The objective of this module is to provide an introduction to coding through the R language and its associated environment, RStudio. This objective is met by first detailing instructions with corresponding screenshots describing how to download/install both of these programs. An introduction on installing and loading packages in R is then provided. Scripting basics are detailed, including setting a working directory, importing and exporting files, and viewing data within the R console/RStudio environment. The importance of this module is that it provides the foundation needed for participants to become acclimated and set-up for running R programming on their computing systems.

The objective of this module is to provide an introduction on data organization methods. This objective is met by presenting basic data organization methods using an example environmentally relevant human cohort dataset. This cohort was generated by creating data distributions randomly pulled from our previous publications (Rager et al., 2014a; Clark et al., 2019; Payton et al., 2020; Clark et al., 2021), resulting in a bespoke dataset for these training purposes. Data include subject information/demographic data, as well as environmental exposure data, focusing on metals concentrations in drinking water and human urine samples. Data organization methods that are demonstrated in this training module include merging, filtering, subsetting, melting, and casting. These important methods are demonstrated using base R functions, as well as the commonly implemented package, Tidyverse, that allows users to more efficiently organize and manipulate datasets in R (CRAN, 2021b). The importance of this module is that it provides basic skills needed to organize data, in general, within the coding environment, representing a foundational skill that must be acquired prior to running any scripted analysis.

The objective of this module is to provide an overview of basic statistical tests and data visualizations. This objective is met leveraging the same example cohort with environmentally relevant data introduced in Section 3.1.2. Tests for normality are first presented, alongside methods to plot histograms and boxplots to view data distributions. Basic statistical tests are then presented, including the t-test, analysis of variance, regression modeling, chi-squared test, and Fisher’s exact test. Additional example visualizations are provided alongside these statistical tests, including boxplots, scatterplots, and regression lines. These statistical tests are introductory-level, with more extensive examples and associated descriptions of statistical models in the proceeding applications-based training modules (e.g., modules 2.4, 2.5, 2.6, 3.2, and 3.3). The importance of this module is that it provides an overview of statistical methods that are very routinely employed within environmental health studies, and thus learning how to carry out these basic statistics represents a foundational skillset for anyone in this field of study.

The objective of this module is to provide an introduction to methods that can be used to visualize high dimensional data. Approaches described in this training module include data formatting, data scaling, and the visualization of prepared datasets through density plots, GGally plots, boxplots, correlation plots, hierarchical clustering, and heatmaps. Visualization approaches are demonstrated using a large environmental chemistry dataset, based off a chemical analysis of smoke samples collected during lab-based simulations of wildfire events. These data have been previously published (Kim et al., 2018; Rager et al., 2021) and are used here as an example of an environmental dataset relevant to environmental health. These visualization methods are provided here at an introductory-level, with many other examples detailed throughout the majority of the next training modules. The importance of this module is that it provides ideas and techniques that can be used to visualize data relevant to environmental health, which are becoming increasingly high dimensional and thus, require these more sophisticated methods to adequately illustrate important data trends.

The objective of this module is to introduce trainees to best data management practices in environmental health research. A method to ensure proper data management is the implementation of Findability, Accessibility, Interoperability, and Reusability (FAIR) practices (Wilkinson et al., 2016). This topic is receiving much attention in recent years through workshops, government reports, and publications which are published within the online training module. The following questions are addressed throughout this training module:

1) What is FAIR?

This module first provides an introduction to FAIR (Figure 2), including a history of how this term was first developed and implemented. Trainees are then guided through each component of FAIR, organized by letter. To detail, the F in FAIR identifies components needed to make the meta(data) findable. These components include unique persistent identifiers and descriptive information (i.e., metadata) that can be searched by both humans and computer systems. The A components are designed to enable that meta(data) be available long-term, and accessed by humans and machines using standard communication protocols with clearly described limitations on reuse. The I components of the principles address needs for data exchange and interpretation by humans and machines which includes the use of controlled vocabularies or ontologies to describe meta(data) and to describe provenance relationships through appropriate data citation. The R components highlight needs for meta(data) to be reused and support integration such as sufficient description of the data and data use limitations. The training module then reviews different types of data repositories that can be used to publish datasets in exposure science, toxicology, and environmental health research. Lastly, this module provides participants with additional training resources, workshops, government reports, and example publications surrounding the use of FAIR data management practices. The importance of this module is that effective data management, organization, and longevity are becoming increasingly critical in ensuring studies are scientifically sound and reproducible, and thus, all scientists that are involved in the analysis of data for a project should be aware of these issues and implement them within their ongoing studies. Research funding agencies are additionally requiring increased attention surrounding data sharing and FAIR practices (NIH, 2022).

The objective of this module is to provide an overview on analyzing toxicological response data in relation to exposure concentrations (or doses), resulting in the derivation of benchmark doses (BMDs). This topic is of high relevance to the field of environmental health, as BMDs represent values that are commonly used as the basis for evaluating risk in chemical safety evaluations, informing the levels at which chemicals may be regulated. This module specifically analyzes animal tumor incidence rates in response to exposure to a mock chemical tested across 12 different concentrations in drinking water. This dataset was generated for the specific purposes of this exercise, to allow for some interesting curve fits and a comparison between tissue site sensitivity to an example chemical insult. Several environmental health questions are posed throughout this module, including:

1) Which target tissue demonstrated the highest incidence of tumor formation from any single exposure dose?

This module first provides a high-level introduction to BMD modeling, and then guides trainees through the process of downloading/loading required packages and example data used in this exercise. These data are then viewed, such that trainees can see the four different tissue sites evaluated for carcinogenicity in response to exposure (i.e., kidney, liver, intestinal, and stomach tissues) and also obtain information on the overall distributions of tissue-specific tumor incidence. Then, data are plotted in dose-response using standard scatter plots with exposure concentrations along the x-axis and tumor incidence along the y-axis. With these foundation plots generated, trainees are then guided through methods to fit various model curves to these dose-response data, spanning log-logistic, Weibull, and asymptotic regression models as core examples available through the drc package (Ritz et al., 2015). The best fitting curves are then identified through 1) visual inspection of curve fits, and 2) calculation of Akaike Information Criterion (AIC) values. These examples highlight the importance of evaluating model fit to ultimately determine which model should be used to derive final BMD estimates (Figure 3). Trainees are lastly pointed to example dose-response publications that have addressed environmental health questions (Rager et al., 2017; Auerbach and Paules, 2018; Thompson et al., 2018; Johnson et al., 2020), as well as additional modeling tools and guidance documents surrounding dose-response assessments. The importance of this module is that BMD modeling represents a foundational topic in environmental health, where methods can be used to better understand which exposure concentrations/doses are required to elicit toxicity by evaluating trends in datasets.

The objective of this module is to provide an overview of machine learning (ML) approaches to evaluate high dimensional data relevant to environmental health applications. This module begins by introducing the need for predictive modeling, defining its use in the context of toxicology and environmental health, then establishing a working distinction between ML and traditional statistical methods. Recognizing the wide variety of machine learning methods currently available to researchers, this training module presents introductory-level information on two commonly employed methods, principal component analysis (PCA) and k-means clustering. Their use is illustrated on real data obtained from the National Toxicology Program’s Integrated Chemical Environment (ICE) resource (https://ice.ntp.niehs.nih.gov/). This module analyzes an example dataset of physicochemical property information for chemicals spanning two classes: per- and polyfluoroalkyl substances (PFAS) and statins. PFAS represent a ubiquitous and pervasive class of man-made industrial chemicals of high environmental relevance due to their persistence in the environment after contamination events (Fenton et al., 2021). Statins represent a class of lipid-lowering pharmaceuticals used for patients at risk of cardiovascular disease, and statins have been identified as present in water/wastewater effluent (Tete et al., 2020). The applied data example in this training module was designed to illustrate the concept of using ML methods to differentiate chemical class and “predict” (in this case, the group membership is known) chemical groupings that can inform a variety of environmental and toxicological applications. The following environmental health questions are addressed throughout this training module:

1) Can we differentiate between PFAS and statin chemical classes, when considering just the raw physicochemical property variables without applying machine learning techniques?

This module first provides a high-level introduction to the topics of machine learning and predictive modeling, and then guides trainees through the process of downloading/loading required packages and example data used in this exercise. These data are then viewed by plotting chemicals along their native physicochemical scales (e.g., boiling point versus molecular weight) for all 144 different chemicals, colored according to the two classes of PFAS and statins. Visualizing these data through two bivariate plots demonstrates that there is signal in the data, but substantial overlap between classes for most properties, i.e., individual physicochemical properties may not clearly differentiate between chemical classes. This limitation substantiates the need to employ machine learning methods to better describe group-level trends in these data. K-means clustering is then performed across all physicochemical property data in their native scale and visualized using a heat map. Next, PCA is carried out across all physicochemical property data and visualized to illustrate the concept of dimensionality reduction. The code is provided to calculate results from this PCA, including eigenvalues, percent variance captured by each principal component, and loading scores for the input variables. Finally, PCA is combined with k-means to generate predictions of two chemical groupings that almost entirely capture real-world classifications (Figure 4). Lastly, these methods are discussed in relation to additional applications, including the evaluation of other outcomes such as environmental fate and transport and disease outcome predictions. Trainees are provided additional resources including recent example studies that incorporate machine learning to address environmental health questions (To et al., 2019; Clark et al., 2021; Green et al., 2021; Odenkirk et al., 2021; Ring et al., 2021). The importance of this module is that it provides a helpful introduction to foundational ML concepts, and upon receiving this training, participants should be positioned to apply these methods to make predictions within their own high dimensional datasets.

The objective of this module is to provide an overview of chemical composition signatures and toxicological responses that can be evaluated to inform whether complex mixtures are “sufficiently similar.” Results from these analyses, referred to as sufficient similarity analyses, can be used to inform data extrapolation from a data-rich mixture to a data-poor mixture during a chemical safety/risk assessment, to adequately protect human health. In this example, data are re-analyzed from a study evaluating the chemical composition and toxicological effects of Ginkgo biloba extract, a common dietary supplement ingredient that is commercially available in the U.S. (Catlin et al., 2018). Here, 29 different sample lots of G. biloba extract were collected from several suppliers and analyzed. The chemical components of these sample extracts were evaluated using targeted methods, and associated toxicity was evaluated using gene-specific in vitro response assays. These data are leveraged in this training module to inform which of the G. biloba samples are sufficiently similar (and thus could use the same toxicological data for risk evaluation), and which are different (and thus would require additional testing). Several questions are posed throughout this module, including:

1) When viewing the variability between chemical profiles, how many groupings of potentially “sufficiently similar” G. biloba samples do you see?

This module specifically guides trainees through the loading of required packages and data, and then carries out an example sufficient similarity analysis first using the chemistry data. Trainees are guided through data processing and scaling, leading to two different grouping and visualization approaches: 1) PCA and associated scatter plot, and 2) hierarchical clustering and associated heat map visualization. Results are used to inform which G. biloba extracts display similar chemical composition profiles and which do not. Participants are also guided through the evaluation of potential outlier samples, gauging whether these impact overall data distributions. Similar methods are then used to evaluate the toxicological response data. This analysis concludes with a side-by-side comparison of the sample groupings that result when considering chemical composition vs. toxicity profile data (Figure 5), highlighting the importance of considering both data streams when determining sufficient similarity in the evaluation of complex mixtures. Trainees are then provided additional resources and references for further information on sufficient similarity analyses in environmental health research (Rice et al., 2009; Catlin et al., 2018; Ryan et al., 2019; Collins et al., 2020). The importance of this module stems from real-world exposures to complex mixtures that often have incomplete toxicity data, which necessitate training to determine when data from a reference mixture can be extrapolated to a mixture-of-concern for risk evaluation.

The objective of this module is to provide an overview of the -omics field and its relation to environmental health research, highlighting transcriptomics as an important -omics endpoint to analyze as a scripted example. The field of -omics initiated from genome-wide information obtained through the Human Genome Project, and has since then expanded to include -omic endpoints spanning the genome, epigenome, transcriptome, proteome, metabolome, microbiome, and the exposome (Cho and Blaser, 2012; Wild, 2012; Rager and Fry, 2013; Clark and Rager, 2020). Stressors within the environment have the potential to alter -omic signatures, impacting downstream biological processes, cellular function, tissue phenotypes, and overall health (Rager and Fry, 2013; Clark and Rager, 2020; Rager et al., 2020). When interpreting the potential consequences of -omic alterations, it is often helpful to place findings into the context of systems biology. In these systems-level analyses, molecules can be overlaid onto molecular networks to uncover biological pathways and cellular functions that are altered under the condition being tested (Rager and Fry, 2013; Meisner and Reif, 2015). This training module provides an overview of these strategies, using an example transcriptomics dataset acquired from lung tissues of mice exposed to biomass burn conditions indicative of the potential wildfire exposure scenarios (Kim et al., 2018; Rager et al., 2021). Several questions are posed through this module, including:

1) What two input data files are commonly needed in the analysis of -omics (e.g., transcriptomics) data?

This training module specifically guides users through the loading, viewing, and formatting of the example transcriptomics datasets and associated metadata. Methods to carry out QA/QC of the transcriptomics data are then detailed, including background filtering, sample filtering, and identification of potential sample outliers. Data are adjusted for potential sources of heterogeneity, including mixed cell population distributions that are commonly present when analyzing bulk tissue samples. Statistical models are then designed and implemented to identify genes that were significantly differentially expressed by the evaluated biomass burn scenarios, as enabled through the commonly implemented DESeq2 statistical pipeline (Love et al., 2014). We find that exposure to both flaming and smoldering of pine needles caused substantial disruptions in gene expression profiles. LPS serves as a positive control for inflammation and produced the greatest transcriptomic response. Gene expression alterations are then summarized via visualizations using MA and volcano plots (Figure 6). Resulting lists of differentially expressed genes are lastly evaluated in the context of systems biology, through pathway enrichment analysis based off relationships to KEGG pathways (KEGG, 2021) using gene set analysis enabled through the PIANO package (Varemo et al., 2013). We find that pathways involved in cardiopulmonary function, carcinogenesis, and hormone signaling were altered in response to these wildfire-relevant exposure scenarios. Trainees are lastly pointed to additional resources, including further information on -omics and systems biology, as well as additional research examples that have evaluated -omic alterations occurring in relation to the environment and involved in disease (Smeester et al., 2011; Lu et al., 2014; Rager et al., 2016; Chappell and Rager, 2017; Balik-Meisner et al., 2018; Chappell et al., 2019; Manuck et al., 2021b; Chang et al., 2021). The importance of this module lies in the training of systems biology concepts and analysis of -omics data, including RNA sequencing data, which are becoming increasingly standard molecular endpoints used in the evaluation of exposure-induced biological responses and disease etiologies.

The objective of this module is to provide an overview of the basics of toxicokinetic (TK) modeling and how this type of modeling can be used in the high-throughput setting for environmental health research applications. TK modeling refers to the evaluation of the uptake and disposition of a chemical in the body. In this activity, the capabilities of the high-throughput TK modeling package, “httk,” are demonstrated on a suite of environmentally relevant chemicals. The httk R package implements high-throughput TK modeling, including a generic physiologically based toxicokinetic model, as well as chemical-specific parameters needed to solve the model for hundreds of chemicals (Pearce et al., 2017). Several questions are posed through this module, including:

1) What is the maximum concentration of bisphenol-A estimated to occur in human plasma, after one exposure dose of 1 mg/kg/day?

This module specifically guides trainees through a general introduction to TK, TK modeling, and the types of TK modeling that can be employed to understand how chemicals travel throughout the body. The model provides scripted examples of TK modeling, starting with the estimation of plasma concentrations over time for a human exposed to bisphenol-A. Then, population variability is considered using information from CDC National Health and Nutrition Examination Survey (NHANES) to inform a distribution of possible plasma concentrations resulting from daily exposure to benzo(a)pyrene. Then, an example high-throughput analysis is carried out over ∼1,000 chemicals, in which population variability is captured to derive estimated quantile distributions of chemical plasma concentrations during steady-state conditions of 1 mg/kg/day exposures. Trainees are then guided through the process of deriving administered equivalent doses that associate with concentrations eliciting toxicity derived through toxicity testing. Equivalent doses are specifically derived across ∼1,000 chemicals that are estimated to elicit toxicity in humans, based on in vitro data, through “reverse TK” calculations. The in vitro dataset used in these derivations is the ToxCast high-throughput screening program. ToxCast activity concentrations that elicit 50% maximal bioactivity (AC₅₀) are uploaded and organized as inputs, and the 10th percentile ToxCast AC₅₀ is calculated for each chemical and carried forward in the analysis as concentration estimates for potency. Bioactivity exposure ratios (BERs) are then calculated to place findings into the context of risk assessment. Here, previously generated exposure estimates that have been inferred from CDC NHANES urinary biomonitoring data are used as estimates of chemical exposures. These hazard and exposure estimates are then visualized (Figure 7). The final BERs are calculated as the ratio of the lower-end hazard equivalent dose (for the most-sensitive 5% of the population) divided by the upper-end estimated exposure (here, the upper bound on the inferred population median exposure). The importance of these BERs in chemical prioritization efforts are lastly discussed in relation to environmental health research and corresponding government regulatory decisions. Trainees are provided additional resources and cases studies that have incorporated TK/httk to address environmental health issues (Wambaugh et al., 2015; Ring et al., 2017; Klaren et al., 2019; Breen et al., 2021; Ring et al., 2021). The importance of this module is that how chemicals travel throughout the body and elicit different toxicities based upon target organs significantly depends upon toxicokinetics, and being able to model these relationships is therefore critical towards understanding chemical-induced impacts throughout the body.

The objective of this module is to provide an overview of read-across methods to computationally predict chemical toxicity based on molecular structure information. This training module represents a timely topic, paralleling increased impetus for reducing reliance upon animal testing (NAS, 2007; EPA, U.S, 2021b). In this example, data are analyzed spanning approximately 7,000 chemicals that have known structure and acute toxicity data. The specific acute toxicity endpoint that is analyzed is LD₅₀, reflecting the dose required to cause lethality in 50% of animals, collected through historical animal testing. These data have been previously summarized and analyzed (Helman et al., 2019a). In this activity, we aimed to estimate an LD₅₀ value for an example target chemical of interest that is commonly used in the production of industrial compounds, 1-chloro-4-nitrobenzene. To achieve this aim, we explore ways in which we can search for structurally similar chemicals that have LD₅₀ data already available. Data on these structurally similar chemicals, termed “analogues,” are then used to predict acute toxicity for the target chemical. The following questions are addressed throughout this module:

1) How many chemicals with acute toxicity data are structurally similar to 1-chloro-4-nitrobenzene?

This module specifically guides trainees through the loading of required packages and example data, and then carries out an example read-across analysis specifically using the generalized read-across method (GenRA) (Shah et al., 2016). Trainees are guided through viewing the distribution of LD₅₀ values across all evaluated chemicals. Steps are then detailed to convert SMILES nomenclature into computed molecular fingerprint data. Using these molecular fingerprint data, the degree to which each chemical is structurally similar to another chemical is evaluated based on the Tanimoto similarity index. This structural similarity analysis yields an overall similarity matrix, containing all possible pairwise similarity values. Data are then filtered to focus on chemicals with Tanimoto similarity values >0.75 to the target chemical, 1-chloro-4-nitrobenzene, resulting in a list of 11 chemical analogues that could then be used to predict toxicity for the target chemical (Figure 8). Finally, generalized read-across was carried out by calculating a similarity-weighted activity score (Shah et al., 2016), using information from the 11 analogues to predict a LD₅₀ for 1-chloro-4-nitrobenzene. This in silico prediction was then compared to the experimentally observed LD₅₀ value for this chemical, which were very similar, highlighting the utility of read-across models to inform and predict toxicity for chemicals lacking data. The importance of this module is that predicting chemical-induced toxicity using entirely in silico approaches represents a highly efficient skillset that scientists can leverage to better understand chemical-toxicity relationships and predict which chemicals may induce harm to public health.

The objective of this module is to provide an exercise on organizing and analyzing chemical-gene lists aggregated through the Comparative Toxicogenomics Database (CTD) (CTD, 2021; Davis et al., 2021). Data were specifically pulled for published chemical-gene relationships mapping to the example environmental contaminant, arsenic. The following environmental health questions were addressed through this training module:

1) Which genes show altered expression in response to arsenic exposure?

This module specifically guides trainees through steps used to query CTD, including the specific selections used in this training dataset to organize chemical-gene interaction data for arsenic. These data are then uploaded into the training module R environment and used as an example for trainees to learn how to view file content and overall dimensions. Then data are filtered to include chemical-gene interactions that map specifically to changes in expression levels, yielding a list of genes that show arsenic-associated expression changes compiled from published literature. Data are additionally filtered using a different approach to yield genes that also show arsenic-associated gene methylation changes. These gene lists are then compared to result in the final elucidation of arsenic-altered genes that have published evidence for epigenetic modifications. Resulting genes represent critical mediators of inflammation and oxidative stress, among other important cellular processes. A visualization of these gene list comparison results is also scripted for using Venn diagram illustrations (Figure 9). Trainees are then provided additional resources, including reference to additional case examples that leveraged data from CTD to identify new mechanisms of environmental exposure-induced disease (Ahir et al., 2013), fill gaps on data poor chemicals to elucidate environmental influences on disease pathways (Kosnik et al., 2019), and derive new chemical risk values for prioritizing links between environmental factors, genetic variants, and human diseases (Kosnik and Reif, 2019). Together, this training module serves as an applications-based example to learn basic data manipulation, filtering, and organization steps in R, while highlighting the utility of CTD to identify novel genomic/epigenomic relationships to environmental exposures. The importance of this module is that analyzing data within CTD represents a powerful skillset within the environmental health field, which can be leveraged to improve the understanding of environmental influences on disease outcomes.

The objective of this module is to provide an overview of pulling, organizing, visualizing, and analyzing -omics data from the GEO database (NCBI, 2021). Data were specifically pulled from an example GEO dataset (accession number GSE42394) representing gene expression data originally used in a publication evaluating the genomic effects of formaldehyde inhalation exposure in the rat (Rager et al., 2014b). The following environmental health questions were addressed through this training module:

1) What kind of molecular identifiers are commonly used in microarray-based -omics technologies?

This module specifically guides trainees through the loading of required packages and data, including the manual upload of GEO data as well as the automated upload of data leveraging the GEO query package. Data are then further organized for downstream analyses. Trainees are then provided an overview of the types of molecular identifiers used in this example dataset, originally centered around microarray-based probeset identifiers. To increase interpretability of analysis findings, methods to merge platform-specific identifiers with gene-level annotation information are carried out. Example visualizations are then produced, including boxplots to evaluate the overall distribution of expression data across samples, as well as heat map visualizations that compare unscaled versus scaled gene expression values to emphasize the utility of scaled values for improved visualization of patterns between samples (Figure 10). Statistical analyses are then included to identify which genes are the most significantly altered in expression upon exposure to formaldehyde. The gene identified with the most significantly increased expression in the rat nose is olfactory receptor 633 (Olr633), demonstrating that formaldehyde inhalation exposure induced olfactory-related signaling. Together, this training module serves as an important example on how scientists can efficiently leverage existing genome-wide datasets to address new environmental health questions. Trainees are also pointed to previous publications applying these methods to existing GEO datasets that address additional environmental health questions (Rager and Fry, 2012; Rager et al., 2019). The importance of this module is that online -omics databases, such as GEO, represent robust resources that can be mined to better understand mechanisms of disease and biological responses to insults, and becoming familiar with such resources will expand data reusability and interpretation in future environmental health studies.2

The objective of this module is to provide an example analysis based on the integration of data across multiple environmental health databases. Specifically, air quality monitoring data from the U.S. EPA’s Air Quality System (AQS) (EPA, U.S, 2021a) were analyzed, focusing on the 2016 EPA Monitoring Data Annual Average database. These data included average measures of particles ≤2.5 μm in diameter (PM_2.5), nitrogen dioxide (NO₂), and sulfur dioxide (SO₂). Health outcome data were also analyzed, specifically from the Center for Disease Control (CDC) Wide-ranging ONline Data for Epidemiologic Research (WONDER) database (CDC, 2021). These data included the 2016 all-cause mortality rates. Population-level variables were additionally analyzed, including race, and included in the statistical modeling as well as the evaluation of population-level information that can be used to examine Environmental Justice issues. All data were pulled and summarized at the county-level across the entire U.S. for the year 2016 (Remington et al., 2015; UWPHI, 2021). The following environmental health questions were addressed through this training module:

1) What areas of the U.S. are most heavily monitored for air quality?

This module specifically guides trainees through an explanation of how the data were downloaded and organized, and then details the loading of required packages and datasets. Then, this module provides code for visualizing county-level air pollution measures obtained through U.S. EPA monitoring stations throughout the U.S. Air pollution measures include PM_2.5, NO₂, and SO₂, and are visualized here as the yearly average (Figure 11A). Air pollution concentrations are then evaluated for potential relationship to the health outcome, mortality (Figure 11B). Specifically, age adjusted mortality rates are organized and associated with PM_2.5 concentrations through linear regression modeling. Crude (univariate) statistical models are first provided that do not take into account the influence of potential confounders. Then, statistical models are used that adjust for potential county-level confounders, including adult smoking, obesity, food environment indicators, physical activity, employment status, rural vs. urban living percentages, sex, ethnicity, and race. Results from these models point to the preliminary finding that PM_2.5 is associated with elevated county-level mortality rates. Previous studies have shown that minority populations reside closer to air pollution sources and as a result are exposed to poorer air quality. This is also seen in these data, with measured air quality differing by percent African-American race in each county. Race is then evaluated further in this analysis as a potential differentiating factor in the models. Here, data distributions are pulled for counties with the highest percentage of African-Americans (top 25%) as well as those with the lowest percentage of African-Americans (bottom 25%). Models associating PM_2.5 with all-cause mortality rates are then re-run in these groups and the PM_2.5-mortality associations are compared. Counties with the highest percentages of African-American race had a significant association with mortality, with magnitudes substantially greater than counties with the lowest percentages of African-American race. This result corresponds with known Environmental Justice concerns, and demonstrates how even a cross-sectional, ecological analysis can highlight differences in environmental health risks. The importance of this module is that it demonstrates ways to integrate disparate health and environmental exposure databases in order to study key questions in environmental public health, including examinations of important Environmental Justice issues.

A final module is included that lists additional resources to aid in the continued training of users on data management and analysis strategies. We specifically include online websites and other training resources that we have found to be useful towards programming and data analysis approaches. These resources are sorted into the following four categories: 1) R programming resources; 2) R packages resources; 3) community discussions on R and R packages; 4) R interfaces; and 5) data science and statistical analysis resources.