to obtain the continuous NHANES 1999 to 2008, resulting in 12,309 additional women in the age and birth-year range for our analysis. We adjusted the survey weights for the 10-year period (ref. 8, page 282) and harmonized variables as appropriate. Age at first live birth was not included in NHANES I and II but was obtained from NHANES III and continuous NHANES.

The National Health Interview Survey (NHIS) is a cross-sectional, national survey that measures the health of the U.S. civilian noninstitutionalized population. We used data and corresponding weights from 10,691 women in the eligible age and birth cohorts from the NHIS 1987 (10). From NHIS 2000 (11), we used data from 32,463 women, and from NHIS 2005 (11), we used data from 17,283 women. From all three NHIS surveys, we obtained information on BMI, age at menarche, number of births, and age at first live birth.

We defined 71 partially overlapping 5-year birth-cohorts, from 1917 to 1988. The last two birth cohorts only include data up to 1988. We created 13 nonoverlapping 5-year age categories (20–24, 25–29, …,80–84). We estimated proportions, means and standard errors (SEs) for each risk factor for each age-cohort group (PROC SURVEYMEANS, PROC SURVEYFREQ, SAS 9.3), using appropriate sampling weights. Missing data and observations with nonpositive weights were excluded from analysis. We repeated the computations restricted to white and black women.

We categorized mean exposure values using quintiles from the overall population. For plots that used coarser age categorizations (<40 years, 40–<50 years, 50–<60 years, and 60 years or older), mean values of risk factors for 5-year age groups were combined using weights proportional to total population size [from the NCI Surveillance, Epidemiology and End Results (SEER) mortality database] in 5-year age groups and calendar years.

We used breast cancer incidence data for the same thirteen 5-year age groups and six 5-year calendar periods (1982–1986, 1987–1991, …, 2007–2011) spanning 18 partially overlapping 10-year birth cohorts from the NCI's SEER 9 and 13 Registries databases (SEER 9–13).

We modeled breast cancer incidence separately for women aged <50 years (“premenopausal”) and ≥50 years (“postmenopausal”) and for white and black women to investigate different effects of risk factors and period by age group (“menopausal status”) and race.

We analyzed SEER incidence counts, population sizes at risk, and categorized mean risk factor levels in ecologic units defined by age and period groups.

We assumed that the numbers of breast cancer cases |${Y_{a,p}}|⁠, for specific 5-year age groups |$a = 0,1,2,...A$|⁠, and periods |$p = 0,1,2,...P,$| are independent Poisson counts. This working model leads to estimates of the expected counts. However, we relax the independence assumption by allowing for over-dispersion when computing standard errors. For each age and period group, c is the corresponding birth-cohort value, with level |${X_{a,c}},$| for risk factor X. Let |${Z_{1a}}$| and |${Z_{2p}}$| be indicator variables that are 1 for age group a and period p, respectively, and 0 otherwise, for |$a = 1,2,...,A,$| and |$p = 1,2,...,P.$| Reference levels were |$a = 0$| and |$p = 0$|⁠. First, we fit a Poisson model to the data that included age and period and the natural logarithm of the person-years, |${\eta _{a,p}},$|in cell (a,p), as an offset (PROC GENMOD, SAS 9.3), such that the expected count is |${\eta _{a,p}}\lambda _{a,p}^0$|⁠, where

For the second model, we added the risk factor |${X_{a,c}},$|

We allowed for overdispersion in estimating standard errors (option SCALE = PEARSON in PROC GENMOD), as described in detail in the Supplementary Material.

The risk factor categories and corresponding parameter estimates from fitting each risk factor X marginally are shown in Supplementary Tables S2 and S3.

We compared the unadjusted estimated rates |$\lambda _{a,p}^0$| from model (1) with the rates |${\lambda _{a,p,{X_0}}}$| obtained from model (2) with |${X_{a,c}}$|set to the reference level |${X_0}$|⁠. To assess the impact of adjustment for a risk factor over time, we averaged values of |$\lambda _{a,p}^0$| and of |${\lambda _{a,p,{X_0}}}$| for periods p between 1980 and 1994 and between 1995 and 2011. The division at 1995 was chosen because it is the middle period. We converted these average rates into the percent relative difference (analogous to attributable risk),

The variance computation for |$\Delta (a,p{\rm{ - interval}})$| is described in the Supplementary Material.

We plotted |${\log _{10}}(\lambda _{a,p}^0)$| and |${\log _{10}}(\lambda _{a,p,{X_0}}^{})$|against calendar time to display more refined information on the effect of adjustment for risk factors over time. In particular, we chose two age groups, 40 to 44 years and 60 to 64 years to represent pre- and postmenopausal women in these plots. However, all premenopausal age groups have the same log-incidence plots, apart from vertical displacements, as do all postmenopausal age groups. Thus, it suffices to present only two age groups. Percent relative differences, |\Delta ,$| do depend on age group (Table 1). “Statistical significance” refers to two-sided 0.05 level tests.

Sample sizes for risk factor surveys overall and by race are in Supplementary Table S1.

Figure 1A–D plots the mean levels of four risk factors over time, weighted to the U.S. female population, separately for age groups <40, 40–<50, 50–<60 and ≥60 years.

Mean BMI increased for all age groups, and by 1995 mean BMI was and remained elevated compared with the 1980s (Fig. 1A). For example, for women ages <40 years, the mean BMI was 23.9 in 1980, 25.7 in 1995, and 25.9 kg/m² in 2009. Women under age 50 years had lower mean BMIs. Women ages 50 to 59 had comparable mean BMIs with women ages 60+ years before 1993, but higher BMIs thereafter.

For each calendar time point, mean age at menarche was lowest in women under 50 years old and highest in women ages ≥60 years (Fig. 1B). For all age groups, however, mean age at menarche declined. The strongest drop was seen for women ages 50 to 60, whose mean age at menarche was 13.0 years in 1980, and 12.6 years in 2009.

The mean number of live births also declined overall for all women between 1980 and 2009 (Fig. 1C). However, an increase between 1990 and 1997, that is most pronounced among the oldest women, reflects the baby boom after World War II.

Mean age at first live birth among parous women did not change appreciably over time for women <40 or ≥60 years old (Fig. 1D). It increased from 21.9 in 1983 to 24.0 in 2009 for women ages 40 to 50 and from 22.2 to 23.1 for women ages 50 to 60, with the steepest increase after 1990.

The risk factor patterns for white women (Supplementary Fig. S1) largely agree with the overall risk factor patterns (Fig. 1). Black women had higher mean BMI levels (Fig. 2A) than white women for all age groups for each calendar year. For example, mean BMI in 1980 was 23.6, 25.3, 26.6 and 26.9 kg/m² for white women ages <40, 40–50, 50–60, and 60+, respectively, and the corresponding values were 25.6, 29.0, 28.4, and 29.5 kg/m² for black women. Among black women, mean BMI increased starting in 1980 for all age groups; the increase was much stronger than for white women (compare Fig. 2A with Supplementary Fig. S1A). The strongest increase was seen for black women aged <40 years, for whom the mean BMI was 25.6 in 1980, 28.7 in 1995, and 31.2 kg/m² in 2009, while the corresponding values were 23.6, 25.03, and 25.7 kg/m² for white women aged <40 years. Among black women in age groups <40, 50–59, and ≥60 years, mean age at menarche decreased from 12.7 in 1980 to 12.2 in 2009, from 13.5 to 12.7, and from 13.8 to 12.7, respectively (Fig. 2B). For black women ages 40 ≤ 50 years, mean age at menarche barely changed from 12.6 years in 1980 to 12.8 years in 2009. The changes for white women were somewhat smaller than for black women (Supplementary Fig. S1B). For white women, the mean number of births decreased for all age groups, with the strongest decline among women ages 40 to 50 years (Supplementary Fig. S1C). Black women had higher mean number of births for all age groups and calendar years compared with white women. From 1980 to 2009, the mean number of births in black women aged <40 years did not change, but it decreased from 4.0 to 2.3 for black women ages 40 to 50, from 4.8 to 2.2 for black women ages 50 to 59 and from 4.1 to 2.9 for black women ages ≥60 years (Fig. 2C). These decreases are comparable with those among white women. White women of all ages had later mean age at first live birth than black women (Supplementary Fig. S1; Fig. 2D). The largest change was seen for women ages 40 to 49 years, for whom mean age at first live birth rose from 22.4 in 1983 to 23.2 in 2009. As for white women, a change in mean age at first live birth was only seen for black women ages 40 to 49 years, for whom mean age at first live birth was 20.6 in 1983 and 22.2 in 2009 (Supplementary Fig. S1; Fig. 2D).

We compared incidence rates computed from the unadjusted model (1) to those from the adjusted model (2) with the risk factor set to its lowest (reference) risk level: <25.3 kg/m² for BMI, >13.1 years for age at menarche, >2.9 for number of births. The exception was age at first live birth, for which the referent category was second to lowest (21.6–22.1; Supplementary Table S2). We graphed results for a woman ages 40 to 44 years and for a woman ages 60 to 64 years by adding specific age effects to the intercept and period coefficients. Thus, the results can be interpreted as unadjusted log-incidence rates and log-incidence rates adjusted to the reference level for an ecologic risk factor, for women in those two age groups. However, the same log-incidence patterns would be observed for other pre- and post-age 50 age groups, respectively, apart from vertical displacements.

First, we considered all U.S. women ages <50 years (“premenopausal”). For a woman in the 40 to 44 years age group, breast cancer increased slightly from 1980 to 1985, then dropped slightly and remained constant until 2011. Reference-level incidence from the BMI-adjusted model is higher than unadjusted incidence after 1995, as in Fig. 3A, that plots log₁₀ incidence against calendar period. This reflects the facts that higher BMI is associated with lower breast cancer risk for younger women and that population BMI was higher after 1995 for all ages. Before 1995, when BMI levels were lower, adjustment to reference-level BMI increased incidence statistically significantly by 5.9%, and after 1995, by 9.1% (Table 1), consistent with a protective effect of increased BMI in premenopausal women. Additional detail on 5-year age–specific percentage relative differences in incidence (with SEs) is in Table 1. For age-at-menarche (Fig. 3B) the adjusted and unadjusted curves differed less. From 1980 to 1994, adjustment to reference-level age at menarche increased incidence nonstatistically significantly by 5.4%, and from 1995 to 2011 by 5.7%. This nonstatistically significant result is not consistent with the facts that age-at-menarche decreased over time and that early age-at-menarche increases breast cancer risk. There is a strong impact of the number of births on incidence (Fig. 3C). Adjustment to reference-level for numbers of births statistically significantly reduced incidence by 21.1% before 1995 and by 23.1% between 1995 and 2011 (Fig. 3C; Table 1), reflecting the protective effect of more births (and less nulliparity) on breast cancer risk. However, adjustment to reference-level age at first live birth among parous women had a much smaller nonstatistically significant impact; it decreased incidence by 1.8% from 1980 to 1994 and by 1.3% between 1995 and 2011 (Fig. 3D; Table 1).

Next, we considered all U.S. women ages ≥50 years (“postmenopausal”). Figure 4A–D shows the log₁₀-incidence trends for women ages 60 to 64 years. Incidence increased strongly until 2002, then fell slightly and remained nearly constant, but still higher than in 1980. Adjustment to reference-level BMI decreased incidence nonstatistically significantly by 3.4% after 1995 and by 1.9% before then (Fig. 4A; Table 1), consistent with BMI's positive association with breast cancer risk in postmenopausal women. Adjustment to reference-level age at menarche had virtually no impact on incidence; it increased incidence by 0.6% between 1980 and 1995 and decreased it by 0.3% after 1995 (Fig. 4B; Table 1). Adjustment to reference-level numbers of births decreased incidence statistically significantly by 4.2% between 1980 and 1994 and by 4.3% between 1995 and 2011 (Fig. 4C; Table 1). Changes in age at first live birth had virtually no impact on breast cancer incidence (Fig. 4D; Table 1).

For women ages <50 years, unadjusted breast cancer incidence was similar in white and black women (Supplementary Fig. S2). However, adjustment to reference-level for BMI and number of births had a stronger impact on incidence in black than in white women (Supplementary Fig. S2A and S2C). For a black woman ages 40 to 44 years, adjustment to reference-level BMI increased incidence by 14.7% before 1995 and by 21.0% thereafter (Supplementary Table S4). For a white woman ages 40 to 44 years, adjustment increased incidence only by 4% both before and after 1995 (Supplementary Table S5). From 1980 to 1994, adjustment to reference-level number of live births (Supplementary Fig. S2C) decreased risk by 21.6% for white women and by 37.1% for black women ages 40 to 44 years, and from 1995 to 2011 the decrease was 24.2% for white women and 38.2% for black women (Supplementary Tables S4 and S5).

Among women ages 60 to 64 years, black women had noticeably lower unadjusted breast cancer incidence than white women (Supplementary Fig. S3). Adjustment to reference-level BMI decreased incidence among black women by a nonstatistically significant 1% before 1995 and 2.4% after 1995. For white women, the decrease was 2% (nonstatistically significant) before 1995, but 5.1% (statistically significant) after 1995 (Supplementary Fig. S3A; Supplementary Tables S4 and S5). Adjustment for age at menarche did not impact rates appreciably in either racial group. Adjusting to the reference-level number of live births reduced breast cancer incidence among 60- to 64-year-old black women by 2.1% (statistically significant) before 1995 and by 0.4% (not statistically significant) after 1995. Among white women statistically significant decreases were found: 3.3% before 1995 and 2.9% after 1995. For both time periods, adjusting to reference-level age at first live birth resulted in <1% decrease in incidence for white women, but in statistically significant 2% decreases for black women ages 60 to 64 years (Supplementary Fig. S3C and S3D; Supplementary Tables S4 and S5).

We presented methods for ecologic analyses (also see Supplementary Material) and describe how secular changes in the distributions of well-established breast cancer risk factors are associated with trends in breast cancer incidence in U.S. women. We used these associations to estimate what the incidence rates might have been if BMI, age at menarche, number of births, and age at first live birth had been fixed at reference levels from 1980 to 2011. Although there have been several studies of trends in risk factor prevalence, we are unaware of previous efforts to estimate their ecologic associations with breast cancer incidence trends. We concluded that trends in numbers of live births and BMI were strongly associated with breast cancer incidence in all U.S. women. These associations were especially strong in premenopausal women in whom decreases in numbers of births may have increased breast cancer incidence by 20%, and increases in BMI may have may have decreased incidence by about 7% from 1980 to 2011, with stronger effects after 1995, when BMI levels were particularly high. In women ages ≥50 years, adjustment to reference level BMI reduced breast cancer incidence by approximately 3%, consistent with the known positive association of BMI with breast cancer incidence in postmenopausal women. Adjustment for the risk factors we studied produced smaller percentage changes in incidence in postmenopausal women, but because rates are higher in postmenopausal women, the public health impact could nonetheless be appreciable. The ecologic associations were stronger in black than in white women, possibly reflecting stronger secular trends in their risk factor distributions.

We used nationally representative survey data to estimate average risk factor levels over time. Others have studied these risk factors, sometimes using the same data sources. For example, Komlos and colleagues (12) estimated BMI trends by birth cohort (1882–1986), ethnicity and gender using NHANES data. McDowell and colleagues (13) used NHANES data from 1999 to 2004 to show declines in mean age-at-menarche in women born before 1920 compared with women born in 1980 to 84, with larger declines seen among black women. Krieger and colleagues (14) used NHANES and National Health Examination Surveys (NHES) to show that BMI increased and age-at-menarche fell from 1959 to 2008, with patterns varying by race/ethnicity and socioeconomic status. Using census data, Martin and colleagues (15) demonstrated declines in the birth rate among U.S. women from 1980 to 2011, with stronger decreases for black women than for white women, similar to the pattern we observed. Compared with these articles, our analysis covers a longer time window, and by combining different surveys provides more precise standard errors for risk factor prevalence.

We found ecologic associations with risk factors that agreed qualitatively, with a few exceptions, with findings from analytic epidemiologic studies, although some of the ecologic associations were not statistically significant and the sign of the association varied in some categories (Supplementary Tables S2 and S3). BMI was inversely associated with breast cancer risk in women aged <50 years, and positively associated in women aged ≥50, in accord with analytic studies (e.g., refs. 16–20). Early age-at-first-live-birth was also protective in our ecologic study, in line with analytic studies (e.g., ref. 21). We did not find significant associations with age-at-menarche, however, unlike analytic studies (e.g., ref. 22). However, we want to stress that although the general agreement of our results with that from analytic studies is reassuring, we do not estimate individual level risk factor associations for breast cancer in our ecologic investigation. To do that, much stronger assumptions on a model and additional individual level risk factor information on subsets of individuals in the ecologic unit would be needed for some analyses, for example, those described in ref. 23.

In other settings, ecologic associations are often confounded by factors that differ among the various populations studied (e.g., ref. 24). We are studying a single large region, the United States. Our ecologic units are defined by age and period and may be less subject to unmeasured confounding than studies based on regional units. Nonetheless, an unmeasured risk factor that varied with period and was associated with, for example, BMI could confound the ecologic association of BMI with breast cancer incidence. For example, changes in screening practice or unmeasured changes in the racial or ethnic composition of the population could confound our results. Several facts argue against confounding by screening, however. First, increased BMI is protective in premenopausal but adverse in postmenopausal women, and the effects of BMI are stronger after 1995, whereas screening increased in the 1980s (19, 20). Moreover, protective ecologic associations with number of births were stronger in women under age 50 years, who were less likely to be screened, and were nearly constant across calendar time (Figs. 3 and 4). Although secular change in the racial/ethnic composition of the population is a potential confounder, the key features we found were confirmed in separate analyses of white and black women.

Few articles have attempted formal analyses to estimate associations of time trends in disease incidence with ecologic exposure information. One key article by Holford and colleagues (25) showed that most of the period and cohort contributions to lung cancer incidence trends in Connecticut could be explained by a multiplicative ecologic model that included the proportion of current and ex-smokers and the mean duration of smoking. Here again, a single region was studied, and the ecologic units were defined by period and cohort (or equivalently by age and period).

These considerations do not eliminate the possibility that other unmeasured confounders may account for the plausible ecologic associations we found, but they suggest robustness to confounding in studies with ecologic units defined by time, compared with regional ecologic units.

Strengths of our study include use of large population-based surveys for risk factors and large numbers of breast cancer cases in registry-based incidence data that allow us to stratify the analysis by age and race. As we adjust for 1-year period effects, our model accommodates abrupt changes in breast cancer incidence, like the drop in 2002, following the publication of results on hormone replacement therapy from the Women's Health Initiative (2). This study also has limitations. As previously discussed, our ecologic associations may be influenced by unmeasured confounders that are associated with various age-, period-, and cohort units. We modeled each risk factor marginally and did not assess joint effects. We did not comprehensively evaluate some other important risk factors, such as hormone replacement therapy use or physical activity. We did also not utilize information on mammographic screening, which was widely implemented in the 1980s and contributed variably to increased breast cancer incidence for different age groups (26). We also limited race-specific analyses to white and black women, because there were insufficient data for other groups. Finally, comparison with women at the reference level of risk factors is a hypothetical counterfactual construct, analogous to that used for attributable risk calculations.

In summary, we provided data on trends in breast cancer risk factors, found ecologic associations with breast cancer in line with results from analytic studies, and used data on risk factor prevalence and ecologic associations to estimate the impact that these risk factors might have had on population incidence rates.

Y. Webb-Vargas is a statistical scientist at Genentech, Inc., and the work described in this article was conducted while she was a predoctoral fellow at NCI. No potential conflicts of interest were disclosed by the other authors.

Conception and design: R.M. Pfeiffer

Development of methodology: R.M. Pfeiffer, Y. Webb-Vargas, M.H. Gail

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): R.M. Pfeiffer, Y. Webb-Vargas

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): R.M. Pfeiffer, Y. Webb-Vargas, W. Wheeler, M.H. Gail

Writing, review, and/or revision of the manuscript: R.M. Pfeiffer, Y. Webb-Vargas, M.H. Gail

This research was supported by the Intramural Research Program of the NCI. We thank Mary Alice Anderson for help with identifying risk factor variables from the surveys and Dr. William Anderson for providing the SEER incidence data.

The costs of publication of this article were defrayed in part by the payment of page charges. This article must therefore be hereby marked advertisement in accordance with 18 U.S.C. Section 1734 solely to indicate this fact.