Detecting an Association between Socioeconomic Status and Late Stage Breast Cancer Using Spatial Analysis and Area-Based Measures

One of the fundamental principles of public health is that socioeconomic status (SES) is a strong determinant of health. This strong relationship has been documented for centuries, dating back to ancient Greece, Egypt, and China (1-3). More and more health researchers believe that a narrow focus on individuals outside of historical, social, and biophysical contexts limits the understanding of disease etiology, health, and intervention modes.

Breast carcinoma is the most common cancer in women in the United States, particularly in Florida, and the second leading cause of cancer death in women (4). Approximately 64,000 Florida women were diagnosed with a primary breast cancer between 1998 and 2002.

The relationship between SES and breast cancer is complex. Overall incidence rates increase with increasing SES (5). However, the incidence of distant breast cancer is higher among women of low SES (6). This disproportionate distribution is consistent with reports indicating that women of lower SES have lower rates of mammography screening and lower adherence to recommended follow-up after abnormal mammograms (6).

Racial differences have been documented in breast carcinoma incidence, mortality, and survival rates. Although breast cancer incidence is somewhat lower among African American women than among White women in the United States, breast cancer mortality is consistently higher among African Americans (7, 8). These racial disparities in breast carcinoma mortality and survival rates can be partially explained by cancer stage distribution at the time of diagnosis, treatment quality, treatment adherence, etc., which may in turn be related to SES.

In addition to race, social and economic factors have been associated with cancer mortality and survival. Baquet et al. found that SES, more than race, predicts the likelihood of a group's access to education, certain occupations, and health insurance, as well as income level and living conditions, all of which are associated with a person's chance of being diagnosed with late stage disease and surviving cancer (9).

The lack of readily available patient level SES information in surveillance systems limits large scale research efforts to fully study this phenomenon, subsequently limiting our total understanding of the contributions of SES to breast cancer incidence and mortality. Despite growing recognition of the magnitude of SES inequalities in health, very few, if any, SES data are available in most U.S. public health surveillance systems (10, 11).

One solution to the absence of SES data in surveillance systems gaining increased recognition involves geocoding and the use of area-based SES measures to characterize persons in both public health data bases and total population (12). These census-derived area-based measures can be conceptualized as meaningful indicators of SES context in their own right and not merely “proxies” for individual level data. These measures provide information on not only the area's residents (i.e., its composition) but also area level characteristics not reducible to the individual level (e.g., concentration of poverty, absence of a nearby clinic, adjacency of toxic waste site, environmental attributes, etc.; ref. 13).

Assignment of SES from ecological data can be challenging. The literature is replete with different SES indices that are composed of various census variables. Krieger et al. suggest that efforts to monitor the U.S. socioeconomic inequalities in health using area-based measures will be best served by those tract or block group measures that are (a) most attuned to capturing economic deprivation, (b) meaningful across regions and over time, and (c) easily understood and hence based on readily interpretable variables with a priori categorical cut points. The U.S. census variable “percentage of persons living below the U.S. poverty line” meets all of these criteria (14).

This research uses only late stage disease because late stage disease is most often implicated with lower SES (15, 16). Using cancer surveillance data, this study incorporated spatial statistical methodology and U.S. census–derived area-based socioeconomic measures as an initial evaluation of the degree to which SES is associated with the incidence of late stage breast cancer in Florida. This research shows that the lack of individual SES data need not limit research efforts directed at studying the effects of SES on disease. Additionally, the methods used in this research can “map” geographically defined subpopulations of excess disease, allowing cancer control professionals to target interventions. Furthermore, these geographically defined areas can serve as a baseline for evaluating the effectiveness of future interventions.

This was a cross-sectional study of Florida women diagnosed with late stage (regional and distant stage) breast cancer between 1998 and 2002, as identified by the State of Florida incident cancer registry, the Florida Cancer Data System (FCDS). A woman's SES was based from the census variable “ratio of income to poverty” of her block group of residence at diagnosis (obtained from the individual patient data). Block group was chosen because it is the smallest geographic unit for which U.S. census data are available and allows for a more precise SES assignment than either census tracts or zip codes (which can change over time and are more likely to contain heterogeneous populations due to their larger size). Of note, the words “neighborhoods” and “area” used throughout this paper are intended to be synonymous with the term “block group(s).”

The unit of analysis for this study was the block group. As depicted in Fig. 1 and discussed in detail below, the patient level data from Florida's statewide, population-based registry were aggregated by race and age group at the block group level. The aggregated patient data served as the numerators and the aggregated census population (also stratified by race and age group) as the denominators for the spatial analysis. The results of the spatial analysis identified block groups of higher-than-expected late stage breast cancer incidence. Logistic regression analysis was used to model the probability of a block group having a higher-than-expected incidence of late stage breast cancer as a function of the block group SES. The dependent variable was the incidence of late stage breast cancer (i.e., block groups categorized as having either an excess or expected incidence by the cluster detection software described below). The independent variables included area-based measures: (a) census-derived SES, (b) late stage breast cancer patient–derived insurance status, (c) county level urban/rural residence, and (d) county level mammography prevalence.

The study was conducted using data collected for female breast cancer patients, residents of Florida, diagnosed, from January 1, 1998 through December 31, 2002, with late stage breast cancer disease (regional or distant) by FCDS. FCDS is the legislatively mandated, population-based central cancer registry for the State of Florida.

Cases in the FCDS are abstracted from patient medical records by hospitals, freestanding ambulatory surgical facilities, radiation therapy facilities, private physicians, and death certificates and contain basic demographic, cancer diagnosis (primary site and cancer histology), tumor staging, and treatment information. External estimates designate FCDS completeness percentage to be >95% (as determined by external quality control audits). FCDS is part of the Centers for Disease Control National Program of Cancer Registries and is nationally certified by the North American Association of Central Cancer Registries at its highest level, “gold certification,” for meeting/exceeding the standards for completeness, timeliness, and quality.

Data collected and coded by FCDS are in accordance with national standards as put forth by North American Association of Central Cancer Registries. International Classification of Diseases for Oncology, 3rd edition was used to code primary site and morphology. During the study time period, tumors were staged using the SEER General Summary Stage, which categorizes stage of disease as in-situ, local, regional, or distant. For the purpose of this analysis, only cases coded as regional or distant stage at diagnosis were included in the study. Regional and distant stage tumors were combined in one “late stage” category (of note, because regional and late stage data were combined into one category of late stage, the revised criteria to the summary stage schema in 2001 had no effect on this analysis).

The file used for this analysis was a de-identified file containing over two million patient level cancer records diagnosed between 1981 and 2002. Although there were no patient identifiers on the file, the file contained geocoded information for each case (census tract, block group, block, longitude, and latitude), geocoded to the 2000 U.S. census by a contract vendor. This study was approved by the University of Miami Institutional Review Board.

The criteria for selecting unduplicated primary breast cancer cases included International Classification of Diseases for Oncology, 3rd edition primary site codes C50.0 through C50.9, Florida resident at time of diagnosis, diagnosed between 1998 and 2002, and stage of disease at diagnosis recorded as regional or distant. The dataset used for this analysis was unduplicated by FCDS.

There were 19,023 late stage female breast cancers recorded among Florida residents for the study period. A total of 340 cases were eliminated from the extracted data set because the records contained an “other/unknown” race code (n = 31, 0.2%) or were not assigned to a block group (n = 309, 1.6%). The final analytic data set contained 18,683 cases.

Population. Block group level population was downloaded from the U.S. 2000 Census Bureau Web site. The population data were stratified at the block group level by gender, race, and age group. Because of difficulties obtaining mutually exclusive race/ethnicity data from the U.S. census at the level of specificity necessary for the spatial software, the analysis was done using race only, without regard to Hispanic ethnicity.

There are 9,112 block groups in Florida. The individual block group annual 2000 population estimates were multiplied by five to provide 5-year sex/race/age–specific denominators for the 1998 to 2002 study period.

Socioeconomic status. For this project, the census variable ratio of income to poverty was used for the assignment of SES. The characteristics of the family used to determine poverty status are the number of people in the household, the number of related children under 18, and whether the primary householder is over age 65. An income threshold is determined given a particular family's set of characteristics; if that family's income is below that threshold, the family is considered to be in poverty. The household income divided by this threshold (as established by the U.S. government) is called the ratio of income to poverty. By way of example, the 2000 poverty threshold for a family of four was $17,463 (17).

The U.S. census provides the number of persons within each of the nine ratio of income to poverty categories (<0.50, 0.50-0.74, 0.75-0.99, 1.0-1.24, 1.25-1.49, 1.50-1.74, 1.75-1.84, 1.85-1.99, and >2.0) at the block group level (U.S. census variables, P088002-P088010). The category specific percentages were computed in SPSS.

Based on methodology by Krieger et al., the nine ratio of income to poverty categories were recoded into three categories and classified as “severe poverty” (ratio, <0.99), “near poverty” (ratio, between 1.00 and 1.99), and “nonpoverty” (ratio, >1.99; ref. 18). For each block group, the percentage of residents in severe poverty, near poverty, and nonpoverty were calculated (note these percentages add to 100%). The total block group was then assigned the SES description of severe poverty, near poverty, or nonpoverty according to the largest value among the three categories. There were 16 block groups for which two of the ratio of income to poverty categories were identical. For these 16 block groups, SES was assigned by a flip of a coin.

Insurance. Because neither the U.S. census nor any other population-based Florida survey contains block group level insurance data, the area-based measure for “insurance status” was obtained from the FCDS database. The woman's insurance status at the time of late stage cancer diagnosis in the FCDS file was recoded from 17 different values into four mutually exclusive categories: uninsured, private, medicare, and medicaid. The insurance information obtained from the FCDS data was aggregated by block group, and the proportion for each of the four values was calculated. The block group insurance status was assigned based on plurality as described for SES.

Urbanization. Urbanization was based on the county of residence at diagnosis. Urban/rural continuum codes (Beale codes; ref. 19) were used to characterize counties based on their population size, degree of urbanization, and proximity to a large metropolitan area. Beale codes were recoded into two categories: urban counties and rural counties. Block groups were classified based on the county in which the block group was located.

Mammography Prevalence. Mammography usage was based on the 2002 county level data collected by the Behavioral Risk Factor Surveillance System for the State of Florida. The race specific prevalence was not complete for each county. Therefore, individual county level prevalence data for all races were recoded into quartiles. The block group was assigned the quartile computed for the county in which the block group was located.

SaTScan version 5.0 was used to identify areas within Florida that had an excess of late stage breast cancer. The spatial scan statistic used by SaTScan is a cluster detection test that is able to detect both the location of clusters and evaluate their statistical significance. The SaTScan software can analyze spatial, temporal, and space-time data using the spatial, temporal, or space-time scan statistics (20, 21). For this study, only the spatial analysis was done.

Properties that make the spatial scan statistic suitable for geographic surveillance include (a) its ability to take into account the uneven geographic distribution of cases and population densities, (b) its lack of assumptions about cluster size or location, (c) its ability to adjust for multiple testing (a common problem in testing multiple combinations of cluster locations and sizes), (d) its ability to identify the spatial locations where the null hypothesis is rejected, and (e) its ability to detect multiple clusters (22, 23).

Using Monte Carlo techniques, SaTScan assigned relative risk probabilities to defined block groups. Race and the standard age (18 years) groups were used as covariates in this analysis. Under the null hypothesis, the incidence of breast cancer follows a Poisson distribution, and the probability of a case being diagnosed in a particular location is proportional to the covariate-adjusted population in the location.

For hypothesis testing, the SaTScan program generated 999 random replications of the data set under the null hypothesis. The test statistic was calculated for each random replication as well as for the real data set. When the latter was among the 5% highest, the test was significant at the 0.05 level (24).

ArcGis version 9.0 was the geographic information system used for this analysis to view, analyze, and relate data from a spatial (geographic) prospective.

All area-based measures derived from either the U.S. census data or the FCDS were merged with the master block group data file based on the block group key (n = 9,112). Additionally, the results of the SaTScan analysis were merged with the master block group data file, identifying each block group as having a higher-than-expected incidence of late stage breast cancer or an expected incidence of late stage breast cancer.

All analysis was done on the master block group file using SPSS version 11.0.1 to analyze the degree to which SES was associated with breast cancer stage in Florida. Multivariate logistic regression was used to assess potential confounding variables (such as insurance, urban/rural location, and mammography usage) across groups.

The incident late stage breast cancer data set represented women from each county in Florida, with an age distribution of 24 through 85+ years old. The racial distribution of the cases was similar to the racial distribution of the state, which is 87.6% White, 10.7% Black, and 1.7% others. Once the data were aggregated by block group, women from 6,211 of the 9,112 block groups in Florida were represented in the analysis.

The study population was composed of 16,626 White women and 2,057 Black women. At the time of diagnosis, insurance status varied by racial group. Black women were approximately thrice more likely to be uninsured than White women (12.3% versus 4.7%) and 3.5 times more likely to receive Medicaid (11.2% versus 3.3%). Based on block group area–based socioeconomic measures, the median household income for Blacks was approximately $7,000 per annum less than the White population. Nearly one fourth of the Black women diagnosed with late stage breast cancer lived in areas of severe or near poverty as opposed to 7.5% of White women. The overall prevalence of mammography usage in Florida between Whites and Blacks is not statistically significant. The prevalence for White women is 79.7 [95% confidence interval (95% CI), 78.3-81.2] and 76.7 (95% CI, 71.1-82.3) in Black women.

SaTScan identified 767 block groups in which the incidence of late stage breast cancer cases was higher than expected. By default, all other block groups in the analysis file (n = 5,444) were designated as having an expected incidence. The map of Florida (Fig. 2A) presents the areas that SaTScan identified as having a higher-than-expected incidence of late stage breast cancer.

To understand Florida's population and the nature of the different areas within Florida, SPSS and the spatial analytic properties of ArcGis were used to compare and contrast the area-based characteristics of the different neighborhoods.

As seen in Fig. 2B, the distribution of urban and rural areas within the areas identified by SaTScan as having higher-than-expected incidence or expected incidence was striking. Over 95% of the areas identified as having higher-than-expected rates were located in urban counties (χ² = 7.2, 1df, P = 0.007). Although the areas identified by SaTScan as having a higher-than-expected incidence covered a relatively large geographic area, the population within these areas represented ∼1.5 million (∼9% of the total Florida population).

The logistic regression results (Table 1) present the odds ratios (OR) of SES on the areas with higher-than-expected incidence versus expected incidence of late stage breast cancer. The table presents five models assessing the effect of SES on the diagnosis of late stage breast cancer, both as a simple regression and when area-based variables were sequentially added to the model as possible confounders. There were no significant interactions present in any of the models.

The effect of SES on the higher-than-expected incidence of late stage breast cancer was seen in women living in severely poor areas (OR = 2.6; 95% CI, 1.9-3.4) and women living in areas of near poverty (OR = 1.4; 95% CI, 1.0-2.2; model 1). Insurance and/or the urbanization of the area had little or no effect on this gradient (models 2-4). Whereas the magnitude of the effect is less for women living in near poverty, the implications were the same. Women living in near poverty were 50% more likely to be diagnosed with a late stage breast cancer than women residing in nonpoor areas.

The addition of mammography prevalence to the final model (model 5) had a significant effect on several variables in the model. The magnitude of the effect for both the severe and near-poverty categories increased. Areas of near poverty reached statistical significance (OR = 1.6; 95% CI, 1.1-2.6) and areas designated as severe poverty were thrice more likely to be an area of higher-than-expected incidence of late stage breast cancer (OR = 3.0; 95% CI, 2.2-4.0). Another striking result of this model was the gradient among the mammography prevalence quartiles. Areas with the lowest prevalence of mammography usage (<73.0) were 6.5 times more likely to be identified as having a higher-than-expected incidence of late stage breast cancer than areas with the highest mammography usage (>80.6; Fig. 2C). Likewise, areas in the second lowest quartile (73.0-77.2) were over 3.5 times more likely to be identified as having a higher-than-expected incidence of late stage breast cancer (OR = 3.6; 95% CI, 3.0-4.5).

Unlike the previous four models, medicare was statistically significant in the final model (relative to uninsured), indicating that areas of higher-than-expected incidence of late stage breast cancer are 40% less likely to contain medicare recipients (OR = 0.6; 95% CI, 0.4-0.9).

The results of this research, indicating that residing in areas with high levels of poverty is significantly associated with late stage diagnosis of breast cancer, are noteworthy and consistent with other studies (25-30). These findings remained after attempting to control for important risk factors, such as insurance and mammography prevalence. In fact, the magnitude of the main effect was strengthened with the addition of mammography prevalence.

Insurance status was not significant in any of the models until the inclusion of mammography prevalence. With the inclusion of mammography prevalence estimates, medicare recipients were 40% less likely to live in areas of higher-than-expected incidence of late stage breast cancer. This may be due to the loss of specificity when the individual insurance variable was converted to an area-based measure.

The areas of higher-than-expected incidence of late stage breast cancer were almost seven times more likely to be the same areas with the lowest mammography usage. These findings are consistent with the well-documented benefits of mammography for the early detection of breast cancer. Despite this high association, it is also important to note that the area-based measure of poverty used in the present analysis remained statistically significant in our multivariate models, suggesting that mammography usage alone does not explain all of the variability in late stage breast cancer incidence rates in Florida.

The results of this study indicate that SES plays an independent role in the late stage diagnosis of breast cancer beyond the effects on mammography access. There are poorly understood (and sometimes controversial) underlying factors for increased late stage breast cancer incidence among women of lower SES that include access and or usage of health care (including cultural and social beliefs which may influence mammography usage; refs. 31, 32).

The “upstream” factors defining an individual's socioeconomic status are complex and multidimensional. Women of lower SES may be more likely to ignore symptoms for a variety of economic, social, or cultural reasons. These women tend to work in nonsalaried positions, therefore taking time off for a doctor's appointment more than likely results in a “no work-no pay” situation. Additionally, women of lower SES tend to be lesser educated and may not understand the elements of disease progression, and, as a result, they do not respond as quickly when feeling a palpable mass.

SES factors alone are often not sufficient to explain the dramatic variation in breast cancer stage among certain women. However, SES variables, in conjunction with cultural beliefs and attitudes, have been shown to account for a large portion of the observed effect (28). There is a growing body of literature suggesting an association between cultural, religious, and other psycho/social factors related to breast cancer screening (28, 33, 34). A dissertation of these factors is beyond the scope of this paper. However, the methods used in this research can assist cancer control professionals in identifying geographically defined subpopulations to survey for the identification of possible cultural, religious, and other psycho/social factor barriers to screening which could eventually result in the development of culturally sensitive interventions. Moreover, this methodology can offer researchers a baseline from which to compare the effectiveness of the interventions.

Results from the present study suggest that “where” you live plays an important role in defining the risk of presenting with late stage breast cancer. It is essential that additional research be conducted to assess and quantify barriers to health care, including, but not limited to, access and usage of primary health care in communities with higher-than-expected late stage cancer cases. Additional research is urgently needed to leverage the strengths and resources present in all communities to lower the late stage breast cancer burden, especially in high-poverty communities.

There were several study limitations that should be noted, starting with an acknowledgement that cross-sectional study designs cannot be used to establish causal associations between SES and delayed breast cancer diagnoses. Also, because this is an ecological study, we must acknowledge the possibility of ecological fallacy.

The area-based population and SES classifications were assigned based on data collected in 1999 for the 2000 U.S. census. The population data were extrapolated for the 5-year study period by multiplying the single 2000 census year population by 5. To minimize the potential bias from this methodology, the study period was selected to include 2 years before and 2 years after the censal year.

Estimates of mammography usage were available at the county, but not the block group level, from the Florida Behavioral Risk Factor Surveillance System. Use of county level data to estimate area level prevalence rates almost certainly resulted in some loss of precision given the likelihood of intracounty variability in these rates. Additionally, the Behavioral Risk Factor Surveillance System data may be overestimating recent mammography usage, particularly among low-income women (35).

A potential limitation could be attributed to the accuracy and completeness of the geocoded address data. Approximately, 94% of the cases were geocoded based on the complete address and street number. Only 5.3% of the cases were geocoded at only the zip code level (the centroid of the zip code was used to assign the geocoded variables). There were 309 cases excluded from the study because they were geocoded only to the county level, not at the block group level. There was nothing to indicate that any of the SES strata were disproportionately affected by either of these issues. If such a bias was to exist, it would be expected to effect each stratum equally.

Another potential limitation could have been the use of a single variable indicator versus composite indicators for the assignment of SES. However, based on data from Public Health Disparities Geocoding Project, Krieger et al. found that several of the single-variable measures, especially those intended to measure poverty, detected the same magnitude of socioeconomic inequality in health as the composite measures (36).

Finally, future analysis of these data should use multilevel modeling techniques that will take into account predictor and outcome variables that can occur at multiple levels of aggregation (i.e., the individual, block group, county, and state).

There are also several study strengths, including the use of spatial statistical technology to identify areas of interest below the county level and to incorporate race (in addition to age) as a covariate. Additionally, the inclusion of area-based socioeconomic measures allowed for the enhancement of cancer surveillance data, thus allowing for the inclusion of multifaceted indicators that are not routinely captured in surveillance systems.

When analyzing overall breast cancer incidence, women of higher SES have a higher overall incidence of breast cancer. However, as was found in this study, when the analyses were stratified by the stage of the disease, there is a clear and significant gradient between late stage breast cancer and residence in high-poverty areas. Research has shown that usage of mammography screening in economically disadvantaged areas is less than wealthier areas for a variety of practical and economic reasons (37, 38). Therefore, additional research must be done to assess and understand the barriers to screening and/or the follow-up on positive screening examinations in these populations.

Health care access and usage may not be related solely to income and/or education. Social and economic factors, as well as cultural/spiritual beliefs, may be barriers to health care access and/or usage. To obtain this information, researchers must use emerging spatial methodology and survey techniques that will enable them to augment the current surveillance systems. Additionally, the survey component would enable the research community to “verify” the area-based socioeconomic data.

There are many important reasons to continue enhancement of existing surveillance systems with area-based socioeconomic measures. The methodology used in this study not only gives public health practitioners the ability to identify communities that may benefit from targeted interventions to reduce their burden of delayed breast cancer diagnosis, but it is also a powerful tool for furthering our understanding of the role that SES indicators may play in complex disease processes. Additionally, this methodology has the ability to use modeling techniques to prospectively identify communities that have the potential to become areas of higher-than-expected incidence of cancers. Ultimately, this information can be used to develop effective and culturally sensitive interventions designed to reduce the cancer burden in all communities.