International Journal of Cancer Therapy and Oncology www.ijcto.org Corresponding author: Getachew Dagne; Department of Epidemiology and Biostatistics, University of South Florida, Tampa, USA. Cite this article as: Dagne G, Odedina F, Aime N, Young M. Area-level factors associated with spatial variation of prostate cancer incidence for black men.. 2017; 5(1):5123. DOI: 10.14319/ijcto.51.23 © Dagne et al. ISSN 2330-4049 Area-level factors associated with spatial variation of prostate cancer incidence for black men Getachew Dagne1, Folakemi Odedina2, Nickyjeanna Aime3, Mary Ellen Young4 1Department of Epidemiology and Biostatistics, University of South Florida, Tampa, USA 2Department of Pharmacotherapy & Translational Research, University of Florida Research and Academic Center, Orlando, USA 3Department of Radiation Oncology, Florida A&M University, Tallahassee, USA 4Department of Occupational Therapy, University of Florida, Gainesville, USAReceived August 23, 2017; Revised December 17, 2017; Accepted December 20, 2017; Published Online December 24, 2017 Original Article Abstract Purpose: Black men are disproportionately affected by prostate cancer (CaP)compared to any other racial/ethnic groups within the United States. IdentifyingCaP hotspots along with associated local area-level risk factors is crucial to tacklingthe significant burden of CaP and the disparity seen in Black men. The objective ofthis study was to determine the scope of geographical variation in CaP incidencesand to assess the degree to which this variation is associated with county -level riskand protective factors. Methods: The study population was Black men diagnosedwith prostate cancer between 2006-2010 in Florida. County-level CaP incidencerates were computed as the ratios of the numbers of new CaP cases diagnosedbetween 2006 and 2010 to the corresponding 2000 US census population of Blackmen 20 and over years old data (US Census 2000). Other county-levelenvironmental and health care factors were also obtained. A random effectsPoisson model and Geographical Information System (GIS) were used to map andassess the spatial patterns of CaP incidences in 67 Florida counties. Thesestatistical techniques involved a Bayesian approach for estimating the underlyingcounty-specific CaP risk since the data are very sparse. Results: The findingsshowed that an increasing CaP incidence of Black Men in Florida was significantlyassociated with an increasing unemployment rate ( 2 = .1379 with 95% CI:(.0025, .2703), does not include zero suggesting significance) and with increasingnumber of physicians per capita after controlling for other county characteristics.There was a negative association between poverty and CaP incidence. Regardingspatial distribution of CaP incidence, we observed that there are clustering andhotspots of high CaP incidence rates in Palm Beach county in South Florida, andAlachua and Marion counties in north Florida. Conclusion: Our findings showedthat indicators of socioeconomic status and accessibility of health care servicessuch as poverty, unemployment and health care providers are important variable sthat explain spatial variation of prostate cancer incidence rates of Black Men.Better understanding of such risk factors and identifying specific counties with adisproportionate burden of CaP disease may help formulate targeted interventionsand resource allocation by state and local public officials. Keywords: Bayesian inference, Health disparity, Prostate cancer, Poisson model. 1. IntroductionProstate cancer (CaP) is one of the most commoncancers experienced by men in the United States (US),and the second leading cause of cancer-related deaths.1Black men are disproportionately affected by CaPcompared to any other racial/ethnic groups in the US.Compared to US White men, Black men are about twotimes more likely to develop CaP and die from the disease.1 Although the causes for these disparities arenot yet completely known, genetic heritage, variation inlife styles, health care availability, environmental riskfactors have been suggested as plausible explanations.2-8To examine the influence of environmental risk factorson CaP incidence in a geographic context, the studyobjectives were: (1) to estimate the association between 2 Dagne et al.: Spatial Variation of Prostate Cancer Incidence International Journal of Cancer Therapy and Oncology www.ijcto.org © Dagne et al. ISSN 2330-4049 county-specific relative risk for prostate cancer andcounty-level characteristics such as socio-economicstatus, health care access, poverty, unemployment andwater supply, and (2) to develop spatial mapping of CaPincidence for Black men in Florida.Although some studies have examined the relationshipbetween environmental factors and cancer incidencespatial variations9-14, there is limited publications on thespatial pattern variations of CaP incidence in FloridaBlack men. Knowledge of the spatial distribution of CaPincidence has significant public health implication. Forexample, the CaP burden can be mitigated throughidentifying major health determinants, and allocatingproper public health resources and policies at the locallevel. For this study, we examined the associationbetween spatial variations in CaP incidence and thefollowing county-level environmental and health carefactors: availability of physicians, body weight,environmental exposures, demographic indicators andsocio-economic indicators.15 Specifically, determiningthe role of spatial, environmental, and socio-economicheterogeneity in prostate cancer disparities provides abasis for developing public health interventions that willprevent and control prostate cancer in affectedcommunities.In this paper, we used Bayesian spatial models todescribe the spatial pattern of CaP incidence for Blackmen in Florida's 67 counties. In addition, we assessedthe contribution of socioeconomic, environmental, andhealth care availability in explaining area-levelvariations. 2. Methods and Materials 2.1. Study setting and sources of dataThe study setting was Florida and the targetedpopulation was all Black men diagnosed with CaPbetween 2006 and 2010. County-level CaP cases wereobtained from the Florida Cancer Data System (FCDS)database which is Florida's legislatively mandated,population-based, statewide cancer registry.16County-level CaP incidence rates are computed as theratios of the numbers of new CaP cases diagnosedbetween 2006 and 2010 to the corresponding 2000 UScensus population of Black men 20 and over years olddata (US Census 2000). The 2000 US Census is chosen sothat presumed exposures occurred before CaP diagnosisin 2006-2010, the study period. The Countycharacteristics that may be associated with CaPincidence were identified from the Florida Departmentof Health Division of Public Health Statistics &Performance Management (see Table 1). Some of thesecharacteristics are socio-economic indicators (e.g.percentage of unemployed adults, high schoolgraduation), demographic indicators (e.g. percentage ofindividuals with rural residence), health care resources(e.g. licensed Florida physicians; adults who could notsee a doctor at least once in the past year due to cost),environment indicator (e.g. community water supply).We used the most representative data available for thesecounty-level characteristics for the period of 2006-2010. Table 1: Description statistics of county-level characteristicsCharacteristics Mean SD Min MaxProstate cancer cases 163.6 347.512 1.0 1963.0Unemployed for Yr. 2008 (%) 6.209 1.332 4.000 10.20Median income for Yr. 2009 43960 7554.062 29640 63630Number of physicians for Yr. 2008 (per 100,000)High school graduate for Yr. 2009 (%)Below poverty level for Yr. 2009 (%)Two or more servings of fruit for Yr. 2007 (%)Current smoker for Yr. 2007 (%)Medical checkup for Yr. 2007 (%)Overweight for Yr. 2010 (%)Community water supply rate for Yr. 2010Black population for Yr. 2010 (%)Not seek medical due to cost for Yr. 2007 (%)Rural resident for Yr. 2010 (%) 139.780.5815.4832.4922.2366.8766.892.189014.5915.2442.02 98.3618.0554.8615.6104.7757.0255.7230.90869.3466.03733.762 12.658.607.4018.5014.2047.3054.300.51683.106.200.10 615.296.5026.4046.1033.6079.8082.005.30555.2043.30100.00 Volume 5• Number 1 • 2017 International Journal of Cancer Therapy and Oncology 3 www.ijcto.org © Dagne et al. ISSN 2330-4049 2.2. Statistical methodsWe considered a geographical region divided into Gcontiguous small areas (e.g., counties) represented as=1,…, . Let denote observed counts of diseasecases (e.g., prostate cancer) and a q-dimensional vectorcontains county-level covariates with associatedparameters . We assumed that follows a Poissondistribution with mean satisfyinglog( )= log( )+ log( ) (1)where is the expected number of cases in the ithcounty, and calculated as = (∑∑ ); is numberof individuals at risk of prostate cancer; and is anunknown county-specific relative risk of prostate cancerand further decomposed aslog( )= + + (2)In Model (2), the county-specific random effects, =+ , was further decomposed into an unstructuredheterogeneity and a spatially structured localrandom effects to account for the tendency ofneighboring counties to have similar relative risksbecause of sharing common risk factors.17Specifically, for the Florida prostate cancer cases forBlack men, the log of the relative risk was modeled as    1 2 3 4 5 6 7 8 9 10 log log         i i i i i i i i i i Income Unemploy Poverty Overweight Smoker WaterSupply PrcntBlack MedicalCheckup FruitConsumption Edu                        11 12                 (3)                    i i i i i cation Physician Rural u v       The covariates in (3) were defined in Section 2.1 andTable 1. The above random-effect Poisson regression modelswere used to produce smoothed spatial maps of CaPincidence rates by incorporating the associationsbetween incidence and county-level covariates. Therelative risk in each county was estimated using aBayesian approach based on Markov chain Monte Carlo(MCMC) methods which were implemented in WinBUGSsoftware.18 WinBUGS has a built-in conditionalautoregressive (CAR) distribution for handling spatialautocorrelation. Non-informative prior distributionswere used for the unknown parameters of (3), andsensitivity analyses with different prior specifications were conducted to assess the effect of choices of vaguepriors. 3. ResultsBased on the FCDS, a total of 10,799 Black men werediagnosed with prostate cancer between 2006 and 2010in Florida. The map in Figure 1 shows the number ofprostate cancer cases per County, with the lowest inDixie County and highest in Miami-Dade and BrowardCounties. There is a strong variation in geographicaldistributions of these CaP cases. The variation may bedue to some counties having low cases, sparse sizes ofpopulation of adult Black men, or both. To incorporatethe variation in population sizes across counties, wecalculated the expected number of CaP cases for eachcounties as = (∑ /∑ ). Then, thestandardized morbidity ratio (SMR) was computed asthe ratio of the number of observed cases ( ) toexpected number of cases ( ) for each county. TheseSMRs were mapped in Figure 2. The changes fromobserved cases to SMR are most striking in Charlotteand Levy counties, showing that the CaP cases (4 and 21,respectively) in these counties are very small. Thespatial pattern variation across the counties suggeststhat there is local instability in both observed counts andSMR since they do not take into consideration forsampling errors. A solution for filtering the signal fromthe random noise is to use statistical methods byintroducing random effects and county-level covariatesto explain such strong heterogeneity across counties.Random-effect Poisson regression models described in(Ref# 1,3) were fitted to the observed data to getgeographical maps of county-specific relative risks ofCaP and assess the associations between county-specificrelative risks and county-level covariates given in Table1. The posterior means, standard deviations and 95%credible interval (CI) of the coefficients of the covariatesare presented in Table 2. The results show that anincreasing CaP incidence of Black Men in Florida issignificantly associated with increasing unemploymentrate ( 2 =.1379 with 95% CI: (.0025, .2703), which doesnot include zero) and with increasing number ofphysicians per capita in a county ( 11 =.00212 with 95%CI: (.00006, .0042) after controlling for other countycharacteristics. This implies that the more the number ofphysicians in a county, the higher CaP diagnosed casesdue to accessibility to health services. In the case ofpoverty, however, there is an inverse relationshipbetween CaP incidence and percent of adult individualswho were below poverty level in 2009 in a county. Thatis, a decreasing CaP incidence of Black Men in Florida issignificantly associated with increasing percentage ofpersons below poverty level ( 3 =-.0583 with 95% CI:(-.1039, -.0132), which confirms findings of otherstudies.19 4 Dagne et al.: Spatial Variation of Prostate Cancer Incidence International Journal of Cancer Therapy and Oncology www.ijcto.org © Dagne et al. ISSN 2330-4049 Figure 1: Spatial distribution of prostate cancer cases for Black Men in Florida (2006 -2010) Figure 2: Spatial distribution of ratios of observed and expected cases for Black Men in Florida (2006 -2010). Volume 5• Number 1 • 2017 International Journal of Cancer Therapy and Oncology 5 www.ijcto.org © Dagne et al. ISSN 2330-4049 Table 2: A summary of the estimated posterior mean (PM) and standard deviation (SD) of population parameters andlower limit ( ) and upper limit ( ) of $95% equal-tail credible interval (CI).Predictor Parameter PM SD )InterceptMedian Income -.1538.1469 .04399.4574 -.2434-.7498 -.0681.045Unemployment .1379 .0677 .0025 .2703Poverty -.0583 .0232 -.1039 -.0132OverweightCurrent SmokerCommunity Water SupplyBlack PopulationMedical CheckupTwo or more FruitHigh School GraduationPhysicianRural Resident .0079.0304.0096.0104.0026.0080-.0131.0021.0023 .0144.0167.0895.0091.0115.0143.0099.0010.0048 -.0205-.0027-.1661-.0073-.0204-.0198-.0326.00006-.0074 .0362.0635.1846.0283.0251.0364.0063.0042.0116 Figure 3: Spatial distribution of posterior medians of standardized morbidity ratios of prostate cancer for BlackMen in Florida (2006-2010). A byproduct of the random-effect Poisson regressionmodel is the estimated CaP relative risk in each countyafter adjusting for the effect of county-levelcharacteristics. The posterior median of the smoothedCaP relative risk was mapped in Figure 3 which showsthe spatial pattern inherent in the observed cases (seeFigure 1). Looking at the map in Figure 3, we observethat there are clustering and hotspots of high CaPincidence rates in Palm Beach county in South Florida, and Alachua and Marion counties in north Florida. Atleast 60% of the counties in Florida exhibitdisproportional burden of prostate cancer by havingmore than expected relative risk ( >1 ). Thus, furtherinvestigation into identifying and understandingunderlying causal mechanisms in the communities isparamountly significant for reducing the burden of thisdisease. Specifically, targeted interventions can also be 6 Dagne et al.: Spatial Variation of Prostate Cancer Incidence International Journal of Cancer Therapy and Oncology www.ijcto.org © Dagne et al. ISSN 2330-4049 designed for those counties with high prostate cancerrelative risks. 4. DiscussionIn this spatial study, we assessed the link between thegeographical variation of CaP incidence for Black men inFlorida and potential County-level risk factors. Theresults show that County- specific CaP relative ratios arehigher in counties where there are higher proportion ofunemployed, higher number of Florida licensedphysician, and lower proportion of persons belowpoverty level. Although not statistically significant atcounty level, median income, percentage of overweight,percentage of current smoking status, community watersupply per capita, percentage of Blacks, percentage ofmedical checkup, percentage of persons consuming twoor more fruits daily, high school graduation, andpercentage of rural residents have positive associationwith prostate cancer incidence. These findings are alsoshown in some other studies.20,21After adjusting for County-level characteristics, thesmoothed CaP incidence for Black men was used toidentify Counties with higher or lower than expectedratios (see Figure 3) if every County is equally likely tohave CaP cases. Accordingly, some Counties in northeast,central and south Florida tend to have higher CaPincidence than expected. These findings suggest thatmore detailed study of CaP incidence in Counties withhigher concentration of cases is warranted. In addition,looking into variation within the black ethnicity such asUS-born, Caribbean-born and Africa-born may throwlight on endogenous and exogenous healthdeterminants, which are unique to each subgroup.It is noted that, as in any ecological study, caution needsto be taken when interpreting ecological analysisresults.22 This is because associations assessed betweenrisk factors and CaP incidence at a county level may notnecessarily imply that the risk factors are associatedwith an individual's chance of having CaP. Unmeasuredconfounders (e.g., prostate-specific antigen (PSA) ordigital rectal exam (DRE) screening) are potentialsources of discrepancies between results of county leveland individual level analysis.23,24 Thus, the goal of thisarticle is to investigate risk factors that may contributeto the geographic pattern of CaP incidence of Black menwithin Florida using a Bayesian approach.The Bayesian method was chosen since it is flexible toincorporate a spatially structured variation via aconditional autoregressive function, accounting forspatial dependence of adjacent neighbors, andheterogeneity.25,26 The Bayesian method uses MCMC toestimate the parameters of the Poisson random effectsmodel based on non-informative prior distributions forcoefficients of covariates, spatial and heterogeneityparameters. Furthermore, the estimation process can be easily carried out using the publicly available WinBUGSpackage.18 This makes our approach quite powerful andaccessible to practitioners in the field.There are some limitations to our study. The currentstudy has a spatial dimension only since aggregate dataover the 2006-2010 study periods were used butignores the temporal feature of the observed cases. Thereason is that the observed cases are very sparse atcounty level for each year in the study period and thusnot enough data for analyzing temporal trend. Forexample, 14 out of the 67 counties have less or equal to10 cases aggregated over the 5 year period. Thecounty-level covariates chosen for the analysis arelimited by the availability of data on importantprotective and risk factors for CaP. Other measures ofenvironmental exposure, diet intake, socio-economicand demographic characteristics of the 67 countiesshould be considered in future analysis. 5. ConclusionThis study shows that county-level indicators ofsocioeconomic background and health care servicessuch as number of physicians explain spatialheterogeneity of prostate cancer incidence rates. Betterunderstanding of such risk factors and identifyingspecific counties with a disproportionate burden of CaPdisease may help formulate targeted interventions andresource allocation by state and local public officials. Infuture, given availability of data, further analysisfocusing on geographic variation of treatment modalityand mortality will be useful. Conflict of interestThe following teams and people are acknowledged fortheir support on this study: Florida Department ofHealth Cancer Epidemiology Office, University of FloridaCollege of Pharmacy IT team, Florida Cancer DataSystem, Mr. Cameron Schiller, Dr. Jenn Nguyen and MsHannah Asfaw. FundingThis study is funded by the Department of Defense PCRPAward W81XWH1310473. References1. American Cancer Society (ACS). Cancer facts &figures 2016. American Cancer Society, Atlanta;2016.2. Powell IJ. Prostate cancer and African-Americanmen. Oncol. 1977;11:599-618.3. Powell IJ. Prostate cancer in theAfrican-American: is this a different disease?Semin Urol. 1988;16:221-6. Volume 5• Number 1 • 2017 International Journal of Cancer Therapy and Oncology 7 www.ijcto.org © Dagne et al. ISSN 2330-4049 4. Merrill RM, Brawley OW. Prostate cancerincidence and mortality rates among white andblack men. Epidemiology. 1997;8:126-31.5. Brawley OW, Knopf K, Thompson I. Theepidemiology of prostate cancer Part II: the riskfactors. Semin Urol Oncol. 1988;16:193-201.6. Montie JE, Pienta KJ. A unifying model toexplain the increased incidence and highermortality of prostate cancer in black men.Urology. 1999;53:1073-6.7. Du XL, Fang S, Coker AL et al. Racial disparityand socioeconomic status in association withsurvival in older men with local/regional stageprostate carcinoma: findings from a largecommunity-based cohort. Cancer.2006;106:1276-85.8. Goovaerts P, Xiao H, Gwede CK et al. Impact ofage, race and socio-economic status ontemporal trends in late-stage prostate cancerdiagnosis in Florida. Spat Stat. 2015;14:321-37.9. Toledano M, Jarup L, Best N, et al. Spatial andtemporal trends of testicular cancer in greatBritain. Br J Cancer. 2001;84:1482-7.10. Jarup L, Toledano MB, Best N, et al.Geographical epidemiology of prostate cancerin great Britain. Int J Cancer. 2002;97:695-9.11. Boffetta P, Nyberg F. Contribution ofenvironmental factors to cancer risk. Br MedBull. 2003;68:71-94.12. Waller LA , Gotway CA. Applied spatial statisticsfor public health data. New York: John Wiley,2004.13. Goovaerts P, Xiao H. Geographical, temporaland racial disparities in late-stage prostatecancer incidence across Florida: a multiscalejoinpoint regression analysis. Int J HealthGeogr. 2011;10:63.14. Xiao H, Tan F, Goovaerts P. Racial andgeographic disparities in late-stage prostatecancer diagnosis in Florida. J Health Care PoorUnderserved. 2011;22(4):187-99.15. Signorello LB, Adami H. Prostate Cancer. InTextbook of Cancer Epidemiology Edited by: Adami H, Hunter H, Trichopoulos D. Oxford,Oxford University Press; 400-428. 2002.16. Hernandez MN, Fleming LE, MacKinnon JA, et al. Cancer in Florida Persons of African Descent1988-2007. Miami: Florida Cancer Data System;2010.17. Lawson A, Bohning D, Biggeri A. DiseaseMapping and Risk Assessment for PublicHealth. 2010.18. Lunn DJ, Thomas A, Best N et al. WinBUGS -aBayesian modelling framework: concepts,structure, an extensibility. Stat and Computing.2000;10:325-7.19. Boscoe F P, Johnson C J, Sherman RL et al. TheRelationship between area poverty rate andsite-specific cancer incidence in the UnitedStates. Cancer. 2014;120(14):2191-8.20. Liu L, Cozen W, Bernstein L, et al. Changingrelationship between socioeconomic status andprostate cancer incidence. J Natl Cancer Inst.2001;93(9):705-9.21. Rundle A, Neckerman KM, Sheehan D et al. Aprospective study of socioeconomic status,prostate cancer screening and incidence amongmen at high risk for prostate cancer. CancerCauses Control. 2013; 24(2):297-303.22. Freedman DA. Ecological inference and theecological fallacy. International Encyclopediafor the Social and Behavioral Sciences.2001;6:4027-30.23. Greenland S, Morgenstern H. Ecological bias,confounding and effect modification. Int JEpidemiology. 1989; 18:269-74.24. Prentice R, Sheppard L. Aggregate data studiesof disease risk factors. Biometrika. 1995;82:113-25.25. Lawson A. Statistical methods in spatialepidemiology. 2001.26. Best N, Richardson S, Thomson A. A comparisonof Bayesian spatial models for disease mapping.Stat Methods Med Res. 2005;14:35-59.