Bureaucrat incentives reduce crop burning and child mortality in South Asia

Abstract

Air pollution in South Asia is a health emergency, responsible for 2 million deaths every year¹. Crop residue burning accounts for 40–60% of peak pollution during the winter harvest months^2,3. Despite being illegal, this practice remains widespread^4,5. Any solution to curb the problem necessitates government action at scale. Here we study whether leveraging the incentives of bureaucrats tasked with controlling burning can mitigate this phenomenon. Using a decade of wind, fire and health data from satellites and surveys from the Demographic and Health Surveys Program, we show that crop burning responds to bureaucrat incentives: fires increase by 15% when wind is most likely to direct pollution to neighbouring jurisdictions, and decrease by 14.5% when it pollutes their own. These effects intensify with stronger bureaucratic incentives and capacity. We also find that bureaucrat action against burning deters future polluters, further reducing fires by 13%. Finally, using an atmospheric model, we estimate that one log increase in in utero exposure to pollution from burning raises child mortality by 30–36 deaths per 1,000 births, underscoring the importance of bureaucrat action. Contrary to the growing beliefs that the problem of crop burning is intractable^6,7, these findings highlight specific ways in which existing bureaucrats, when properly incentivized, can improve environmental management and public health outcomes.

Around 1.8–2.7 deaths per 1,000 births (4.4–6.6% of the average child mortality) could be prevented in Pakistan and India if bureaucrats control crop burning across all areas of their jurisdiction as they do in places where fires would pollute their own district.

Main

Globally, air pollution kills over 7 million people annually⁸ and causes wide-ranging health problems that impair child development, educational attainment and worker productivity^9–11. Air pollution in South Asia is one of the largest public health emergencies on the planet¹: one in four humans live in this area, and they are exposed to hazardously unclean air (air pollution in Lahore is shown in Fig. 1). Estimates suggest that crop residue burning accounts for between 40% and 60% of air pollution during the winter harvest months^2,3. Although it is illegal, crop burning continues to be widely practiced in the region^4,5.

Fig. 1. Lahore, Pakistan on clean (left) and polluted (right) air days.

The photograph on the left is from 12 April 2020, when the air quality index was 37. The photograph on the right is from 2 January 2020, when the air quality index was 496. Photograph credit: Dawar Hameed Butt.

Despite the critical role of government in environmental management at scale, how institutions and state administrators control crop burning remains poorly studied. First, previous research has paid little attention to the conditions under which governments implement successful policies to manage the environment¹². Second, the literature on bureaucracy and development is only in the early stages of examining “whether bureaucrats can innovate and adapt to future challenges”, such as environmental degradation¹³. Theoretically, overloaded bureaucrats might have weak incentives to reduce crop burning, as fires are caused by thousands dispersed farmers and punishing any one of them will only reduce pollution minimally. In practice, South Asian state administrators are thought to have little to no impact over crop burning, which is increasingly considered intractable^6,7.

We studied whether and under what conditions bureaucrats exert effort to reduce pollution. We present a systematic large-scale analysis of the political economy determinants and consequences of crop burning in India and Pakistan, combining evidence from a region in which 500 million, or 7% of people, live and whose decisions impact the air quality of over 2 billion people. We use 10 years of satellite and administrative data at the micro level (5 km² grid cells) for  around 18 million observations across India and Pakistan to understand how state administrators manage crop burning. Our analysis affords new insights on how control over crop burning might be established.

We first examine whether bureaucrats allow crop burning when the pollution impacts are borne by neighbouring jurisdictions—as shown for industrial emissions in previous work^14–16—but also whether bureaucrats reduce crop burning when their own jurisdiction gets polluted. This second behaviour would run contrary to the idea of bureaucratic inaction, and underscore the importance of an underappreciated channel of control: the incentives of bureaucrats already tasked with managing the problem.

Second, we examine whether punitive action by bureaucrats against polluters creates deterrence among future polluters precisely in areas where bureaucrats have a greater likelihood of acting in the future. If this is indeed the case, bureaucrats could be motivated to act even in the presence of weak incentives and small returns from targeting any one farmer as their action can translate to larger impacts.

Finally, we link bureaucratic behaviour with infant and child mortality to examine whether bureaucratic decisions matter for pollution and public health. A large body of literature estimates the impacts of air pollution on infant and child mortality^3,11,17–21. Our approach to estimate the mortality impacts of crop burning resolves key challenges in the literature by combining variation in timing of pregnancies with spatial exposure to a concentration of particulate matter ≤2.5 µm in diameter (PM_2.5), in an instrumental variables framework. With estimates of the impact of fires on mortality, we propose a back-of-the-envelope calculation to show how much mortality could be reduced if bureaucrats control crop burning across all areas of their jurisdiction as they do in places where fires would pollute their own district.

Air pollution management in South Asia

The cost of poor air for governments and society is immense and estimated at US $2.9 trillion, corresponding to 3.3% of the global gross domestic product¹⁹. The levels of particulate matter in the air are highest in the developing world, particularly in South Asia.

Crop burning in South Asia

One of the largest contributors to PM_2.5 is agricultural emissions and, in particular, crop residue burning—its impact on air quality is estimated to range from 9% when considering annual averages in the entire subcontinent²² to between 40% and 60% when focusing on northern India and the winter harvest months, the time of the year with the highest pollution levels^2,3 (Fig. 1). Burning is used as a fast and inexpensive method for clearing crop residue after harvest and before the next cropping cycle. Farmers echo this sentiment: “If I can clear my farm using a one-rupee matchbox, why will I spend thousands?”²³, where the latter refers to an alternative method of clearing stubble, such as renting mechanized harvesters.

Several factors have contributed to the severity of the problem. First, the Green Revolution spurred the ‘rice-wheat cropping system’²⁴, which is characterized by rapid harvest and sowing cycles so that crops can catch prime conditions for their growth. Second, burning increased by 39% after the introduction of water conservation legislation that compressed the harvest period for rice⁴. Third, greater use of mechanization technologies produced larger crop residues that needed to be cleared²⁵. Finally, a high labour wage due to sustained out-migration from rural areas further increased the use of these mechanized tools^5,24,26.

Unfortunately, the social costs of burning far outweigh its convenience^27,28. Given the size of the problem, many have deemed the control of crop burning intractable or a ‘mammoth task’⁷.

How the state controls crop burning

The focus of our analysis is on how crop burning is controlled by the district administration which is headed by an officer recruited competitively through state or federal administrative services exams. We examined the incentives for the district administrator and their subordinates to control the problem of crop burning.

First, as the primary unit of state organization in South Asia, district bureaucrats are responsible for enforcing the laws of the land^29–31. Both India and Pakistan have outlawed crop burning, punishable by jail and fines. In India, burning crop residue has been a crime since the Indian Penal Code and the Air (Prevention and Control of Pollution) Act of 1981 (Section 188), but states are ultimately responsible for creating laws to prevent and ban this practice. For example, Punjab, Haryana and Uttar Pradesh were among the states that implemented laws to enforce the control of crop burning in 2015. The Pakistani Punjab adopted similar measures in 2018 (Section 144 of the Code of Criminal Procedure).

Second, pressure to act has also appeared through the courts. In M. C. Mehta versus Union India, the Supreme Court directed state governments to stop farmers from burning stubble, following which several initiatives were undertaken by the bureaucracy. In 2019, the Supreme Court of India further reinforced this stance by declaring the act of setting fires for agricultural residue disposal illegal.

Third, beyond implementing the writ of the law, government officials also face pressure from citizens’ organizations. Given the large externalities from crop burning, increased availability and use of data by citizens and news organizations, and the visibility of smog during the winter season, there is a higher public outcry from failures to act on crop burning by the government. However, citizen pressure can often be cyclical, building in particular during the winter months when conditions are particularly bad, and waning during the rest of the year^2,3,7.

Finally, district administrators also face routine scrutiny on their control over crop burning from senior officials, as was suggested to the authors during an interview with a Pakistan Administrative Service Officer on 20 March 2022. Newspaper accounts reflect this—for example, “Chief secretary [of Indian Punjab] Anurag Verma has asked district magistrates, commissioners of police and senior superintendents of police to jointly hold a review meeting on the issue [of crop burning], adding that action will be taken against the station house officer in-charge of an area if stubble burning takes place in their jurisdiction”³².

A lack of action can also lead to censure by senior administration: the State Chief Secretary issued a formal complaint against the Ludhiana district administrator for failing to stop crop burning in her district³³. More broadly, district administrators’ promotion and transfer decisions incorporate feedback from seniors in their confidential annual evaluations and, more recently, peers and members of the public through 360-degree evaluations³⁴.

Notwithstanding the increased concerns over the control of crop residue burning, there continues to be large variation in law enforcement and public initiatives against this phenomenon, indicating a role for bureaucratic discretion in enforcement. This can originate from several sources: first, beyond extraneous incentives to act on crop burning specifically, the district bureaucracy is also in charge of implementing a host of other schemes and programs, which has stretched their ability to deliver effective governance³⁵. They must therefore consider trade-offs when acting on crop burning. Second, district personnel includes individuals recruited locally, at the state level or federally, suggesting that their intrinsic incentives shaped by local embeddedness may also determine governance outcomes^29,36.

Policy tools to control crop burning

The district administrator supervises and coordinates work across several departments and ministries both inside the district machinery, as well as across the bureaucratic hierarchy^29–31,37. Policy levers to control crop burning can be divided into two categories: those that are exercised before and those that are exercised after crop burning takes place.

Before the crop burning season, the district administration organizes information campaigns through Agriculture Department extension agents and Revenue officials such as patwaris and lekhpals, to dissuade farmers from burning, and also, when feasible, to make subsidies and tools for renting machinery to clear crop stubble available^31,38,39. Reports document how these efforts are labour intensive and require visits by administrators to individual farmers and villages. Below, we provide examples of Agriculture Department officials profiled in the Washington Post: “Gurdial Kumar is one of the foot soldiers in the statewide effort. […] [H]e and his colleagues began visiting each of the 182 villages in their territory last month. […] [They] put up posters illustrating the impact of air pollution, including images of diseased lungs and small children wheezing. They talked about the virtues of the new subsidized machines, including one called the Happy Seeder”⁴⁰; “A senior bureaucrat [...], Pannu is leading the anti-burning efforts in Punjab. He has mostly carrots, not sticks, at his disposal. For the first time this year, India’s central government earmarked money to help farmers buy machinery that turns the straw into mulch. It has also contributed to a large-scale awareness campaign, complete with songs on social media, television advertisements and village-by-village meetings all urging farmers not to burn”⁴⁰.

Beyond specific actions before stubble burning, the district bureaucracy also actively monitors crop burning in coordination with State Pollution Control Boards, which increasingly use satellite data³¹. District administrators routinely use Agriculture and Revenue Department personnel to dissuade farmers from burning⁴¹.

The district bureaucracy is also empowered to carry out coercive and punitive action against polluting farmers by directing the police in their jurisdiction. For example, a Police Senior Superintendent in District Mansa stated that they “will ensure that orders on stubble burning, issued by the District Magistrate, are implemented”⁴². Once farmers are cited for burning, their case passes to the State Pollution Control Board, which decides whether to impose fines or jail term. For example, reports from 2020 note that the “PCB [Indian Punjab’s state pollution control board] [...] imposed fines of Rs 25.75 lakh in 961 cases of stubble-burning, and also made ‘red entries’ in farmers’ land records, which limit farmers ability to secure future loans and subsidies”⁴³.

However, as burning is a one-off event during the harvest season, it is unclear how bureaucratic punishment after the fact could matter in controlling stubble burning. In the following vignette, we highlight the possibility that these actions matter by deterring other farmers who have not yet burned their stubble: “Conversations with farmers indicated that those in the Haryana district shun stubble burning [...] [because of] strict vigilance by the district administration [...] “Farmers were afraid of penalties”, said Ajmer Singh, a farmer in Jaswanti village in Kaithal. “In our village, at least two farmers were fined for stubble fire. So, the fear of penalty also discouraged farmers from burning”. Dr Karam Chand, the deputy director of Kaithal district’s agriculture department, told Newslaundry the district administration had collected Rs 9.10 lakh as penalties from 402 farmers in the district”⁴⁴.

Bureaucratic control of fires

Our empirical analysis examines how bureaucrats manage crop burning. We hypothesize that changes in the expected wind direction impact bureaucratic incentives to control burning. Wind direction affects whether smoke from a specific crop burning incidence travels to a neighbour or pollutes the home district of a bureaucrat. This setup enables us to approximate where bureaucrats may decide to target their efforts to reduce fires.

We combine fire and wind satellite data with the administrative borders of districts. We focus on the period from 2012 to 2022, when high-resolution fire data started being available, and on the north of the Indian subcontinent—the area that is most affected by crop burning. Data are spatially aggregated into a gridded dataset of 5 km² observed in each month for a total of 17,979,600 observations, as illustrated in Fig. 2a.

Fig. 2. Crop burning decreases when wind pollutes the home district instead of neighbouring districts.

a, The 5 km² gridded dataset is shown in green and the district boundaries are shown in blue. The district Hardoi in Uttar Pradesh is enlarged as an example (inset). b, Coding of treatment definition and variation for the district Sitapur. Blue arrows represent the wind direction in a given month and year. Left, the area downwind from a potential fire is larger than the area upwind from the fire. We define this grid cell as treated. Centre, the area downwind from the fire is smaller than the area upwind. We define this grid cell as control. Right, a change in the wind direction converts the same grid cell as at left (middle diagram; orange carat) into the treatment group, because a potential crop burning would now become a source of pollution for the majority of the home district. c, Spatial distribution of this treatment and its temporal variation in two different months. The dark shading shows treated grid cells, and lighter cells are control grid cells. d, Crop burning decreases when wind changes pollute the home district instead of neighbouring districts. The treatment effects  are estimated with an event study design as described in equation (1) in the Methods. We estimate the effect of treatment on the number of fires in each period before and after a grid cell switches to treatment. The reference category is t = 0, when the switch takes place. The shaded area indicates the 95% confidence intervals. We restrict the analyses to a 6-month window around the switch to maximize support. The sample size is 12,883,548. Pooled pre-period coefficient = 0.0003, P = 0.929; pooled post-period coefficient = −0.0135, P < 0.000.

We study bureaucratic incentives by combining temporal variation in the expected wind direction with cross-sectional variation in where a location lies within a bureaucrat’s district. Our approach to defining treatment is illustrated in Fig. 2b: we examine the same grid during two different months—first, when wind is expected to blow smoke from a potential fire towards the grid’s home district; and, second, when wind would carry smoke towards a neighbouring district. In the former scenario, bureaucrats should be motivated to take action to suppress crop burning, whereas, in the latter scenario, intervention is less warranted (a representation of treatment in two months is shown in Fig. 2c; see the ‘Treatment definition’ section of the Methods for more details).

We use a difference-in-differences model to estimate the effect of a grid switching from being a source of exposure mostly for a neighbouring jurisdiction (control condition) to becoming a source of exposure for the home district of the bureaucrat (treatment condition).  Treatment is assigned on the basis of exogenous changes in the expected wind direction interacted with the shape of the district polygon. Because this treatment is probably exogenous, the parallel trends assumption is more credible than it is in a standard difference-in-differences setting. Further details are provided in the ‘Estimation’ section of the Methods.

We examined the results by plotting an event study (Fig. 2d), which captures the difference between treated and control units in each relative time from when a grid cell switches from the control to the treatment state. The flat pre-period line indicates that there is no difference in the number of fires between grid cells that will be treated and the control grid cells. Although establishing causality without experimental manipulation is difficult, these patterns are consistent with the plausibility of the parallel trends assumption. Instead, immediately after grid cells switch to treatment, we observe a 9% decrease in the number of fires (treatment effect = −0.0086, P = 0.024). This reduction reaches its lowest point, with a decrease of 22.24% (treatment effect = −0.0214, P < 0.000), two months after the treatment.

Overall, the difference-in-differences analysis indicates that the number of fires decreases by 10–13% (P < 0.000) after wind switches a location from polluting mostly a neighbouring district to a bureaucrat’s own district (Supplementary Information B.1). This corresponds to an annual reduction of about 54–72 fires per district. The effects are four times larger during the harvest months and in rice-producing areas (treatment effect = −0.046, P = 0.003), which is the primary crop associated with residue burning. These findings are robust to relaxing the difference-in-differences assumptions, various ways of specifying the estimation equation, as well as transformations of the dependent and the independent variables (Supplementary Information B.2). We also present evidence against the potential alternative interpretation that the results are explained by farmers’ pro-sociality towards others instead of bureaucratic effort (Supplementary Information B.3).

Action and inaction in pollution control

Having established that fires decrease when pollution affects the home district, we now examine how crop-burning patterns respond to the strength of incentives faced by bureaucrats. Areas lying closer to a jurisdiction border should be the ones where incentives for action and inaction are the strongest. We illustrate this logic in Fig. 3a: when wind blows from the north, grid cells closer to the upwind border (closest quintile in darkest blue) are those in which bureaucrats have the strongest incentives to curb crop burning (action). This is because these areas have the maximum pollution impact on their home district. When the wind direction shifts and it blows from the south, the same locations (now in darkest red) pollute neighbouring districts. Here bureaucrats now have the largest incentive to allow crop burning (inaction).

Fig. 3. Bureaucrat incentives and action reduce crop burning.

a, Illustration of how changes in the wind direction combined with the distance from the border strengthen or reduce the incentives for bureaucrats to curb crop burning. Left, when the wind blows from the north, bureaucrats have the strongest incentives to curb crop burning in grid cells closer to the upwind border (in darker shades of blue), as these areas have larger pollution impact on their own district. Right, when the wind blows instead from the south, the same locations now pollute neighbouring districts, reversing the incentives; bureaucrats now have the largest incentive to allow crop burning in grid cells closer to the downwind border (in darker shades of red). b, The effect of being close to the upwind (blue) and downwind (red) border on crop burning. Coefficients from equation (3) are plotted, estimated on a grid cell by border pair monthly database. The blue line shows that fires decrease in locations closer to the upwind border, where crop burning would pollute the home district. The red line plots the differential increase in fires approaching the downwind border (relative to the upwind border effect), where smoke would pollute the neighbouring district. All effects and 95% confidence intervals are plotted by quintiles of distance to the border, where the fifth quintile (q5), the farthest, is the omitted category. The sample size is 115,149,943. The effect size for each quintile is shown in column 1 of Supplementary Table C.1. c, The impact of penalties on farmer deterrence. The triple difference result and 95% confidence intervals are plotted for how a bureaucratic penalty reduces the number of future fires in district areas that are a source of pollution to the home district (versus neighbouring districts) using the event study specification in equation (2). The omitted period is the month of the penalty and the month before it. The sample size is 7,934,760. Pooled pre-period coefficient = −0.0111, P = 0.297; pooled post-period coefficient = −0.0284, P = 0.008.

We test for these effects systematically in Fig. 3b (equation (3)). First, as one considers areas closer to the district border relative to the areas farthest away, two opposite effects occur: crop burning decreases when the border is upwind (coefficients in blue), exposing the home district to pollution, and it increases when the same border becomes downwind (coefficients in red), exposing neighbouring jurisdictions to pollution. These effects are large: fires decrease by 14.47% when the border is upwind (treatment effect = −0.0113, P < 0.000) and increase by 15.11% when it is downwind (differential effect versus upwind effect = 0.0231, P < 0.000) with respect to the mean of the dependent variable in the quintile farthest from the border (Supplementary Table C.1).

Second, both effects intensify monotonically with decreasing distance to the border, which we interpret as strengthening bureaucratic incentives towards action and inaction. Third, we show that action and inaction in regulating pollution are five times larger at the India–Pakistan border, with a reduction in fires of 56.22% in the areas closest to the upwind border (P < 0.000) and an increase of 146.37% in areas closest to downwind border (P < 0.000). These are areas where the ability to coordinate with neighbouring bureaucrats is absent (Supplementary Table C.1). Moreover, these effects also occur more generally at international borders. Finally, effects are also larger in the presence of capacity constraints (operationalized as a larger share of the district under rice cultivation), where there is an additional change in fires of 12% in the quintile closest to the upwind border (P = 0.003), and are conditional on bureaucrats’ legibility over crop burning (operationalized through proximity to main roads) (Supplementary Figs. C.3 and C.4).

Bureaucrat action and farmer deterrence

The ‘Air pollution management in South Asia’ section above discussed qualitatively how bureaucratic action against crop burning through fines can create a deterrence against future burning by farmers. We now test for the possibility of this more systematically. We collect data at the district-month level on the most severe punishment that farmers can receive for burning their crops: cases involving a criminal procedure for the violation of air pollution regulations (Methods).

We conducted two analyses. First, we confirmed that bureaucrats’ actions are concentrated in periods and locations where the problem is most visible: penalties are 61.63% (P = 0.000) higher during the winter harvest season, which is when air pollution is most visible due to atmospheric conditions. By contrast, there is no corresponding increase in bureaucratic punishments in the summer harvest months when air pollution is less visible (Supplementary Table D.1).

Second, we study whether bureaucratic actions create a deterrence against future crop burning among farmers by examining the differential impacts of punishment in areas of the district where bureaucrats pay more attention—that is, the areas where crop burning pollutes their home district—versus other areas that mostly pollute neighbouring districts. Figure 3c shows that there are parallel trends in the number of fires in these two areas before bureaucratic punishment. By contrast, we see a decrease in crop burning by farmers after another farmer is punished in their district versus not. This decrease, concentrated in the first 3 months since the punishment, corresponds to a 9 to 13% additional decrease in fires over and above the effect of belonging to an area that is a source of pollution exposure for the home district. This finding is important because it indicates that the problem of controlling pollution by dispersed farmers can be ameliorated through the channel of deterrence: even one-off actions by bureaucrats can create positive spillovers in reducing crop burning by other farmers.

The impacts of crop burning on mortality

We next examined how meaningful the patterns in bureaucratic behaviour that we identified are for public health. We focused on a particularly important dimension of population welfare: infant and child mortality. Despite progress, child mortality in India and Pakistan remains far from the Sustainable Development Goals target of under 25 deaths per 1,000 live births, with some estimates suggesting that 8.8% of child mortality is due to air-pollution-related causes⁴⁵. We examined the impacts of crop-burning-induced air pollution shocks experienced in utero on infant and child mortality.

We used three sources of data: Demographic and Health Surveys (DHS) geolocalized data on births; remote-sensed air-quality data (Copernicus Atmosphere Monitoring Service (CAMS)); and data on fires. Our approach builds on the finding that in utero shocks can impact life outcomes for children⁴⁶. We first estimated the total amount of local air pollution that a child is exposed to in utero using CAMS data. As air quality is affected by multiple sources, we also isolated the impact of crop burning by calculating, for each fire in our data, the path that pollution particles take across space using a meteorological model (Hybrid Single-Particle Lagrangian Integrated Trajectory (HYSPLIT); https://www.ready.noaa.gov/HYSPLIT.php) and aggregating the total exposure in utero. These two variables give us the total exposure to air pollution, as well as an estimate of the amount of pollution from crop burning experienced over the 9 months of the pregnancy.

We used two approaches to estimate the effect of crop burning on mortality. In our first approach, we tested the direct impact of upwind crop burning pollution particles estimated through the HYSPLIT model on infant and child mortality (our approach is shown in Fig. 4a). Examining these reduced-form impacts, we found that a 1-log increase in pollution particles from crop burning causes 0.17 additional deaths among infants in their first year (P = 0.008) and 0.21 (P = 0.001) in the first five years from birth (Supplementary Table E.1 (second panel)).

Fig. 4. PM_2.5 particles from crop burning increase infant and child mortality.

a, The nonlinear dispersion of particles from 8 fires in our data over a 24 h period as estimated by HYSPLIT. We use all such particles for all fires that pass a DHS cluster (grey circles in the inset) as a measure of in utero exposure to crop burning pollution. District lines are marked in black for the sample from which we consider fires and in grey for the sample that receives fires. b, The average infant and child mortality (grey bars) and the impact of a 1-log increase in in utero exposure to PM_2.5 from fires on infant and child mortality, from an instrumental variables model, as described in equation (5) (blue bars). Dark blue lines represent the 95% confidence interval. The independent variable is the natural logarithm of exposure to air pollution for each child in utero, as predicted by the natural logarithm of the total number of particles from all fires affecting the pregnancy calculated using HYSPLIT. The dependent variable is the rate of infant (above) and child (below) mortality in 1,000 births. All regressions include district, birth month and year, and birth order fixed effects, and control for the number of fires in a 10 km radius from the DHS cluster. Standard error values are clustered at the district and birth date level. The effect of crop burning on infant mortality has a P value of 0.012 (n = 542,150), while child mortality has a P value of 0.005 (n = 542,150).

However, measuring how many particles can affect a child in utero might not return an accurate picture of the amount of pollution, as exposure depends also on contextual factors at the location of birth, such as geography or humidity. In our second analysis, we therefore focused on the specific impact of air pollution from crop burning using an instrumental variables approach. We instrument the local PM_2.5 concentration experienced during the 9 months of the pregnancy with pollution particles sent from upwind fires. This strategy relies on the idea that children will experience different levels of in utero pollution depending on the extent to which wind carries crop burning smoke into their area, as well as the period of the year when their mother is pregnant. This setup therefore enables us to focus on a single source of pollution, which is necessary to understand the extent to which government action could reduce mortality burdens.

We show that, across Pakistan and India, exposure to a 1-log increase in in utero exposure to PM_2.5 concentration caused by crop-burning pollution leads to an additional 24 (P = 0.016) to 26 (P = 0.010) infant mortalities and 30 (P = 0.005) to 36 (P = 0.004) child mortalities per 1,000 births (Fig. 4b and Supplementary Table E.1 (first panel)). Crop burning increases the relative risk of infant mortality by 64.3% to 69.8%, and increases the relative risk of child mortality by 74.2% to 86.2%. This instrumental variables approach not only accounts for confounders such as variables that increase pollution but improve economic activity (for example, industrial production and transportation networks) but also reduces measurement error in PM_2.5 (ref. ¹¹). Notice that these estimates measure the local average treatment effect of PM_2.5 on child health for those individuals for whom changes in fire exposure cause changes in local PM_2.5 exposure. With respect to the average mortality in our sample, our estimates suggest that a 10 μg m⁻³ increase in average daily exposure to PM_2.5 raises infant mortality by 7.8% and child mortality by 8.5%, effects that are consistent with previous empirical findings on air-pollution-related mortality^21,47,48.

These estimates enable us to do a back-of-the-envelope calculation to retrieve how much the strategic behaviour of bureaucrats contributes to (or could reduce) health deterioration in this region. Our estimates suggest that 1.8 to 2.7 deaths in 1,000 (or 4.4 to 6.6% of the average child mortality) could be preventable if bureaucrats were able to reduce crop burning such that all areas in their jurisdiction behaved the way as those that pollute their own jurisdictions.

Discussion

States have an increasingly important role in fighting climate change and environmental degradation. Public opinion and media attention are often focused on international and national laws that govern the environment. However, on a daily basis, the apparatus of the state is in charge of enforcing laws and regulations of which the successful enforcement has substantial consequences for environmental management and the health outcomes of citizens.

Here we focus on one such case—air pollution from crop burning in South Asia. We present a systematic large-scale analysis of the political economy determinants and consequences of crop burning in India and Pakistan. Though legislation outlaws crop burning, uneven enforcement continues to make this phenomenon one of the largest health emergencies on the planet. In contrast to other types of pollution, air pollution from crop burning is generated by millions of dispersed farmers, each contributing minimally to the problem, such that punishing any of them might have close to no impact on the total problem.

Our findings suggest that controlling air pollution from crop burning may not be completely intractable as is commonly assumed. We show that burning patterns respond to the incentives bureaucrats face in their own jurisdictions. We demonstrate that, even under the weak incentives currently in place, bureaucrats can bring meaningful reduction to crop burning, especially when they internalize the externalities of polluting behaviour. Importantly, our results show that bureaucratic punishment of farmers can create deterrence among future polluters, suggesting that bureaucratic action against crop burning may not need to entail punishing thousands of farmers, which is likely to be politically infeasible. Overall, these findings indicate that the government infrastructure that already exists is capable of exerting better control over crop burning despite not being considered to be a lever of change.

Future research can further unpack the dynamics that we highlight. First, our results suggest that bureaucrats can calibrate their actions at high frequency in the face of changing incentives, suggesting promising avenues for further work relating to how bureaucrats manage the environment and climate change. Second, scholars should examine interactions inside and across the district bureaucracy to understand which actors and actions are critical for success. They should study the trade-offs involved in focusing on crop burning versus other tasks, and how various departments—such as agriculture, environment and police—work together to exert control. Third, an important next step is to study the incentives of farmers. This is essential for the success of any solution at scale. While we showcase the importance of deterrence, we do not study how best to apply this lever—how many farmers to target, in which period, for which penalty—nor which alternatives to offer farmers who agree to avoid burning^24,27 or how these incentives interact with other major policies, such as agricultural support prices. Scholars should focus on how to use a combination of carrots²⁸ and sticks in a sustainable, scalable and cost-effective way.

Methods

Data

For our empirical analysis, we use data from a variety of sources.

Fires: we measure crop burning with data on fires from NASA satellites Moderate Resolution Imaging Spectroradiometer (MODIS) (1 km × 1 km resolution) and Visible Infrared Imaging Radiometer Suite (VIIRS) (375 m × 375 m resolution) from September 2012, the first year in which VIIRS starts operating. These data are generated by detecting thermal signatures that are not impeded in the presence of cloud cover or smoke in the atmosphere. We plot the total number of fires over time in Supplementary Fig. A.1: each year, there are two peaks in the months of April through May and October through November, in correspondence with the crop burning seasons of wheat and rice, respectively⁴⁹.

Crops: we gather data on average crop area, production and yield in 2010—the latest year in which micro-level crop data are available before the start of our analysis sample (2012–2022). These data are retrieved from Food and Agriculture Organization of United Nations (FAO) maps (https://mapspam.info/index.php/methodology/). They have a resolution of 5 × 5 arcmin, corresponding to approximately the size of a grid cell in our data.

Winds: the data on wind direction and speed are from the European Centre for Medium-Range Weather Forecasts (ECMWF; https://www.ecmwf.int/en/forecasts/datasets) and are precise at a 25 × 25 km resolution. We proxy for the expected wind direction of bureaucrats in a given month–year by computing ten-year rolling averages of wind direction in each month and year by using wind data starting in the 2000s. We plot actual and rolling averages of wind directions in Supplementary Fig. A.1 (right). We can see that rolling averages closely follow the actual wind direction. The two major wind switches happen on average in June and August each year.

Administrative borders: we use the latest 2017 census boundaries for Pakistan, which are available on the Humanitarian Data Exchange website⁵⁰. For the Indian administrative boundaries, we rely on the latest version generated by Community Created Maps of India⁵¹. These maps compile information from various government websites to create district-level polygons. During our analysis period, some districts underwent splits: Charki Dadri was carved out from Bhiwani in Haryana on 1 December 2016, and Malerkotla district was bifurcated from Sangrur in Punjab on 14 May 2021. We account for these splits by modifying our shapefiles accordingly. Our main sample of analyses comprises states in which the problem of crop stubble burning is particularly salient. These states are Punjab in Pakistan, and Punjab, Haryana, Uttar Pradesh, Delhi and Bihar in India. These are the same states considered in other papers focusing on air pollution in South Asia⁵.

Combining all the data at a 5 km² grid-cell level: we combine the above datasets into a single gridded dataset of 5 km². Wind and fire grids do not overlap precisely. To reduce errors in the assignment of fires to wind grids, we split the 25 × 25 km wind grid cells into squares of 5 km² (around 2.25 km × 2.25 km size). This guarantees that, even if wind and fire grids do not perfectly overlap, data on fires are assigned to the wind grid cell that contains most of the fire grid cell. Splitting wind grids in smaller parts also enables us to address a second measurement concern—some grid cells might overlap across jurisdiction boundaries. By reducing their size, we minimize the amount of misattribution and we are able to assign each grid cell to the appropriate jurisdiction. Notice that any remaining error should bias our findings towards zero, as errors are uncorrelated with the dependent and independent variables. In total, 11% of the grid cells fall between more than one district. These are assigned to the district that contains a larger overlapping area with them (and we show that our results are robust to dropping these from the analysis). For crop data, we assign the pre-period crop production to each 5 km² grid cell. This enables us to classify what crops are produced in each grid-cell location. Figure 2a illustrates our final gridded dataset for Hardoi District in Uttar Pradesh, where each square refers to a 5 km² grid cell. Within the sample area, we are able to observe 149,830 grid cells in each month and year from September 2012 to August 2022, for a total of 17,979,600 observations. Supplementary Table A.1 presents descriptive statistics on these data.

Air pollution: we use data on air pollution from the Copernicus Atmosphere Monitoring Service (CAMS; https://ads.atmosphere.copernicus.eu/cdsapp#!/dataset/cams-global-reanalysis-eac4?tab=overview), measuring the concentration in the air of particulate matter with a diameter of 2.5 μm or less (PM_2.5). The data are available at a resolution of approximately 80 km. In Supplementary Fig. A.2, we map the total concentration of PM_2.5 in the air in 2014. The north of the subcontinent is covered by a dense cloud of air pollution (darker red), unlike any other parts of the globe.

PM_2.5 data from CAMS are valuable as they offer high-resolution information on air quality in all the regions of our sample. However, these data are estimates obtained by combining satellite observations, ground-based measurements and atmospheric models. To validate these estimates, we gather data that are purely ground based: measurement of air pollution from air monitors⁵². The database contains data from 2,110 monitors operated by the Central Pollution Control Board (CPCB) in the period 2015–2019. These data cover 76 CAMS grids out of the 2,156 in our entire study area, and not in all months, making it too sparse for use as a regressor. We instead use air monitors to validate CAMS estimates in Supplementary Fig. E.2, which shows a positive correlation (0.726) between the total pollution recorded each month by the monitoring stations and CAMS estimates.

Bureaucratic punishment: for India, penalty data come from the Socioeconomic High-resolution Rural-Urban Geographic Platform for India (SHRUG) database (https://www.devdatalab.org/judicial-data), which contains all judicial case records from the Indian eCourts platform. From this database, we extracted all judicial cases in which there was a violation of the Air (Prevention and Control of Pollution) Act of 1981, which includes crop burning offences from 2010–2018, the years during which the data overlap with our sample. Measurement error in these data should bias our results downwards. In the case of Pakistan, we obtained new data on police records of criminal activity from the Central Police Office in Lahore, called First Information Reports (FIRs), that relate specifically to crop burning offences for the years 2017–2022.

Bureaucratic control of fires

Treatment definition

As discussed above, the treatment turns on when smoke from a potential fire in a grid cell would pollute the home district of the administrator. For each month–year, we proxy for administrators’ expectations on actual wind directions by using 10 month–year rolling average of wind direction. In practice, these beliefs are formed through long-term local knowledge, as well as through weather data that are made available as part of bureaucrats’ routine work⁵³.

Once we have this measure of wind direction, we use it to calculate the area of the jurisdiction that is downwind, as well as upwind, from each grid cell in each month. We do this by sketching a 180° line that passes through the grid-cell centroid and is perpendicular to the wind direction vector such that the district is split into a downwind area and an upwind area⁵⁴. Using this definition, a grid cell is treated in a month–year if the area of the jurisdiction downwind from the grid cell is larger than the area upwind. Figure 2b provides a visual illustration of this treatment definition. Figure 2c illustrates the actual representation of treatment status variation in different months.

Estimation

$$\mathrm{Fire}{\mathrm{s}}_{it}=\sum _{r=-s}^z\zeta D_{ir}+{\alpha }_i+{\gamma }_j\times {\theta }_t+\delta {\mathbf{X}}_{it}+{\nu }_{it}$$ 1

where our dependent variable Fires_it is the number of fires in a grid cell and month. The coefficient of interest ζ is the effect of treatment in each period before and after treatment, relative to the switching month–year (the omitted category). For example, if changes in wind direction convert grid-cell i from control to treatment in May, relative time will take value 0 in May, +1 in June, −1 in April, and so on. We restrict attention to observations in a six-month window around treatment, where the vast majority of our sample lies (Supplementary Fig. A.3) and drop treated observations in the first period in the sample, September 2012, as their relative time cannot be established.

The high dimensionality of our data allows for the inclusion of fine-grained fixed effects, although as we show in the Supplementary Information, the treatment effect size is quite stable when these are included. First, α_i are 5 km² grid-cell fixed effects that absorb time-invariant factors for over 149,000 grid cells in our data, such as crop suitability or economic characteristics of small locations. Second, γ_j × θ_t are jurisdiction by month–year fixed effects that account for time-varying shocks at the district level, such as changes in administrative personnel. X is a vector of controls that includes average wind speed and direction. Standard errors are clustered at the grid cell and month–year by district levels to account for temporal dependence of observations as well as spatial correlation in wind.

In addition to the event study design, we also run a difference in differences strategy. For a 5 km² grid-cell i in jurisdiction (district) j and month–year t, we estimate:

$${\mathrm{Fires}}_{it}=\beta {\mathrm{Downwind}}_{it}+{\alpha }_i+{\gamma }_j\times {\theta }_t+\delta {\mathbf{X}}_{it}+{\mathit{ϵ}}_{it}$$ 2

The coefficient of interest β captures the overall effect of treatment on the number of fires in that grid cell. Downwind represents the treatment variable: a grid switches to treatment when, depending on wind direction, it would expose most of the home district to smoke. Fixed effects, controls and clustering are as in the event study specification described above.

Action and inaction in the control of fires

We build on the above specification to separate the effect of action and inaction towards the control of fires. For each grid cell, we calculate the distance from its respective borders and code whether these borders are upwind or downwind in a given month. This yields a grid-border–grid-cell pair database with spatial variation in the distance of each grid cell to the border and monthly variation in whether pollution exits or enters the border corresponding to each grid cell.

The key advantage of this strategy is that we can separately estimate the effort bureaucrats exert to prevent fires in areas close to upwind borders from the effort not exerted to prevent fires in areas close to downwind borders. Calling a border segment b, we estimate, for each segment b corresponding to grid-cell i in month–year t:

$$\begin{array}{l}{\mathrm{F}\mathrm{i}\mathrm{r}\mathrm{e}\mathrm{s}}_{ibt}=\sum _{q=1}^5\kappa {\mathrm{D}\mathrm{i}\mathrm{s}\mathrm{t}\mathrm{a}\mathrm{n}\mathrm{c}\mathrm{e}}_{it}+\sum _{q=1}^5\lambda \mathrm{D}\mathrm{i}\mathrm{s}\mathrm{t}\mathrm{a}\mathrm{n}\mathrm{c}\mathrm{e}\hfill \\ \times {\mathrm{D}\mathrm{o}\mathrm{w}\mathrm{n}\mathrm{w}\mathrm{i}\mathrm{n}\mathrm{d}\mathrm{N}\mathrm{e}\mathrm{i}\mathrm{g}\mathrm{h}\mathrm{b}\mathrm{o}\mathrm{u}\mathrm{r}}_{ibt}+{\mu }_i+{\eta }_b\times {\theta }_t+{\nu }_{ibt}\hfill \end{array}$$ 3

where q are quintiles of the distance of each grid cell to the bordering district segment b, and we control for grid cell and for border times month–year fixed effects. We use the fifth quintile (the farthest) as the absorbed category. We are interested in two coefficients: κ captures the effect of proximity to the upwind border, an area where bureaucrats have the largest incentives to curb fires, as those would affect their own district the most. Instead, λ captures the differential effect of what happens as we approach the downwind border, which might be the same location as an upwind area in other periods. Our expectation is that bureaucrats have little incentive to act on fires when those are located close to downwind borders, as smoke would exit their district. The logic of this estimation is further illustrated in Fig. 3a, where κ coefficients are represented in blue and λ in red.

Bureaucratic action and farmer deterrence

We code a new district-level variable called PostPunishment, taking value one if a farmer was punished in a district in the previous 12 months. To stay consistent with our calendar month analysis, we start the 12-month duration from the filing month itself if the penalty was filed before the 15th day of the month. Otherwise, it starts the following month. We use this variable in a triple difference-in-differences event study design by interacting it with the down > up dummy referenced as D_ir in equation (1). We show the event study results in Fig. 3c. We also plot these results in 1-month windows and find consistent, but noisier, results (Supplementary Fig. D.1). Finally, we show that these results are robust using a simple triple difference-in-differences setup in Supplementary Table D.2.

Mortality impacts of crop burning

Linking fires to downstream recipients of pollution

We measure infant and child mortality using geolocalized individual-level health information from surveys collected by the Demographic and Health Surveys (DHS) program (https://dhsprogram.com/data/) in India (2015–2016 and 2019–2021) and Pakistan (2016–2017). The DHS data enable us to match each pregnancy to a specific small location, called a DHS cluster, across India and Pakistan. In total, we have 542,150 pregnancies across 58,103 DHS clusters for which we observed child, mother and family characteristics.

For each birth, we calculate the pollution exposure from crop burning fires as the total number of particles coming from fires started in our analysis sample (Punjab in India and Pakistan, Haryana, Delhi, Uttar Pradesh and Bihar) to the home DHS cluster location of the birth, for the expected length of the pregnancy. This process enables us to link pollution particles from each upwind fire to pregnancies, thereby generating a precise measure of the expected air pollution from crop burning in our sample for each pregnant mother. Besides the number of particles received in the home DHS cluster location of the pregnancy, we also calculate the total PM_2.5 pollution experienced in that location using satellite data from the CAMS.

We proceed to collapse the data at the pregnancy level, calculating the total local air pollution, as well as total particles from crop burning fires received by each child in utero. Children born in the same location and from a similar family will experience different levels of in utero pollution from PM_2.5 depending on the timing of the pregnancy.

Estimation strategy

Our quantity of interest is the impact of air pollution (PM_2.5 concentration) caused by crop burning on child and infant mortality. This quantity is generally confounded by characteristics of the location, period of birth and of the family, which might be correlated with both air pollution and worse health outcomes. We can address these concerns in our setting using time variation in exposure to pollution. We consider two models. First, we use a reduced form regression of the impact of total crop burning fire particles on mortality outcomes. Second, we capture the impact that fires have on overall levels of air pollution using an instrumental variable strategy. For this, we first estimate the impact of total fire particles on PM_2.5 concentration experienced by each child in utero (first stage). We then estimate the impact of the predicted air pollution from fires on children’s mortality (second stage). For each child (c) born in district d during month–year t, we estimate:

$$\left(\text{first stage}\right)\mathrm{PM}{2.5}_{ct}=\pi {\mathrm{FireParticles}}_{ct}+{\xi }_d+{\tau }_t+{\varphi }_b+\sigma X_{t,dhs}+{\upsilon }_{ct}$$ 4

$$\left(\text{second stage}\right){\mathrm{M}\mathrm{o}\mathrm{r}\mathrm{t}\mathrm{a}\mathrm{l}\mathrm{i}\mathrm{t}\mathrm{y}}_{ct}=\rho \widehat{\mathrm{P}\mathrm{M}{2.5}_{ct}}+{\xi }_d+{\tau }_t+{\varphi }_b+\sigma X_{t,dhs}+{\omega }_{ct}$$ 5

Both the first and second stage control for district ξ, month–year from birth τ, and birth order ϕ fixed effects. All regressions also control for the total number of fires during the pregnancy within a range of 10 km from the DHS cell, X—while most of the particles affecting a pregnancy will be carried from different and farther locations, fires originating around the DHS cluster are likely to have a larger impact on exposure, and locations with more fires might be different on average than other places. Our coefficient of interest, ρ, identifies the effect of exposure to fire-related air pollution in utero on child and infant mortality. We provide evidence consistent with the validity of the IV identifying assumptions in Supplementary Table E.1 (c), which shows a strong first stage of fire particles on PM_2.5 concentration, and Supplementary Fig. E.3, which shows that their relationship is monotonic. The extent to which a pregnancy is affected by pollution particles from a specific fire is likely as if random, because this is determined by a host of upwind meteorological factors interacted with the timing of the pregnancy. Finally, the case for exclusion restriction in our setting is strong because it is unlikely that particles from fires have an impact on mortality from channels other than air pollution.

Linking bureaucratic incentives with mortality outcomes

In the main section, we have shown that (1) bureaucratic strategic behaviour impacts the incidence of fires, and that (2) fires have a substantial impact on child and infant mortality. We then connect these two findings to assess how much of the increase in mortality determined by fires can be attributed to jurisdictional incentives. We do a back-of-the-envelope calculation to understand what would happen if bureaucrats had the same incentives and capacity across the entire district area, and exerted control in the downwind parts of their districts the same way they do in the upwind parts. Here we describe the details of our calculation.

We have shown in Supplementary Table B.1 that treatment reduces fires by 10–13% on average in the upwind part of the district, or by approximately 5–7.5% in the full district. We have also estimated that a 1-log increase in particles from fires over the 9 months of the pregnancy raises child mortality by up to 36 deaths in 1,000 births (Supplementary Table E.1). Now, consider what would happen if exposure to fires in utero dropped by the extent that bureaucrats can affect fires in the district. If bureaucrats, assuming proper incentives and capacity, reduced crop burning in the downwind parts of their district to match the upwind parts, child mortality would reduce by between 1.8 to 2.7 fewer deaths in 1,000, which corresponds to a 4.4 to 6.6% reduction over the average child mortality in the sample. Of course, a problem with bureaucratic action is the lack of capacity, but this scenario illustrates one way in which effects could be obtained by creating slack in bureaucratic capacity through the prioritization of environmental management.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Online content

Any methods, additional references, Nature Portfolio reporting summaries, source data, extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at 10.1038/s41586-024-08046-z.

Author contributions

Both authors contributed equally to the manuscript.

Peer review

Peer review information

Nature thanks Marshall Burke, Haidong Kan and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. Peer reviewer reports are available.

Data availability

The replication material used for the analysis are available in the Code Ocean repository (10.24433/CO.6383087.v1). MODIS and VIIRS data on fires were obtained from the Fire Information for Resource Management System (https://firms.modaps.eosdis.nasa.gov/download/). Crop production data are available from the FAO (https://mapspam.info/index.php/methodology/). The ECMWF dataset on wind direction and wind speed is publicly available (10.24381/cds.adbb2d47). For administrative borders, maps were taken from the latest 2017 census boundaries for Pakistan, available on the Humanitarian Data Exchange website (https://data.humdata.org/dataset/cod-ab-pak). For India, district maps are provided by DataMeet Community Maps Project (http://projects.datameet.org/maps/), made available under the Creative Commons Attribution 2.5 India (http://creativecommons.org/licenses/by/2.5/in/). The dataset on air pollution is the ECMWF Atmospheric Composition Reanalysis 4, and it is available from the CAMS (https://ads.atmosphere.copernicus.eu). Data on bureaucratic punishment are from the SHRUG database for India (https://devdatalab.org/shrug), while, for Pakistan, we obtained data on criminal activity from the Central Police Office in Lahore. Further information on input data and their sources is provided in the ‘Data’ section of the Methods.

Code availability

Replication files for the analysis can be found in the Code Ocean Repository (10.24433/CO.6383087.v1). Stata MP (v.18), Stata SE (v.17), Rstudio (v.4.3.1), Python (Anaconda 3 2021.11) and QGIS (v.3.26) were used for the analyses.

Competing interests

The authors declare no competing interests.

Supplementary information

The online version contains supplementary material available at 10.1038/s41586-024-08046-z.