The Effect of Patient and Contextual Characteristics on Racial/Ethnic Disparity in Breast Cancer Mortality Free

There are well-described differences in breast cancer mortality by race and ethnicity in the United States. For example, breast cancer mortality rates were similar for African Americans and non-Latina Whites (NL Whites) in the United States until the late 1970s, but these rates began to diverge in the early 1980s (1–3). Reasons for racial/ethnic differences in mortality after breast cancer diagnosis are not completely understood. Demographic, screening, disease characteristics (4), socioeconomic and lifestyle factors (5–7), adequacy of treatment (8–13), presence of comorbidities (14, 15), neighborhood factors (16), differential influence of tumor biology on outcome (17, 18), and patient hematologic traits (19) have been cited or investigated as possibly accounting for these observed disparities. Irrespective of race and ethnicity, history of comorbidity (20, 21), and lifestyle factors such as body size and physical activity (22–27) have been shown to influence breast cancer–specific mortality, albeit not in all studies (28). Most of these previous studies were conducted in NL Whites only. Few studies have investigated the combined roles of clinical, lifestyle, and contextual factors [i.e., those related to socioeconomic and manmade (“built”) physical attributes of an individual's surroundings; refs. 29, 30] in relation to breast cancer–specific mortality.

We have previously reported on the role of institutional and neighborhood contextual factors (31, 32), low physical activity (33), large body size (34), and comorbidities (35) in relation to mortality following breast cancer diagnosis in the large California Breast Cancer Survivorship Consortium (CBCSC) study, which comprised 12,210 invasive breast cancer cases diagnosed between 1993 and 2007 (36). Our objective in the current analysis was to investigate the contribution of each of these four domains of risk factors to the apparent racial/ethnic disparity in breast cancer–specific mortality within the context of a single baseline model that included common demographic, clinical, and disease characteristics.

The CBCSC, established in 2011, pooled data from six California-based studies to investigate racial/ethnic disparities in mortality (36). They included the Asian American Breast Cancer Study (AABCS), the Women's Contraceptive and Reproductive Experiences study (CARE), the San Francisco Bay Area Breast Cancer Study (SFBCS), the Life after Cancer Epidemiology study (LACE), the California Teachers Study (CTS), and the Multiethnic Cohort study (MEC). The CBCSC harmonized and pooled interview information from breast cancer cases and linked this information with neighborhood data from the California Neighborhoods Data System (29), data on characteristics of the primary reporting hospitals, and corresponding clinical characteristics and mortality data from the California Cancer Registry for cases diagnosed with primary invasive breast cancer between 1993 and 2007 (36). Additional methodologic details of the construction of the cohort are included in Supplementary Methods. The protocols for the CBCSC study were approved by the institutional review boards of all participating institutions and by the California State Committee for the Protection of Human Subjects.

We used a previously developed (36) Cox multiple regression model as the starting point for examining the contribution of individual disease and patient characteristics to the observed racial/ethnic disparity in breast cancer–specific mortality. This model, which we refer to as the baseline model, used the attained age time scale, was stratified by study, and included racial/ethnic group, age at diagnosis as a continuous variable on both the natural and log scales, and clinical and demographic variables that contributed significantly to predicting breast cancer–specific mortality. The baseline model variables are listed in Supplementary Table S1 and further details of its development are described in the Supplementary Methods. The CBCSC included four exposure domains: contextual factors (31, 32), physical activity (33), body size (34), and comorbidity (35) to investigate their roles in breast cancer-specific and overall mortality. For this analysis, we selected representative variables that were found to be predictive of outcome in these analyses as exposures of interest.

We performed several types of analyses in the baseline model to assess the influence of individual clinical and lifestyle factors on the racial/ethnic disparity in breast cancer–specific mortality. A univariable analysis started with the racial/ethnic group variable alone in a Cox regression model adjusted for continuous age at diagnosis and stratified by study. The remaining variables in the baseline model (Supplementary Table S1) were included individually, one by one in turn, and the resulting changes in the HR between African Americans, Latinas, and Asian Americans versus NL Whites were described. A multivariable analysis started with all variables in the model, and the change in HR of each racial/ethnic group was noted when variables were removed, individually, one by one in turn. These two analyses reveal slightly different information about the influence of the different variables. The univariable analysis identifies variables that by themselves have a large influence on HR, although if they are highly correlated with other variables these also would have similar influence on HR in this analysis. The multivariable analysis identifies variables that have, by themselves and without regard to other variables, a large influence on racial/ethnic disparity. In this analysis, removing a variable that is correlated with other variables in the model will have a smaller influence on HR, because some of the information it contains remains in the model through the remaining correlated variables. In addition, a sequential analysis was performed, starting again with the racial/ethnic group variable alone in the model, and adding individual variables in a stepwise fashion in the order of their univariable significance. This analysis assesses the change in HR as more and more variables are added, with larger changes in HR occurring whenever the first influential variable in a correlated set or an individual influential uncorrelated variable is added. Tests of significance of a single variable or sets of variables of interest when added to a reference Cox model were based on the Cox partial likelihood ratio test.

We repeated these three analyses to investigate the further influence of contextual, physical activity, body size, and comorbidity-related factors. The difference in these additional analyses was that the starting, reference model included all of the variables in the original baseline model. We also restricted these analyses to subgroups of the cohort that included only breast cancer cases from individual studies that collected data relevant to each of these four domains, as will be described below.

The measure that we devised to summarize the racial/ethnic disparity in breast cancer–specific mortality for a particular model was |$D = \sqrt {\frac{{\sum {{n_i}} {{\{{{\beta_i} - {{\overbar \beta{_\bullet}}}} \}}^2}}}{{\vskip2pt\sum {{n_i}}}}}$|⁠, the sample-size weighted standard deviation of log_e HR estimates for racial/ethnic groups from the model. Here β_i is the log_e HR estimate (i.e., the Cox regression coefficient) for the ith racial/ethnic group, n_i is the racial/ethnic group sample size, and |${\overbar \beta_{\bullet}}$| is the sample-size weighted mean of the β_i. D is independent of which group is chosen as the reference group and will be closer to zero for models where there is a smaller difference in breast cancer–specific mortality HR across the racial/ethnic groups.

To assess relative changes in this disparity measure, define D₀ as the total disparity in a reference model. For the analyses of variables in the original baseline model, D₀ was computed from a model that included only the racial/ethnic group variable, stratified by study and adjusted for age, but with no other variables from the baseline model included. For the analysis of variables related to contextual, physical activity, body size, and comorbidity factors, D₀ was computed from a model that included all of the original baseline model variables. The relative influence of a particular variable on the disparity across racial/ethnic groups in the context of one of these reference models was defined as |${\frac{{{D_-}-{D_+}}}{{{D_0}}}}\times 100$|⁠, that is, the percent of the total disparity that is contributed by the variable, where D₊ and D₋ denote the disparity measures for models that do and do not include a variable of interest, respectively. Note that D is a general measure of disparity that will not by itself reflect complex changes in the HR between racial/ethnic groups. Therefore, we also examined changes in HR for individual racial/ethnic groups. We also devised an approximate likelihood ratio-based test of significance of change in the disparity measure D. This test, termed “disparity χ²,” is described in detail in Supplementary Methods.

All reported P values are based on two-sided tests, with P < 0.05 generally regarded as significant.

The current analysis included 12,098 of the original 12,210 cases in CBCSC cohort—112 cases designated as “other” race/ethnicity were excluded. The cohort comprised 6,501 (54%) NL Whites, 2,060 African Americans (17%), 2,032 Latinas (17%), and 1,505 (12%) Asian Americans. Table 1 summarizes the frequency of selected variables and their distribution by racial/ethnic group, and Supplementary Table S1 shows all variables that are included in the baseline model.

Figure 1 shows visually the analysis we conducted to examine the influence of the variables in the baseline model on the disparity in breast cancer–specific mortality across racial/ethnic groups. The top panel of Fig. 1 shows the HR for African Americans, Latinas, and Asian Americans relative to NL Whites as variables were added to the model. The left-most point, labeled “RACE” is the model with racial/ethnic group alone, stratified by study and adjusted for age at diagnosis, but without the remaining variables. In this starting model, breast cancer–specific mortality in African Americans was higher compared with NL Whites, with an HR of 1.69 (95% CI, 1.46–1.96; see Table 2, column 1), whereas Latinas had similar mortality to NL Whites (HR 1.00; 95% CI, 0.84–1.19), and Asian Americans had lower mortality compared with NL Whites (HR 0.52; 95% CI, 0.33–0.85).

The succession of points from left to right in the top panel of Fig. 1 resulted from adding each of the variables sequentially to the model in the order of their univariable significance, that is, their significance in a model that otherwise included only racial/ethnic groups with stratification by study and adjustment for age at diagnosis. The right-most point is the multivariable model that included all 14 variables shown in Supplementary Table S1 (the variables corresponding to the labels on the figures are shown in Table 3). The disparity in breast cancer–specific mortality between African Americans and NL Whites was substantially reduced by adjusting for all of these variables (HR 1.13; 95% CI, 0.96–1.33; Table 2, column 2). Although adding AJCC stage changed the HR somewhat for Latinas and Asian Americans, adjusting for all of the variables in the model had a much smaller impact than for African Americans. For Latinas, the HR decreased to 0.82 (95% CI, 0.69–0.99) in the fully adjusted model, whereas it increased slightly to 0.59 (95% CI, 0.38–0.96) in Asian Americans.

The bottom panel of Fig. 1 shows the univariable and multivariable percent relative influence of individual variables in the model, providing more information about which variables were most influential in affecting racial/ethnic disparity in breast cancer–specific mortality. The variables that appeared to have the most influence are those related to disease characteristics and presentation (specifically AJCC stage, nodal involvement, tumor size, and grade, as well as surgical treatment, which is, in large part, dictated by disease presentation), neighborhood socioeconomic status (NSES), and smoking status at diagnosis. From the top panel, one sees that as variables are added to the model, the largest changes occur with AJCC stage, NSES, grade, and smoking, with AJCC stage affecting HR for all three racial/ethnic groups, NSES and grade affecting mostly African Americans and Latinas, and smoking affecting mostly Latinas and Asian Americans.

The large disparity in breast cancer–specific mortality between African Americans and NL Whites was explained to a large extent, but not completely, by the variables in this model, with comparatively smaller influence of these variables on the Latinas versus NL Whites and Asian Americans versus NL Whites disparity. The measure of disparity (D) was 40% smaller in the full model compared with the racial/ethnic group only model (Table 2, column 2, bottom), and this reduction in disparity was significant [χ²(3) = 40, P < 0.0001]. To examine whether this large disparity was due to the small number of cases with later-stage tumors (AJCC Stage III–IV), we repeated these analysis restricting only to the 10,789 cases with AJCC Stage I–II disease. Supplementary Table S3 and Supplementary Fig. S1 summarize these results. Interestingly, although the initial racial/ethnic groups disparity was similar to the full cohort analysis, the decrease in D from adjusting for baseline model covariates was 30% rather than 40%, due primarily to a somewhat larger remaining HR for African Americans versus NL Whites (HR 1.28; 95% CI, 1.05–1.55) in the full baseline model. HRs for Latinas and Asian Americans differed by a much smaller amount compared with the full cohort analysis. Clinical variables, NSES, and smoking once more had the largest influence on reducing overall disparity.

The variables that were included for each of the four domains of interest were, for contextual factors, measures of total business count, housing crowding, urban/rural categorization, population density, commuting via public transportation, restaurant environment, traffic density, and hospital socioeconomic composition (31, 32); for physical activity, measures of activity between the ages of 10 and 19, between age 10 and diagnosis of breast cancer, and in the 10 years prior to diagnosis of breast cancer (33); for body size, prediagnosis body mass index (BMI; ref. 34); and for comorbidity, history of hypertension, diabetes, or myocardial infarction (35). Because not all of the CBCSC studies collected data relevant to the four domains of interest, the domain-specific analyses were restricted to subcohorts comprising cases from the studies with data relevant to these domains. These variables and their distribution among racial/ethnic groups within these subcohorts are described in Supplementary Table S2. Note that the subcohorts with data on contextual factors, body size, and comorbidity had similar although not identical racial/ethnic composition. The physical activity subcohort had nearly equal distributions in the racial/ethnic groups and differed from the other three subcohorts and the overall cohort because two of the three studies (CTS, LACE) not included had predominantly NL Whites. Also, this subcohort included only one study (AABCS) with Asian Americans, which prohibited estimating the HR for Asian Americans in this analysis as this category was confounded with the study stratification variable.

Table 4 shows the HR estimates for breast cancer–specific mortality for the four domains of interest (contextual, physical activity, body size, and comorbidity), both under the complete baseline model (left column in each section) and with the relevant domain-specific variables included (right column in each section). Note that the HR estimates in the baseline models were similar in magnitude and pattern, although not identical to, the HR estimates in the baseline model that included the full cohort (Table 2, columns 1). Hence, we think it is reasonable to use these subcohort baseline models as a reference to investigate the influence of domain-specific variables on racial/ethnic groups disparity.

The right column of each of the four sections of Table 4 shows the HR estimates when the domain-specific variables are added to the corresponding baseline model. Adding these variables had minimal further influence on the disparity in breast cancer-specific mortality across the four racial/ethnic groups. Contextual factors (those aside from NSES) and physical activity factors had a negligible effect (percent change in disparity was <±1%, both overall P > 0.20). The effects were larger when body size (percent change in disparity = −3.70%, overall P = 0.20, disparity P > 0.95) and comorbidity (percent change in disparity = −2.30%, overall P = 0.003, disparity P > 0.95) factors were added to the baseline models. Figure 2 shows the HR changes as variables are added sequentially to the full baseline model. This figure visually reinforces the observation that accounting for these variables has minimal influence on the residual disparity in breast cancer–specific mortality across racial/ethnic groups, even though, as in the case of the comorbidity variables, they may be significantly associated with breast cancer-specific mortality. The parallel analysis restricted to cases with stage I to II disease showed similar results (Supplementary Table S4).

We show in this analysis of more than 12,000 breast cancer cases that individual-level patient and disease characteristics and NSES explain part (but not all) of the apparent disparity in breast cancer–specific mortality across different racial/ethnic groups, with disease characteristics and presentation, NSES, and smoking the most influential in explaining this disparity. Moreover, these factors have the most influence on the disparity between African Americans and NL Whites and comparatively little influence on disparities between Latinas and Asian Americans and NL Whites. In addition, we show that other contextual factors, physical activity, body size, and comorbidities apparently play a negligible role in explaining the observed racial/ethnic disparity in breast cancer–specific mortality.

Our results are in agreement with those of Curtis and colleagues (4), who used Surveillance, Epidemiology, and End Results (SEER) Medicare data on 41,020 women over 68 years of age with incident breast cancer to examine the influence of mammographic screening, tumor characteristics, tumor biology, treatment, comorbidities, and community income level on breast cancer–specific mortality between Whites, African Americans, Hispanics, and Asian and Pacific Islanders (API). The HR estimates for cancer-specific mortality from their baseline model, adjusted only for age and SEER site, for African Americans, Hispanics, and API (1.63, 1.24, and 0.59, respectively) and from their fully adjusted model (1.08, 0.88, and 0.61, respectively) are similar to the estimates from our baseline and fully adjusted models, respectively. We considered additional factors, including smoking and other lifestyle and contextual factors that were not available in their study. The association between smoking and breast cancer outcome has been reported (37, 38), and our observation that adjusting for smoking reduced the disparity between both Latinas and Asian Americans compared with NL Whites is consistent with the smaller fraction of Latinas and Asian Americans in our cohort who were smokers. The findings from both Curtis and colleagues (4) and our study suggest that the most important determinants of mortality disparity are tumor and patient characteristics and NSES, that these factors mostly affect the African Americans versus NL Whites disparity, and that comorbidities and other contextual factors, while possibly related to mortality, play a small role in explaining mortality disparity. To our knowledge, our current analysis is the first large population-based study to examine four domains of risk factors, including contextual factors, modifiable lifestyle factors (body size, physical activity), and comorbidities that appear to affect breast cancer outcome in NL Whites but that have not been studied in large numbers of women from other racial/ethnic groups.

An important strength of our analysis is that we have investigated the role of contextual factors, physical activity, body size, and comorbidity on racial/ethnic disparity in a harmonized cohort of breast cancer cases and within the context of a single multivariable baseline model. The contextual factors we studied included many additional parameters such as variables of crowding, urbanicity, street connectivity, number of businesses, restaurants, and parks (29, 32). Our study also had the limitations that the domain-specific variables were not all available in all six studies in the CBCSC. In addition, these variables were limited to exposures that occurred before diagnosis, and they may not capture complete exposure (e.g., not all possible comorbidities are included, physical activity measures did not include occupational or housework-related activity).

There are other variables that could contribute to the racial/ethnic disparity in breast cancer–specific mortality that we were not able to consider. Although we had some data on treatment, data on the appropriateness of chemotherapy, radiotherapy, or surgical treatments administered, or patient compliance with prescribed treatments and clinical follow-up, were not available. Relevant to this, within an equal-access health system with standardized practice guidelines, racial/ethnic-specific differences in the initiation of adjuvant hormonal therapy (39) and in the use of chemotherapy (11) have been reported. Results on compliance from single institution studies are less clear. Sharma and colleagues (40) evaluated compliance with radiotherapy in a large institutional series of White and African-American patients who had undergone breast conservation surgery and found no difference in the rate of compliance. In a much smaller series, Bhatta and colleagues (41) found no difference in compliance to adjuvant hormone therapy between White and Black patients with estrogen receptor–positive breast cancer. Our inclusion of factors related to NSES, which may indirectly reflect access to care, possibly captured part of any therapy compliance effect.

Genetic differences between racial/ethnic groups also have been considered. Bach and colleagues (17) synthesized a very large cohort of White and Black breast cancer patients and concluded from a meta-analysis that in comparably and appropriately treated patients, there was little difference in cancer-specific mortality between races. They concluded, therefore, that differences in cancer biology are unlikely to explain any disparity in outcome between Blacks and Whites. In contrast, Albain and colleagues (42) analyzed outcome in African Americans versus other races in more than 19,000 patients treated on Southwest Oncology Group clinical trials and concluded that, for sex-specific cancers, African Americans had worse outcome despite uniform staging, treatment, and follow-up. Genetic differences potentially can explain other disparities, such as the better outcome in Asian Americans. For example, Shimizu and colleagues (43) have reported a much lower frequency in Asians than in Caucasians of genetic polymorphisms in CYP2D6 associated with poor drug metabolism, and these have been associated with lower endoxifen levels (44) and with poorer outcome (45) in breast cancer patients treated with tamoxifen.

Dietary factors may also contribute to the observed racial/ethnic disparity in outcome. For example, Nechuta and colleagues (46) have shown that soy intake was inversely associated with breast cancer recurrence in both U.S. and Chinese women.

Bias in assessing breast cancer–specific mortality also may be a factor. For example, Gomez and colleagues (personal communication) investigated completeness of follow-up (alive and more than 2 years from the reference date December 31, 2012) in more than 500,000 cases in the CCR during the years 2000 to 2009. They found higher rates of incomplete follow-up in Asian Americans (9%) and Hispanics (6%) compared with non-Hispanic Whites (2%). This admits the possibility that the observed better outcome in these patient groups could in part be due to bias resulting from informative censoring. In our cohort, using the same definition with our cut-off date of December 31, 2012, incomplete follow-up rates in non-Hispanic whites, Latinas, Asians, and African Americans were 2%, 4%, 5.2%, and 4.3%, respectively. Hence, the overall rate of incomplete follow-up is small in our cohort, as is the difference in these rates between racial/ethnic categories. Although this does not eliminate the theoretical possibility that differences in informative censoring may be causing bias in our results, these generally low rates of incomplete follow-up suggest that whatever bias would be very small or negligible.

In conclusion, in our analysis, the racial/ethnic disparity in breast cancer–specific mortality remains after accounting for clinical, patient, contextual, and modifiable lifestyle factors. Although some of these factors may have prevalence that is different across racial/ethnic groups, and may also be associated with breast cancer–specific mortality, this association is not strong enough to explain all of the observed racial/ethnic disparity in outcome. Other unmeasured factors, such as yet unknown underlying genetic differences, unappreciated biases in measuring racial/ethnic differences in outcome, remain to be investigated to fully explain the observed mortality differences between NL Whites, African Americans Latinas, and Asian Americans.

No potential conflicts of interest were disclosed.

The ideas and opinions expressed herein are those of the author(s) and endorsement by the State of California, Department of Public Health, the National Cancer Institute, and the Centers for Disease Control and Prevention or their contractors and subcontractors is not intended nor should be inferred.

Conception and design: R. Sposto, T.H.M. Keegan, C. Vigen, E.M. John, A.W. Kurian, K.R. Monroe, S.L. Gomez, A.H. Wu

Development of methodology: R. Sposto, C. Vigen, K.R. Monroe, A.H. Wu

Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): C. Vigen, M.L. Kwan, L. Bernstein, E.M. John, J. Koo, B.J. Caan, Y. Lu, K.R. Monroe, S. Shariff-Marco, A.H. Wu

Analysis and interpretation of data (e.g., statistical analysis, biostatistics, computational analysis): R. Sposto, C. Vigen, M.L. Kwan, I. Cheng, J. Yang, J. Koo, B.J. Caan, Y. Lu, K.R. Monroe

Writing, review, and/or revision of the manuscript: R. Sposto, T.H.M. Keegan, C. Vigen, M.L. Kwan, L. Bernstein, E.M. John, I. Cheng, J. Koo, A.W. Kurian, B.J. Caan, Y. Lu, K.R. Monroe, S. Shariff-Marco, S.L. Gomez, A.H. Wu

Administrative, technical, or material support (i.e., reporting or organizing data, constructing databases): C. Vigen, L. Bernstein, J. Yang, S.L. Gomez

The authors gratefully acknowledge Dr. Brian Henderson's contribution to this research.

This work was supported by grants 16ZB-8001 (to A.H. Wu, R. Sposto, and C. Vigen), 16ZB-8002 (to S.L. Gomez, E.M. John, T.H.M. Keegan, A.W. Kurian, and S. Shariff-Marco), 16ZB-8003 (to L. Bernstein and Y. Lu), 16ZB-8004 (to M.L. Kwan and B.J. Caan), and 16ZB-8005 (to K. Monroe) from the California Breast Cancer Research Program. The AABCS was supported by the California Breast Research Program (CBCRP) grants 1RB-0287, 3PB-0120, and 5PB-0018 (to A.H. Wu). The SFBCS was supported by National Cancer Institute grants R01 CA63446 and R01 CA77305, by the U.S. Department of Defense (DOD) grant DAMD17-96-1-6071, and by the CBCRP grants 4JB-1106 and 7PB-0068 (to E.M. John). The Women's CARE Study was funded by the National Institute of Child Health and Human Development (NICHD), through a contract with USC (N01-HD-3-3175), and the CTS was funded by the California Breast Cancer Act of 1993; National Cancer Institute grants (R01 CA77398 and K05 CA136967; to L. Bernstein); and the California Breast Cancer Research Fund (contract 97-10500; to L. Bernstein and Y. Lu). The Multiethnic Cohort Study was supported by National Cancer Institute grants R01 CA54281, R37CA54281, and UM1 CA164973. The LACE Study is supported by National Cancer Institute grant R01 CA129059 (to B.J. Caan and M.L. Kwan). The collection of cancer incidence data used in this study was supported by the California Department of Health Services as part of the statewide cancer reporting program mandated by California Health and Safety Code Section 103885; the National Cancer Institute's Surveillance, Epidemiology, and End Results Program under contract HHSN261201000140C awarded to the Cancer Prevention Institute of California, contract HHSN261201000035C awarded to the University of Southern California, and contract HHSN261201000034C awarded to the Public Health Institute; and the Centers for Disease Control and Prevention's National Program of Cancer Registries, under agreement #1U58 DP000807-01 awarded to the Public Health Institute.