Air pollution mortality from India’s coal power plants: unit-level estimates for targeted policy

Air pollution from coal-fired electricity generation is an important cause of premature mortality in India. Although pollution-related mortality from the sector has been extensively studied, the relative contribution of individual coal-fired units to the fleet-wide mortality burden remains unclear. Here, we find that emissions from a small number of units drive overall mortality. Units producing just 3.5% of total generation and constituting less than 3% of total capacity result in 25% of annual premature mortality from coal-fired generation. This is a direct consequence of the 200-fold variation that we find in the mortality intensity of electricity generation across units. We use a detailed emissions inventory, a reduced complexity air quality model, and non-linear PM2.5 concentration-response functions to estimate marginal premature mortality for over 500 units operational in 2019. Absolute annual mortality ranges from less than 1 to over 650 deaths/year across units, and the mortality intensity of generation varies from under 0.002 to 0.43 deaths/GWh. Our findings suggest the potential for large social benefits in the form of reduced PM2.5-related premature mortality in India if the highest mortality intensity units are prioritized for the implementation of pollution control technologies or accelerated retirement.


Introduction
Exposure to ambient fine particulate matter (PM 2.5 ) pollution is estimated to account for 0.98 million (95% uncertainty interval: 0.77-1.19 million) premature deaths per year in India, approximately 10% of total annual deaths in the country [1].Emissions of PM 2.5 and its precursor pollutants from coal-fired power plants are a major contributor to this mortality [2].Numerous studies have estimated the aggregate premature mortality burden of India's coal fleet by combining air quality models with concentration-response functions [2][3][4][5][6], as well as quasi-experimental methods [7], and their estimates range from just under 50 000 to approximately 100 000 deaths per year.Previous research has also investigated the role of implementation of air quality norms, finding that inadequate implementation may lead to an additional 59 000 PM 2.5 -related premature deaths per year by 2040 [8].
Understanding the total premature mortality attributable to all coal-fired power plants is valuable for estimating overall damages from the sector, but insufficient for developing targeted policies to improve public health.For that, we require plant or unit-level estimates of marginal premature mortality.Such estimates can, for example, allow policymakers who are constrained to retiring a limited amount of capacity for cost and reliability reasons to prioritize units that deliver the greatest mortality-reduction benefit.Our study builds on recent work that has begun to address plant and unit-level variability in health impacts from the perspectives of emissions intensities [9] and population exposure to primary PM 2.5 emissions [10].
Previous studies of fleet-wide premature mortality impact typically make use of full-scale Eulerian chemical transport models (CTMs) such as the Comprehensive Air Quality Model with Extensions (CAMx) [3,4,6], WRF-Chem [2], and GEOS-Chem [11].CTMs are sophisticated tools considered state-of-the-science for estimating the impact of emissions reductions on human health.The sophistication comes at the cost of implementation complexity and computational expense, with a single simulation taking several days to run even on a high-performance computing system [12,13].Consequently, studies that use full-scale CTMs are typically able to run only a handful of simulations.The general approach for estimating mortality impacts of a set of power plants involves a CTM simulation that includes a complete emissions inventory inclusive of the power plants and one that excludes emissions from those power plants.The difference in spatially resolved total PM 2.5 concentrations is attributed to the operation of those power plants.To estimate PM 2.5 attributable to each plant would be computationally prohibitive because it would involve as many simulations as there are plants, and to do so for individual units (a single power plant can have multiple electricity generating units) would be even more intractable.
Here, we use a reduced complexity model (RCM), the global Intervention Model for Air Pollution (Global InMAP) to estimate the relationship between each unit's estimated emissions and the change in total PM 2.5 concentrations attributable to those emissions.RCMs were originally developed by air quality researchers to address the challenges posed by CTMs.Although they use simpler representations of atmospheric chemistry and physics than full-scale CTMs, RCMs are more accessible and require considerably less computational power, which enables many more scenarios to be analyzed [14].Global InMAP combines annual average physical and chemical data from a full-scale CTM (GEOS-Chem) with a simplified model of atmospheric chemistry to estimate changes in PM 2.5 concentrations from marginal changes in spatially resolved emissions [13].Global InMAP outputs spatially resolved annual average pollutant concentrations by estimating a steady-state solution to a simplified reaction-advection-diffusion equation and accounts for both the long-distance transport of PM 2.5 and the formation of secondary PM 2.5 .
InMAP has previously been extensively used in studies focused on the magnitude and distribution of emissions-related health damages from a variety of sources in the United States [15][16][17][18][19][20], and its extension to the global domain has been used to address similar questions in India [21].
We produce unit-level, spatially resolved estimates of annual average PM 2.5 concentrations attributable to each unit by combining daily electricity generation reports from India's Central Electricity Authority (CEA) for the calendar year 2019 with unitspecific emission factors for SO 2 , NO 2 , and PM 2.5 from prior research [22,23], and running one Global InMAP simulation for emissions from each of 510 units, reflecting a combined capacity of just under 175 GW.This represents just under 90% of India's total of 197 GW of coal and lignite-based capacity in 2019 [24,25].
The marginal premature mortality for each unit in each InMAP grid cell from a given disease endpoint is estimated by combining baseline ambient PM 2.5 concentration estimates and the integrated exposure response (IER) curves from the 2019 Global Burden of Disease (GBD) study [1,26].Through our study, we provide unit-level marginal mortality estimates for a substantial subset (∼90%) of India's coal-fired generators that can help identify high-priority units for the implementation of pollution control technologies or retirement.

Integration of generation data and emission factors
We combine daily generation reports provided by the CEA, unit-specific emission factors for primary PM 2.5 , SO 2 , and NO 2 , and unit-level metadata from the Global Power Plant Database (GPP) v1.3.0 to produce unit-level, daily resolution time series for 510 units [22,27].Power plant names and unit identifiers can vary across the data sources, so only units that can be identified in all three data sources are included in the final analysis.

Global InMAP simulations
We create one shapefile for every unit that contains its annual emissions of PM 2.5 , SO 2 and NO 2 in 2019 (see figure S14 for an overview of the process).Since InMAP allows for elevated emissions, we also provide estimates of stack height, diameter, flue gas velocity, and flue gas temperature.In the absence of a database containing these parameters for Indian coalfired units, we use approximate values derived from a combination of regulations, prior research and publicly available environmental impact assessments for some coal power plants [3,[27][28][29].Given the high degree of uncertainty, we conduct extensive sensitivity analysis to understand the potential impact of errors in these parameters on attributable PM 2.5 and marginal mortality (figure S15).
We run InMAP once for each unit in our dataset, providing the unit-level shapefile as the input.Each simulation takes approximately 6 h on our server.Across the 510 units, this takes approximately 3060 h but we were able to run up to four simulations simultaneously depending on server availability and completed the simulations in approximately twelve weeks.Although time-intensive, this approach is necessary to estimate the marginal increase in PM 2.5 in each grid cell that is attributable to each unit.Each simulation yields an estimate of total PM 2.5 attributable to a given unit across a global computational grid containing ∼1.5 million grid cells, which we clip down to the ∼27 000 grid cells in India.

Estimating counterfactual PM 2.5 concentrations
To estimate the mortality associated with the operations of a unit, we need an estimate of the total PM 2.5 concentrations in the absence of emissions from that unit.This is relevant because the dose-response functions used here are non-linear, and therefore the change in relative risk attributable to a 1 µg m −3 increase in PM 2.5 exposure depends on the prevailing level of exposure.
We estimate this counterfactual quantity by subtracting the PM 2.5 attributable to a unit in each grid cell from an estimate of the baseline PM 2.5 in that grid cell.For the baseline PM 2.5 in each grid cell, we rely on estimates that combine satellite measurements of aerosol optical depth, chemical transport modeling, and a geographically-weighted regression [26].We use estimates at the 0.05 • × 0.05 • resolution and use the mean of all available estimates in each InMAP grid cell, or match to the nearest available estimate when there are none available in the InMAP grid cell.

Mortality estimation
We consider six mortality endpoints: chronic obstructive pulmonary disease, lung cancer, lower respiratory tract infections, type 2 diabetes, stroke, and ischemic heart disease.The relative risk functions from GBD 2019 are modified to incorporate the theoretical minimum risk exposure level [21,30].We use premature mortality as the outcome of interest in this study, but we note that there are other potential outcome metrics that could be considered, such as disability-adjusted life-years, years of life lost due to premature mortality and years lived with disability [1].We choose premature mortality with the belief that it is more easily understood by policymakers, and it remains highly correlated with other adverse health outcomes.
We estimate the marginal mortality from each endpoint for each coal-fired unit in each InMAP grid cell according to the following equation [31]: ) where: 1. M i,j : change in premature mortality in grid cell i from endpoint j associated with a change in average PM 2.5 concentration from C * to C. 2. I j,k : baseline annual mortality rate from endpoint j in region k, where 'region' is defined as the state in which the grid cell is located.We use the state-level baseline mortality rate rather than a cell-level rate because this data is only available for states [32].3. P i : estimated population in grid cell i, from the Gridded Population of the World (GPW), v4 [33].4. RR j (C * i ) : relative risk of mortality from endpoint j in cell i at the baseline PM 2.5 concentration C * . 5. RR j (C i ): relative risk of mortality from endpoint j in cell i at the counterfactual PM 2.5 concentration C. 6. RR j,k : average population-weighted relative risk of mortality from endpoint j in region k.
For stroke and ischemic heart disease, the mappings from PM 2.5 exposure to relative risk vary by age group.We construct state-specific functions using the latest estimated demographic data available from the GBD 2019 study.In doing so, we assume that the age distribution of the population is homogenous within states.
To facilitate matching estimated exposures to the corresponding relative risk values, the relative risk values are linearly interpolated to 0.01 µg m −3 increments between 1 and 200 µg m −3 , the relevant range for baseline and counterfactual PM 2.5 concentrations over India (see figure S16 for an illustration of the effect of varying the interpolation step-size).
To estimate the absolute marginal mortality for a unit, we sum over all endpoints and India grid cells: The mortality intensity of generation for a unit is computed by dividing the absolute marginal mortality by the amount of electricity produced in 2019, and the mortality intensity of capacity is computed by dividing absolute marginal mortality by the installed capacity of the unit.
We estimate the marginal premature mortality for each unit rather than the average attributable mortality.The former refers to the premature mortality reduction that we would expect to see if a given unit is turned off, while the latter refers to the premature mortality that can be attributed to its ongoing operations.We choose the marginal approach to illustrate the consequences of removing emissions from individual units (figure S17).Approaches that aim to estimate average attributable mortality utilize the average slope of the non-linear curve [34].

Marginal mortality varies by ∼500× and marginal mortality intensity ∼200× across units
Marginal mortality displays considerable variation across units, ranging from less than 1 death/year to 670 deaths/year (588-719), with a median of 54 deaths/year (47-60).The range provided in parentheses corresponds to the range of estimated marginal mortality, computed by using the 5th and 95th percentiles of relative risk of mortality in the 2019 GBD study [30].The central estimate of marginal mortality is derived using the mean estimate of relative risk at any given exposure level.The range should not be interpreted as a 95% confidence interval for the marginal mortality estimate, since that is subject to multiple sources of uncertainty beyond the uncertainty in the relative risk functions.Since the uncertainty in several inputs has not been quantified, we conduct extensive robustness analyses to demonstrate the sensitivity of our results to errors in key inputs (section 3.3).
In all four of the main grid regions, most units have a marginal mortality below 200 deaths/year (figure 1).In the Northern grid, no units have a marginal mortality of more than 200 deaths/year.The highest marginal mortality is found at lignite-fired units in the Southern grid region in the state of Tamil Nadu (see figure S1 for spatial distribution of marginal mortality).
The cumulative marginal mortality estimated across the 510 units in this study in calendar year 2019 is ∼43 600 (38 000-47 600) deaths per year.Our estimates are on the lower end of estimates from previous studies on fleet-wide mortality [3,4,6,7], in part because we estimate marginal mortality rather than average attributable mortality, and we apply non-linear dose-response functions that are sensitive to estimates of ambient PM 2.5 concentrations (see figure S3 for a quantification of the sensitivity to ambient PM 2.5 estimates).
Across all units, marginal mortality intensity ranges from a low of around 0.0017 deaths/GWh (0.0015-0.002), to a high of around 0.43 deaths/GWh (0.38-0.46) (figure 2).The low and high estimates in parentheses correspond to intensity estimates computed using the low and high estimates of marginal mortality (deaths/year) for each unit.In contrast with the spatial pattern of marginal mortality, some units in the Northern region are among those with the highest marginal mortality intensity (see figure S2 for spatial distribution of marginal mortality intensity).
The age of the coal-fired unit does not appear to correlate with the mortality intensity of generation, as might be expected if newer units had systematically lower emissions intensities or were more favorably sited than older units from a population exposure perspective (figure S4).There is also no clear relationship between the load factor of a unit and its marginal mortality intensity (figure S5).The size, generation and emission factors are correlated with marginal mortality intensity.Mortality intensity above 0.2 deaths/GWh is only found in units with a capacity of less than 300 MW (figure S6) and those that produced less than 2 TWh in 2019 (figure S7).Lignite-burning units with extremely high SO 2 emission factors have extremely high mortality intensity (figure S8).
Mortality intensity varies considerably among units with similar levels of output and emission factors, underscoring the role of location (figure S8).For instance, we simulate the emissions and resulting PM 2.5 for the units in Mundra in Gujarat, and in Sasan in Madhya Pradesh which are assumed to have nearly identical emission factors for SO 2 (approximately 5 kg MWh -1 ) and primary PM 2.5 (0.6 kg MWh −1 ).Units in both locations produce 5500-6000 GWh yr −1 .However, the mortality intensity of generation for the units in Mundra is 0.063 deaths per GWh (0.055-0.07), more than twice as high as the 0.024 deaths per GWh (0.02-0.026) for those in Sasan.This is linked to the difference in population-weighted average PM 2.5 exposure from the emissions of both plants.The emissions from Mundra result in an increased annual average exposure of 0.028 µg/person, while those from Sasan increase exposure by only 0.012 µg/person because of substantially lower population in grid cells affected by the emissions from the latter compared to the former.
The marginal mortality intensity estimates in India are much greater than even the highest comparable estimates in the United States.An earlier study estimated a similar intensity metric across different US Regional Transmission Organizations (RTOs) using data from 2014 [16].The highest intensity was estimated in the midcontinent independent system operator (MISO) RTO at 0.0082 deaths per GWh.508 out of the 510 units in this study are estimated to have a higher mortality intensity.The median mortality intensity of 0.032 deaths/GWh is 3.8× the MISO intensity estimate.These differences are driven by a combination of differences in emissions intensities as well as population density in regions exposed to those emissions since both of those are substantially higher in India, but quantifying the relative importance of each of those drivers is beyond the scope of this analysis.

Eliminating emissions from a small number of units yields large reductions in mortality
An important consequence of the approximately 200fold variation in mortality intensity across units is the  highly non-linear relationship between the fraction of total electricity generation that a unit produces and its share of cumulative marginal mortality.In 2019, approximately 25% of the cumulative marginal mortality was linked to units producing just 3.4% of cumulative generation (figure S9).In absolute terms, this equates to avoiding approximately 11 000 premature deaths/year (9600-11 800) from eliminating emissions from 33 units that collectively produced 31.3TWh in 2019.These units had a combined capacity of 5.3 GW spread across 4 states.
Approximately 22 000 premature deaths (19 000-23 800), 50% of the cumulative marginal mortality, may have been avoided by removing emissions from units that produced less than 20% of total output in 2019.If mortality intensity is defined as deaths/GW of capacity, similar conclusions are reached (figure S10).
The highly disproportionate contribution of a small number of coal-fired units to overall mortality means that removing emissions from a small fraction of units may yield very large public health Figure 3. Small fractions of both cumulative generation and cumulative capacity are linked to large proportions of cumulative marginal mortality.Panel A shows cumulative marginal mortality against cumulative generation, both as a percentage of the total for 2019 units that continue to remain operational in 2022, with units ranked in descending order of marginal mortality intensity of generation (deaths/GWh).Panel B shows cumulative marginal mortality against cumulative capacity, with units instead ordered according to their marginal mortality intensity of capacity (deaths/GW).A quarter of cumulative marginal mortality (marked (a) in both panels) can be linked to approximately 3.5% of generation and 3% of capacity, and half of it can be linked to around 20% of generation and capacity (marked (b) in both panels).The units are colored by their load factor in 2019, with blue units operating with load factors below 40%, grey units between 40% and 80%, and red units above 80%.As seen in Panel B, selecting units for pollution control technology or retirement based on the mortality intensity of capacity deprioritizes interventions for units that have very low utilization rates.benefits.Figure 3(A) shows that even after excluding units that have been decommissioned since 2019, the non-linear relationship between cumulative marginal mortality and cumulative generation persists.Among the units that remain operational, onequarter of cumulative marginal mortality-around 10 600 deaths/year (9300-11 400)-may be avoided by eliminating emissions from 30 units that account for 3.5% of generation.This result holds with the alternate definition of mortality intensity in terms of deaths/GW (figure 3(B)) as well: around 25% of cumulative mortality may be avoided by retiring, or otherwise eliminating emissions from the 2.9% (5.03 GW) of still-operational capacity.
The units identified as high priority by the two metrics differ, however, and have different load factor profiles as shown in figures 3(A) and (B).Note: these changes do not account for new coal-based power plants that have been constructed since 2019 or those that may be commissioned in the future, additional retirements on technical, economic or environmental grounds, or changes in non-coal capacity such as the build-up of renewable energy (RE) capacity.These figures should, therefore, not be construed as net changes in states' electricity generation capacity.
The generation metric does not prioritize units with specific utilization profiles.The capacity metric deprioritizes units with low load factors since for a given amount of capacity, an extremely low load factor generally results in low generation, emissions, marginal PM 2.5 changes, and premature mortality compared with units operating at higher load factors.
A potential concern with prioritizing the retirement of high mortality intensity units is the impact on generation capacity in individual states.Although these units constitute only 3.5% of generation at a national level, the geographical clustering of high mortality intensity units may disproportionately affect generation capacity in certain areas.
These units are dispersed across four states, one in each of the four main grid regions (table 1).In Tamil Nadu, these units comprise 24% of installed coal capacity, but in the other three states, they comprise small fractions of total coal capacity.If retrofit or retirement costs are borne by states, it may be useful to also assess where the mortality reduction benefits will occur.We note that these are concentrated in-state.Of the 10 600 premature deaths that may be avoided each year, 80% are avoided in the four states where the units are located.The largest number of deaths are avoided in Tamil Nadu and West Bengal.
We expect meaningful mortality reduction even in the extreme case where generation is shifted away from the highest mortality intensity units to other units in the same state that had surplus capacity when considering total generation and capacity in 2019 (table S1).

Robustness of ordering by mortality intensity
Simulation-based approaches to attributing premature mortality to specific emission sources must contend with multiple sources of uncertainty, including emission factors, air quality models, and IER curves.Our primary finding regarding the highly disproportionate contribution of a small number of units to cumulative mortality emerges from the heterogeneity in marginal mortality intensity across units.The units identified as having disproportionately large mortality impacts are the ones with high mortality intensity relative to other units.Here we report the results of extensive sensitivity analyses that test the robustness of this ordering of units by mortality intensity.
To address the uncertainty in the IER curves, we check the relationships between marginal mortality and the marginal population-weighted increase in PM 2.5 , and their corresponding generation intensity metrics.Marginal mortality and populationweighted change in PM 2.5 are strongly positively correlated (figure 4(A)).Importantly, of the 30 units with the highest mortality of generation linked to 25% of cumulative mortality but only 3.5% of generation, 29 are those with the highest marginal increase in population-weighted PM 2.5 per GWh of generation (units marked in red in figure 4(B)).This indicates that the identification of units driving cumulative mortality is robust to uncertainty in the shape of the IER curves.
To assess the impact of uncertain emission factors, we re-simulate emissions and the resulting PM 2.5 concentrations from the 30 units with the highest mortality intensity of generation after adjusting each of their SO 2 emission factors by 20%.We focus on SO 2 emission intensity because its impact on marginal PM 2.5 and mortality dominates that of NO 2 and direct PM 2.5 (figure S11).For 21 out of the top 30 units, reducing the SO 2 emission factor by 20% still results in a mortality intensity of generation greater than that of the lowest intensity among these 30 units in the base case (figure S12).
Although uncertainty in the PM 2.5 estimates from Global InMAP cannot directly be addressed through such sensitivity analyses, we note that there remains a considerable margin of safety between the highest mortality intensity units and the rest.(figure S13).The variability in marginal mortality for a given level of PM2.5 exposure is a consequence of spatial variability in baseline PM2.5 concentrations and demographics.Panel B highlights in red the 30 units with the highest mortality intensity of generation that together account for 25% of cumulative marginal mortality while only producing 3.5% of total in generation.These units (with one exception) are also those that result in the highest population-weighted PM2.5 exposure.

Discussion
Our results underscore the importance of considering unit-level heterogeneity in policies that aim to reduce the health effects of electricity generation in India.This is relevant for policies for deploying emission control technologies such as flue gas desulfurization (FGD), as well as for the emerging policy discourse and research on the retirement of coal-fired units in India [10,[35][36][37][38].
Our results extend existing literature by providing unit-level mortality estimates that account for longdistance pollutant transport and secondary pollutant formation.This is valuable because attempts to identify the highest priority units for enforcing emissions norms or for retirement that use simple heuristics may miss several highly damaging units.
Arbitrary distance-based heuristics, for example, may miss units with high mortality intensity.This is relevant because current policy imposes the most stringent emissions standards on units within 10 km of a city with more than a million residents [39].As one example, regulators acceded to a request from Budge Budge Thermal Power Station in West Bengal to delay reducing its emissions on the grounds that it was not within 10 km of such a city [40].Although Budge Budge may be slightly more than 10 km from Kolkata, it has among the highest estimated mortality intensities when, as in our results, the long-distance transport of pollutants is considered.While increasing the threshold to a distance larger than 10 km will reduce the number of units that are missed, it would not address the underlying problem with a heuristic approach.There is no single distance threshold that is optimal across the country given differences in emission intensities, population density, and meteorology, and approaches such as the one taken in this study incorporate all those criteria.
Although our estimates reveal important variation, we highlight four avenues for improvement.First, the absence of publicly available continuous emissions monitoring system (CEMS) data for power plants in India necessitates the use of emission factors that are themselves estimates based on heat rates and weighted average coal characteristics [22], as well as approximations for stack characteristics.Future research may be able to validate the estimated emission factors and stack features if CEMS data from selected units becomes available.
Second, there is uncertainty in the PM 2.5 attributed to each unit.Future research may consider CTM simulations for selected units predicted to have high marginal mortality or mortality intensity, because CTMs capture more atmospheric phenomena and generally have greater accuracy than RCMs at predicting pollutant concentrations.Despite lower accuracy versus CTMs, the use of RCMs likely represents an improvement over heuristic approaches that miss long-distance pollutant transport and secondary pollutant formation.
Third, there is uncertainty in the estimation of changes in premature mortality risk from a change in PM 2.5 concentrations on three counts.Although the latest IER curves improve upon prior iterations by incorporating non-linearity across the range of exposures, there remains uncertainty, for example, regarding this relationship at very high levels of exposure as is the case across most of India [41].Given the numerous coal unit retirements between 2016 and 2021, future research in this area may consider evaluating the observed changes in mortality and other health outcomes in regions where the retired units contributed a meaningful share of ambient PM 2.5 .Recent studies have highlighted the potential for meaningful differences in the susceptibility of different populations to PM 2.5 [42].As higher resolution demographic data, and population-specific dose response functions become available, future research may seek to improve the accuracy of the mortality estimates presented here.
Recent research has also found that PM 2.5 from coal increases mortality risk more than other types of PM 2.5 [43].This suggests that using a coal PM 2.5 -specific IER may produce substantially higher mortality estimates.In the absence of such IERs for India, we are restricted to IERs that account for the exposure ranges seen in India.We note, however, that concerns about the IER do not affect the key result regarding the disproportionate impact of a small number of units on total mortality.As shown in section 3.3, these results are driven the populationweighted average change in PM 2.5 per unit, which is not affected by the choice of IER.
Fourth, our study looks only at premature mortality from PM 2.5 exposure via six disease endpoints, and thus is likely an underestimate of premature mortality from coal power plants.We also do not consider nonmortality, morbidity-related health damages including those in children, and other environmental damages such as climate change damages from greenhouse gas emissions.
While the uncertainties highlighted indicate the need for continued research, this study provides highly granular, unit-specific, estimates of marginal mortality and PM 2.5 for a comprehensive portion of the current Indian coal fleet.The estimates show valuable and robust heterogeneity that can be exploited to achieve meaningful public health gains from developing targeted policies in a country where sector-wide policies have previously proved difficult to enforce [44].The availability of unit-level mortality intensity estimates also enables future capacity expansion studies to consider the mortality reduction co-benefits of renewable energy (RE) expansion.Similar studies in the United States have shown considerable additional gains from explicitly considering mortality, and analogous research for India may enable it to maximize social benefits as it builds towards one of the world's most ambitious RE capacity targets [45].

Figure 1 .
Figure 1.Unit-level marginal mortality (deaths/year) varies ∼500-fold, with the highest mortality units located in southern and eastern India.Most units have an estimated marginal mortality below 200, with a median marginal mortality of approximately 50 deaths/year.

Figure 2 .
Figure 2. Marginal mortality intensity of generation (deaths/GWh) varies ∼200-fold across individual units.As with marginal mortality, the Southern grid region has most of the high mortality intensity of generation units.In contrast to the spatial pattern of marginal mortality, however, units with high marginal mortality intensity are not limited to just the Southern and Eastern regions but are present across the country.

Figure 4 .
Figure 4. Marginal mortality (Panel A) and marginal mortality intensity of generation (Panel B) are both strongly correlated with unit-level marginal population-weighted PM2.5 and marginal population-weighted PM2.5 intensity of generation, respectively.The variability in marginal mortality for a given level of PM2.5 exposure is a consequence of spatial variability in baseline PM2.5 concentrations and demographics.Panel B highlights in red the 30 units with the highest mortality intensity of generation that together account for 25% of cumulative marginal mortality while only producing 3.5% of total in generation.These units (with one exception) are also those that result in the highest population-weighted PM2.5 exposure.

Table 1 .
The 30 units with the highest mortality intensity of generation are concentrated in four states.While for Tamil Nadu, this could affect almost a quarter of the operational coal capacity, the units in West Bengal, Rajasthan, and Gujarat constitute a much smaller fraction of in-state coal generation capacity.