Geographic and Racial Disparities in Breast Cancer–Related Outcomes in Georgia

Abstract

Objective

To measure the effects of race/ethnicity, area measures of socioeconomic status (SES) and geographic residency status, and health care supply (HCS) characteristics on breast cancer (BC)-related outcomes.

Data Sources/Study Setting

Female patients in Georgia diagnosed with BC in the years 2000–2009.

Study Design

Multilevel regression analysis with adjustment for variables at the county, census tract (CT), and individual level. The county represents the spatial unit of analysis for HCS. SES and geographic residency status were grouped at the CT level.

Principal Findings

Even after controlling for area-level characteristics, racial and ethnic minority women suffered an unequal BC burden. Despite inferior outcomes for disease stage and receipt of treatment, Hispanics had a marginally significant decreased risk of death compared with non-Hispanics. Higher CT poverty was associated with worse BC-related outcomes. Residing in small, isolated rural areas increased the odds of receiving surgery, decreased the odds of receiving radiotherapy, and decreased the risk of death. A higher per-capita availability of BC care physicians was significantly associated with decreased risk of death.

Conclusions

Race/ethnicity and area-level measures of SES, geographic residency status, and HCS contribute to disparities in BC-related outcomes.

Despite continuous investment by public and private agencies to improve breast cancer (BC)-related outcomes, particularly among undeserved women, BC is the second-ranking cause of cancer death in women (Siegel et al. 2011). The encouraging decline in BC-related mortality since 1990 (American Cancer Society [ACS] 2011) has not been observed equally among all segments of the U.S. population. The literature associates disparities in access to cancer care and cancer-related outcomes with such nonmedical factors as race/ethnicity (Agency for Healthcare Research and Quality [AHRQ] 2009; U.S. Cancer Statistics Working Group 2010; Jacobellis and Cutter 2002), geographic residency status (rural, suburban, or urban; Liff, Chow, and Greenberg 1991; Higginbotham, Moulder, and Currier 2001; Coughlin et al. 2002; McLafferty and Wang 2009), area-level socioeconomic status (SES; Bradley, Given, and Roberts 2002; Barry and Breen 2005; Abe, Martin, and Roche 2006; Buchholz et al. 2006; Foley et al. 2007; Barrett et al. 2008; Byers et al. 2008; Lamont et al. 2012), and health care supply (HCS) of generalist physicians, specialist physicians, including surgeons, and hospital beds (Ferrante et al. 2000; Baicker et al. 2004; Baicker, Chandra, and Skinner 2005; Starfield, Shi, and Macinko 2005b; Starfield et al. 2005a; Klabunde et al. 2009; Lamont et al. 2012).

Two overarching components contribute to overall health and health care disparities. The first is unequal treatment within the health care setting or by a provider, possibly due to provider bias, poor communication, failure to consider patient preferences, health insurance status, or patient-level SES. The second component is unequal treatment due to the person's place of residence, based on such factors like area-level SES, HCS, and geographic region (Baicker, Chandra, and Skinner 2005). However, research explaining how these individual and contextual components interact to determine disparities in cancer outcomes and the role of HCS as a mediator of SES variation in patient outcomes is lacking. Some studies have suggested that in some areas of the United States, a substantial component of observed racial disparities in care can be explained by geographic variation in health care quality (Baicker, Chandra, and Skinner 2005).

Using data from the Georgia Comprehensive Cancer Registry (GCCR), US Census Bureau, and the Health Resources and Services Administration's Area Resource File (ARF), this study aimed to quantify prognostic factors for BC-related outcomes, including race/ethnicity, an area measure of SES, geographic residency status, and HCS. We hypothesized that racial and ethnic disparities in use of BC-directed care and outcomes were, in part, explained by these contextual characteristics. The work builds on our previous study in Georgia suggested that African American (AA) race and rural residence independently affected BC-related outcomes (Markossian and Hines 2012). However, we looked at only 15 counties from the Atlanta and Rural Georgia Cancer Registries that report to the Surveillance, Epidemiology, and End Results (SEER) Program of the National Cancer Institute, and Hispanic women were too few to include. The small number of counties also made it impossible to assess the significance of area-level SES on BC-related outcomes while controlling for geographic residency status.

The new study includes all cases of women diagnosed with BC in Georgia between 2000 and 2009 who met the study's eligibility criteria and assesses area measures of SES and geographic residence at the census tract (CT) level. Results call attention to the specific effects of both race/ethnicity and area-level characteristics on disparate cancer-related outcomes. They can inform the design and development of programs and policies that target the populations most in need who will receive the greatest benefit.

Materials and Methods

Data Acquisition and Cohort Definition

Study participants were recruited from the GCCR, which collects all demographic, tumor, and treatment-related information about the patient at the time of reporting. Georgia legally mandated cancer reporting in 1995, and standardized data collection and quality-control procedures have been in place since then (Georgia Division of Public Health [GDPH] 2006). Registry staff and funding agency contractors periodically conduct re-abstracting and completeness studies on the data (Clarkson 2007).

All female patients diagnosed with invasive BC from 2000 to 2009 were included in this study (N = 55,140). The termination date was December 31, 2009. Men were excluded, and other cases were excluded if they were diagnosed with in situ cancer; if the BC diagnosis was not the first lifetime diagnosis of cancer (excluding nonmelanoma skin cancer); if they had more than one primary tumor (n = 11,913 excluded); and if their CT information was missing (n = 111). The final study population was N = 43,116. The research was approved by the Institutional Review Boards of the Georgia Department of Public Health and Georgia Southern University.

Selection of Study Variables

Study variables were selected based on the constructs for which data were available from Andersen's (1995) revised behavioral model for access to care and health status outcomes. According to this model, various influences at the level of the individual and community resources affect patients' use of health services and subsequent health status.

Patient-Level Variables

Patient-level variables were obtained from the GCCR and included race (white/AA/Other), ethnicity (Hispanic: yes/no), age at diagnosis, date of diagnosis, tumor-related information, type of first-course treatment received, last date of follow-up, and vital status at last follow-up. The majority of “Other” races were Asians, including Koreans, Chinese, and Filipinos. For tumor stage, the SEER summary (2000–2003) and collaborative (2004–2009) staging classification system (Young et al. 2001) used by the GCCR defined tumors as localized, regional, distant, and unstaged. For tumor grade, well and moderately differentiated tumors were combined as lowgrade, and poorly and undifferentiated tumors were combined as highgrade. Tumors were also classified as to estrogen receptor (ER) status (positive/negative) and progesterone receptor (PR) status (positive/negative). For tumors, missing information and unknown status were combined.

Area-Level Variables

The two geographic units under study are CTs and counties. CTs have an average population of 4,000 and are nested within county boundaries (Krieger et al. 2002). For all cancer patients, the GCCR records the CT corresponding to their residential address. We merged the CT of residence for each patient with area-level measures of SES and geographic residency status.

The percentage of the population with a household income below the federal poverty level (FPL) is the most important indicator of area-level SES and correlates well with other SES measures (Singh et al. 2004; Krieger 2005). To characterize CT-level SES, we obtained that percentage from the US Census 2000 Summary File 3. As described previously (Hausauer et al. 2009), subjects were categorized as residing in low-(≤9.9 percent of households below FPL), middle-(10.0–19.9 percent of households below FPL), or high-poverty CTs (20.0+ percent of households below FPL).

Using the US Department of Agriculture's Rural-Urban Commuting Area (RUCA) primary and secondary codes, categorization B (Rural Health Research Center (RHRC) 2012), we classified subjects as residing in urban areas, large rural cities or towns, or small and isolated rural towns.

To characterize HCS at the county level, we obtained ARF data about hospitals in the American Hospital Association (AHA) for the year 2000 and American Medical Association physicians for the year 2005. We transformed the HCS variables into counts per 1,000 county residents. We selected health care variables that measure the availability of services required for providing the BC care continuum and that might be associated with study outcomes (Lamont et al. 2012), including the following: (1) for stage of diagnosis analysis, the number of BC-screening physicians per capita; (2) for receipt of treatment, the number of BC-treating physicians and community hospital beds per capita; and (3) for survival analysis, the number of BC care physicians and community hospital beds per capita. BC-screening physicians represents the number of physicians who could perform BC screening and assist specialty physicians in posttreatment follow-up: physicians of general practice, general family medicine and its subspecialties, general internal medicine, general obstetrics and gynecology, general preventive medicine, and public health. BC-treating physicians represented all general surgeons and radiation oncologists. The ARF groups oncologists with other internal medicine subspecialists, and they could not be subtracted for inclusion as cancer-treating physicians. BC care physicians are the sum of BC screening and treating physicians. Community hospitals defined by the AHA as “all non–federal short–term general and other special hospitals, excluding hospital units of institutions” (U.S. Department of Health and Human Services [DHHS] 2011) were used to calculate hospital beds. Study participants were assigned the value for the HCS characteristics corresponding to their county of residence.

BC-Related Outcomes

Outcomes of interest included the odds of being diagnosed with unstaged and late-stage disease, the odds of receiving first-course treatment, and overall survival following a diagnosis of BC. Localized tumors were categorized as SEER summary early stage. Distant and regional tumors were categorized as late stage. Tumor stage was also dichotomized as unstaged or staged. Unstaged tumors were excluded only from the analysis of late-stage disease as the outcome. The registry records the date on which first-course treatment (surgery, radiation, or chemotherapy) began at any facility. For each treatment type, information was dichotomized as yes/no with a separate category for missing/unknown. Each treatment type was analyzed separately. Survival time was calculated as the number of months from the date of initial BC diagnosis to either the date of death due to any cause, last date of follow-up, or termination of the study.

Statistical Analysis

To assess the association among patient characteristics, area-level attributes, clinical characteristics, and BC-related outcomes, we compared frequencies and measures of central tendency using the chi-squared test of significance and the t-test. Multilevel mixed-effect logistic regression models were constructed for each dichotomous outcome, where patients were nested within counties, and county was treated as a random effect. Regardless of the level of significance for disease stage and treatment modality in the bivariate and multivariate analyses, all models controlled for these variables.

In this study, subjects originated from 1,602 CTs; 25 percent of the CTs had fewer than 7 subjects; 50 percent, fewer than 16 subjects; and 75 percent, fewer than 28 subjects. Due to the small number of subjects in similar CTs, hierarchical models controlling for random CT-level effects in addition to random county-level effects were computationally demanding and, for most outcomes for which convergence was achieved, produced results were similar to those of models that controlled for county random effects only. Therefore, the models kept only random county effects.

Statistical interaction between area-level variables and patient demographic characteristics were tested after obtaining the final model. As the correlation coefficient between the HCS variables (per-capita rate of physicians and hospital beds) were highly significant, separate models were run for each. Statistical interactions between the HCS variables were also tested in models containing both.

For the analysis of survival time, the outcome was all-cause 5-year survival. Therefore, study participants were censored at 5 years if they were followed up to that point or longer. The Cox proportional hazards model with shared frailty was used to obtain adjusted-hazard rate ratios (aHRs) with 95 percent CIs for the relative risk of death. Shared gamma frailty was used to model the within-county correlation. The proportional hazards (PH) assumption was evaluated for area-level variables and patient characteristics by testing their interaction with time and for the nonzero slope of the scaled Schoenfeld residuals on functions of time. One variable, hospital beds, did not meet the PH assumption; an interaction term of the variable with a function of study time was added to the final model. A model-building procedure similar to the one described for the multilevel logistic regression was conducted. All analyses were performed using STATA/MP statistical software version 11.2 (StataCorp, College Station, TX, USA).

To analyze the odds of receiving treatment and survival, we looked only at subjects who were diagnosed from 2000 to 2007 based on the acknowledged limitation that subjects more recently diagnosed were less likely to have complete treatment information (R. Bayakly, personal communication).

Results

Table 1 shows patient demographics, tumor characteristics, and receipt of first-course treatment. Most patients were white (72.6 percent), non-Hispanic (98 percent), and between the ages of 45 and 64 years (49.9 percent). Most resided in low-poverty (52.5 percent) and urban (77.6 percent) CTs. Breast cancer tumors were most commonly localized (59.0 percent), well or moderately differentiated (53.4 percent), and hormone-receptor positive (66.0 percent). Most patients received surgery; 47.3 percent received breast-conserving surgery, and 42.8 percent mastectomy. Nearly half of all patients received chemotherapy (45.6 percent) and/or radiation (46.8 percent). Most were alive 5 years after the date of diagnosis (81.3 percent).

Table 1.

Characteristics of Study Population, Years 2000–2009 (N = 43,116)

Variable	No.	%
Age at diagnosis, years
14–44	6,746	15.6
45–54	10,800	25.0
55–64	10,759	24.9
65–74	7,785	18.1
≥75	7,026	16.3
Race
White	31,296	72.6
African American	11,044	25.6
Other	678	1.6
Unknown	98	0.2
Ethnicity
Non-Hispanic	42,265	98.0
Hispanic	851	2.0
Tumor grade
Well differentiated	7,643	17.7
Moderately differentiated	15,384	35.7
Poorly differentiated	14,942	34.7
Undifferentiated	484	1.1
Unknown	4,663	10.8
Stage
Localized	25,422	59.0
Regional	14,228	33.0
Distant	2,219	5.1
Unstaged	1,247	2.9
ER/PR status
ER+/PR+	23,180	53.8
ER+ or PR+	5,249	12.2
ER−/PR−	8,849	20.5
Unknown	5,838	13.5
Surgery
No	3,619	8.4
BCS	20,417	47.3
Mastectomy	18,442	42.8
Surgery, NOS	75	0.2
Unknown	563	1.3
Chemotherapy
No	21,793	50.5
Yes	19,674	45.6
Unknown	1,649	3.8
Radiation
No	21,389	49.6
Yes	20,168	46.8
Unknown	1,559	3.6
Five-year survival
Dead	8,065	18.7
Alive	35,051	81.3
Census tract poverty level
Low poverty	22,644	52.5
Middle poverty	13,408	31.1
High poverty	7,051	16.4
Census tract RUCA categorization
Urban	33,459	77.6
Large rural city or town	4,644	10.8
Small and isolated rural	5,013	11.6

BCS, breast-conserving surgery; ER, estrogen receptor; NOS, not otherwise specified; PR, progesterone receptor; RUCA, rural–urban commuting area.

In 2005, 19 counties in Georgia did not have any BC care physicians as defined by this study. In 2000, 48 counties did not have any community hospitals. Table 2 shows the HCS characteristics of counties in Georgia. The average number of BC screening, treating, and care physicians were 0.26, 0.06, and 0.32 per 1,000 capita, respectively. There were 3.01 hospital beds per 1,000 capita.

Table 2.

Health Care Supply Characteristics of Counties in Georgia (n = 159)

Health Services	Counts per 1,000 County Residents
	Range (min–max)	Range (min–max)
BC-screening physicians*	0.26 (0.22)	0–1.15
BC-treating physicians*	0.06 (0.07)	0–0.33
BC care physicians*	0.32 (0.28)	0–1.49
Hospital beds†	3.01 (3.53)	0–18.58

^*American Medical Association data, 2005.

^†American Hospital Association data, 2000.

BS, breast cancer; BC care physicians is the sum of BC-screening and treating physicians.

Table 3 shows the adjusted odds ratios (aORs) for having unstaged and late-stage disease. Hispanics were 42 percent (aOR = 1.42, 95 percent CI = 0.93–2.17) more likely than all other patients to have unstaged cancer, although this association failed to reach statistical significance due to the small number of Hispanics in our study. They were 26 percent (aOR = 1.26, 95 percent CI = 1.10–1.45) more likely to be diagnosed with late-stage cancer, which is statistically significant. AAs were 28 percent (aOR = 1.28, 95 percent CI = 1.11–1.49) more likely to have unstaged cancer and 51 percent (aOR = 1.51, 95 percent CI = 1.44–1.59) more likely to be diagnosed with late-stage cancer than whites; other races were 22 percent (aOR = 1.22, 95 percent CI = 1.04–1.43) more likely to be diagnosed with late-stage cancer than whites. Women residing in middle-poverty CTs and high-poverty CTs were 12 percent (aOR = 1.12, 95 percent CI = 1.07–1.18) and 21 percent (aOR = 1.21, 95 percent CI = 1.13–1.30), respectively, more likely to be diagnosed with late-stage cancer than women residing in low-poverty CTs. Women residing in small and isolated rural areas were 27 percent (aOR = 1.27, 95 percent CI = 1.02–1.58) more likely to have unstaged cancer than urban women.

Table 3.

Adjusted^† Associations with Unstaged and Late-Stage Breast Cancer Diagnosis

	Unstaged Cancer	Late-stage‡Cancer
	Adjusted OR (95% CI) N = 43,116	Adjusted OR (95% CI) N = 41,869§
	Events = 1,247 (2.9%)	Events = 16,447 (39.3%)
Individual-level variables
Race
White	Ref	Ref
African American	1.28** (1.11–1.49)	1.51** (1.44–1.59)
Other	1.13 (0.64–1.98)	1.22* (1.04–1.43)
Ethnicity
Non-Hispanic	Ref	Ref
Hispanic	1.42 (0.93–2.17)	1.26** (1.10–1.45)
Area-level variables
CT poverty level
Low poverty	Ref	Ref
Middle poverty	1.03 (0.88–1.20)	1.12** (1.07–1.18)
High poverty	0.93 (0.76–1.14)	1.21** (1.13–1.30)
CT RUCA categorization
Urban	Ref	Ref
Large rural	1.03 (0.82–1.31)	0.95 (0.88–1.03)
Small and isolated rural	1.27* (1.02–1.58)	1.00 (0.92–1.08)
BC-screening physicians	0.72 (0.51–1.03)	0.93 (0.84–1.03)

^*p < .05

^**p < .01.

^†Adjusted for the variables listed and age at diagnosis, year of diagnosis, tumor grade, and random county effect.

^‡Late-stage cancer is the combination of regional and distant tumors.

^§Unstaged cancers are excluded from this analysis.

BC, breast cancer; CI, confidence interval; CT, census tract; OR, odds ratio; Ref, referent group; RUCA, rural–urban commuting area.

Table 4 shows the results of the analysis of first-course treatment. AAs were less likely to receive surgery (aOR = 0.62, 95 percent CI = 0.55–0.70) and radiotherapy (aOR = 0.91, 95 percent CI = 0.85–0.97) than whites. Women who self-identified as “other race” were less likely to receive surgery (aOR = 0.63, 95 percent CI = 0.43–0.91) than whites. Hispanics were 35 percent (aOR = 0.65, 95 percent CI = 0.47–0.89) less likely to receive surgery and 22 percent (aOR = 0.78, 95 percent CI = 0.66–0.94) less likely to receive radiotherapy than all other groups. Women residing in high-poverty CTs were 21 percent (aOR = 0.79, 95 percent CI = 0.68–0.93) less likely to receive surgery and 17 percent (aOR = 0.83, 95 percent CI = 0.75–0.91) less likely to receive radiotherapy than women in low-poverty CTs. Women in middle-poverty CTs were 11 percent (aOR = 0.89, 95 percent CI = 0.83–0.95) less likely to receive radiotherapy than low-poverty CT residents. Compared with women residing in urban locations, women residing in small and isolated rural areas were 30 percent (aOR = 1.30, 95 percent CI = 1.06–1.60) more likely to receive surgery but 17 percent (aOR = 0.83, 95 percent CI = 0.71–0.96) less likely to receive radiotherapy.

Table 4.

Adjusted^† Associations with Odds of Receiving Surgery, Radiotherapy, and Chemotherapy for Breast Cancer Patients (N^‡ = 33,347)

	Surgery	Radiotherapy	Chemotherapy
	Adjusted OR (95% CI)	Adjusted OR (95% CI)	Adjusted OR (95% CI)
	Events‡ = 30,404 (92.4%)	Events‡ = 15,404 (47.9%)	Events‡ = 15,099 (47.2%)
Individual-level variables
Race
White	Ref	Ref	Ref
African American	0.62** (0.55–0.70)	0.91** (0.85–0.97)	1.01 (0.94–1.09)
Other	0.63* (0.43–0.91)	0.83 (0.68–1.02)	1.06 (0.84–1.33)
Ethnicity
Non-Hispanic	Ref	Ref	Ref
Hispanic	0.65** (0.47–0.89)	0.78** (0.66–0.94)	1.15 (0.93–1.4)
Area-level variables
CT poverty level
Low poverty	Ref	Ref	Ref
Middle poverty	0.88 (0.78–1.00)	0.89** (0.83–0.95)	0.97 (0.90–1.05)
High poverty	0.79** (0.68–0.93)	0.83** (0.75–0.91)	0.91 (0.82–1.02)
CT RUCA categorization
Urban	Ref	Ref	Ref
Large rural	1.23 (0.99–1.53)	0.97 (0.82–1.15)	1.01 (0.87–1.18)
Small and isolated rural	1.30* (1.06–1.60)	0.83* (0.71–0.96)	0.99 (0.86–1.14)
BC-treating physicians	0.74 (0.27–2.06)	1.17 (0.40–3.41)	0.70 (0.32–1.54)

^*p < .05

^**p < .01.

^†Adjusted for the variables listed and age at diagnosis, year of diagnosis, tumor grade, hormone receptor status, tumor stage, other treatment, and random county effect.

^‡Patients diagnosed 2008–2009 are excluded, as well as patients with respective missing treatment information.

BC, breast cancer; CI, confidence interval; CT, census tract; OR, odds ratio; Ref, referent group; RUCA, rural–urban commuting area.

Table 5 shows the results of the survival analysis. Compared to all other groups, AA race was associated with a 27 percent increased risk of death (aHR = 1.27, 95 percent CI = 1.19–1.34) and Hispanic ethnicity with a 20 percent decreased risk of death (aHR = 0.80, 95 percent CI = 0.63–1.00), which was borderline statistically significant. Compared to women residing in low-poverty CTs, women residing in middle-and high-poverty CTs were 21 percent (aHR = 1.21, 95 percent CI = 1.13–1.29) and 30 percent (aHR = 1.30, 95 percent CI = 1.20–1.40), respectively, more likely to die. Compared to urban residents, residents of small and isolated rural areas were 11 percent (aHR = 0.89, 95 percent CI = 0.81–0.98) less likely to die. A one-unit increase in per-capita rate of BC care physicians was associated with a 13 percent decrease in risk of death (aHR = 0.87, 95 percent CI = 0.79–0.97). As such, for a county of 25,000 residents, doubling the number of BC care physicians from 10 to 20 physicians corresponding to a 0.4 unit increase in per-capita rate would be associated with a 5 percent decreased risk of death (aHR = 0.95, 95 percent CI = 0.91–0.99). For a county of 50,000 residents, doubling the number of BC care physicians from 10 to 20 physicians corresponding to a 0.2 unit increase in per-capita rate would be associated with a 3 percent decreased risk of death (aHR = 0.97, 95 percent CI = 0.95–0.99).

Table 5.

Adjusted^† Associations with All-Cause Death (5-year) for Breast Cancer Patients (N^‡ = 33,347)

	Adjusted HR (95% CI)
	Events‡ = 6,902 (20.7%)
Individual-level variables
Race
White	Ref
African American	1.27** (1.19–1.34)
Other	0.79 (0.60–1.04)
Ethnicity
Non-Hispanic	Ref
Hispanic	0.80 (0.63–1.00)
Area-level variables
CT poverty level
Low poverty	Ref
Middle poverty	1.21** (1.13–1.29)
High poverty	1.30** (1.20–1.40)
CT RUCA categorization
Urban	Ref
Large rural	0.95 (0.87–1.04)
Small and isolated rural	0.89* (0.81–0.98)
BC care physicians	0.87** (0.79–0.97)

^*p < .05

^**p < .01.

^†Adjusted for the variables listed and age at diagnosis, year of diagnosis, tumor grade, hormone receptor status, tumor stage, treatment (surgery/chemotherapy/radiotherapy), and random county effect.

^‡Patients diagnosed 2008–2009 are excluded.

BC, breast cancer; CI, confidence interval; CT, census tract; HR, hazard ratio; Ref, referent group; RUCA, rural–urban commuting area.

Per-capita availability of hospital beds was not significantly associated with any of the study's outcomes (results from this analysis are not presented). Also, the interaction terms between the per-capita rate of physicians and hospital beds were not statistically significant in the models that included both variables. For all the outcomes we evaluated in this study, the interaction terms between CT poverty levels and rural/urban designations did not yield significant findings. For receipt of surgery and all-cause survival, we found slight effect modification for the effect of AA race when stratified according to geography. AAs living in large rural and urban areas were 43 percent (aOR = 0.57, 95 percent CI = 0.41–0.79) and 32 percent (aOR = 0.68, 95 percent CI = 0.59–0.78), respectively, less likely to receive surgery than whites. However, AAs living in isolated rural areas were 57 percent (aOR = 0.43, 95 percent CI = 0.32–0.59) less likely to receive surgery than whites. For the survival outcome, AAs living in large rural and urban areas were 24 percent (aHR = 1.24, large rural 95 percent CI = 1.05–1.46 and urban 95 percent CI = 1.16–1.33) more likely to die than whites. However, AAs living in isolated rural areas were 45 percent (aHR = 1.45, 95 percent CI = 1.25–1.69) more likely to die than whites. There was no evidence of effect modification by geography for the effect of HCS; results from stratified analysis by geographic residency status yielded similar findings regarding the associations between HCS and BC-related outcomes.

Discussion

Results from this study that examined the effects of race/ethnicity and area-level characteristics on BC-related outcomes in Georgia contribute in several ways to advancing the science of cancer disparities. Our findings quantify prognostic factors for BC burden that persist despite years of research and efforts to alleviate the problem. Notably, the multilevel data analysis approach was used to adjust for variables at the county, CT, and individual level. The county was used as the spatial unit of analysis for HCS. The CT was used as the spatial unit of analysis for SES and geographic residency status. Our findings suggest that individual and contextual factors have similar effects on disparities in BC-related outcomes.

Unlike previous studies that found in some areas of the United States, many of the observed racial disparities in care were due to geographic variations in health care quality (Baicker, Chandra, and Skinner 2005), our results in Georgia suggest that even after controlling for area-level SES and geographic residency status, AA women had worse BC-related outcomes than whites. Also, compared to non-Hispanics, Hispanic women had increased odds of having unstaged and late-stage cancer, and decreased odds of receiving surgery and radiotherapy. However, interestingly, Hispanic women had lower mortality. The literature has proposed several suggestions to the paradox of lower morality for Hispanics given their lower SES and reduced access to health care. The potential reasons for this paradox include the following: cultural factors, family dynamics, and social support (LeClere, Rogers, and Peters 1997; Palloni and Arias 2004); data reporting and data collection problems (Elo and Preston 1997; Rosenberg et al. 1999; Elo et al. 2004); and healthy in-migrant and unhealthy out-migrant selection effects (Abraido-Lanza et al. 1999; Franzini, Ribble, and Keddie 2001; Jasso et al. 2004; Palloni and Arias 2004; Turra and Elo 2008).

In this study, each treatment type was analyzed separately because determining the optimal proportion of patients who should receive a particular BC treatment modality according to clinical indicators using the GCCR data is not possible. If patients with different characteristics or living in different neighborhoods were presented with similar information and opportunities, there is little reason to believe that they would make significantly different treatment choices or had disparate survival outcomes. Earlier studies have shown that patients who live in isolated areas face major barriers to access care and, therefore, these patients utilize health care less frequently and inadequately (Ricketts and Savitz 1994; Ricketts 1999). This might explain why, in our study, rural patients received more surgery, which probably requires fewer follow-up visits, and less adjuvant therapy, which probably requires more frequent visits to the provider. These findings are similar to the results from our previous study in Georgia in which we found rural women received more mastectomy and less radiotherapy (Markossian and Hines 2012).

The evidence is inconclusive regarding the association between rural residency and BC-specific or overall cancer mortality. Some studies have found no significant association between rural residency and BC mortality (Higginbotham, Moulder, and Currier 2001; Markossian and Hines 2012). The results of a study of Appalachian women suggested smaller BC death rates but larger overall cancer death rates for rural women compared to all Appalachian women and women in the United States (Centers for Disease Control and Prevention (CDC) 2002). Our results regarding the association between area-level poverty status and BC outcomes were similar to many studies that found residents of low SES areas had more advanced disease stage, received less aggressive treatment, and had higher risk of all-cause mortality compared to residents of high SES areas (Bradley, Given, and Roberts 2002; Barry and Breen 2005; Abe, Martin, and Roche 2006; Buchholz et al. 2006; Foley et al. 2007; Barrett et al. 2008; Byers et al. 2008; Lamont et al. 2012).

The results of research using diverse sources of data and methods have reported different findings regarding the association between HCS and cancer-related outcomes. According to some studies, the differential supply of specialists and primary care physicians at the county level was a possible cause of geographic variation for late-stage cancer risk and mortality (Ferrante et al. 2000; Campbell et al. 2003; Shi et al. 2005; Starfield, Shi, and Macinko 2005b; Starfield et al. 2005a). Mays and Smith (2011) found that increases in counties' public health spending were associated with decreases in cancer mortality rates (Mays and Smith 2009). However, a recent study (Lamont et al. 2012) restricted to urban areas suggested that the association between social area deprivation and individual residents' poorer cancer outcomes was not mediated by inefficient supply of cancer care. Areas with high levels of social deprivation did not have decreased access to cancer care. A possible explanation to their findings was that low-income and AA populations generally resided next to inner-city academic medical centers and faced other barriers to accessing adequate cancer care, which they did not evaluate, that resulted in unfavorable cancer outcomes. In contrast, the whiter and higher SES populations tended to reside in fringe areas that are more suburban (Kahn et al. 1994; Lamont et al. 2012).

Among all the different associations we tested between the HCS variables and BC-related outcomes, only the association between the per-capita availability of BC care physicians and the risk of death was statistically significant. As such, higher per-capita availability of BC care physicians was associated with lower risk of death. A possible explanation for the lack of strong associations between per-capita availability of BC-treating physicians, hospital beds, and BC-related outcomes might be because people travel outside their county of residence to receive health care. Our results also suggested an association between per-capita availability of BC-screening physicians and decreased odds of having unstaged disease, although it failed to reach statistical significance. One possible explanation to our findings is more vigilant diagnosis in areas with higher HCS.

Study subjects originated from the GCCR, and hence our research was restricted to data elements collected by the registry. As a result, the most significant limitation was limiting the analysis to all-cause mortality instead of using cause-specific survival analysis because the GCCR does not collect information on cause of death, which makes it difficult to rule out deaths due to other confounding individual and contextual factors. Also, the GCCR does not collect information about patient comorbidities which are normally associated with receipt of BC-directed treatment and outcomes. To reduce biases, all regression models controlled for age at diagnosis and excluded all subjects previously diagnosed with other cancers. We also included adjustments for random area of residence effects. Regarding treatment-related information, the GCCR collects data on first-course treatment from hospital inpatient and outpatient settings, as well as other sources, including physician's offices, laboratories, and nursing homes. However, the receipt of subsequent or delayed treatment is not collected by the registry and, therefore, characterizing all dimensions of treatment received was not possible. Also, underreporting of treatment modality—mainly chemotherapy, which is largely administered in outpatient settings—might be a study limitation.

Another source of inaccuracy to our analysis was using data for area-level attributes and HCS which were reported over a single year for a 10-year period of analysis. As individual-level SES is not routinely collected by cancer registries, differentiating between individual-and CT-level SES contributions to BC disparities was not possible. Using a more precise measure for HCS characteristics (e.g., CT) or geocoding to account for the effect of HCS of neighboring counties might have produced more significant findings regarding the contribution of per-capita availability of treating physicians and hospital beds on BC-related outcomes, which this study was unable to detect. Also, this study did not include oncologists with cancer-treating physicians. However, we are somewhat reassured by the fact that most patients with BC receive surgery. As such, 94 percent of BC diagnosed patients in SEER between 1998 and 2002 received surgery as the first course of treatment (Yu 2009), and we would not expect that this would have diminished the potential impact of per-capita availability of cancer-treating physicians on BC-related outcomes. Finally, although cancer registry data (Zippin, Lum, and Hankey 1995) including the GCCR have high completeness for incident cases (Clarkson 2007), research has shown that the quality and completeness of treatment data need improvement (Malin et al. 2002). We attempted to minimize missing or incomplete treatment information by removing the most recent years of data as more recently diagnosed cases were less likely to have complete treatment information.

Results from our study suggest that both race/ethnicity and area-level measures of SES and geographic residency status contribute to disparities in BC-related outcomes. Based on the magnitudes of the effect sizes (aOR and aHR), individual and contextual factors are equally important predictors of BC disparities. National and local policies and programs designed to improve the quality of health care for all patients are necessary to eliminate disparities, as disadvantaged communities have a disproportionately larger percentage of minorities, understanding and answering their needs will have multiplicative effects in reducing racial, ethnic, and geographic disparities in health care and health outcomes.

SUPPORTING INFORMATION

Additional supporting information may be found in the online version of this article

Appendix SA1

Author Matrix.