Breast Cancer Risk after Recent Childbirth
Hazel B. Nichols,Minouk J. Schoemaker,Jianwen Cai,Jiawei Xu,Lauren B. Wright,Mark N. Brook,Michael E. Jones,Hans‐Olov Adami,Laura Baglietto,Kimberly A. Bertrand,William J. Blot,Marie‐Christine Boutron‐Ruault,Miren Dorronsoro,Laure Dossus,A. Heather Eliassen,Graham G. Giles,Inger Torhild Gram,Susan E. Hankinson,Judy Hoffman-Bolton,Rudolf Kaaks,Timothy J. Key,Cari M. Kitahara,Susanna C. Larsson,Martha S. Linet,Melissa A. Merritt,Roger L. Milne,Valeria Pala,Julie R. Palmer,Petra H. Peeters,Elio Riboli,Malin Sund,Rulla M. Tamimi,Anne Tjønneland,A. Trichopoulou,Giske Ursin,Lars J. Vatten,Kala Visvanathan,Elisabete Weiderpass,Alicja Wolk,Wei Zheng,Clarice R. Weinberg,Anthony J. Swerdlow,Dale P. Sandler
DOI: https://doi.org/10.7326/m18-1323
IF: 39.2
2018-01-01
Annals of Internal Medicine
Abstract:Breast cancer is the leading cancer diagnosis among reproductive-aged women worldwide (1). Parity is recognized as a protective factor for breast cancer overall, but this may largely apply to the peak ages of incidence (after age 60 years) and may not be true for younger women. Previous studies have reported that recent childbirth confers a short-term increase in breast cancer risk (213), which may last 10 or more years (6, 11, 1416) and be amplified in women who are older at first birth (6, 11, 15, 16). Evidence for this increased risk often comes from national registry linkage studies in Scandinavian countries (2, 4, 6, 17). Information about such behaviors as breastfeeding is often not available or comes from casecontrol studies (811), where potential risk factors are assessed after diagnosis and parenting responsibilities could differentially deter study participation. We used combined data from 15 cohort studies to assess breast cancer risk after childbirth. The use of international, prospective data offers a new opportunity to assess the strength and duration of associations between recent childbirth and breast cancer risk while considering the effect of such factors as breastfeeding and family history of breast cancer (5, 18). It also enables evaluation of risk that is specific to breast cancer subtypes that may be differentially influenced by reproductive history (12, 13, 19). Understanding these patterns may have implications for identifying risk-reducing strategies and vulnerable subgroups. Methods We used data from the Premenopausal Breast Cancer Collaborative Group, a pooling project involving 20 prospective cohort studies (20). This work was approved by the relevant institutional review boards. In brief, participating studies contributed data from women aged younger than 55 years who did not have breast cancer at enrollment; these women were followed prospectively through direct contact or linkage with cancer registries (described previously [20]). Studies contributed (as available) age at enrollment and end of follow-up, demographic characteristics, lifestyle factors, reproductive history, medical conditions, and first-degree family history of breast cancer at enrollment and each round of follow-up. Data harmonization and quality control were done by the study coordinating centers in North Carolina and London. Across 15 cohorts that provided information on women's ages at childbirth (2135), 890269 women (96% of participants) had available information on total number of births and age at most recent birth or were nulliparous. We excluded women who reported a first birth before age 13 years (n= 82), were 50 years or older at study entry and at most recent birth (n= 60), or reached parity greater than 10 births before enrollment (n= 183). These events were considered to have greater potential for data errors. This left 889944 women for analysis (Supplement Figure 1). Supplement. Supplementary Material Attained age, ages at first and most recent births, and parity at study enrollment were available in all 15 studies (2135). Twelve studies (2127, 29, 31, 3335) assessed pregnancy history in at least 1 follow-up questionnaire after enrollment; the remaining 3 provided pregnancy information at enrollment only. Breastfeeding status was available in 12 studies (2123, 2527, 2934) and family history in 13 (2127, 29, 3135). Thirteen studies (2123, 2527, 2935) reported breast cancer stage and estrogen receptor (ER) status. Statistical Analysis Parity, time since most recent birth, breastfeeding, and family history of breast cancer were analyzed as time-varying exposures over follow-up. We used Cox proportional hazards regression to calculate hazard ratios (HRs) and 95% CIs for the association between time since most recent birth and breast cancer, with attained age as the underlying time scale (17, 36). Follow-up started at age at study enrollment or the first available follow-up round with information on age at most recent birth and ended at breast cancer diagnosis, death, last follow-up, or age 55 years, whichever occurred first. During follow-up, women were censored at the age at which they reached parity greater than 10 births (n= 22) or the age at which they had a birth at age 50 years or older (n= 9) (Supplement Figure 1). Proportional hazards assumptions were assessed by Schoenfeld residuals (37) and were not significantly violated. We first examined study-specific estimates and calculated a pooled estimate across studies using a random-effects model that weighted estimates by the inverse of the study-specific variance (3840). Because we detected no significant heterogeneity between studies with the Cochran Q test or I 2 statistic (4143) (Supplement Figure 2), we pooled individual-level data and did an aggregated analysis stratified by study cohort. Supplement Table 1 gives characteristics of the individual cohorts. Time since most recent birth and parity were modeled as time-varying exposures in 1-year and 1-birth increments, respectively. We accounted for additional births during follow-up by resetting time since birth to 0 at the time of each birth. The Supplement gives additional detail on these methods. Quadratic splines (44) were used to examine time since birth as a continuous, nonlinear exposure, defining time with knots at the 5th, 25th, 50th, 75th, and 95th percentiles for the distribution of time since most recent birth for women with a breast cancer diagnosis before age 55 years. In spline models, time since most recent birth was set to 0 for nulliparous women and an indicator term for parity allowed the risk at 0 years since most recent birth to differ between nulliparous and parous women. As an approximation of the 95% CI (in years since most recent birth) for the point where the HR crossed 1.0, we used the points where the lower and upper bounds of the 95% CI for the spline regression crossed 1.0. In categorical models, exposure was defined as nulliparous or 0 to 2.9, 3 to 4.9, 5 to 9.9, 10 to 14.9, 15.0 to 19.9, 20.0 to 24.9, 25.0 to 29.9, or 30 or more years since most recent birth. Covariates considered as potential confounders were parity, age at first birth, breastfeeding, infertility, education, oral contraceptive use, and birth cohort. We identified confounding variables using a directed acyclic graph (45, 46) and the prior literature (Supplement Figure 3); the minimally sufficient adjustment set was parity and breastfeeding. All models were adjusted for attained age (as the time scale; continuous), study, and parity (1 to 10 births; time-varying). Adjustment for breastfeeding was possible only in analyses limited to the 12 studies with available breastfeeding data. We evaluated potential effect modification by parity (primiparous [1 birth], biparous [2 births], or multiparous [3 births]), age at first birth (<25, 25 to 34, or 35 to 39 years), breastfeeding, and family history of breast cancer. Interactions between these factors and time since most recent birth were assessed using likelihood ratio tests (47). We examined risk for invasive nonmetastatic disease (stage I to III breast cancer) by treating breast cancer of stage 0 (in situ) or IV as a censoring event. Augmentation models were used to assess differences in HRs by ER status by using the Wald test (48). In additional analyses, we restricted the cohort to parous women only, censored follow-up at age 45 years or the last age at which pregnancy history was assessed if younger than 45 years (to minimize the potential for additional pregnancies after the most recent questionnaire), excluded each study in turn to identify potentially influential studies, and excluded women with multiple births (for example, twins). We also calculated the weighted cumulative incidence of breast cancer according to attained age for categories of time since most recent childbirth (nulliparous and 0 to 2.9, 3 to 6.9, 7 to 14.9, 15 to 24.9, and 25 or more years), adjusted for the distribution of parity in the overall pooled sample using an inverse probability of exposure approach, described further in Supplement Figure 4 (49). Because our data were left-truncated, we also provided a standardized weighted cumulative incidence function calculated over a common age interval that had participants in each category of time since most recent birth. The standardized weighted cumulative incidence function starts at 0 at the beginning of the common age interval and cumulates throughout the interval, which allows comparison of the cumulative incidence across categories of time since most recent birth (Supplement Figure 5). Analyses were done with SAS, version 9.3 (SAS Institute); figures were produced in SAS and R (R Foundation for Statistical Computing). Role of the Funding Source The funding sources for this analysis had no role in the design, conduct, or interpretation of the study. Results During 9625727 person-years of follow-up (mean, 10.8 years [SD, 6.4]), 18826 incident cases of breast cancer were diagnosed before age 55 years among 889944 women. At enrollment, 720555 women were parous; 71609 women contributed 1 or more births during follow-up. The mean age at study entry was 41.8 years (range, 16.0 to 54.9 years). The last update of pregnancy information occurred at a mean age of 50.0 years (range, 16.0 to 76.7 years). Overall, 12.4% of person-years were contributed by women who reported a family history of breast cancer (Table). For parous women, 72.9% of person-years were contributed by women who reported breastfeeding. Table. Distribution of Person-Years Among 889944 Women in 15 Prospective Cohort Studies, by Exposure Characteristics* Figure 1 shows the association between time since most recent birth and breast cancer risk, modeled nonlinearly as a continuous exposure. Compared with nulliparous women, parous women had an HR for breast cancer associated with time since most recent birth that peaked 4.6 years after birth (HR, 1.80 [95% CI, 1.63 to