Longitudinal Association of Anthropometry with Mammographic Breast Density in the Study of Women's Health Across the Nation (Swan)

Abstract

High percent mammographic breast density is strongly associated with increased breast cancer risk. Though body mass index (BMI) is positively associated with risk of postmenopausal breast cancer, BMI is negatively associated with percent breast density in cross-sectional studies. Few longitudinal studies have evaluated associations between BMI and weight and mammographic breast density. We studied the longitudinal relationships between anthropometry and breast density in a prospective cohort of 834 pre- and perimenopausal women enrolled in an ancillary study to the Study of Women's Health Across the Nation (SWAN). Routine screening mammograms were collected and read for breast density. Random intercept regression models were used to evaluate whether annual BMI change was associated with changes over time in dense breast area and percent density. The study population was 7.4% African American, 48.8% Caucasian, 21.8% Chinese, and 21.9% Japanese. Mean follow-up was 4.8 years. Mean annual weight change was +0.32 kg/year, mean change in dense area was -0.77 cm²/year, and mean change in percent density was -1.14%/year. In fully adjusted models, annual change in BMI was not significantly associated with changes in dense breast area (-0.17 cm², 95% CI -0.64, 0.29). Borderline significant negative associations were observed between annual BMI change and annual percent density change, with percent density decreasing 0.36% (95% CI -0.74, 0.02) for a one unit increase in BMI over a year. This longitudinal study provides modest evidence that changes in BMI are not associated with changes in dense area, yet may be negatively associated with percent density.

Introduction

In 2008 alone an estimated 182,460 U.S. women will be diagnosed with breast cancer.¹ Though substantial efforts have been devoted to studying breast cancer etiology and prevention, the pace of this research is often slow due to the decades needed for breast cancer to develop. Therefore, prospective cancer epidemiology studies are incorporating surrogate endpoints, which allow for studies to be conducted with fewer subjects over a shorter time period while aiding in the understanding of the mechanisms of cancer development.²

Mammographic breast density has been used as a surrogate endpoint for breast cancer in a number of studies.^3-5 The breast is composed of different types of tissues, and the composition varies from woman to woman; fat appears dark on a mammogram because it is radiologically lucent, while epithelial and connective tissues appear bright because they are radiologically dense. Mammographic breast density is determined by the relative proportions of fat and structural tissues in the breast. Two measures of mammographic density are dense breast area and percent density. The areas of density can be measured and summed to quantify the dense breast area. Percent breast density is calculated as the dense breast area divided by the total breast area expressed as a percentage. High percent breast density is associated with a fourfold risk of breast cancer compared to women with low breast density.⁶ In the few studies that have examined the association between area of dense tissue and breast cancer risk, associations have been found to be slightly weaker than the associations with percent density.⁷

Body mass index (BMI) bears an idiosyncratic relationship to breast cancer and breast density. BMI is related to an increased risk of postmenopausal breast cancer and a decreased risk of or no association with premenopausal breast cancer.⁸ However, cross-sectional studies of mammographic breast density in both pre- and postmenopausal women consistently report that increased weight or BMI is associated with lower percent breast density.^{7, 9-15} For example, one study reported percent breast density was 5.2% lower among premenopausal women and 4.7% lower among postmenopausal women in the 3^rd quartile versus 1^st quartile of BMI.¹¹

The amount of fat present in the breast is strongly related to BMI.¹⁶ Recent studies have evaluated the association between anthropometric measures and the dense and non-dense areas of the breast.^{13, 17-20} Several studies have shown that weight and BMI are positively associated with the size of the non-dense area;^{13, 17-20} yet for the dense breast area, three studies reported negative correlations,^{17, 18, 20} one reported a non-significant positive association,¹³ and another conducted in the same SWAN study population as the present study reported that the direction of the association varied by race/ethnicity.¹⁹ Only a few studies have used longitudinal data to evaluate associations between anthropometric measures and breast density.^{3, 15, 21} These longitudinal studies have been limited by small numbers, self-reported anthropometric measurements, or use of a non-quantitative measure of breast density.

In the current analysis we analyzed data from a prospective cohort study of 834 women enrolled in an ancillary study to the Study of Women's Health Across the Nation (SWAN). Our primary hypothesis was that changes in BMI and weight would be positively associated with changes in dense breast area and negatively associated with changes in percent density over time.

Materials and Methods

Study population

SWAN began in 1994 to study the health of women as they transition from premenopause to postmenopause.²² Briefly, women eligible for SWAN were age 42-52, had ≥1 menstrual period in the previous three months, had a uterus and ≥1 intact ovary, and were not currently taking hormone therapy (HT) or oral contraceptives (OCs). Participants in SWAN attended a baseline clinic visit followed by annual clinic examinations and questionnaires. The SWAN Mammographic Density Substudy is an ancillary study to SWAN with the goal of examining factors related to mammographic breast density and how mammographic breast density changes as women progress through the menopausal transition. Separate written informed consent and institutional review board approvals were obtained for this ancillary study. Participants from three SWAN sites (Los Angeles, CA, Oakland, CA, and Pittsburgh, PA) representing four races/ethnicities (African American, Caucasian, Chinese, and Japanese) were enrolled at their 5^th or 6^th annual follow-up visit. Caucasian participants were enrolled from all three sites, while all African Americans were from the Pittsburgh site, all Chinese from the Oakland site, and all Japanese from the Los Angeles site. Eligible mammograms included routine screening mammograms that were taken two years prior to the baseline SWAN visit through two years after the 6^th annual SWAN follow-up visit. Mammograms of breasts that had undergone a biopsy or more extensive surgery were ineligible for the SWAN Mammographic Density Substudy. Of those eligible for the SWAN Mammographic Density Substudy, 86.1% (N=1,055) consented and at least one mammogram was obtained from 95.3% of these women (N=1,007).

Participants were excluded from this analysis if they reported a history of breast cancer at SWAN enrollment (N=6) or had only one available mammogram (N=139). Women who were diagnosed with breast cancer during their SWAN follow-up (N=21) were censored at the visit they reported this diagnosis. Ten of these women had no mammograms prior to their breast cancer diagnosis and were excluded from this analysis. After these exclusions 852 participants and 3,784 mammograms remained in the analysis. Three participants reported being either pregnant or breastfeeding at the time of a SWAN visit, and these specific visits were not included in the analysis.

Mammographic density readings

The mammographic density assessments were performed by a single expert reader using craniocaudal views of the right breast.^23-25 Films of the left breast were used if a participant had a previous surgery to the right breast or if the right breast films were of poor quality. Prior to performing the density assessment, the reader rated the quality of each film as excellent, good, fair, or poor.

Quantitative measures of density were made by tracing the total area of the breast and outlining the areas of density (excluding biopsy scars, Cooper's ligaments, and breast masses) onto clear acetate placed over the mammogram. A compensating polar planimeter (LASICO, Los Angeles, CA) was used to measure the total breast area and the dense breast area in cm². A blinded random sample of films (N=449) was used to assess the reproducibility of the density assessments. This re-review showed good association between the initial and repeat readings of percent density (within-person Spearman correlation coefficient=0.96).

Anthropometric measures

Height was measured without shoes using either a metric folding wooden ruler or measuring tape (home and some clinic visits), or a fixed stadiometer (clinic visits). Weight was measured without shoes and in light indoor clothing, using a portable digital scale (home and some clinic visits) or either a digital or balance beam scale in the clinic. Portable scales were calibrated weekly, and stationary clinic devices were calibrated monthly. BMI was calculated as the weight in kilograms divided by the square of the height in meters. In a few instances, height was not measured at a visit (N=12) or the measured height was considered to be unreliable (N=5) due to significant deviation from prior or later measured heights. In these cases, we used the mode of the participant's other height measurements to estimate the missing or unreliable height value so that BMI could be calculated.

Additional variables

Additional data were collected at the clinic visits by either interviewer- or self-administered questionnaires. These data included demographic information and personal and family history of breast cancer. Cancer history was updated at each SWAN visit. Reproductive variables ascertained at baseline included age at menarche, age at first birth, number of births, and history of breastfeeding. Number of births and breastfeeding history were updated at each SWAN visit. History and number of previous breast biopsies were reported at the time of enrollment into the SWAN Mammographic Density Substudy. These data were used to calculate the Gail score for each participant.^{26, 27} Gail scores of 1.66 or higher indicate a high 5-year risk of breast cancer. History of atypical hyperplasia was not collected in this study and was considered to be “unknown” for each woman in the Gail score calculation.

At baseline women reported their history of HT and OC use, as current users at baseline were excluded; current HT and OC use was updated at each SWAN visit. Menopausal status was ascertained using an interviewer-administered questionnaire at each visit and defined according to SWAN protocol.²² Women were asked about the frequency and regularity of their menstrual bleeding. Women reporting no decrease in their menstrual regularity over the past year were classified as premenopausal. Early perimenopausal was classified as decreased menstrual regularity within the previous three months, and late perimenopausal was defined as no menstrual bleeding in the 3-11 months prior to the interview. Women reporting no menstrual bleeding for at least 12 months prior to the interview were classified as postmenopausal. Women reporting a bilateral oophorectomy and/or hysterectomy were defined as having a surgical menopause. Those women reporting use of HT with some bleeding within the past 12 months were classified as unknown menopausal status. Women classified as postmenopausal remained classified as such thereafter, regardless of their HT use.

Breast density estimation

Mammogram data and SWAN visit data for each participant were ordered chronologically. Because these routine screening mammograms were not performed as part of the SWAN study, the mammogram dates did not match the dates of SWAN visits. Approximately half of all mammograms matched to SWAN visits 1-6 were taken within 90 days of the visit. Of mammograms matched to SWAN visits 0 and 7, 28.4% and 36.9% were taken within 90 days of the visit, respectively (data not shown). We developed an algorithm to match retrieved screening mammograms to the nearest SWAN study visit, regardless of whether the mammogram preceded or followed the study visit. We did not adopt an alternative approach, interpolating BMI and weight at the time of the mammograms, because other variables needed for the analysis (such as HT use) were collected at the SWAN visits. To reduce potential error associated with variable time between mammograms and visits, only mammograms taken within 90 days of a SWAN visit were used as matches. Mammographic breast density variables (total breast area and dense breast area) at other study visits were estimated based on linear interpolation using the “ipolate” command in Stata version 10.0 (Stata Corporation, College Station, TX). Separately for each participant, the interpolated estimates were obtained by constructing straight line segments between measurements from consecutive mammograms. The interpolated values for unmatched SWAN visits were obtained assuming that the changes in total and dense breast areas were linear between the two measurement dates that defined the line segment. For visits where density was estimated using interpolation, the average number of days between the visit and the nearest mammogram was 177 (standard deviation [SD] 118.7). Visits without mammogram data both before and after the visit time were not included in the analysis because we did not extrapolate breast density beyond the available mammogram data.

We added a noise term to each interpolated value of total or dense breast area to account for the error introduced by estimating the breast density measurements. These noise terms were randomly generated from a normal distribution with a mean of 0 and SD corresponding to the person-specific SD of the observed measurements for each participant. We multiply imputed ten analytic datasets. This is a novel application of multiple imputation, which is typically used to account for missing data. Percent breast density was calculated by dividing the final (i.e. observed data if mammogram within 90 days, imputed data otherwise) dense area measurement by the final total breast area and expressing this value as a percentage. Due to the addition of the random noise terms, some interpolated values (<1%) were considered to be implausible (i.e. total area<0, dense area<0, or dense area>total area); such values were discarded and additional random noise terms were generated until acceptable imputed values were obtained.

After implementation of the matching and interpolation algorithms, 18 participants had fewer than two mammographic density measurements due to the timing of their mammograms. These participants were excluded from further analyses; the remaining 834 participants had 3,746 eligible mammograms. Women enrolled in the SWAN Mammographic Density Substudy but excluded from this analysis were of similar age, educational level, and menopausal status as those included. Included participants were of slightly lower BMI (25.4 versus 26.4 kg/m², p=0.06), less likely to be African American (7.4% versus 16.8%, p<0.001), and less likely to be from the Pittsburgh clinical site (23.7% versus 43.9%, p<0.001; data not shown).

Statistical analysis

Summary statistics were calculated for demographic, anthropometric, reproductive history, and mammographic breast density variables. The averages of the mammographic density variables across the ten multiply imputed datasets were used for the descriptive statistics. Analysis of variance (ANOVA) and Kruskal-Wallis tests were used to test for baseline differences by site in normally and non-normally distributed continuous variables, respectively, and chi-square tests were used to test for differences by site in categorical variables. Cross-sectional associations between BMI category or weight quartile and the breast density were assessed using ANOVA.

The length of time between the participants' study visits and their nearest mammograms was calculated. Annual change in the anthropometry and breast density variables was calculated by ascertaining the change in each variable between each participant's consecutive visits and standardizing the time period to one year as follows: (ΔBMI, kg/m²/Δtime, days)*(365.24 days/year). Averages of the annual change variables were calculated for the study population.

Random intercept linear models were fit with the participants' age in days as the time scale using the Stata “xtreg” command. Four separate regressions were performed for the two primary mammographic breast density outcomes (annual changes in dense breast area and percent density) for each of the two primary independent variables of interest (annual changes in BMI and weight). Possible covariates included: age (continuous), combined race and site (Caucasian/Pittsburgh, African American/Pittsburgh, Caucasian/Oakland, Chinese/Oakland, Caucasian/Los Angeles, Japanese/Los Angeles), family income (<$35,000, $35,000-$49,999, $50,000-$74,999, $75,000-$99,999, ≥$100,000), education (≤high school, >high school, college, post-college), age at menarche (<12, 12, 13, ≥14), age at first birth (<20, 20-24, 25-29, 30-34, ≥35), breastfeeding history (nulliparous, parous/never, 1-4 months, 5-11 months, 12-22 months, ≥22 months), number of births (0, 1, 2, ≥3), menopausal status (pre-/early perimenopausal, late perimenopausal/postmenopausal/hysterectomy with bilateral oophorectomy, unknown due to HT use), ever use of OCs prior to baseline (no, yes), ever use of HT prior to baseline (no, yes), current HT/OC use (no, yes), ever HT/OC use at each visit (no, yes), number of 1^st degree relatives with breast cancer (0, ≥1), number of 2^nd degree relatives with breast cancer (0, 1, ≥2), history of breast biopsy (no, yes), number of breast biopsies (0, 1, ≥2), and Gail score (continuous). Continuous variables were centered at the population mean. Time-varying variables were defined as appropriate (e.g. menopausal status, hormone use).

Model building using the first imputed dataset followed a backward selection of covariates, retaining covariates that were significant at the 0.10 level. The “mijoin” and “micombine” commands in Stata were used to estimate the overall regression models and to estimate the multiply imputed regression coefficients and their variances following the method of Rubin.^{28, 29} We report the results of two different models: Model 1, annual change in BMI (or weight) and age; and Model 2, annual change in BMI (or weight) and additional covariates.

Analyses were repeated separately for subgroups based on race/site, baseline BMI category, and menopausal status. Interaction terms between menopausal status and the anthropometry change variables were included in separate fully adjusted models to test for effect modification by menopausal status. Similarly, interaction terms between baseline BMI category and the anthropometry change variables were used to test for effect modification by initial BMI among Caucasians. The significance of the interaction terms was assessed using Wald tests. Sensitivity analyses also were performed on subgroups of women with only excellent or good mammograms, never HT users throughout follow-up, and women with ≥80% of mammograms within 90 days of the nearest study visit. All tests performed were two-sided with a p≤0.05 considered statistically significant. All analyses were conducted using Stata version 10.0.

Results

Characteristics of study population

The 834 participants comprising the study population are described in Table 1. The average age of the participants at SWAN enrollment was 46.5 years (SD 2.7). By design, this population was racially diverse, with 62 (7.4%) African American, 407 (48.8%) Caucasian, 182 (21.8%) Chinese, and 183 (21.9%) Japanese participants. The majority of participants (57.5%) was categorized as normal weight at SWAN enrollment. On average, participants had a low risk of being diagnosed with breast cancer within the next 5 years, with a mean Gail score of 1.06 (SD 0.5). By design all women were either premenopausal (58.3%) or early perimenopausal (41.7%) at enrollment. The vast majority reported a previous use of OCs (75.4%) but no previous use of other exogenous hormones (86.8%) at enrollment.

Table 1. Characteristics of the study population at SWAN enrollment, N=834.

Baseline characteristic	Total N=834	Pittsburgh N=198	Los Angeles N=321	Oakland N=315	P value*
General characteristics
Age, years; mean (SD)	46.5 (2.7)	46.1 (2.5)	46.7 (2.7)	46.5 (2.7)	0.09
Race/ethnicity; N (%)					<0.001
African American	62 (7.4)	62 (31.3)	0 (0.0)	0 (0.0)
Caucasian	407 (48.8)	136 (68.7)	138 (43.0)	133 (42.2)
Chinese	182 (21.8)	0 (0.0)	0 (0.0)	182 (57.8)
Japanese	183 (21.9)	0 (0.0)	183 (57.0)	0 (0.0)
Family income; N (%)					<0.001
<$35,000	126 (15.4)	50 (25.5)	25 (8.0)	51 (16.4)
$35,000-$49,999	140 (17.1)	39 (19.9)	38 (12.2)	63 (20.2)
$50,000-$74,999	212 (25.9)	53 (27.0)	74 (23.8)	85 (27.2)
$75,000-$99,999	141 (17.2)	32 (16.3)	62 (19.9)	47 (15.1)
≥ $100,000	200 (24.4)	22 (11.2)	112 (36.0)	66 (21.2)
Education; N (%)					0.001
≤ High school	136 (16.3)	38 (19.2)	38 (11.8)	60 (19.1)
>High school	250 (30.0)	65 (32.8)	112 (34.9)	73 (23.2)
College	223 (26.7)	38 (19.2)	95 (29.6)	90 (28.6)
Post-college	225 (27.0)	57 (28.8)	76 (23.7)	92 (29.2)
History of any cancer; N (%)	5 (0.6)	0 (0.0)	1 (0.3)	4 (1.3)	0.26†
Anthropometric characteristics
Body mass index, kg/m2; mean (SD)**	25.5 (5.9)	28.5 (6.0)	23.9 (4.9)	25.2 (6.2)	<0.001
Underweight: <18.5 kg/m2; N (%)	16 (1.9)	0 (0.0)	9 (2.8)	7 (2.2)	<0.001
Normal: 18.5 - <25.0 kg/m2; N (%)	478 (57.5)	63 (32.9)	219 (68.4)	196 (62.4)
Overweight: 25.0 - <30.0 kg/m2; N (%)	203 (24.4)	71 (35.0)	64 (20.0)	68 (21.7)
Obese: ≥30.0 kg/m2; N (%)	135 (16.2)	64 (32.3)	28 (8.8)	43 (13.7)
Weight, kg; mean (SD)	66.3 (17.1)	75.6 (17.1)	61.5 (14.0)	65.4 (17.9)	<0.001
Reproductive history
Age at menarche, years; N (%)					0.004
<12	171 (20.6)	53 (26.9)	74 (23.2)	44 (14.0)
12	234 (28.2)	55 (27.9)	91 (28.5)	88 (28.0)
13	251 (30.2)	59 (30.0)	92 (28.8)	100 (31.9)
≥14	174 (21.0)	30 (15.2)	62 (19.4)	82 (26.1)
Age at first birth, years; N (%)					<0.001
Not applicable	148 (17.8)	28 (14.1)	58 (18.1)	62 (19.8)
<20	55 (6.6)	29 (14.7)	13 (4.1)	13 (4.1)
20-24	160 (19.2)	52 (26.3)	55 (17.2)	53 (16.9)
25-29	235 (28.3)	46 (23.2)	95 (29.7)	94 (29.9)
30-34	145 (17.4)	28 (14.1)	61 (19.1)	56 (17.8)
≥ 35	89 (10.7)	15 (7.6)	38 (11.9)	36 (11.5)
Cumulative breastfeeding, months; N (%)					<0.001
Nulliparous, never	148 (17.8)	28 (14.1)	58 (18.1)	62 (19.8)
Parous, never	138 (16.6)	53 (26.8)	30 (9.4)	55 (17.5)
1-4 months	145 (17.4)	32 (16.2)	59 (18.4)	54 (17.2)
5-11 months	135 (16.2)	37 (18.7)	57 (17.8)	41 (13.1)
12-22 months	141 (17.0)	24 (12.1)	63 (19.7)	54 (17.2)
≥ 23 months	125 (15.0)	24 (12.1)	53 (16.6)	48 (15.3)
Number of births; N (%)					0.22
0	148 (17.8)	28 (14.1)	58 (18.1)	62 (19.7)
1	138 (16.6)	32 (16.2)	57 (17.8)	49 (15.6)
2	351 (42.1)	78 (39.4)	138 (43.1)	135 (42.9)
≥ 3	196 (23.5)	60 (30.3)	67 (20.9)	69 (21.9)
Menopausal status; N (%)					0.91
Premenopausal	483 (58.3)	112 (57.1)	189 (59.1)	182 (58.2)
Early Perimenopausal	346 (41.7)	84 (42.9)	131 (40.9)	131 (41.9)
Ever used birth control pills; N (%)	627 (75.4)	152 (77.2)	85 (26.5)	239 (76.1)	0.60
Ever used hormones other than birth control pills; N (%)	110 (13.2)	23 (13.3)	42 (13.2)	42 (14.4)	0.99
Other characteristics
Number of 1st degree relatives with breast cancer; N (%)					0.99
0	755 (91.2)	180 (90.9)	291 (91.2)	284 (91.3)
≥ 1	73 (8.8)	18 (9.1)	28 (8.8)	27 (8.7)
Number of 2nd degree relatives with breast cancer; N (%)					0.12
0	657 (79.4)	145 (73.2)	255 (79.9)	257 (82.6)
1	132 (15.9)	39 (19.7)	50 (15.7)	43 (13.8)
≥ 2	39 (4.7)	14 (7.1)	14 (4.4)	11 (3.5)
Number of breast biopsies; N (%)					0.02
0	730 (87.5)	171 (86.4)	279 (86.9)	280 (88.9)
1	76 (9.1)	20 (10.1)	24 (7.5)	32 (10.2)
≥ 2	28 (3.4)	7 (3.5)	18 (5.6)	3 (1.0)
Gail score; mean (SD)	1.06 (0.45)	0.95 (0.5)	1.13 (0.5)	1.05 (0.4)	<0.001
<1.66; N (%)	756 (91.8)	182 (92.9)	284 (89.6)	290 (93.3)	0.20‡
≥1.66; N (%)	68 (8.3)	14 (7.1)	33 (10.4)	21 (6.8)

^*P values from two-sample t tests or Kruskal-Wallis tests for continuous variables and chi-square tests for categorical variables

^†P value from Fisher's Exact Test

^**Two participants were missing weight measurements at enrollment, therefore BMI at enrollment could not be determined

^‡P value for test of Gail Score as a dichotomous variable

Use of OCs or HT since the previous visit increased during follow-up, reaching a maximum of 27.4% at visit 6. At visit 7, 26.2% of participants were premenopausal/early perimenopausal, 68.0% were late perimenopausal/postmenopausal, and 5.9% were of unknown menopausal status due to HT use (data not shown).

Breast density characteristics and cross-sectional baseline associations with anthropometry

From 2 to 10 mammograms were collected on the 834 participants in the present analysis, with a median of 4. The median time between mammograms was 469 days (interquartile range 385 – 728). The mean dense breast area from participants' first available mammogram was 46.2 cm² (SD 26.7), and the mean percent breast density was 42.3% (SD 19.6; Table 2). Dense breast area and percent density were positively correlated (r = 0.48, p<0.001). When participants were cross-classified by quartiles of dense breast area and percent density from their first mammogram, 39.8% were ranked in the same quartile of both dense breast area and percent density (data not shown).

Table 2. Cross-sectional associations between participant characteristics and mammographic breast density measurements, N=834^*.

		Dense Breast Area (cm2)			Percent Density (%)
	N	Mean (SD)	25th – 75th Percentile	P value	Mean (SD)	25th – 75th Percentile	P value
Total population	834	46.2 (26.7)	28.9 – 59.2		42.3 (19.6)	29.3 – 57.4
Race/ethnicity				<0.001			<0.001
African American	62	57.0 (35.0)	34.2 – 68.5		33.0 (20.0)	15.4 – 48.5
Caucasian	407	51.3 (29.8)	31.3 – 66.9		41.0 (20.5)	25.5 – 54.7
Chinese	182	39.0 (17.7)	26.2 – 48.1		51.0 (17.7)	38.1 – 63.2
Japanese	183	38.6 (18.7)	25.2 – 47.7		44.3 (16.7)	32.8 (55.5)
Body mass index category				0.003			<0.001
Underweight: <18.5 kg/m2	15	32.5 (12.9)	21.3 – 42.2		61.9 (19.0)	45.4 – 77.6
Normal: 18.5 – <25.0 kg/m2	456	43.9 (23.3)	28.9 – 54.3		50.8 (16.8)	39.2 – 63.3
Overweight: 25.0 – <30.0 kg/m2	207	49.9 (25.3)	32.6 – 62.9		39.1 (16.1)	27.2 – 50.1
Obese : ≥30.0 kg/m2	143	49.6 (36.7)	22.6 – 73.1		23.6 (16.6)	9.7 – 35.0
Weight, kg				<0.001			<0.001
1st Quartile: 39 – <55.0	207	39.0 (19.9)	25.4 – 48.7		54.6 (16.5)	42.8 – 68.0
2nd Quartile: 55.0 – <63.2	208	47.5 (24.6)	32.8 – 58.9		49.0 (16.3)	38.3 – 61.3
3rd Quartile: 63.2 – <73.8	201	48.3 (24.4)	31.2 – 60.9		41.9 (17.1)	28.8 – 53.9
4th Quartile: 73.8 – 153.9	205	50.1 (34.4)	25.1 – 67.4		27.5 (17.4)	12.8 – 41.2

^*Mammographic and personal characteristics are from the first timepoint at which the participant has mammographic density values, averaged across all imputations; in some cases the first timepoint did not correspond to the enrollment visit, thus the distributions of BMI and weight presented here differ from those in Table 1

^†P values from ANOVA across groups

In cross-sectional analyses of the participants' first visit with mammogram data and their BMI and weight at that time (Table 2), mean dense breast area generally increased with BMI category (p=0.003) and weight quartile (p<0.001). Percent density was inversely associated with both BMI category (p<0.001) and weight quartile (p<0.001).

Longitudinal associations between anthropometry and breast density

The average follow-up time between the first and last SWAN visits included in this analysis was 4.8 years (SD 1.8). Overall, participants tended to gain weight over follow-up (Table 3), with a mean annual BMI increase of 0.13 kg/m² (between-subject SD 0.54) and a mean annual weight increase of 0.32 kg (between-subject SD 1.46). The average annual change over follow-up was -0.77 cm² (between-subject SD 4.49) for dense area and -1.14% (between-subject SD 3.60) for percent density (Table 3).

Table 3. Summary statistics for annual change in anthropometric and breast density measures over follow-up, by race/ethnicity and baseline BMI category.

	N	Mean (SD*)	P Value†
Annual change in BMI (kg/m2)**
Total population	829	0.13 (0.54)
Race/ethnicity			0.81
African American	62	0.10 (0.86)
Caucasian	402	0.14 (0.61)	0.24‡
Pittsburgh	132	0.13 (0.53)
Oakland	133	0.08 (0.58)
Los Angeles	137	0.22 (0.69)
Chinese	182	0.10 (0.38)
Japanese	183	0.13 (0.38)
BMI category at SWAN enrollment§			0.34
Underweight: <18.5 kg/m2	16	0.20 (0.27)
Normal: 18.5 - <25.0 kg/m2	476	0.15 (0.37)
Overweight: 25.0 - <30.0 kg/m2	201	0.12 (0.51)
Obese: ≥30.0 kg/m2	134	0.05 (0.97)
Annual change in weight (kg) **
Total population	829	0.32 (1.46)
Race/ethnicity			0.79
African American	62	0.25 (2.26)
Caucasian	402	0.37 (1.67)	0.29‡
Pittsburgh	132	0.34 (1.46)
Oakland	133	0.22 (1.47)
Los Angeles	137	0.55 (2.02)
Chinese	182	0.24 (0.96)
Japanese	183	0.32 (0.93)
BMI category at SWAN enrollment§			0.28
Underweight: <18.5 kg/m2	16	0.55 (0.77)
Normal: 18.5 - <25.0 kg/m2	476	0.38 (0.96)
Overweight: 25.0 - <30.0 kg/m2	201	0.30 (1.37)
Obese: ≥30.0 kg/m2	134	0.10 (2.64)
Annual change in dense area (cm2)
Total population	834	-0.77 (4.49)
Race/ethnicity			0.57
African American	62	-1.33 (6.94)
Caucasian	407	-0.93 (5.33)	0.80‡
Pittsburgh	136	-1.04 (7.50)
Oakland	133	-1.08 (3.92)
Los Angeles	138	-0.67 (3.68)
Chinese	182	-0.56 (2.17)
Japanese	183	-0.48 (2.74)
BMI category at SWAN enrollment§			0.89
Underweight: <18.5 kg/m2	16	-0.25 (1.23)
Normal: 18.5 - <25.0 kg/m2	478	-0.80 (3.14)
Overweight: 25.0 - <30.0 kg/m2	203	-0.63 (5.69)
Obese: ≥30.0 kg/m2	135	-1.00 (6.42)
Annual change in percent density (%)
Total population	834	-1.14 (3.60)
Race/ethnicity			0.96
African American	62	-0.98 (3.92)
Caucasian	407	-1.23 (3.51)	0.91‡
Pittsburgh	136	-1.19 (4.56)
Oakland	133	-1.35 (3.07)
Los Angeles	138	-1.13 (2.62)
Chinese	182	-1.06 (3.50)
Japanese	183	-1.08 (3.78)
BMI category at SWAN enrollment§			0.53
Underweight: <18.5 kg/m2	16	-1.65 (3.03)
Normal: 18.5 - <25.0 kg/m2	478	-1.32 (3.70)
Overweight: 25.0 - <30.0 kg/m2	203	-0.98 (3.61)
Obese: ≥30.0 kg/m2	135	-0.70 (3.26)

^*Standard deviation reported is the between-subject standard deviation

^†P values from random effects model with race/ethnicity or BMI category as the independent variable

^**Number of observations for annual change in BMI and weight are <834 because some participants were missing height and/or weight data at study visits with mammogram data

^‡P values indicate tests among Caucasians across study sites

^§Some participants were missing weight measurements at enrollment, therefore BMI category at enrollment could not be determined

No statistically significant longitudinal associations between annual changes in BMI or weight and annual changes in the dense breast area were apparent (Table 4). Age-adjusted regressions (Model 1) resulted in non-significant negative associations between changes in BMI (β=-0.17 cm²/(kg/m²), p=0.46) or weight (β=-0.06 cm²/kg, p=0.50) and change in dense breast area. In models adjusted for age, race/site, menopausal status, family history of breast cancer, number of previous breast biopsies, and hormone use since previous visit (Model 2), similar, non-significant, negative associations were observed for BMI (β=-0.18 %/(kg/m²), p=0.44) and weight (β=-0.06 %/kg, p=0.48). We observed no statistically significant interaction between BMI (p=0.64) or weight (p=0.68) and menopausal status in regressions involving dense breast area.

Table 4. Regressions of annual change in (i) dense breast area and (ii) percent density on annual change in anthropometry variables, based on random intercept models and multiply imputed data^*.

	N	β†	Standard Error	95% CI	P Value
Annual change in dense breast area
BMI, kg/m2
Model 1: BMI + age	825	-0.17	0.22	-0.63, 0.29	0.46
Model 2: Fully adjusted†	820	-0.18	0.22	-0.64, 0.29	0.44
Weight, kg
Model 1: Weight + age	825	-0.06	0.09	-0.24, 0.12	0.50
Model 2: Fully adjusted†	820	-0.06	0.09	-0.25, 0.12	0.48
Annual change in percent breast density
BMI, kg/m2
Model 1: BMI + age	825	-0.35	0.19	-0.72, 0.03	0.07
Model 2: Fully adjusted**	818	-0.36	0.19	-0.74, 0.02	0.07
Weight, kg
Model 1: Weight + age	825	-0.13	0.07	-0.28, 0.02	0.09
Model 2: Fully adjusted**	818	-0.13	0.09	-0.29, 0.02	0.09

^*Regression coefficients have the following units: cm²/(kg/m²) for regression of dense breast area on body mass index; cm²/kg for regression of dense breast area on weight; %/(kg/m²) for regression of percent density on body mass index; %/kg for regression of percent density on weight

^†Model 2 for dense breast area is adjusted for age, race/site, menopausal status, 1^st degree relative with history of breast cancer, number of previous breast biopsies, hormone use since previous visit

^**Model 2 for percent density is adjusted for age, race/site, education, menopausal status, number of previous breast biopsies, age at menarche, age at first birth, number of births, history of oral contraceptive use at baseline, hormone use since previous visit

Annual changes in BMI and weight were negatively associated with annual changes in percent breast density (Table 4), though these associations were of borderline statistical significance. A one unit increase in BMI was associated with a 0.35% decrease in percent density over one year in age-adjusted analyses (p=0.07) (Model 1), and a similar association was observed when the model was adjusted for age, race/site, education level, menopausal status, number of previous breast biopsies, age at menarche, age at first birth, number of births, history of OC use at baseline, and hormone use since previous visit (Model 2, β=-0.36, p=0.07). Similar relationships were observed for annual weight change, with a decrease of 0.13% in percent breast density per kilogram increase in weight over one year in a fully adjusted model (p=0.09) (Model 2). We observed no statistically significant interactions between BMI (p=0.96) and weight (p=0.98) and menopausal status in regressions involving annual change in percent density.

Longitudinal associations across initial BMI categories

To assess whether the longitudinal associations between BMI and weight and breast density variables varied by initial BMI, regression coefficients for Caucasian participants were compared across categories of BMI at SWAN enrollment (data not shown). The Caucasian subgroup was the only one that included enough participants in the normal, overweight, and obese categories to provide reliable estimates. For dense breast area, the regression coefficients for BMI remained small and non-significant (normal β=-0.002 cm²/(kg/m²); overweight β=0.12 cm²/(kg/m²); obese β=-0.12 cm²/(kg/m²)). The interaction between baseline BMI category and BMI on the outcome of dense area among Caucasians was not statistically significant (p=0.95). For percent density the regression coefficients remained negative yet became smaller as baseline BMI category increased: normal β=-0.52 %/(kg/m²); overweight β=-0.29 %/(kg/m²); obese β=-0.04 %/(kg/m²); the interaction between baseline BMI category and BMI on the outcome of percent density among Caucasians was not statistically significant (p=0.64). Similar relationships were observed for regressions with weight as the primary independent variable.

Longitudinal associations across other subgroups

The stability and consistency of these relationships were investigated through a series of sensitivity analyses grouping on potentially confounding factors. Similar associations, as judged by the magnitude and direction of regression coefficients, to those observed in the complete cohort were observed in analyses restricted to racial subgroups, to women who did not use any exogenous hormones throughout follow-up (N=439), and to observations with mammograms of good or excellent quality (N=802; data not shown). Somewhat stronger associations of borderline statistical significance were observed for annual change in dense area among women who remained premenopausal/early perimenopausal throughout follow-up (N=178; β_BMI=-0.55 cm²/(kg/m²), p=0.10), and among women with ≥80% of mammograms taken within 90 days of a SWAN visit (N=53; β_BMI=-1.24 cm²/(kg/m²), p=0.08). Among these subgroups, regression results for annual change in percent density were similar to those observed in the complete cohort.

Discussion

This analysis of longitudinal data from a prospective, multi-ethnic cohort of 834 women revealed no statistically significant association between annual changes in anthropometry and dense breast area, yet borderline statistically significant negative associations between annual changes in anthropometry and percent breast density. This study is among the first to prospectively evaluate the relationships between anthropometry and both relative and absolute measures of breast density. Over one year, a one unit increase in BMI was associated with a decrease of 0.36% in percent breast density and a one kilogram increase in weight was associated with a decrease of 0.13% in percent breast density. Though these results did not achieve statistical significance, they are suggestive of negative associations between changes in anthropometry and percent breast density. We also found no significant effect modification by menopausal status, despite the fact that the association between BMI and breast cancer varies markedly by menopausal status.

A previous study reported that over a 5-year period percent density decreased 7% among women who remained premenopausal and 8% among women who became postmenopausal; this translates into an annual decrease of 1.4% for the premenopausal women and 1.6% for the postmenopausal women.³⁰ Thus our observation that a one unit increase in BMI is related to a decrease of 0.35% over a one year period is likely to be clinically meaningful, as it represents nearly a quarter of the normally observed annual change in percent density. The lack of a relationship with dense breast area was counter to our hypothesis. We did observe highly statistically significant positive cross-sectional associations between BMI and weight and the dense breast area, but these associations did not persist in longitudinal analyses.

Our results are largely in agreement with the many previous cross-sectional studies reporting significant inverse relationships between BMI or weight and percent density,^{9-14, 17, 18, 20, 31} including a cross-sectional analysis conducted in this same SWAN cohort.¹⁹ We did observe a significant positive cross-sectional relationship between BMI and weight and dense area, in agreement with some,¹³ but not all,^{17, 18, 20} previous studies. In one study, however, the association became positive after adjustment for the non-dense breast area.¹⁷

Few studies have used longitudinal data to analyze associations between anthropometry and mammographic breast density. McCormack et al. reported that women with larger increases in BMI between ages 43 and 53 had an increased risk of having high-risk Wolfe patterns.¹⁵ Their study, however, is not directly comparable to ours due to their use of a qualitative breast density measurement and their use of longitudinal change in BMI yet only a single mammographic density assessment at follow-up. Boyd et al.³ related weight change over two years to change in breast density measurements over the same time period. Similar to our analysis, Boyd et al. reported a significant negative association between weight change and percent density, such that percent density was increased in those who lost weight; however, their study also reported a significant positive association between weight change and the size of the dense breast area, such that dense breast area was decreased in those who lost weight.³ It is unclear why our findings differed from these latter results. The characteristics of the two populations differ markedly in their age, ethnicities, baseline breast cancer risk, and observed weight change. These differences may explain at least some of the discordant results of these two studies. Indeed, a nested case-control study of Native Hawaiian, Japanese, and Caucasian women recruited from the general population reported results similar to those we observed. Maskarinec et al.²¹ reported that overweight and obese women had a more gradual decline in percent density over time as compared to women of normal weight. Likewise, we observed that the regression coefficients for both BMI and weight with the outcome of percent density were more strongly negative among women of normal weight at study enrollment than among those who were overweight or obese at that time, although formal tests of the interaction terms were not statistically significant. Maskarinec et al. did not report on the outcome of dense breast area which precludes a direct comparison with our results.

The observation that anthropometry is related to the dense breast area in cross-sectional studies but not in longitudinal studies appears to be paradoxical at first consideration. Overweight and obese women may have a larger dense breast area than underweight or normal weight women simply because the total breast size is generally larger in women of greater weight; this explains the highly significant cross-sectional associations. Indeed, a previous cross-sectional analysis of mammographic data from SWAN demonstrated significant positive correlations between BMI and the total breast area and also between the total breast area and the dense breast area.¹⁹ After these cross-sectional differences are accounted for, further increases in weight and BMI appear to result in the accumulation of fat in the breast rather than altering the dense breast tissue. Thus the total breast area increases while the dense breast area remains relatively constant. As total breast area is the denominator when calculating percent breast density, increased total breast area results in a decrease in percent density.

Overall, our results provide evidence that the consistently demonstrated relationship between anthropometry and breast cancer risk may not proceed through a direct association between anthropometry and the dense breast area. The borderline statistically significant negative association between increases in BMI and weight and percent breast density most likely reflects the effect of anthropometry on the non-dense area. Indeed, this effect on the non-dense area may explain the relationship between anthropometry and breast cancer risk. In adipose tissue, such as that comprising the non-dense area of the breast, androstenedione is converted to estrogen.^{32, 33} Higher non-dense breast area therefore may result in increased estrogen exposure to the nearby dense breast tissue due to this peripheral production of estrogen.^{13, 21} This increased estrogen exposure of the ducts and lobules where cancers arise may result in increased risk of breast cancer. Therefore, observing a longitudinal decrease in percent density may actually reflect an increase in breast cancer risk if the decreased percent density results from an increase in non-dense tissue rather than a decrease in the dense breast area.

Haars et al.¹³ noted that percent breast density may not be valid for etiologic inference because this measure incorporates information about both the dense breast area, believed to represent cells at risk for developing cancer, and BMI, an independent risk factor for breast cancer. Further, Haars et al. reported that only 37% of their participants were ranked in the same quartile of both percent density and dense breast area; in other words, a group of women with equivalent percent density may actually have a wide range in the size of their dense breast areas.¹³ Likewise, 39.8% of our participants had concordant classifications for quartiles of dense breast area and percent density. Thus when one studies percent breast density as an intermediate endpoint, the results are also reflective of associations with BMI and do not necessarily reflect unique effects of the exposure being evaluated on the dense breast area.¹³ Our results, which show no statistically significant relationship between annual changes in anthropometry and dense breast area, yet are suggestive of negative associations between annual changes in anthropometry and percent density, support the recommendation by Haars et al. that the dense breast area be used as the outcome in studies using mammographic breast density to make inference to breast cancer etiology.¹³ Studying dense breast area as the outcome reduces the likelihood that BMI confounds the results. We add to their recommendation that the non-dense area should be evaluated for its role in breast cancer etiology.

Strengths of this study include its large sample size and multi-ethnic, population-based cohort. Also, menopausal status and use of HT were carefully monitored in SWAN. Our quantitative measurements of mammographic breast density are preferable to the qualitative and subjective measurements used in many previous studies. The high reliability of the single expert reader of the mammograms also is a substantial strength. Finally, we demonstrated consistent findings when across subgroups defined by race, exogenous hormone use, and menopausal status.

Limitations to this study include potential residual confounding, despite careful adjustment for confounders. Also, the participants in the SWAN Mammographic Density Substudy are not a representative sample of the areas from which they were recruited, and this may limit the external validity of these results. The most significant limitation is that the mammograms were not taken at the same time as the SWAN visits. Therefore we used linear interpolation with multiple imputation of random noise terms to estimate the participant's breast density at the time of her SWAN visit if the mammogram was not taken within 90 days of the nearest SWAN visit. We have described and validated our interpolation and imputation method elsewhere.³⁴ In this validation study, we found that the original values were highly correlated (r=0.96) with values estimated from linear interpolation with multiple imputation.³⁴ Further, we observed similar results to those observed in the entire cohort when we repeated analyses among women with the majority of their mammograms occurring within 90 days of a SWAN visit. It is possible that addition of the noise terms resulted in overly conservative tests of significance. This might explain our borderline significant results for change in percent density. Future studies may benefit from incorporating mammograms into their study visits to avoid the imputation that was required in this study.

This study provides evidence that anthropometry is not longitudinally associated with changes in the dense breast area, yet is associated with percent breast density, at least among women transitioning through menopause. Our findings suggest that as a surrogate for breast cancer, the absolute dense breast area is likely to be the most relevant outcome, though the non-dense area could be important to disease etiology as well. Choosing dense breast area as the outcome is preferable because BMI would not confound the observed results, as it would if percent density were used instead.

Abbreviations Used

ANOVA

analysis of variance

BMI

body mass index

HT

hormone therapy

OC

oral contraceptive

SD

standard deviation

SWAN

Study of Women's Health Across the Nation