Body Mass Index and Mammographic Density in a Multiracial and Multiethnic Population-Based Study

This article is featured in Highlights of This Issue, p. 1247

Breast cancer is the second most common cause of cancer-related death among women in the United States (1). Breast cancer incidence is highest among non-Hispanic White women and non-Hispanic Black women, and lowest among the combined Asian and Native Hawaiian/Pacific Islander (NHPI) populations (1). However, data from the Surveillance Epidemiology and End Results (SEER) cancer registries suggest that when data from Asian and NHPI women are analyzed separately, some of the highest breast cancer incidence and death rates are among NHPI women (2–4).

Mammographic density (MD) is one of the strongest breast cancer risk factors. MD is defined as the extent of fibroglandular versus fatty breast tissue, and can be measured as the percentage of the breast containing fibroglandular tissue, or as a visual estimate by radiologists [Breast Imaging Reporting and Database System (BI-RADS) score] (5). Women with percent MD greater than 75% are estimated to have a risk of breast cancer that is 4 to 6 times greater than women with percent MD less than 5% (6, 7), and more than 30% of breast cancers are attributed to the presence of dense tissue in more than 50% of the breast (8).

Body mass index (BMI), a metric correlated with measures of body fat, is correlated with the amount of adipose tissue in the breast (9) and, therefore, associated with lower MD (7). While BMI is associated with lower MD in both pre- and postmenopausal women, BMI is inversely associated with breast cancer among premenopausal women, but positively associated with breast cancer among postmenopausal women (7). Hormone therapy (HT) use is also a strong breast cancer risk factor among postmenopausal women (10, 11), and prior studies have reported stronger positive associations between MD and breast cancer among women taking HT (12, 13).

The associations between BMI and MD, and BMI, MD, and breast cancer are well-established; however, the extent to which these associations vary by race and ethnicity is still being investigated. Here, we estimated the prevalence of high BMI and high MD within racial and ethnic subgroups in Utah and sought to understand the proportion of MD attributable to low BMI, overall and by race and ethnicity. Then, putting the two together, we estimated the proportion of breast cancer cases that would not have occurred if Utah women had lower MD, and the proportion of breast cancer cases that could have been avoided if postmenopausal Utah women had lower BMI.

The Utah Population Database (UPDB) is a population-based resource that contains information on more than 11 million individuals who are living or have lived in Utah, or are ancestors of current and former residents. Data are captured from multiple sources, including vital records, U.S. census data, the Utah Cancer Registry (UCR), statewide claims databases and ambulatory surgery records, hospital-level electronic medical records, and Utah Driver License Division data. Within the UPDB, we assembled a mammography cohort including all women who received at least one digital mammogram at Intermountain Healthcare (Salt Lake City, UT) or University of Utah Health Care (Salt Lake City, UT) between 2005 and 2019. As of October 2020, this cohort included 235,520 women, ages 18 to 70 years at baseline mammogram (i.e., first screening mammogram between 2005 and 2019). Data were collected under a waiver of informed consent, and study protocols were approved by the Resource for Genetic Epidemiology ethics committee and the University of Utah Institutional Review Board (IRB), following the guidelines in the Belmont Report.

Information on baseline MD was extracted from medical records at Intermountain and University of Utah Health Care, which together, capture more than 75% of health care in Utah. Radiologists recorded BI-RADS breast density as (a) almost entirely fatty; (b) scattered fibroglandular densities; (c) heterogeneously dense; or (d) extremely dense. Breast cancer cases, defined by ICD-O-2/3 codes of C500–509, were identified through linkage to the UCR, a SEER site since 1973.

Information on race and ethnicity was available through the UPDB demographic dataset which standardizes data from multiple self-reported and observational data sources. We cross-classified the five racial categories used by the Federal Office of Management and Budget (OMB)—American Indian or Alaska Native (AIAN), Asian, Black or African American, Native Hawaiian or Other Pacific Islander, and White (14)—with Hispanic ethnicity to get six racial-ethnic categories: non-Hispanic AIAN; non-Hispanic Asian; non-Hispanic Black or African American; non-Hispanic Native Hawaiian or Other Pacific Islander; non-Hispanic White, and Hispanic. We also evaluated NHPI, and Asian subgroups. For Native Hawaiian and Pacific Islander women, these included Native Hawaiian and “Other Pacific Islander” (i.e., Samoan, Guamanian, Micronesian, Tahitian, Tongan, or Pacific Islander not otherwise specified). For Asian women, subgroupings included Chinese, Japanese, Filipino, and “Other Asian” (i.e., from Korea, Vietnam, India, and additional countries in Asia). Recognizing that all of the above are social constructs, we followed reporting guidelines to encourage clear, consistent, and equitable consideration of race and ethnicity (15).

BMI (in kg/m²) was calculated using height and weight as self-reported on issuance or renewal of a Utah driver license, a method that systematically underestimates true BMI but effectively ranks people by BMI and consistently identifies those who are obese (BMI ≥ 30; ref. 16). To generate a single value for use in analyses, we dropped unrealistic BMI values (BMI < 14 or BMI > 100) then averaged each participant's first and most recent BMI. First BMI often reflected height and weight many years prior to the baseline mammogram, while the most recent BMI was calculated using height and weight information reported around the time of, or many years after, the baseline mammogram. When only one measurement was recorded we used the single, available value in place of an average.

Consistent with studies of metabolic health and breast cancer risk, we categorized BMI using NIH and World Health Organization (WHO) cutoffs: underweight (BMI ≤ 18.5 kg/m²); normal (BMI 18.6–24.9 kg/m²); overweight (BMI 25–29.9 kg/m²); and obese (BMI ≥ 30 kg/m²; refs. 17–19). Due to the very small number of women in the underweight category, especially within subgroups, we collapsed underweight and normal weight into a single reference category. In sensitivity analyses, we included categorizations for Asian individuals as 23 (overweight) and 27.5 (obese); and 26 (overweight) and 32 (obese) for NHPI women.

Data on menopausal status were not available, so, consistent with other cohorts, we used age ≥55 years as a proxy for postmenopausal status (20). We also considered age less than 45 years as a proxy for premenopausal status, though prior data suggest that 11% to 22% of women experience menopause prior to age 45 (21). Data on past and current (within 3 years of baseline mammogram) use of tamoxifen, aromatase inhibitors, and HT were collected from inpatient and outpatient orders and the All-Payer Claims Database. Information on parity, including age at first birth and number of children prior to 2018, was gathered from multiple sources including UPDB ancestry data and birth certificates. Utah birth certificates capture both the current birth and the mother's prior births, so, when combined with ancestry data, parity in Utah is captured very well and out-of-state births are partially captured. Urbanicity was calculated using Rural-Urban Commuting Area (RUCA) coding of the home address in closest temporal proximity to the baseline mammogram (codes <7 were considered urban), and educational status was estimated from the UPDB demographic dataset, also using data in close temporal proximity to the baseline mammogram.

Given our interest in MD as observed on screening mammograms, we excluded women whose initial mammograms within the study period were diagnostic (n = 17,619), women with a UCR breast cancer diagnosis before baseline (n = 5,374), women with a history of tamoxifen or aromatase inhibitor use (n = 238), and women with breast implants (n = 17,285). We further excluded women under age of 30 (n = 408), women missing race or ethnicity data (n = 29,355), and women without BMI data (n = 4,437). This left 160,804 women (1,962,299 person-years) eligible for analyses.

To estimate multivariable-adjusted relative risks for the association between BMI and high MD we used modified Poisson generalized estimating equations with robust error estimates (22). To estimate multivariable-adjusted relative risks for the associations between BMI and breast cancer, and MD and breast cancer, we used Cox proportional hazards models mutually adjusted for BMI/MD and allowed for differing baseline hazards by increasing age (as individuals contributed person-time). For each analysis, we reported point estimates and 95% confidence intervals (CI) overall and jointly stratified by race and ethnicity and by age (<45, 45–<55, and ≥55 years). Each model accounted for age at mammogram, parity (0, 1, 2+ births in Utah, parous with number of births not known), HT use (current, past, never), and educational status (less than high school, high school, some college, college, missing). Because the distributions of BMI and MD differed across racial and ethnic groups, and these comparisons were central to our research, we created a categorical BMI variable (<25, ≥25-<30, and ≥30 kg/m²) and binary MD variable (low MD defined as almost entirely fatty or scattered fibroglandular densities and high MD defined as heterogeneously dense or extremely dense) for our analyses. Tests for trend were conducted using BMI as a continuous variable.

To test for effect modification by race and ethnicity and by HT, we ran models with and without an interaction between MD and the modifier of interest and used likelihood ratio tests to evaluate statistical significance. We were also interested in potential effect modification by urban versus rural residency; however, the vast majority of women resided in urban locations so we were only able to consider an analysis restricted to women in urban neighborhoods.

We calculated population attributable risk percents (PAR%) to understand the proportion of women with high MD who would have had low MD if they had a BMI ≥ 25, and we calculated the proportion of breast cancer that would not have occurred if high BMI (BMI ≥ 25) or high MD (BI-RADS C/D) had been removed from our population. PAR%s were calculated using the following formula in which P is the proportion of exposed individuals in the population and RR is the relative risk.

We estimated adjusted relative risks using the modified Poisson generalized estimating equation modeling approach, as described above, when calculating the percent of high MD explained by low BMI (22), and we estimated adjusted relative risks using pooled conditional regression when calculating the percent of incident breast cancer explained by BMI or MD (23). Analyses were completed using SAS 9.4.1, and statistical tests assumed a two-sided α of 0.05. Consistent with UPDB confidentiality policies, we masked all counts and percentages that reflect less than 11 cases. The data analyzed in this study are available from the UPDB. UPDB data usage is governed by the Utah Resource for Genetic and Epidemiologic Research (RGE). Data are available from the authors upon reasonable request and with approval from the RGE.

Our study included 631 American Indian or Alaska Native women, 1,828 Asian women, 821 Black women, 8,791 Hispanic women, 271 NHPI women, and 148,462 White women. There were dramatic differences in the prevalence of obesity and high MD by race and ethnicity (Table 1). For example, we observed the highest prevalence of obesity among NHPI women (i.e., 39.1% of Native Hawaiian, and 52.2% of Other Pacific Islander women were obese) and the lowest prevalence of obesity among Asian women (i.e., ranging from 4.6% of Other Asian women to 6.8% of Chinese women). We observed the lowest reportable prevalence of “extremely dense” MD among American Indian and Alaska Native women (5.1% categorized as BIRADS-D), and the greatest prevalence of “extremely dense” MD among Chinese (27.5%) and “Other Asian” (16.7%) women.

Higher BMI was strongly associated with lower MD (Table 2); the relative risk of high MD among individuals with BMI ≥ 30 kg/m² compared with BMI < 25 kg/m² ranged from 0.52 (Asian; 95% CI, 0.39–0.68) to 0.35 (AIAN; 95% CI, 0.26–0.46). There was statistically significant evidence of heterogeneity (P = 0.009) by race and ethnicity among women ages less than 45 years with estimates ranging from 0.56 (Asian; 95% CI, 0.34–0.94) to 0.34 (NHPI; 95% CI, 0.21–0.54). We did not observe statistically significant evidence of heterogeneity by race and ethnicity among women ages 45 to <55 years (P = 0.14) or ≥55 (P = 0.24), though among women ages ≥55 years, magnitudes of association ranged from RR_{BMI≥30 vs. BMI<25} = 0.55 (95% CI, 0.33–0.91) for Asian women to RR_{BMI≥30 vs. BMI<25} = 0.24 (95% CI, 0.09–0.64) for NHPI women.

We conducted sensitivity analyses to explore the heterogeneity within our study population. First, breast cancer screening is not recommended among average risk women before age 40 (24), so we conducted an analysis restricted to women ages ≥40 years. Results were nearly identical to the main analysis (Supplementary Table S1). Second, to acknowledge the racial diversity within our Hispanic population, we considered alternate cross-classifications of race and ethnicity such that all non-White Hispanic women were moved from the Hispanic category to their racially defined category (Supplementary Table S2). Using this classification, we observed similar associations between BMI and MD (Supplementary Table S3) as were observed in the main analysis (Table 2). Third, as prior studies have suggested that Asian women may experience adverse health effects at lower BMI and Pacific Islander women may experience adverse health effects at higher BMI, we also considered the associations between BMI and MD using racial and ethnic-specific BMI cutoffs (9, 25–28); associations were slightly attenuated, particularly among Asian women (Supplementary Table S4). Fourth, as Utah's urban and rural/frontier populations often have different cultural and lived experiences and different access to health care we ran analyses of BMI and MD that were restricted to women living in more urban areas (RUCA < 7; Supplementary Table S5) and observed similar results to the main analysis. Finally, in analyses stratified by HT use, the associations between BMI and MD among never users of HT were similar to the full cohort (RR_{BMI≥30 vs. BMI<25} = 0.43; 95% CI, 0.42–0.44), while results among current HT users were slightly weaker in magnitude (RR_{BMI≥30 vs. BMI<25} = 0.50; 95% CI, 0.47–0.54; Supplementary Table S6).

The counts of breast cancer cases among AIAN, Asian, Black, and NHPI women in this cohort were too small to evaluate associations between BMI or MD and breast cancer risk, so we evaluated these associations only among White and Hispanic women (Table 3). Higher BMI was statistically significantly associated with greater breast cancer risk among women ages ≥55 years (HR_{BMI≥30 vs. BMI<25} = 1.41; 95% CI, 1.30–1.52), but not among women ages <45 years (HR_{BMI≥30 vs. BMI<25} = 1.09; 95% CI, 0.72–1.64), or 45 to <55 years (HR_{BMI≥30 vs. BMI<25} = 1.16; 95% CI, 0.99–1.36). There was no evidence of heterogeneity in the magnitude of associations between Hispanic women and White women for the associations between BMI and breast cancer for women ages <45 years (P = 0.22), 45 to <55 years (P = 0.14), or ≥55 years (P = 0.65).

High MD was associated with greater breast cancer risk among women ages <45 years (HR = 1.73; 95% CI, 1.24–2.40), 45 to <55 (HR = 2.01; 95% CI, 1.76–2.31), and ≥55 years (HR = 1.71; 95% CI, 1.61–1.82; Table 4). There was no statistically significant evidence of heterogeneity comparing Hispanic and White women ages <45 years (P = 0.98), 45 to 55 (P = 0.98), or ≥55 years (P = 0.20).

PAR%s are presented in Table 5. The proportion of high MD explained by low BMI was 28.4% (95% CI, 27.7–29.1) among women ages <45 years, and 22.9% (95% CI, 22.2–23.5) among women ages ≥55 years (after adjusting for HT). For women ages <45 years, PAR%s were highest among White women (PAR% = 29.2; 95% CI, 28.4–29.9) and lowest among Asian (PAR% = 17.2; 95% CI,= 8.5–25.8) and Black (PAR% = 17.3; 95% CI, 11.6–22.9) women. For women ages ≥55 years, after additionally adjusting for HT, PAR%s were lowest for Hispanic (PAR% = 23.2; 95% CI, 19.6–26.7) and White (PAR% = 22.5; 95% CI, 21.8–23.1) women, and highest for AIAN women (PAR% = 37.5; 95% CI, 28.1–46.9).

There was no evidence to suggest that breast cancer was explained by BMI among women ages <45 years (PAR% < 0), yet 6.2% (95% CI, 4.7–7.7) of breast cancer was explained by high BMI among women ages ≥55 years. The percent of breast cancer explained by high MD was 29.2% (95% CI, −17.2 to 65.0%) among Hispanic women ages <45 years, and 28.0% (95% CI, 13.3–41.4) among White women ages <45 years, while the percent of breast cancer explained by high MD was 3.5% (95% CI, −3.0 to 10.0%) among Hispanic women ages ≥55 years, and 22.1% (95% CI, 19.6–24.5%) among White women ages ≥55 years.

In this large, population-based study we evaluated distributions of BMI and MD by race and ethnicity and estimated associations between BMI and MD overall and within racial and ethnic subgroups. We observed strong evidence of variation in the distributions of both BMI and MD by race and ethnicity, with the highest BMI among NHPI women and the highest MD among Asian women. Consistent with prior studies, we observed an association between high BMI and low MD among women of all ages (5, 7, 29–31). We noted heterogeneity in the magnitude of this association by race and ethnicity, particularly among women ages <45 years, with the weakest association among Asian women and the strongest association among NHPI women.

Racial and ethnic variations in BMI and MD have been reported previously. Consistent with many of our findings in Utah, the burden of obesity in the United States has been described as especially strong among AIAN, Black, Hispanic, and NHPI populations, and lowest among Asian populations (3, 4, 17, 32). MD is strongly correlated with age, menopausal status, and BMI, and before accounting for these factors many studies observe the greatest prevalence of high MD among Asian women (17, 18, 29, 33), just as we observed. However, after accounting for these factors, the prevalence of high MD is similar across racial and ethnic subgroups (18, 29–31, 33).

More research is needed to understand how menopausal status, BMI, estrogens, and factors associated with race and ethnicity (e.g., cultural practices, experience of racism and psychosocial stressors, genetics) interact to influence MD. Postmenopausal BMI has been positively correlated with estrone and estradiol levels (34, 35), and these estrogens have been inversely associated with MD in some, but not all, studies of postmenopausal women (36). In the Multiethnic Cohort Study, Japanese American, African American, and Native Hawaiian women had higher levels of estrogens than non-Hispanic White or Latina women (37), suggesting that, if the inverse association between BMI and MD in postmenopausal women is mediated by estrogen levels, the strength of the association should be greater among Asian, Black, and NHPI women than in White or Hispanic women. This is consistent with our finding of nonstatistically significantly stronger magnitudes of association between BMI and MD among Black, NHPI, and American Indian or Alaska Native women (Table 2).

We also evaluated breast cancer PAR%s which reflect the prevalence of the exposure, the cut-off points used to define exposure levels, and the strength of association between the exposure and outcome within the population of interest (38, 39). Our breast cancer PAR%s for BMI were lower than expected for the percent of breast cancer explained by low BMI among premenopausal women (10, 17), but similar to existing literature for the percent of breast cancer explained by high BMI among postmenopausal women (10, 17, 40, 41). While our study had limited power to consider PAR%s for BMI with breast cancer risk within racial and ethnic subgroups, our finding of similar postmenopausal PAR%s across Hispanic (9.1%; 95% CI, −1.5 to 19.4) and White (6.1%; 95% CI, 4.6–7.7) groups was generally consistent with data from seven U.S.-based Breast Cancer Surveillance Consortium (BCSC) registries, though BCSC postmenopausal PAR%s were higher than in our report (12.0% Hispanic, and 15.4% White; ref. 17).

Breast cancer PAR%s for MD were also similar to prior U.S.-based studies. Here, we estimated that if women with heterogeneously dense or extremely dense breasts had achieved scattered fibroglandular or fatty breast density, breast cancer incidence would have been reduced by 28.3% (95% CI, 14.3–41.1) among women ages <45 years, and by 21.3% (18.9%–23.7%) among women ages ≥55 years. Similar reductions were reported in U.S.-based BCSC registry sites (17, 41). For example, one BCSC study estimated that 29% (95% CI, 25–33) of premenopausal and 14% (95% CI, 13–16) of postmenopausal breast cancers could have been avoided if all women with heterogeneously or extremely dense breasts had scattered fibroglandular breast density (41). Interestingly, the PAR%s varied by ethnicity in both our study and the U.S.-based BCSC studies, but the magnitude of variation was not consistent (17).

While racial and ethnic differences in breast cancer incidence can reflect differences in access to mammography, genetic variation, or other factors, findings from the present study suggest that the prevalence of lifestyle and other risk factors should be considered when evaluating the causes of disparities in breast cancer incidence and strategies for reducing those disparities. For example, most breast cancers are diagnosed among postmenopausal women, so considering the prevalence of high BMI, a modifiable risk factor for postmenopausal breast cancer and many other chronic diseases (e.g., cardiovascular disease, diabetes, chronic kidney disease; refs. 7, 42), may be important when conceptualizing community-based prevention strategies. In this Utah-based study, we observed that BMI is highest in NHPI women and lowest in Asian women. In other U.S.-based studies, the burden of obesity has been described as highest among AIAN, Black, Hispanic, and NHPI populations, and lowest among Asian populations (3, 4, 17, 32). Given these differences in the distribution of BMI and that sustained weight loss has been associated with lower breast cancer risk (43), weight management interventions that take into consideration structural inequities and the diverse cultural, religious, and language preferences of the AIAN, Black, Hispanic, and NHPI communities are important.

MD is also considered a modifiable breast cancer risk factor, and reductions in MD have been associated with lower breast cancer risk (44). As aromatase inhibitors, tamoxifen and other interventions continue to be assessed as possible modifiers of MD and breast cancer risk, our data suggest that it will be important to consider effect modification by factors that are differentially distributed by race and ethnicity.

Key strengths of our study include the population-based design and the inclusion of granular data on race and ethnicity. The large, population-based design allowed us to consider the importance of obesity and MD to breast cancer risk among NIH-defined racial and ethnic groups, and allowed us to estimate PAR%s for obesity on MD. Important limitations of the study included low power to consider breast cancer incidence, the use of binary BI-RADS scores, limited data on reproductive factors (e.g., age at menarche, menopausal status, parity outside of Utah, oral contraceptive use), and the use of covariate data measured before or after the timing of the baseline mammogram. While we were unable to estimate the associations of BMI and MD with breast cancer risk for Asian, AIAN, Black, and NHPI women, we were able to estimate the associations between BMI and MD in all of these, often understudied, racial and ethnic groups. As more breast cancer cases accrue within the Utah Mammography Cohort, further evaluation of breast cancer risk may become possible. This study also could have been improved by consideration of percent density; however continuous MD measures are not yet available in this cohort. Evaluation of MD using dichotomized BI-RADS scores resulted in estimates of association that were strong and in the expected direction, but follow-up studies using more granular measures may detect more nuanced differences in the strengths of the associations between BMI and MD, BMI and breast cancer, or MD and breast cancer by race or ethnicity. Measurement error in BMI and residual confounding by unmeasured reproductive factors were also a concern; however, we were able to incorporate BMI and parity data from multiple time points and sources. Furthermore, the median age at menopause in the United States is approximately 52.5 years (45), so, by stratifying at ages 45 and 55, we ensured that our younger age group was mostly premenopausal and our older age group almost entirely postmenopausal.

In conclusion, we have reported that BMI ≥ 25 and high MD account for a large proportion of breast cancers in Utah, and we have generated preliminary data to suggest that the extent to which each of these factors influences breast cancer risk may vary by race and ethnicity. Future efforts to reduce breast cancer risk will need to consider racial and ethnic differences in the contributions of BMI and MD to breast cancer to ensure that any novel prevention strategies reduce breast cancer incidence both successfully and equitably.

M.E. Barnard reports grants from NIH/NCI during the conduct of the study; and personal fees from Epi Excellence LLC outside the submitted work. No disclosures were reported by the other authors.

M.E. Barnard: Conceptualization, formal analysis, supervision, funding acquisition, methodology, writing–original draft, writing–review and editing. T. Martheswaran: Conceptualization, formal analysis, writing–original draft. M. Van Meter: Conceptualization, resources, data curation, writing–review and editing. S.S. Buys: Conceptualization, resources, data curation, writing–review and editing. K. Curtin: Conceptualization, resources, data curation, supervision, funding acquisition, methodology, project administration, writing–review and editing. J.A. Doherty: Conceptualization, supervision, funding acquisition, writing–review and editing.

We thank the Pedigree and Population Resource (PPR) of Huntsman Cancer Institute, University of Utah (funded in part by the Huntsman Cancer Foundation) for its role in the ongoing collection, maintenance, and support of the Utah Population Database (UPDB). We also acknowledge support for the UPDB through grant P30CA042014 from the NCI, through the University of Utah, and from the University of Utah's program in Personalized Health and Center for Clinical and Translational Science. We thank the University of Utah Center for Clinical and Translational Science, the PPR, University of Utah Information Technology Services, and Biomedical Informatics Core for establishing the Master Subject Index between the Utah Population Database, the University of Utah Health Sciences Center, and Intermountain Health Care. The Utah Cancer Registry is funded by the NCI's SEER Program contract no. 75N91018D000016, the U.S. Center for Disease Control and Prevention's National Program of Cancer Registries, cooperative agreement no. NU58DP006320, with additional support from the University of Utah and Huntsman Cancer Foundation. Research was supported by the NCRR grant, “Sharing Statewide Health Data for Genetic Research” [grant no. R01RR021746, G. Mineau, principal investigator (PI)] with additional support from the Utah Department of Health and the University of Utah. Research was also supported by the Susan Cooper Jones Fellowship award (to M.E. Barnard. PI), and the NCI of the NIH (K00CA212222 to M.E. Barnard, PI). We wish to thank Ms. Emily Guinto (PPR), Mr. Mike Newman (University of Utah IT), and Mr. Jesse Gygi (Intermountain IT) for their data extraction support. The authors take full responsibility for the analysis and interpretation of data. The content of this manuscript is solely the responsibility of the authors and does not represent the official views of the NIH or other research sponsors.

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.