Introduction

Female breast cancer mortality rates have, on the whole, been decreasing in the United States (U.S.) for well over two decades, due in part to advancements in early detection and treatment [1]. However, similar to the secular declines of other disease-specific mortality rates, breast cancer mortality rate reductions have not been experienced equally by all racial/ethnic groups, resulting in persistent disparities [2, 3]. Interestingly, disparities vary by geography, including counties and cities [2, 4, 5]. For instance, between 1990 and 2009, in most US cities, white women experienced greater decreases in breast cancer mortality than did their black counterparts [2]. In addition, although white women experienced lower mortality from breast cancer compared with their black counterparts, the magnitude of disparity varied across cities, and in some cases even favored black women [2].

The etiology of racial/ethnic disparities in breast cancer survival is multifactorial and challenging to unravel [6]. Nonetheless, extensive research over the past two decades has identified some contributing factors such as racial/ethnic differences in tumor stage and biology [7,8,9,10,11], comorbidities [12, 13], timeliness of treatment [14, 15], and receipt of guideline-concordant treatment [16, 17]. Although studies have largely focused on individual-level factors that can impact disparities, research suggests that area-level factors, including those related to socioeconomic and healthcare supply factors, also contribute to individual-level racial disparities in breast cancer mortality [18, 19]. However, very few studies have examined factors that might explain geographic variation in race-specific population-level breast cancer mortality rates [4, 5].

Descriptive studies that document local-area racial disparities in breast cancer mortality are critical for prompting action at the local level [20,21,22,23]. However, additional research is needed to further our understanding of what drives these differences, and, specifically, to explore city-level predictors when comparing mortality in black and white women. Examining data at the city level allows for a more localized and tailored response. Further, focusing on race-specific mortality rates as an outcome rather than on racial/ethnic disparities provides a more nuanced approach because disparity measures obscure race-specific outcomes. Overall, this type of analysis can help to more accurately guide the efforts of public health leaders as they determine where and how to invest limited resources.

Ecologic studies are particularly well suited to examining city-level drivers of population-level breast cancer mortality, as ecological studies can identify and illuminate context-specific differences that may be contributing to disparate outcomes. They are also useful for generating hypotheses of possible determinants of disease distribution, particularly at the group level. For instance, ecologic analyses enable the exploration of whether breast cancer mortality disparities across cities are related to healthcare resource availability or mammography uptake across cities, and whether such differences impact white versus black female breast cancer mortality rates. Several recent studies have demonstrated the utility of employing city-level demographic and socioeconomic variables to help explain the variation in mortality disparities related to lung and prostate cancer, diabetes, and heart disease [24,25,26,27]. In the current study, we employ an ecologic, cross-sectional approach, building on this body of work while furthering the large body of research that has primarily used individual-level data to examine the factors contributing to disparities in breast cancer stage at diagnosis, treatment, and survival [18, 28,29,30,31,32]. Our unique contributions include using a city-level ecologic approach; focusing on age-adjusted race-specific breast cancer mortality rates (as opposed to stage at diagnosis, treatment, and survival) as the outcome variable; and identifying black/white differences in important city-level breast cancer mortality predictors (i.e., sociodemographic inequality, health insurance, segregation, religiosity, healthcare resources, and breast care behaviors).

Methods

Sample

The 50 most populous US cities were determined from 2005 US Census Bureau data [33]. The methods used in the calculation of the non-Hispanic black (Black) and non-Hispanic white (White) breast cancer mortality rates have been described at length elsewhere [3] and are summarized below.

Main outcome and data sources

The outcome variable for this analysis was the age-adjusted, race-specific breast cancer mortality rate. The numerator for each mortality rate was composed of the number of breast cancer deaths in a given race/ethnic/age group for each city. Using data from the National Center for Health Statistics (NCHS) [34] for 2010 to 2014, we extracted race-specific deaths by age group (25 to 34, 35 to 44, 45 to 54, 55 to 64, 65 to 74, 75 to 84, and 85+ years) among women in whom the cause of death was malignant neoplasm of the breast (ICD-10 code C50) for each US city [33]. Race- and age-specific population-based denominators were also obtained for each city from the US Census Bureau and the American Community Survey. The process for calculating denominators is described in detail elsewhere [3].

Predictor variables and data sources

The potential predictors of the breast cancer mortality rate are summarized in Table 1 and, with a few exceptions, are city-level variables that provide information on sociodemographic characteristics, socioeconomic inequality, health insurance, segregation, religiosity, healthcare resources, and breast care behaviors. These categories of variables have been shown to impact breast cancer outcomes such as stage at diagnosis, treatment, survival, and disparities in mortality [18, 28,29,30, 32, 35, 36]. Notably, race- and/or sex-specific predictors were used when available.

Table 1 Predictor and outcome variables with descriptions/definitions, population level, years covered, and source
Full size table

All data are publicly available and variables were selected for the analysis based on two factors: (1) there was evidence that they may impact breast cancer mortality; and (2) the data were available at the city or similarly local level. The data for the present analysis were gathered in 2016. We attempted to obtain data for the predictors that were collected prior to the time period for breast cancer mortality (2010–2014). Data on sociodemographic variables (median household income, percentage poverty, and education level) and income inequality (Gini index) were obtained from the 2005–2007 American Community Survey. The sociodemographic variables were also used to calculate race-specific (absolute) differences in median household income (white minus black median household income; white higher in all cases), as well as in poverty (black minus white poverty; black higher in all cases) and education levels (white minus black education; white higher in all cases). These calculated measures were created to capture the magnitude of socioeconomic inequities. Health insurance information (percentage female uninsured) was obtained from the 2008–2010 American Community Survey. Measures of segregation (black and white indices of isolation) were obtained from Brown University’s Diversity and Disparities Project (2005–2009). Data on religiosity/community resources (percentage with religious affiliation and number of religious congregations per 10,000 population) were acquired from the Association of Statisticians of American Religious Bodies (ASARB) and measured at the metropolitan statistical area level. Healthcare resources were city-level variables and included the number of federally qualified health centers (the FQHCs per Population variable used in this study), family physicians, Medicare primary care physicians (PCPs), and Medicare facilities per 100,000 population. Additionally, information on the number of Food and Drug Administration (FDA)-certified mammography facilities (per female population) was obtained. Healthcare resource data originated from a variety of sources, including the Health Resources & Services Administration (HRSA), American Board of Family Medicine (ABFM), Centers for Medicare & Medicaid Services (CMS), and FDA. Breast care behavior information (current mammography use from 1997 to 1999 and 2000 to 2003) originated from the Small Area Estimates for Cancer-Related Measures website provided by the National Cancer Institute (NCI) at the National Institutes of Health (NIH).

Statistical analysis

In order to identify potentially significant predictors, we ran age-adjusted models between each predictor and race-specific breast cancer mortality. Potential predictors were examined as continuous and dichotomous variables. Variables were dichotomized at the 50th as well as 75th percentile levels. Variables that were associated with breast cancer mortality at the 0.20 α-level were included in the full multivariable models. Preference was given to the continuous version of a variable when it was associated (α = 0.20) with the outcomes in both its continuous and dichotomous forms. The overall purpose was to achieve the optimal predictive model, using population-level data to obtain the best estimates of the most highly predictive variables. Consequently, a backward-selection approach was used to drop non-significant (α = 0.05) variables from the full models. Variables were dropped, one at a time, based on the magnitude of their p value. Only statistically significant variables remained in the final model.

Multivariable negative binomial regression models were constructed to examine the association between area-level predictors and race-specific breast cancer mortality rates. Multivariable analyses were restricted to women aged 25 years and older due to no deaths having occurred in the datasets among women aged 0 to 24 years. Additionally, three cities were excluded from the multivariable analyses. Due to previously cited uncertainties concerning data reliability on Miami, Florida [2, 25, 27], we performed analyses (e.g., boxplots) on all the variables and found that the city’s data points were consistently outliers. As a result, Miami was excluded from the analysis. Two additional cities were excluded due to lack of city-level population data availability (Louisville/Jefferson County, Kentucky and Nashville/Davidson County, Tennessee), resulting in a total of 47 cities included in the analysis. All models were age-adjusted and separately estimated for each racial group, with the number of breast cancer deaths as the outcome and the natural log of the population size (by city, race/ethnicity, and age group) as the offset.

In the multivariable negative binomial regression model, the exponentiated coefficients represent breast cancer mortality rate ratios (RRs) while simultaneously adjusting for all variables in the model. Statistical analyses were carried out using SAS 9.3 (SAS Institute, Cary, North Carolina). The GLIMMIX procedure was used for modeling to account for the nested structure (age group data within cities) of the data.

Results

Table 2 summarizes the characteristics of the 47 US cities included in the analysis. A more detailed table containing all the data used is available in the Appendix, including the age-adjusted, race-specific breast cancer mortality rates for all 47 cities as well as their sociodemographic characteristics. With few exceptions, Table 2 reveals that the race-specific summary measures were less favorable for black women compared with white. For instance, the 5-year average age-adjusted breast cancer mortality rate was 37% higher (33.0 vs. 24.1 per 100,000) for black than for white women. Furthermore, the median household income was 44% lower for black than for white women ($31,208 vs. $55,512). Other characteristics related to health insurance, healthcare resources, and mammography use varied widely across cities. For example, the median percentage of uninsured women was 15.2%, although it ranged from 3.8 to 30.6%.

Table 2 Summary statistics of characteristics of the 47 largest U.S. cities
Full size table

The age-adjusted associations between each predictor and the outcome of interest revealed that potential predictors of breast cancer mortality rates generally differed by race (Table 3). For instance, for every $5,000 white/black (absolute) difference in the city-level household income there was an associated increase of 2% in the city-level black rate (RR = 1.02; p = 0.10); however, this measure was not associated with the white rate (RR 0.99; p = 0.52). On the other hand, for every five-point white/black difference in the city-level proportion of high school graduates there was an associated 4% decrease in the white rate (RR 0.96; p = 0.11), but no corresponding association with the black rate (RR 1.02; p = 0.39). Evidence also showed that (2000–2003) higher current mammography use was associated with lower black (RR 0.93; p = 0.02) and white (RR 0.96 = 0.20) mortality rates.

Table 3 Age-adjusted associations between potential predictors and race-specific breast cancer mortality
Full size table

Table 4 illustrated that the full model for the white mortality rate differed from that of the black rate in terms of the potential predictors. The full model for the white rate included the following seven variables: white/black absolute difference in the percent of high school graduates, index of isolation for the white population, level of Medicare primary care providers, availability of Medicare facilities, proportion of those with a religious affiliation, presence of congregations, and current mammography use (2000–2003). The full model for the black mortality rate included eight potential predictors. It included three measures of socioeconomic disparity: white/black absolute difference in the median household income and percent of high school graduates, and black/white difference in percent in poverty. In addition, the following five variables were included: percent of uninsured females, availability of family physicians, level of mammogram facilities, and current mammography for 1997–1999 and 2000–2003.

Table 4 Multivariable negative binomial regression of factors associated with breast cancer mortality rate
Full size table

After the stepwise elimination of non-significant predictors, each final model only included three predictors. The final adjusted model for the white rate, included white/black absolute difference in the percent of high school graduates, level of Medicare primary care providers, and the presence of congregations. A five-point increase in the white/black difference in the percentage of high school graduates was associated with a 5% lower white mortality rate (RR = 0.95; 95% CI 0.91–0.99). Cities that ranked in the 75th percentile in terms of Medicare PCPs per population had a white mortality rate that was 15% higher than that of lower ranking cities (RR = 1.15; 95% CI 1.04–1.28). Finally, a five-point increase in the number of congregations per population was associated with a 13% (RR = 0.87; 95% CI 0.77–0.97) lower white mortality rate. The final model for the black rate only included the white/black absolute difference in the median household income, level of mammogram facilities, and current mammography (1997–1999). In this final adjusted model, a $5,000 increase in the white/black difference in household income was associated with a 3% (RR = 1.03; 95% CI 1.01–1.05) higher black mortality rate. A five-point increase in the number of mammogram facilities per population was associated with an increase in mortality (RR = 1.07; 95% CI 1.03–1.12). However, a five-point increase in the percentage of women with a current mammogram from 1997 to 1999 was associated with a 7% reduction (RR = 0.93; 95% CI 0.89–0.97) in the black mortality rate.

Discussion

The current study employed a city-level ecologic approach to investigating predictors of breast cancer mortality rates for black and white women. We found that larger white/black differences in level of education and increased number of religious congregations were associated with a lower breast cancer mortality rate for white women, whereas a higher availability of Medicare PCPs was associated with higher mortality. For black women, larger white/black differences in household income and number of mammography facilities were associated with a higher breast cancer mortality rate, whereas the percentage of women with a current mammogram was associated with lower mortality.

Ecologic studies are useful for hypothesis generating and providing insight into group-level differences, such as city-level variation in breast cancer mortality rates. Within this context, the study findings prompt two immediate questions: (1) Why might an education gap favoring white women be associated with lower white mortality rates for breast cancer? (2) Why might an income gap favoring white women be associated with higher black mortality rates for breast cancer? Previous research suggests that advances in technology related to both screening and treatment may be more available to those with more resources, and less available to those with fewer [37,38,39,40]. As Phelan and Link posited in their fundamental causes theory, “When we develop the ability to control disease and death, the benefits of this newfound ability are distributed according to resources of knowledge, money, power, prestige, and beneficial social connections’’ [41, p. 27]. The findings from the present analysis lend additional support to the notion that women with more resources—education and income in this case—may be better equipped to take advantage of technologic advances than their counterparts with fewer resources, and this may have a direct impact on breast cancer mortality.

The fundamental causes framework also helps provide insight into two other significant predictors of black female breast cancer mortality (mammogram facilities and mammography usage). The present analysis suggested that the number of mammography facilities per population was positively associated with black female breast cancer mortality, a finding that seems counterintuitive. We must first acknowledge that, given the cross-sectional nature of our study (in which determining temporality and causality are not possible), this finding may simply be a product of higher breast cancer mortality rates leading to increased mammography facilities. On the other hand, in the context of the fundamental causes framework, this finding may be an example of black women’s relative inability to take advantage of technologic advances due to their lower social standing. Thus, even when residing in areas with relatively high numbers of available mammography facilities, black women may still be unable to access these resources for reasons directly tied to their comparatively lower income levels. In addition, the locations of these facilities may make them inaccessible, particularly if transportation is not readily available. Other barriers, such as inflexible employers and child/elder care responsibilities, may also prevent access. Moreover, competing survival priorities may make it difficult for black women to prioritize breast cancer screening and prevention activities [35].

Evidence also suggests that the quality of the mammography facility is an important factor in survival outcomes, and black women generally have access to facilities of lower quality compared with white women [42, 43]. Thus, if a large number of facilities are available, but they are of low quality, it may negatively impact black female breast cancer mortality. However, when black women are able to access advances in screening technology, as measured in the present analysis by the variable of mammography within the past 2 years (i.e., current use), lower breast cancer mortality rates are observed.

The fundamental causes theory also proves useful in examining the relationship between the number of religious congregations per population and white breast cancer mortality rates. The theory establishes white women as the group best positioned to take advantage of technologic advancements in breast cancer screening and treatment. Consequently, white women who have benefited from these advancements subsequently become beneficial social connections, diffusing the information to their peers (often other white women). To the extent that religious congregations provide opportunity for information exchange, larger numbers of congregations may serve to increase exposure to information among white women. These women, in turn, may be more likely to make use of the technologic advancements, potentially decreasing the overall white breast cancer mortality rate. Although black women may also be exposed to this information, particularly in more integrated settings, it is plausible that the history of medical mistrust in the black community would inhibit action on their part [35].

The finding of a positive relationship between Medicare PCPs per population and white female breast cancer mortality is not in the expected direction. As with the previously discussed finding of a positive association between mammography facilities and black female breast cancer mortality, we must again acknowledge that the cross-sectional nature of the data prevents us from establishing causality or temporality. Previous research has demonstrated that, following the 1991 introduction of Medicare reimbursement for mammography, both black and white elderly women experienced declines in breast cancer mortality, with white rates declining more rapidly than black [40]. Whereas the findings from the present analysis do not directly contradict this evidence, they also do not support it. Unless our measure of Medicare PCPs is assessing treatment more than it is screening, and/or these facilities require supplemental insurance for treatment, we would expect our findings to be in the opposite direction.

Implications

The results of this analysis could help inform interventions at the local level. Although reducing income inequality and increasing investment in education would be ideal next steps, these may be beyond the control of local health officials, who may be looking for more actionable approaches to reducing breast cancer mortality. We urge local government leaders to adopt a “Health in All Policies” approach in finding ways to promote healthy environments through intersectoral collaboration [44]. For example, traditional navigation programs have long helped to assist women in overcoming barriers to access such as lack of insurance and transportation [45,46,47]. To the extent that these programs include funding for outreach, there will also be opportunities to educate women about existing guidelines and recommendations related to breast health and mammography, about which black women are less likely to be informed than are white women [35, 48]. Navigators can also be a great resource for helping women discuss treatment options with their physicians.

Strengths and limitations

The main strength of this analysis is that, to our knowledge, it is the first to solely use population-level data to identify the variables most highly predictive of breast cancer mortality for black and white American women, at the city level. We obtained mortality data from the NCHS, a reliable and consistent source for this type of information. For city-level predictor variables, data sources included the American Community Survey from the US Census Bureau, as well as several other national sources such as the HRSA, FDA, NCI, and CMS.

The main limitations of this analysis are the lack of causality and temporality inherent to its cross-sectional design and the limited number of study variables. Because this is an ecologic, cross-sectional study and the outcome (mortality) occurs late in the continuum, it is not possible to establish causality. Due to the lack of individual-level data, it is not possible to account for factors such as stage at diagnosis, tumor subtype, treatment uptake and adherence, physical activity, or comorbidities. Relatedly, our results do not extend to the individual as the observations were at the city level. Additionally, a handful of variables (e.g., health insurance and mammography use) were not available in race-specific terms. In some cases, data may come from the same source, but were not available for the same years (e.g., ACS data for percent uninsured (2008–2010) and income (2005–2007). Finally, we did not have information on provider/facility location relative to individuals.

In conclusion, our ecologic analysis found that predictors of breast cancer mortality differ for black and white women and that the relatively greater resources available to white women facilitate their access to technologic advancements in screening and treatment, which may ultimately drive differences in black and white breast cancer mortality rates. Our hope is that these findings can be used to inform local-level responses.