Study participants
The present analysis was conducted within two studies: a case control study from the UK and a cohort study from Norway.
UK study
The study methodology of the UK case control study is described in detail elsewhere [
9]. In short, cases (
n = 414) were women with newly diagnosed breast cancer at the Royal Marsden Hospital (RMH), London, between April 2010 and July 2012. Controls (
n = 685) were women screened and found to be breast cancer-free at the Central and East London Breast Screening Service (CELBSS) in the same time period. The CELBSS invites women between 50 and 70 years of age for mammographic screening once every 3 years as part of the English National Health Service Breast Screening Programme. Women over 70 years can optionally contact the service for a self-referred appointment every 3 years.
Data on breast cancer risk factors, including age, ethnicity, parity, menopausal status, use of oral contraceptives, use of hormone therapy and self-reported height and weight, were collected by a self-administered questionnaire at the time of screening for controls and within 15.5 months of the diagnostic mammography for cases. BMI was calculated as weight in kg/(height in m)
2. Ethnicity was categorised in accordance with the census classification as “White”, “Black” (African or Caribbean), “Asian” (Indian, Pakistani or Bangladeshi) and “Other” [
10].
Participants underwent full-field digital mammography, with two views (cranio-caudal (CC) and mediolateral-oblique (MLO)), of both breasts. The images were taken on Senographe DS machines (GE Healthcare, Slough, UK). The anonymised raw images were analysed by using Volpara version 1.0 (Matakina Technology Limited, Wellington, New Zealand) [
11]. This algorithm provided fully automated estimates of the volumes (all in cm
3) of the total breast (BV), non-dense (fat) tissue (NDV) and dense (fibro-glandular) tissue (DV) separately for each one of the four breast/view images, and percent mammographic density (%MD) was estimated as DV/BV×100.
Norwegian study
The Cancer Registry of Norway is responsible for the administration of BreastScreen Norway (the Norwegian Breast Cancer Screening Program). All women within a targeted age-range of 50–69 years resident in the country are invited to undergo mammography screening every 2 years. From August 2006 to 2014, women who underwent mammographic screening in the nationwide programme were asked to complete a questionnaire on a number of standard breast cancer risk factors and a second questionnaire on current exposure to risk factors. Included in the present study were women who participated in BreastScreen Norway in four counties, had information on volumetric mammographic density from their first mammographic screening between 2007 and 2014, and had completed both questionnaires. However, for the second questionnaire on current exposures, if the questionnaire or certain values were missing, information from the questionnaire completed at a previous screening round was used (approximately 16.5%). The cohort consisted of 61,716 women, including 657 women who were diagnosed with a first occurrence of breast cancer during a median follow-up from date of screening of 3.84 (interquartile range 2.08, 4.83) years. Women with a previous diagnosis of breast cancer (n = 970), a ductal carcinoma in situ (DCIS) diagnosis up to 6 months after the screening date (n = 224) and a bilateral breast cancer (n = 11) were excluded.
In a similar manner to the UK study, all women had standard two-view full-field digital mammography of each breast with Senographe DS or Senographe Essential machines (GE Healthcare) or MDM L50 or MDM L30 machines (Phillips). The raw images were read by Volpara version 1.5.0 (Volpara Health Technologies Limited, Wellington, New Zealand) to obtain, similarly to the UK study, volumetric estimates of BV, NDV, DV and %MD.
Statistical methods
Descriptive analysis of UK controls and the full Norwegian cohort included examination of the distributions of BMI and volumetric mammographic measurements. For these analyses, measurements were averaged over the four images (that is, left and right CC and MLO images). Natural-log transformations were applied to average %MD, DV, NDV and BMI to normalise their distributions. Scatter plots and Pearson’s correlation coefficients were used to examine BMI associations with %MD, DV, NDV and BV. BMI and each mammographic measure were regressed on age at mammogram, parity and menopausal status using linear regression (including controls only in the UK study and the full cohort in the Norwegian study). Pearson’s correlation coefficients between the residuals derived from these models were then calculated (and denoted r) to allow examination of correlations that are not influenced by these variables.
For the UK case control analysis, the average density measures from the CC and MLO images from the unaffected breasts for cases and for a randomly selected breast for controls were used. In order to compare the association of %MD and DV with the odds of breast cancer, after adjusting for different sets of confounders, three different logistic regression models were fitted where these two exposures were first standardised as recommended previously [
12]. The resulting estimates are referred to as OPERA ORs (“odds ratios per adjusted standard deviation”) and are effects per residual standard deviation of the exposure once its association with the confounders is accounted for. Estimation requires first fitting a linear regression model of the exposure on the confounders and then using the standardised residuals derived from this model as the exposure of interest in logistic regression models that include the same confounders. Fifty-one cases and 38 controls (8.1% of the study participants) were excluded from all logistic regression analyses because they were missing at least one of the variables used in the modelling.
The first (minimally adjusted) model controlled for age (continuous), menopausal status (pre-, peri/post-) and parity (yes/no). (Further adjustment for ethnicity, use of exogenous hormones and the other variables listed in Table
1 was also considered but it is not shown as it yielded similar results.) Second, a model was fitted that additionally adjusted for self-reported BMI. Finally, an alternative model was fitted that additionally adjusted for log-transformed NDV in place of BMI. Adjustment for BV instead of NDV was not considered because, albeit this variable is highly correlated with NDV (
r = 0.99;
P = 0.001), its interpretation is made more difficult by the fact that it reflects both DV and NDV.
Table 1
Baseline characteristics of the participants by status in the UK and Norwegian studiesa
Age at mammography |
Mean (SD) | 59.5 (6.6) | 67.5 (12.7) | 56.9 (5.74) | 57.7 (5.43) |
Number | 679 | 412 | 61,059 | 657 |
BMIb |
Mean (SD) | 26.1 (5.6) | 26.4 (4.9) | 25.6 (4.2) | 25.8 (4.1) |
Number | 656 | 368 | 54,345 | 589 |
Ethnicity (UK)/Country of birth (Norway), n (%) |
White/Norway | 520 (76.5) | 370 (89.4) | 56,234 (93.8) | 612 (94.2) |
Non-white/Outside Norway | 160 (23.5) | 39 (9.6) | 3693 (6.2) | 38 (5.8) |
Missing | 5 | 5 | 1132 | 7 |
Family history of BC, n (%) |
No | N/A | N/A | 45,168 (77.1) | 447 (70.0) |
Yes | N/A | N/A | 13,390 (22.9) | 192 (30.0) |
Missing | | | 2501 | 18 |
Menopausal statusc, n (%) |
Pre- + peri-menopausal | 91 (13.3) | 55 (13.3) | 14,776 (25.2) | 141 (22.1) |
Post-menopausal | 591 (86.7) | 358 (86.7) | 43,856 (74.8) | 496 (77.9) |
Missing | 3 | 1 | 2427 | 20 |
Parity, n (%) |
Nulliparous | 209 (30.9) | 65 (15.9) | 4946 (8.5) | 57 (9.0) |
Parous | 467 (69.1) | 343 (84.1) | 53,563 (91.5) | 577 (91.0) |
Missing | 9 | 6 | 2550 | 23 |
Age at menarche in years, n (%) |
<13 | 271 (53.9) | 159 (54.1) | 16,764 (40.9) | 186 (41.9) |
14+ | 232 (46.1) | 135 (45.9) | 24,202 (59.1) | 258 (58.1) |
Missing | 14 | 33 | 4107 | 43 |
Hormone therapy use, n (%) |
No | 459 (68.8) | 246 (63.2) | 34,150 (66.2) | 305 (55.6) |
Yes | 208 (31.2) | 143 (36.8) | 17,418 (33.8) | 244 (44.4) |
Missing | 18 | 25 | 9491 | 108 |
Educational level, n (%) |
None/primary school | 35 (5.2) | 17 (6.2) | | |
Lower secondary | | | 13,772 (23.3) | 164 (25.9) |
Secondary or higher | 641 (94.8) | 225 (93.8) | 45,457 (76.7) | 470 (74.1) |
Missing | 9 | 142 | 1830 | 23 |
Breastfeeding among parous women, n (%) |
Yes | 358 (76.7) | 224 (74.7) | 46,107 (99.9) | 497 (100) |
Missing | 3 | 43 | 9929 | 103 |
In the Norwegian cohort study, average density measures were based on log-transformed average values of the CC and MLO readings from the unaffected breast for cases and from a randomly selected breast for non-cases. Cox regression proportional hazards models were fitted to the cohort data, using age as the time-scale, to evaluate the associations of (log-transformed and standardised as described above for the UK study) %MD and DV with breast cancer risk, expressed in terms of hazard ratios and referred to as OPERA HRs.
Three different models were fitted as in the UK study; the first was minimally adjusted for screening year (categorised using 2-year intervals), menopausal status (pre-, peri-, post-) and parity (yes/no) (further adjustment for country of birth as a proxy for ethnicity did not affect the findings). A second model was additionally adjusted for BMI, and a third model was additionally adjusted for NDV in place of BMI. In all, 10,288 participants, including 99 cases, were excluded from all three models because they missed data for at least one of the variables listed.
Three further models were also fitted to the Norwegian data using the full reproductive and lifestyle risk factor questionnaire data collected in this study (that is, screening year category, menopausal status, parity, age at menopause, age at menarche, age at first birth, duration of breastfeeding, use of hormone therapy, family history of breast cancer, education, smoking, alcohol use and physical activity level). In the first model, BMI was omitted; in the second model, BMI was included; in the third model, NDV was used instead of BMI. In total, 25,833 (41.9% of the original cohort) women with missing data on any of the variables examined were excluded to ensure that these additional models were fitted to the same subset of women. Departure from the proportional hazards assumption underlying each of these fitted models was evaluated by using tests based on Schoenfeld residuals. The Akaike information criterion (AIC) corresponding to each multivariable model from the two countries is also reported.
Similar analytical steps were followed to study the associations between BMI and breast cancer risk, and then NDV and breast cancer risk, in both studies, in each case adjusting for age, menopausal status and parity.
Fixed-effects models were used to obtain pooled summary OPERA relative risk (RR) estimates from the two studies. Between-study heterogeneity was assessed by the Q statistic and the
I2 statistic [
13].
In all the analyses, we considered statistical significance (two-sided) at a
P value of less than 0.05. All analyses were conducted in Stata (IC 14 for the statistical analysis of the UK data and the meta-analysis and IC 15 for the analysis of the Norwegian data) [
14].