Introduction

Cancer likely results from complex interactions of factors at the macro-environmental, individual, and biologic levels[1]. Identifying relevant factors within each level for joint studies is a challenge, particularly at the macro-environmental level, defined here by the neighborhood in which a person lives. Studies of cancer that evaluate neighborhood consider a limited number of variables based on prior knowledge; this affects comparability and consistency across studies and makes etiologic inferences difficult. A lack of empirical assessment is a well-cited limitation of neighborhood studies[2], and empiric approaches from biology could be applied to the macro-environmental level. For example, genome-wide association studies (GWAS) have driven population-based cancer research for the past several years [3]. GWAS use high-throughput, low-cost technology and readily available genome-mapping to evaluate the role of millions of genetic markers for a variety of diseases using an agnostic approach[4]. These approaches are hypothesis-generating, and the clinical implications of GWAS are starting to have translational impact[5].

Applying concepts from GWAS, environmental-wide association studies (EWAS) were subsequently developed to study the effect of exposures at the individual level (e.g., pesticides), and to provide insights for gene-environment interaction studies[6]. However, neighborhood factors have not been comprehensively studied using these approaches. Borrowing concepts from GWAS and EWAS, we propose the neighborhood-wide association study (NWAS) as a novel, empirical approach to evaluate the effect of multiple neighborhood-level exposures on disease outcomes and to address gaps in neighborhood research. The objective of this method is to apply informatics approaches to the study of neighborhood through the systematic identification of neighborhood factors that may be associated with disease phenotypes. With NWAS, we aim to generate hypotheses in order to inform gene-environment studies and potentially more precisely identify neighborhoods at high risk for poor cancer outcomes.

We introduce the NWAS approach and demonstrate how agnostic, high-dimensional data analyses can be used to identify neighborhood characteristics associated with high grade/high stage, aggressive prostate cancer. There are at least two hypotheses that may explain the role of neighborhood in prostate cancer aggressiveness. First, unfavorable neighborhood environments may exert a biological effect on prostate cancer aggressiveness. Neighborhood environment could affect prostate cancer severity under a chronic stress hypothesis, in which residents from disadvantaged neighborhoods experience greater emotional stress and constant “wear and tear” on the body that can affect cancer initiation and progression[7] [8, 9]. Second, unfavorable neighborhood environments may be correlated with factors related to health care access, particularly screening behaviors and practices. Because screening can detect cancer at earlier stages, people living in less favorable neighborhoods may have less access to care that lead to later (i.e., more aggressive) cancers at the time of diagnosis[10–13]. These two hypotheses are not mutually exclusive of one another and could both be acting through neighborhood-level influences. Given few individual-level risk factors for prostate cancer have been identified[14] and only a few studies have investigated neighborhood effects on prostate cancer using a priori variable selection approaches[10–13], empiric assessments of the effect of neighborhood on aggressive prostate cancer are warranted.

Materials and methods

Results

Of the reported 3,135 census tracts in PA in 2000, 3,037 (97%) census tracts are represented in our study sample (S1 File). Aggressive prostate cancer cases were clustered in urban areas, namely Pittsburgh and Philadelphia (S1 File). The average age of the study population was 69.2 (standard deviation (sd): 9.4) and mean year of diagnosis was 2000. The average age of aggressive cases was 69.8 (sd: 10.4) and of nonaggressive cases was 68.8 (sd: 9.0).

Fig 1 summarizes Phases 1–3 study methods (Fig 1a) and findings (Fig 1b). In Phase 1, from 14, 663 variables, we identified 434 unique census variables significantly associated with prostate cancer aggressiveness at Bonferroni-corrected significance levels (S1 and S4 Files). After Phase two, 217 unique variables were still significant at Bonferroni-corrected credible intervals (S5 File). The average amount of residual variability in aggressive prostate cancer risk across the 217 Bayesian models in Phase 2 was 34% (range: 14%-50%), which is considered large. In Phase 3, 17 uncorrelated principal components were identified from these 217 neighborhood variables. Components 1–8 explain 80% of the variance, and related to poverty (Component 1), white only neighborhood characteristics (Component 2), household/ housing unit poverty status (Component 3), households and living alone (component 4), rented houses built before 1939 (Component 5), civilian population (Component 6), household income above $60K (Component 7), and immigration (Component 8) (Fig 1b). 76 of 217 Phase 2 variables loaded on Component 1 and 51 loaded on Component 2 after the Phase 3 principal components analysis. The top 10% of significant variables from Phase 2 loaded on Component 1 (S5 File).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. a. Study flow chart of NWAS statistical methods b. Overview of study findings by methodological phase.

^a Logit(p) = α+β_i0x_age+ β_i1x_{year of diagnosis} + β_i2x_{neighborhood variable (i, j)} + ε_ij;

where i = individual cancer cases; j = census tracts (Phase 1)

^b Logit(p) = α+β_i0x_age+β_i1x_{year of diagnosis}+β₂x_{neighborhood variable (i, j)} +V_(j) +U_(j)

where i = individual prostate cancer cases; j = county, V_(j) are independent non-spatial random effects and U_(j) are spatially structured random effects (Phase 2).

*These 17 components explain 90% of the variance

https://doi.org/10.1371/journal.pone.0174548.g001

The top 17 most significant variables within each of the 17 principal components are presented in Figs 2 and 3. The most significant variable from Component 1 represented Non-Hispanic Whites aged 6–11 for whom poverty status was determined (OR = 1.07, CI = 1.01–1.12). Seven of the top 17 hits or variables from Components 1–3, 7, 9–10, 15 were related to socioeconomic status. One top hit related to employment, specifically male protective service occupations such as fire-fighting and law enforcement (OR = 0.94, CI = 0.89–0.99), and one related to immigration status (OR = 0.93, CI = 0.87–0.99). Two were associated with physical environment (aggregate income of occupied, rented housing units built 1940–1949 with a householder aged 15–34 (OR = 1.06, CI = 1.01–1.11) and percent (%) renter occupied housing units built 1939 or earlier with householder aged 15–24 years (OR = 1.07, CI = 1.02–1.11). Two variables related to social support (%male householder living alone (OR = 1.06, CI = 1.01–1.11) and %male householder over 65 living alone in nonfamily household (OR = 1.07, CI = 1.02–1.13)). The top hit (most significant variable from Phase 2) was %workers >16 years taking trolley or street car public transportation to work (OR = 1.05, CI = 1.001–1.09).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Phase 3-Principal components and fine mapping analysis to identify top hits.

Dots represent single neighborhood variables from Phase 2 (n = 217 total dots). Open dots are color-coded to their respective component (from Phase 3-Principal Components analysis). Closed-colored dots represent the most significant variable within each component (Phase 3-Fine Mapping) and corresponding statistics are provided by component number in Fig 3. *Top hit based on statistical significance from Phase 2 data. ^a Statistical significance determined by Bonferroni-corrected confidence intervals from the Phase 2 Bayesian model, i.e. smaller credible interval length indicates greater statistical significance.

https://doi.org/10.1371/journal.pone.0174548.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 3. Summary of neighborhood variable “top hits” associated with aggressive prostate cancer by phase.

^aStandard deviation (sd); ^bConfidence or Credible Interval (CI).

https://doi.org/10.1371/journal.pone.0174548.g003

Discussion

We used a novel NWAS to assess the association of 14,663 neighborhood variables with prostate cancer aggressiveness from the PA Cancer Registry. Through a series of progressively more stringent phases, model adjustments, and dimension reduction techniques, we identified the top 17 neighborhood variables associated with aggressive prostate cancer. These findings confirm some previous associations, but also provide new insights into the role of neighborhood in prostate cancer and suggest the potential value of NWAS to inform public health interventions and multilevel studies.

Previous studies of neighborhood and prostate cancer suggest that neighborhoods with poor socioeconomic (SES) circumstances are related to high-grade prostate cancer[29], independent of individual-level exposures[12, 20]. In these studies, SES was measured with deprivation scores and single, a priori selected U.S. census variables related to education, income, poverty, housing[30] and employment (S1 File). Our findings support that neighborhood income and poverty (Components 1, 2, 3, 7, 8, 9, 10, 15), employment (Component 14) and housing (Components 3, 4, 5, 9, 11, 13, 16, 17) relate to prostate cancer aggressiveness. However, neighborhood education was not an important determinant of aggressive prostate cancer here, suggesting education could be a confounder rather than a main effect in neighborhood studies, but more study is warranted.

Immigration (Component 9; %foreign-born naturalized citizens at or above poverty) was a significant NWAS finding. Studies of neighborhoods with higher rates of foreign-born immigrants have shown associations with decreased risk for cancer[31]. Even if individuals are diagnosed with late-stage prostate cancer, survival is improved for those who live in high ethnically homogeneous enclaves, suggesting the strong role social support, alone and in conjunction with poverty, may play in prostate cancer progression[31, 32].

The top hit in this analysis related to taking public transportation to work. This variable, as well as not owning a vehicle, relate more to urban, as compared to rural settings, and are also often used as surrogates for access to medical care[33, 34]. Access to care is often cited as a cause of disparity in prostate cancer treatment[34] and survival[33] across both urban and rural settings. Higher cancer incidence and mortality rates are noted in more urban settings, and cases arising from rural environments often are diagnosed at later disease stages[35]. Thus, NWAS findings are plausible and consistent with previously identified sociodemographic domains[2, 36].

From a methodologic standpoint, NWAS provides a new, agnostic approach to neighborhood and contextual variable selection[2]. In the past, one study might define poverty as proportion of households below poverty, another as percentage receiving public assistance. This inconsistency in the choice of neighborhood or contextual variables has limited the ability to make etiologic inferences across studies[27]. Further, previous neighborhood studies often select census variables that represent fewer socioeconomic parameters, for instance, % population below poverty[17][10]. Our NWAS identified more complex, joint effect variables that combined race, age, and poverty information with household or renter status. These more complex variables could provide insights into disease etiology and suggest that interactions may exist among demographic domains that are often considered individually in current neighborhood studies[10]. For example, percentage of male nonfamily householders living alone AND percentage of male nonfamily householders living alone over 65 appear to represent similar social concepts. However, NWAS separated these variables into two different components. Variables related to single resident households are used as markers of social support[37, 38], and it is possible that they could represent separate or potentially dynamic changes in the role of social support across the lifespan. Thus, given the specificity of NWAS top hits, it is possible they could be used alone or in combination, in future multilevel investigations and to more precisely identify and target geographic areas that are associated with one or more unfavorable NWAS characteristics for disease interventions. However, NWAS is a new methodologic approach and the etiologic significance of the NWAS hits would need to be investigated within those neighborhoods that exhibit these unfavorable characteristics. Further, the utility of NWAS findings for neighborhood risk assessments will need to be determined through comparisons with existing variable selection methods in future studies.

While NWAS methods can be extended in a variety of ways, the current formulation described here has limitations. Area-level data analyses assume individuals residing within the same geographic area experience similar circumstances. In reality, non-residential experience (e.g., work) and individual-level characteristics (e.g., biology, behavior, risk factors) also impact health states. Thus, future NWAS should be conducted in study populations that can adjust for or directly study individual-level SES factors[29]. In addition, standardized data processing, aggregation methods, and geographic boundaries used in administrative datasets can suffer from systematic reporting bias and missingness [6, 16]. Based on our missing data assessments, bias is likely non-differential (S1 File), but future NWAS studies should investigate missingness using both spatial autocorrelation and imputation techniques[39], as well as evaluate the effects of aggregation and geospatial boundary selection using interpolation and point-based, boundary-free approaches [40, 41] [42, 43].

The NWAS approach described here features many of the methodologic requirements previously proposed for GWAS or EWAS studies[23]. A hallmark of GWAS has been replication of discovery findings in comparable study populations. The NWAS presented here focuses on discovery and minimization of false positives through statistical adjustments, without a separate replication population. Under certain circumstances, a single discovery phase[44] and other biologic or functional-based approaches may be favored over statistical replication[45]. For example, the frequency or percentages of census variables may vary by geography, which can bias estimates of association. Interactions between variables (as indicated by the more complex, joint effect variables in the NWAS) are also likely to vary by geography[45]. While comparisons across geographical areas may be undertaken, use of independent datasets to validate findings may mask real differences between these geographies, and may not be appropriate in the NWAS setting. This is a topic that requires further exploration.

This NWAS study demonstrates that high-dimensional data analysis can be applied to large, publically-available datasets and can yield biologically plausible results. This is the first study to systematically, agnostically, and comprehensively evaluate the role of neighborhood-level factors in prostate cancer using “big data” methods. Although NWAS approaches should be tested in other study populations, NWAS addresses methodological limitations in current neighborhood studies, while capitalizing on methodologic approaches used for precision medicine[46] [47]. Further, coupling an NWAS approach with individual-level risk factor information could have implications for multilevel, health disparity studies, as well as precision public health initiatives aimed at identifying and targeting geographic areas in need of intervention efforts across disease sites.

Supporting information

Acknowledgments

These data were supplied by the Bureau of Health Statistics and Research, Pennsylvania Department of Health, Harrisburg, Pennsylvania. The Pennsylvania Department of Health specifically disclaims responsibility for any analyses, interpretations, or conclusions.