A. General method. We carried out a cross-sectional observational study in which we recruited young women aged 15-18 years. Breast water and fat in participants were measured with MR, and anthropometric variables were measured. Diet was assessed by food frequency questionnaire, and glucose and insulin were assayed in fasting serum.
B. Recruitment of subjects. Recruitment of young women was from Toronto high schools, and family practices. Recruitment took place in two phases, Phase 1 between 2003 and 2007 and Phase 2 between 2010 and 2015. Subjects in Phase 1 have been included in a previous publication 8. Ethics approvals were obtained from the University Health Network, Sunnybrook Hospital, and Women’s College Hospital (all in Toronto) and from the Toronto District School Board, the Toronto Catholic District School Board, the York Region District School Board and the York Catholic District School Board.
1. School recruitment. With the consent of school principals we contacted their health and/or science departments, which were asked to agree to our approaching their students in their classrooms, to present a cancer educational talk and introduce the research study. Eligible students interested in participating were given an information package to take home that described the study and included consent forms for both mother and daughter. On receipt of written consent from both mother and daughter we contacted them to set up appointments.
2. Family practice recruitment. Participating family practices contacted young women aged 15-18 on their patient lists, introduced the study and invited their participation. Those interested were mailed the same information package and consent forms that were used in schools.
3. Inclusion and exclusion of subjects. Subjects were excluded if they had not established menses, had used oral contraceptives within the previous 6 months, had been pregnant, had breast implants, augmentation or reduction mammoplasty, or previous breast cancer. To avoid population stratification, we recruited only subjects who by self-report were white Caucasians. Exclusion criteria for the safe use of MR, included recent breast surgery, metal implants of any sort, known claustrophobia, and weight more than 200 lbs.
C. Measurements.
1. Menstrual and reproductive characteristics. Information on demographic and risk factors for breast cancer was obtained by questionnaire. This included age at the onset of menses, and details of prior exposure to oral contraceptives.
2. Anthropometric measures. A research assistant, trained by an instructor from the Department of Physical Education, University of Toronto, measured height, weight and waist and hip circumferences. Skinfold thickness was measured using calipers at subscapular, triceps and supra-iliac sites. Percent body fat was calculated using methods validated in adolescent girls 9,10, that use skinfold thickness measurements at triceps (T) and subscapular sites (S).
3. Diet. Dietary intake was assessed using a food frequency questionnaire adapted from the US DHQ II to optimise the capture of foods consumed by Canadians22.
4.Magnetic resonance (MR) measures of breast water and fat. The first 181 young women included in the analysis were examined using a 1.5T Signa Cvi MR system (GE, Waukesha WI)8. The remaining 776 subjects were examined in a 3.0T scanner (Phillips). All scans were carried out in the prone position with commercially available breast coils from the respective vendors.
To determine if the data from these two scanners could be pooled we scanned 12 healthy volunteers on both scanners on the same day. Total fat and water measures from the two scanners were strongly correlated (R2>0.99) but showed some underestimation of fat in the 3.0T protocol relative to the 1.5T protocol. A correction applied to the T3.0 protocol decreased the discrepancy 11.
With both scanners the sequence was calibrated using a series of home-built phantoms. Bi-monthly scans of three phantoms with known water/oil concentrations and various volumes using the same MR imaging protocol confirmed volume accuracy within 2% and water/oil content accuracy within 3%.
The output of the MR examination was a series of “slices” at 7-mm intervals through both breasts. The breast was distinguished from surrounding tissues on each slice by an observer using a semi-automated image analysis program, and the water and fat within each slice calculated and summed over all slices which acquires the water and fat signals with phase shifts of (0, pi, 2pi). The results shown are measurements in the right breast only and are expressed as percent water, total breast water and total fat. All measurements have been shown to be bilaterally symmetrical 12. A small amount of water is also present in fat and allowance is made for this in calculating the water content of the breast used here.
5. Blood samples for measurement of glucose and insulin. Fasting blood samples were collected in the early morning after a 12-hour fast on the day of the MR examination and within 10 days of the first day of the most recent menstrual period. Serum was separated within 2 hours of collection and stored in 2ml aliquots at -70C until analysis.
6. Assays ofglucose and insulin. All assays in subjects from Phases 1 and 2 of recruitment were carried out in fasting serum at the same time in the Immunochemical Core Lab (ICL), Mayo Clinic Research Core Labs. The assays used, their inter- and intra- assay coefficients of variation, and lower limits of detection are shown in Supplementary material.
7. Calculation of insulin measures. We used fasting serum levels of glucose and insulin for each subject to calculate measures of insulin production, sensitivity and resistance using on-line software: https://www.dtu.ox.ac.uk/homacalculator/download.php version 2.2.3 for the HOMA model, that calculates13 measures of insulin sensitivity (HOMA2-S), beta-cell function (HOMA2-B), and insulin resistance (HOMA2-IR).
D. Statistical methods. Nine hundred and fifty-seven young women were recruited and had breast MR and complete anthropometric measures. One hundred and sixty-eight subjects did not provide a blood sample. An additional five subjects whose fasting insulin or glucose values fell outside the limits set by the HOMA software, were excluded from the analysis of blood samples. Eighteen subjects had values for insulin or glucose that were outliers suggesting they were not fasting samples, leaving 766 subjects, of whom 25 did not complete a food frequency questionnaire. The remaining 741 subjects were included in the analysis. Details of these exclusions are given in the Supplementary Figure.
Means and standard deviations were calculated for selected characteristics. Total water and total fat by MR were log transformed for analysis.
The associations of MR breast tissue characteristics with dietary intakes and insulin were examined using univariable and multivariable regression models. The multivariable models included age, height and weight, percent body fat and waist-to-hip ratio at the time of the MR examination.
To illustrate graphically the main findings of regression analysis, we ran multivariable regression analyses with selected variables divided into quintiles. The statistical significance of the differences in breast measures of percent water, total water and total fat associated with increasing levels of these variables were assessed using tests for linear trend.
In mediation analysis, we examine the effect of an exposure on an outcome, which may be mediated by an intermediate variable. The exposure influences the intermediate variable, which in turn, affects the outcome14. The intermediate variable is generally referred to as the ‘mediator’ and the set of rules to quantify such effects is known as ‘mediation analysis’ 14,15 proposed a set of techniques to quantify the effect of an exposure on an outcome which is mediated by a mediator, namely the indirect effect, based on the assumption that there is no interaction between the exposure and the mediator. In the present context, we found that the interaction terms were non-significant and thus, we followed this approach. The details are described elsewhere (Baron and Kenny, 1986). This approach has been extended by several authors to allow for interaction between the exposure and the mediator 16,17.
The analyses were conducted using statistical software R version 2.15.1. All p-values were calculated from two-tailed tests of statistical significance. Statistical significance was declared at the 5% level. All p-values are adjusted for the variables shown in the Table footnotes.