ARTICLE IN PRESS

Health & Place 16 (2010) 209–218

Contents lists available at ScienceDirect

Health & Place

1353-82

doi:10.1

n Corr

E-m

Wilson)1 Te2 Te

journal homepage: www.elsevier.com/locate/healthplace

Female breast cancer mortality clusters within racial groupsin the United States

Nancy Tian a,1, J. Gaines Wilson b,n, F. Benjamin Zhan a,c,2

a Texas Center for Geographic Information Science, Department of Geography, Texas State University-San Marcos, 601 University Drive, San Marcos, Texas 78666, USAb Department of Chemistry and Environmental Sciences, University of Texas at Brownsville, 80 Fort Brown —MO1.114, Brownsville, Texas 78520, USAc School of Resource and Environmental Science, Wuhan University, Wuhan, Hubei 430079, China

a r t i c l e i n f o

Article history:

Received 20 May 2009

Received in revised form

21 September 2009

Accepted 23 September 2009

Keywords:

GIS

Clusters

Spatial epidemiology

Race

Breast cancer

92/$ - see front matter & 2009 Elsevier Ltd. A

016/j.healthplace.2009.09.012

esponding author. Tel.: +1 956 882 6646; fax

ail addresses: jt1277@txstate.edu (N. Tian), je

, zhan@txstate.edu (F. Benjamin Zhan).

l.: +1 512 245 2170; fax: +1 512 245 8353.

l.: +1 512 245 8846 fax: +1 512 245 8353.

a b s t r a c t

Although breast cancer is the second leading cause of cancer deaths among women in the Unites States,

to date there have been no nationwide studies systematically analyzing geographic variation and

clustering. An assessment of spatial–temporal clusters of cancer mortality by age and race at the county

level in the lower 48 United States indicated a primary cluster in the Northeast US for both younger

(RR=1.349; all RR are pr0.001) and older (RR=1.283) women in the all-race category. Similar cluster

patterns in the North were detected for younger (RR=1.390) and older (RR=1.292) white women. The

cluster for both younger (RR=1.337) and older (RR=1.251) black women was found in the Midwest. The

clusters for all other racial groups combined were in the West for both younger (RR=1.682) and older

(RR=1.542) groups. Regression model results suggest that lower socioeconomic status (SES) was more

protective than higher status at every quartile step (Medium-high SES, OR=0.374; Medium-low,

OR=0.137; Low, OR=0.061). This study may provide insight to aid in identifying geographic areas and

subpopulations at increased risk for breast cancer.

& 2009 Elsevier Ltd. All rights reserved.

1. Introduction

Since the early 1990s, cancer morbidity and mortality rateshave declined for both men and women, largely due to preventiveprograms such as mammography screening and anti-smokingcampaigns. However, breast cancer continues to be a seriousconcern; breast cancer mortality only declined approximately 5%between 1995 and 2006 to around 41,000 deaths (Greenlee et al.,2001; Jemal et al., 2006). Breast cancer is ranked as the secondhighest cause of cancer death in women of all ages, preceded onlyby lung cancer, and for women between forty and fifty-nine yearsold, breast cancer is the greatest cause of cancer death (Greenleeet al., 2001). Moreover, breast cancer outcomes significantly varyamong racial/ethnic groups (Sarker et al., 2007). African-Amer-icans have the second highest incidence for breast cancer afterwhite women, and African-Americans are 33% more likely to diefrom breast cancer than whites and more than twice as likely asAsians to die from the disease (Ries et al., 2000). However, womenwithin other racial groups (consisting of American Indian/Alaska

ll rights reserved.

: +1 956 882 6693.

ff.wilson@utb.edu (J. Gaines

Native and Asian/Pacific Islander) have 50% lower incidence andmortality from breast cancer as compared with white and blackwomen (Smigal et al., 2006).

A limited number of studies assessing regional, state-specificand even localized cluster investigations have been performed inorder to assess geographic variations of breast cancer (Kulldorff,1997; Zhan, 2002; Hsu et al., 2004; Jacquez and Greiling, 2003).Kulldorff and colleagues found that the New York City–Philadel-phia Metropolitan Area had 7.4% higher breast cancer mortalitythan the rest of the Northeast, after adjusting risk factors for race,urbanicity and parity (Kulldorff, 1997). The well-known LongIsland Breast Cancer Project has warranted further researchattention on environmental exposures such as traffic pollutionand pesticide use at residence (Jenks, 1994; Lewis-Michl et al.,1996). The Atlas of Cancer Mortality, published by the NationalCancer Institute, suggests that there is a higher concentration ofdeath in women from breast cancer in the Northeastern UnitedStates relative to the rest of the country, although no systematicstatistical analysis was performed to investigate those patterns(Devesa et al., 1999; Pickle et al., 1987; Riggan et al., 1983).Sturgeon et al. (2004) have argued that the geographic concen-tration of breast cancer mortality in the Northeast may have beenattenuated due to the less favorable socioeconomic trends in theSouth, although their study did not include any statistical tests forthe geographic variability by age group. Research that assesses

ARTICLE IN PRESS

N. Tian et al. / Health & Place 16 (2010) 209–218210

geographic variations in disease area remains a priority as it mayoffer evidence for enacting programs that enhance the quality ofhealth care in underserved minority groups and allow for moreeffective allocation of health care resources and education (Farrowet al., 1992; Garland et al., 1990; Jenks, 1994).

To date there have been no national breast cancer clusterdetection studies conducted within the United States. In thispaper we intend to identify statistically significant clusters ofbreast cancer mortality for women in various age and racecategories at the county level for the entire contiguous UnitedStates. The broader goal of this work is to provide baselineresearch on geographic areas of focus for future work attemptingto understand the factors that affect spatial clustering of breastcancer.

2. Data and methods

2.1. Geographic, mortality and population data

The study area consists of the 3088 counties within the 48contiguous states in four geographic regions: the Northeast(n=213), the Midwest (n=1055), the South (n=1408) and theWest (n=412). The county geographical boundaries for the 2000census were obtained from the National Cancer Institute (NCI).Geographic position of each county is required as input for thespatial analyses, so we derived the position of the centroid (ameasure for the geographical center point of a polygon) for eachcounty using a zonal geometry function in a GeographicInformation System (GIS). All GIS techniques and map layoutswere performed using ArcMap v. 9.3 from ESRI.

The mortality and population data were from the NationalCancer Institute (NCI) at the county level for the period of 1984–2004 (e.g., 1984–1986, 1987–1989, 1990–1992, 1993–1995, 1996–1999, 2000–2004). The county level is the finest level at whichbreast cancer mortality is available from the NCI. Number of casesby age groups were automatically generated by NCI’s SEERnStatsoftware (Version 6.4.4). The county population estimatesimplemented in SEERnStat software were based on the US CensusBureau’s Population Estimates Program. The count for femalebreast cancer mortality was designated as one of three racialgroups (white, black and other), by 19 age groups with five-yearintervals (0=0–4, 1=5–9, y, 19=85+) and linked to the county ofresidence. Age was included as a covariant for all the spatialcluster analyses within racial groups by taking into account agedistribution heterogeneity. The literature has demonstrated thatmenopausal effects may influence women to develop breastcancer differently because of the estrogen level changes (Carey etal., 2006). When the menopausal history is not available, 50 yearsof age is suggested as the best proxy to take into accountmenopausal effects on breast cancer for women (Anderson et al.,2002; Morabia and Flandre, 1992). The resulting analysis from thisstudy intends to identify the significant clusters for the sixcombinations of tests built from three racial groups and twocategories of menopausal status while adjusting for age.

Socioeconomic status (SES) has been reported as a risk factorfor breast cancer mortality (Madigan et al., 1995; Singh et al.,2003). SES in this study was recoded into quartiles using thepercentage of population in each county under the povertythreshold ($11,250 for two-person household) based on the US2000 Census (Yost et al., 2001). The SES were categorized ashighest, medium-high, medium-low and low according to thequartile levels. Beyond SES, mammography facilities have alsobeen shown to explain breast cancer mortality (Bradley et al.,2002). The US Food and Drug Administration (FDA) maintain amammography facility database according to information from

the FDA-approved accreditation bodies. The count of all US FDA-certified mammography facilities in 2009 was available by zipcode (historical data at the sub-state level was not available) andaggregated at the county level using GIS. Our analysis alsointegrated a measure of environmental exposure to furtherenhance understanding hotspots with elevated cancer mortalityburden (Brody and Rudel, 2003; Brody et al., 2007). The USEnvironmental Protection Agency (EPA) provides data on thelocations and characteristics of facilities related to air, water,toxics and waste that may adversely affect human health and theenvironment. The nationwide coordinates for regulated facilitieswere obtained from the EPA, geocoded and then aggregated at thecounty level using GIS.

2.2. Statistical methods

In this study, a ‘‘cluster’’ is defined as a geographic area withsignificantly elevated risk within the study region as compared toother regions. Here, we will use the terms ‘‘cluster’’ and ‘‘excessmortality’’ interchangeably. Various statistical methods have beenproposed to verify if reported disease clusters are significant ornot (Knox and Elliot, 1989; Kulldorff, 1997; Jacquez and Greiling,2003; Openshaw et al., 2006). The cluster detection method weuse in this paper, the space-time scan statistic, SaTScan, byKulldorff (1998), has been applied to a variety of disease clusterdetections including breast cancer (Jemal et al., 2002; Hsu et al.,2004), brain cancer (Kulldorff et al., 1998) and leukemia (Hjalmarset al., 1996). The spatial scan statistic is an appropriate approachfor spatial surveillance detecting elevated cancer burden becauseof the following characteristics. First, it takes into account theheterogeneous distribution of cases and population at space andaccounts for confounding variables flexibly. Second, it avoids theassumption of the size and location of clusters by searchingvarying size of scanning windows, which prevents pre-selectionbias issues. In contrast with the exploratory analysis methods(Fotheringham and Zhan, 1996; Kulldorff, 1998; Openshaw et al.,2006; Rushton and Lolonis, 1996), the spatial scan statistic adjustsmultiple testing by only evaluating statistical significance of themost likely and secondary clusters. This method has been appliedin various epidemiological studies for identifying disease clustersand performing significance tests (Kulldorff, 1997; Zhan, 2002;Hsu et al., 2004) and there is a substantial body of literatureshowing that the spatial scan statistic effectively facilitatesdisease surveillance by detecting preliminary clusters (Hjalmarset al., 1996; Hsu et al., 2004; Sheehan et al., 2004; Zhan and Lin,2003).

We adopted a Poisson-based model for the mortality countdata under the assumption that the expected cases of breastcancer mortality were proportional to the age-adjusted popula-tion in each county. Hsu et al. (2004) utilized a similar method inspatiotemporal analysis of breast cancer mortality in Texas,controlling for age while examining racial groups in separatecategories. In this study, SaTScan imposes a scanning ellipticwindow to compare the observed cases with those expectedaccording to the corresponding co-variable adjusted population.The cylindrical scanning windows have an elliptic geographic basewith height corresponding to years under consideration. We alsotested the sensitivity of cluster results to a temporal effect byincluding purely spatial cluster detection over the entire studyperiod. Scanning windows had flexible radii from zero up to amaximum value of 50% of the total population under considera-tion. Fifty percent of population at risk is recommended tooptimize the parameter setting in order to detect all the potentialareas with elevated cancer risk (Kulldorff, 2006). SaTScangenerates a likelihood function with a value that is proportional

ARTICLE IN PRESS

N. Tian et al. / Health & Place 16 (2010) 209–218 211

to the ratio of observed cases by expected cases within a windowcompared with the ratio outside the window. The Monte Carlosimulation was implemented to generate the replication ofdataset under the null hypothesis to assess the significance ofdetected clusters. Nine hundred and ninty-nine simulations werechosen in order to obtain the simulated p-values. The most likelywere determined based on the likelihood ratio with the p-value.We only report statistically significant clusters with an indicatedp-value less than 0.05. The SaTScan program reported observedcases, expected cases, relative risk, log likelihood ratio and FederalInformation Processing Standard (FIPS) of the counties in aspecific cluster. Further details about the space-time scan statisticcan be found in the user guide of the SaTScan software package(version 7.0.3) that was designed to implement the statisticalmethod (Kulldorff, 2006). A custom C++ program was coded toprocess the format of county FIPS within clusters in order torepresent these clusters using GIS techniques.

A multivariate logistic regression model was developed, withpresence (1) or absence (0) of a cluster in the county as thedependent categorical variable based on the output from thecluster analysis. The model evaluated the influence of followingpredictor variables at the county level: (i) socioeconomic status;(ii) mammography facilities and (iii) US EPA-regulated facilities.The independent variable of SES was categorized as high,modestly high, medium and low according to quartile of thepoverty level for each respective county according to the 2000Census (Du et al., 2007; Wang et al., 2008). The number ofmammography facilities and EPA-regulated facilities in eachcounty were continuous variables. The ANOVA and logisticregression analyses were performed in SPSS 16.0.

3. Results

During the years 1984–2004, the average annual populationfor women was 133,767,531 in 48 continuous states and the totalbreast cancer mortality was 879,002. Fifteen percent of womenwho died of breast cancer were younger than 50 (n=130,298) andeighty-five percent were age 50 and older (n=748,704). The space

Fig. 1. All races: the potential statistically significant clusters for female breast cancer m

1984–2004.

time scan statistic was applied to identify significant clusters ofbreast cancer mortality for women after adjusting for race and agecovariates at the county level. In order to identify spatial andtemporal concentrations of cancer mortality, the cluster detectionapproach was also implemented for six combinations consistingof three racial groups (white, black and other) by two categories ofmenopausal status while adjusting for age covariates.

The overall annual mortality over the study period from breastcancer was 30.4 per 100,000, while the annual mortality rates forwomen younger than 50 and 50 and older were 6.1 and 91.1 per100,000, respectively. For women younger than 50, the annualage-adjusted mortality rates were 6.0, 8.1 and 2.6 per 100,000 forwhite, black and other racial groups, respectively. For women age50 and over, the annual age-adjusted mortality rates were 92.8,92.9 and 26.9 per 100,000 for white, black and other racial groups.For each racial group, the mortality rates for women 50 and olderwere ten times greater than for women younger than 50. Incontrast with whites and blacks, the other racial group had 3.5times lower mortality rates across both age groups.

Fig. 1 illustrates the significant clusters for female breastcancer mortality after adjusting for all races and age groups. Themost likely cluster (Relative Risk (RR)=1.288, p=0.001) covers asubstantial amount of the counties (n=1192) in the Northeast,South and Midwest Regions between 1984 and 1995. Whenadjusting for all race and age covariables, Fig. 2(a) displays themost likely clusters for women younger than 50 with a relativerisk of 1.349 (p=0.001). Women age 50 and over (RR=1.283,p=0.001) (Fig. 2(b)), have similar geographic patterns as identifiedfor all race and age groups in the years of 1984–1995 (refer toFig. 1). Part A of Table 1 lists the detailed statistics of the mostlikely clusters for all races combined in addition to younger andolder groups, which include age-adjusted rates, cluster years,observed cases, expected cases, relative risk and p-value.

For white women, Fig. 3 summarizes the clusters withstatistical significance (po0.05). We conducted the spatialcluster analyses for younger than 50 and age 50 and olderseparately to account for menopausal effect on breast cancermortality (Paffenbarger et al., 1980). For women younger than 50(Fig. 3(a)), the most likely clusters were found in 196 counties in

Most Likely Cluster for All RacesYears: 1984-1995Relative Risk: 1.288P Value: 0.001

ortality adjusted for race and age covarariates at the county level, the United States,

ARTICLE IN PRESS

Most Likely Cluster for All RacesYears:1984-1995RelativeRisk:1.349P Value:0.001

Most Likely Cluster for All RacesYears: 1984-1995Relative Risk: 1.283P Value: 0.001

Fig. 2. All races: the potential statistically significant clusters for female breast cancer mortality among younger than 50 as well as 50 and over adjusted for race and age

covarariates at the county level, the United States, 1984–2004.

N. Tian et al. / Health & Place 16 (2010) 209–218212

the Northeast, 303 counties in the Midwest and 716 counties inthe South with a relative risk of 1.390 (p=0.001). The group ofwhite women age 50 and over (Fig. 3) has a similar geographicand temporal mortality concentration (RR=1.292, p=0.001) as thecluster detected for the older age group of all races combined(refer to Fig. 2). Clusters identified for both younger and oldergroup within whites occurred between 1984 and 1995. Part B ofTable 1 refers to the detailed statistics of the most likely clustersfor white females.

Fig. 4 illustrates the statistically significant clusters for blackwomen younger than 50 as well as 50 and older. For youngerwomen (Fig. 4(a)), the most likely cluster was concentrated in 384counties in the Midwest (n=384) and 1008 counties in the Southwith the temporal concentration of 1984–1995. There was a 33.7%higher risk (RR=1.337, p=0.001) in these areas relative to the restof the study area. For black women 50 and over (Fig. 4(b)),substantial clusters were identified in the Northeast, the Southand the Midwest for the period of 1984–1995. The most likely

cluster included 1376 total counties extending from the Northeast(n=131) through the Midwest (n=641) to the South (n=604). Thecluster detected had as high as 1.251 of relative risk with asignificance level of 0.001. Part C of Table 1 describes the variousstatistics about the most likely clusters for black women for boththe age groups.

For women younger than 50 within the other racial group(consisting of American Indian/Alaska Native and Asian/PacificIslander), the most likely cluster was found in California, whichincluded 42 counties (Fig. 5(a)). The area had 68.2% (RR=1.682,p=0.001) more cases observed than would be expected. Forwomen age 50 and over in the other racial group (Fig. 5(b)), themost likely cluster extended its size to include many morecounties (408) in the West (RR=1.542, p=0.001). Clusters detectedfor both younger and older groups persisted in the whole studyperiod of 1984–2004. Part D of Table 1 shows the statisticalproperties for both younger and older women in the other racialgroup.

ARTICLE IN PRESS

Most Likely Cluster for WhiteYears: 1984-1995Relative Risk: 1.390P Value: 0.001

Most Likely Cluster for WhiteYears: 1984-1995Relative Risk: 1.292P Value: 0.001

Fig. 3. White females: the potential statistically significant clusters for female breast cancer mortality among younger than 50 as well as 50 and over adjusted by age at the

county level, the United States, 1984–2004.

Table 1Breast cancer mortality spatial–temporal clusters for women in the United States at the county level, using Spatial Scan Statistic with statistical significance.

Race Age Age-adjusted rates Year Observed cases Expected cases Relative risk p-value

A All All 35.9 1984–1995 268,396 227,263 1.288 0.001

o50 7.5 1984–1995 39,231 32,218 1.349 0.001

Z50 107.3 1984–1995 231,552 196,677 1.283 0.001

B White

o50 7.4 1984–1995 32,004 25,837 1.390 0.001

Z50 109.8 1984–1995 207,413 175,318 1.292 0.001

C Black

o50 9.9 1984–1995 7047 5792 1.337 0.001

Z50 107.9 1984–1995 21,504 18,508 1.251 0.001

D Other

o50 3.5 1984–2004 981 724 1.682 0.001

Z50 32.8 1984–2004 2823 2318 1.542 0.001

N. Tian et al. / Health & Place 16 (2010) 209–218 213

ARTICLE IN PRESS

Most Likely Cluster for BlackYears: 1984-1995Relative Risk: 1.337P Value: 0.001

Most Likely Cluster for BlackYears: 1984-1995Relative Risk: 1.251P Value: 0.001

Fig. 4. Black Females: the potential statistically significant clusters for female breast cancer mortality among younger than 50 as well as 50 and over adjusted by age at the

county level, the United States, 1984–2004.

N. Tian et al. / Health & Place 16 (2010) 209–218214

A multivariate logistic regression model was used to assess therelationship between a county belonging to a cluster or not andother risk factors including SES, the number of EPA-regulatedfacilities and the number mammography facilities for each county(Table 2). The dependent variable of a county belonging to acluster or not was detected for all races combined was age- andrace-adjusted. Table 2 details the odds ratio and statisticalsignificance reported for socioeconomic status, number ofmammography facilities in each county and number of EPA-regulated facilities which was used as an indicator ofenvironmental exposure. The R2 for the model was 0.268. Whencompared with the first quartile (reference) of SES (higher SES),second, third and fourth quartiles had odds ratios of 0.374, 0.137and 0.061 (all po0.001), respectively. The odds ratio for thecontinuous variable of number of EPA-regulated facilities wasslightly positive at 1.001 (p=0.478; 95% C.I. 0.999–1.003). Theexplanatory variable for the number of mammography facilitieshad an odds ratio equal to 1.003 and p-value greater than 0.05(p=0.808).

4. Discussion

The spatial analysis presented in this paper provides a generalpicture of where the excess mortality of female breast cancer waslocated in the United States and whether the variance wasstatistically significant for each racial group by grouping thesubjects into age groups of younger than 50 as well as 50 andover, as 50 is the typical age for the onset of menopause (Andersonet al., 2002; Morabia and Flandre, 1992). Moreover, the aboveresults imply that for both age groups, breast cancer mortalityvaries significantly across races and has a temporal concentrationin the years 1984–1995 except for women falling into the otherracial group. Detected clusters disappeared in the years 1996–2004relative to the period of 1984–1995. While this does not necessarilyindicate a downward trend in breast cancer mortality per se,previous studies have noted a declining trend in breast cancermortality since the 1990s due to more effective screening andimproved treatment techniques (Berry et al., 2005). We demon-strated that the cluster in the Northeast Region was statistically

ARTICLE IN PRESS

Most Likely Cluster for OtherYears: 1984-2004Relative Risk: 1.682P Value: 0.001

Most Likely Cluster for OtherYears: 1984-2004Relative Risk: 1.542P Value: 0.001

Fig. 5. Other females (consisting of Alaska Native, American Indian, Asian and Pacific Islander): the potential statistically significant clusters for female breast cancer

mortality among younger than 50 as well as 50 and over adjusted by age at the county level, the United States, 1984–2004.

Table 2Results of logistic regression for United States county-level clusters of female breast cancer mortality adjusted by race and risk factor.

Variables Wald statistic Odds ratio 95% CI Significance

Socioeconomic status

1st quartile (highest) 540.305 – – –

2nd quartile (medium-high) 84.034 0.374 0.304–0.462 o0.001

3rd quartile (medium-low) 297.790 0.137 0.109–0.171 o0.001

4th quartile (low) 429.434 0.061 0.047–0.080 o0.001

EPA-regulated facilities 0.504 1.001 0.999–1.003 0.478

Mammography facilities 0.059 1.003 0.980–1.027 0.808

(–) represents not applicable for the cell.

N. Tian et al. / Health & Place 16 (2010) 209–218 215

significant, confirming previous research that reported elevatedmortality rates in the Northeast (Garland et al., 1990; Kulldorf et al.,1997). Furthermore, adjustment for an age-proxy of menopausaleffects on elevated risk of breast cancer mortality influenced theidentified geographic clusters for all three racial groups. Thisevaluation showing higher risk for breast cancer mortality in

women by region, race and age provides insight for public healthplanning officials and may help identify targeted subpopulations.

The higher mortality rates in the Northeast could be attributedto the considerable amount of white women who died of breastcancer in that area, because the white racial group for bothyounger and older women shared a similar geographic and

ARTICLE IN PRESS

N. Tian et al. / Health & Place 16 (2010) 209–218216

temporal variation in breast cancer mortality as compared to theall races population. We obtained more supporting evidenceconcerning excessive breast cancer mortality on Long Island froma nationwide perspective (Jacquez and Greiling, 2003; Lewis-Michl et al., 1996). The two counties of Nassau and Suffolk on LongIsland fall into the most likely clusters for white females includingboth the younger and older age groups. Our results challengeanother hypothesis that substantial geographic variations couldbe explained by poor survival rates from elderly women whodevelop breast cancer in the Northeast region (Farrow et al., 1992;Goodwin et al., 1998). Goodwin and his colleagues proposed thatthe elevated mortality was primarily concentrated in olderwomen in the Northeast. However, our statistical analysisdemonstrates that there were persistent significant clusters inthe Northeast for both younger and older women among the all-race combined group (Fig. 2), which extends to the South and theMidwest. The socioeconomic status results from the logisticregression model suggest that lower SES is more protective thanhigher status at every quartile step (Table 2: Medium-high SES,OR=0.374; Medium-low, OR=0.137; Low, OR=0.061). The North-east had higher mortality rates compared to other regions before1995, even though the higher SES in the Northeast affordsresidents easier access to health care and thus, earlier detectionof breast cancer. This finding further illustrates the paradoxicalinfluence of SES on breast cancer—that higher incidence rates areoften associated with higher SES in other studies, (Clarke et al.,2002; Hsu et al., 1997), which has been attributed due to riskfactors such as late-age childbearing or lifestyle behavior amonghigher-SES populations (Kelsey et al., 1993; McPherson et al.,2000). Clarke et al. (2002) found that Marin County had 42.6%higher probability for women age 45–64 to die of breast cancercomparing to other areas in California, which is another exampleof agreement with studies that show higher breast cancermortality rates in areas with a higher level of affluence (Clarkeet al., 2002; Hsu et al., 1997). Overall, increased mortality rates inthe South, an area with historically lower SES narrow geographicdisparities in breast cancer mortality (compared to the Northeast),further supports the protective nature of higher SES (Sturgeon etal., 1995; Gordon, 2003). This consequently deserves furtherexamination in order to more effectively understand the under-lying causes of elevated cancer burden on these areas. Thenumber of EPA-regulated facilities in each county was not shownto be a significant predictor of higher mortality rates (p=0.478).Despite this finding, further investigation is warranted in order touncover the nature of environmental exposure mechanisms onbreast cancer outcomes (Saunders et al., 1997). A case-controlstudy designed for environmental exposure assessment at finerscales may also provide more convincing evidence for toxicchemical release associated with variability in female estrogenlevels (Kettles et al., 1997; Brody et al., 2006). The number ofmammography facilities in each county was not associated withgeographic cluster of excessive cancer burden. Nevertheless,issues with data may explain the insignificant p-value (p40.05).First, the data on locations of mammography facilities was notavailable on an annual basis. Even though the numbers were likelyconsistent since 1994 when the Mammography Quality StandardsAct and Program was initiated, there may be inconsistencies in theearly years of our study prior to 1994. Second, in order to obtainthe number of mammography facilities at county level, a file forthe individual clinic was joined to a geographic file from 2000census at zip code level. In the process, 96.7% of zip codes weresuccessfully joined. Inaccuracies in the number of mammographyfacilities in counties could result from implementing the centroidof zip code during aggregation due to some zip codes crossingcounty boundaries. For mammography facilities, annual zip codelevel data were only available for 2009, and not historically.

However, we were able to obtain and test the variation in numberof facilities at the state level over time and found that the numberof mammography facilities at the state level were highlycorrelated (r=0.994) between the years 1994 and 2009. Basedupon this finding, the authors made the assumption (supportedby a mammography facility expert stating that ‘‘the number ofmammography facilities has remained very stable’’ and thatfacilities at the county level ‘‘were somewhat stable’’, personalcommunication: Frier, 2009). Although the number of mammo-graphy facilities was not found to be a significant variableprotecting geographic areas from experiencing cancer burden,this does not mean that intensive mammography screening doesnot play an important role in preventing late-stage breast cancerdiagnosis (Taplin et al., 2004).

The robustness of clusters detected by SaTScan could beassessed by using finer geographic scales such as zip codes andcensus tracts, given that the data were available (they were not). Ifthe detected clusters were to persist at all geographic levels, theevidence of actual clusters would be further strengthened. It isimportant to note that patterns of disease may change at variouslevels of aggregation and scale, a problem related to themodifiable areal unit problem (MAUP) (Openshaw, 1984). Thearea-based approach utilized in this study is commonly used toderive the socioeconomic status using education, income andpoverty level as measurements due to lack of socioeconomicinformation in the current surveillance system (Krieger et al.,2002). However, it may be informative to estimate the MAUP-scale and aggregation effect by performing the cluster analysis atdifferent scales of ecological aggregation using multilevel analysisin tandem with the methods presented here.

There are several limitations to this work. First, there are issuesresulting from the inherent issues with the spatial scan statisticapproach. Because the spatial scan statistic detects clustersutilizing an elliptic scanning window, the resulting clusters mayhave difficulties in detecting irregular-shaped clusters. Thiscreates a higher false positive rate as compared with other clusteralgorithms such as Besag & Newell and Local Moran (Besag andNewell, 1991; Jacquez and Greiling, 2003). This method could beimproved by using flexible scanning windows to detect irregular-shaped clusters (Tango and Takahashi, 2005), which is not feasiblefor larger cluster sizes. The spatial scan statistic utilizes centroidsto represent polygon features which could inflate the cluster sizedue to the edge effects. Instead of employing centroids ofgeographic units, cases and population in a county could beredistributed according to the percentage area included inscanning windows. In lieu of a small number of cases in the otherracial group, the significant cluster may be falsely exaggeratedusing the spatial scan statistic approach, even under the Poissondistribution. Despite these limitations, the scan statistic stillprovides an appropriate and efficient exploratory tool to identifythe geographic variation in disease patterns. In the cases wherewe detected clusters that can not be explained by randomoccurrences, questions about the underlying causal mechanismsneed to be answered. Second, in order to assess the effect of SESon clusters of breast cancer mortality, we were required to makeassumptions about the temporal nature of the data. We usedquartiles of SES from the 2000 census as the indicator variable, asthe logistic regression model did not adjust for time–time hadalready been accounted for in the space-time cluster analysis andthe dependent variable of the model was coded as a clusterpresent (1) or absent (0) over the time period when the clusterwas detected. In assessing the effect of SES on clusters of breastcancer mortality, we were required to make several assumptionsabout the temporal nature of the data. We used quartiles of SESfrom the 2000 census as the indicator variable over the entirestudy period. To test the validity of this measure over other years,

ARTICLE IN PRESS

N. Tian et al. / Health & Place 16 (2010) 209–218 217

we ran a non-parametric ANOVA test that indicated no significantdifference (p=0.975) between the two years (1990 and 2000)when the full data was available for all counties in the US. Lastly,there are a range of established risk factors for breast cancer thatwe did not control for, namely of income, education, alcoholconsumption, obesity and environmental exposure (Dupont andPage, 1985; Madigan et al., 1995). The above risk factors may leadcertain areas to be more susceptible to higher risk. The currentstudy did not control for these additional factors with theexception of age and race, as the data were not available.However, we did take into account risk factors such as socio-economic status, mammography facilities and regulated facilitiesby EPA in the multivariate logistic regression model. Theadjustment for additional factors would be an intriguing arenaof future research exploration on breast cancer outcomes,especially in more localized studies where the data may be easierto obtain.

5. Conclusion

This study was focused on detecting statistically significantclusters with excessive mortality rates of female breast canceracross age and racial groups at the county level in the UnitedStates for the years 1984–2004. In addition to the confirmation ofexcessive mortality in the Northeast Region, a multivariate logisticregression model was developed to offer further insight into thedetected cluster by incorporating other risk factors. The geo-graphic variation of breast cancer mortality may provide greaterquantitative evidence for optimal intervention programs andeffective health resources allocation so that the underservedsubpopulations may approach similar health outcomes as thoseexperienced by whites. Future research could be focused onincorporating spatial patterns of racial disparities into clusterswithin racial groups with the aim of strengthening our under-standing of the nature of these hotspots. Overall efforts should beput into place to explore and investigate modifiable factorsresponsible for racial/ethnic and geographic disparities under-pinning various breast cancer continuums ranging from incidence,survival, stage-diagnosis, through to mortality. Due to thecomplexities of risk factors associated with mortality, futureresearch on incidence and survival may provide more informativeinsights into modifying risk factors of developing breast cancerand offer solid evidence of attributing primary prevention, earlierdiagnosis and better treatment to improved survival. Ultimately,the aim of this continued work would be improving breast cancermortality rates among all women, regardless of age, socio-economic status or race.

Acknowledgements

Data used in the analysis were from Surveillance, Epidemiol-ogy and End Results (SEER); program (www.seer.cancer.gov)SEERnStat Database: National Cancer Institute, DCCPS, Surveil-lance Research Program, Cancer Statistics Branch, released April2007. Underlying mortality data provided by NCHS (www.cdc.gov/nchs). We appreciate the thoughtful input on the manuscript fromPierre Goovaerts at BioMedare. F. Benjamin Zhan’s work was inpart supported by the Chang Jiang Scholar Awards Program.

References

Anderson, W.F., Chatterjee, N., Ershler, W.B., Brawley, O.W., 2002. Estrogen receptorbreast cancer phenotypes in the Surveillance, Epidemiology, and End Resultsdatabase. Breast Cancer Research and Treatment 76 (1), 27–36.

Berry, D.A., Cronin, K.A., Plevritis, S.K., Fryback, D.G., Clarke, L., Zelen, M.,Mandelblatt, J.S., Yakovlev, A.Y., Habbema, J.D.F., Feuer, E.J., 2005. Effect ofscreening and adjuvant therapy on mortality from breast cancer. The NewEngland Journal of Medicine 353 (17), 1784–1792.

Besag, J., Newell, J., 1991. The detection of clusters in rare diseases. Journal of theRoyal Statistical Society Series A 154 (1), 143–155.

Bradley, C.J., Given, C.W., Roberts, C., 2002. Race, socioeconomic status, and breastcancer treatment and survival. Journal of the International Cancer Institute 94(7), 490–496.

Brody, J.G., Rudel, R.A., Michels, K.B., Moysich, K.G., Bernstein, L., Attfield, K.R., Gray,S., 2007. Environmental pollutants, diet, physical activity, body size, and breastcancer. Where do we stand in research to identify opportunities forprevention?. Cancer 109 (12), 2627–2634.

Brody, J.G., Aschengrau, A., McKelvey, W., Swartz, C., Kennedy, T., Rudel, R.A., 2006.Breast cancer risk and drinking water contaminated by wastewater: a casecontrol study. Environmental Health 5, 28.

Brody, J.G., Rudel, R.A., 2003. Environmental pollutants and breast cancer.Environmental Health Perspectives 111 (8), 1007–1019.

Carey, L.A., Perou, C.M., Livasy, C.A., Dressler, L.G., Cowan, D., Conway, K., Karaca, G.,Troester, M.A., Tse, C.K., Edmiston, S., 2006. Race, breast cancer subtypes, andsurvival in the Carolina Breast Cancer Study. Journal of the American MedicalAssociation 295 (21), 2492–2502.

Clarke, C.A., Glaser, S.L., West, D.W., Ereman, R.R., Erdmann, C.A., Barlow, J.M.,Wrensch, M.R., 2002. Breast cancer incidence and mortality trends in anaffluent population: Marin County, California, USA, 1990–1999. Breast CancerResearch Treat 4 (6), R13.

Devesa, S.S., Grauman, D.J., Bolt, W.J., Pennello, G.A., Hoover, R.N., Fraumeni, J.F.,1999. In: Atlas of Cancer Mortality in the United States, 1950–1994. NCI,Bethesda.

Du, X.L., Meyer, T.E., Franzini, L., 2007. Meta-analysis of racial disparities in survivalin association with socioeconomic status among men and women with coloncancer. Cancer 109 (11), 2161–2170.

Dupont, W.D., Page, D.L., 1985. Risk factors for breast cancer in women withproliferative breast disease. The New England Journal of Medicine 312 (3),146–151.

Farrow, D.C., Hunt, W.C., Samet, J.M., 1992. Geographic variation in the treatment oflocalized breast cancer. The New England Journal of Medicine 326 (17),1097–1101.

Fotheringham, A.S., Zhan, F.B., 1996. A comparison of three exploratory methodsfor cluster detection in spatial point patterns. Geographical Analysis 28 (3),200–218.

Frier, Warrier. Environmental Scientist at the Department of Health RadiationProgram, North Dakota. E-mail personal communication with Nancy Tian.September 1, 2009. Environmental Scientist at Department of Health RadiationProgram, North Dakota.

Garland, C.F., Gorham, E.D., Young, J.F., 1990. Geographic variation in breast cancermortality in the United States: a hypothesis involving exposure to solarradiation. Preventive Medicine 19 (6), 614–622.

Goodwin, J., Freeman, J., Freeman, D., Nattinger, A., 1998. Geographic variations inbreast cancer mortality: do higher rates imply elevated incidence or poorersurvival?. American Journal of Public Health 88 (3), 458–460.

Gordon, N.H., 2003. Socioeconomic factors and breast cancer in black and whiteamericans. Cancer and Metastasis Reviews 22 (1), 55–65.

Greenlee, R.T., Hill-Harmon, M.B., Murray, T., Thun, M., 2001. Cancer statistics,2001. CA: A Cancer Journal for Clinicians 51 (1), 15–36.

Hjalmars, U., Kulldorff, M., Gustafsson, G., Nagarwalla, N., 1996. Childhoodleukemia in Sweden: using GIS and a spatial scan statistic for cluster detection.Statistics in Medicine 15 (7-9), 707–715.

Hsu, C.W.E., Jacobson, H., Mas, F.S., 2004. Evaluating the disparity of female breastcancer mortality among racial groups—a spatiotemporal analysis. Interna-tional Journal of Health Geographics 3 (4), 1–11.

Hsu, J.L., Glaser, S.L., West, D.W., 1997. Racial/Ethnic differences in breast cancersurvival among San Francisco Bay Area women. Journal of the National CancerInstitute 89 (17), 1311–1312.

Jacquez, G.M., Greiling, D.A., 2003. Local clustering in breast, lung and colorectalcancer in Long Island, New York. International Journal of Health Geographics 2(3).

Jemal, A., Siegel, R., Ward, E., Murray, T., Xu, J., Smigal, C., Thun, M.J., 2006. Cancerstatistics, 2006. CA: A Cancer Journal for Clinicians 56 (2), 106–130.

Jemal, A., Kulldorff, M., Devesa, S.S., Hayes, R.B., Fraumeni, J.F., 2002. A geographicanalysis of prostate cancer mortality in the United States, 1970–89. Interna-tional Journal of Cancer 101 (2), 168–174.

Jenks, S., 1994. Researchers to comb Long Island for potential cancer factors.Journal of the National Cancer Institute 86 (2), 88–89.

Kelsey, J.L., Gammon, M.D., John, E.M., 1993. Reproductive factors and breastcancer. Epidemiologic Reviews 15 (1), 36–47.

Kettles, M.K., Browning, S.R., Prince, T.S., Horstman, S.W., 1997. Triazine herbicideexposure and breast cancer incidence: an ecologic study of Kentucky counties.Environment Health Perspectives 105, 1222–1227.

Krieger, N., Chen, J.T., Waterman, P.D., Soobader, M.J., Subramanian, S.V., Carson, R.,2002. Geocoding and monitoring of US socioeconomic inequalities in mortalityand cancer incidence: does the choice of area-based measure and geographiclevel matter? The public health disparities geocoding project. AmericanJournal of Epidemiology 156 (no. 5), 471.

Kulldorff, M., 2006. SaTScan User Guide. SaTScan Version 7.0: Software for theSpatial and Space-time Scan Statistics.

ARTICLE IN PRESS

N. Tian et al. / Health & Place 16 (2010) 209–218218

Kulldorff, M., 1998. Statistical methods for spatial epidemiology: tests forrandomness. In: Gatrell, A.C., Loytonen, M. (Eds.), GIS and Health. Taylor &Francis, London, pp. 49–62.

Kulldorff, M., Athas, W.F., Feurer, E.J., Miller, B.A., Key, C.R., 1998. Evaluating clusteralarms: a space-time scan statistic and brain cancer in Los Alamos, NewMexico. American Journal of Public Health 88 (9), 1377–1380.

Kulldorff, M., 1997. A spatial scan statistic. Communications in Statistics—Theoryand Methods 26 (6), 1481–1496.

Kulldorff, M., Feuer, E.J., Miller, B.A., Freedman, L.S., 1997. Breast cancer clusters inthe Northeast United States: a geographic analysis. American Journal ofEpidemiology 146 (2), 161–170.

Knox, E.G., Elliot, P., 1989. In: Methodology of Enquiries into Disease Clustering.Small Area Health Statistics Unit, London, pp. 17–20.

Lewis-Michl, E.L., Melius, J.M., Kallenbach, L.R., Ju, C.L., Talbot, T.O., Orr, M.F.,Lauridsen, P.E., 1996. Breast cancer risk and residence near industry or traffic inNassau and Suffolk Counties, Long Island, New York. Archives of EnvironmentalHealth 51 (4), 255–265.

Madigan, M.P., Ziegler, R.G., Benichou, J., Byrne, C., Hoover, R.N., 1995. Proportion ofbreast cancer cases in the United States explained by well-established riskfactors. Journal of the National Cancer Institute 87 (22), 1681–1685.

McPherson, K., Steel, C.M., Dixon, J.M., 2000. Breast cancer-epidemiology, riskfactors, and genetics. British Medical Journal 321 (7261), 624–628.

Morabia, A., Flandre, P., 1992. Misclassification bias related to definition ofmenopausal status in case-control studies of breast cancer. InternationalJournal of Epidemiology 21 (2), 222–228.

Openshaw, S., Charlton, M.E., Wymer, C., Craft, A., 2006. A Mark 1 GeographicalAnalysis Machine for the automated analysis of point data sets. Classics fromIJGIS: Twenty Years of the International Journal of Geographical InformationScience and Systems 1 (4), 335–358.

Openshaw, S., 1984. In: The Modifiable Areal Unit Problem. Geo Books, Norwich,UK.

Paffenbarger, R.S., Kampert, J.B., Chang, H.G., 1980. Characteristics that predict riskof breast cancer before and after the menopause. American Journal ofEpidemiology 112 (2), 258–268.

Pickle, L.W., Mason, T.J., Howard, N., 1987. Atlas of US Cancer Mortality amongWhites: 1950–1980. Department of Health and Human Services, Washington,DC.

Ries, L., Eisner, M.P., Kosary, C.L., 2000. In: Cancer Statistics Review 1973–1997.National Cancer Institute, Bethesda.

Riggan, W., Bruggen, J.V., Acquavella, J., Beaubier, J., Mason, T., 1983. In: US CancerMortality Rates and Trends, 1950–1979. NCI/EPA Interagency Agreement onEnvironmental Carcinogenesis, Washington, DC.

Rushton, G., Lolonis, P., 1996. Exploratory spatial analysis of birth defect rates in anurban population. Statistics in Medicine 15 (7-9), 717–726.

Sarker, M., Jatoi, I., Becher, H., 2007. Racial differences in breast cancer survival inwomen under age 60. Breast Cancer Research and Treatment 106, 135–141.

Saunders, L.D., Chen, W., Hrudey, S.E., 1997. Studies of hazardous waste sites andhuman health: evaluation of internal validity. Environment Review 5 (3-4),167–180.

Sheehan, T.J., DeChello, L.M., Kulldorff, M., Gregorio, D.I., Gershman, S., Mroszczyk,M., 2004. The geographic distribution of breast cancer incidence inMassachusetts 1988 to 1997, adjusted for covariates. International Journal ofHealth Geographics 3 (17).

Singh, G.K., Miller, B.A., Hankey, B.F., Edwards, B.K., 2003. In: Area SocioeconomicVariations in U.S. Cancer Incidence, Mortality, Stage, Treatment, and Survival,1975–1999. NCI Cancer Surveillance Monograph Series, No. 4. NCI, Bethesda.

Smigal, C., Jemal, A., Ward, E., Cokkinides, V., Smith, R., Howe, H.L., Thun, M., 2006.Trends in breast cancer by race and ethnicity: update 2006. CA: A CancerJournal for Clinicians 56, 168–183.

Sturgeon, S.R., Schairer, C., Grauman, D., Ghormli, E.l., Devesa, S., 2004. Trends inbreast cancer mortality rates by region of the United States, 1950-1999. CancerCauses Control 15 (10), 987–995.

Sturgeon, S.R., Schairer, C., Gail, M., McAdams, M., Brinton, L.A., Hoover, N.R., 1995.Geographic variation in mortality from breast cancer among white women inthe United States. Journal of National Cancer Institute 87, 1853–1864.

Surveillance, Epidemiology, and End Results (SEER) Program (www.seer.cancer.gov) SEERnStat Database: Mortality—All COD, Aggregated With County, TotalUS (1969–2004)—Linked To County Attributes—Total US, 1969–2004 Counties,National Cancer Institute, DCCPS, Surveillance Research Program, CancerStatistics Branch, released April 2007. Underlying mortality data provided byNCHS (www.cdc.gov/nchs).

Surveillance Research Program. 2008. National Cancer Institute SEERnStat software(www.seer.cancer.gov/seerstat) version 6.4.4.

Tango, T., Takahashi, K., 2005. A flexibly shaped spatial scan statistic for detectingclusters. International Journal of Health Geographics 4, n.p.

Taplin, S.H., Ichikawa, L., Buist, D.S., Seger, D., White, E., 2004. Evaluating organizedbreast cancer screening implementation: the prevention of late-stage disease?.Cancer Epidemiology Biomarkers & Prevention 13, 225–234.

Wang, M., Burau, K.D., Fang, S., Wang, H., Du, X.L., 2008. Ethnic variationsin diagnosis, treatment, socioeconomic status, and survival in a largepopulation-based cohort of elderly patients with non-hodgkin lymphoma.Cancer 113 (11).

Yost, K., Perkins, C., Cohen, R., Morris, C., Wright, W., 2001. Socioeconomic statusand breast cancer incidence in California for different race/ethnic groups.Cancer Causes Control 12, 703–711.

Zhan, F.B., Lin, H., 2003. Geographic patterns of cancer mortality clusters in Texas,1990 to 1997. Texas Medicine 99, 58–64.

Zhan, F.B., 2002. Monitoring geographic concentration of female breast cancerusing cluster analysis: the case of New Mexico. Papers and Proceedings ofApplied Geography Conferences 25, pp. 1–8.