Abstract
Objectives
The high incidence of female breast cancer that has been consistently reported in urban areas could be mediated by breast density, which is considered to reflect the cumulative exposure of breast tissues to hormones. The aim of this study was to assess how mammographic density varies by the degree of urbanization.
Setting
The population consisted of 55,597 cancer-free women, aged 50–59 years, who participated in a French breast cancer screening programme (Franche-Comté region) between 2005 and 2009.
Methods
Ordered logistic regression was run with mammographic density as the outcome, and degree of urbanization as the independent variable, while adjusting for some known confounding factors. Multiple imputation was used to deal with missing data.
Results
A significant positive linear trend with urbanization was found in a univariate approach (P trend <10−3), and after adjusting for risk factors (P trend = 10−3). A negative and highly significant association with mammographic density was highlighted both for age at the time of mammography (odds ratio (OR) 0.41, 95% confidence interval (CI) 0.39–0.43, per 10 years), and for low socioeconomic status (OR 0.71, 95% CI 0.67–0.75). The OR for hormone replacement therapy use was 1.51 (95% CI 1.43–1.58).
Conclusions
Knowledge of this urbanization gradient in density (whatever its mechanism) may help to identify women who may require full-field digital mammography for the early detection of breast cancer, and could assist primary care providers in recommending the best screening strategy in a risk factor-based approach.
INTRODUCTION
Female breast cancer incidence has consistently been reported to be higher in urban areas than in rural areas in the United States, Canada, and European countries. 1,2 This association could be mediated by breast density, as suggested by a recent British study including women who underwent mammography in London. 3 Women aged 45–54 years who lived in central London were twice as likely to have very dense breasts as women who lived in outlying suburban and rural areas. Thus, women living in urban areas potentially have a greater risk of breast cancer, as mammographic density is one of the strongest predictors of breast cancer, independent of age, menopausal status, or ethnicity. 4,5 The biological mechanisms through which it may influence breast cancer risk are, however, not fully understood. 6
Mammographic density is considered to reflect the cumulative exposure of breast stroma and epithelium to hormones and growth factors that stimulate cell division. 7 The biological basis for the association of breast density with urban residence could rely, at least partially, on environmental factors, such as oestrogenic particles present in traffic emissions. 8 The latter include certain polycyclic aromatic hydrocarbons (PAHs) that mimic the effects of oestrogen. 9 This pathway is supported by the growing evidence of an association between breast cancer and occupational exposure to PAHs, 10,11 or traffic-related air pollution. 12–14
Identifying lifestyle factors associated with increasing mammographic density (such as residence in urban settings) may have important implications for breast cancer prevention and may result in new cancer screening guidelines. The aim of this study was therefore to assess how mammographic density varies by the degree of urbanization, using population-based screening data and adjusting for some established risk factors for breast cancer at the individual level.
MATERIALS AND METHODS
Study population
The study population consisted of women participating in a mass breast cancer screening programme introduced in the Franche-Comté region (France) in September 2003.
The French screening programme is based on existing radiologic facilities in both the private and public sectors to give equitable access to mammography and close the gap between health status disparities. Local screening authorities invite the target population, women aged 50 to 74 years, by surface mail to undergo a free mammography once every two years. The invitation underlines the objectives of screening by mammography and distinguishes screening from diagnostic testing. Women may choose from a list of participating radiologists provided with the invitation. The screening procedure consists of a clinical breast examination and a two-view mammography (mediolateral oblique and craniocaudal) for each breast.
Participating radiologists sign a binding agreement with the programme co-ordinator in which they commit themselves to submitting films to a colleague or expert committee for review, following specific training, and periodically checking their equipment for quality control. Although the European guidelines recommend the double reading of mammograms to improve the quality of the interpretation, 15 in France double reading is only done for women with negative findings after the first reading.
Only those women who participated in the first two incident rounds (September 2005 – August 2009) were considered for this study (the prevalent round, September 2003 – August 2005, was discarded to avoid incomplete data and potentially biased results). For the sake of comparison with previous results 3 , we have restricted our study to women aged 50–59 years. Of the 79,405 targeted women at risk (2007), 62,298 underwent at least one mammography during the study period (78.46% attendance rate) (Figure 1).
Flow diagram of the analytical selection criteria for a study of mammographic density and urbanization (women aged 50–59 years, 2005–2009, Franche-Comté region, France)
Mammographic density measurement
The radiographic appearance of the breast on mammography varies among women and reflects variations in breast tissue composition and the different X-ray attenuation characteristics of these tissues. Fat is radiologically lucent and appears dark on a mammogram, whereas connective and epithelial tissues are radiologically dense and appear light.
If a woman had received several mammograms during the study period, only the earliest one was considered. Estimation of density was based on both mediolateral oblique and craniocaudal views. The proportion of the breast that comprises connective and epithelial tissues was expressed as a percentage of total area of the breast, and classified by the use of four categories in the Breast Imaging Reporting and Data System (BI-RADS): 1) the breast is almost entirely fat (<25% glandular), 2) there are scattered fibroglandular densities (approximately 25–50% glandular), 3) the breast tissue is heterogeneously dense (approximately 51–75% glandular), and 4) the breast tissue is extremely dense (>75% glandular). 16 This predominant classification for reporting breast density among radiologists has shown adequate interobserver validity. 17,18 For each woman the measurements were performed independently for her left and right breasts, and when the densities differed between the two breasts, the higher score was retained.
Risk factor data
The degree of urbanization was based on the French Census Bureau rural-urban continuum code for municipalities. This typology takes account of the suburbanization phenomenon on the basis of employment attractiveness. An ‘urban’ municipality is a continuously built-up area providing at least 5000 salaried jobs. The adjacent municipalities that send at least 40% of their working population to this single urban municipality, forming a periurban belt, are classified ‘unipolar suburban’. A ‘multipolar suburban’ municipality is simultaneously attracted to a number of urban municipalities, the 40% threshold not being reached by just one of them. The remaining municipalities are considered ‘rural’. Women's municipalities of residence, provided by the social security administration, were assigned a degree of urbanization according to this classification scheme.
Age at the time of mammography was calculated from the date of birth (provided by the social security administration) and the date of mammography.
Socio-economic status (SES) is a proxy for both the uptake of mammography screening and breast percent density (with positive associations). 19 To circumvent the lack of readily available SES information in the screening database, ‘dependency’ (according to the French social security terminology) was used as a measure of SES. This situation mainly concerns one-income families, because a so-called ‘dependent’ woman (spouse or unmarried partner) is unemployed but eligible for a refund of medical expenses on the same basis as the insured husband or partner. Dependency was derived from the first digit of the social security identification number.
Information of the use of hormone replacement therapy (HRT) at the time of mammography was obtained from women during the clinical breast examination by the radiologist. Unfortunately, no information regarding HRT type, menopausal status (although all of the women were older than 50 years), adiposity, or reproductive factors was available in the screening database.
Statistical analysis
As in any large epidemiological study, there was some incorrect or missing data. We first performed complete-case analyses by excluding all women for whom the outcome or any of the inputs were missing (and thereby obtaining likely biased estimates because the sample of observations that have no missing data might not be representative of the full sample). As a sensitivity analysis, this naïve approach was compared with multiple imputation using a chained equation algorithm for handling missing data. 20 We specified distributions for all variables with missing values conditioned on other variables in the data, and the imputation algorithm sequentially iterated through the variables to impute the missing values using the specified model. We ordered the variables in the iterative imputation process to have increasing numbers of the missing value so that we built the models with as much information as possible. BI-RADS density and urbanization were defined as ordered-categorical variables. Age was introduced as a continuous variable. We generated 10 imputed datasets, and checked the fit of the conditional imputation models by looking at diagnostic plots.
Ordered logistic regression was run on the imputed datasets to test the association between breast density categories (dependent variable) and urbanization degree (independent variable) while adjusting for three confounding factors (age at time of mammography, dependency, hormone replacement therapy). Logits for the response were defined as cumulative logits. We used the proportional-odds model, constraining the regression coefficients to be equal for all cut-off points of the ordinal outcome and fitting therefore a common odds ratio (OR) estimate for all response groups, analogous to a trend test. 21 This common OR therefore reflects the association of a given risk factor with an immediately higher BI-RADS category. Importantly, this interpretation holds across the entire BI-RADS range (from 1 to 4). Urbanization degree was introduced in turn into the models as a categorical or an ordered variable to obtain OR estimates and trend P values, respectively. Analyses performed on each generated dataset were combined to deliver pooled estimates. All statistical tests were two-tailed, and P values below 0.05 were considered statistically significant.
All data were analysed using the R 2.12.1 statistical software (MASS and mi packages) (The R Foundation for Statistical Computing,
Ethics
The procedures at the Franche-Comté cancer screening programme (France) were approved by the National Commission for the Confidentiality of Computerized Data. Only anonymous data were used in the analyses.
RESULTS
To ensure that our results could be evaluated and compared, we restricted our analysis to the 55,597 cancer-free women whose mammogram findings were classified as American College of Radiology (ACR) 1 (negative) (n = 16,611) or ACR 2 (benign breast condition) (n = 38,986), and who had not undergone any previous breast surgery (Figure 1).
We therefore excluded 567 ACR 0 (additional imaging required), 1107 ACR 3 (probably benign), 565 ACR 4 (suspicious abnormality), and 157 ACR 5 (highly suggestive of malignancy) women, as well as 4305 women with previous breast surgery (of whom 107 had undergone surgery for breast cancer).
Baseline characteristics of the study population are summarized in Table 1. Age at the time of mammography was available for all women. Missing rates for the remaining variables ranged from 0.06% (dependency) to 5.40% (HRT use). As a result, 50,165 women had complete data on all variables (90.23% completeness rate). The mean age at mammography was 54.5 (standard deviation, 3.0) years. The majority of women were assigned a 25–50% breast density (BI-RADS grade = 2). Most of the women lived in predominantly urban areas (67.48%). A lower SES (according to the dependency proxy) was found for 13.96% of the women, and the proportion of current HRT users was low (12.58%).
Baseline characteristics of the study participants (55,597 women aged 50–59 years, 2005–2009, Franche-Comté region, France)
Table 2 shows the cross-tabulation of BI-RADS categories and urbanization degrees. When modelling this association in a univariate approach, a statistically significant linear trend was found (P trend <10−3). Compared with women living in rural areas, women living in suburban unipolar settings (OR 1.06, 95% confidence interval [CI] 1.01–1.11) or urban settings (OR 1.07, 95% CI 1.03–1.12) had higher mammographic densities (Table 3).
Cross-tabulation of mammographic density categories with urbanization degrees (counts and column percentages in parenthesis, 52,957 complete cases, women aged 50–59 years, 2005–2009, Franche-Comté region, France)
*Breast Imaging Reporting and Data System
Association of mammographic density with urbanization degree (55,597 women aged 50–59 years, 2005–2009, Franche-Comté region, France)
OR, odds ratio; CI, confidence interval
*52,957 cases in the univariate approach, 50,165 cases in the multivariate approach
†Per 10 years
The linear trend remained unchanged in the multivariate approach (P trend = 10−3). In this fully-adjusted model, women living in urban areas had a significantly increased risk of higher mammographic density (OR: 1.06, 95% CI: 1.02–1.11) (Table 3). A negative and highly significant association with BI-RADS density was highlighted for both age at mammography (OR: 0.41, 95% CI: 0.39–0.43, per 10 years), and dependency (OR: 0.71, 95% CI: 0.67–0.75). Finally, the adjusted OR for HRT was 1.51 (95% CI: 1.43–1.58). No significant interaction between age and urbanization was found (P = 0.11).
The sensitivity analysis using an alternative modelling strategy (multiple imputation) yielded very similar results (Table 3).
DISCUSSION
Geographic variation and poor prediction of individual risk provide evidence that additional risk factors for female breast cancer are still to be identified. 22 Thus the study of important risk factors, such as breast density, may be informative. We found positive trends between the degree of urbanization and mammographic density, the effect of urbanization being mainly apparent in the highest degree (urban settings).
The main strengths of this study are a large population base, a high screening attendance rate, and a very high completeness rate. There are many ways to quantify urbanization (population size, population density, proximity to urban settings, economic activities, etc.). The classification scheme we used has the advantage of recognizing the continuum and integration between urban and rural (going beyond the urban/rural dichotomy) and taking into account both economic activities and geographic patterns.
Although care was taken to reduce the potential for bias, this study was subject to some limitations. The BI-RADS classification does not take into account the thickness of the breast and is based on the projected area rather than on the volume of breast tissue. It has shown lower associations with the breast cancer risk compared with more quantitative methods. 4 There are several sources of error when measuring mammographic density, regardless of the method used. Different film types, film projections, mammography machines, radiologist techniques, and levels of compression (depending on radiologist technique or patient sensitivity due to menstrual phase or hormone use) can all potentially alter the appearance of dense tissue in a mammogram. 23 Because numerous radiologists were involved in this screening programme, these potential sources of error in measurement are likely to attenuate the observed associations between mammographic density and urbanization.
The French breast cancer screening programme relies on self-report regarding HRT use. Assessing data reliability would require the linkage of individual medical records or prescribing databases to screening records, which was unfortunately impossible because of strict French privacy protection laws.
Another limitation lies in the lack of individual adiposity measures, such as the body mass index (BMI). Because the breast is one of the fat deposits in the body, adiposity is positively associated with non-dense and total breast areas and is therefore negatively associated with percent dense area. If obesity prevalence was to be higher in rural settings of the study area, then it could confound the observed findings. This potential bias is, however, not supported by a French nationwide survey performed in 2009. The prevalence of obesity (BMI ≥30) was found to be 15.5% in rural villages (<2000 inhabitants), 16.1% in small size towns (2000 to 20,000 inhabitants), and 14.3% in medium-sized towns (20,000 to 100,000 inhabitants). 24 These rates are in line with a recent study conducted in England and Wales that showed no BMI gradient with a urban/rural indicator. 19
Complete-case methods, which simply discard observations with any missing data, require stronger assumptions than does imputation. 25 Conversely, the use of a multiple imputation algorithm enables us to consider the sample of the observations as fairly representative of the target population by allowing inclusion of all individuals in analyses. However, both approaches performed equally, most likely because of the low overall number of missing values.
We found a significant linear gradient (P trend <10−3) between mammographic density and urbanization in a univariate approach, in agreement with Perry et al. who reported a significant trend across BI-RADS categories (P trend = 0.04). 3 Moreover, in our study this trend remained unchanged after adjusting for known risk factors (age, SES, and HRT use). The ORs for suburban unipolar and urban settings are small in magnitude; the statistical significance most likely is driven by the large sample size. Conversely, Aitken et al. found no evidence of an association between a binary urban/rural indicator and percent density, but reported heterogeneous results consisting of a negative absolute difference with a minimally-adjusted model and a positive absolute difference with a fully-adjusted model. 19
Regarding confounding factors, a decrease in percent breast density with increasing age confirms previous reports, 26,27 reflecting the age-differences in breast tissue composition. 6 HRT use, strongly socially patterned, can also be considered a proxy for SES. HRT is associated with an increased mammographic density among current users that is in line with several observational and intervention studies, 28 and consistent with oestrogen stimulation of breast tissues. However, there is heterogeneity in a woman's breast tissue response to the use of these exogenous hormones, with only 20–35% of women experiencing increases in breast density upon initiation and continuation of HRT. 29 This could explain the relatively modest risk associated with HRT use (OR 1.51), which is very similar to the OR reported for HRT users in an Italian study (OR 1.49). 30 Dependency was used as a socioeconomic indicator. Our results are in agreement with the very few studies that have addressed this issue, which have showed an association between higher education and high breast percent density 19,29–31 reflecting a lower mean BMI in women of higher SES. 19
We found a strong link between mammographic density and urbanization, although the possibility of residual confounding due to unmeasured variables (such as reproductive factors) cannot be entirely dismissed. Because of their higher breast density, women living in urban settings will not only have a higher risk of developing breast cancer but also a lower probability of having it detected earlier through screen-film mammography. Knowledge of this urbanization gradient in density (whatever its mechanism) may therefore help to identify groups of women who may require full-field digital mammography for the early detection of breast cancer.
Current screening guidelines from many groups (such as the American Cancer Society among others) recommend mammography every one or two years starting at age 40 or 50 years. This one-size-fits-all approach is challenged by recent findings, suggesting that mammography should be tailored to risk factors. Shousboe et al. showed that the most cost-effective frequency of mammography depends on a woman's age, breast density, family history, and history of breast biopsy. 32 However, to move toward a more personalized mammography recommendation, women will need to know their breast density, and will therefore have undergone an initial mammography and subsequently received information on their breast density. In the absence of these prerequisites, urbanization (although its association with breast density is small in absolute terms) could assist primary care providers in recommending the best screening strategy in a risk factor-based approach.
Footnotes
ACKNOWLEDGMENTS
The authors thank Dr. Arlette Le Mouel from the Franche-Comté cancer screening program for allowing access to data and Albert Hoareau from EpiConcept Company for data extraction.