Final Report: Effects-Based Cumulative Risk Assessment in a Low-Income Urban Community near a Superfund Site

EPA Grant Number: R834577
Title: Effects-Based Cumulative Risk Assessment in a Low-Income Urban Community near a Superfund Site
Investigators: Levy, Jonathan , Ayala, Arlene , Fabian, Maria Patricia , French, Robert , Korrick, Susan A. , Peters, Junenette , Pina, Jordan , Rosario, Maria
Institution: Boston University , NorthStar Learning Centers
EPA Project Officer: Hahn, Intaek
Project Period: September 1, 2010 through February 28, 2014 (Extended to February 28, 2015)
Project Amount: $749,662
RFA: Understanding the Role of Nonchemical Stressors and Developing Analytic Methods for Cumulative Risk Assessments (2009) RFA Text |  Recipients Lists
Research Category: Health Effects , Human Health Risk Assessment , Health

Objective:

Our study had the objective of characterizing the contributions of multiple chemical and non-chemical stressors to ADHD-related behavior and elevated blood pressure in New Bedford, Massachusetts. Our initial objective was to characterize exposures to relevant stressors across the population, with an emphasis on the population attributes that contributed to high exposures to multiple stressors. However, multivariable attribute data at high geographic resolution do not exist because of privacy concerns. To address this limitation, we first developed a simulated population of New Bedford that included census tract of residence as well as extensive individual attributes relevant to chemical and non-chemical stressor exposures (Levy, et al., Risk Analysis, 2015). The simulated population was derived by applying probabilistic reweighting using simulated annealing to data from the 2006-2010 American Community Survey. We then developed a series of multivariable regression models to predict key behaviors (e.g., smoking) using New Bedford-specific data from the Behavioral Risk Factor Surveillance System. These behaviors and individual attributes were used as predictors of chemical exposures as measured within a New Bedford cohort study (NBC) focused on ADHD-related behavior and from the National Health and Nutrition Examination Survey (NHANES), which included blood pressure measures.

Summary/Accomplishments (Outputs/Outcomes):

To link these exposures with health outcomes, we needed epidemiological studies that applied methods relevant to cumulative risk assessment, given the interest in understanding the joint effects of multiple stressors. As there were no published studies for the stressors and outcomes of interest that used methods relevant to cumulative risk assessment, we conducted new epidemiological analyses using methods intended for this complex exposure environment. For example, we used structural equation modeling (SEM) to simultaneously evaluate predictors of chemical stressor exposures and predictors of four different blood pressure outcomes using NHANES data (Peters, et al., Environmental Research, 2014). This allowed us to separate the influence of demographic and behavioral factors on chemical exposures from the direct influence of these factors on health outcomes. We developed SEMs for chemical and non-chemical stressors associated with ADHD-like behavior in the NBC, adapting the exposure models to publicly available data in order to apply in a cumulative risk model. We also analyzed data from the NBC using generalized additive modeling approaches that allow us to examine the joint influence of multiple stressors on ADHD-related behavior and determine whether potential interactions are present. This included combinations of chemical stressors as well as chemical and non-chemical stressors.
 
The final cumulative risk assessment relied on the combination of the simulated population characteristics, the exposure models linked to those characteristics, and the epidemiological analyses considering both chemical and non-chemical stressors. To ensure that we captured the ranges of exposures anticipated in the population, we applied the exposure regression models with inclusion of the residual model errors, and we generated multiple simulated populations given the probabilistic assignment of individual attributes (like smoking). To make the cumulative risk assessment relevant to decision-making by community members and other stakeholders, we focused on the health benefits associated with counterfactual exposure scenarios or hypothetical interventions, centering on modifiable human behavior interventions such as changes in food consumption practices. In other words, we examined how the distribution of ADHD-related behavior or blood pressure would shift as a result of reducing chemical stressor exposures to a low level throughout the population. The final output from these analyses, therefore, included detailed characterization of the populations at risk for elevated exposure to stressors influencing ADHD-related behavior or elevated blood pressure, as well as the magnitude and distribution of benefits from reducing exposures to these stressors. These analyses met our stated project objectives.
 
Within our project, university researchers worked closely with community representatives, including our primary community partner (NorthStar Learning Centers) and individuals from the City of New Bedford. NorthStar staff provided key input regarding interpretability of models, helped to identify data streams for community characterization, helped to focus the analytical efforts on models that would produce tangible guidance to community members, and proposed and implemented a community survey to evaluate exposure patterns and predictors in New Bedford among diverse populations. The community survey was administered to 382 New Bedford residents in 2012, with the overarching goal to evaluate exposure patterns and predictors in current New Bedford and determine whether trends were similar to those found in the NBC in the mid-1990s. The surveys covered all neighborhoods of New Bedford and reflected the demographic diversity of the city (72% had income under $40,000, 26% were Hispanic, 27% were Cape Verdean, and 20% Portuguese/Azorean). In our core analyses, we developed predictive regression models of exposure risk factors of interest with the 2012 survey data to parallel analyses performed on the NBC data. These analyses helped us to establish that demographic predictors of food consumption and other behaviors were generally consistent with the prior models, providing reassurance regarding their application, while offering insights about populations (e.g., Hispanics) who were underrepresented in the NBC. In general, the community-based participatory research (CBPR) aspect of our study was crucial to the success of the analytical models.
 
Beyond our collaborative work with NorthStar, we met with other governmental and community organizations and presented some study findings at a community event held at the public library. We will be holding additional outreach and dissemination events in the summer of 2015 and beyond; while this is subsequent to the end of the project period, all study partners are committed to comprehensive dissemination of study findings, including those that were not peer reviewed and published by the end of the study period.
 
Generation of Synthetic Microdata
Cumulative risk modeling relies on the ability to apply exposure and health risk models to an entire community, and thus requires extensive individual-level or cross-tabulated data with high geographic resolution. However, these data are not available because of the need to limit potential identifiability. As such, to estimate cumulative exposures and risks in New Bedford, we needed to generate synthetic microdata that reasonably represented attributes of New Bedford relevant to chemical and non-chemical stressor exposures.
 
As described in our publication (Levy, et al., Risk Analysis, 2015), two general approaches have been used in the literature to generate synthetic microdata—reweighting, in which the existing microdata are assigned sets of weights for each small area to best match the aggregate distributions, and synthetic reconstruction, in which the microdata are probabilistically generated given the aggregate distributions. We focus on reweighting approaches in this analysis, given our interest in characterizing a large number of population characteristics that may predict exposure or vulnerability. Probabilistic reweighting using simulated annealing has been shown to perform slightly better than alternative reweighting approaches, such as deterministic reweighting or conditional probability modeling. This entails selecting a random subset of households from microdata and comparing to aggregate area-level constraints, sequentially replacing individual households to see if the fit improves. The simulated annealing approach helps ensure that global rather than local optima are obtained. We applied this methodology using the software package CO, which has been tested in multiple applications and is considered to be more robust than alternative algorithms.
 
We first applied probabilistic reweighting using simulated annealing to data from the 2006-2010 American Community Survey, combining 9,135 microdata samples from the New Bedford area with census tract-level constraints for individual and household characteristics. We then evaluated the synthetic microdata using goodness-of-fit tests and by examining spatial patterns of microdata fields not used as constraints. As a demonstration, we developed a multivariable regression model predicting smoking behavior as a function of individual-level microdata fields using New Bedford-specific data from the 2006-2010 Behavioral Risk Factor Surveillance System (BRFSS), linking this model with the synthetic microdata to predict demographic and geographic smoking patterns in New Bedford.
 
Our simulation produced microdata representing all 94,944 individuals living in New Bedford in 2006-2010. Variables in the synthetic population matched the constraints well at the census tract level (e.g., ancestry, gender, age, education, household income) and reproduced the census-derived spatial patterns of non-constraint microdata. Smoking in New Bedford was significantly associated with numerous demographic variables found in the microdata, with estimated tract-level smoking rates varying from 20% (95% CI: 17%, 22%) to 37% (95% CI: 30%, 45%). While both the microdata generation and regression models had some statistical uncertainty, as shown in more detail in our publication (Levy, et al., Risk Analysis, 2015), the errors were generally low and the underlying patterns were consistent with observational data. Therefore, we were able to show that our simulation approach could successfully create geographically resolved individual-level microdata that can be used in community-wide exposure and risk assessment studies. We currently are working with community partners to develop and apply other community-relevant BRFSS models to the synthetic database in order to identify New Bedford subpopulations at high risk for multiple behaviors and characteristics.
 
Structural Equation Modeling Using Data from NHANES and the NBC
As described elsewhere (Peters, et al., Environmental Research, 2014), we used data from the National Health and Nutrition Examination Survey (NHANES) 1999-2008 for adults ages 20 years and older to investigate the effect of chemical exposures on blood pressure accounting for non-chemical stressors and for the interrelationships among these stressors, chemical exposures and blood pressure. We identified multiple candidate chemical stressors from a literature review—arsenic, BPA, cadmium, lead, mercury, PCB, PCDD and PCDF—but ultimately focused on cadmium, lead, and PCB given the results of preliminary regression models. We then constructed SEMs to simultaneously evaluate predictors of chemical stressor exposures and predictors of four different blood pressure outcomes. As described elsewhere (Levy, et al., PLoS One, 2014), structural equation modeling combines linear regression, path analysis, and factor analysis, and is particularly well suited to assess the relative importance of multiple stressors and how they interact to affect a health outcome in the context of cumulative risk assessment. Appropriate application of SEM requires a clear prior specification to avoid illogical results, so we used the literature and first principles to explicitly consider the set of factors that could both predict chemical exposures and health outcomes.
 
Predictive models of chemical and non-chemical exposures within NHANES illustrated the complexity of the multi-stressor exposure environment. Lead was most significantly elevated among older men who were current or past smokers. Smokers and older individuals also had high blood cadmium levels, but exposures were higher among women. PCB exposure was very strongly associated with increasing age, consistent with its persistence, and also was positively associated with lipid levels. We also developed predictive models for BMI, lipid levels and menopausal status, as other non-chemical risk factors could be ascertained directly from public datasets. In our SEMs, lead appears to have greater influence on the blood pressure measures that are important predictors in younger populations (i.e., diastolic blood pressure and mean arterial pressure), with PCBs having greater influence on the blood pressure measures more important in later life (i.e., systolic blood pressure and pulse pressure). All complete models are summarized in Peters, et al., Environmental Research (2014), with full parameterization available in the Supporting Information section.
 
We also developed SEMs linking demographic and geographic predictors with exposures measured within the NBC. Based on our literature review and preliminary modeling, we focused on prenatal Pb, Hg, PCB, and DDE as chemical stressors, and we also considered non-chemical stressors such as sociodemographic factors and quality of the home environment, measured with the Home Observation for Measurement of the Environment (HOME). As a first step, we built comprehensive exposure models based on covariates available from the NBC. We then linked the covariates to variables available from public databases. For those covariates not available in public databases, we built models predicting the NBC-specific variables, based on publicly available data. For example, our initial regression model for cord serum PCBs included predictors reflecting multiple aspects of maternal diet, information not available from public data resources. Therefore, we used SEMs to include sociodemographic predictors of food consumption patterns and other key behaviors influencing exposures. For the NBC, we did not apply structural equation modeling to jointly evaluate predictors of exposures and health outcomes, in part because the sample size was insufficient to develop robust models, as well as the limitations that SEMs have to study interactions.
 
Generalized Additive Modeling to Evaluate Effects of Chemical and Non-chemical Stressors on ADHD-related Behavior in the NBC Cohort
To evaluate effects of simultaneous exposure to multiple chemical and non-chemical stressors, we used generalized additive modeling (GAM) approaches recently developed to explore effects of chemical mixtures. The general concept involves constructing two-dimensional smoothed functions of the health response associated with combinations of stressors, applying smoothing techniques more commonly used for spatial data. Examination of one-dimensional transects at varying levels of exposure of one stressor can allow for evaluation of whether the other stressor exhibits health effects, and examination of the overall surface can provide insight about potential non-linear associations or complex interactions. While this GAM approach has been more traditionally used to formally evaluate the likelihood of concentration addition or to model toxicological data, we use it for data visualization, model-building, and exploratory analysis given our model application. We examine pairs of chemical and non-chemical stressors determined a priori to be predictors of ADHD-related behavior, and only include two-dimensional GAM terms for the subset where significant effects are observed and could not be captured through conventional covariates. The resulting multivariable regression model, therefore, provides insight about the stressors significantly associated with ADHD-related behavior in the NBC, as well as about potential departures from linearity and important interactions. GAMs were applied directly to model the association of continuous chemical and non-chemical stressors on ADHD-like behavior, and in stratified analyses to analyze categorical or binary non-chemical stressors (e.g., income categories and education).
 
Modeling of Cumulative Risks and the Benefits of Hypothetical Interventions
Our final blood pressure and ADHD-related behavior cumulative risk assessment models combined the synthetic microdata, the exposure models, and the new epidemiological models. For each health outcome, we included the subset of individuals representing the populations included in our epidemiological analyses: adults age 20 and older for blood pressure, and women of childbearing age for child ADHD-related behaviors (given exposure metrics related to cord blood and prenatal maternal characteristics). Each individual in the synthetic microdata had numerous individual attributes and predicted exposures to chemical and non-chemical stressors given our derived regression models. In order to incorporate the uncertainty in our regression models, we predicted exposures and added a random error term with a mean of zero and the error-reported variance. We followed the same approach to predict either ADHD-related behavior or blood pressure. We used Monte Carlo analysis in SAS to create multiple populations reflecting each realization of the random number draws for exposure and outcomes, and for each population, we examined the change in the distribution of health outcomes for defined counterfactual scenarios. Our counterfactual scenarios focused on chemical stressors and considered hypothetical changes to distributions (including reductions of stressor exposures down to a low level throughout the population, for calculation of attributable burden, and reductions of stressor exposures reflecting defined exposure pathways or quantified benefits from interventions documented in the literature). We quantified the benefits for each chemical stressor individually and also considered the implications of joint reductions. We examined the shift in the distribution of the continuous measures of ADHD-related behavior and blood pressure, and also calculated the number of individuals moving from above to below a level considered to have clinical relevance (e.g., 140/90 as a measure of hypertension). The synthetic microdata allowed us to describe the demographics and geographic locations of the high-risk individuals exposed to multiple stressors, as well as identify those who would benefit most from reductions in exposures.

Conclusions:

Our methods were both technically effective and would be economically feasible in future cumulative risk assessment applications. Most of our analyses involved datasets freely available to the public (such as the Census, NHANES, or BRFSS) and use of statistical software that is either free or widely available in the research community, and we used statistical techniques that are applicable to numerous health endpoints. Specifically, while we generated synthetic microdata for New Bedford, analogous simulation modeling could be done for any city across the United States, including the development of multivariable regression models to predict the probability of smoking and other behaviors. The application of SEMs to jointly evaluate predictors of exposures and health outcomes was highly effective in a context where sample size was adequate, and for studies with smaller sample sizes, developing separate regression models for exposures and outcomes proved to be technically sound. In addition, using GAMs to examine potential interactions between stressors was a powerful approach for both data visualization and statistical evaluation of interactions without making strong assumptions about functional form. The core modules for these analyses are freely available within the R statistical software package. The only primary field data collected was a community survey, and working with community collaborators and collecting surveys both in-person and online was an economically efficient approach to reach our target sample size. We did have the opportunity to directly access and analyze data from a longstanding cohort study conducted in the community of interest for cumulative risk assessment, which strengthened our study considerably but would not necessarily be available for other communities. That said, our NHANES-based analyses reinforced how local population attributes could be used to develop local insights from national-scale exposure models and epidemiological analyses.
 
This research adds significantly to our understanding of how multiple stressors in a high-risk urban environment can work in combination to influence important health outcomes. Both ADHD-related behavior and elevated blood pressure have significant public health burdens and are influenced by a complex array of stressors. Our research study helps to characterize populations that are highly exposed to multiple influential stressors, allowing for more targeted intervention strategies, and also quantifies the public health benefits of an array of exposure reductions. Outcomes like blood pressure and ADHD-related behavior are measured on a continuum, where a small shift in the population distribution could lead to a significant change in morbidity and a decrease in the number of people reaching clinical disease definitions.
 
Computer Models
Our project relied entirely on modeling techniques previously described in the literature, including probabilistic reweighting using simulated annealing, structural equation modeling, generalized additive models, and Monte Carlo analysis. Therefore, we have not included any model development elements, but have discussed above how we applied these existing models to generate novel insight in the context of cumulative risk assessment.


Journal Articles on this Report : 4 Displayed | Download in RIS Format

Other project views: All 20 publications 4 publications in selected types All 4 journal articles
Type Citation Project Document Sources
Journal Article Levy JI, Fabian MP, Peters JL. Community-wide health risk assessment using geographically resolved demographic data: a synthetic population approach. PLoS One 2014;9(1):e87144 (10 pp.). R834577 (Final)
  • Full-text from PubMed
  • Abstract from PubMed
  • Associated PubMed link
  • Full-text: ResearchGate-Full Text PDF
    Exit
  • Abstract: PLoS One-Abstract
    Exit
  • Journal Article Levy JI, Fabian MP, Peters JL. Meta-analytic approaches for multistressor dose-response function development: strengths, limitations, and case studies. Risk Analysis 2015;35(6):1040-1049. R834577 (Final)
    R834798 (Final)
  • Abstract from PubMed
  • Abstract: Wiley Online-Abstract
    Exit
  • Other: Harvard University-Prepublication PDF
    Exit
  • Journal Article Payne-Sturges DC, Korfmacher KS, Cory-Slechta DA, Jimenez M, Symanski E, Carr Shmool JL, Dotson-Newman O, Cloughtery JE, French R, Levy JI, Laumbach R, Rodgers K, Bongiovanni R, Scammell MK. Engaging communities in research on cumulative risk and social stress-environment interactions: lessons learned from EPA's STAR Program. Environmental Justice 2015;8(6):203-212. R834577 (Final)
    R834576 (Final)
    R834578 (Final)
    R834579 (Final)
    R834580 (Final)
    R834581 (Final)
    R834582 (Final)
  • Full-text from PubMed
  • Abstract from PubMed
  • Associated PubMed link
  • Full-text: Mary Ann Liebert, Inc. Publishers-Full Text-HTML
    Exit
  • Abstract: Mary Ann Liebert, Inc. Publishers-Abstract
    Exit
  • Other: Mary Ann Liebert, Inc. Publishers-Full Text-PDF
    Exit
  • Journal Article Peters JL, Fabian MP, Levy JI. Combined impact of lead, cadmium, polychlorinated biphenyls and non-chemical risk factors on blood pressure in NHANES. Environmental Research 2014;132:93-99. R834577 (Final)
  • Abstract from PubMed
  • Full-text: ScienceDirect-Full Text HTML
    Exit
  • Abstract: ScienceDirect-Abstract
    Exit
  • Other: ScienceDirect-Full Text PDF
    Exit
  • Supplemental Keywords:

    ADHD, blood pressure, cadmium, cumulative risk assessment, lead, mercury, non-chemical stressor, PCBs, Superfund

    Progress and Final Reports:

    Original Abstract
  • 2010 Progress Report
  • 2011 Progress Report
  • 2012 Progress Report
  • 2013 Progress Report
  • 2014