Grantee Research Project Results

Development and Evaluation of Air Quality Models

A major focus of CACES was the development and evaluation of both mechanistic and empirical air quality models. Mechanistic models use knowledge of chemical and physical processes to predict concentrations of PM_2.5 and other air pollutants. They complement purely empirical (statistical) models that are derived solely from observations. We used data collected by CACES (and archived data) to extensively evaluate model predictions.

The primary purpose of CACES modeling activities was to quantify the influence of modifiable factors on pollutant concentrations; to derive exposure estimates for health analyses; and to investigate disparities in pollutant exposure by race-ethnicity and income. The results from these applications are described in other sections of this report. This section focuses on CACES model development and evaluation activities

Mechanistic Modeling: CACES used two broad types of mechanistic models: chemical transport models (CTMs) and reduced-complexity models (RCMs). CTMs are highly detailed models requiring specialized expertise, large computing clusters, preparation of detailed input files, and generally months of time to analyze a set of scenarios. Reduced complexity models (RCMs) are newer and much simpler tools than CTMs. They are much faster and more accessible than CTMs and can be used to help link policy decisions to emissions to air quality and health outcomes. They offer the opportunity to bring rigorous consideration of air quality and public health impacts to decision analyses that previously omitted them, treated them simply, or considered only changes in emissions rather than changes in concentration and health endpoints. However, they are newer than CTMs and the necessary simplifications required to create them mean that they must be evaluated and used with care.

The CACES mechanistic modeling activities focused on RCMs. CACES investigators developed three RCMs: APEEP, EASIUR, and InMAP. Predictions from all three RCMs are available online (www.caces.us).

CACES RCM development activities included creation of very high spatial resolution (300 m) models for environmental justice analyses; development of a model that accounts for recent advances in our knowledge of organic particulate matter sources and chemistry (semi-volatile primary organic aerosol (POA), intermediate-volatility organic compounds (IVOCs), and multi-generational aging); and development of RCMs that can be used for applications outside the US. This work demonstrated the need for very high resolution (~1 km) analyses to capture race-ethnicity and other demographic differences in PM_2.5 exposure. It also demonstrates that health damages per ton of VOC emitted vary strongly across source categories. Finally, we explored using machine learning approaches to build RCMs that better preserve a CTM-like structure while speeding up the computations via neural network emulators of key processes such as chemistry.

Given the simplifications required to derive RCM, robustness of their predictions is a concern. To evaluate this issue, we compared the predicted marginal social costs of the three RCMs. That intercomparison showed that, despite being derived in fundamentally different ways, the three RCMs predicted similar marginal social costs. For example, nationally averaged social costs agreed to within 10-40% depending on species and emissions height. This range is indicative of the uncertainty one might expect when doing a cost-benefit assessment of a broad (spanning the country) policy. This level of uncertainty is manageable and boosted confidence in RCMs for use in policy analyses. An important conclusion for managing uncertainty is to apply multiple models that are based on fundamentally different simplifying assumptions to the same policy scenarios. Consistency between predictions provides confidence in conclusions; inconsistency highlight areas that require additional study.

The CACES CTM development focused on evaluating the value-added associated with higher spatial resolution (1 km horizontal resolution) simulations. Perhaps not surprisingly, higher resolution simulations significantly improved predicted concentrations of primary and/or short-lived species such as NO₂ and ultrafine particles (UFPs). In contrast, CTM predictions for total PM_2.5 were only marginally improved by increasing CTM spatial resolution. This is not surprising given that much of PM_2.5 mass is secondary and regional in nature. Probably the most important effect of resolution is better prediction of urban-rural gradients in PM_2.5. Total population exposure and, therefore, total health damages associated with PM_2.5 are very similar across 36 km, 12 km, 4 km, and 1 km model resolutions supporting the use of coarser model resolutions in environmental impact studies and other applications that prioritize total damages across the population.

Empirical Modeling: CACES developed a suite of national-scale, high spatial resolution empirical models to predict outdoor concentrations of multiple pollutants at census tract and census block group levels. We developed models for both criteria and non-criteria pollutants, including the first empirical models for UFPs, COA, HOA, and BC in the contiguous US. We also investigated model development using low-cost and mobile air quality data. Finally, we investigated novel land use covariates and different modeling frameworks. Predictions of many of these models are available for download via the CACES website (www.caces.us/data).

We developed a suite of empirical models to predict outdoor concentrations of criteria pollutant (O₃, CO, SO₂, NO₂, PM_2.5, PM₁₀) concentrations throughout the contiguous U.S. Model estimates are annual-average values for the years 1979 – 2015 (O₃, SO₂, NO₂), 1988 – 2015 (PM₁₀), 1990-2015 (CO), and 1999-2015 (PM_2.5). These models were derived using regulatory monitoring data, satellite information, publicly available land use variables, and a partial least squares – universal kriging modeling approach. We combined predictions of the CTM with the PM_2.5 empirical model to derive species and source-resolved PM_2.5 estimates at the county-scale for health analyses.

As part of a CACES-led ACE-center collaborative project, we compared predictions of our base criteria pollutant empirical models against six other (i.e., non-CACES) national empirical models. There is a relatively high level of agreement among the six national empirical models. For example, for year-2010 PM_2.5, pairwise correlation (“r”) values are between 0.84 and 0.92 and RMSD (root-mean-square-difference) values are 0.83 – 1.43 μg/m³.

We also developed national empirical exposure models for non-criteria pollutants, including ultrafine particles (UFP), traffic organic aerosol (HOA), cooking organic aerosol (COA), and black carbon (BC). This is a challenging problem because of the lack of routine monitoring data for non-criteria pollutants. We demonstrated that a nationally representative dataset can be generated for even sparsely monitored pollutants through a careful and targeted sampling design. We achieved this by using a hybrid approach that combines both mobile and fixed site measurements. The mobile measurements were used to characterize the intraurban spatial patterns in pollutant concentrations; the fixed site measurements from urban and rural locations were used to characterize the interurban and regional patterns in pollutant concentrations. We used this hybrid approach to create the first national exposure models that predict UFP and source-resolved PM_2.5 at high spatial resolution (census block group). In addition, the hybrid approach we developed is suitable for use for other pollutants not measured by national networks and/or require very expensive instrumentation. The models reveal that the spatial patterns of non-criteria pollutants were distinct from criteria pollutants with clear links to sources (e.g., restaurants for COA). Predictions of these models were used for health analyses at the census tract level.

We investigated the use of low-cost sensing data from the open-source PurpleAir PM_2.5 sensor network to derive national models. We found that hybrid land use regression (LUR) models that include the low-cost network may better capture within-city variation than models derived using data from regulatory networks and satellites.

Finally, we investigated the value of novel land use variables and algorithms for empirical exposure modeling. This included: (1) Landsat satellite-derived Local Climate Zones (LCZs) with spatiotemporally varying landcover variables with historical coverage (back to ~1980s); (2) Google Point of Interest (POI) data that can provide microscale information on sources (e.g., restaurants, gas stations) that are not well covered by existing covariates; (3) Yelp data that adds detailed information on restaurant type and location; and (4) object identification from Google Street View (GSV) imagery. We found that integrating the Google POI data improves performance of empirical exposure models. We also found that machine learning (ML) models outperformed traditional approaches (e.g., stepwise forward selection models). In addition, ML approaches allow for creation of high performing exposure models with only the new microscale covariates (POI, GSV, and LCZ) and satellite-based air pollution measurements, highlighting the utility of a flexible ML approach and supporting our use of new covariates for exposure model building.

Effects of Modifiable Factors

Reducing the air pollutant health impacts effectively, efficiently, and equitably requires attributing them to specific emission sources and other modifiable factors. CACES used both measurements and models to quantify the contribution of different sources air pollutant concentrations. We used measured data to quantify the influence of important urban sources (traffic and restaurant cooking) on intraurban spatial patterns of PM_2.5. We used RCMs to estimate the impact of different source categories in both economy-wide and sector-specific frameworks.

We used our spatially and source resolved measurements to investigate the influence of modifiable factors such as land use in air pollution concentration. In all three focus cities (Pittsburgh, Oakland, and Baltimore), traffic and cooking created more than half of the intraurban variability of PM_2.5. Restaurant density and commercial land use-related variables are important predictors for the spatial variability of COA. Transportation and urbanicity-related variables are important predictors for the spatial variability of HOA. The results highlight the the potential importance of controlling commercial cooking emissions for air quality management in the US.

Using RCMs, we estimated air quality-related health effects of 428 distinct sectors of the United States economy. All sectors are major contributors to excess mortality risk, not just those dominated by tailpipes and smokestacks. However, half of air pollution-related deaths are attributable to just five activities: (i) electricity generation, (ii) passenger vehicle use, (iii) industrial boiler and combustion engine use, (iv) residential cooking and heating, and (v) livestock rearing. The diversity of sectors demonstrates the need for economy-wide regulatory strategies.

We also used integrated assessment models (IAMs) to compute marginal damages for PM_2.5-related emissions for each county in the contiguous United States and matched location-specific emissions with these marginal damages to compute economy-wide gross external damage (GED) due to premature mortality. We found that economy-wide, GED has decreased by more than 20% from 2008 to 2014, underscoring the benefits of improved air quality. Four sectors, comprising less than 20% of the national gross domestic product (GDP), are responsible for ∼75% of GED attributable to economic activities. Given the substantial emission reductions in transportation and electricity generation, farms have become the largest contributor to air pollution damages from PM_2.5-related emissions.

CACES also performed sector specific analyses. Given the importance of agricultural in our economy-wide analyses, much of the CACES sector-specific analysis focused on agriculture and food production. Using RCMs, we estimated that agricultural production in the United States results in 17,900 annual air quality–related deaths, 15,900 of which are from food production. Of those, 80% are attributable to animal-based foods, both directly from animal production and indirectly from growing animal feed. On-farm interventions can reduce PM_2.5-related mortality by 50%, including improved livestock waste management and fertilizer application practices that reduce emissions of ammonia. Dietary shifts toward plant-based foods that maintain protein intake and other nutritional needs could reduce agricultural air quality–related mortality by 68 to 83%.

Food production also emits greenhouse gases. We found that even if fossil fuel emissions were immediately halted, current trends in global food systems would prevent the achievement of the 1.5°C target and, by the end of the century, threatening the achievement of the 2°C target. Therefore meeting climate targets will require rapid and ambitious changes to food systems as well as to all nonfood sectors.

In addition to agriculture and food production, we investigated the effects of changes in the electricity, transportation and natural gas sectors. Examples of this sector specific work include an analysis the implications of integrating health and climate when determining the best locations for replacing power plants with new wind, solar, or natural gas to meet a CO₂ reduction target in the United States; quantifying the effects of changing emissions standard and atmospheric conditions on secondary organic aerosol for gasoline vehicle exhaust; and quantification of the cumulative effects of the shale gas boom in the Appalachian basin from 2004 to 2016 on air quality, climate change and employment. The breadth and extent of these analysis required collaboration across the entire CACES team.

Environmental Justice

Environmental justice was an important cross-cutting theme of CACES. We used the wealth of spatially-resolved data along with our empirical exposure models to investigate exposure disparities by race-ethnicity and income at both the urban and national scales. Our analyses show that exposure disparities vary by pollutant, year, and U.S. state; however, there are no examples of states that do not experience disparities. Where there are disparities, non-Hispanic White people experience lower-than-average exposures, and the most-exposed group is a racial- ethnic minority. In addition, concentration differences among racial-ethnic groups are larger than among income categories. An important goal of CACES was to move beyond simply characterizing exposure disparities and to begin attributing these disparities to modifiable factors. For example, using source-resolved PM_2.5 measurements and RCMs, we investigated the contribution of different sources to the observed exposure disparities.

Using our national exposure models, we investigated exposure disparities by race-ethnicity and income. Our work considered changes in concentrations of six criteria pollutants from 1990 to 2010. We found that for all years and pollutants, the racial-ethnic group with the highest national average exposure was a racial-ethnic minority group. In 2010, the disparity between the racial-ethnic group with the highest versus lowest national-average exposure was largest for NO₂ (64%) and smallest for O₃ (4%). Absolute exposure disparities were much larger among racial- ethnic groups than among income categories. From 1990 to 2010, absolute racial-ethnic exposure disparities declined by between 35% (PM_2.5) and 88% (CO). However, in 2010, racial-ethnic exposure disparities remained across income levels, in urban and rural areas, and in all states, for multiple pollutants.

At the urban scale, measured spatial patterns in source-resolved PM_2.5 associated with modifiable factors leads to higher exposures for people of color. We showed that the magnitude of these disparities can vary from city-to-city, but the same disparities exist in all of our sampled cities and are driven by the same set of sources. Somewhat unexpectedly, commercial cooking created more disparities than traffic.

Analysis of air pollution data collected with Google Street View cars in San Francisco Bay Area revealed the complex interplay of the different spatial scales of air pollution and demographics. The data revealed that in the Bay Area systematic disparities by race-ethnicity are influenced by regional concentration gradients due to sharp contrasts in demographic composition among cities and urban districts, while within-group extremes arise from local peaks in pollutant concentrations. These findings could help inform the spatial scales for regulatory action.

Using RCMs, we investigated the specific emissions sources that contribute most to PM_2.5 exposure and disparities in exposure by race-ethnicity and income across the contiguous United States. We linked PM2.5 exposure to the human activities responsible for PM_2.5 pollution. We used these results to explore “pollution inequity”: the difference between the environmental health damage caused by a racial–ethnic group and the damage that group experiences. We showed that, in the United States, PM2.5 exposure is disproportionately caused by consumption of goods and services mainly by the non-Hispanic white majority, but disproportionately inhaled by black and Hispanic minorities. We found substantial geographic variation in the relative importance of individual source categories for exposure and exposure disparities. This suggests that there is no clear one-size-fits-all policy approach to mitigate inequitable exposure to air pollution across the United States. Instead, these findings and RCM models can help states prioritize efforts to further study and address the unique causes of environmental injustice within their borders.

Human Health

Using predictions of the CACES exposure models, we estimated multi-pollutant mortality risk surfaces using two large, unique, population-based U.S. datasets. We also investigated regional and temporal variability in risk and inequalities based on income, poverty, education and race. Several review/meta-analysis were also performed, including a review and discussion of 25 plus years of cohort studies and a formal meta-analytic analysis that was used to generate health impact functions to assess marginal changes in PM_2.5.

National Health Interview Survey (NHIS) Cohort Studies: We conducted a series of studies using cohorts constructed from the U.S. National Health Interview Survey (NHIS) data linked to the national mortality death index. We performed analyses using both unrestricted (publicly available) data for metropolitan statistical areas (MSA) and restricted data at the census-tract level. The restricted data were accessed through the NCHS Research Data Center (RDC).

We found elevated risks of mortality, including cardiopulmonary mortality and lung cancer, associated with PM_2.5 (Pope et al., Environ. Health Persp.,  2019). Causal inference techniques were used to test the sensitivity of these associations to modelling approaches. The estimates were generally robust to model selection. The analysis of the restricted NHIS data was extended to include multiple pollutants, including PM_2.5, PM_2.5-10, and SO2, NO₂, and O₃ (Lefler et al. Environ Health 2019). The association between PM_2.5 and mortality risk was the most robust and least sensitive to controlling for other pollutants. Smaller, less consistent associations were also observed for coarse fraction particulate matter (PM_2.5-10) and sulfur dioxide (SO₂).

We also investigated all-cause, cardiopulmonary, and cancer mortality associations with PM_2.5 species and sources, including ultrafine particle concentrations. Cardiopulmonary mortality associations varied by species and source with evidence that elemental carbon, secondary organic aerosols, vehicle sources and cooking may be important contributors to mortality risk. With further validation, these findings could facilitate targeted pollution regulations that more efficiently reduce air pollution mortality.

In single-pollutant models, UFP was significantly associated with all-cause and cancer mortality. However, in two-pollutant models, mortality associations varied based on co-pollutant adjustment. UFP mortality associations were robust to controlling for PM_2.5-10 and SO2, but not PM_2.5.

To further investigate associations between air pollution and cancer risk, a series of analyses were conducted using the SEER Cancer Registry county-level cross-sectional data. All cancer and lung cancer were associated with PM2.5. PM2.5 exposures were also adversely associated with cardiopulmonary mortality for cancer patients and survivors, especially those who received chemotherapy or radiation treatment. Greenness was associated with decreased risk of cancer mortality; PM_2.5 was associated with increased cardiopulmonary mortality (Coleman et al., Environ. Int., 2021).

County Mortality Data Studies Data on all deaths in the contiguous US from 1999 to 2015, with information on county of residence, were used to directly estimate the number of deaths, by age group and sex, and loss of life expectancy due to current PM_2.5 using time-series of PM concentrations from LUR models in Project 3. Estimates were based on novel Bayesian spatiotemporal models that directly estimate the proportional (percent) increase in county-level age-specific death rates from cardiorespiratory diseases associated with the county’s annual PM_2.5 concentration. We estimated that current air pollution concentrations are associated with a significant health burden after any reasonable control for other local determinants of mortality. Life expectancy loss due to PM2.5 was larger in counties with lower income than in wealthier counties, counties with a higher proportion of population whose family income is below the poverty threshold, with a higher proportion of population who are of Black or African American race, or with a lower proportion of population who graduated from high school, contributing to nationwide health inequalities. We also estimated that reductions in air pollution since 1999 have contributed to the longevity gains in the US population, with larger benefits especially in California and some southern states such as Alabama and Georgia, where PM2.5 declined more than the national average.

A thorough understanding of the long-term dynamics of seasonality of mortality, and its geographical and demographic patterns, is needed to identify at-risk groups, plan responses at the present time as well as under changing climate conditions. To address this need, we analyzed the seasonality of mortality by age group and sex from 1980 to 2016 in the USA and its subnational climatic regions, and investigated the impacts of temperature anomalies that are expected to increase with global climate change. We identified distinct seasonal patterns in relation to age group and sex, including elevated risks among young adults in the summertime. This led to an investigation if deaths from various unintentional (transport, falls and drownings) and intentional (assault and suicide) injuries might be affected by anomalously warm temperatures that occur today and are expected to become increasingly common as a result of global climate change. We defined a novel measure of anomalous temperature for each county and month, which represents the deviation from the county’s long-term average temperature in that month, which mimics the conditions that may arise with global climate change. We used a spatiotemporal model to estimate how cold and warm temperature anomalies are associated with deaths from various injuries. We found that a 1.5 °C anomalously warm year, as envisioned under the Paris Climate Agreement, would be associated with an estimated ~1,600 additional injury deaths mostly in adolescence to middle age.

Data democratization

A final goal of CACES was data democratization. Through the CACES support center, we provided multiple tools, models, and datasets in a publicly-available website, www.caces.us.

Download rates have grown over time. Currently download rates average ~70 data-downloads per month (i.e., more than 2 data-downloads per day), a rate that far exceeded our goals and expectations when we established the website.