Motivating Uptake and Utility of Geographically Sensitive Neighborhood-wide Association Studies of Cancer

A substantial body of literature is devoted to relationships between myriad environmental factors and cancer survival, yet the role of geographic operational scale has been paid little attention. If representing a causal effect within a defined population of people with cancer, geographic operational scale can be defined as the boundary around a typical person within which an environmental exposure exerts an influence on cancer survival. Boyle and colleagues recently modified extant Neighborhood Wide Association Study (NWAS) techniques to identify the most relevant geographic operational scale within a cohort of people diagnosed at a mid-Atlantic Comprehensive Cancer Center with female breast cancer between 2014 and 2024 (N=3041) and who had a valid residential address at diagnosis geocoded to the parcel level.¹ They considered three, pre-defined sets of geographic boundaries to determine operational scale: 1) census tract only, 2) census tract portions within 1 kilometer from geocoded address, and 3) census tract portions within 5 kilometers from geocoded address. Readily accessible neighborhood measures (n = 5553) were from various sources and represented constructs across domains spanning demographics, socioeconomic status (SES), discrimination, social networks, health behaviors, and air pollution. To avoid extensive missingness due to limited geographic coverage of some neighborhood data (e.g., Home Owner’s Loan Corporation historical redlining only available in specific housing markets), decisions were made to only consider those neighborhood variables with at least 70% non-zero and non-missing values, which excluded hundreds of variables including some air pollution and historical redlining data from consideration. NWAS techniques – false discovery rate-corrected variable selection and elastic net regression model tuning via cross validation – were applied to generate hazard ratios and 95% CIs from Cox proportional hazards frailty models (i.e., census tract random intercept) of all-cause death, adjusted for individual-level race, ethnicity, tobacco use, alcohol use, and body mass index.

Results indicated that hazard of death increased with increases in four neighborhood variables across each of the three geographic scales considered: census tract percentage of renter-occupied households making $35,000-$49,999 annually who pay >30% of household income on housing costs, 5 km buffer percentage of households moving within the past year, census tract percentage of renter-occupied households containing 6+ people, and 1 km buffer of percentage of 3+ years old enrolled in public pre-school. The authors appropriately make no claim of causality, but that the financial and housing domains which these four variables represent should be the focus of further investigation. This geographically sensitive NWAS (gs-NWAS) approach put forth by Boyle and colleagues has the potential to improve measurement of environmental factors in cancer epidemiology by critically considering and investigating the most relevant areas through which exposures might impact cancer outcomes. What follows is an overview of four main considerations that could motivate greater uptake and utility of a gs-NWAS approach: 1) specificity based on theory and evidence, 2) alignment of exposome data with the research question, 3) incorporating exposure estimate uncertainty, and 4) operationalizing temporal dimensions of exposures.

Care should be taken when defining ‘neighborhoods’ of breast cancer cases, or cases of other cancer types. Measuring and empirically testing associations between potential neighborhood exposures – oftentimes derived from responses to annual surveys – and cancer outcomes should be motivated by current theory and evidence and with specificity towards the study population (e.g., cancer type/subtype, sociodemographic composition, spatial and temporal contexts). Influential work by Krieger et al., published over 20 years ago examined cancer mortality rates by census bureau-based SES measures calculated at the block group, census tract, and zip code levels.² Cancer mortality rates were higher among those residing in low compared to high SES areas, regardless of geographic scale. SES-based trends in cancer incidence, however, were more sensitive to geographic unit at which SES was measured and varied across various cancer types and geographic regions, such that census tract-level measures exhibited the most consistent and theoretically expected trends. This and other work over the decades largely solidified census tracts as the preferred unit of analysis when investigating associations between area-level SES and many cancer outcomes, including incidence, mortality, or survival. Two other groups of environmental measures with substantial bodies of literature around measurement and geographic scale are racial-ethnic residential segregation and air pollution,^3,4 both considered in Boyle at al., and theoretically linked to breast cancer outcomes.⁵ The spatial scales at which specific indices of racial-ethnic segregation (e.g., dissimilarity, isolation, Theil/entropy) or air pollutants (e.g., particulate matter, polycyclic aromatic hydrocarbons) impact cancer may vary by cancer types and, in the case of air pollution, might operate at multiple spatial scales which do not conform to administrative units such as census tracts.³

Neighborhood exposome measures can include individual domains of social determinants of health, environmental exposures or composite indices derived from them.⁶ Such indices are typically highly correlated with their components, and may themselves introduce additional sources of variation. Therefore, care should be taken to align the choice of individual measures or indices with the study objective (e.g., risk stratification, community-level intervention design, or etiologic research), which can improve both interpretability and methodological rigor.

Another important statistical consideration is that many neighborhood exposome measures are estimates, as opposed to direct measurements, with associated variability. For example, American Community Survey or CDC PLACES provides neighborhood-level measures from model-based estimates, which include published margins of error or confidence intervals. However, commonly used variable selection techniques (e.g., elastic net) typically do not account for measurement error during model fitting. As a result, future research should also consider the development of methods that incorporate uncertainty estimates into the selection and evaluation of exposome variables, to improve robustness and inference.

Lastly, there may be opportunities to extend gs-NWAS methods by accounting for temporal dimensions of exposures and outcomes. Large-scale ascertainment of residential histories of people with cancer has recently been made possible through efforts of many central cancer registries within the US. Such data provide critically important information on when in addition to where someone with cancer resided. Coupled with temporally resolved exposure data – oftentimes spanning decades and indexed at least annually – can enable investigation of spatio-temporal scales or lags most relevant for cancer outcomes. If a gs-NWAS is intended to inform the relevant stakeholders of where to focus efforts, a g-t(temporally)-s-NWAS could provide critical details about when such exposures might impact cancer outcomes and how long it could take before any intervention might improve cancer health outcomes among a population.