Distributional impacts of fleet-wide change in light duty transportation: mortality risks of PM_2.5 emissions from electric vehicles and Tier 3 conventional vehicles

For all scenarios, we use estimates of VMT at the census tract level using multiple data sources [40, 59, 60] and as explained in detail in SI section S.1. We assume that census tract level miles driven are the same across scenarios. In the SI section S.6c, we also run our analysis using county-level emissions as the National Emissions Inventory, one of our primary sources of emissions from the current LDV fleet, reports total emissions at the county level and discuss the differences in the results. Across all scenarios, dust suspension and tire break-and-wear emissions of ICVs and EVs are not included. However, emerging evidence shows EVs may have smaller brake wear emissions due to regenerative braking but larger tire wear emissions due to higher weight [61, 62]. Internal combustion engines equipped with selective catalytic reactors are an emerging source of ammonia due to 'ammonia slipping,' which occurs due to non-optimal temperatures in the exhaust chamber [63–65]. We do not include this effect in the main results as emissions standards currently do not regulate ammonia, but it could be an emerging source of secondary PM_2.5.

We use population, race, and ethnicity data from ACS 2016–2020, obtained via NHGIS IPUMS [53]. We use eight racial-ethnic groups. People who self-identify as Hispanic or Latino ethnicity are included here as Latino (all races). The seven racial groups in ACS (i.e. Black, White, Asian, Native American & American Indian (Native), Hawaiian & Pacific Islander (HPI), Other, and Mixed here refer to non-Hispanic individuals.

Census block group-level population data are distributed to the InMAP grid as an area-weighted average (figure 1). The total population is 324.41 million, out of which 195.5 million are White, 59.1 million are Hispanic/Latino, 39.94 million Black, 17.6 million Asian, 8.8 Mixed, 1.96 million Native American and American Indian, and 0.4 HPIs.

National Emissions Inventory (NEI 2017) reports that total on-road non-diesel LDVs drove 2.65 trillion miles, emitting 1.9 million short tons of nitrogen oxides (NO_x ), 1.5 million short tons of VOCs, and 0.05 million short tons of primary PM_2.5 [40]. We use county-level emissions of NO_x , NMOG (VOC), and primary PM_2.5 reported by NEI and redistribute them to census tracts based on VMT estimates described above to model the health consequences of emissions as our current LDV fleet. Ammonia emissions for the current LDV fleet are not included in the analyses in the main text to enable comparison with emissions standards that do not regulate ammonia. Change in pollution exposure with changes in emissions is modeled using InMAP, a reduced complexity air quality model.

Emissions standards set quantitative limits of pollutant emissions that new vehicles can emit per mile. Between Tier 1 (1994) and Tier 3 standards (2022), per-mile gasoline NO_x + NMOG and PM_2.5 limits under regulatory test conditions have decreased by 91% and 97%, respectively. For this work, we rely on model year 2022 Tier 3 emissions standards in the FTP drive cycle and supplemental FTP drive cycle as two possible scenarios (SI section S2a). We use fleet-wide average emissions standard values for NO_X and NMOG. Since Tier 3 standards regulate total NO_x + NMOG emissions, we assume different ratios of NO_X and NMOG (50:50, 30:70, or 70:30) to account for uncertainty [66] (SI sections S2b and S6b). We assume that all vehicles meet the PM_2.5 mandated standard. Because the emissions of new Tier 3 vehicles will change with age and cumulative mileage, our estimates represent a lower bound of health damages for ICVs. Table 1 shows the emissions per mile for Tier 3 emissions standards (model year 2022) on FTP and SFTP drive cycles with different NO_x and NMOG ratios.

We use results from previous works to estimate a range of electricity consumption if the entire existing stock of vehicles were to be substituted by electric vehicles [54]. In brief, the energy consumption of electric vehicles is significantly impacted by both temperature and type of driving. We estimate the electricity required by an electric vehicle by using publicly available laboratory-tested data on energy consumption per mile at different temperatures, drive cycles, and vehicle miles traveled, as described above. Our estimates account for the effect of hourly ambient temperature at the county level and whether the type of driving in a county is predominantly city, highway, or combined driving. We use the Nissan Leaf (economy car, 40 kWh battery) and the Tesla Model S (luxury car, 100 kWh battery) as reasonable low and high bounds of the energy requirements of EVs (SI section S.3a) [67]. We assume vehicles are charged with the fleet of electricity generators in their NERC regions (SI section S.5b). NERC regions roughly divide the contiguous United States into six regional reliability entities and differ in power systems characteristics and resources. In table 2, we report the range of average energy requirements for Nissan Leaf and Tesla Model S across counties in different NERC regions, as estimated in our previous work, and the rough geographic representation of each NERC region by sub-regions of the United States.

We assume the electricity generated to meet the electricity demand from charging the vehicles will be distributed across power plants proportionally to their annual 2019 generation. We ignore the potential redispatch of power plants due to the demand for EV electricity. We also do not account for potential electricity generation capacity limits for a plant that may occur at high charging levels or for marginal generator characteristics. This assumption is suitable given that we are considering scenarios where many vehicles are being replaced, and we are using annual emissions from the electricity grid. We estimate plant-level emissions of SO₂, and NO_X using NEI 2017 and primary PM_2.5 e-GRID 2019 data [40, 50, 68]. Our dataset includes 3342 fossil power plants out of a total of 3400. Our estimates of total emissions are 0.13 million short tons for PM_2.5, 1.08 million short tons for SO₂, and 1.16 million short tons of NO_x . We exclude from our analysis 30 renewable plants that have a small percentage of oil, coal, or gas. To our knowledge, this is the most exhaustive and up-to-date characterization of the current electricity system for evaluating air pollution-related health impacts.

For our analysis of the future clean grid where 50 most SO₂ polluting power plants were retired or retrofitted with CCS, we assume that they are either replaced by renewable energy or did not have a reduction in their generation due to CCS. All retired or retrofitted plants in the continental US, except one, are coal power plants (SI section S.3d). These 50 power plants constituted 78 GW in nameplate capacity, 335 TWh of generation, and emitted 253 kilotons NO_x , 663 kilotons SO₂, 20 kilotons PM_2.5, and 364 447 kilotons CO₂ emissions. For context, cumulative wind and solar capacity in the US are nearly 136 GW and 73.5 GW [69, 70] and 300 GW of wind and 947 GW of solar are currently in transmission interconnection queues [71]. A complete list of power plants, state-wise capacity, generation removed or retrofitted, and emissions avoided are given in SI section S.3d.

Table 3 shows the total electricity generation in each NERC region, the percentage of renewable generation, and the total emissions of SO₂, NO_x and, PM_2.5, and CO₂ for 2019 (more details in SI sections S.3b and c).

We use InMAP, a reduced complexity air quality model [39, 45] to estimate the change in concentration of PM_2.5 across our different scenarios. InMAP uses an Eulerian grid model to estimate the annual average PM_2.5 concentration change attributed to a change in annual emissions based on steady-state solutions to equations representing pollution emission, transport, transformation, and removal. InMAP's representation of chemistry and meteorology uses spatially varying parameters obtained from a detailed chemical transport model simulation (the WRF-Chem model coupled with the National Emission Inventory). The InMAP source-receptor matrix (ISRM) provides a relationship between source damages (i.e. damages across the country that are attributable to emissions in a specific grid cell) and receptor damages (i.e. the damages suffered by people in a particular grid cell, regardless of where the emissions occurred). This ISRM was built by running more than 150 000 runs of InMAP, each time inputting a 1-t emission change from a single grid (out of ∼50 000 grid cells) at three different heights (ground, medium, and high height). ISRM as a dataset contains estimates of a linear relationship between marginal changes in emissions at every source location and marginal changes in annual-average PM_2.5 concentration at a receptor location [44, 72]. The grid-cell size in InMAP varies from 1 km × 1 km (typically in urban areas) to 48 km × 48 km (typically in rural areas), depending on the gradient in the population density and pollutant concentrations. Fine grid in populated areas is critical to accurately estimate air pollution-related health disparities [43, 73]. Our study explicitly looks at receptor damages (where the damages occur) along with a granular characterization of where damages originate (sources).

Premature mortality due to PM_2.5 concentration changes is calculated using the county-wide hypothetical 'underlying incidence,' the mortality hazard ratio derived from the concentration-response function, and the population in the grid cell. We use the approach of Apte et al [74] to calculate the hypothetical underlying incidence rate if there were no PM_2.5 emissions, ${I_{o,{\text{ }}c}}$ as:

$$\begin{align}{I_{o,c}} = \frac{{{I_c}}}{{\overline {{\text{H}}{{\text{R}}_c}} }}\end{align}$$

where ${I_c}$ is the reported county-level all-cause mortality, $\overline {{\text{H}}{{\text{R}}_c}}$, is the average mortality hazard ratio caused by PM_2.5 in county c. $\overline {{\text{H}}{{\text{R}}_c}}$, in turn, is calculated as:

$$\begin{align}\overline {{\text{HR}}_c} = {\text{ }}\frac{{\mathop \sum \nolimits _{i = 1}^n{P_i} \times {\text{ HR}}\left( {{C_i}} \right){\text{ }}{f_{i,c}}}}{{\mathop \sum \nolimits _{i = 1}^n{P_i}}}\end{align}$$

where ${P_i}$ is the population in grid cell i; n is the number of grid cells in county c; ${\text{HR}}\left( {{C_i}} \right)$ is mortality hazard ratio (GEMM in this study) resulting from ambient PM_2.5 concentration ${C_i}$ i, and ${f_{i,c}}$ is the area fraction of grid cell i that overlaps with county c. We use ambient PM_2.5 concentration from Meng et al for 2019 [75]. The authors estimate ambient PM_2.5 concentration using chemical transport modeling, satellite remote sensing, and ground-based measurements. We use baseline population-wide all-cause mortality rates at the county level from the US Center for Disease Control (CDC) [76] for 2019. Our unit of analysis is limited to the county level as this is the smallest geographic unit with publicly available health data from the CDC. While this CDC data has been used extensively in studies to examine health outcomes and environmental disparities in disadvantaged communities [77–79], health data at finer spatial resolutions would increase the certainty and accuracy in exposure and health disparity estimates. Unfortunately, there is no such data in a publicly available form. The average mortality rate ${I_c}$ is 833.71 deaths per 100 000 people per year, and our ${I_{o,{\text{ }}c}}$ estimate is 790 deaths per 100 000 people per year (SI section S5d).

Throughout this work, we use the mortality hazard ratio ${\text{HR}}\left( {{C_i}} \right)$ function from the Global Exposure Mortality for non-accidental mortality (GEMM) to calculate both the underlying incidence rate mentioned above and change in premature mortality due to change in PM_2.5 concentration from our scenarios. Non-accidental mortality in GEMM corresponds to non-communicable diseases and Lower Respiratory Infections and uses data from 41 cohort studies from 16 countries to estimate the shape of the association of ambient PM_2.5 exposure and non-accidental mortality. GEMM function is supralinear at lower concentrations, near linear at high concentrations, and applies a counterfactual threshold of 2.4 μg m⁻³, the lowest concentration observed in any of the cohort studies. GEMM function assumes that PM_2.5 exposure does not affect health below this level. Some versions of the GEMM function are parameterized differently depending on whether the function is segmented by age and whether a cohort of Chinese men with a wider PM_2.5 exposure range than the other studies is included. Our version applies a single function to estimates of non-accidental mortality for all ages above 25 and includes the Chinese male cohort [80]. We use GEMM's equation that uses non-accidental mortality for all ages, which is as follows:

$$\begin{align}{\text{HR }}\left( {{C_i}} \right) = {e^{0.143 \times \left( {\frac{{\text{ln} \left( {\frac{{\text{max} \left( {0,{\text{ }}{C_i} - 2.4} \right)}}{{1.6}} + 1} \right)}}{{1 + {e^{\frac{{ - \left( {\text{max} \left( {0,{\text{ }}{C_i} - 2.4} \right) - 15.5} \right)}}{{36.8}}}}}}} \right)}}\end{align}$$

where ${\text{HR}}\left( {{C_i}} \right)$ is the hazard ratio of mortality incidence at PM_2.5 concentration ${C_i}$ in units of micrograms per cubic meter—compared with a hypothetical underlying incidence rate, ${I_o}$ in the absence of ambient PM_2.5. Comparison of GEMM with other dose-response functions is given in SI section S.4.

Mortality associated with the concentration of PM_2.5 in a grid cell is computed as:

$$\begin{align}M\left( {{C_i}} \right) = {P_i}\mathop \sum \limits_c^n {I_{o,{\text{ }}c{\text{ }}}}{f_{i,c}}{\text{HR}}\left({C_i}\right)\end{align}$$

where $M\left( {{C_i}} \right){\text{ }}$ are premature deaths caused by the concentration of PM_2.5 at location i, ${P_i}$ is the population in grid cell i, ${I_{o,{\text{ }}c}}{\text{ }}$ underlying incidence rate for one $n$ counties overlapping grid $i$ and ${f_{i,c{\text{ }}}}$ is the fraction of grid cell $i$ that overlaps county $c$.

The transportation-attributable mortality rate is aggregated at the state and MSA levels as follows:

$$\begin{align} & {\text{Transportation attributable mortality rat}}{{\text{e}}_{{\text{pop or race}}}} \notag\\ & \quad = \frac{{M{{\left( C \right)}_{{\text{state or MSA}}}}{\text{ }}}}{{{{\left( {{\text{Populatio}}{{\text{n}}_{{\text{pop or race}}}}} \right)}_{{\text{state or MSA}}}}}}{\text{ }}\times{\text{ }}100,00\end{align}$$

For results across different scenarios, the concentration-response function, all-cause mortality, the underlying incidence rate, and the population counts remain the same. The changes in premature mortality are only due to changes in air emissions from different transportation choices. Tier-3 ICVs and EVs reduce total mortality by 80%–92% compared to the current fleet (SI section S6b). While the underlying assumptions and risk function differ, our estimates of premature mortality due to electric vehicles are in agreement with the most recent study by Peters et al [35]. We also present changes in PM_2.5 exposure from different transportation choices in SI section S6a.

The health benefits are particularly large for states and MSAs with large populations. Figure 2(a) shows the state-wide mortality rate of the current LDV fleet compared to two alternative scenarios: a fleet-wide transition to Tier-3 ICVs and EVs charged on the current electricity grid. Estimates for Tier-3 ICVs may, in practice, be higher in real-world conditions and will increase with age and cumulative mileage. In the future, electric vehicle emissions will be lower than our estimates suggest since the grid is expected to continue decarbonizing.

Zoom In Zoom Out Reset image size

Black circles and squares indicate Tier 3 emissions standards for FTP and SFTP driving schedules, respectively, with a ratio of 50:50 between NO_x and NMOG. Red and maroon circles indicate the mortality attributable to low-range and high-range EVs, respectively. The energy efficiency of EVs is calculated using temperature and urbanization level characteristics at the county level, as described in our previous work [54]. The carbon intensity of electricity for each state is obtained from the Power Sector Carbon Index [81, 82]. In figure 2(b), we show estimates of CO₂ emissions from the current LDV fleet, of Tier-3 ICVs (gray bars), and of the EVs considered in this study, using fuel economy of the latest Corporate Average Fuel Economy (CAFE) standards for passenger cars and trucks for 2021 (46.1, 32.6 MPG) and 2022 (48.2, 34.2 MPG) [83] and the fuel economy of current LDV fleet (23 MPG) [84].

We assume that the vehicle miles traveled are constant for both technologies and are derived from NEI 2017, as explained in the methods and SI section S.1. The secondary Y-axis in subfigures indicates the size of the population of each state (2019) [52] and state's power sector carbon intensity (2021) [82]. The states are ordered based on the decreasing mortality attributable to the current LDV fleet for each census region, i.e. South, Midwest, West, and Northeast (SI section S.5a). In the Western US, EVs have a lower mortality rate than Tier-3 ICVs, except for Wyoming, which still relies on a significant coal fleet. In the Northeast, EVs have similar health damages to Tier-3 ICVs in all states except Pennsylvania. In the Midwest, EVs have higher mortality than Tier-3 ICVs in most states, particularly in the Ohio Valley. In the Southern US, EVs have higher mortality rates than Tier-3 ICVs in most states but perform better in populous states like Florida, Texas, and Georgia. Altering NOX/NMOG ratios to 70:30 from 50:50 did not significantly change the results, with a 3% increase in total deaths (41 deaths). On the other hand, EVs have lower CO₂ emissions in all states except West Virginia (state carbon intensity of 876 gCO₂/kWh) (figure 2(b)). Removing or retrofitting 50 most SO₂ plants achieves health damages parity between EVs and Tier-3 ICVs for almost all states except West Virginia and Kentucky (figure 3).

Zoom In Zoom Out Reset image size

Figure 4 compares damages from changing the US LDV fleet to EVs and Tier 3 ICVs for the 50 most populous MSAs in the U.S. 55% of health damages (∼9 k out of 16 k deaths) from the current LDV fleet occur in the 50 most populous MSAs. Fleet-wide change to Tier-3 ICVs and EVs reduces mortality in all. EVs provide more health benefits than Tier 3 standards in all MSAs except for a few MSAs in the Ohio River Valley (figure 4(a)), where power plants are the predominant source of PM_2.5 [85, 86]. A future decarbonized grid can reduce electrification mortality in the region (figure 4(b)).

Zoom In Zoom Out Reset image size

Figures 5 and 6 present data on passenger vehicles' state-level mortality rates by race and ethnicity across scenarios. Bar height represents the US-wide population average mortality rate for each racial and ethnic group, while black lines indicate the population-level mortality rates by state. Previous research has shown that the current LDV fleet has a greater impact on people of color than on White Americans, with Blacks, Latinos, and Asians experiencing higher mortality rates than the population average [5, 29, 42, 87]. Our estimates are consistent with these earlier findings. Mixed-race also have higher mortality rates than the population average, while Hawaiian and Pacific Island groups have lower. Tier 3 vehicles reduce mortality rates across all groups, but differences between racial and ethnic groups persist. EVs have lower relative disparities than ICVs (figures 6 and 7). White Americans, on average, face higher health consequences than the population average for electric vehicles charged on the current grid, particularly in states in and near the Midwest (Pennsylvania, Indiana, Illinois, Virginia, Maryland) (figure 6, SI section S.6d). Black and Mixed-race Americans also face higher mortality than the population average with EVs charged on the current grid. With a future grid, Black and Mixed-race Americans continue to face higher mortality compared to the population average, but the relative disparity for White Americans declines. Latinos, the second largest ethnic group in the US, face lower transport-attributable mortality compared to the population average with electrification, both with current and future electricity grids (figure 6).

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

Zoom In Zoom Out Reset image size

To further explore the potential of each technology to reduce pollution disparities, we show the risk gap for states and MSAs in figure 7. The risk gap is the difference between the mortality rate of the most burdened race or ethnicity and that of the overall population of a state or MSA. It is a race-agnostic term used only to capture the differences in pollution disparities between transportation choices. The states and top 50 most populous MSAs are arranged in increasing order of risk gap for EVs. The risk gap decreases with fleet-wide shifts from the current fleet to either Tier 3 ICVs or EVs. An overall switch to EVs (current grid) leads to a lower or almost comparable risk gap with a few exceptions (St. Louis, New York, Houston, and Cincinnati MSAs, and Wyoming, Arkansas, Delaware, West Virginia, and New York states).

Electrifying transportation moves the air pollution from the tailpipe in urban areas to smokestacks of power plants, usually located in areas far from cities [24]. Our results show that while EVs and Tier 3 ICVs both reduce transportation-attributable mortality, they affect rural and urban populations, races and ethnicities, and income groups differently. Figure 8 displays the dependence between income and transport-related mortality categorized by various races and ethnicities for urban and rural populations. The mortality rates are compared across the median household income data for census tracts from ACS 2016–2020 for the current LDV fleet (top row), full electrification using a high-range EV charged on the current grid, and Tier 3 vehicles (FTP drive cycle, 50/50 ratio assumed). The colored lines and marker size correspond to race/ethnicity and their population in the income brackets, and the black line represents the population's average mortality rate. Tracts with a population density above 500 per square mile are defined as urban, while those less are rural as per the 2020 Census urban areas criteria [88].

Zoom In Zoom Out Reset image size

Both technologies reduced mortality rates for all income groups, races, and ethnicities compared to the current fleet. White Americans have lower mortality compared to the population average in all scenarios except in low-income census tracts. Transport-attributable mortality decreases with increasing income for Asians in urban areas. However, similar trends do not hold for Black and Latino Americans, who see an increase in mortality with income, with a striking increase in urban areas. Our results suggest that within richer urban census tracts, Black and Latino residents have significantly higher transportation-attributable pollution mortality than the population average. Latinos in rural low-income census tracts, though not in high numbers, have disproportionately high mortality with conventional vehicle technologies. Similar figures for tier 3 ICVs on SFTP drive cycle and high-range EVs charged on the future grid where the top 50 SO₂ polluting plants are retired are available in SI section S6e.