Quantifying the Role of Circulating Unconjugated Estradiol in Mediating the Body Mass Index–Breast Cancer Association

In a case–control study nested within the Prostate, Lung, Colorectal, and Ovarian Cancer Screening Trial (PLCO) cohort, we for the first time used liquid chromatography-tandem mass spectrometry (LC/MS-MS) to measure and prospectively evaluate serum concentrations of the parent estrogens, estrone and estradiol, and 13 estrogen metabolites with breast cancer risk (1). We found, as expected, that the serum concentration of unconjugated estradiol was statistically significantly associated with breast cancer risk, with risk doubling across extreme deciles. We also reported that less extensive methylation of potentially genotoxic 4-hydroxylation pathway catechols was associated with higher breast cancer risk and that more extensive 2-hydroxylation of parent estrogens was associated with lower breast cancer risk among women who were postmenopausal nonhormone users at the time of serum collection.

Higher body mass index (BMI) has also been associated with increased breast cancer risk among postmenopausal women, particularly among never hormone users (2–6) and those with estrogen and progesterone receptor–positive breast cancer (7). Elevated estrogen production through aromatization of androgens in adipose tissue is hypothesized to mediate, at least in part, the influence of BMI (8). Previous efforts to determine whether endogenous serum estrogen levels and BMI are independent breast cancer risk factors have led to mixed results (9–12). To our knowledge, the one study that quantified the extent of estrogen mediation of the BMI–ER⁺ breast cancer association reported that 49% of the effect of BMI was mediated through circulating estradiol (12).

In fiscal year 2012, the U.S. National Cancer Institute (NCI) introduced an initiative to determine how obesity promotes cancer development (13), with the hope that determining the molecular mechanisms could identify new ways to prevent and treat cancer. In this analysis, we quantify the mediating effects of circulating unconjugated estradiol on the BMI–ER⁺ breast cancer association. Our use of LC/MS-MS to measure serum concentrations of the parent estrogens, estrone and estradiol, and 13 estrogen metabolites (1) also allowed us to evaluate whether estrogen metabolism profiles predictive of breast cancer risk contribute to the mediating effect of BMI on ER⁺ breast cancer risk.

We described the methods for this case–control study nested in the PLCO cohort in our previous publication on estrogen metabolism and risk of breast cancer in postmenopausal women (1). In brief, the PLCO is a multicenter trial in which participants randomized to the screening arm, during 1993 through 2001, completed self-administered questionnaires about their medical history, height and weight, personal characteristics, and health-related behaviors, and had blood samples drawn at baseline. Vital status and incident breast and other cancers were ascertained primarily via annual mailed questionnaires, but also through state cancer registries, the National Death Index, and physician and next-of-kin reports. Ninety-six percent of breast cancers were subsequently confirmed by pathology reports. Institutional Review Boards at the US NCI and the 10 participating screening centers approved this study.

Our previous analysis of estrogen metabolism and breast cancer risk (1) included 277 histologically confirmed invasive breast cancer cases diagnosed between January 1, 1993 and June 30, 2005 and 423 controls who were free of breast cancer as of June 30, 2005. All were postmenopausal and not currently using hormone therapy at baseline, had no other cancers other than nonmelanoma skin cancer diagnosed during follow-up, had sufficient baseline serum for biochemical analyses, and did not have extreme values for the sum of all estrogens and estrogen metabolites. Specifically, we excluded subjects with summed estrogen/estrogen metabolite values outside ± 1.5*interquartile range (n = 27). The present analysis includes only those cases (N = 193) and controls (N = 268) in the previous analysis who reported at baseline that they had never used menopausal hormone therapy, and had baseline weight of 60+ pounds, heights between 48 and 78 inches, and BMI values of 15+ kg/m². We excluded former hormone users because the associations of postmenopausal breast cancer with BMI are seen most clearly in never hormone users (2–6). A total of 143 cases (74%) were ER⁺ (>1% of cells positive) and 31 (16%) ER⁻. The 19 cases with unknown ER status were included in the overall analyses, but excluded from the ER-specific analyses.

Serum samples were stored at −80°C within 2 hours of blood collection. Fifteen estrogens and estrogen metabolites were measured concurrently with stable isotope dilution LC/MS-MS (1). Serum samples from six postmenopausal women selected to cover the range of circulating estrogen/estrogen metabolite concentrations served as blinded quality control samples inserted into the assay batches. Total laboratory coefficients of variation, including both within- and between-batch variation and all steps of the analytic procedure, were less than 5% for all individual estrogens and estrogen metabolites measured and less than 3% for estradiol and estrone, in both unconjugated and conjugated form. The lower limit of quantitation for each estrogen/estrogen metabolite was estimated to be 1 to 2 pmol/L (1). No assays of estrogens or estrogen metabolites in this study resulted in nondetectable readings.

We analyzed estrogens and estrogen metabolites individually, in groups representing metabolic pathways, and as ratios of metabolic pathway groups. We used base_1.2 to log-transform individual and grouped estrogens and estrogen metabolites because serum concentrations increased by approximately 20% between the 10th and 90th percentiles. Metabolic pathway ratios increased by approximately 4% between the 10th and 90th percentiles; therefore, we log-transformed them using base_1.04. Log-transformed values were used in all analyses presented in the article.

BMI was defined as weight (in kilograms)/height (in meters) squared (kg/m²). We classified participants as normal weight (<25 kg/m²), overweight (25–29.9 kg/m²), and obese (≥30 kg/m²), according to World Health Organization standards (14, 15). Three subjects with BMI <18.5 kg/m², which is considered underweight, were included in the normal weight category.

We calculated Pearson correlation coefficients (r) and coefficients of determination (r²) between log-transformed estrogens and estrogen metabolites and measures of BMI using control subjects weighted by the inverse of the sampling fraction to represent the study cohort. Percentages of control subjects at each level of selected risk factors were also based on weighted counts as were χ² test P values for the comparison of all cases and controls. We calculated person-years for the cohort by subtracting entry age into the cohort from exit age and multiplying this by the weighting factor.

We based mediation analyses on a counterfactual framework in a survival setting using two methods: one based on the Aalen additive hazards model (16) and the other based on the Cox proportional hazards model (assuming the outcome is rare; ref. 17). The additive hazards model assumes that the hazard is linear in the covariates and the proportional hazards model assumes that the log of the HR is linear in the covariates. Both of these methods require the following additional assumptions for direct causal interpretation: conditional on covariates, (i) there is no confounding of the BMI–breast cancer relationship, (ii) no confounding of the unconjugated estradiol–breast cancer relationship, (iii) no confounding of the BMI–unconjugated estradiol relationship, (iv) no confounding of the unconjugated estrogen–breast cancer relationship in which the confounder is affected by BMI; and (v) the consistency assumption that a person's potential outcome under counterfactual conditions is the outcome experienced by that person.

We estimated the mediating effect of unconjugated estradiol (E2) on the association between BMI and ER⁺ breast cancer in the following manner. Suppose the Aalen or Cox regression models for ER⁺ breast cancer with attained age as the time variable have linear predictor δ × BMI + ϵ × log(E2) + other covariates, and that the linear regression of log(E2) as predicted by BMI is log(E2) = β × BMI + other covariates.

Then the direct effect of BMI on ER⁺ breast cancer risk, the effect not mediated by estradiol, is δ, the indirect effect of BMI on ER⁺ breast cancer risk, the effect that is mediated by estradiol, is ϵβ, and the total effect of BMI on ER⁺ breast cancer risk is τ = δ + ϵβ. The total effect can also be calculated as the coefficient for the exposure in a model unadjusted for the mediator. The mediated proportion is ϵβ/τ. The percent relative change in E2 per unit change in BMI is 100(exp(β)−1).

We weighted controls in the Aalen and Cox regression models to represent the study cohort and gave case subjects a weight of 1 since no sampling occurred. We adjusted all regression models for study design factors and standard breast cancer risk factors as described in Table 1.

We obtained 95% confidence intervals (CI) for the direct effects of BMI from the parameter estimates and SEs from the Aalen additive risk and Cox regression models. We calculated HRs for a 5 kg/m² increase in BMI to be consistent with a previous report (12). We used parametric bootstrap methods to obtain CIs for indirect and total effects as well as for the mediated proportion from the Aalen model. In particular, we assumed that the original estimates were normally distributed with means and covariances obtained from the original analysis. Then we simulated 100,000 replications of the estimated parameters from this normal distribution and calculated total and indirect effects and the mediated proportion for each replication. We derived 95% CIs from the 2.5th and 97.5th percentiles of the simulated values. We used the same approach to obtain CIs for the mediated proportion based on the Cox model.

We conducted regression analyses in R version 3.0.1. We tested the Aalen additive risk model for time-dependent effects and there were none. We tested the Cox proportional hazards assumption for estradiol and BMI using the cox.zph module in R. The slopes for the βs across attained age were not statistically significantly different from 0. An interaction term between unconjugated estradiol and BMI (classified as normal weight, overweight, and obesity) in the Cox model was not statistically significantly different from 0.

Females have been reported to over report their height and generally under report their weight (18); thus, regression models have been developed in which measured height and weight were predicted by self-reported height and weight (19). We applied parameters from these models to the PLCO data for analyses of continuous BMI in 5 kg/m² units to correct for potential errors in self-reported weight (weight_kg^SR) and height (height_meter^SR) (19):

Weight_kg^C = −4.9776 + 1.1625 × Weight_kg^SR − 0.0008 × (Weight_kg^SR)² − 0.0202 × Age − 0.1648 × Non-Hispanic Black −0.6189 × Hispanic + 0.0952 × Other races.

Height_meter^C = 2.31537 − 1.69849 × Height_meter^SR + 0.77468 × (Height_meter^SR)² + 0.00131 × Age − 0.00002 × Age² −0.00418 × Non-Hispanic Black − 0.01943 × Hispanic − 0.01359 × Other races.

We then repeated selected statistical analyses with the corrected data to determine whether reporting bias might impact our estimated mediation effects.

The incidence rate of breast cancer weighted to the PLCO cohort was 336 cases per 100,000 woman-years. The median time from serum collection to breast cancer diagnosis was 3.6 years (interdecile range = 0.9–7.9 years) for cases. Weighted median follow-up among all participants in the study was 7.7 years (interdecile range = 5.1–10.3 years). The median age (5th–95th percentile) at study baseline was 64 (56–73) years for cases and 62 (55–72) years for controls. Median BMI values (5th–95th percentile) were 27.3 (21.5–37.2) kg/m² and 26.6 (20.9–37.4) kg/m² among cases and controls, respectively; median serum levels (5th–95th percentile) of unconjugated estradiol were 15.7 (14.5–16.9) pmol/L and 15.6 (14.5–16.9) pmol/L among cases and controls, respectively. Other characteristics of cases and controls, weighted to represent the study cohort, are shown in Table 1.

Serum concentrations of estrogens, estrogen metabolites, and metabolic pathway groups were positively and statistically significantly correlated with baseline BMI (Table 2); the strongest correlation was with unconjugated estradiol (r = 0.45). Of the metabolic pathway ratios, only two, the ratios of the 2-hydroxylation pathway to parent estrogens and the 2-hydroxylation pathway to the 16-hydroxylation pathway, were inversely and statistically significantly associated with baseline BMI measures (r = −0.22 and −0.23, respectively; Table 2). The r for the log_1.04 (ratio of conjugated to unconjugated estradiol) with baseline BMI was −0.12 (P = 0.06; data not shown in Table). Estrogens, estrogen metabolites, and metabolic pathway groups were not correlated with height (data not shown).

HRs (without adjustment for BMI) for total and ER⁺ breast cancer associated with the change in risk across interdecile ranges of unconjugated estradiol and the three ratios that showed evidence of estradiol-independent associations with the risk of breast cancer (i.e., 2-hydroxylation pathway:parent estrogens, 2-hydroxylation pathway:16-hydroxylation pathway, and 4-hydroxylation pathway catechols:methylated catechols) (1) are shown in Table 3. The HR for unconjugated estradiol was greater for ER⁺ breast cancer (HR = 2.06) than for total breast cancers (1.66). HRs for the ratios were similar for ER⁺ and total breast cancer.

We focus on the mediating effect of unconjugated estradiol on ER⁺ breast cancer for the following reasons: although several metabolic pathway ratios confounded the BMI–breast cancer association to a similar degree as unconjugated estradiol when unconjugated estradiol was not included in the model (Table 4), unconjugated estradiol was the strongest predictor of ER⁺ breast cancer in our analysis, it was more highly correlated with BMI than other estrogens, estrogen metabolites, and metabolic pathway ratios, and additional adjustment of the BMI–breast cancer association for metabolic pathway ratios after adjustment for unconjugated estradiol did not change the BMI risk estimate (Table 4). In addition, the HR for a 5 kg/m² increase in BMI was greater for ER⁺ breast cancer (HR = 1.28; 95% CI, 1.10–1.49; Tables 4 and 6) than for total breast cancer (1.15; 95% CI, 1.01–1.32), without adjustment for unconjugated estradiol.

The relative percent increases in unconjugated estradiol among overweight and obese cases and controls, relative to women of normal weight, were 17.5% and 32.6%, respectively; with each 5 kg/m² increase in BMI the relative percent increase in unconjugated estradiol was 12.8% (Table 5).

Total, direct (unconjugated estradiol-independent), and indirect (unconjugated estradiol-mediated) effects of overweight, obesity, and a 5 kg/m² increase in BMI on ER⁺ breast cancer risk derived from the Aalen additive risk model and Cox proportional hazards model are shown in Table 6. Also shown is the percentage of the BMI effect mediated through unconjugated estradiol. These analyses were based on 37 ER⁺ cases classified as normal weight, 56 ER⁺ cases classified as overweight, and 50 ER⁺ cases classified as obese. The corresponding number of controls (unweighted) was 103 (38.4%), 92 (34.3%), and 73 (27.2%).

Under the Aalen model, overweight and obesity were associated with 18.3 and 20.2 additional breast cancer cases per 100,000 woman-years, respectively, compared with normal weight; a 5 kg/m² increase in BMI was associated with 7.0 cases of breast cancer per 100,000 woman years. Circulating unconjugated estradiol accounted for 10% (1.8 cases per 100,000) and 15% (3.0 cases per 100,000) of the cases associated with overweight and obesity, respectively, and 20% (1.4 cases per 100,000) of the effect of a 5 kg/m² increase in BMI. The corresponding percentages of indirect effects to total effects based on the Cox proportional hazards model were 7%, 12%, and 19%.

The mean difference between corrected and self-reported weight was 1.18 kg for all subjects combined. The range in the difference was −0.81 to 2.20. The mean difference between corrected and self-reported height was −0.02 meters, with the range from −0.05 to 0.15.

Regression of unconjugated estradiol on corrected BMI and other covariates resulted in a relative increase in unconjugated estradiol of 12.2% for each 5 kg/m² increase in BMI (as compared with 12.8% for uncorrected data, Table 5). Using corrected BMI, the percentage of the BMI effect mediated by unconjugated estradiol was 19 for the Aalen model and 17 for the Cox model. These results are quite similar to those based on the original data (Table 6).

In this study of postmenopausal women who reported never having used menopausal hormone therapy, we found statistically significant positive correlations at baseline between BMI and circulating levels of estrogens and estrogen metabolites, and statistically significant negative correlations between BMI and pathway-related ratios, specifically 2-hydroxylation pathway:parent estrogens and 2-hydroxylation pathway:16-hydroxylation pathway. Of the estrogens and estrogen metabolites, unconjugated estradiol was most strongly associated with increased risk of ER⁺ breast cancer. Approximately 7% to 10% of the effect of overweight and 12% to 15% of the effect of obesity on ER⁺ breast cancer risk was mediated through unconjugated estradiol. Using a continuous BMI measure, approximately 19% to 20% of the increased risk associated with a 5 kg/m² increase in BMI was mediated through unconjugated estradiol. Estrogen metabolism pathways did not appear to explain more of the BMI–breast cancer relationship than that already explained by unconjugated estradiol.

Our results are consistent with other literature in finding positive correlations between circulating estrogen levels and BMI in postmenopausal women (12, 20–22); others have not examined correlations for the full range of estrogens and estrogen metabolites.

In an analysis from the WHI-OS, the only other study to our knowledge which has quantified the mediating effect of estrogens using a counterfactual model, 49% of the effect of a 5 kg/m² increase in BMI on ER⁺ breast cancer was mediated through estrogens (12). Our estimate of mediation (19%–20%) was noticeably lower, but was within the lower bound of the 95% CI for ER⁺ breast cancer (18.8%) in the Hvidfeldt study (12). Our upper bound 95% CI was well within their upper bound of 161.1% (12). We note several differences between our study and the Hvidfeldt study: in our study former hormone users of menopausal hormone therapy were excluded, the median age at baseline was approximately 3 years younger, the unadjusted absolute rate of breast cancer was substantially lower, a 5 kg/m² increase in BMI was associated with a smaller relative increase in estradiol levels, the covariates adjusted for were not identical, and different estrogen assays were used. Median BMIs were comparable in the two studies.

Several studies have assessed the degree to which estrogens mediate the relationship between BMI and risk of all breast cancer by adjusting the BMI-breast cancer association for circulating estradiol levels (9–11). An earlier analysis from the WHI-OS showed only a modest (10%) reduction of the BMI–breast cancer association (10), whereas two other studies reported substantial attenuations (9, 11). Unlike our study and the WHI-OS study (10), these latter two studies reported lower breast cancer risk in the highest category of BMI than in the next lower category. Thus, the degree of attenuation of the BMI–breast cancer association by adjustment for estradiol based on the highest versus lowest categories of BMI and/or changes in risk per 5 kg/m² of BMI, which assumes a linear relationship between BMI and breast cancer risk, may appear more pronounced in these studies.

Enhanced production of estrogens has been hypothesized to be a major mediator of the obesity/breast cancer association. In postmenopausal women androgens produced by the adrenal cortex and ovaries are converted into estrogens by aromatase, which is a complex of enzymes found in adipose tissue. However, obesity is associated with other exposures that may work independently as well as in concert with estrogens in the development of breast cancer. For instance, patterns of adipokine secretion occurring with increased adiposity, namely increased leptin secretion and decreased adiponectin secretion, have been associated with breast cancer risk (23, 24), as have biologic markers of insulin resistance, namely increased insulin and C-peptide levels (24). In fact, in the analysis by Hvidtfeldt and colleagues (12), approximately 66% of the effect of a 5 kg/m² increase in BMI could be attributed to insulin pathways. We did not measure insulin or C-peptide in our samples, and are thus unable to examine these results. We did find that higher levels of insulin-like growth factor (IGF)-I were positively, but not significantly, associated with breast cancer risk in the PLCO dataset (25).

A major strength of our analysis is the sensitive, specific, precise LC/MS-MS assay used to measure unconjugated estradiol, other parent estrogens, and all the estrogen metabolites (26). Other analyses of the mediating effect of estrogens on the BMI–breast cancer association have largely used direct and indirect radioimmunoassays (9–12, 27). In recent years, it has become widely recognized that many radioimmunoassays for unconjugated estradiol, particularly direct radioimmunoassays (28, 29), are not sufficiently specific, especially at the low concentrations characteristic of postmenopausal women (30). Circulating unconjugated estradiol measured by LC/MS-MS or indirect radioimmunoassay (assays which include a prior purification step) has been more highly correlated with BMI than estradiol measured by direct radioimmunoassay (27). In addition, absolute concentrations of unconjugated estradiol were highest for direct assays, lowest for mass spectrometry assays, and intermediate for indirect assays, suggesting possible cross-reactivity between steroids in the direct assays.

A limitation of the current analysis is the moderate number of breast cancer cases, particularly according to ER status of the tumors. Thus, the CIs around the indirect estradiol-mediated effects of BMI on ER⁺ breast cancer in the mediation analysis are large. We also had estrogen measurements at only one point in time. The intraclass correlation coefficient for total estradiol over a 6-year period was 0.73 in one study (31), although the authors note that intraclass correlation coefficients in prior studies ranged from 0.36 to 0.69 over shorter time periods. They suggest that their intraclass correlation coefficient may be higher due to their stringent inclusion criteria to insure that women were truly postmenopausal and not perimenopausal or on exogenous hormones, storage of the samples at lower than recommended temperatures to reduce sample degradation, use of the more sensitive indirect extraction-based radioimmunoassay, and separation of the laboratory assay measurement error from the within and between women variations to avoid attenuation of the intraclass correlation coefficient. In any case, any variation in estradiol due to measurement error and possible long-term changes likely led to some underestimation of the extent to which estradiol mediates the effect of BMI in our analysis.

We also lacked data on sex hormone–binding globulin (SHBG), a glycoprotein that binds tightly most of the unconjugated estradiol in circulation. The fraction of unconjugated estradiol not bound to SHBG is considered bioavailable; the fraction not bound to SHBG or albumin, which has a substantially lower affinity for unconjugated estradiol, is considered free. Obesity is associated with higher estrogen levels and lower SHBG levels (9, 11), leading to marked increases in circulating levels of bioavailable estradiol and free estradiol. In two previous studies, adjusting for free and/or bioavailable estradiol reduced the BMI–breast cancer association more than did adjusting for total estradiol (9, 11). Thus, we most likely underestimated the mediating effect of estrogens on the BMI–breast cancer association by not evaluating bioavailable or free unconjugated estradiol. Notably, information on circulating SHBG was also not available in the Hvidfeldt study (12).

In postmenopausal women, circulating androgens, specifically dehydroepiandrosterone sulfate and testosterone, are positively associated with circulating estradiol and with breast cancer risk. Adjustment for androgens in a collaborative pooled reanalysis (9) and in the EPIC cohort (20) mildly attenuated the estradiol–breast cancer association, but estrogens may be on the causal pathway for the androgen association. Alcohol consumption is related to breast cancer risk, but again estrogens may be on the causal pathway. We are not aware of possible confounders of the BMI–estradiol relationship. We cannot rule out, however, the possibility that the assumptions for mediation analysis were not met.

Our study is one of two analyses of BMI and postmenopausal ER⁺ breast cancer risk to attempt to quantify the proportion of the BMI association that is mediated by circulating concentrations of estradiol. Both studies suggest that a portion of the BMI–ER⁺ breast cancer association is mediated by estradiol levels, but that a substantial portion may be mediated by other factors. Neither study, however, measured free or bioavailable estradiol, which may be more important mediators of the BMI–breast cancer association. Furthermore, neither study addressed the possibility that it is tissue levels of estradiol, rather than circulating levels, that are important. Additional investigation of the mediating effect of estradiol and identification of other biologic pathways that may mediate the BMI–breast cancer association are important for both etiology and prevention.

No potential conflicts of interest were disclosed.

Conception and design: C. Schairer, B.J. Fuhrman, R.N. Hoover, R.G. Ziegler

Development of methodology: M.H. Gail, R.G. Ziegler

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): R.N. Hoover

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): C. Schairer, B.J. Fuhrman, J. Boyd-Morin, M.H. Gail, R.G. Ziegler

Writing, review, and/or revision of the manuscript: C. Schairer, B.J. Fuhrman, J. Boyd-Morin, J.M. Genkinger, M.H. Gail, R.N. Hoover, R.G. Ziegler

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): J. Boyd-Morin

Study supervision: R.G. Ziegler

The authors thank the study participants, Screening Center investigators, and staff of PLCO.

This work was supported by the Intramural Research Programs of the Division of Cancer Epidemiology and Genetics of the NCI, NIH; the Division of Cancer Prevention of the NCI, NIH; and contract HHSN261200800001E to Leidos, Inc., from NCI, NIH.

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.