Statistical methods

We analysed PD and dense, non-dense, and breast area after square-root transforming each measure to improve normality of residuals. The use of the square-root transformation of MD in regression models means that absolute differences in areas are smaller at the lower end of the PD range. For example, an increase in √PD from 4 to 5 (change in √PD of 1) is, on the original scale, an increase from 4² to 5², or an absolute increase of 9 (from 16 to 25), but the same change in √PD of 1 from 7 to 8 is, on the original scale, an increase from 7² to 8², or an absolute increase of 15 (from 49 to 64). To aid the interpretation on the square-root scale, √MD can be interpreted by considering areas as squares, thus regression model beta-coefficients represent differences in the width, in centimetres, of the side of the square. Similarly, as PD (percent) is the dense area per 100 cm², √PD is the width of a dense square within a 10 cm × 10 cm square, and therefore beta-coefficients for √PD represent differences in the dense square width, within a 10 cm × 10 cm square.

We examined cross-sectional differences in √MD measures by age and menopausal status using 2 approaches. First, we examined associations by population group and combined estimates using a random effects meta-analysis. Population group estimates were obtained from a multi-level linear regression model to account for repeat MD readings between readers and between batches for the same woman. The percentage of variation in regression estimates across population groups that was due to heterogeneity was examined using the I² statistic [21], for which percentages of 0%, <30%, 30%–60%, and >60% were roughly considered as no, low, moderate, and large heterogeneity, respectively. Second, we pooled all ICMD data and analysed them in a single model by extending the multi-level model to a third level to account for within-population-group correlations. For both approaches, explanatory variables included fixed effects for the 2 primary variables of interest, age (linear) and menopausal status (binary), as well as BMI (linear + quadratic), ever use of HRT (coded as never, past, current, ever [some studies did not distinguish past and current use], or unknown), parity (continuous), age at first birth (categorical: <20, 20 to <25, 25 to <30, ≥30 years), and reader (categorical). All analyses were based on participants with non-missing data for all variables, as completeness was high (>99%), as per eligibility for inclusion. For pooled analyses (the second approach), factors with between-study variation were additionally adjusted for, i.e., mammogram view and image type. The pooled analysis was also used to examine associations with age in 5-year bands, and linear associations with age were compared to the best-fitting fractional polynomial models.

Interactions of age and menopausal status with demographic and lifestyle factors were tested using likelihood ratio tests. We also examined effect modification by ‘breast area for BMI’, a proxy for a woman’s breast size independent of her BMI, which we generated as the residual of the regression of √breast area on BMI (linear and quadratic terms) and mammography view. Sensitivity analyses excluding women with very early or late menopause, those under 40 or over 70 years, and women with self-reported versus measured BMI were also conducted.

Menopause

Menopausal differences showed no or low inconsistency across population groups for PD (I² = 16.5%) and dense (I² = 26.5%), non-dense (I² = 0%), and breast area (I² = 0%) (Figs 2A, S4A, S5A and S6A). The combined effect estimates for the menopausal difference showed lower PD and smaller dense area in postmenopausal compared to premenopausal women of the same age, but larger non-dense area and slightly larger breast area. These differences did not differ by ASR category (Fig 2A; p = 0.31 for PD). The findings from the pooled analyses closely matched the meta-analytic results, but confidence intervals (CIs) were narrower (Tables 2 and S2): the √PD difference was −0.46 cm (95% CI: −0.53, −0.39), which was driven by a smaller √dense area (difference −0.55 cm [95% CI: −0.65, −0.45]) and larger √non-dense area (difference +0.32 cm [95% CI: 0.21, 0.43]). However, in contrast to the meta-analysis combined estimate, breast area was not significantly larger in postmenopausal compared to premenopausal women (0.04 cm [95% CI: −0.07, 0.14]) (S2 Table).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Association of square-root percent mammographic density with menopausal status and age.

Associations of square-root percent density, by population group, with (a) menopausal status, (b) age among premenopausal women, and (c) age among postmenopausal women, meta-analysed overall and by ASR in source population (low, medium, high). Associations are adjusted for age (for [a] only), BMI, BMI2, parity, age at first birth, HRT use (never, current, former, ever, unknown), mammography view, and MD reader. Full names and details of studies/population groups presented in this figure are provided in S1 Text. Chile is excluded from (a) and (c) as all women were premenopausal. Norway, Australia (Greek), and Australia (Italian) were not included in (a) and (b) as all women were postmenopausal. Turkey was excluded from (a) as selection of women implied that age completely determined menopausal status. ASR, age-standardised incidence rate; BMI, body mass index; CI, confidence interval; HRT, hormone replacement therapy; MD, mammographic density; PD, percent mammographic density.

https://doi.org/10.1371/journal.pmed.1002335.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Difference in square-root mammographic density measures in postmenopausal compared to premenopausal women and by time since menopause: Overall and in subgroups (pooled analyses).

https://doi.org/10.1371/journal.pmed.1002335.t002

The menopause-associated PD difference held across the subgroups of women examined. Where effect modification was present, the associations differed in magnitude, but not direction (Tables 2 and S2). For PD and dense area, interactions were found with parity, age at first birth, HRT use, BMI, and breast area for BMI (e.g., tests for interaction with PD, p = 0.11, 0.04, 0.08, 0.005, and 0.06, respectively; Table 2), and were independent of each other. The magnitude of the menopause-associated difference in dense area, and consequently in PD, was larger in women with lower breast cancer risk profiles, i.e., larger in women who had had an early age at first birth (difference in √dense area of −0.84 cm if <20 years and −0.30 cm if ≥30 years at first birth; Table 2), in women with higher parity, and in never and past HRT users (and null in current users). Across the broad ethnic groups, menopause–MD associations were of a similar magnitude, except for a larger decline in √PD (−0.75 cm [95% CI: −0.93, −0.57]) in the South Asian and Malay group. For effect modification by BMI, in contrast to the aforementioned interactions, whose effects on PD and dense area were in similar directions, the largest menopause-associated differences in dense area occurred in women with a BMI ≥ 30 kg/m², whilst these women tended to have the smallest menopause-associated differences in PD (Table 2). Amongst women with a BMI ≥ 35 kg/m², postmenopausal women had smaller breast areas than premenopausal women, in contrast to findings of no difference in other categories (S2 Table).

Age

Meta-analyses (Fig 2B and 2C) and pooled analyses (Table 3) showed very similar age–MD associations. For a 10-year increase in age, the difference in √PD was −0.24 cm among premenopausal women and −0.38 cm among postmenopausal women (p for interaction by menopausal status = 0.15). These associations reflect similar inverse √dense area–age associations (−0.27 and −0.32 cm for pre- and postmenopausal women). However, whereas in premenopausal women PD differences occurred without a difference in breast area, in postmenopausal women √breast area was 0.34 cm (95% CI: 0.24, 0.39) larger per 10-year age increase (S3 and S4 Tables; S5B and S5C Fig). PD–age associations had low inconsistency across population groups (I² ~ 30%) and across ASR categories; in postmenopausal women, all point estimates were negative.

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Difference in square-root mammographic density measures with a 10-year difference in age, in pre- and postmenopausal women.

https://doi.org/10.1371/journal.pmed.1002335.t003

Fig 3 shows overall and group-specific fitted associations of MD measures with age, including a main effect of menopausal status and a menopause-by-age interaction term to allow for different age associations (slopes) in pre- and postmenopausal women. Predictions were made by age, assuming premenopausal up to age 50 years and postmenopausal thereafter; thus, the change at age 50 years represents menopausal difference. In contrast to the multiple effect modifiers found for the PD–menopause difference, MD–age associations did not differ substantially between women grouped by most reproductive factors (S3 and S4 Tables; Fig 3). In Fig 3 the few exceptions include a smaller age-associated PD difference in current and ever users of HRT compared to never users (p for interaction = 0.045; S4 Table), which appeared to be driven by a differential association with dense area (p = 0.007) and not with non-dense or breast area (p = 0.46 and 0.23, respectively). Additionally, among premenopausal women, there was some evidence that the inverse √dense area–age association was stronger in nulliparous than parous women (−0.68 versus −0.25 overall, p for interaction = 0.06), and whilst there were no differences by ethnicity in premenopausal women, in postmenopausal women the PD–age slope was shallowest for black women (Fig 3).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Modelled associations of square-root percent density, dense area, and total breast area with age and menopausal status, overall and by subgroups.

Square-root dense/breast area is the width of a square representing the dense/breast area; square-root PD is the width of a dense-area square within a 10 cm × 10 cm square. All models are adjusted for age, BMI, BMI2, parity, age at first birth, HRT use (never, ever, past, current, not known), MD reader, image type, and mammography view. BMI, body mass index; CI, confidence interval; ICMD, International Consortium on Mammographic Density;HRT, hormone replacement therapy; MD, mammographic density.

https://doi.org/10.1371/journal.pmed.1002335.g003

Further examination of MD–age associations using the best-fitting flexible fractional polynomials revealed that such non-linear models were not a better fit than the simplest linear models for PD (p = 0.42), dense area (p = 0.90), and breast area (p = 0.09) in premenopausal women and for all 4 MD measures in postmenopausal women (p = 0.23, 0.66, 0.53, and 0.91 for PD, dense, non-dense, and total breast area, respectively), thus suggesting that there is no plateauing of MD within this age range. For non-dense area in premenopausal women, a fractional polynomial model had a better fit than a linear model (p = 0.008), as the positive association with age was not seen until around age 45 years. In sensitivity analyses, the PD–age association in premenopausal women was attenuated from −0.24 to −0.18 if older premenopausal women (>50 years) were excluded, and postmenopausal PD–age associations were stronger (increased from −0.38 to −0.47) when women over age 65 years were excluded (estimates are provided here and do not appear in tables).

Ageing and the menopause

In postmenopausal women, there is sufficient variability in age at menopause (IQR 45 to 52 years) to analyse the impact of age-adjusted years since menopause, which was independently associated with √PD (−0.10 per 10 years [95% CI: −0.17, −0.04]), an effect that was present up to 15 years after menopause (Table 2). After adjusting for years since menopause, the PD–age association reduced from −0.38 to −0.29 per 10 years, but remained significant.

Main findings

Across diverse populations of breast-cancer-free women worldwide, we assessed how MD is related to age and menopausal status. Overall, despite inevitable differences between contributing studies, we found consistency in the direction of associations of age and menopausal status with PD and its component tissues across ICMD women. Assuming that inferences from these cross-sectional data reflect within-woman changes, the greatest reduction in MD occurred upon menopausal transition, when dense area declined from a mean of 16.0 to 12.5 cm², equivalent to 15 to 20 years of ageing. This difference in PD between post- and premenopausal women of the same age was present up to 15 years after the menopause and may be greater in women with lower breast cancer risk profiles (early age at first birth, high parity). PD also decreased with increasing age both among premenopausal and postmenopausal women. In premenopausal women, these declines were entirely due to breast compositional changes, i.e., due to a decline in dense area and an equal but opposite increase in non-dense area. However, in postmenopausal women, PD reductions were also accompanied by increasing breast area and continued beyond 70 years of age.

Comparison to other studies

The large decrease in MD at menopause after adjusting for age is largely consistent with previous observations in westernised countries, either from within-woman or between-woman comparisons [2,8–10,12,15]. Boyd et al. [2] examined differences in breast area before and after menopause and found no significant change, indicating that lower PD after menopause may be due a compositional change alone. Only 1 longitudinal study reported a decline in PD at menopause that was non-significant, although in this study the largest age-related decline did occur around the age of menopause (between ages 45 and 55 years) [11].

Comparisons of age-related changes have been more variable. In ICMD, we found no strong evidence of different PD–age or dense area–age slopes in pre- and postmenopausal women, in agreement with the full longitudinal data from 1 of the ICMD studies, the Melbourne Collaborative Cohort Study/Australian Breast Cancer Family Registry [22], but differing from studies that reported steeper MD declines in either premenopausal [14] or postmenopausal women [9,11]. In addition, there was no strong evidence of non-linear MD–age trends in ICMD, but 3 longitudinal studies have found that PD (and dense area, where examined) declines most rapidly around ages 45 to 55 or 60 years [8,11,12], including the Melbourne study, which also observed a possible plateau after 65 years of age [11,12,22].

Biological plausibility

The remarkable consistency of the age and menopausal effects on MD across diverse female populations spanning all continents suggests that these associations reflect a biological process common to all women. The extent of the decline in MD may be modified by ‘external’ factors, such as reproductive or other lifestyle or environmental factors. These observations lend credibility to the hypothesis that the sharp decline in MD over the menopause is a phenotypic marker of the decline in the rate of breast tissue ageing [4], which results in Clemmesen’s hook in the age–breast cancer incidence curves across the world. However, this simple parallel ignores the well-established molecular and aetiological heterogeneity of breast cancer [5]. Notably, although it has previously been hypothesized that the hormonal drivers of PD would result in it being a stronger risk marker for oestrogen receptor (ER)–positive than ER-negative breast cancer, a large differential in the MD-breast cancer association by tumour subtype does not appear to be present [23]. This finding should perhaps not be surprising, if the menopausal decline in PD, observed universally here, is related to the incidence curves of breast cancer with age, which feature a much stronger menopausal decline in the age-acceleration of incidence rates for ER-negative tumours than for ER-positive tumours [5,24]. At the tissue level, changes involving the dense area arise from changes in stromal tissue and/or in epithelial tissue, whilst changes in PD are additionally affected by changes in adipose tissue in the breast. Declining dense area likely reflects involution of terminal duct lobular units. X-rays of histological tissue show that these units have raised concentrations of radiologically dense areas and that these areas also decline with age [25]. Two additional observations link MD, terminal duct lobular units, and breast cancer risk: breast carcinoma arises within epithelial tissue and the radiologically dense tissue seen on a mammogram [26], and, within cohorts with benign breast disease, greater terminal duct lobular unit involution is predictive of subsequent breast cancer risk [27].

At a systemic level, a hormonal pathway is most certainly implicated in MD changes that occur during the menopausal transition. Several reproductive hormones decline at different stages of the transition [28], some of which may contribute to the MD decline. Combined exogenous oestrogen and progesterone are known to increase PD [29], whilst in premenopausal women, endogenous oestrogen levels are positively associated with PD [30]. The shallower age-related declines in PD in premenopausal women compared to postmenopausal women in our study may reflect similar hormonal pathways. Further, in postmenopausal women only, an increase in breast area also contributed to the decreasing PD. This is likely to reflect a true increase in breast volume as breast area and thickness were positively correlated in ICMD (r = 0.18).

Strengths and limitations

The present study is greatly enhanced by the heterogeneity of the study populations included in ICMD, in terms of a wide range of BMI values, reproductive characteristics, ages (40-year age range), and ethnicities and their associated range of breast sizes and composition, making this, to our knowledge, the largest multi-ethnic, multi-country study of MD to date. The consistent findings are likely to reflect intrinsic biological processes and should therefore apply to the general population of women worldwide. Our separate analyses of tissue components revealed important differences in the drivers of changes in breast tissue composition at different life stages. However, inherent to the consortial nature of ICMD is the fact that we relied on existing data; thus, some sample sizes were smaller, not all studies included women spanning the full age range, and there is variability in image types and exposure variable definitions. Menopause was included as binary variable due to the limitations of the data, but may occur over many months or years. The definition of menopausal status also varied, from simple self-reports to complex algorithms, measured at or some time before or after mammography. Potential misclassification bias might attenuate the estimate of the menopausal effect but is unlikely to affect within-population group estimates of the menopause difference. Additionally, including some perimenopausal women would also underestimate a menopausal effect. In the 31% of women with self-reported anthropometry, possible underreporting of weight and overestimation of height, particularly in older women, would lead to an underestimation of BMI [31]; thus, part of the increase of breast area postmenopausally may be due to incomplete adjustment for BMI. In addition, the validity of BMI as a measure of total body adiposity across such diverse populations is uncertain.

Women in ICMD were recruited from a range of settings, including some that were not population-based screening; thus, there is likely an overrepresentation of symptomatic or high-risk women. Nevertheless, the present findings rely on the representativeness of the associations and not of the sample populations; thus, any lack of representativeness of study populations would affect results only if there was an age-related pattern to representativeness. In a similar manner, cross-sectional associations may only partially reflect within-woman changes, as birth cohort effects on MD, through factors unadjusted for, may confound associations. These findings should be verified with longitudinal data in which within-woman changes can be examined.

Implications

There are broader implications of the profile of MD with ageing. Similar to the reduced sensitivity of mammography in younger women than in older women in organised screening programs [32], because of the former group’s higher average breast density, this problem is also likely to affect the performance of mammography as part of the diagnostic work-up of symptomatic breast cancers. Low- and middle-income countries may be disproportionately affected by this issue, because of the young population structures and thus younger average age at breast cancer diagnosis. Further, if MD reflects the rate of breast tissue ageing, the parallel to Pike’s model suggests that cumulative MD to a given age is the pertinent marker of breast cancer risk. Indeed, cumulative density has been shown to be highly correlated with age-related breast cancer risk [11], and women whose MD decreases substantially over time may be at lower risk of breast cancer than those who maintain the same density [10,33]. More efficient and cost-effective stratified approaches to population-based breast cancer screening using MD-based risk scores have been suggested [34]. Placing a woman in a particular MD-based risk stratum would be aided by the fact that MD displays a high level of tracking within a woman over time, at least postmenopausally [12], but it is not known whether risk predictions might be improved if earlier life periods, before MD tracking is established, were included.

Finally, based on the current findings that MD is in general highest earlier in life, primary prevention efforts might best be targeted to younger ages, reducing breast density earlier in life, leading to a lower cumulative MD across life. Studies of breast density in girls or younger women that use MRI or dual-energy X-ray absorptiometry (DXA) imaging are fewer. They have found suggestions of increased breast density associated with greater birth size [35], duration of use of hormonal contraceptives [36], higher testosterone levels [37], and lower childhood BMI, independent of current BMI [38]. At adult ages, the potential actions needed to lower breast density for a reduction in breast cancer risk, apart from the known pregnancy-associated effects, include lowering alcohol intake [39], whilst the protective effect of physical activity does not appear to be mediated through MD [40]. Further, in high-risk women and in women with ductal carcinoma in situ, tamoxifen reduces breast cancer risk through a reduction in MD [41]. Given the relatively low uptake of such oral chemoprevention therapies, other risk-lowering options for women with high breast cancer risk and raised MD may come from the ongoing trials of oral low dose and topical tamoxifen [42]. The consistency of the MD–ageing associations internationally is suggestive of an intrinsic shared biology, which might also extend to other modifiable determinants of MD.