Social stressors and air pollution across New York City communities: a spatial approach for assessing correlations among multiple exposures

Within the field of environmental health, there is substantial interest in the combined effects of chemical and non-chemical exposures on human health [1–4]. Recent epidemiologic and toxicologic evidence indicates significant modification of pollution effects on health by chronic psychosocial stress [5–12]. For investigators interested in understanding the relationship between the social and physical environment, there is a growing need for refined, replicable methods for: a) measuring social stressor exposures across large cohorts, and b) reducing confounding between social and chemical exposures in environmental epidemiology [13].

Recent research on this topic has considered psychosocial stress as a possible key factor modifying the relationship between chemical exposures, including air pollution or lead, and adverse health outcomes [14]. Measurement of psychosocial stress differs from chemical pollution exposure assessment, because the physiologic impact of non-chemical stressors is mediated through individual perception [15]. Psychosocial stress is often a result of exposure to social stressors (i.e., an event, condition, or external stimuli posing a physical or psychological challenge). When individuals evaluate stressors as imposing demands that are beyond their ability to cope, a sense of distress results; with repeated exposure to such stressors this sense of distress can become chronic. Chronic psychosocial stress is associated with negative emotional states and maladaptive behaviors that influence immune, endocrine, and metabolic function to produce cumulative wear-and-tear – often referred to as allostatic load[16]. These physiologic changes may alter individuals’ reactivity to chemical exposures (e.g., pathogens, pollutants) and increase risk for multiple disease etiologies [17]. As such, individuals and communities who are chronically exposed to social stressors may be more susceptible to adverse health effects of environmental chemicals. The field of stress measurement primarily relies on individual questionnaire or biomarker data to assess the occurrence of stressful events [18], conditions that might produce stressful experiences [19], recent perceptions of stress [20], or the mental health sequelae of chronic stress (i.e., depression, anxiety).

In contrast, large epidemiological studies that seek to evaluate whether chronic psychosocial stress increases susceptibility to chemical exposures are often unable to assess stress at the individual-level. As a result, they often rely on administrative indicators (e.g., crime, poverty rates) uniformly assessed across heterogeneous communities, as proxy measures to capture the presence of social stressors (e.g., lack of neighborhood safety, financial stress), and, by extension, psychosocial stress. However, because the impact of any stressor depends on an individual’s appraisal (i.e., how one experiences, perceives, or interprets the event), measuring social stressors can result in imprecise assessments of psychosocial stress. Moreover, based on evidence that psychosocial stress levels are high in low SEP areas [21, 22], most epidemiological studies of combined social and environmental effects have primarily used census-derived socioeconomic position (SEP) and demographic measures as a proxy for both a range of social stressors and for psychosocial stress per se[14]. Few studies have tested the assumption that SEP indicators are an appropriate proxy. As a result, it remains unclear how well SEP indicators capture exposure to social stressors and psychosocial stress; if these indicators are, in fact, weak proxies, it would limit the interpretability of contextual SEP effects, and hamper identification of possible causal mechanisms. As an alternative approach, some studies aiming to focus on psychosocial stress have examined other single social stressors, choosing stressors that are unlikely to be appraised positively (e.g., exposure to violence [5]). Both approaches suffer from unmeasured confounding insofar as they cannot account for, or distinguish amongst, the constellation of social stressors that can contribute to differential physiological susceptibility to chemical exposures.

Spatial correlation, or common clustering between distinct exposures – and discerning its impact on possible confounding and effect modification – is a key measurement challenge for social-environmental epidemiology. For example, traffic-related air pollution may be inherently confounded by traffic-related noise [23, 24], complicating the interpretability of effects for either exposure. Combining data on multiple social stressors addresses some of the concerns identified above, but a further methodological challenge is that publicly-available indicators are often aggregated to different administrative spatial scales by data source and type (i.e., police precincts, census tracts). Moreover, a number of different stressor indicators for the same construct may be available (e.g., multiple felony crime indicators – assault, robbery, or burglary), and it remains under-explored how well each of these various stressor indicators captures the intended psychosocial construct. As such, using only a single indicator of that construct may or may not be sufficient for capturing spatial distributions in these exposures. Thus, with reproducible geo-statistical methods to elucidate common spatial variation in social stressors and chemical exposures across large cohorts, we will improve our ability to reduce confounding and design studies appropriately powered to disentangle separate and combined effects.

Here, we present a spatial approach for characterizing co-varying social and environmental exposures. To demonstrate this approach, we use refined geographic analyses to examine intra-urban relationships across multiple exposures in New York City (NYC), where social, economic, and physical environmental conditions vary widely. Exposure data are drawn from multiple publicly-available administrative databases to capture dimensions of the social environment, and air pollution data are from the New York City Community Air Survey (NYCCAS). We quantify spatial relationships across this broad set of social stressor indicators, and between these stressor indicators and air pollution. We use geographic information systems (GIS)-based methods to: a) facilitate comparisons across different, incongruent administrative areal units, and b) explore potential effects of areal unit and spatial autocorrelation on observed associations between stressors and air pollution. Finally, we present an exploratory ecologic analysis of spatial confounding and effect modification by social stressors in the relationship between nitrogen dioxide (NO₂) and childhood asthma exacerbations, to illustrate the risks associated with mis-specification of spatially-patterned exposures and susceptibility.

We used GIS-based techniques to quantify relationships across social stressor indicators, and between these potential social stressors and air pollutants across NYC. Our findings call attention to complex spatial patterning across diverse social stressor and SEP indicators, and emphasize the importance of refined social exposure assessment for environmental health research. This spatial approach enables the disentangling of potentially correlated, yet conceptually distinct, chemical and non-chemical exposures – towards better quantifying spatial confounding and effect modification in social-environmental epidemiology.

Importantly, we found that a diverse set of social stressors across NYC are: 1) not consistently correlated, even among indicators that appear to be measuring similar aspects of the social environment (e.g., crime indicators), 2) not consistently correlated with area-level SEP, and 3) not consistently correlated with air pollution. The complexity of relationships among stressor indicators was borne out in factor analysis, which identified three spatially-distinct suites of stressors – ‘violent crime and physical disorder’, ‘crowding and poor access to resources’, and ‘noise complaints and property crime’ – suggesting that co-variation might be driven more by common spatial patterning than by shared meaning. Importantly, these three spatial factors did not represent different levels of socioeconomic position; areas that were similar with respect to area-level SEP did not necessarily have similar prevalence and combinations of other social stressors. As such, using any single stressor (including SEP) measure to serve as a proxy for psychosocial stress may be misleading; because areas that may be similar with respect to area-level SEP measures may differ regarding social stressors, single measures may inadvertently lead to confounding, and fail to capture important nuances of the social environment. It is also worth noting that some communities had high factor scores for more than one stressor factor, underscoring the potential for cumulative effects of multiple exposures in those communities. While the spatial patterning empirically summarized by stressor factors would likely differ between cities and regions, this reproducible approach may be helpful in developing locally appropriate composite social stressor measures.

Leveraging common spatial patterns among social stressors across communities enabled a more comprehensive characterization of social exposures, and perhaps psychosocial stress, and potential interactions with air pollution, which may contribute to social disparities in health. For example, in our ecologic analysis, air pollution was strongly correlated only with the spatial factor corresponding to ‘noise complaints and property crime’ (Factor 3), not with the other factors, or with indicators of area-level SEP. This is noteworthy, as communities with relatively high SEP and better healthcare access loaded relatively strongly on Factor 3 – a result which counters the common assumptions that air pollution would be highest in low-SEP communities, leading to positive confounding in air pollution epidemiology. It is also of note that our only indicator of perceived pollution – air quality complaint rates – loaded strongly on Factor 3, suggesting correlation between spatial patterns in modeled pollution concentrations and perceived air poor quality. The ability of pollution (or its sources) to act as both a chemical and non-chemical stressor is increasingly recognized as an important source of confounding [14].

In our ecologic analysis, we illustrated how modification in the NO₂-asthma exacerbation association may vary substantively by the selection of social stressors – represented here by our three stressor factors. Conceptually, this ecologic analysis underscores the need for thoughtful selection of stressor indicators, as mis-specification of stressors, which are hypothesized to impart physiologic susceptibility, can substantially alter observed effect modification. Further, empirically grouping social stressors according to spatial relationships may better capture potential physiologic susceptibility patterns, relative to using a single stressor indicator – an observation which is reinforced by our result that area-level SEP indicators did not strongly correlate with stressor factors, and thus are likely inadequate proxies for stressor exposures and psychosocial stress.

Though there are few examples in environmental epidemiology for refined social exposure assessment, our findings recall notions of “unpatterned inequality” in urban resource distribution, wherein communities may be favored in the allocation of some resources, while deprived in others [37]. In a recent study of individual- and area-level associations between SEP and air pollution, Hajat et al. [38] identified regional and intra-urban heterogeneity in the strength and direction of associations between area-level SEP and air pollution using spatially-informed regression models, wherein area-level SEP was positively associated with PM_2.5 and NO_X exposures across a geographic subset of NYC communities. More work is necessary, however, to replicate and refine salient social stressor measures, especially for large geographic cohorts wherein individual-level survey assessments of stress experience (e.g., [12]) or on-foot built environment assessments (e.g., [39]) are generally infeasible.

An alternative explanation for our empirically-derived findings include spurious associations due to unit of analysis (i.e., administrative areas are highly imperfect proxies for communities), measurement error in administrative data, or construct misspecification. Generally, larger administrative areas yield less precise metrics [40]; thus, while we aimed to include the widest variety of administrative indicators of social stressors possible, each indicator was examined at the finest resolution available, and areal units were robust to MAUP effects. Likewise, some stressor indicators may capture aspects of both chemical and non-chemical exposure constructs. For example, some physical disorder indicators are linked with poor mental health, but also with allergen and chemical exposures (e.g., cockroaches, pesticides) – both implicated in asthma etiology. Here, we attempted to minimize such confounding by focusing on stressors hypothesized to act predominantly through psychosocial stress pathways. These interpretation challenges are not, however, unique to this analysis, as administrative indicators are widely employed in social and environmental epidemiology. As such, mixed qualitative and quantitative methods for identifying salient stressors across spatially heterogeneous domains, and for validating administrative indicators against community- and individual-level stress experience (e.g., [41]), are important for improving reliability of administrative indicators for environmental epidemiology.

We aimed to develop and validate broadly applicable methods for quantifying common spatial patterning across urban chemical and non-chemical exposures. The NYCCAS fine-scale air pollution data enabled examination of spatial correlations across pollutants – and between pollution and social stressors – and provided fine-scale surfaces for validation of areal re-aggregations. GIS-based sensitivity analyses lend confidence to our quantitative findings. First, our validation method for areal weighting of incongruent spatial units could utilize any smooth surface supplying a known underlying distribution (e.g., elevation raster, kernel density surface), ideally with a scale of variability similar to (or more refined than) the re-aggregated exposure of interest. Though areal reformulation may induce local exposure misclassification, due to unknown within-area variability, our approach is useful for exploring global spatial confounding patterns. Second, sensitivity testing for MAUP effects and autocorrelation improved our understanding of spatial correlations, providing insights for future spatially-informed multi-variable modeling of social-environmental interactions.

In conclusion, our city-wide examination of social stressors and air pollution in one U.S. city highlight the utility of spatial analysis for disentangling the separate and combined effects of chemical and non-chemical exposures. The process presented for systematically identifying and assimilating area-based administrative indicators of social stressors, and deriving empirical spatially-covariant composites can minimize of confounding among social stressors, and between social stressors and air pollution. Our findings demonstrate that selection of social stressors and geographic scale may substantially alter observed effect modification, caution against using single SEP indicators as proxies for social stressors, and demonstrate the risks associated with mis-specification of social stressor exposures. Empirical studies with stronger validated and spatially-informed measures of social stressor exposures are needed to better understand spatial confounding and joint effects between chemical and non-chemical stressors.