Policy Research Working Paper 10376 Man or Machine? Environmental Consequences of Wage Driven Mechanization in Indian Agriculture A. Patrick Behrer Development Economics Development Research Group March 2023 Policy Research Working Paper 10376 Abstract This paper uses an exogenous shock to wages from the districts that appear more likely to mechanize the harvest. world’s largest anti-poverty program to show that higher MNREGA did not lead to changes in area planted or ton- wages can lead to increased air pollution, likely by inducing nage produced in fire intensive crops. The estimates show farmers to shift into a labor-saving and mechanized produc- that nationally, the shock increased the rate of particulate tion process. Using a difference-in-differences approach on emissions from biomass burning by 30 to 50 percent. The the staggered roll-out of India’s Mahatma Gandhi National results suggest that absent policies to correct for envi- Rural Employment Guarantee Act (MNREGA), combined ronmental externalities of mechanization at all stages of with data on nearly 1 million fires, the paper shows that development, labor market shocks may lead to inefficient the frequency of agricultural fires increases by 21 percent levels of mechanization. after the shock. The increase in fires is concentrated in This paper is a product of the Development Research Group, Development Economics. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/prwp. The author may be contacted at abehrer@worldbank.org. The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. Produced by the Research Support Team Man or Machine? Environmental consequences of wage driven mechanization in Indian agriculture A. Patrick Behrer JEL Codes: Q1, Q53, D2, O13, O44, J43 Keywords: Air pollution, Mahatma Gandhi National Rural Employment Guarantee Act, poverty reduction, structural change, agriculture, India ∗ I would like to thank Edward Glaeser, Rema Hanna, Gabriel Kreindler, Robert Stavins, Gernot Wagner, Joseph Aldy, Jisung Park, Teevrat Garg, Michael-David Mangini, Kristen McCormack, Kibrom Tafere, Gabriel Englander, Valentin Bolotnyy and numerous seminar participants at Harvard, Stanford, The University of Maryland, The World Bank, and the University of Chicago for comments and suggestions. Simon Schröder gave excellent research assistance and both Rakesh Kumar and Ryan Lee provided invaluable technical assistance collecting data. All remaining errors are my own. A previous version of this paper was circulated as “Earth, Wind and Fire: The impact of anti-poverty efforts on Indian agriculture and air pollution.” Funding from the Harvard Environmental Economics Program, the Environmental Protection Agency and the Harvard Climate Change Solutions Fund is gratefully acknowledged. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the World Bank and its affiliated organizations, or those of the Executive Directors of the World Bank or the governments they represent. 1 Introduction Shifting labor from low-productivity, labor-intensive sectors to high-productivity, often capital in- tensive, sectors is an essential component of economic development (Huneeus and Rogerson, 2020; Kuznets, 1973; Arrow et al., 1995). This occurs as workers transition out of agriculture into indus- try (Gollin et al., 2002) and again as workers transition out of industry into services or higher value added industry. In both cases, higher wages in the labor-intensive sector, at least in part, help to drive the transition by encouraging firms to invest in labor-saving mechanization. But, unlike their impacts on structural transformation, the effects of higher wages on environmental damage and pollution are theoretically ambiguous (Jayachandran, 2021) and depend on which type of produc- tion has larger environmental externalities. In settings in which mechanization relies on fossil-fuel generated power – electrical or otherwise – this mechanization will be associated with a pollution externality.1 If firms that rely on fossil-fuel generated power do not internalize those externalities, their investments in mechanization may be inefficiently high. There is a large literature on the potentially adverse impacts of the adoption of “inappropriate technology” on development, espe- cially when the technology is labor-saving in a labor-rich setting (Stewart, 1978; Acemoglu and Zilibotti, 2001; Acemoglu, 2021). This adverse effect, in the context of mechanization of manufac- turing, has even been implicated in premature deindustrialization in some contexts (Huneeus and Rogerson, 2020). The hypothesized relationship between wages and pollution here suggests that proper Pigouvian taxation of environmental externalities might lead to a more efficient transition from labor to capital intensive production and a healthier process of economic development. I show empirically that an anti-poverty policy that exogenously raised wages had the unintended consequence of also increasing pollution. This increase in pollution appears to be driven by mech- anization in response to the wage increases. A central challenge in demonstrating that increasing wages has a causal impact on pollution by increasing mechanization has been the simultaneity of wage growth and mechanization (Ebenstein et al., 2015). Rising wages could drive increases in 1 The transition from human powered cotton production to steam powered production in England is a classic example: as firms replaced human labor with coal-fired powered cotton looms, production increased and shifted to Manchester which suffered notable declines in air quality (Rodgers, 1960; Longhurst and Conlan, 1970). Longhurst and Conlan (1970) quotes the Manchester police commission in the 1800s: “the increase of steam engines as well as smoak issuing from chimnies used over stoves, foundries, dressers, dyehouses and bakehouses are become a great nuisance to the town.” 2 mechanization by raising labor wages and leading firms to invest in labor-saving mechanization.2 However, exogenous innovation in production techniques also leads to labor-saving mechanization that raises labor productivity, with wage increases coming as a consequence of higher productivity (Solow, 1957).3 I address the problem of simultaneity by measuring how air pollution changes after an exogenous shock to wages generated by India’s Mahatma Gandhi National Rural Employment Guarantee Act (MNREGA), the world’s largest anti-poverty program. After the implementation of MNREGA, low-skill wages increased by between 4% and 8% (Imbert and Papp, 2015) in rural areas of India. I show, using a difference-in-differences framework taking advantage of the staggered roll-out of MNREGA coupled with data on nearly 1 million fires, that agricultural fires increased by between 9% and 21% after implementation of MNREGA. To explain the mechanism, I outline a simple model showing that rising incomes could lead firms to invest in labor-saving, but polluting, mechanization. My model suggests that farmers responded to MNREGA by mechanizing harvest. This is consistent with both Hornbeck and Naidu (2014) and Clemens et al. (2018), who show that farmers mechanize after shocks to the low-skill labor market in the United States. Mechanization leads to an increase in agricultural fires in India because it leaves between 80% and 120% more biomass on a field relative to manual harvesting (Yang et al., 2008; Jitendra et al., 2017). Biomass must be removed prior to planting and fires are the least expensive way to remove biomass (Ministry of Finance, 2018). I show that fires increase by 27% in districts with the highest probability of mechanization compared to no statistically or economically significant change in districts with low probability of mechanization. To further isolate the mechanization channel, I test for evidence of two alternative mechanisms: (1) an increase in production that may have led to more fires and (2) a shift in cropping patterns due to MNREGA’s role as implicit insurance. Increases in income driven by MNREGA may have led to changes in local demand that raised production. I show reduced form estimates of MNREGA’s impact on total hectares planted and total tonnage produced that suggest there were not large 2 Allen (2011) argues that firms’ choice to invest in labor-saving production techniques in response to higher relative wages in England is the central reason the Industrial Revolution began there. 3 Humphries (2013) and Kelly et al. (2014) argue that wages were not in fact higher in England at the time of the Industrial Revolution and adoption of labor-saving technology was driven by concerns other than wages. See also Voth et al. (2022). Higher subsequent wages were then a side-effect of increased labor productivity. Zheng and Kahn (2017) point out that the causality could also go in the opposite direction: firms in polluted cities in 19th England might have had to pay higher wages to attract workers because of the pollution. 3 changes in the area or tonnage of crops specifically associated with the use of fires. Alternatively, MNREGA provided implicit insurance that may have allowed farmers to shift into higher value but higher variance crops. While others have found that MNREGA induced shifts into higher value crops (Raghunathan and Hari, 2014), I do not find evidence of an increase in the overall volume of agricultural output due to MNREGA. Further, the crops that are shifted into are not associated with the use of fires. I show that air emissions from biomass burning also increase after the implementation of MN- REGA. I show that the rate of emissions from biomass burning of black carbon, organic carbon, and SO2 – pre-cursor pollutants to PM2.5 and PM10 – increased by between 30% and 50% after the implementation of MNREGA. I also show that the number of months in which the Indian ambient standard for PM2.5 is exceeded increases by 11% after implementation of MNREGA.4 My results speak to the growing body of work on “envirodevonomics” outlined by Greenstone and Jack (2013) and Jayachandran (2021) and the general question of how economic development – and concurrent changes in incomes, consumption, and production – interacts with environmental quality. I contribute most directly to the work examining the environmental impacts of raising incomes and wages (Gertler et al., 2006; Alix-Garcia et al., 2013).5 Pollution, in particular air pollution, tends to be substantially above recommended limits in developing countries (Alpert et al., 2012; Liu et al., 2018) and these elevated levels of pollution lead to meaningful negative impacts on health and other economic outcomes (Cropper et al., 2012; Ebenstein et al., 2017; Barrows et al., 2018). Work examining the causes of these elevated levels of pollution has focused on institutional failures (Greenstone and Hanna, 2014) and potentially lower willingness to pay for 4 The contribution of agricultural fires to air pollution is an acute policy challenge in India. As much as 40% of the pollution in Delhi during the winter may be due to crop residue burning (Bikkina et al., 2019). Increases in agricultural fires are believed to be a major reason that winter air quality in Delhi is among the worst in the world (Shyamsundar et al., 2019) with levels of PM2.5 that exceed WHO standards by as much as 1,000% (Liu et al., 2018). Aside from the Indian context, the use of agricultural fires to clear residue after harvest is a long-standing practice globally and contributes substantially to local air pollution. Agricultural burning is widely used in Pakistan and China and is used more intensively in Africa than anywhere else (Korontzi et al., 2006). The drivers of agricultural fires appear to be similar across countries (Cassou, 2018; Andini et al., 2018) thus understanding these drivers is important beyond India. 5 This work has often focused explicitly on the impact of anti-poverty programs. Alix-Garcia et al. (2013) show evidence of an anti-poverty and environment trade-off in places with little market access operating via a consumption channel. I show evidence that the consumption mechanism does not appear to be operating in my context. The same low levels of market access that drive the negative effects they observe may generate positive environmental change in other settings (Barbier, 2010). Determining the extent of this trade-off is especially important because the poor disproportionately live in more environmentally degraded areas (Dogo et al., 2017). 4 environmental quality (Kremer et al., 2011).6 I move this work forward by showing, using an exogenous shock, that raising wages led to an increase in pollution and introducing evidence that wage increases can causally increase pollution by inducing mechanization.The opening example of England in the Industrial Revolution provides a non-agrarian example where increased labor costs led to increases in mechanization, and power demand, in manufacturing that results in higher pollution. Proper Pigouvian policies have long been recognized as an efficient mechanism for internalizing environmental or other externalities (Weitzman, 1974; Stavins, 1996; Blackman, 2010; Kremer and Willis, 2016). My results suggest that the implementation of anti-poverty programs that distort labor markets in settings without proper Pigouvian taxation can lead to externalities that reduce the gains of the program itself. MNREGA led to an increase in pollution that likely had negative health consequences for rural residents that were the target of the anti-poverty efforts. Proper Pigouvian taxation of pollution might have offset some of these negative effects. Further, and in line with existing work on the role of tax policy in indirectly driving automation (Acemoglu et al., 2020), this paper suggests that by forcing firms to internalize the environmental costs of mechanization, Pigouvian policies might also have implications for the adoption of “inappropriate technologies.” There is a large literature on the drivers of structural transformation and a smaller, but growing literature on the environmental consequences of this transformation. There are two recent papers in that literature that are closely related to the current work. Garg et al. (2022) examine the expansion of the roads network in India that they argue facilitates structural transformation by enabling the movement of labor out of agriculture and find that crop burning increases substantially in villages after they receive a road. Like this paper, they demonstrate a causal link between elements of structural transformation and environmental damage. Caunedo and Kala (2021) separately examine the impact of facilitating access to mechanization technology in an RCT among farmers in India and find that, consistent with the results here, when the cost of mechanization is lowered relative to that of labor, farmers substitute mechanization for labor and release labor to other sectors of the economy. 6 In showing that MNREGA increases pollution emissions I contribute to the nascent literature on the health impacts of MNREGA (Thomas, 2015; Dasgupta, 2017; Nair et al., 2013). Emissions from agricultural fires have been shown to have negative health consequences (Rangel and Vogl, 2016; Pullabhotla, 2018) and so my results suggest that there may have been important health impacts from MNREGA, particularly in downwind districts, that are not captured by existing work. 5 This work also contributes to the on-going examination of what has driven the increase in agricultural burning in North India over the last 20 years. In addition to Garg et al. (2022), recent work has shown that shifting the beginning of the cropping season to minimize groundwater extraction has contributed to the increase in burning (Kant et al., 2022). I show that wage changes driven by MNREGA appear to be an additional driver of the increase in burning on top of changes induced by regulations around the timing of planting and the construction of rural roads.7 2 Agricultural fires in India and MNREGA 2.1 Fires in Indian agriculture While slash-and-burn agriculture is still widely used in some parts of the world (notably Africa and Indonesia) (Korontzi et al., 2006; Andini et al., 2018), the predominant use of fire in agriculture in India today is to clear harvest residue off of fields in order to prepare the field for the subsequent season’s planting (Jain et al., 2014; Bhuvaneshwari et al., 2019).8 Fires are widely used despite being nominally illegal since the mid-1990s (Lohan et al., 2018). While governments in some states have begun to enforce these laws, the expected cost of violation during my period of study is small. In 2012 the state government of Haryana handed out a total of roughly $12,000 in fines (Anand, 2016). That works out to an expected fine of 0.75$USD per fire in the state. Farmers face an average marginal cost of roughly 50$USD to clear their fields of residue without using fire (Ministry of Finance, 2018). While fires offer an inexpensive means of clearing crop residue, they have substantial negative effects. The primary negative effect is the increase of both local and global air pollutants. The largest source of carbonaceous particulate matter globally is crop residue burning (Cassou, 2018). More significant than their contribution to climate change is the impact that agricultural fires have on local air pollution and health. Source-apportionment studies have suggested that pollution from agricultural fires can raise local concentrations of PM2.5 to more than 1,000% above the WHO 24-hour guideline of 15µg/m3 (Bikkina et al., 2019; Balakrishnan et al., 2019; Liu et al., 7 While not examining the drivers of burning directly, Jack et al. (2022) show that an up-front payment for ecosys- tem services (PES) program that provide farmers payments not to burn their fields can reduce burning, underlining the fact that burning behavior is responsive to economic incentives. 8 For a brief discussion of the history of fire use and details of its global use, see Appendix 1.1. 6 2018). Exposure to these elevated levels of pollution leads to reductions in child height for age and weight for age scores (Singh et al., 2019; Rangel and Vogl, 2016) and increased infant mortality (Pullabhotla, 2018).9 The use of fire is particularly prevalent in the parts of India that grow crops in a coupled rice- wheat cropping system (Jain et al., 2014; Prasad et al., 1999) because of the short turn-around time between harvest of rice and planting of wheat. In this system farmers plant rice during the monsoon season (“kharif”), roughly from August to December, and wheat immediately following rice harvest during the pre-monsoon (“rabi”) season from January to March or April. This system of agriculture is particularly widespread in Punjab, Haryana, Uttar Pradesh and Uttarakhand (NAAS, 2017).10 Figure 1 shows the general pattern of agricultural fires across India in the years from 2003-2005, prior to the implementation of MNREGA. Consistent with the expected pattern of fire use by crop type, fires are concentrated in the northwest and Indo-Gangetic plain. The sugarcane producing areas of the country also show some local hot spots. Areas that predominantly produce oilseeds, namely Rajasthan and Maharashtra, have low levels of fire. Districts that plant more rice, wheat, and sugarcane have more fires in the pre-MNREGA period. Areas with more land in larger farms also appear to have more fires.11 In Table 1, I show the correlation between district characteristics in the pre-MNREGA period and the level of fires over the same time period. Column one shows the univariate relationship while column two shows the results including all predictors in the same regression. 9 There is a large body of literature that shows exposure to high levels of general particulate pollution – not solely pollution from crop burning – has negative economic consequences aside from health effects. Hanna and Oliva (2015) show reductions in pollution levels in Mexico City due to a refinery closure increased weekly hours worked. Chang et al. (2016) show substantial declines in worker productivity as exposure to PM2.5 increases in California and Deschenes et al. (2017) show that individuals make substantial defensive investments to avoid the consequences of air pollution. Aguilar-Gomez et al. (2022) provides a recent overview of the literature on human capital outcomes. There is also evidence that elevated levels of particulate pollution have direct negative consequences for agricultural yields (Burney and Ramanathan, 2014). 10 In Appendix Figures A4a-A4c I show the distribution of wheat, rice and sugarcane production across India in the pre-MNREGA period. I also show maps of the average crop coverage on October 31st each year as an approximation of which areas most heavily engage in rice-wheat production in Appendix Figure A5. 11 This may be due to the fact that larger farms enjoy more economies of scale from mechanizing the harvest. Evidence from both China (Wang et al., 2018) and Indonesia (Yamauchi, 2016) supports the claim that larger farms are more able to substitute mechanization for labor. 7 2.2 MNREGA 2.2.1 Roll-out of MNREGA The objective of MNREGA is explicitly to provide employment to rural households on projects that create public assets that “address causes of chronic poverty...so that the process of employ- ment generation is on a sustainable basis” (GOI, 2007). The essential feature of the program is a guarantee of 100 days of employment for rural households in a given financial year. The first districts received MNREGA in February of 2006 (“phase 1”). The targeting formula used to select these first districts is unknown (Sukhtankar, 2016).12 However, the government had an explicit goal of ensuring that poor districts were included in the first wave and every state had to have at least one district in the first wave (Shah and Steinberg, 2015). After the initial roll-out, another 130 districts received the program in April 2007 (“phase 2”) with the remaining roughly 270 districts receiving the program in April 2008 (“phase 3”). Table 2 summarizes the pre-MNREGA (pre-2006) level of a number of measures of economic development and the primary outcome variables in this study by MNREGA phase. The table highlights that earlier districts were on average poorer, more rural, slightly more agricultural and had less land in cash crops than later districts. Differences between each phase are not equal however, with the largest differences between districts in the first two phases and those in phase three. 2.2.2 Impact of MNREGA in the literature Despite, or perhaps because of, the size of MNREGA, program implementation has been incon- sistent. This inconsistency has resulted in heterogeneous impacts across states and difficulty in precisely assessing the true impact of MNREGA.13 In the most comprehensive review of the re- search on MNREGA, Sukhtankar (2016) suggests that there are four aspects of MNREGA that the substantial literature agrees on: (1) heterogeneity in impact, (2) despite the statutory requirement that employment be provided on demand, there has been meaningful rationing of employment pro- vision, (3) MNREGA has increased rural, private sector wages, and (4) the overall impact on rural 12 For details on the history of workfare in India as well as background on the structure of MNREGA see Appendix 1.2. 13 One reason for the heterogeneity in implementation has been substantial corruption and rent-seeking by imple- menting officials (Jha et al., 2009; Niehaus and Sukhtankar, 2013). 8 productivity is ambiguous. There is wide consensus in the published estimates that MNREGA increased unskilled wages by between 5% and 8%, increased labor participation in public works, and may have led to declines in the supply of labor to the private sector (Azam, 2011; Berg et al., 2012; Imbert and Papp, 2015). Imbert and Papp (2015) show that MNREGA increased wages of low-skill workers by 5% on average and this increase is concentrated in the dry season, when they show the bulk of MNREGA work occurs, and is accompanied by a decline in private sector labor supply of 1.3%. When they focus on the states in which the fraction of time spent on public works projects by rural, prime age adults was above 1 percent (“star” states), they find wages increased by 9% and private sector employment fell by 3%. Deininger and Liu (2013) show that MNREGA led to an increase in the accumulation of non-financial assets in the medium run. Raghunathan and Hari (2014) show that farmers plant riskier crops after the implementation of MNREGA which increases their incomes above the direct wage support of the program. Berg et al. (2012) suggests that the wage increase takes between 6 and 11 months after program implementation to materialize and is biased towards low-skill labor. In the only large-scale RCT on MNREGA to-date (Muralidharan et al., 2016) (MNS) show that improving the implementation of MNREGA increases wages by 7%.14 They focus only on Andhra Pradesh but note that the similarity in the size of their estimates on the impact of improving MNREGA to the impacts of initial implementation highlight the importance of implementation heterogeneity. They also do not find substantial evidence of impacts on migration but note that their migration data differs from the data used in previous studies.15 The impact of MNREGA on education also suggests MNREGA tightened local labor markets. Shah and Steinberg (2015) show that MNREGA decreased the educational attainment of 13-16 year- olds by decreasing school enrollment. The effects are similar for boys and girls. They suggest that 14 I describe a replication exercise of my main results using treatment in this RCT in Appendix 6. 15 Other attempts to estimate the impact of MNREGA on migration have been limited but suggest it may have had important impacts. A government review of the impacts of MNREGA claims that the program led to a 27% decline in cross-district migration caused by economic distress. One of the few academic studies of the impact of MNREGA on migration finds a reduction in rural to urban migration of around 8% (Imbert and Papp, 2014). While not direct evidence of reductions in migration, several studies have estimated a positive spillover of MNREGA on wages in neighboring districts that is hypothesized to operate via reductions in inter-district migration (Prasann, 2016; Muralidharan et al., 2017). These estimates suggest that wages in neighboring districts may have increased by as much as 9% and that the estimated impact of MNREGA on wages in districts in which it was implemented may be substantially underestimated due to these spillovers. 9 MNREGA induces these changes by tightening local labor markets which raises the opportunity cost of schooling. As a result, boys leave school to provide labor in the market and girls substitute into domestic work. Overall the existing literature on MNREGA suggests that it had positive impacts on wages, incomes, and potentially health outcomes. It is clear that MNREGA had meaningful impacts on local labor markets. In particular it appears to have tightened labor markets, especially for unskilled labor, by providing an outside option for unskilled labor in the form of public works at a wage that may have been above the prevailing agricultural wage and therefore made unskilled labor more expensive. 3 Research design and implementation I follow much of the existing literature analyzing MNREGA and utilize a difference-in-differences framework that takes advantage of the phased roll-out of MNREGA across the country to examine its overall impact on the use of fires in agriculture. As a robustness check, I also utilize the treatment pattern from the MNS experiment to determine if improving the implementation of MNREGA leads to an increase in agricultural fires. The design and results of this robustness check are described in Appendix 6. 3.1 Data Agricultural fires: The source of the raw data on fire presence comes from NASA’s Fire Information for Resource Management System (FIRMS).16 The FIRMS data provides the latitude and longitude and detection time of fires around the world using imagery from the MODIS and VIIRS imaging platforms. For all the primary analysis I use only data from MODIS because VIIRS does not provide sufficient temporal coverage.17 MODIS Aqua imagery is available from mid-2002 to the present. Imagery from the satellites is collected every 6-12 hours around the world and processed using NASA’s image processing algorithm to identify fires based on the emissions of mid-range infrared radiation. The algorithm is designed to filter out spurious signals (e.g. solar glare and gas flaring). NASA suggests that the imagery can detect fires as small as 16 I primarily use data from MODIS, available here: https://earthdata.nasa.gov/firms 17 See Appendix 1.5 for additional details on the difference between MODIS and VIIRS imagery. 10 50m2 if conditions are ideal and at sizes around 100m2 under average conditions. They also report that fires are located at the correct location with a spatial margin of error of less than 100m on average. Unfortunately, the resolution of MODIS is such that the data available in FIRMS only measures whether at least one fire exists in a given square kilometer. As a result, MODIS does not provide any information about the size of the detected fire or the total burned area. Further, it does not distinguish between pixels with only one fire and those with multiple fires. I combine the FIRMS data on fire presence with remotely sensed landcover data from the European Space Agency’s Copernicus system to determine the land uses on which fires occur.18 Copernicus land cover data assigns each pixel to a land class based on imaging that measures its reflectiv- ity. Classes include water/ice, urban, wetland, irrigated cropland, non-irrigated cropland, various classes of forest, and various classes of dry shrubland. I use the cropland classes to determine which fires occur on cropland and focus my analysis on these agricultural fires. I assign agricultural fires based on land use in 2006. I supplement the land use data from Copernicus with additional polygon layers from the Harvard Center for Geographic Analysis (CGA) that identify roads and urban areas in India. I verify that my agricultural fires do not include any fires that occur in locations that the CGA defines as urban. Finally, I aggregate the assigned agricultural fires to the district by month level. Figure 1 shows the annual average number of monthly fires per subdistrict prior to MNREGA. Fires are concentrated in the northwest of the country and along the Indo-Gangetic Plain. However, it is clear that burning occurs on cropland throughout India. Both of these facts are confirmed by Figure 2 showing the distribution of states by mean monthly fires in subdistricts. While fires are used in most states in India, their use in Punjab and Haryana is far more widespread than in other states (see notes on Figure 2). Weather Controls: I collect weather re-analysis data from ERA5. ERA5 is a weather re-analysis product produced by the European Commission’s Copernicus Climate Change Service.19 When working with weather data, there is a trade-off between using re-analysis data, which combines observed data with a physics model to provide data at a fine resolution over long 18 Copernicus data is available here:https://land.copernicus.eu/imagery-in-situ 19 Data available here: https://cds.climate.copernicus.eu/cdsapp!/dataset/reanalysis-era5-land?tab=overview 11 time periods, and data collected from weather stations. Station data has the advantage of being based only on observation and not including a modeled component. However, station networks often lack complete geographic coverage of a given area and station records may be incomplete, introducing temporal gaps in coverage as well. Re-analysis solves these problems but relies on models to do so. Despite the reliance on models, re-analysis data is broadly believed to provide a reasonable best estimate of weather variables (Auffhammer et al., 2013). As a result, it is widely used in both environmental and development economics (Schlenker and Lobell, 2010; Hsiang, 2016; Emerick, 2018). I chose re-analysis data for this project because comprehensive station data is not available. The ERA5 Land hourly product provides data at an hourly level on a grid of 0.1◦ ×0.1◦ , which translates to a 9km resolution. I collect data on cloud cover, temperature, and precipitation over the full sample from 2003 to 2014. I aggregate these weather variables to the district level in the primary analysis. Agricultural Data: I collect data on Indian agriculture from a number of sources. The first is the ICRISAT Meso data (Rao et al., 2012). This data is collected by the International Crops Research Institute for the Semi-Arid Tropics and measures the performance, structure, and behavior of the agricultural economy at the district level in India since 1966. I use data from 2003 to 2014. Because Indian district boundaries have changed over time, they provide both apportioned, where they adjust data for boundary changes, and unapportioned data in which data is not re-apportioned based on changes. I use the unapportioned data and manually re-apportion data to the district boundaries as they were recorded in the 2001 census to align with my other data sources. I use data on district level cropping patterns and land holdings from ICRISAT. I supplement agricultural data from ICRISAT with data from the Indian Ministry of Agriculture. I scrape agricultural census data from 2001, 2005, and 2011 at the district and subdistrict levels. This provides additional data on characteristics of agricultural holdings and planting patterns at the district level to supplement ICRISAT data and provides resolution at the subdistrict level not available in the ICRISAT data. I also scrape the agricultural input survey data from 2001, 2006, and 2011 to collect data on machine inputs to production.20 I also scrape data from the cost of 20 The agricultural census can be found here: http://agcensus.dacnet.nic.in/districtsummarytype.aspx. The agri- cultural input survey can be found here: http://inputsurvey.dacnet.nic.in/districttables.aspx. Cost of cultivation data is here: https://eands.dacnet.nic.in/Cost_of_Cultivation.htm. 12 cultivation survey to measure trends in input costs at the state level. SHRUG Data: To measure baseline conditions in the districts in the primary analysis, I use data from the 2001 Census compiled in the new Socioeconomic High-resolution Rural-Urban Geographic panel for India (SHRUG) dataset (Asher et al., 2019). I use the night lights data provided in SHRUG in robustness checks as well. Pollution Data: Previous work on the relationship between agricultural fires and pollution has relied on data from air quality monitoring stations (Pullabhotla, 2018). Like with weather data, this has the disadvantage of limiting the analysis of pollution to areas with monitors that have been active over the full time period. To get around this issue, I use satellite re-analysis data from the Modern-Era Retrospective analysis for Research and Applications (MERRA) database provided by NASA (Rienecker et al., 2011). This is a satellite based product used by economists to study air pollution from coal fired power in India (Barrows et al., 2018) that provides data on the monthly average emissions rates for black carbon, organic carbon, and sulfur dioxide (SO2 ) on a 0.5◦ ×0.625◦ grid. Importantly for my study, MERRA separately identifies the emissions of the above pollutants by source, including biomass burning. I also collect concentrations of black carbon, organic carbon, sulfur dioxide, and sulfate, which allows me to calculate the concentration of PM2.5 (He et al., 2019). 3.2 Empirical framework Following the difference-in-differences approach of Shah and Steinberg (2015), I estimate the effect of MNREGA on fire use as the change in fires before and after implementation of MNREGA within a district, controlling for month by year and district fixed effects. In the primary framework, districts are treated when MNREGA becomes statutorily effective in the district. Appendix Figure A1 shows the pattern of MNREGA roll-out. I assume that the number of fires Fimy in district i in month m of year y follows a Poisson distribution. This is appropriate given both the count nature of the data and the skewness of the distribution of monthly fires. Many district-months do not have any fires as a result of the nature 13 of agricultural production. I estimate variants of: ( ) log µ(Ximy ) = βNimy + ωi Wimy + δmy + ψi (1) ( ) where µ Ximy is the conditional mean of fires in district i in month m and year y , Nimy is an indicator that MNREGA had been implemented in district i in month m and year y . Wimy is a vector of weather controls including minimum and maximum temperature, cloud cover, and total precipitation. δmy is a month × year fixed effect and ψi is a district fixed effect. I estimate versions of equation 1 with and without Wimy . Because of the impact that cloud cover has on the ability of satellites to detect fires and the impact that temperatures and precipitation can have on the presence of fire, my preferred specification includes Wimy .21 β is the estimate of interest and measures the approximate percentage change in monthly fires after the implementation of MNREGA.I cluster standard errors at the district level (Abadie et al., 2017). In any difference-in-differences study, the crucial identifying assumption is that the trends in the outcome variable, in this case agricultural fires, would have been similar across all groups without the treatment. This is a fundamentally untestable assumption. Instead, common practice is to test whether the trends were parallel prior to the implementation of the policy being evaluated. Here that requires that the trend in agricultural fires in the years leading up to treatment in phase 1 districts was similar to that in phase 2 and phase 3 districts. Like Imbert and Papp (2015) and Shah and Steinberg (2015), I am relying on the assumption that the assignment of districts to MNREGA phases was based on features of the district that do not include the trend in fires and are not correlated with trends in fires. I show evidence in Figure 3 that the trends in pre-MNREGA fires did not differ across the phases. Figure 3 shows the results of an event study on the year of MNREGA implementation where the outcome is monthly agricultural fires. There is a clear and significant increase in fires after the implementation of MNREGA but little evidence of trends in the number of fires prior to MNREGA’s implementation. I show event studies for a number of other outcomes (e.g. area 21 Cloud cover introduces non-classical measurement error into my estimates and failing to control for it may lead to attenuation bias (see appendix section Appendix 2). There is also evidence that large-scale burning induces cloud creation (Fromm et al., 2010; Gatebe et al., 2012; Jain et al., 2014; Liu et al., 2020). This will exacerbate the attenuation effect. 14 planted in various crops) in the appendix. 4 Main Results I present the main results in two subsections. First, I discuss the impact of implementing MNREGA on fires across the entire country using my primary difference-in-differences specification and the full country sample. Second, I present results that show how the effect of MNREGA varies when districts are divided based on their score in an index measuring ease of mechanization. 4.1 Nationwide mean impacts Table 3 shows that MNREGA increased fires by approximately 21% after implementation (column 2 of Table 3). Although the confidence interval is wide – I cannot rule out an increase of between 11% and 30% at 95% confidence – I can easily reject a zero effect at 99% confidence. This suggests that MNREGA had a sizable impact on the frequency of agricultural fires. To put these estimates in perspective, the average number of monthly fires increased by approximately 40% from the beginning of my sample in 2003 to the end in 2014. The estimates here suggest that between 25% and 75% of that increase can be explained by MNREGA. The estimates of MNREGA’s impact when I do not to control for the weather conditions in the district at the time of fire detection are substantially smaller than the estimates in the preferred specification (column 1 of Table 3). This is consistent with the expectation that failure to control for cloud cover biases the estimates towards zero because it introduces non-classical measurement error.22 Even with this potential bias, when I do not control for weather, I estimate that MNREGA increased the frequency of agricultural fires by approximately 10% with a range from 1% to 18% at 95% confidence. While the estimated impact when I do not control for weather is smaller than my preferred specification, the expected impact is still meaningful and there is reason to believe that this estimate may be biased downwards. 22 Because the measurement error is not classical, the standard result that measurement error in the dependent variable only reduces precision, and does not introduce attenuation bias (Hausman, 2001), does not apply. See Appendix 2 for a discussion of this measurement error. 15 4.2 Robustness 4.2.1 TWFE robustness There has been an enormous amount of recent econometric research on the use of difference- in-differences in settings with staggered treatment where all units are eventually treated (see De Chaisemartin and D’Haultfoeuille (2022)), raising concerns about the validity of difference-in- differences estimators in these contexts. I summarize the most relevant portions of this literature and implement several of the robustness checks it describes in Appendix 3.1. I find that the results from my primary specification are not sensitive to varying weights in the aggregation of the indi- vidual 2x2 comparisons (de Chaisemartin and d’Haultfoeuille, 2020; Goodman-Bacon, 2021) nor to restricting comparisons to those between treated and untreated units (Callaway and Sant’Anna, 2019). Across a range of weights, I estimate treatment effects of between 0.18 and 0.23, compared to 0.21 in my primary specification. When I focus only on the treated to untreated comparison I estimate a treatment effect of 0.23 using the phase 1 to phase 2 and 3 comparison and one of 0.7 using the phase 2 to phase 3 comparison. In the latter comparison only 8% of the overall sample period is in the treated group. 4.2.2 Results in “star” states Imbert and Papp (2015) identify several states that were particularly effective at implementing MNREGA. In these states they find wage effects that are roughly double their average effects. If the impact of MNREGA on fires is driven by changes in wages, one would expect effects on fires to be larger in these states as well. In Appendix 3.2 I replicate their specification and show that effects are substantially larger in “star” states relative to both the average impact and effects in non-star states. I examine whether these differences simply reflect higher state capacity in these “star” states in Appendix 3.3. The RGGVY program is another government program, implemented by states, whose implementation timing may reflect general state capacity independent from capacity to implement MNREGA specifically. I do not find any difference in effects across states that were more proactive in implementing RGGVY compared to those that were less proactive. 16 4.2.3 Other robustness I show in Appendix 3.4 that implementation of MNREGA only increased agricultural fires and had no detectable impact on shrubland fires, forest fires, or fires on plantations. This suggests that my results are not driven by secular trends in the incidence of fires or changes in fire detection technology. In Appendix 3.5, I examine the seasonality of MNREGA’s impact. Fires are used after harvest in both the late fall (post-kharif ) and in the spring (post-rabi ). MNREGA had larger impacts on labor markets in the spring (Sukhtankar, 2016) and consistent with this I find larger impacts on fires after the rabi season. This also helps alleviate concerns that my estimation is simply picking up increases in fires that occur after the kharif season due to changes in the timing of planting kharif crops. While that mechanism surely explains some of the increase in fires after 2009, when the law driving changes in planting times was implemented (Kant et al., 2022), those changes cannot explain my overall results that are driven by changes prior to 2009 nor the effect I observe after the rabi season. In Appendix 3.6, I conduct a standard placebo test where I move treatment forward by one year for all phases. In this placebo treatment, I find that “MNREGA treatment” had no detectable impact on fires. In Appendix 3.7, I conduct something akin to a randomization inference test where I keep treatment timing the same but randomly assign districts to phases. This imposes a null hypothesis of no effects and provides an estimate of the likelihood that my results are due to random chance in the assignment of districts to treatment status. I find that my estimated effects are substantially different than the average effect estimated (clustered around a zero impact) of the randomization test. 5 Heterogeneity of MNREGA’s impact on agricultural fires Fires are used to manage crop residue differentially across India as the predominant crops vary with agricultural conditions. The ease of switching from manual harvesting to mechanical harvesting in response to wage changes also varies as average farm characteristics and the presence of the necessary mechanical equipment vary across districts. As a result, we should not expect the impact of MNREGA on fires to be consistent across districts. Rather, MNREGA should have had large impacts on the use of fires to clear crop residue in those districts that (a) tended to plant crops 17 whose harvest can be easily mechanized, (b) where combines were present to facilitate the hiring of these combines in place of manual labor, and (c) where, as I show in the model described in Appendix 4, farms tended to be larger. This expected variation in the impact of MNREGA provides an opportunity to test the hypoth- esis that it was increases in mechanization that linked MNREGA to increases in agricultural fires. To do so, I construct an index of mechanization for each district. The construction of this index is driven by the features of my model and the relationship between fire use and mechanization in Indian agriculture as described in the existing literature (e.g. among others Jain et al. (2014) and Bhuvaneshwari et al. (2019)). I follow Asher and Novosad (2018) and use an index rather than testing multiple measures of mechanization to limit concerns of multiple hypothesis testing. 23 To construct the index, I consider the district averages of the following variables in the years prior to 2006: the share of agricultural land in both marginal and medium or large holdings (less than 1 HA or more than 4 HA), the number of combines in 2006, and the production areas of wheat, rice, and sugarcane (the ability to mechanize varies across crops in India (Solomon, 2016)).24 To make the index, I turn these averages into Z scores. Each district receives a Z score for each individual measure. I then add them together to determine an index measuring the likelihood of mechanization in a given district. Higher index scores indicate a higher predicted level of mechanization. In the appendix, I show that this score is positively correlated with the best available measure of post-MNREGA mechanization – the number of combines in a district in 2011. I then divide districts into quartiles based on their mechanization index score and run my difference-in-differences regression from equation 1 on districts in each quartile.25 I find in the full, nationwide sample, districts with the highest mechanization scores have sub- 23 For details of each component of the index, as well as heterogeneity results by the individual components, see Appendix 3.8. 24 Proposition 2 of my model indicates the impact of MNREGA on agricultural fires will vary across districts based on the median farm size, AD . Specifically, given a distribution of A that is single-peaked and where the threshold farm size for adoption of mechanization (A ˆ) is above AD districts with initially higher AD will see a larger increase in fires after the implementation of MNREGA. Figure A2 indicates that the average distribution of farm sizes across India does appear to be single-peaked. Further, Foster and Rosenzweig (2011) report that in the pre-MNREGA period (the late 1990s) fewer than 10% of farms were mechanized, suggesting a level of A ˆ substantially above AD . 25 Li (2017) suggests that areas with higher land concentrations, an important piece of my mechanization score, had lower pre-MNREGA agricultural wages than other districts, which suggests my mechanization score is not simply capturing districts with high wages prior to MNREGA. Using, imperfect, wage data from ICRISAT I find that the correlation between the Z score of district wages in 2005 and the mechanization index is -0.11 (p = 0.71). I directly measure the correlation between the Z score of per capita GDP and my mechanization score and find that it is -0.07 (p = 0.45) which suggests that I am also not simply identifying wealthier districts. 18 stantially larger increases in the frequency of fire after the implementation of MNREGA. Districts in the highest quartile of the index see a 27% increase in monthly fires after the implementation of MNREGA (table 4). This is compared to an effect that is not different from zero in all other quartiles of the index. I show in Appendix 3.8 that this relationship is broadly consistent using each of the components individually as well. Further, in Appendix 3.9, I show that the importance of farm size in the heterogeneity analysis does not appear to be driven by differential likelihood of detecting fires in districts with larger farms. This suggests that districts in which more farmers could feasibly respond to the agricultural labor market shock caused by MNREGA by mechanizing their harvests saw larger increases in fires. Larger increases in the use of fire in these districts is consistent with the predictions of my model and suggests that it was an increase in mechanization in response to the labor shock imposed by MNREGA that drove the increase in fires. However, it is possible that MNREGA impacted the use of fires through other channels. I explore two of these in the next section. 6 MNREGA’s impact on agricultural output MNREGA’s shock to agricultural labor markets may have done more than simply changing low skill labor wages. Changing the cost and availability of labor may have caused farmers to change which crops they planted as well. Alternatively, by increasing incomes in local markets, MNREGA may have increased demand, and prices, for agricultural products and incentivized farmers to increase their production of existing crops. If this increased production occurred in crops that used fire as part of the production process the increase in fires may have been driven by the increased crop production as opposed to being driven directly by changes in the labor market. I call this a consumption effect. MNREGA may have also acted as an implicit insurance program for farmers (Sukhtankar, 2016). By guaranteeing the availability of outside employment in the event of a crop failure, MNREGA may have encouraged farmers to plant higher value but higher risk crops. If these crops are associated with more fire use than the previously planted, lower value crops, it could be that the insurance aspects of the MNREGA program are the driving force behind the change in fire frequency. I explore both of these potential explanations now. 19 6.1 MNREGA had little impact on production I find that MNREGA had little impact on the area planted (in 000s of HA) or total tonnage produced of crops most associated with fire production (wheat, rice, and sugarcane) (Jain et al., 2014). When I use the full, national sample I find no effect on total area planted or total tonnage of fire associated crops or in total area planted and tonnage in non-fire associated crops. In the appendix I show event studies for each of these crop outcomes. In all cases the trend after the implementation of MNREGA is flat or declining and in no case is it different from zero. Using the same difference-in-differences approach as described by equation 1, replacing agri- cultural fires (Fimy ) with the area in each crop (Aimy ) and tonnage produced (Timy ) in each crop as the outcome, I confirm the lack of impact suggested by the event studies (see Table 5). In all crops I examine (total crops, other non-fire associated crops, wheat, and sugarcane), except rice, I estimate an effect of MNREGA that is not statistically different from zero and is, in most cases, estimated to be close to zero with relatively high precision. The only crop for which I estimate a statistically significant effect at standard levels is rice, which sees a small (approximately 3%-4%) increase in area under production and total tons produced. This increase is small relative to the estimated change in fires and relatively imprecisely estimated. Based on these estimates I can reject an increase in fires of more than 0.40 % due to increasing area in wheat production at 95% confidence. Similarly I can reject changes greater than 1.60 % and 0.27 % for changes in the area in rice and sugarcane production.26 MNREGA may have led to small increases in the area planted and total tons produced of rice. I find no evidence that it led to meaningful increases in area planted or tonnage of other crops. The estimated increase in fires due to the changes in area produced is small both in absolute terms and relative to the estimated overall impact of MNREGA on the frequency of fires. No more than 30% of the smallest estimated increase in fires is estimated to be due to changes in area under production and the estimated changes in production would account for no more than approximately 10% of 26 To determine these bounds, I take the estimated change in area planted due to MNREGA from Table 5 and convert that estimated change in area into standard deviation units based on the distribution of area planted in each crop across all of India. This makes the units comparable to the units in Table 1 where I estimate the correlation between area planted in wheat, rice, and sugarcane and the frequency of fires. Based on the correlations in Table 1, I estimate the predicted change in fires for each crop based on the predicted change in area planted from Table 5. I use the delta method to calculate standard errors for these estimates. I report the upper end of the 95% confidence interval for the estimated change in fires. 20 the estimated change in fires in the preferred specification. 6.2 Changes in crop choice induced by MNREGA’s role as insurance Table 5 provides no evidence that the overall production of non-fire associated crops, which includes higher value, higher variance crops, increases as measured either by area in production or total tons. Existing work (Gehrke, 2013; Raghunathan and Hari, 2014) has suggested that farmers do in fact shift into higher variance, higher value crops such as cotton after the implementation of MNREGA. I find no evidence of such a change but more importantly, the production process of the crops that farmers may have shifted into (Gehrke, 2013) does not typically include fire (Jain et al., 2014; Bhuvaneshwari et al., 2019). It is therefore difficult to explain the observed increase in fires as being driven by MNREGA’s implicit insurance provision leading to transitions in crop type. In Appendix 3.11, I show that the absence of an impact of MNREGA on crop production in the average persists when examining outcomes by quartiles of the mechanization index. That is, there is no differential impact of MNREGA on crop production in different quartiles of the mechanization index. 7 Implications for air pollution The primary concern regarding the use of fires to clear crop residue stems from concern that this practice may increase air pollution. Previous estimates of the contribution of crop burning to pollution in Delhi suggest that substantial amounts, from 17% to 60% of particulate emissions in Delhi in the winter months are the result of upwind crop burning (Liu et al., 2018; Bikkina et al., 2019). In a direct examination of the impact of upwind crop burning on infant mortality in India, Pullabhotla (2018) suggests that an increase of five upwind fires in a given year increases the infant mortality rate by approximately 10%. The availability of satellite measures of emissions due specifically to biomass burning enables me to directly examine how the increase in agricultural fires caused by MNREGA translates to an increase in the emissions of three precursor pollutants of PM10 and PM2.5 : black carbon, organic carbon, and sulfur dioxide (SO2 ). I use data from the MERRA-2 satellite platform that measures the monthly emissions rates from biomass burning of these three pollutants by district to determine the impact of MNREGA on emissions. To do so I replace agricultural fires (Fimy ) as the outcome of 21 equation 1 with the monthly average emissions rate of each of these three pollutants and re-estimate the same difference-in-differences model on the national sample. 7.1 Direct effect of MNREGA on pollutant emissions To begin I present the event study for the emissions rate of each of the three pollutants in Figure 4. Each panel shows the trend in emissions rates for black carbon, organic carbon, and SO2 before and after the implementation of MNREGA. In all three panels there is a clear increase in emissions rates in the year of MNREGA implementation that continues to grow in the year after implementation before leveling off and remaining above pre-MNREGA levels. For both black carbon and organic carbon the pre-trends are relatively flat and not-distinguishable from zero. SO2 shows more of an increasing trend but this is driven primarily by the estimates on the earliest years in the sample.27 For all three pollutants that I examine, the implementation of MNREGA significantly increases the emissions rate from biomass burning. Beginning with black carbon, column 1 of Table 6 indicates that MNREGA increased the emissions rate from biomass burning by 38%. Organic carbon emissions also increase by 37% while SO2 emissions increase by an estimated 49%. These estimated effects indicate substantial increases in the average monthly emissions rates as a result of the implementation of MNREGA. They provide some confirmatory evidence that MNREGA increased the number of fires; if fires increase one would expect to see a corresponding increase in emissions from biomass burning. The size of the change in emissions rates is substantially larger than the estimated change in the number of fires. There are two explanations for this. The first is that the quantity of biomass burned per fire may have increased. This would be consistent with the model explored below of increased fires being driven by increased combine use. Previous estimates suggest fields cleared with combines have roughly 100% more biomass than fields cleared manually (Yang et al., 2008). To the extent that fires that consume more biomass have higher emissions (Smil, 1999) this increase in biomass is comparable to the difference between the estimated increase in emissions rates and the increase in fires. 27 I pool the early and late years of the sample because I do not have a balanced sample in event time. In other words, the estimates of the coefficient on event time -4 are only identified by districts in Phase 2 or Phase 3. As a result, I pool event time -5 and -4 with -3 in Figure 4. If I do not do this pooling, and drop event time -4 and -5 instead, the slight trend disappears. 22 Fires being used more intensively following MNREGA would also explain a larger increase in emissions than raw fires. While the MODIS satellite can detect fires as small as 100m2 it does not distinguish between pixels that have one fire and those that have many fires (Korontzi et al., 2006). As a result, if the implementation of MNREGA induces the use of fires in areas that already had frequent fires, MODIS will underestimate the increase in fires. Using more finely resolved data on fires in the period after 2014 from the VIIRS satellite platform, I show in the appendix that MODIS does indeed under-count fires relative to VIIRS (appendix Figure A14a and A14b). Translating these increases in emissions rates into precise changes in pollutant concentrations is difficult and would require a model of pollution dispersion. I do not create such a model here. Rather, I estimate the impact of the change in emissions on pollution concentrations in two ways. First, I estimate the correlation between emissions rates of black carbon, organic carbon, and SO2 using data on emissions concentrations from MERRA at the district × month level.28 These estimated correlations suggest that in months with emissions rates that are higher than average, PM2.5 concentrations are also higher than average. Consistent with this correlational evidence, I show that the implementation of MNREGA is associated with an increase of approximately 11% in the number of months in which PM2.5 levels exceed the annual ambient standard set by the Indian government.29 I look specifically at PM2.5 as it has the largest negative impacts on health (Behrer and Mauter, 2017; Muller and Mendelsohn, 2007; Nel, 2005; Chen et al., 2016). The annual threshold for ambient concentrations is set by the Central Pollution Control Board at 40 µg/m3 . I calculate the share of district × months that have an average PM2.5 level that exceeds this threshold and use a linear fixed effects regression in the difference-in-differences framework to determine if the implementation of MNREGA increased the share of months in which the threshold was exceeded. Column 4 of Table 6 shows that implementation of MNREGA increases share of months that exceed the threshold by 0.014 percentage points. This represents an 11% increase relative to the pre- MNREGA baseline rate.30 The magnitude of this effect is consistent with pollution levels exceeding 28 MERRA provides estimates of the monthly concentration of black carbon, organic carbon and SO4 . I convert these into a measure of PM2.5 using the formula described in He et al. (2019). 29 The Air (Prevention and Control of Pollution) Act 1981 set standards for Annual and 24- hour concentrations in the ambient air of a number of pollutants. They are detailed here: http://www.arthapedia.in/index.php?title=Ambient_Air_Quality_Standards_in_India 30 When I use the 24-hour threshold of 60 µg/m3 I observe a similar increase, 0.007 percentage points or roughly 10% on the pre-MNREGA baseline rate. 23 the standard for one additional month - the month of harvest - after the implementation of NREGA. Figure 5 shows the event study of the share of district×months that exceed the standard. There are no obvious pre-trends and a clear increase after the implementation of MNREGA. This discussion of pollution concentration has focused on the level of pollutant concentrations in the districts in which MNREGA is implemented. That leaves out the change in concentration levels that are downwind of the implementing district. These downwind effects may be more severe than the impacts in implementing districts (Behrer and Mauter, 2017). Even if they are less severe, considering pollution only in the implementing district presents only a partial picture of the impact of MNREGA on pollutant concentrations. 7.2 Welfare consequences of increased pollution There are a variety of negative economic and health consequences associated with exposure to air pollution. Mani et al. (2022) review many of these consequences in the context of South Asia while Aguilar-Gomez et al. (2022) provide a more comprehensive review. While existing studies of MNREGA’s impact on health find generally positive effects, it is possible that these effects would have been larger absent the increase in pollution. Previous work has estimated substantial infant mortality and reductions in birth weight from exposure to cropland fire smoke (Pullabhotla, 2018; Rangel and Vogl, 2016; Singh et al., 2019; Garg et al., 2022) and forest fire smoke (Jayachandran, 2009; Pullabhotla et al., 2022).31 Rangel and Vogl (2016) in particular find meaningful negative impacts on infant health based on changes in concentrations similar to what I observe after the implementation of MNREGA. My results combined with previous literature raise a natural question: what was the magnitude of the monetized damages due to the increase in agricultural fires caused by MNREGA? The difference-in-differences framework I use cannot estimate the impact of the increase in agricultural fires on health outcomes separately from the total impact of MNREGA. MNREGA likely impacted both pollution and health outcomes through channels in addition to changing agricultural fires. Estimating the total health impacts of MNREGA is well beyond the scope of this paper. Instead, I can roughly benchmark the health costs against the total spending on MNREGA to provide a 31 Additional work suggests that those same forest fires reduced the later in life wages of those children who were exposed but survived and that the pollution exposure rates they experienced may be comparable to the exposure rates residents of Delhi experience from crop burning (Tan-Soo and Pattanayak, 2019). 24 sense of the magnitude of the effects of MNREGA’s increase in crop burning. Jack et al. (2022) report that the monetized damages from deaths due to crop burning was USD$66 billion in 2018.32 The average quantity of dry matter burned annually from 2015-2018 was approximately 65% higher than in 2006 at the beginning of MNREGA implementation (Liu et al., 2020). That suggests that mortality costs in 2006 were roughly USD$27 billion prior to MNREGA implementation and my estimates indicate that MNREGA increased burning by 21%, implying an increase in mortality costs of around USD$5 billion. Cropper et al. (2019) calculate a much smaller VSL for India than used by Jack et al. (2022), of between USD$256,000 and $84,000 (USD$2015). Using the mid-point of these VSL estimates lowers the welfare costs to roughly USD$0.85 billion. This is a very crude benchmarking exercise, intended not to provide a precise estimate of the welfare costs of the increase in pollution but rather to provide a rough benchmark of their magnitude. The estimated magnitude of the welfare losses is large but plausible given the scope of burning, the size of the populations impacted, and the well-documented and severe costs of exposure to high levels of air pollution. The scale of the impacts is of the same order of magnitude as the annual federal expenditures on MNREGA, which has averaged between USD$4 billon and USD$11 billion between 2014 and 2021 An important dimension of these potential health effects is their potentially unequal distribution. Previous work (Lipscomb and Mobarak, 2016; Rangel and Vogl, 2016; Behrer and Mauter, 2017) suggests that while the region containing the source of emissions suffers negative consequences from those emissions, the majority of the damages may occur outside the region containing the polluting source. This is particularly true if there are major cities downwind of the emitting region. Rangel and Vogl (2016) find clear evidence that crop burning has negative effects in downwind areas but, because crop burning is driven by economic activity in the burning areas, increased burning may be associated with slight improvements in health in the areas where burning occurs. 8 Conclusion In this paper, I examine whether wage increases led to corresponding increases in pollution using the exogenous shock to wages generated by MNREGA. Because farmers appear to have responded 32 They arrive at this figure using a value of a statistical life (VSL) of USD$1 million. It is worth noting that this is roughly 1/10 of the value used by the U.S. EPA in cost benefit calculations. 25 to the wage increases generated by MNREGA by mechanizing, an indirect consequence of the program was an increase in the frequency of agricultural fires of between 9% and 21%. It also led to large increases, between 30% and 50%, in the emissions rate of black carbon, organic carbon, and SO2 from biomass burning. These pollutants are important contributors to both PM2.5 and PM10 pollution. In districts that had higher levels of a number of indicators of mechanization prior to MNREGA, I find an increase in agricultural fires of 27%. I test empirically for alternative explanations for the increase in agricultural fires but find the results to be most consistent with a model that suggests districts with greater ability to mechanize the harvest saw larger effects on fires. I cannot rule out that the increase in fires was driven by some other aspect of MNREGA or a feature of districts that is correlated with my mechanization index. However, higher pre-MNREGA wages are not highly correlated with higher scores in the mechanization index nor is district level wealth. I also observe low correlation between the mechanization index and pre-MNREGA fires, suggesting that the mechanization index is not simply identifying areas with high levels of fire prior to MNREGA. Despite this, it remains possible that the observed increase was driven by some factor other than changes in mechanization and these results should be interpreted with caution. With that in mind, it is important to remember that this is not an analysis of MNREGA in its entirety. Estimating the welfare consequences of MNREGA inclusive of the measured increase in agricultural fires is beyond the scope of this paper. Others have found meaningful increases in income, consumption, and health as a result of MNREGA that may offset any negative effects of increased emissions from increases in the number of fires. How the distribution of these benefits compares to the distribution of negative impacts from the emissions increase I measure deserves the attention of future work. The results presented here do not call for, nor justify, wholesale changes to MNREGA. Nor should they be interpreted as casting doubt on the value of anti-poverty programs generally. Rather, they suggest that policy makers should be cognizant of the potential consequences of large policy changes that raise wages and incomes for environmental quality and consider ways to mitigate negative impacts. Another potential policy response is to expand existing programs to encourage the adoption of the agricultural practices that do not require residue removal or mechanize such removal (Shyam- 26 sundar et al., 2019). Burning has declined as an agricultural practice in much of North America and Europe as the result of the adoption of “no till” agricultural practices (Korontzi et al., 2006; Marlon et al., 2008). Recent work has suggested that re-allocation of existing subsidies to more promising technology might encourage more widespread adoption of capital that reduces the need to burn (Shyamsundar et al., 2019). Alternatively, PES programs that encourage farmers not to burn with up-front payments have shown some promise in reducing burning (Jack et al., 2022). There are lessons here for policy makers beyond India as well. Many countries around the world have a goal of raising incomes with some form of a work guarantee scheme. The results here highlight that these kinds of interventions in labor markets, and programs to raise incomes generally, can have environmental impacts by changing production decisions. Governments may be able to reduce these environmental impacts by incorporating Pigouvian tax and subsidy policies, which encourage firms to internalize environmental externalities, into income raising policies. . Most generally, the results here indicate that in many settings Pigouvian policies that correct for environmental externalities may have implications for the rate of structural transformation. Specifically, if labor-saving mechanization has environmental externalities, firms may over-invest in mechanization in response to wage shocks relative to the efficient level of investment with regard to the pollution externality. Policies that correct for these pollution externalities slow the rate of mechanization, which may have positive spillovers in the context of arguments about inappropriate technology adoption. 27 References Abadie, Alberto, Susan Athey, Guido W Imbens, and Jeffrey Wooldridge (2017), “When should you adjust standard errors for clustering?” Technical report, National Bureau of Economic Research. Abay, Kibrom A, Gashaw T Abate, Christopher B Barrett, and Tanguy Bernard (2019), “Correlated non- classical measurement errors,‘second best’policy inference, and the inverse size-productivity relationship in agriculture.” Journal of Development Economics, 139, 171–184. Abraham, Sarah and Liyang Sun (2018), “Estimating dynamic treatment effects in event studies with het- erogeneous treatment effects.” Available at SSRN 3158747. Acemoglu, Daron (2021), “Harms of ai.” Technical report, National Bureau of Economic Research. Acemoglu, Daron, Andrea Manera, and Pascual Restrepo (2020), “Does the us tax code favor automation?” Technical report, National Bureau of Economic Research. Acemoglu, Daron and Fabrizio Zilibotti (2001), “Productivity differences.” The Quarterly Journal of Eco- nomics, 116, 563–606. Aguilar-Gomez, Sandra, Holt Dwyer, Joshua S Graff Zivin, and Matthew J Neidell (2022), “This is air: The ”non-health” effects of air pollution.” Working Paper 29848, National Bureau of Economic Research, URL http://www.nber.org/papers/w29848. Alix-Garcia, Jennifer, Craig McIntosh, Katharine RE Sims, and Jarrod R Welch (2013), “The ecological footprint of poverty alleviation: evidence from mexico’s oportunidades program.” Review of Economics and Statistics, 95, 417–435. Allen, Robert C (2011), “Why the industrial revolution was british: commerce, induced invention, and the scientific revolution 1.” The Economic History Review, 64, 357–384. Almås, Ingvild, Maximilian Auffhammer, Tessa Bold, Ian Bolliger, Aluma Dembo, Solomon M Hsiang, Shuhei Kitamura, Edward Miguel, and Robert Pickmans (2019), “Destructive behavior, judgment, and economic decision-making under thermal stress.” Working Paper 25785, National Bureau of Economic Research, URL http://www.nber.org/papers/w25785. Alpert, Pinhas, Olga Shvainshtein, and Pavel Kishcha (2012), “Aod trends over megacities based on space monitoring using modis and misr.” American Journal of Climate Change, 1, 117. Ambasta, Pramathesh, PS Vijay Shankar, and Mihir Shah (2008), “Two years of nrega: The road ahead.” Economic and Political Weekly, 41–50. Anand, Greeta (2016), “Farmers’ unchecked crop burning fuels india’s air pollution.” The New York Times, URL https://www.nytimes.com/2016/11/03/world/asia/ farmers-unchecked-crop-burning-fuels-indias-air-pollution.html. Andini, Ade, Sebastien Bonnet, Patrick Rousset, and Udin Hasanudin (2018), “Impact of open burning of crop residues on air pollution and climate change in indonesia.” Current Science, 115, 2259–2266. Arrow, Kenneth, Bert Bolin, Robert Costanza, Partha Dasgupta, Carl Folke, Crawford S Holling, Bengt- Owe Jansson, Simon Levin, Karl-Göran Mäler, Charles Perrings, et al. (1995), “Economic growth, carrying capacity, and the environment.” Ecological economics, 15, 91–95. Asher, Sam, Tobias Lunt, Ryunosuke Matsuura, and Paul Novosad (2019), “The Socioeconomic High- resolution Rural-Urban Geographic Dataset on India (SHRUG).” Working paper. Asher, Sam and Paul Novosad (2018), Rural roads and local economic development. The World Bank. 28 Auffhammer, Maximilian, Solomon M Hsiang, Wolfram Schlenker, and Adam Sobel (2013), “Using weather data and climate model output in economic analyses of climate change.” Review of Environmental Eco- nomics and Policy, 7, 181–198. Azam, Mehtabul (2011), “The impact of indian job guarantee scheme on labor market outcomes: Evidence from a natural experiment.” Available at SSRN 1941959. Balakrishnan, Kalpana, Sagnik Dey, Tarun Gupta, RS Dhaliwal, Michael Brauer, Aaron J Cohen, Jeffrey D Stanaway, Gufran Beig, Tushar K Joshi, Ashutosh N Aggarwal, et al. (2019), “The impact of air pollution on deaths, disease burden, and life expectancy across the states of india: the global burden of disease study 2017.” The Lancet Planetary Health, 3, e26–e39. Barbier, Edward B (2010), “Poverty, development, and environment.” Environment and Development Eco- nomics, 15, 635–660. Barrows, Geoffrey, Teevrat Garg, and Akshaya Jha (2018), “The economic benefits versus environmental costs of india’s coal fired power plants.” Available at SSRN 3281904. Behrer, Arnold Patrick and Meagan Mauter (2017), “Allocating damage compensation in a federalist system: Lessons from spatially resolved air emissions in the marcellus.” Environmental Science and Technology, 51. Berg, Erlend, Sambit Bhattacharyya, Rajasekhar Durgam, and Manjula Ramachandra (2012), “Can rural public works affect agricultural wages? evidence from india.” Bhargava, Anil K (2014), “The impact of india’s rural employment guarantee on demand for agricultural technology.” Bhuvaneshwari, S, Hiroshan Hettiarachchi, and Jay N Meegoda (2019), “Crop residue burning in india: Policy challenges and potential solutions.” International journal of environmental research and public health, 16, 832. Bikkina, Srinivas, August Andersson, Elena N Kirillova, Henry Holmstrand, Suresh Tiwari, AK Srivastava, DS Bisht, and Örjan Gustafsson (2019), “Air quality in megacity delhi affected by countryside biomass burning.” Nature Sustainability, 1. Blackman, Allen (2010), “Alternative pollution control policies in developing countries.” Borusyak, Kirill and Xavier Jaravel (2017), “Revisiting event study designs.” Available at SSRN 2826228. Burlig, Fiona and Louis Preonas (2016), “Out of the darkness and into the light? development effects of rural electrification.” Energy Institute at Haas WP, 268. Burney, Jennifer and V Ramanathan (2014), “Recent climate and air pollution impacts on indian agriculture.” Proceedings of the National Academy of Sciences, 111, 16319–16324. Callaway, Brantly and Pedro HC Sant’Anna (2019), “Difference-in-differences with multiple time periods.” Available at SSRN 3148250. Cassou, Emilie (2018), “Field burning.” Technical report, World Bank, Washington, DC. Caunedo, Julieta and Namrata Kala (2021), “Mechanizing agriculture.” Technical report, National Bureau of Economic Research. Chang, Tom, Joshua Graff Zivin, Tal Gross, and Matthew Neidell (2016), “Particulate pollution and the productivity of pear packers.” American Economic Journal: Economic Policy, 8, 141–69. Chen, Rui, Bin Hu, Ying Liu, Jianxun Xu, Guosheng Yang, Diandou Xu, and Chunying Chen (2016), “Beyond pm2. 5: the role of ultrafine particles on adverse health effects of air pollution.” Biochimica et Biophysica Acta (BBA)-General Subjects, 1860, 2844–2855. 29 Clemens, Michael A, Ethan G Lewis, and Hannah M Postel (2018), “Immigration restrictions as active labor market policy: Evidence from the mexican bracero exclusion.” American Economic Review, 108, 1468–87. Cook, Justin and Manisha Shah (2019), “Aggregate Effects from Public Works: Evidence from India.” Working paper. Cropper, Maureen, Shama Gamkhar, Kabir Malik, Alexander Limonov, and Ian Partridge (2012), “The health effects of coal electricity generation in india.” Resources for the Future Discussion Paper. Cropper, Maureen L, Sarath Guttikunda, Puja Jawahar, Zachary Lazri, Kabir Malik, Xiao-Peng Song, and Xinlu Yao (2019), “Applying benefit-cost analysis to air pollution control in the indian power sector.” Journal of Benefit-Cost Analysis, 10, 185–205. Dasgupta, Aparajita (2017), “Can the major public works policy buffer negative shocks in early childhood? evidence from andhra pradesh, india.” Economic Development and Cultural Change, 65, 767–804. de Chaisemartin, Clément and Xavier d’Haultfoeuille (2020), “Two-way fixed effects regressions with several treatments.” arXiv preprint arXiv:2012.10077. De Chaisemartin, Clément and Xavier D’Haultfoeuille (2022), “Two-way fixed effects and differences-in- differences with heterogeneous treatment effects: A survey.” Technical report, National Bureau of Economic Research. de Chaisemartin, Clement, Xavier D’Haultfoeuille, and Antoine Deeb (2020), “Twowayfeweights: Stata module to estimate the weights and measure of robustness to treatment effect heterogeneity attached to two-way fixed effects regressions.” de Chaisemartin, Clément, Xavier D’Haultfoeuille, and Yannick Guyonvarch (2019), “Did_multiplegt: Stata module to estimate sharp difference-in-difference designs with multiple groups and periods.” Deininger, Klaus and Yanyan Liu (2013), Welfare and poverty impacts of India’s national rural employment guarantee scheme: evidence from Andhra Pradesh. The World Bank. Deschenes, Olivier, Michael Greenstone, and Joseph S Shapiro (2017), “Defensive investments and the de- mand for air quality: Evidence from the nox budget program.” American Economic Review, 107, 2958– 2989. Deshingkar, Priya, Craig Johnson, and John Farrington (2005), “State transfers to the poor and back: The case of the food-for-work program in india.” World Development, 33, 575–591. Dogo, Harun, Carter Brandon, Martin Heger, Shun Chonabayashi, Joanne Gaskell, Mook Bangalore, Therese Norman, Jia Jun Lee, Pheobe Spencer, and Andres Chamorro Elizondo (2017), “Exploring hidden dimen- sions: Environmental and natural resource aspects of poverty.” Technical report. Ebenstein, Avraham, Maoyong Fan, Michael Greenstone, Guojun He, Peng Yin, and Maigeng Zhou (2015), “Growth, pollution, and life expectancy: China from 1991-2012.” American Economic Review, 105, 226– 31. Ebenstein, Avraham, Maoyong Fan, Michael Greenstone, Guojun He, and Maigeng Zhou (2017), “New evidence on the impact of sustained exposure to air pollution on life expectancy from china’s huai river policy.” Proceedings of the National Academy of Sciences, 114, 10384–10389. Emerick, Kyle (2018), “Agricultural productivity and the sectoral reallocation of labor in rural india.” Journal of Development Economics, 135, 488–503. FLA (2012), “Task and risk mapping of sugarcane production in india.” Technical report, The Fair La- bor Association, URL https://www.fairlabor.org/sites/default/files/documents/reports/task_ and_risk_mapping_of_sugarcane_production_in_india.pdf. 30 Foster, Andrew D and Mark R Rosenzweig (2011), “Are indian farms too small? mechanization, agency costs, and farm efficiency.” Fromm, Michael, Daniel T Lindsey, René Servranckx, Glenn Yue, Thomas Trickl, Robert Sica, Paul Doucet, and Sophie Godin-Beekmann (2010), “The untold story of pyrocumulonimbus.” Bulletin of the American Meteorological Society, 91, 1193–1210. Gaiha, Raghav (1997), “Do rural public works influence agricultural wages? the case of the employment guarantee scheme in india.” Oxford Development Studies, 25, 301–314. Garg, Teevrat, Maulik Jagnani, and Hemant K Pullabhotla (2022), “Structural transformation and environ- mental externalities.” arXiv preprint arXiv:2212.02664. Gatebe, CK, T Varnai, R Poudyal, C Ichoku, and MD King (2012), “Taking the pulse of pyrocumulus clouds.” Atmospheric environment, 52, 121–130. Gehrke, Esther (2013), “Does the indian employment guarantee reduce households’ risk exposure? assessing the effects of the nrega on crop choice.” Unpublished manuscript, IZA–Institute of Labor Economics. Retrieved from http://www. iza. org/conference_files/worldb2013/gehrke_e8393. pdf. Gertler, Paul, Sebastian Martinez, and Marta Rubio-Codina (2006), Investing cash transfers to raise long term living standards. The World Bank. GOI (2007), “National rural employment guarantee act 2005 (nrega): Report of the second year april 2006 – march 2007.” Technical report, Ministry of Rural Development, Government of India, URL https: //nrega.nic.in/NREGArpt_eng.pdf. GOI (2009), “Guidelines for implementation of works on individual land under nrega.” Technical report, Ministry of Rural Development, Government of India, URL https://nrega.nic.in/draft_guidelines. pdf. Gollin, Douglas, Stephen Parente, and Richard Rogerson (2002), “The role of agriculture in development.” American economic review, 92, 160–164. Goodman-Bacon, Andrew (2018), “Difference-in-differences with variation in treatment timing.” Technical report, National Bureau of Economic Research. Goodman-Bacon, Andrew (2021), “Difference-in-differences with variation in treatment timing.” Journal of Econometrics, 225, 254–277. Goodman-Bacon, Andrew, Thomas Goldring, and Austin Nichols (2019), “Bacondecomp: Stata module to perform a bacon decomposition of difference-in-differences estimation.” Greenstone, Michael and Rema Hanna (2014), “Environmental regulations, air and water pollution, and infant mortality in india.” American Economic Review, 104, 3038–72. Greenstone, Michael and B Kelsey Jack (2013), “Envirodevonomics: A research agenda for a young field.” National Bureau of Economic Research. Hanna, Rema and Paulina Oliva (2015), “The effect of pollution on labor supply: Evidence from a natural experiment in mexico city.” Journal of Public Economics, 122, 68–79. Hausman, Jerry (2001), “Mismeasured variables in econometric analysis: problems from the right and prob- lems from the left.” Journal of Economic perspectives, 15, 57–67. He, Lijie, Aiwen Lin, Xinxin Chen, Hao Zhou, Zhigao Zhou, and Peipei He (2019), “Assessment of merra-2 surface pm2. 5 over the yangtze river basin: Ground-based verification, spatiotemporal distribution and meteorological dependence.” Remote Sensing, 11, 460. 31 Hornbeck, Richard and Suresh Naidu (2014), “When the levee breaks: black migration and economic devel- opment in the american south.” American Economic Review, 104, 963–90. Hsiang, Solomon (2016), “Climate econometrics.” Annual Review of Resource Economics, 8, 43–75. Humphries, Jane (2013), “The lure of aggregates and the pitfalls of the patriarchal perspective: a critique of the high wage economy interpretation of the british industrial revolution.” The Economic History Review, 66, 693–714. Huneeus, Federico and Richard Rogerson (2020), “Heterogeneous paths of industrialization.” Technical re- port, National Bureau of Economic Research. Imbert, Clément and John Papp (2014), “Short-term migration and rural workfare programs: Evidence from india.” Journal of the European Economic Association. Imbert, Clement and John Papp (2015), “Labor market effects of social programs: Evidence from india’s employment guarantee.” American Economic Journal: Applied Economics, 7, 233–63. Jack, B Kelsey, Seema Jayachandran, Namrata Kala, and Rohini Pande (2022), “Money (not) to burn: Payments for ecosystem services to reduce crop residue burning.” Technical report, National Bureau of Economic Research. Jain, Meha, Pinki Mondal, Gillian Galford, Greg Fiske, and Ruth DeFries (2017), “An automated approach to map winter cropped area of smallholder farms across large scales using modis imagery.” Remote Sensing, 9, 566. Jain, Niveta, Arti Bhatia, and Himanshu Pathak (2014), “Emission of air pollutants from crop residue burning in india.” Aerosol and Air Quality Research, 14, 422–430. Jakiela, Pamela (2021), “Simple diagnostics for two-way fixed effects.” arXiv preprint arXiv:2103.13229. Jayachandran, Seema (2009), “Air quality and early-life mortality evidence from indonesia’s wildfires.” Jour- nal of Human resources, 44, 916–954. Jayachandran, Seema (2021), “How economic development influences the environment.” Working Paper 29191, National Bureau of Economic Research, URL http://www.nber.org/papers/w29191. Jha, Raghbendra, Sambit Bhattacharyya, Raghav Gaiha, and Shylashri Shankar (2009), ““capture” of anti- poverty programs: An analysis of the national rural employment guarantee program in india.” Journal of Asian Economics, 20, 456–464. Jitendra, Shreeshan Venkatesh, Ishan Kukreti, Kundan Pandey, Deepanwita Gita Niyogi, and Polash Mukerjee (2017), “India’s burning issue of crop burning takes a new turn.” Down to Earth, URL https://www.downtoearth.org.in/coverage/agriculture/river-of-fire-57924. Kant, Yogesh, Prakash Chauhan, Aryan Natwariya, Suresh Kannaujiya, and Debashis Mitra (2022), “Long term influence of groundwater preservation policy on stubble burning and air pollution over north-west india.” Scientific Reports, 12, 1–11. Kelly, Morgan, Joel Mokyr, and Cormac Ó Gráda (2014), “Precocious albion: a new interpretation of the british industrial revolution.” Annu. Rev. Econ., 6, 363–389. Korontzi, Stefania, Jessica McCarty, Tatiana Loboda, Suresh Kumar, and Chris Justice (2006), “Global dis- tribution of agricultural fires in croplands from 3 years of moderate resolution imaging spectroradiometer (modis) data.” Global Biogeochemical Cycles, 20. Kremer, Michael, Jessica Leino, Edward Miguel, and Alix Peterson Zwane (2011), “Spring cleaning: Rural water impacts, valuation, and property rights institutions.” The Quarterly Journal of Economics, 126, 145–205. 32 Kremer, Michael and Jack Willis (2016), “Guns, latrines, and land reform: Dynamic pigouvian taxation.” American Economic Review, 106, 83–88. Kuznets, Simon (1973), “Modern economic growth: findings and reflections.” The American economic review, 63, 247–258. Li, Yanan (2017), “Land inequality and the provision of public works-evidence from national rural employ- ment guarantee scheme.” Technical report, Unpublished Job Market Paper. Lipscomb, Molly and Ahmed Mushfiq Mobarak (2016), “Decentralization and pollution spillovers: evidence from the re-drawing of county borders in brazil.” The Review of Economic Studies, 84, 464–502. Liu, Tianjia, Miriam E Marlier, Ruth S DeFries, Daniel M Westervelt, Karen R Xia, Arlene M Fiore, Loretta J Mickley, Daniel H Cusworth, and George Milly (2018), “Seasonal impact of regional outdoor biomass burning on air pollution in three indian cities: Delhi, bengaluru, and pune.” Atmospheric environment, 172, 83–92. Liu, Tianjia, Loretta J Mickley, Sukhwinder Singh, Meha Jain, Ruth S DeFries, and Miriam E Marlier (2020), “Crop residue burning practices across north india inferred from household survey data: bridging gaps in satellite observations.” Atmospheric Environment: X, 8, 100091. Lohan, Shiv Kumar, HS Jat, Arvind Kumar Yadav, HS Sidhu, ML Jat, Madhu Choudhary, Jyotsna Kiran Peter, and PC Sharma (2018), “Burning issues of paddy residue management in north-west states of india.” Renewable and Sustainable Energy Reviews, 81, 693–706. Longhurst, JWS and DE Conlan (1970), “Changing air quality in the greater manchester conurbation.” WIT Transactions on Ecology and the Environment, 3. Mandal, Krishna G, Arun K Misra, Kuntal M Hati, Kali K Bandyopadhyay, Prabir K Ghosh, and Manoran- jan Mohanty (2004), “Rice residue-management options and effects on soil properties and crop productiv- ity.” Journal of Food Agriculture and Environment, 2, 224–231. Mani, Muthukumara, Hans Timmer, Martin Raiser, Jostein Nygard, Aiga Stokenberga, Nadia Sharmin, Margaret Triyana, Siddharth Sharma, Martin Heger, Adanna Chukwuma, Michael Norton, Shamemmur Rahman, Isabel Maria Ramos Tellez, Zara Ali, Utkarsh Saxena, and Ahmad Imran Aslam (2022), “Striving for clean air; air pollution and public health in south asia.” Technical report. Marlon, Jennifer R, Patrick J Bartlein, Christopher Carcaillet, Daniel G Gavin, Sandy P Harrison, Philip E Higuera, Fortunat Joos, MJ Power, and IC Prentice (2008), “Climate and human influences on global biomass burning over the past two millennia.” Nature Geoscience, 1, 697. Ministry of Finance, G.I. (2018), Economic Survey 2017-18 (Volume I and Volume II). OUP India, URL https://books.google.com/books?id=kiZZDwAAQBAJ. Morrison, Kathleen D (1994), “Monitoring regional fire history through size-specific analysis of microscopic charcoal: The last 600 years in south india.” Journal of Archaeological Science, 21, 675–685. Muller, Nicholas Z and Robert Mendelsohn (2007), “Measuring the damages of air pollution in the united states.” Journal of Environmental Economics and Management, 54, 1–14. Muralidharan, Karthik, Paul Niehaus, and Sandip Sukhtankar (2016), “Building state capacity: Evidence from biometric smartcards in india.” American Economic Review, 106, 2895–2929. Muralidharan, Karthik, Paul Niehaus, and Sandip Sukhtankar (2017), “General equilibrium effects of (im- proving) public employment programs: Experimental evidence from india.” Working Paper 23838, National Bureau of Economic Research, URL http://www.nber.org/papers/w23838. 33 NAAS (2017), “Innovative viable solution to rice residue burning in rice-wheat cropping system through concurrent use of super straw management system-fitted combines and turbo happy seeder.” Technical report, National Academy of Agricultural Sciences, New Delhi. Nair, Manisha, Proochista Ariana, Eric O Ohuma, Ron Gray, Bianca De Stavola, and Premila Webster (2013), “Effect of the mahatma gandhi national rural employment guarantee act (mgnrega) on malnutrition of infants in rajasthan, india: a mixed methods study.” PloS one, 8, e75089. NASA (2017), “India annual winter cropped area, 2001-2016.” Technical report, Center for International Earth Science Information Network (CIESIN), Columbia University, URL https://doi.org/10.7927/ H47D2S3W. Nel, André (2005), “Air pollution-related illness: effects of particles.” Science, 308, 804–806. Niehaus, Paul and Sandip Sukhtankar (2013), “Corruption dynamics: The golden goose effect.” American Economic Journal: Economic Policy, 5, 230–69. Prasad, R, B Gangaiah, and KC Aipe (1999), “Effect of crop residue management in a rice–wheat cropping system on growth and yield of crops and on soil fertility.” Experimental Agriculture, 35, 427–435. Prasann, Ashesh (2016), “The spillover effects of public works on labor allocation and wages: Evidence from national rural employment guarantee scheme, india.” Technical report. Pullabhotla, H (2018), “Fires, wind, and smoke: Air pollution and infant mortality.” Job Market Paper. Available here. Pullabhotla, Hemant Kumar, Mustafa Zahid, Sam Heft-Neal, Vaibhav Rathi, and Marshall Burke (2022), “Global biomass fires and infant mortality.” Pyne, Stephen J (2019), Fire: a brief history. University of Washington Press. Pyne, Stephen J and Johann G Goldammer (1997), “The culture of fire: an introduction to anthropogenic fire history.” In Sediment Records of Biomass Burning and Global Change, 71–114, Springer. Rabotyagov, Sergey S, Todd D Campbell, Michael White, Jeffrey G Arnold, Jay Atwood, M Lee Norfleet, Catherine L Kling, Philip W Gassman, Adriana Valcu, Jeffrey Richardson, et al. (2014), “Cost-effective targeting of conservation investments to reduce the northern gulf of mexico hypoxic zone.” Proceedings of the National Academy of Sciences, 111, 18530–18535. Raghunathan, K and S Hari (2014), “Providing more than just employment? evidence from the nrega in india.” Unpublished Working Paper. Rangel, Marcos A and Tom Vogl (2016), “Agricultural fires and infant health.” Technical report, National Bureau of Economic Research. Rao, P Parthasarathy, Lalmani Pandey, E Jagadeesh, Uttam Kumar Deb, Rajani Jain, and Kaushik Basu (2012), “Meso-level database coverage and insights village dynamics in south asia.” Rienecker, Michele M, Max J Suarez, Ronald Gelaro, Ricardo Todling, Julio Bacmeister, Emily Liu, Michael G Bosilovich, Siegfried D Schubert, Lawrence Takacs, Gi-Kong Kim, et al. (2011), “Merra: Nasa’s modern-era retrospective analysis for research and applications.” Journal of climate, 24, 3624–3648. Rodgers, HB (1960), “The lancashire cotton industry in 1840.” Transactions and Papers (Institute of British Geographers), 135–153. Rue, David, David Webster, and Alfred Traverse (2002), “Late holocene fire and agriculture in the copan valley, honduras.” Ancient Mesoamerica, 13, 267–272. 34 Sawyer, John (2019), “Estimating losses when cornstalk fields are accidentally burnt.” Tech- nical report, Iowa State University Extension, URL https://crops.extension.iastate.edu/ estimating-losses-when-cornstalk-fields-are-accidentally-burnt. Schlenker, Wolfram and David B Lobell (2010), “Robust negative impacts of climate change on african agriculture.” Environmental Research Letters, 5, 014010. Shah, Amita and Aasha Kapur Mehta (2008), “Experience of the maharashtra employment guarantee scheme: are there lessons for nregs?” Available at SSRN 1538914. Shah, Manisha and Bryce Millett Steinberg (2015), “Workfare and human capital investment: Evidence from india.” Technical report, National Bureau of Economic Research. Shyamsundar, P, NP Springer, H Tallis, S Polasky, ML Jat, HS Sidhu, PP Krishnapriya, N Skiba, W Ginn, V Ahuja, et al. (2019), “Fields on fire: Alternatives to crop residue burning in india.” Science, 365, 536–538. Singh, Prachi, Sagnik Dey, Sourangsu Chowdhury, and Kunal Bali (2019), “Early life exposure to outdoor air pollution: Effect on child health in india.” Smil, Vaclav (1999), “Crop residues: Agriculture’s largest harvest: Crop residues incorporate more than half of the world’s agricultural phytomass.” Bioscience, 49, 299–308. Solomon, S (2016), “Sugarcane production and development of sugar industry in india.” Sugar Tech, 18, 588–602. Solow, Robert M (1957), “Technical change and the aggregate production function.” The review of Economics and Statistics, 312–320. Stavins, Robert N (1996), “Correlated uncertainty and policy instrument choice.” Journal of environmental economics and management, 30, 218–232. Stewart, Frances (1978), Technology and underdevelopment. Macmillan. Sukhtankar, Sandip (2016), “India’s national rural employment guarantee scheme: What do we really know about the world’s largest workfare program?” Tan-Soo, Jie-Sheng and Subhrendu K Pattanayak (2019), “Seeking natural capital projects: Forest fires, haze, and early-life exposure in indonesia.” Proceedings of the National Academy of Sciences, 116, 5239– 5245. Thomas, MD (2015), “Guaranteeing children a better life? the impact of the mahatma gandhi national rural employment guarantee act on child health and learning in rural india.” University of Texas Austin Working Paper. Vasilica, STAN, Gina Fîntîneru, and Mircea Mihalache (2014), “Multicriteria analysis of the effects of field burning crop residues.” Notulae Botanicae Horti Agrobotanici Cluj-Napoca, 42, 255–262. Voth, Hans-Joachim, Bruno Caprettini, Alex Trew, et al. (2022), “Fighting for growth: Labor scarcity and technological progress during the british industrial revolution.” Technical report. Wang, Xiaobing, Futoshi Yamauchi, Jikun Huang, and Scott Rozelle (2018), “What constrains mechanization in chinese agriculture? role of farm size and fragmentation.” China Economic Review. Weitzman, Martin L (1974), “Prices vs. quantities.” The review of economic studies, 41, 477–491. Yadav, RNS (2007), “Mechanization of sugar cane production in india.” In Proceedings of the XXVI ISSCT Congress held in Durban in July/August, volume 7. 35 Yamauchi, Futoshi (2016), “Rising real wages, mechanization and growing advantage of large farms: Evidence from indonesia.” Food Policy, 58, 62–69. Yang, Shijian, Hongping He, Shangling Lu, Dong Chen, and Jianxi Zhu (2008), “Quantification of crop residue burning in the field and its influence on ambient air quality in suqian, china.” Atmospheric Envi- ronment, 42, 1961–1969. Zhang, Tianran and Martin Wooster (2016), “Small fire detection algorithm development using viirs 375m imagery: application to agricultural fires in eastern china.” In EGU General Assembly Conference Ab- stracts, volume 18. Zheng, Siqi and Matthew E Kahn (2017), “A new era of pollution progress in urban china?” Journal of Economic Perspectives, 31, 71–92. 36 9 Tables Table 1: What predicts average monthly fires over 2003-2005? Outcome: Average cropland fires by district (1) (2) Avg. Per Capita GDP .099 .124 ( .115) ( .132) Avg. Per Capita GDP, Ag .323*** .13 ( .078) ( .093) Avg. Area Planted in Rice (000s HA) .44*** .237* ( .099) ( .121) Avg. Area Planted in Wheat (000s HA) .389*** .189*** ( .034) ( .07) Avg. Area Planted in Sugarcane (000s HA) .238*** .23*** ( .046) ( .014) Avg. Area Planted in Other Crops (000s HA) -.025 -.122 ( .062) ( .105) Combines in 2006 (000s) .119*** .018 ( 9.0e-03) ( .015) Share of holdings >4 HA .5*** .471*** ( .064) ( .091) Covariates regressed separately X State FE X X Notes: In all columns the outcome is the average number of monthly fires in a district averaged over the years 2003-2005. In column 1 each row is a seperate regression. In column 2 all covariates listed down the left are included in the same regression. In all cases the specification is a fixed effects poisson regression. All independent variables are measured as Z scores so that units are comparable. Per Capita GDP (Ag) reports the average per capita value of (agricultural) GDP in the district in from 2003 to 2005 measured in Lakh |. Average Area Planted in Wheat, Rice, and Sugarcane measures the district average area in each crop from 2003-2005 in 1,000s of hectares as reported in the ICRISAT Meso data (Rao et al., 2012). Total area reports the total area in all other crops also in 000s of HA from the same data. Combines reports the number (in 000s) of self-propelled combines in the district as recorded in the 2006 Agricultural Input Survey. The Share of Holdings in each size class reports the share of acreage in the district in holdings within each size class in 2005 as reported in the ICRISAT Meso data. All columns include state fixed effects. Standard errors clustered at the state level are reported in parentheses. 37 Table 2: Summary of Pre-MNREGA Economic, Agricultural and Fire data Phase 1 Phase 2 Phase 3 (1) (2) (3) Data from 2001 Census Total Population (000s) 1,712.12 1,772.15 1,693.61 Total Households (000s) 320.12 326.19 320.43 % Rural 0.84 0.83 0.73 % Urban 0.16 0.17 0.27 % Scheduled Castes 0.15 0.15 0.15 % Literate 0.48 0.52 0.57 % with domestic electricity 0.88 0.87 0.91 % with agricultural electricity 0.74 0.59 0.72 % with electricty 0.75 0.65 0.74 % Paved road 0.49 0.56 0.68 % Mud road 0.82 0.82 0.76 Data from 2011 Census, 2012 SECC and VCF Data Distance to nearest place with >100k population 50.78 41.87 40.74 Per capita consumption, rural 15.64 16.59 19.02 % HH with cultivation as main income 0.35 0.37 0.38 Avg. Forest Cover, 2002-05 11.31 15.15 13.34 Avg. Night Lights 29.50 27.93 61.63 Data from NASA FIRMS Monthly Fires prior to MNREGA 3.61 3.73 7.33 Data from ICRISAT, Planning Commission and 2006 Input Survey Avg. Per Capita GDP, Ag 3.70 4.22 4.90 Avg. Per Capita GDP 14.39 15.13 21.26 Avg. Area Planted in Sugarcane (000s HA) 5.86 5.71 10.36 Avg. Area Planted in Rice (000s HA) 102.43 117.19 58.77 Avg. Area Planted in Wheat (000s HA) 39.76 57.85 61.63 Avg. Area Planted in Other Crops (000s HA) 172.67 165.39 207.96 Share of holdings >4 HA 0.26 0.28 0.33 Combines in 2006 (000s) 1.41 1.12 1.41 Notes: Columns 1-3 report the mean of the named variable by district according to the MNREGA phase of that district. Data from the 2001 Census, the 2011 Census, the 2012 SECC and the VCF data come from the SHRUG dataset (Asher et al., 2019). Fire data is downloaded and assembled from the NASA FIRMS data and is derived from imagery from the MODIS satellite. ICRISAT data comes from the ICRISAT meso dataset (Rao et al., 2012). GDP data is scraped from the Indian Planning Commission website and covers the years 2003-2005 for most districts. Information on combines is scraped from the Indian Ministry of Agriculture website and comes from the 2006 Agricultural Input survey. 38 Table 3: Nationwide impact of MNREGA on monthly fires Cropland Fires Post-MNREGA 0.096∗∗ 0.213∗∗∗ (0.044) (0.051) Districts 558 558 Months 144 144 N 80,352 80,352 Avg. monthly fires pre-MNREGA 5.46 5.46 District FE X X Year × Month FE X X Weather Controls X Notes: Each column represents separate regressions. In all columns the outcome is monthly cropland fires. In all columns the coefficient can be interpreted as the approximate percentage change in fires after MNREGA was statutorily implemented in a district. In all columns the base regression is a Poisson of fires in district i in month m in year y that includes district fixed effects and year by month fixed effects. In column 2 I include controls for the monthly average cloud cover, precipitation and minimum and maximum temperature in district i in month, year y. N refers to the number of districts × months included in each regression. The sample is a balanced, monthly panel of districts in India from 2003 to 2014. Heteroskedasticity robust standard errors clustered at the district level are in parentheses. (* p<.10 ** p<.05 *** p<.01). 39 Table 4: Heterogeneity of treatment impact by ease of mechanization index All of India (A)Quartile 1 of Ease of Mechanization Index Post-MNREGA -0.013 (0.087) Districts 140 Months 144 N 20,160 Avg. monthly fires pre-MNREGA 2.82 (B)Quartile 2 of Ease of Mechanization Index Post-MNREGA -0.008 (0.075) Districts 143 Months 144 N 20,592 Avg. monthly fires pre-MNREGA 2.7 (C)Quartile 3 of Ease of Mechanization Index Post-MNREGA 0.144 (0.112) Districts 135 Months 144 N 19,440 Avg. monthly fires pre-MNREGA 3.06 (D)Quartile 4 of Ease of Mechanization Index Post-MNREGA 0.265∗∗∗ (0.084) Districts 140 Months 144 N 20,160 Avg. monthly fires pre-MNREGA 13.28 District FE X Month × Year FE X Weather Controls X Notes: Coefficients can be interpreted as the approximate percentage change in fires after MNREGA was statutorily implemented in a district. The sample is all districts in India. The specification is a fixed effects Poisson where the outcome is monthly cropland fires in district i in month m in year y . The ease of mechanization index is the sum of a district’s Z score across measures of land concentration, combine presence and crop types. The mechanization index is calculated based on levels of each component variable in the district prior to 2006. I include district fixed effects, a year by month fixed effect, and weather controls. Each panel is a different quartile of the mechanization index with quartile 4 corresponding to the places mechanization is predicted to be easiest. N refers to the number of district × months included in each regression. Districts reports districts in each sample. The average number of monthly fires (the outcome) in the pre-treatment period in each quartile are reported. The sample is a balanced, monthly panel of districts in India from 2003 to 2014. Heteroskedasticity robust standard errors clustered at the district level are in parentheses. (* p<.10 ** p<.05 *** p<.01). 40 Table 5: Effect of MNREGA on crop production All Crops Wheat Rice Sugarcane Other Crops Area Tons Area Tons Area Tons Area Tons Area Tons Post-MNREGA 0.011 0.016 -0.001 0.014 0.025∗∗ 0.039∗∗ 0.034 -0.119 0.006 0.032 (0.007) (0.022) (0.019) (0.025) (0.010) (0.019) (0.040) (0.102) (0.008) (0.060) Districts 492 492 441 429 473 473 442 441 492 491 N 56,376 56,376 50,388 48,948 54,180 54,180 51,264 51,132 56,376 56,256 District FE X X X X X X X X X X Year × Month FE X X X X X X X X X X Weather Controls X X X X X X X X X X Notes: Each column represents separate regressions. In all columns the outcome is identified in the column heading. Area is measured in 000s of hectares while quantity produced is measured in 000s of tons. In all columns the coefficient can be interpreted as the approximate percentage change in the outcome after MNREGA was statutorily [ implemented ] ( a district. In all columns in ) the base regression is a fixed effects Poisson of the form E Cimy |Ximy = exp βP ostimy + γi + δmy where Cimy is the outcome in district i in 41 month m in year y . Postimy is a dummy variable equal to one after MNREGA treatment takes effect in district i. γi are district fixed effects while δmy is a year by month fixed effect. All columns include controls for the monthly average cloud cover, precipitation and minimum and maximum temperature in district i in month t. N refers to the number of districts × months included in each regression. The sample is a balanced, monthly panel of districts in India from 2003 to 2014. Heteroskedasticity robust standard errors clustered at the district level are in parentheses. (* p<.10 ** p<.05 *** p<.01). Table 6: Effect of MNREGA on emission rates of pollutants from biomass burning and months exceeding ambient PM2.5 standard Emissions Rates Ambient PM2.5 Standard Black Carbon Organic Carbon SO2 Share of months > standard Post-MNREGA 0.384∗∗∗ 0.378∗∗ 0.491∗∗∗ 0.014∗∗∗ (0.143) (0.163) (0.176) (0.003) Districts 558 558 558 560 Months 144 144 144 144 N 80,352 80,352 73,656 80,640 Pre-MNREGA Mean 45.81 536.09 52.09 .12 District FE X X X X Year × Month FE X X X X Weather Controls X X X X Notes: Each column represents separate regressions. In columns 1-3 the outcome is the monthly emissions rate of the pollutant named at the top of the column measured in ng/m2 s. In column 4 the outcome is the share of months in which the measured PM2.5 concentration exceeds the annual ambient air quality standard for India set by the Air Prevention and Control of Pollution Act (1981). Concentrations are measured in µg/m3 . All data comes from the MERRA-2 satellite system. In columns 1-3 the coefficient can be interpreted as the approximate percentage change in the outcome after MNREGA was statutorily implemented in a district. In column 4 the coefficient is the change in percentage points in the percent of months that exceed the Indian Ambient ( standard) of 40µg/m3 . In columns 1-3 the base regression is a fixed effects Poisson of [ ] the form log E Eimy |Ximy = βP ostimy + γi + δmy where Eimy is the outcome in district i in month m in year y . In column 4 I use a linear fixed effects specification of the form Timy = βP ostimy + γi + δmy where Timy is an indicator for whether district i had PM2.5 levels that exceed ambient standards in month m in year y . Postimy is a dummy variable equal to one after MNREGA treatment takes effect in district i. γi are district fixed effects while δmy is a year by month fixed effect. In all columns I include controls for the monthly average cloud cover, precipitation and minimum and maximum temperature in district i in month t. N refers to the number of districts × months included in each regression. The sample is a balanced, monthly panel of districts in India from 2003 to 2014. The mean of the outcomes prior to MNREGA for each is presented. Heteroskedasticity robust standard errors clustered at the district level are in parentheses. (* p<.10 ** p<.05 *** p<.01). 10 Figures 42 Figure 1: Pre-MNREGA Fires by Subdistrict Notes: The count of total fires by subdistrict in the years 2003-2005, prior to implementation of MNREGA. Darker areas had more fires. White areas had no fires. Data comes from the NASA FIRMS database. 43 Figure 2: Mean monthly agricultural fires by state from 2003-2005 Notes: The mean of the average number of fires within each subdistrict in a state. The mean for Andhra Pradesh, the location of the RCT, are shown with a lighter color bar and highlighted in red. The mean number of fires is calculated by finding the average number of fires by subdistrict across the pre-MNREGA years and taking the mean of these by state. The overall average for Andhra Pradesh, the location of the MNS RCT, is shown as well as the mean for the subdisricts in each quartile of the distribution of the number of pre-treatment fires. These are highlighted in red. The following are omitted from the Figure because their levels are so high including them would make it difficult to see variation in the remaining states: Punjab (Mean: 18.88) and Haryana (Mean:3.37). 44 Figure 3: Impact of MNREGA on Fires Notes: The estimating equation includes district fixed effects, month × year fixed effects, and weather controls. The outcome is the number of agricultural fires in month m in year y in district i. 95% CIs are show in dashed grey lines. The Figure uses the full sample. I pool event years less than -3 and greater than 4 into those boundary values. The base year is the year prior to MNREGA implementation. 45 Figure 4: Impact of MNREGA on emissions from biomass burning of various pollutants (a) Black Carbon (b) Organic Carbon (c) SO2 ( ) [ ] ∑ Notes: Each point is the estimated ωτ coefficient from the regression log E Fimy |Ximy = τ ∈T ωτ Yτ i + Wimy + ψi + δmy , where Yτ is an indicator for event-time year τ in the set T = {−3, −2, −1, 0, 1, 2, 3, 4}, ψi is a district fixed effect , δmy is a month × year fixed effect and Wimy are weather controls. Eimy is the average monthly rate of emissions of the named pollutant in ng/m2 s in month m in 46 year y in district i. 95% CIs are shown in dashed grey lines. The Figure uses the full sample. I pool event years less than -3 and greater than 4 into those boundary values. The base year is the year prior to the implementation of MNREGA. Figure 5: Event studies of the share of districts where PM2.5 exceeds annual Indian ambient standard Notes: Each ∑point is the estimated ωτ coefficient from the linear regres- sion Timy = τ ∈T ωτ Yτ i + Wimy + ψi + δmy , where Yτ is an indicator for event-time year τ in the set T = {−3, −2, −1, 0, 1, 2, 3, 4}, ψi is a district fixed effect , δmy is a month × year fixed effect and Wimy are weather controls. Timy is an indicator for whether district i had PM2.5 levels that exceed ambient standards in month m in year y . 95% CIs are show in dashed grey lines. The Figure uses the full sample. I pool event years less than -3 and greater than 4 into those boundary values. The base year is the year prior to MNREGA implementation. 47 Appendices For Online Publication Appendix 1 Additional Background Discussion Appendix 1.1 Fires & Agriculture Fire has been used by humans to manage landscapes for at least 40,000 years (Pyne and Goldammer, 1997) and in the cultivation of corn for at least 5,000 years (Rue et al., 2002). Sediment cores indicate that fire was used to manage agricultural fields in India at least 600 years ago (Morrison, 1994). The use of fires in agriculture is still widespread today in both developed and developing countries. Historically fires were used to clear land for planting in a swidden (“slash-and-burn”) style of agriculture and its use evolved over time to include preparing fields for harvest and clearing residue to prepare fields for re-planting (Pyne, 2019). Analysis of satellite imagery suggests that globally there are roughly 1.5 million fires annually with the largest number in Russia (Korontzi et al., 2006). The widespread use of fires in Russia, Eastern Europe (namely Ukraine) and North America mean that in absolute terms fire use in agriculture is more common in the developed world than the developing world (Cassou, 2018). On a per hectare basis, however, African countries are the most frequent users of fire. This is in part due to declines in recent years in the use of fires in North America and the European Union (Marlon et al., 2008), driven in part by increasing regulation around the practice. It should be noted, however, that the use of fires to clear crop residue was widespread in California and Western Canada until the mid-1990s when increasing concerns over air pollution led to regulation to eliminate the practice (Cassou, 2018). Frequent and long-term use of crop residue burning may decrease the productivity of agricultural land (Smil, 1999; Vasilica et al., 2014; Sawyer, 2019; Prasad et al., 1999; Mandal et al., 2004). It does so by destroying micro-nutrients in the soil and removing valuable fertilizer including nitrogen and phosphorus. Others have argued, however, that the extent to which burning negatively impacts soil quality is highly dependent on the type of soil. Further, farmers who shift from a production process that includes burning to one that removes residue from the field without burning may suffer short-term yield losses if they fail to adjust their use of fertilizer as well (Jain et al., 2014; Bhargava, 2014). Appendix 1.2 MNREGA Details The MNREGA program is the latest in a succession of work-based anti-poverty programs in India dating back to at least the British Raj (Imbert and Papp, 2015). The most notable program prior to MNREGA was the Maharashtra Employment Guarantee Scheme (MEGS) introduced in 1977 by the Government of Maharashtra (Shah and Mehta, 2008). MEGS was not a national program but much of the design of MEGS was incorporated into MNREGA. Like MNREGA the aim of MEGS was to provide employment to rural residents, focused on labor-intensive work, and targeted the formation of public goods. The number of person-days of work that it generated reached an early peak in 1980 and declined through the 1990s before climbing again through the early 2000s. Variation in the number of person-days supplied through MEGS appears to be tied to changes in the wage schedule and declines in the level of activist support for the program (Shah and Mehta, 2008). Of particular relevance to this study and MNREGA, estimates suggest that MEGS increased agricultural labor wages by around 18%(Gaiha, 1997). After independence, the national government experimented with a number of national rural workfare programs. A series of small-scale and pilot programs in the 1960s and 1970s were rolled into 48 the national Food for Work Programme (FWP) in 1977. Despite receiving significant investment, there is little evidence that the FWP had a meaningful impact on reducing rural unemployment, due, at least in part, to poor implementation and exclusion of the poorest citizens (Deshingkar et al., 2005). The FWP became the National Rural Employment Program in the 1980s. In 2001, the Sampoorn Grameen Rozgar Yojana (SGRY) program combined this with several existing poverty alleviation programs and rural infrastructure programs to consolidate effort and provide additional employment, food security, and infrastructure in rural areas (GOI, 2007). Wages were paid in a combination of cash and food supplies. By 2008 the SGRY program had been fully merged with MNREGA. The law creating MNREGA was passed in September 2005 and the program was implemented in the first districts in February 2006. Figure A1 shows the districts included in each phase. I map districts on 2001 geographies and apportion all data to the districts as they existed in 2001. To participate in MNREGA households obtain job cards from their local districts and then are able to apply for work whenever they would like. The district office is to provide work within 5 km of their house within 15 days of receiving their applications. The district must pay an unemployment allowance in cash if they fail to provide employment. Wages are to be paid at a statutorily set minimum wage that is not less than 60 Rs/day. MNREGA project lists are prepared at the district level and projects must be in one of the permitted categories. Those are: water conservation, drought proofing, flood protection, land development, minor irrigation, and rural connectivity (GOI, 2007).33 All projects must have a ratio of labor expense to material expense of at least 60/40 and the use of contractors and machinery is not allowed.34 The cost of MNREGA is split between the central government and state governments but, crucially, the full cost of unskilled labor is borne by the central government. State governments bear none of the cost of unskilled labor and 25% of the cost of skilled labor and materials, giving states an incentive to prioritize projects that use a greater share of unskilled labor. The scale of the program is remarkable. It is generally agreed that MNREGA is the largest workfare/rural poverty reduction program in the world (Ambasta et al., 2008). By 2014, 121 million job cards had been registered. In 2009-2010 there were 2.8 billion person-days of work conducted under the program (Sukhtankar, 2016). Participation appears to have grown steadily from implementation in 2006 to around 2013. Roughly 11% of the world’s population is covered by the program(Niehaus and Sukhtankar, 2013). In principle, MNREGA marked a shift from existing anti-poverty programs by being demand, as opposed to supply, driven (GOI, 2007). In practice, implementation challenges and state capacity may have limited the extent to which it was able to fully meet demand (Sukhtankar, 2016; Niehaus and Sukhtankar, 2013). By March 2007, demand for work exceeded supply in at least 30% of the states with significant demand (GOI, 2007). Despite this, MNREGA provided an average of three times the number of person-days of employment in its first years as SGRY provided in its final years. Appendix 1.3 Agricultural mechanization in India and fires Rice, wheat, and sugarcane are the three crops most associated with the use of fire on cropland in India. Of these, the harvest of rice and wheat can be mechanized using existing technology present 33 This work can occur on private land if it is owned by a member of a scheduled caste or tribe. The operational guidelines of MNREGA were modified in 2009 to allow for work to also be conducted on private land if the total holdings of the owner placed them in the “small” or “marginal” categories and the owner participated in the work (GOI, 2009). 34 The use of contractors and machinery was understood to be an obstacle to the effectiveness of previous programs in providing pro-poor benefits (Ambasta et al., 2008). 49 Figure A1: Map of MNREGA districts by phase of roll-out Notes: I show here the MNREGA phase of dis- tricts across India. Phase one districts received MNREGA in February 2006. Phase two received it in April 2007. Phase 3 received it in April 2008. 50 in India (Yadav, 2007; Solomon, 2016). Mechanizing the rice and wheat harvest is done using combines, which data from the Agricultural Input Survey shows are present throughout India prior to the implementation of MNREGA but tend to be present at higher levels in areas with higher use of fire. The relationship between mechanization and the use of fires is driven by the fact that harvesting with a combine leaves more residue on the field than harvesting by hand (Yang et al., 2008). Specifically, combine harvesting leaves stocks that tend to be around 30-40cm as opposed to the 10-15cm that harvesting by hand leaves. The higher stalks interfere with the ability of farmers to plant the following season’s crops and must be removed to facilitate planting (Jain et al., 2014; Smil, 1999; Cassou, 2018; Bhuvaneshwari et al., 2019).35 Burning is the least expensive way of dealing with this residue (Cassou, 2018), as interviews with farmers indicate: “Ankit Choyal Jat...offers an answer. ‘If I can clear my farm using a one-rupee matchbox, why will I spend thousands? (Jitendra et al., 2017)’” While combines were used in harvesting prior to the implementation of MNREGA, the shock that MNREGA provided to agricultural labor markets led to a substantial increase in their use. The only existing study of the impact of MNREGA on agricultural production processes (Bhargava, 2014) shows that the use of labor saving animal-based production processes increased after the implementation of MNREGA. Data from the Agricultural Input Survey shows a substantial increase in the average number of combines from 2006 to 2011 (Figure A3) – an increase that is much larger both in absolute and percentage terms than the increase from 2001 to 2006. Such an increase is consistent with broad state level trends from the Cost of Cultivation survey that show a decline in the amount of labor used in agriculture and an increase in the amount spent on machine inputs to production over the period that MNREGA was implemented. The notion that farmers responded to the impact of MNREGA on agricultural labor markets by mechanizing is supported by their own statements as well. Media interviews with farmers from areas that have seen the largest increase in burning frequently include quotes like the following (Jitendra et al., 2017): Hari Ram Karore, a 71-year-old farmer who owns more than 10 hectares (ha) in the same village, says, “We started using combine harvester machines to tide over the labour scarcity. The machine finishes the task of reaping, threshing and winnowing in a few hours and is also economical,” he adds. and Residents of villages around Kota say that mechanisation has killed the practice of using wheat stalk and straw as fodder, and burning is the only way out. “The cuttings left by the machines are too sharp. Not only do they injure us, even animals find it difficult to graze on,” says Shital Devi. These quotes are supported by more formal work interviewing farmers (FLA, 2012) that indi- cates farmers believe the supply of unskilled agricultural labor has declined as a result of MNREGA. This is consistent with survey data from the 2011 India Human Development Survey that suggest a large share (≈33%) of surveyed villages believe there to be fewer agricultural laborers than in the past and individual households report a 30% decline in the hours worked in agriculture. Fewer households in the IHDS report traveling for work after MNREGA as well. 35 There are alternatives to clearing the residue after a combine harvest. Namely tilling the standing residue back into the field. However this requires specialized planting equipment that is both expensive and not widely present in India (Jain et al., 2014; Bhuvaneshwari et al., 2019). 51 Figure A2: Average distribution of farm size by district, 2005 Notes: I plot here the distribution of farms by size class across all districts in my sample in India in 2005. The vast majority of farms are marginal or small with holdings of no more than 1 hectare. Data comes from the Agricultural Input Survey in 2005. Marginal farms are those less than 1 hectare (HA), small farms are between 1 and 2 HA, semi-small are between 2 and 4, medium are between 4 and 10 and large are greater than 10. Appendix 1.4 Cropping patterns across India Figures A4a-A4c show the distribution of planted area in wheat, rice, and sugarcane across India. I show the annual average in 000s of hectares in each crop over the years 2003-2005. Data comes from the ICRISAT meso database (Rao et al., 2012). Wheat production is clearly concentrated in Punjab and along the Indo-Gangetic plain. This broadly aligns with the areas that have the highest frequency of fires as shown in Figure 1 in the main text. The relationship between fires and coupled rice-wheat production is highlighted by comparing the map of rice production with the map of wheat production and fires. The districts with the highest frequency of fires are the districts where high production of rice and wheat appear to overlap. Notably, there are districts of high wheat production with low rice production and vice-versa, these districts do not appear to have as high a frequency of fire. Districts with a high area in sugarcane production also appear to have more frequent fires although the visual correlation is not as strong. The importance of the coupled rice-wheat production for fire use is brought out more clearly in Figure A5. Here I show the average share of a subdistrict’s area that is covered in crops on 52 Figure A3: Presence of combines over time Notes: The average number self-powered combines by district over time. Data scraped from the Indian Agricultural Input survey conducted in 2001, 2006 and 2011. The dashed lines indicate the phases of MNREGA. October 31st where the average is calculated across the years 2003-2005. The area covered by crops is calculated by NASA from remotely sensed imagery (NASA, 2017; Jain et al., 2017) and measures not the area of cropland but the share of a pixel on which crops are actively growing on October 31st each year. This is not a perfect proxy for areas that engage in coupled rice-wheat production but it captures areas that are growing rice crops during the post-monsoon season and the visual correlation between these areas and those that grow wheat in Figure A4a is high. Figure A5 highlights that areas that appear to most intensively engage in coupled rice-wheat production are also the areas that have the highest frequency of fires. Figure A24 shows that much of AP, and many of the sample subdistricts in MNS do not appear to engage in coupled rice-wheat production. 53 Figure A4: Cropping patterns in Wheat, Sugarcane and Rice across India (a) Wheat (b) Rice (c) Sugarcane 54 Notes: The average area planted by district annually in 000s of hectares in wheat, rice and sugarcane in the pre-MNREGA period from 2003-2005. Data comes from the ICRISAT meso dataset (Rao et al., 2012). Appendix 1.5 Relationship between MODIS and VIIRs fire detection VIIRs is similar to the MODIS platform in that it is a source for satellite based imagery. However, VIIRs is newer than MODIS, with the imagery available starting in 2012, and has higher resolution. The lack of data prior to 2012 means I cannot use VIIRs for the primary analysis. VIIRs and MODIS are able to detect roughly the same size fires but VIIRs provides data at a much finer pixel resolution than MODIS. VIIRs resolution is roughly 375m compared to 1km for MODIS. Both classify a pixel as having a fire if at least one fire is detected in that pixel. However, the finer resolution of VIIRs means it is able to count more pixels that contain fires. For example, two fires located 750m apart within a given square km would be counted as only one fire by MODIS but would likely be distinguished as two separate fires by VIIRs. Figure A6 shows clearly that MODIS detects substantially fewer individual fires than VIIRs. Korontzi et al. (2006) shows that MODIS and VIIRs are both highly accurate in counting a pixel that should contain a fire as containing a fire. So while MODIS underestimates fires it does not appear to mis-classify pixels that include low numbers of fires as non-fire. 55 Figure A5: Average crop coverage on October 31st Notes: This Figure shows the average share of pixels in a subdistrict across the areas of India for which data is available that have crop cover- age on October 31st over the years 2003-2005. Crop coverage is measured by reflectivity detected by satellite as described in Jain et al. (2017). Data comes from the Center for International56 Earth Science at Columbia Uni- versity NASA (2017). Figure A6: Monthly fires detected by MODIS and VIIRS from 2012-2017 Notes: I count the number of fires detected by the MODIS and VIIRs platforms in each month from 2012-2017. Each black circle is a month plotted according to the fires detected by MODIS and VIIRs in that month. The dashed blue line is the 45◦ line. If MODIS and VIIRs detected the same number of fires each month would be on the 45◦ line. The observed distribution suggests that MODIS undercounts fires relative to VIIRs. 57 Appendix 2 Measurement error There are at least three types of measurement error present in my measurement of agricultural fires. The first is introduced by the presence of cloud cover, which makes it difficult for satellites to measure the presence of fires. The presence of clouds therefore leads me to systematically under count the true number of fires. This will attenuate my estimates towards zero. To see this consider a district that has 0 “true” fires prior to treatment and 4 fires after treatment. In that case the “true” treatment effect is an increase of 4 fires. Now consider the case with a constant level of cloud cover that reduces my counts of fire by 50%. Now I still measure 0 fires prior to treatment but only 2 fires after treatment and estimate that treatment increased fires by only 2 fires. The same attenuation would exist in the case of a negative treatment effect. The second case relates to the fact that widespread fire use can lead to the creation of clouds (Fromm et al., 2010; Gatebe et al., 2012; Jain et al., 2014). In the Indian context in particular Liu et al. (2020) reports that failure to account for clouds and haze generated by fires leads to substantial under-counting of fires in MODIS data and the undercounting increases over time due to increases in fire frequency. This will exacerbate the attenuation described above because it suggests I under count fires by more when there are more fires. In the example above it suggests that clouds lead to counting 2.5 fires prior to treatment but, due to increased cloud cover driven by the increase in fires, I only count 3 fires after treatment and therefore estimate that treatment only increased fires by 0.5. This attenuation occurs regardless of whether the pre- or post-treatment number of fires is bounded at zero. The final source of measurement error is due to the large resolution of MODIS. As I discuss below, MODIS undercounts the number of fires because its resolution is 1 square kilometer and it cannot distinguish between pixels with 1 or 10 fires. The impact of this measurement error is more ambiguous than the previous two but recent work suggests it may also lead to attenuation (Abay et al., 2019). Appendix 3 Additional Robustness Checks and Discussion Appendix 3.1 Alternative difference-in-differences estimation There has been an explosion in the literature on the use of two-way fixed effects (TWFE) difference- in-differences estimators in the last several years (e.g. (Borusyak and Jaravel, 2017; Abraham and Sun, 2018; Callaway and Sant’Anna, 2019; de Chaisemartin and d’Haultfoeuille, 2020; Goodman- Bacon, 2021; Jakiela, 2021)) that is usefully summarized in De Chaisemartin and D’Haultfoeuille (2022). Most relevant for this paper the new TWFE literature indicates that the TWFE estimator can be biased in settings where: 1. Treatment timing is staggered. 2. Treatment effects are heterogenous over time. 3. All units are ultimately treated. The implementation of MNREGA clearly satisfies conditions #1 and #3. My event study estimates, and economic theory, suggest that it also satisfies #2. This raises two obvious questions: (1) are the TWFE estimates presented here biased and (2) if so to what extent are they biased? In the staggered treatment framework bias arises, in part, because the TWFE coefficient is a weighted average of the ATEs of the many 2x2 DiD comparisons where weights are determined by the relative size of treatment groups and the timing of treatment (Goodman-Bacon, 2021; 58 De Chaisemartin and D’Haultfoeuille, 2022). Because these weights are not constrained to be positive, or the same sign, this weighted average does not satisfy a no sign reversal condition (De Chaisemartin and D’Haultfoeuille, 2022). That means that an estimated TWFE can be neg- ative even if the true treatment effect is universally positive (and vice versa). The potential bias from negative weights is more acute in settings with a long-post treatment period, when treatment groups are unevenly sized, when staggered treatment is far apart in time, and when all units spend a large portion of the sample treated (Jakiela, 2021; Goodman-Bacon, 2021). Treatment in MN- REGA was clustered - all units are treated within a 3 year period - and treated groups are relatively evenly sized. However, there is a long-post treatment period in which all units are treated in my primary sample. Several alternative estimators and estimation approaches have been proposed to deal with the challenges that can arise in TWFE estimation and many have been implemented in common statis- tical packages (e.g. Goodman-Bacon et al. (2019); de Chaisemartin et al. (2019)). However, none of the current implementations are appropriate for TWFE estimation with a Poisson specification. Instead, I conduct several robustness tests similar to those proposed in Jakiela (2021), de Chaise- martin and d’Haultfoeuille (2020) and Callaway and Sant’Anna (2019). First, I examine the sign of the weights using the decomposition proposed in Goodman-Bacon (2018). This decomposition measures the weights placed on each of the 2x2 group comparisons that are aggregated into the TWFE estimator. In my case those are an Early-Late, a Mid-Early/Late, and a Late-Early that correspond to the treatment timing for phase 1, phase 2, and phase 3. The decomposition is possible, despite the non-linearity of my estimator, because the weights depend only on the size of treated groups and the timing of treatment. This decomposition indicates that all of the weights are positive with the largest weights placed on the Late-Early comparison. As highlighted in both Goodman-Bacon (2021) and Borusyak and Jaravel (2017) the comparison of late treated groups to already treated groups can introduce bias in a context with non-constant treatment effects. Goodman-Bacon (2018) argues that estimators based on that 2x2 comparison with treatment effects that grow over time, arguably the setting for MNREGA, the TWFE estimate will be a lower bound of the true treatment effect. de Chaisemartin and d’Haultfoeuille (2020) propose a decomposition of the weights that builds on Goodman-Bacon (2018) and estimates weights individually for every 2x2 comparison in the data. That is, the weight on each group and time cell where groups are treatment units (or some other geographic entity) and time is the relevant time step in the data. In my sample of MNREGA treatment that results in roughly 22,000 group by time cells where groups are Indian districts and time is measured as sample months. Both de Chaisemartin and d’Haultfoeuille (2020) and Goodman-Bacon (2021) note that changing length of the post-period changes the weights used because it changes the relative treatment timing and the share of time that all units are treated (in a universal treatment design). This leads Jakiela (2021) to propose a robustness check where later periods are dropped and the TWFE is re-estimated. I conduct a similar robustness check in Figure A7 where I sequentially drop years out of the post- treatment period and re-estimate the primary specification. The top portion of the figure shows that the point estimates are robust to dropping these years, varying by an average of 0.03 percentage points from the primary results. In the lower portion of the figure I use the decomposition suggested by de Chaisemartin and d’Haultfoeuille (2020) and implemented by de Chaisemartin et al. (2020) to measure the share of the 22,000 2x2 comparisons that receive a negative weight. Unsurprisingly this figure declines as the post-period is shortened but the point estimates remain largely stable despite these changes in the share of negative weights. In addition to bias from non-convex combinations of the individual 2x2 estimates caused by weights of different signs bias in TWFE specifications can arise when estimating treatment effects 59 Figure A7: Point estimate robustness to TWFE weight changes Notes: The top portion reports the point estimates and 95% CI from the primary specification where the sample is all pre-treatment years and sequentially fewer post-treatment years as indicated in the labels on the x-axis. The bars show the share of 2x2 TWFE comparisons that receive negative weights based on the decomposition in de Chaisemartin et al. (2020). by comparing the change at treatment in previously non-treated units to previously treated units (Borusyak and Jaravel, 2017; Callaway and Sant’Anna, 2019; Goodman-Bacon, 2021; De Chaise- martin and D’Haultfoeuille, 2022). This bias arises because, unless treatment effects occur only at the time of treatment and have no lagged effects, the “control” units in this 2x2 comparison do not satisfy the parallel trends assumption at the time of treatment. In the case of positive treatment effects that grow over time that will lead to under- (and potentially wrong-signed) estimated true effects. This is obviously a concern in the MNREGA setting as the 2x2 comparison of phase 2 to phase 1 and phase 3 to both phase 2 and phase 1 districts compares a non-treated unit to a previously treated unit. In an ideal case the TWFE could be estimated by comparing treated units only to those that are never treated. This is the core of the estimator proposed in Callaway and Sant’Anna (2019). In the case of MNREGA units this comparison is not possible because no never treated units exist. Instead, I can estimate the model using only the comparisons between treated and non-treated units. I do this by estimating the primary specification with a sample construction such that the treatment effect is based on a comparison of outcomes in phase 1 districts to those in phase 2 and phase 3 prior to treatment in phase 2 and phase 3 and, separately, a sample where the treatment 60 effect is estimated by comparing treatment in phase 2 districts to untreated phase 3 districts. Because there are no non-treated units when phase 3 districts are treated I cannot estimate the effects in this way using a sample that includes treatment in phase 3. Table A1 shows the results of estimating the impact of MNREGA on fires using these restricted samples. The treatment effect of MNREGA comparing treated phase 1 districts only to non-treated phase 2 and 3 districts is a 23% increase in fires, roughly equivalent to the estimate from the full sample. Comparing treated phase 2 districts to untreated phase 3 districts, a comparison that is only able to estimate an effect using treatment in 8% of the months in the full sample, finds a statistically insignificant 7% increase in fires. Table A1: Nationwide impact of NREGA on monthly fires in selected samples Cropland Fires Treated Phase 1 vs. Treated Phase 2 vs. Untreated Phase 2 & 3 Untreated Phase 3 Post-NREGA 0.236∗∗∗ 0.072 (0.061) (0.088) Districts 551 551 Months 62 56 N 33,968 30,792 District FE X X Year × Month FE X X Weather Controls X X Notes: Each column represents seperate regressions. In all columns the outcome is monthly cropland fires. In all columns the coefficient can be interpreted as the approximate percentage change in fires after NREGA was statutorily implemented in a district. In all columns the base regression is a fixed effects poisson of the form yit = βP ostit + γi + δt where yit is the outcome in district i in month t. Postit is a dummy variable equal to one after NREGA treatment takes effect in district i.γi are district fixed effects while δt is a year by month fixed effect. In both columns I include controls for the monthly average cloud cover, precipitation and minimum and maximum temperature in district i in month t. N refers to the number of districts × months included in each regression. The sample varies across columns. In column 1 all district×months that are treated in either phase 2 or 3 are dropped. In column 2 all district×months that are treated in either phase 1 or 3 are dropped. Heteroskedasticity robust standard errors clustered at the district level are in parentheses. (* p<.10 ** p<.05 *** p<.01). Collectively these results suggest that the results from the primary specification are not driven by the structural biases that can be a challenge in some TWFE settings. The results do not appear to be a consequence of the weighting imposed by the structure of treatment. Nor do they appear driven by comparisons of untreated units to previously treated units. While the estimates are necessarily uncertain due to a shorter post period, effects estimated using only non-treated units as controls yield point estimates that are broadly similar in sign and magnitude to the estimates from the primary specification. This is consistent with the discussion in De Chaisemartin and D’Haultfoeuille (2022) that notes TWFE and heterogeneity-robust DID estimators may result in similar point estimates and our understanding of the conditions under which they diverge is still nascent. 61 Appendix 3.2 Imbert & Papp (2015) Star State Replication Imbert and Papp (2015) show that when they focus on the states in which the fraction of time spent on public works projects by rural, prime age adults was above one percent (“star” states) the effects of MNREGA on wages are roughly double what they estimate in the full sample. They suggest this is due to better implementation in these states. If farmers increase their use of fires to clear their fields is a consequence of higher labor costs one might expect that areas that see larger increases in wages also saw larger increases in fires. I can test this directly by replicating the approach taken by Imbert and Papp (2015) and comparing the change in fires in districts that were located in star states to those located in non-star states. I show the results of this examination in Table A2. I find substantially larger effects in star states compared to non-star states. The impacts in star states are double the impacts I observe in the full sample and impacts in star states are nearly 5x larger than in non-star states. This is consistent with MNREGA having had a much larger impact on labor markets in these states and consequently having a larger impact on fires. Table A2: MNREGA Star Table Cropland Fires (A) Star Status Star × Post 0.343∗∗∗ 0.505∗∗∗ (0.058) (0.069) Non-star × Post 0.032 0.135∗∗ (0.048) (0.055) Districts 558 558 Months 144 144 N 80,352 80,352 District FE X X Year × Month FE X X Weather Controls X Notes: Each column represents separate regressions. In all columns the outcome is monthly cropland fires. In all columns the base regression is a fixed effects Poisson of the form yit = βP ost + ψ [P ost × M N REGA × Star]+ γi + δt where yit is the outcome in district i in month t. Post is a dummy variable equal to one after MNREGA treatment takes effect in a given phase and MNREGA is a dummy indicating the MNREGA phase of district i. Star is an indicator for whether Imbert and Papp (2015) consider the state a Star state, one where MNREGA was particularly well-implemented and where they find the wage effects of MNREGA to be concentrated. γi are district fixed effects while δt is a year by month fixed effect. In column 2 I include controls for the monthly average cloud cover, precipitation and minimum and maximum temperature in district i in month t. N refers to the number of district × months included in each regression. The sample is a balanced, monthly panel of districts in India from 2003 to 2012. Heteroskedasticity robust standard errors clustered at the district level are in parentheses. (* p<.10 ** p<.05 *** p<.01). Appendix 3.3 RGGVY comparison In 2005 the Indian government rolled out a national program (“Rajiv Gandhi Grameen Vidyu- tikaran Yonana” (RGGVY)) intended to electrify those villages that remained un-electrified or were “under-electrified” (see Burlig and Preonas (2016) for more detail). The program had a sim- 62 ilar financing structure as MNREGA – funding came from the Federal government but projects were implemented at a local level. Crucially, funding was dispersed under two different five year plans, the 10th and the 11th and not all districts receive funding under both. In order to receive funding a State had to submit a district specific proposal to the Rural Electrification Corporation (REC), overseen by the Ministry of Power. Proposals were reviewed by the REC and funds were disbursed by them on approval. Submitting a proposal was a costly act by the state, requiring surveys, and a detailed village- by-village implementation plan indicating which households and public places were to be electrified (Burlig and Preonas, 2016). Performing this costly action earlier or faster may be an indicator of a government’s ability to effectively implement programs. Under that assumption, districts that received funding in the 10th Five Year plan may be more effective at implementing government programs than those that did not. However, this assumption may not be valid. It may have been the case that lower capacity or less developed districts were specifically targeted by the REC for assistance in putting together their applications in order to facilitate participation in the 10th five year plan. In that case, participation in the earlier round of funding may indicate lower government capacity. There are a number of reasons district government capacity might impact the fire response to MNREGA. One is simply that if higher capacity districts can better implement MNREGA the labor market shock may be larger (Imbert and Papp, 2015). I divide districts into those that receive funding under RGGVY in the 10th Five Year Plan and the 11th Five Year Plan (districts that receive funding in both are included in the 10th ). I show a map of these districts in A8. I then run the primary specification described in equation 1 of the main text on each sub-sample. The results in Table A3 indicate no difference in the impact of MNREGA on fires between the early implementing and late RGGVY implementing districts. This is consistent with what Burlig and Preonas (2016) find in measuring the direct impacts of the program.36 I also find no evidence of a direct effect of the RGGVY program on fires. This is not evidence that government capacity did not impact the implementation of MNREGA - in part due to the lack of certainty around how district participation in each phase of the RGGVY program was determined in practice - but it is reassuring that I do not find substantial differences in the impact of MNREGA on fires based on the timing of an unrelated government program. 36 They find no differential effects by early or late. However, while they find very strong increases in the rate of electrification and use of electricity they find negligible effects on several measures of income and economic activity. 63 Table A3: Effect of NREGA on by RGGVY Phase Cropland Fires Post-NREGA, RGGVY Phase 1 0.214∗∗∗ (0.070) Post-NREGA, RGGVY Non-phase 1 0.213∗∗∗ (0.053) Districts 558 Months 144 N 80,352 District FE X Year × Month FE X Weather Controls X Notes: Each column represents separate regressions. In all columns the outcome is monthly cropland fires. In all columns the base regression is a fixed effects Poisson of the form yit = βP ost + ψ [P ost × N REGA × RGGV Yi ] + γi + δt where yit is the outcome in district i in month t. Post is a dummy variable equal to one after NREGA treatment takes effect in a given phase and NREGA is a dummy indicating the NREGA phase of district i. RGGVY is an indicator for whether district i was in the first phase of the RGGVY program. γi are district fixed effects while δt is a year by month fixed effect. In all cases N refers to the number of district × months included in each regression. The sample is a balanced, monthly panel of districts in India from 2003 to 2012. Heteroskedasticity robust standard errors clustered at the district level are in parentheses. (* p<.10 ** p<.05 *** p<.01). 64 Figure A8: RGGVY phase 1 districts Phases Notes: Districts that were included in the first phase of the RGGVY electrification program are shown here. The first phase districts are those that received funding from the program in the 10th Five Year plan. Districts that receive funding in both phases are included in the first phase in this Figure. The Figure is based on data from Burlig and Preonas (2016). 65 Appendix 3.4 MNREGA’s impact on other types of fires One might be concerned that there has been improvements in NASA’s algorithms used to detect fires over time that has led to a secular increase in the number of fires in the data and this is correlated with the timing of the implementation of MNREGA, leading to the results I document. While the time fixed effects I use in the primary specification ought to account for any such trend, I can also examine how non-agricultural fires change after the implementation of MNREGA. To do so I use the same measure of fires from the MODIS platform combined with the Copernicus land use data but I focus on fires that occur on land uses that are non-agricultural. Specifically, I focus on fires that occur in forested areas, shrubland areas, and on forest plantations. I find no statistically significant changes in fires after the implementation of MNREGA in any of these areas (Figure A9). In both forested lands and shrublands the point estimates suggest a decline in fires while in plantation fires there may have been a slight increase. This set of results suggests that my main results are not due to general trends in the frequency of fires or changes in the ability of the MODIS platform to detect fires over time that are correlated with the implementation of MNREGA. 66 Figure A9: MNREGA’s impact on non-agricultural fires Forest Fires Shrubland Fires Plantation Fires 67 Notes: In the main text I focus on the impact of MNREGA on agricultural fires, identifying agricultural fires as fires measured by MODIS that occur in an area that the Copernicus land-cover data indicates is agricultural. Using the same approach I can examine fires that occur on other types of land-use in the Copernicus data. I estimate the impact of MNREGA on fires that occur in forested land, shrubland, and on forest plantations using the same specifiation as described in the main text. The only difference is the outcome variable. In Figure A20 I show the average level of night lights appears to decline initially and then recover after the implementation of MNREGA. This is consistent with the findings in Cook and Shah (2019) where they find that night lights decline after MNREGA implementation in phase 1 and 2 districts but to increase in phase 3 districts. If one is concerned that my measure of fires is simply identifying areas with more night lights the decline in night lights after the implementation of MNREGA suggests that measurement error of this type cannot explain my results. I also show in Figure A10 that there does not appear to be a strong relationship between districts with high night lights and high fires. Figure A10: Scatter of districts by fires and total night light luminosity Notes: The scatter of monthly fires and the total annual luminosity of a district as reported in Almås et al. (2019). Appendix 3.5 Changing timing of harvest I find some evidence that the timing of the kharif harvest shifts over time from October to November (Figure A11). Some have argued that shifting the kharif harvest to later in the year is the primary cause of the increased pollution from crop fires in Delhi in December in January. This occurs because of a shift in wind patterns in early December that cause more pollution to be blown from Punjab to Delhi. I cannot rule this out as a potential cause of the increase in pollution in Delhi. However, the major increase in fires occurs in the post-rabi harvest. This can clearly be seen in the monthly patterns. The timing of fires shifts during the kharif harvest but there is no evidence that the overall number of fires declines. Further, as Table 3 shows, MNREGA substantially increases 68 pollution within districts. Thus, while the increase in Delhi’s pollution is likely due to many causes there is still clear evidence that MNREGA increases the frequency of agricultural fires. Figure A11: Monthly pattern of fire use over time Notes: The monthly pattern of fires averaged over all districts by each year in the sample. The count of fires by month by district is averaged over within each year. Lighter lines indicate earlier years. Data comes from the NASA FIRMS database. Appendix 3.6 Placebo tests To examine whether the results I report in the main text are due to underlying differences in districts that are correlated with the implementation timing of MNREGA by chance I run a placebo test where I maintain the order of treatment but move it forward in time by two years for all districts. As a result, phase 1 districts are treated in 2004, phase 2 in 2005 and phase 3 in 2006. As I report in Table A4 I find no treatment effect in this placebo treatment. That suggests that my results are not driven by differences in districts across phases that are correlated with the treatment timing. 69 Table A4: Effect of MNREGA on Fires (Placebo 1) Cropland Fires Post MNREGA 0.032 0.018 (0.048) (0.052) Districts 558 558 Months 144 144 N 80,352 80,352 District FE X X Year × Month FE X X Weather Controls X Notes: Each column represents seperate regressions. In all columns the outcome is monthly cropland fires. In all columns the base regression is a fixed effects poisson of the form yit = βP ost + ψ [P ost × N REGA]+ γi + δt where yit is the outcome in district i in month t. Post is a dummy variable equal to one after NREGA treatment takes effect in a given phase and MNREGA is a dummy indicating the MNREGA phase of district i. γi are district fixed effects while δt is a year by month fixed effect. In columns 2 I include controls for the monthly average cloud cover, precipitation and minimum and maximum temperature in district i in month t. N refers to the number of district × months included in each regression. The sample is a balanced, monthly panel of districts in India from 2003 to 2012. Heteroskedasticity robust standard errors clustered at the district level are in parentheses. (* p<.10 ** p<.05 *** p<.01). Appendix 3.7 Randomization test As a further test of the robustness of the main results I conduct something similar to a ran- domization inference test on the full, national implementation of MNREGA. I keep the timing of implementation the same (i.e. phase 1 districts receive MNREGA in February 2006) but I randomly assign districts to phases. This imposes a null hypothesis of no effect of program implementation. I then run the primary specification, with weather controls, 1,000 times and plot the distribution of the estimated impact of MNREGA implementation in Figure A12. Comparing the distribution of estimated effects under the null of no effect to the effect I estimate in Table 3 in the main text (shown in the dashed red line in Figure A12) shows that it is highly unlikely the effect I estimate is due to chance assignment of districts to phases. The implied p-value on my estimated effect from this randomization exercise is < 0.001. 70 Figure A12: Distribution of estimated impact coefficient with random MNREGA assignment Notes: The distribution of the β coefficient from the primary specification in the paper under the random assignment of MNREGA phase to districts. The dashed vertical line is the actual estimated coefficient in the paper. I randomly assign districts to be in phase 1, 2, or 3 of MNREGA and re-estimate the primary specification, with weather controls, 1,000 times to get the distribution under random assignment. This is in the spirit of a randomization inference exercise. I can reject the null of no effect with a p< 0.001. 71 Appendix 3.8 Heterogeneity by mechanization index components My index of mechanization is the sum of a district’s Z score across several different measures of how easy it may be to mechanize the harvest in a given district. These are all measured over the pre-MNREGA period from 2003-2005 and are: 1. The share of agricultural land in holdings larger than 4 hectares. I consider this because the efficiency of harvesting by combines increases as the area of land harvested increases. As Clemens et al. (2018) shows, mechanization is more efficient at lower levels of labor per unit land. Larger farms have more available land to spread the cost of operating the combine over. 2. The share of land in marginal holdings. This is not mechanically determined by the share of land in large holdings because there are several size classes in between marginal and large. Increasing the share of land in marginal holdings, holding the share in large holdings constant reduces mechanization. This happens for two reasons. The first is simply the inverse of the reason mechanization is more frequently used on larger holdings; combines are less efficient on smaller plots with higher levels of labor per unit land. There is an additional reason why mechanization occurs less used less on marginal land however. MNREGA allows marginal farmers to use labor from MNREGA on projects on their private land (GOI, 2009). While marginal farmers cannot use MNREGA labor on the harvest, to the extent that money is fungible, using MNREGA labor on projects that farmers would have otherwise paid for them- selves frees money to pay for harvest labor.Because more land in marginal plots should reduce mechanization I invert the Z score by multiplying by -1 when I calculate the mechanization index. 3. The number of combines in the district in 2005. Combines are the unit of capital most directly related to mechanization of the harvest and their use is a primary reason for the increase in the use of fires (Yang et al., 2008; Bhuvaneshwari et al., 2019; Shyamsundar et al., 2019). Using combines to harvest leaves more residue on the field than harvesting by hand and it is this increased residue that interferes with the next season’s planting. Farmers often do not own their own combines but rather rent a combine and operator’s time for a specific harvest (Shyamsundar et al., 2019). Having more combines at implementation of MNREGA facilitates mechanization, all else equal, by reducing congestion in this rental market. 4. The area planted in rice and wheat. As discussed above, the use of fires is most intense in areas of coupled rice-wheat production. To account for this I measure the average annual area planted in rice and wheat. 5. The area planted in sugarcane. Farmers who grow primarily sugarcane do not have the ability to easily mechanize in India (Yadav, 2007; Solomon, 2016). Areas that have more land planted in sugarcane should, as a result, see less mechanization than other areas. To account for this I invert the Z score for sugarcane area by multiplying by -1 before I calculate the mechanization index. In Table A5 and Figure A13 I show that across each individual component of the mechanization index the use of fires increases in areas that the component would predict are easier to mechanize. For several of the components the effect seems to begin after the median, as opposed to only in the top quartile as with the overall mechanization index, but others follow the same pattern as the overall mechanization index. The only exception is the area planted in sugarcane. However, this is consistent with the distribution of districts by the area planted in sugarcane. All districts except 72 the first quartile, the areas with the most sugarcane production, have essentially no area planted in sugarcane. As a result, the large increases in fires outside in the second through fourth quartile are consistent with farmers in sugarcane producing areas not being able to mechanize easily while those in non-sugarcane producing areas are more able to mechanize the harvest. Figure A13: Heteogeneity in MNREGA impact by mechanization index components This figure replicates the point estimates presented in Table A5. Light bars report estimates from the first quartile. Dark bars from the fourth. All estimates are from the full MNREGA sample. 73 Table A5: Heterogeneity of treatment impact by components of mechanization index Share of Inverted share Area in Inverse area land >4HA of marginal land Combines Wheat & Rice sugarcane (A)Quartile 1 Post-MNREGA 0.166 -0.074 0.111 0.126 0.008 (0.150) (0.178) (0.091) (0.084) (0.070) Districts 100 100 243 122 123 Months 144 144 144 144 144 N 14,400 14,400 34,992 17,568 17,712 Avg. monthly fires pre-MNREGA 1.73 1.81 2.46 1.93 6.42 (B)Quartile 2 Post-MNREGA -0.044 -0.014 0.111 0.119 0.389∗∗∗ (0.111) (0.107) (0.091) (0.080) (0.120) Districts 100 100 243 123 123 Months 144 144 144 144 144 N 14,400 14,400 34,992 17,712 17,712 Avg. monthly fires pre-MNREGA 4.21 3.58 2.46 3.57 10.21 (C)Quartile 3 Post-MNREGA 0.212 0.351∗∗∗ 0.028 0.209∗∗ 0.333∗∗∗ (0.142) (0.132) (0.120) (0.106) (0.103) Districts 100 100 103 123 123 Months 144 144 144 144 144 N 14,400 14,400 14,832 17,712 17,712 Avg. monthly fires pre-MNREGA 3.57 3.98 4.15 3.86 4.74 (D)Quartile 4 Post-MNREGA 0.289∗∗∗ 0.298∗∗∗ 0.264∗∗∗ 0.176∗∗ 0.266∗ (0.101) (0.102) (0.090) (0.078) (0.154) Districts 100 100 116 122 121 Months 144 144 144 144 144 N 14,400 14,400 16,704 17,568 17,424 Avg. monthly fires pre-MNREGA 17.08 17.22 14.73 14.68 2.57 District FE X X X X X Month × Year FE X X X X X Weather Controls X X X X X Notes: The outcome is monthly cropland fires. The coefficient can be interpreted as the approximate percentage change in fires after MNREGA was statutorily implemented in a district. ( ∑ in[India that were part [ The sample] is all districts ] of the MNREGA ) program. The specification is a fixed effects Poisson of the form E Fimy |Ximy = exp β 4 z =1 P ostimy × M echiz + γi + δmy where Fimy is the outcome in district i in month m in year y . Postimy is a dummy variable equal to one after MNREGA treatment takes effect in district i. Mechiz is an indicator for where district i falls in the distribution of the ease of mechanization index. The ease of mechanization index is the sum of a district’s Z score across measures of land concentration, combine presence and crop types. The mechanization index is calculated based on levels of each component variable in the district prior to 2006. γi are district fixed effects while δmy is a year by month fixed effect. γi are district fixed effects while δt is a year by month fixed effect. Each panel is a different quartile of the mechanization index with quartile 4 corresponding to the places mechanization is predicted to be easiest. N refers to the number of district × months included in each regression. Districts reports districts in each sample. The average number of monthly fires (the outcome) in the pre-treatment period in each quartile are reported. The sample is a balanced, monthly panel of districts in India from 74 2003 to 2014. All columns include controls for weather in the month the outcome number of fires is measured. Heteroskedasticity robust standard errors clustered at the district level are in parentheses. (* p<.10 ** p<.05 *** p<.01). Appendix 3.9 Plot size and fire relationship robustness I show in the main text in Figure 1 that in the cross-section districts that have a larger share of agricultural land in large plots there are more frequent agricultural fires on average. Given the large pixel size of MODIS one might think this is simply due to the increased likelihood that MODIS detects a fire on a large plot because fires are likely to be larger as a result of the larger plot size. I can test this prediction by comparing the relationship between share of agricultural land in large plots and agricultural fires detected by MODIS and agricultural fires detected by VIIRs. If the relationship is driven only by the lower resolution of MODIS then the relationship between size and fires should be weaker when I use fires detected by VIIRs. Figures A14a and A14b show that the relationship appears to be stronger when I use the fires detected by VIIRs. This suggests that it is not being driven by areas with larger plots having larger fires that are easier for MODIS to detect. Figure A14: Comparison of plot size and fires relationship by satellite (a) MODIS (b) VIIRS Notes: Panel a shows the relationship between the number of monthly fires in a district and the share of farmland in that district in plots greater than 4 hectares when the fires were detected with the MODIS platform. Panel b shows the same but when the fires were detected by the VIIRS platform. In both Figures the sample period is 2012-2017 and the sample covers all districts in India. The VIIRS platform can detect fires up to 10x smaller than the MODIS platform (Zhang and Wooster, 2016). Appendix 3.10 Event studies for non-fire outcomes Figures A15-A18 show the event studies of MNREGA’s impact on the area planted in rice, wheat, sugarcane and all other crops in 000s of HA. Data comes from the ICRISAT meso dataset (Rao et al., 2012). In all cases I confirm that the assumption of no differential pre-trends appears to hold. Further, each shows little to no evidence of an increase in area planted after the implementation of MNREGA. Consistent with the results in Table 5 rice shows a small increase after MNREGA’s implementation but this decays quickly and is never statistically different from zero. It is distinctly different from the impact of MNREGA on fires that shows initial increases that persist over time. Figure A16 suggests MNREGA had no impact on area planted in wheat while Figure A17 shows weak evidence that sugarcane production declines after the implementation of MNREGA. A decline in sugarcane production is consistent with both anecdotal evidence (FLA, 2012) and the model presented in the main text. That model suggests that facing higher labor costs farmers can either 75 reduce production or mechanize. However, the technology to mechanize the harvest of sugarcane in India is not widespread (Yadav, 2007; Solomon, 2016). That suggests that mechanization is not as feasible in the short-term, which leads farmers to reduce production in the face of higher labor costs (Clemens et al., 2018). The area planted in other crops (Figure A18) follows a similar pattern to rice production, showing a small initial increase that decays. This is consistent with farmers shifting into other, higher value crops (e.g. cotton) after the implementation of MNREGA (Rabotyagov et al., 2014; Gehrke, 2013). Figure A15: Area planted in rice event study ( : Each point Notes ) is the estimated ωτ coefficient from the regression [ ] ∑ log E Cimy |Ximy = τ ∈T ωτ Yτ i + Wimy + ψi + δmy , where Yτ is an indicator for event-time year τ in the set T = {−3, −2, −1, 0, 1, 2, 3, 4}, ψi is a district fixed effect , δmy is a month × year fixed effect and Wimy are weather controls. Cimy is the area planted in rice in 000s of hectares in month m in year y in district i. 95% CIs are show in dashed grey lines. The Figure uses the full sample. I pool event years less than -3 and greater than 4 into those boundary values. The base year is the year before MNREGA is implemented. In Figure A19 I show the event study of MNREGA’s impact on monthly agricultural fires in the first and fourth quartiles of the mechanization index. This corresponds to the regression results in panel A and panel D of Table 4 in the main text. The event study confirms that there is no 76 evidence of pre-trends within the two quartiles. Further, it shows the same increase in fires in the districts with the highest score in the mechanization index that are reported in the Table 4. 77 Figure A16: Area planted in wheat event study ( : Each point Notes ) is the estimated ωτ coefficient from the regression [ ] ∑ log E Cimy |Ximy = τ ∈T ωτ Yτ i + Wimy + ψi + δmy , where Yτ is an indicator for event-time year τ in the set T = {−3, −2, −1, 0, 1, 2, 3, 4}, ψi is a district fixed effect , δmy is a month × year fixed effect and Wimy are weather controls. Cimy is the area planted in wheat in 000s of hectares in month m in year y in district i. 95% CIs are show in dashed grey lines. The Figure uses the full sample. I pool event years less than -3 and greater than 4 into those boundary values. The base year is the year before MNREGA is implemented. 78 Figure A17: Area planted in sugarcane event study Notes: Each point ( ) the regression is the estimated ωτ coefficient from [ ] ∑ E Cimy |Ximy = exp τ ∈T ωτ Yτ i + Wimy + ψi + δmy , where Yτ is an indicator for event-time year τ in the set T = {−3, −2, −1, 0, 1, 2, 3, 4}, ψi is a district fixed effect , δmy is a month × year fixed effect and Wimy are weather controls. Cimy is the area planted in sugarcane in 000s of hectares in month m in year y in district i. 95% CIs are show in dashed grey lines. The Figure uses the full sample. I pool event years less than -3 and greater than 4 into those boundary values. The base year is the year before MNREGA is implemented. 79 Figure A18: Area planted in other crops event study ( : Each point Notes ) is the estimated ωτ coefficient from the regression [ ] ∑ log E Cimy |Ximy = τ ∈T ωτ Yτ i + Wimy + ψi + δmy , where Yτ is an indicator for event-time year τ in the set T = {−3, −2, −1, 0, 1, 2, 3, 4}, ψi is a district fixed effect , δmy is a month × year fixed effect and Wimy are weather controls. Cimy is the area planted in all other crops in 000s of hectares in month m in year y in district i. 95% CIs are show in dashed grey lines. The Figure uses the full sample. I pool event years less than -3 and greater than 4 into those boundary values. The base year before is the year MNREGA is implemented. 80 Figure A19: Mechanization index event study ( : Each point Notes ) is the estimated ωτ coefficient from the regression [ ] ∑ log E Fimy |Ximy = τ ∈T ωτ Yτ i + Wimy + ψi + δmy , where Yτ is an indicator for event-time year τ in the set T = {−3, −2, −1, 0, 1, 2, 3, 4}, ψi is a district fixed effect , δmy is a month × year fixed effect and Wimy are weather controls. Fimy is number of agricultural fires in month m in year y in district i. 95% CIs are show in dashed grey lines. I run the regression separately on districts in the first quartile of the mechanization index, areas where mechanization is predicted to be more difficult, and the fourth quartile, areas where mechanization is predicted to be easier, to generate each line. The Figure uses the full sample. I pool event years less than -3 and greater than 4 into those boundary values. The base year is the year before MNREGA is implemented. 81 Figure A20: MNREGA’s impact on average night lights event study Notes: ∑ Each point is the estimated ωτ coefficient from the regression Limy = τ ∈T ωτ Yτ i + ψi + δy , where Yτ is an indicator for event-time year τ in the set T = {−3, −2, −1, 0, 1, 2, 3, 4}, ψi is a district fixed effect , δy is a year fixed effect. Liy is average of the night lights in year y in district i. 95% CIs are show in dashed grey lines. Data on night lights comes from (Asher et al., 2019). The Figure uses the full sample. I pool event years less than -3 and greater than 4 into those boundary values. The base year is the year before MNREGA is implemented. 82 Appendix 3.11 Effect of MNREGA on crop production by mechanization index and pre-MNREGA fires To verify that the lack of a mean effect on cropping levels is not masking heterogeneity across districts in cropping responses that is correlated with the mechanization index, and so allowing for changes in cropping levels to drive the results shown in Table 4 in the main text, I show that the impact of MNREGA on cropping levels does not vary by the mechanization index in Table A6. Across all crops and all levels of the mechanization index the results indicate that MNREGA has little impact on cropping levels, consistent with the results in Table 5 in the main text. There appear to be slight increases in production and area planted of rice in the second quartile of the mechanization index and slight increases in area planted in the fourth quartile. Sugarcane shows a slight increase in production in the second quartile as well. However, none of these changes can explain the pattern of results in the main text where the increase in fires appears to be strongly concentrated in the districts with the highest mechanization index score. 83 Table A6: Effect of MNREGA on Crop Production by Mechanization Index Wheat Rice Sugarcane All Crops Area Tons Area Tons Area Tons Area Tons (A) Mech. Index, Q1 Post-MNREGA 0.001 0.007 0.017 0.055 -0.009 -0.246∗ 0.003 -0.032 (0.012) (0.017) (0.019) (0.033) (0.033) (0.129) (0.011) (0.038) Districts Months 120 120 120 120 120 120 120 120 N 13,224 12,384 16,860 16,860 15,972 15,972 16,860 16,860 Mean (B) Mech. Index, Q2 Post-MNREGA 0.011 0.054 0.070∗ 0.097∗ 0.093 0.204∗∗ 0.005 0.122 (0.056) (0.077) (0.037) (0.053) (0.066) (0.096) (0.021) (0.081) Districts Months 111 111 111 111 111 111 111 111 N 8,880 8,760 8,976 8,976 8,112 8,112 9,240 9,240 Mean (C) Mech. Index, Q3 Post-MNREGA -0.031 -0.012 0.006 0.058 0.077 -0.283 0.019 0.103∗∗ (0.042) (0.065) (0.020) (0.043) (0.088) (0.182) (0.015) (0.044) Districts Months 108 108 108 108 108 108 108 108 N 12,456 12,096 13,620 13,620 11,988 11,976 13,980 13,980 Mean (D) Mech. Index, Q4 Post-MNREGA -0.044 -0.058 0.023∗ -0.004 0.056 0.024 0.008 -0.033 (0.032) (0.041) (0.012) (0.023) (0.061) (0.100) (0.010) (0.027) Districts Months 116 116 116 116 116 116 116 116 N 15,828 15,708 14,724 14,724 15,192 15,072 16,296 16,296 Mean District FE X X X X X X X X Year × Month FE X X X X X X X X Weather Controls X X X X X X X X Notes: Each column represents separate regressions. Outcomes for each column are listed in the column headings. Area is measured as 000s of HA and Tons area planted in ( ) measures annual production in Tons. In all columns the base regression is a poisson fixed effects [ ] ∑ [ ] of the form log E Cimy |Ximy = β 4 z =1 P ostimy × M echiz + γi + δmy where Cimy is the outcome in district i in month m in year y . Postimy is a dummy variable equal to one after MNREGA treatment takes effect in district i. Mechiz is an indicator for where district i falls in the distribution of the ease of mechanization index. The ease of mechanization index is the sum of a district’s Z score across measures of land concentration, combine presence and crop types. The mechanization index is calculated based on levels of each component variable in the district prior to 2006. γi are district fixed effects while δt is a year by month fixed effect. The Ease of Mechanization Index is an index that considers, in MNREGA districts, the type of crops planted, the average area of holdings and the number of combines prior to treatment in a given district. Areas that have larger farms, plant more wheat and/or rice, and have more pre-treatment combines are given higher scores. In the Andhra Pradesh subdistricts the index omits combines for lack of data. N refers to the number of district × months included in each regression. The sample is a balanced, monthly panel of districts in India from 2003 to 2014. Heteroskedasticity robust standard errors clustered at the district level are in parentheses. (* p<.10 ** p<.05 *** p<.01). 84 Appendix 4 A model of mechanization driven by wage changes The setting in which I examine the impact of income growth on pollution is the Indian agricultural sector, but the model I present here is more general. It could apply broadly to any setting in which firms face a choice among multiple production technologies, some more labor intensive than others, and where the labor saving production technologies produce at least weakly more pollution externalities than the labor intensive technology. I start by outlining the general model and then discuss the predictions it makes in my specific context of Indian agriculture.37 Farmers produce crops that they sell at fixed prices, normalized to one. Each farmer has a fixed quantity of land A ≥ 1 with fixed productivity. They face a choice between production technologies they can use to produce crop Y . One is a labor intensive production technology G(L) and one is a capital intensive technology F (L). F relies on mechanization to produce output so also includes a pollution externality. There is no pollution externality associated with G. Farmers do not consider the social costs of F and G but I assume that F imposes a cost of emissions e > 0 on society. e is monotonically increasing in the number of farmers who employ F . Workers earn wages w and capital is purchased at a fixed cost. Farmers are price takers in both labor and capital markets. ′ ′ ′′ ′′ I assume that F (L) > G(L) for all L, 0 < F (L) < G (L), and F (L) < 0, G ( (L ) ) < 0. ′ −1 ′ Further, F (x) < G −1 (x). Farmers face the associated profit conditions π G = AG L − wL ( ) and π F = AF L − K − wL and jointly choose a production technology and optimal level of labor L to maximize profits. Farmers will adopt F (L∗∗ ) when the increase in profits exceeds the cost of switching to capital intensive production (K ). The gap in profits between F (L∗∗ ) and G(L∗ ) is larger when farmers have more land, which leads to the following proposition: Proposition 1. There exists an A ˆ then π F = π G , iff A > A ˆ > 0 such that iff A = A ˆ then π F > π G ˆ ˆ −wL∗ +wL∗∗ and farmers choose F , and iff A < A then π < π and farmers choose G. A = K F G F (L∗∗ )−G(L∗ ) ˆ is increasing in K and decreasing in w. Further, e is decreasing in A such that A ˆ. For proof see Appendix 4.1. MNREGA acts as a shock to the agricultural labor market that makes it more difficult for farmers to hire labor. As a result, MNREGA raises the equilibrium wage in the agricultural labor market (Imbert and Papp, 2015). Farmers face the same choice between G and F but the increase in w implies that A ˆP ost−M N REGA < A ˆP re−M N REGA and so the number of farmers using F increases. Finally, as an extension, consider a third production technology J C where C denotes that this is a clean technology. Specifically, eG ≤ eC < eF so that emissions associated with the sustainable technology are lower than with the new technology that farmers adopt. This could be because the new technology does not require the removal of residue (i.e. no till) or mechanizes the removal so that burning is not necessary (Shyamsundar et al., 2019). However, consistent with the experience of Indian farmers, K C > K such that the low emissions technology is more expensive to adopt than the existing labor saving technology. This highlights that without Pigouvian policy setting a price τ on the emissions associated with the existing labor saving technology (so that the farmer faces a choice of K C versus K + τ ) the adoption of J C will be below the social optimum. 37 I draw heavily on both Hornbeck and Naidu (2014) and Clemens et al. (2018). Hornbeck and Naidu (2014) show an increase in out migration of relatively unskilled labor leads to an increase in the use of labor-saving capital in response. Clemens et al. (2018) show how farmers rapidly mechanize agricultural production in crops with a readily available mechanization technology in response to a change in U.S. immigration policy that substantially reduces the supply of migrant labor from Mexico. 85 Heterogeneity in fire response Assume A is distributed such that A = AD + µ where µ has a distribution H (µ) that is constant across space and the median A, AD , varies by district. Then Proposition 2. The level of pollution e in a district is increasing with AD if A ˆ is everywhere greater than AD and µ has a constant mean zero distribution that is single-peaked at 0. For any ′ ˆ will be greater in D′ two districts D and D the increase in pollution for an ϵ shift downwards in A if AD < AD′ . For proof see Appendix 4.1. This implies that for a given wage shock some districts may see larger changes in fires. In particular, those districts where the AD is higher prior to MNREGA will see larger increases. Appendix 4.1 Model proofs Proposition 1: By assumption: 1. G(L) and F (L) are continuous 2. F (L) > G(L) for all L ′ ′ 3. 0 < F (L) < G (L) ′′ ′′ 4. F (L) < 0, G (L) < 0 ′ −1 ′ 5. F (x) < G −1 (x) ( ) ( ) [ ] Farmers face the adoption condition AF L − wL − AG L − wL = K . Solving for A and ˆ in the text: substituting for farmers optimal labor choices in F and G yields the equation for A ∗ ∗∗ ˆ = K − wL + wL A F (L∗∗ ) − G(L∗ ) The assumptions above imply F (L∗∗ ) − G(L∗ ) > 0 and that wL∗∗ − wL∗ < 0 so it is clear that A ˆ is increasing in K and decreasing in w. ˆ is unique note that: To see that A 1. ∂π F /∂A > ∂π G /∂A and ∂ 2 π F /∂A∂A = ∂ 2 π G /∂A∂A 2. K is weakly positive so that there exists some A where π F ≤ π G . ˆ that defines the point at which farmers Then by the intermediate value theorem there exists a A are indifferent between F (L) and G(L) and it is unique. That also implies, because e is monotoni- cally increasing in the number of farmers who choose F (L), that e is decreasing in Aˆ. Proposition 2: By assumption: 1. A = AD + µ 86 2. µ has a constant single-peaked distribution H (µ) with mean zero ˆ is everywhere greater than AD 3. A Since the level of pollution is increasing in the number of farmers above A ˆ the level of pollution e in any given district D will be a function of the distribution of µ and A such that e ≡ e(1 − H [A ˆ ˆ]). As AD approaches A from below for a constant distribution of µ the area 1 − H (A) is strictly ˆ ˆ increasing. Further, I’ve shown that A ˆ is declining in w. Assume some increase in w that results in an ϵ shift down in A. Then for a AD < AD′ , for a single peaked distribution, H (1 − [A ˆ ˆ − ϵ], AD ) − H (1 − A,ˆ AD ) < H (1 − [Aˆ − ϵ], A ′ ) − H (1 − A, D ˆ A ′ ) so the change in pollution moving from A D ˆ to Aˆ − ϵ is increasing in AD . Appendix 5 Predictive power of the mechanization index The goal of the mechanization index is to identify areas in which mechanization may be a more viable option for farmers. While I cannot observe the direct impact of MNREGA on mechanization because of data limitations I do observe combine counts in 2011, five years after the first imple- mentation of MNREGA. As a result, I can test the ability of the mechanization index to predict the number of combines five years after MNREGA’s implementation. In Figure A21 I show that the mechanization index appears to predict combine levels in 2011 reasonably well. I show the binscatter of the district level count of combines in 2011 and the mechanization index score calculated based on data from 2003 to 2005. Data on combines comes from the agricultural input survey. The mechanization index is calculated as described in the main text. I plot the linear best fit line using all the data in the dashed light blue line. Because there is a clear outlier in the binscatter I also show the linear best fit line excluding that data in the darker dotted line. Note that the excluded data is not a single district but rather the binned data for the approximately 30 districts with the highest mechanization index scores and the highest level of combines in 2011. 87 Figure A21: Ability of mechanization index to predict combine levels in 2011 Notes: The binscatter of mechanization index against the average number of combines by district in 2011. Higher values of the mechanization index indicate that mechanization was expected to be easier in that district. The mechanization index is calculated based on data from 2003-2005. The average number of combines by district in 2011 is collected by scraping the Agricultural Input Survey data for 2011. The lighter dashed line is the OLS best fit line including the districts with the highest mechanization score and the greatest number of combines. The darker dotted line is the same OLS best fit line excluding the 30 districts with the highest mechanization index and highest number of combines in 2011. 88 Appendix 6 Replication using MNS data Appendix 6.1 Framework for analyzing the RCT data As an alternative to using the national roll-out of MNREGA to estimate its impact on agricultural fires I can use treatment in an independent RCT that improved MNREGA implementation in a specific state to examine how MNREGA implementation impacted agricultural burning. I provide a brief description of the RCT here. For more details see Muralidharan et al. (2016) and Muralidharan et al. (2017). The RCT conducted by MNS was conducted beginning in 2010 and focused on improving the implementation of MNREGA by improving the provision of biometric smartcards connected to bank accounts that enabled electronic payment of MNREGA wages. Electronic wage payments reduced the opportunity for corruption in the payment process, reduced the time between work and payment, and increased the likelihood that workers received payment for their participation in MNREGA. This in turn increased participation in MNREGA projects and wages received from MNREGA projects. The government of Andhra Pradesh (GoAP) began the initial smartcard program in 2006 with the beginning of MNREGA. However, early implementation was heterogeneous because different banks were used to implement the program in different districts. In 2010 the GoAP restarted the program in eight districts in which initial implementation had been particularly poor. MNS were able to randomize the timing with which subdistricts in these eight districts received the new program. Specifically, 112 of the subdistricts were assigned to a treatment group, 45 to a control group, and the remaining 139 to a buffer group. The treated subdistricts received the program beginning in June 2010 and there was a two year lag between implementation in the treated and control subdistricts. Baseline surveys were conducted in the treated and control subdistricts prior to implementation and endline surveys were conducted in treated subdistricts prior to implementation in control subdistricts. Appendix Figure A22 shows the treatment pattern from Muralidharan et al. (2016). Treated subdistricts received treatment following the completion of baseline surveys in June of 2010. Buffer subdistricts received treatment following treated subdistricts and meant that there was at least a two year gap between the treatment of the treated and control subdistricts. MNS show balance across the treated and control subdistricts on their outcomes variables as well as a range of baseline socioeconomic characteristics. In Table A7 I replicate the balance table from MNS showing balance across treated and control subdistricts on socio-economic covariates. Column 3 reports the coefficient from a regression of treatment on the named covariate in a linear regression with state fixed effects. Unfortunately, the balance does not extend to the frequency of agricultural fires. While the difference in pre-treatment fires between treated and control subdistricts is not significant the relative difference is sizable. Control subdistricts have approximately 30% fewer fires than treated subdistricts. In Table A8 I report summary data for the average number of pre-MNREGA fires by subdistrict across all of India and the treatment and control subdistricts in the MNS RCT. I also report pre- MNREGA levels of several covariates that are important in predicting fires. Except for monthly fires and average share of crop coverage all variables are reported at the district level for the All of India sample and at the subdistrict level for the MNS sample due to data limitations. Figure A23 shows the distribution of pre-MNREGA fires in Andhra Pradesh and the MNS sample. I use the same scale as in Figure 1 in the main text so the Figure is exactly analogous to the Figure showing the distribution of pre-MNREGA fires across all of India. Subdistricts that are included in the MNS RCT are outlined, in black for treated subdisricts and blue for control subdistricts. 89 Figure A22: Map of MNS treated and control subdistricts Notes: This map replicates the map of treatment shown in the appendix of Muralidharan et al. (2017). It shows the treated, control and buffer subdistricts of the RCT that improved the im- plementation of MNREGA by providing biomet- ric smart bank cards to participants as described in Muralidharan et al. (2017) and Muralidharan et al. (2016). 90 Table A7: Balance Table Treatment Control Difference p-value (1) (2) (3) (4) (A) Numbers based on official records from GoAP in 2010 % population working .53 .52 .0072 .41 % male .51 .51 .00043 .67 % literate .45 .45 .0034 .72 % SC .19 .18 .0019 .85 % ST .1 .12 -.014 .48 Jobcards per capita .54 .55 -.0074 .72 Pensions per capita .12 .12 .0012 .76 % old age pensions .48 .49 -.013 .09 % weaver pensions .0088 .011 -.0018 .63 % disabled pensions .1 .1 .0019 .59 % widow pensions .21 .2 .013** .04 (B) Numbers based on 2011 census rural totals Population 45697 45600 392 .85 % population under age 6 .11 .11 -.00074 .66 % agricultural laborers .23 .24 -.0048 .61 % female agri. laborers .12 .12 -.0031 .55 % marginal agri. laborers .071 .063 .0078 .16 (C) Numbers based on 2001 census village directory # primary schools per village 2.9 3.2 -.3 .28 % village with medical facility .67 .7 -.032 .41 % villages with tap water .59 .6 -.0033 .94 % villages with banking facility .12 .16 -.034** .024 % villages with paved road access .8 .82 -.01 .78 Avg. village size in acres 3394 3743 -316 .38 Notes: Table A7 reports baseline characteristics across treated and control subdistricts. Column 3 reports the difference in treatment and control means. Column 4 reports the p-value on the treatment indicator from simple regressions of the outcome with district fixed effects as the only controls. The table exactly replicates that found in Muralidharan et al. (2016). They provide the following notes: “A “jobcard” is a household level official enrollment document for the NREGS program. “SC” (“ST”) refers to Scheduled Castes (Tribes), historically discriminated-against sections of the population now accorded special status and affirmative action benefits under the Indian Constitution. “Old age”, “weaver”, “disabled” and “widow” are different eligibility groups within the SSP administration. “Working” is defined as the participation in any economically productive activity with or without compensation, wages or profit. “Main” workers are defined as those who engaged in any economically productive work for more than 183 days in a year. “Marginal” workers are those for whom the period they engaged in economically productive work does not exceed 182 days. The definitions are from the official census documentation. The last set of variables is taken from 2001 census village directory which records information about various facilities within a census village (the census level of observation). “# primary schools per village” and “Avg. village size in acres” are simple mandal averages - while the others are simple percentages - of the respective variable (sampling weights are not needed since all villages within a mandal are used). Note that we did not have this information available for the 2011 census and hence use the 2001 data. Statistical significance is denoted as: *p < 0.10, **p< 0.05, ***p< 0.01.” 91 Figure A23: Pre-MNS Fires Notes: The count of fires by subdistrict in the years prior to implementation of the MNS RCT in Andhra Pradesh. The subdistricts that partic- ipated in the RCT are outlined. Darker areas had more fires. White areas have no fires. Data comes from the NASA FIRMS database. 92 Table A8: Average level of covariates predictive of fires nationally and by MNS treatment status All of India MNS Treated MNS Control Mean SD Max Mean SD Max Mean SD Max Avg. monthly fires 0.57 2.71 70.38 0.21 0.37 2.22 0.15 0.23 1.00 Avg. share planted in rice 0.32 0.27 0.92 0.32 0.26 1.00 0.30 0.29 0.92 Avg. share planted in wheat 0.17 0.18 0.62 0.00 0.00 0.00 0.00 0.00 0.00 Avg. share planted in sugarcane 0.03 0.07 0.57 0.01 0.03 0.21 0.01 0.02 0.13 Combines in 2005(000s)* 1.35 5.45 64.60 0.02 0.04 0.22 0.02 0.03 0.11 Share of holdings >4 HA 0.30 0.20 0.96 0.25 0.09 0.44 0.27 0.10 0.54 Avg. share of crop coverage 22.00 20.13 84.50 10.79 7.59 40.50 10.82 6.35 30.01 Notes: Statistics for each covariate are calculated for years prior to NREGA implementation. The all of India sample includes all districts in India outside of Nicobar and Jammu & Kashmir. The MNS Treated and MNS Control refer to the subdistricts in the RCT conducted in Andhra Pradesh by Muralidharan et al. (2016) (MNS). Average monthly fires and average share of crop coverage are calculated at the subdistrict for both the All of India and MNS samples. All the remaining covariates are measured at the district level for the All of India sample and at the subdistrict level for the MNS sample because of data limitations. The average district in the MNS sample has 19 subdistricts. Data used to calculate the share planted in each crop in the All of India sample comes from the ICRISAT meso data (Rao et al., 2012) and is the average over the years 2003-2005. Data used to calculate the share planted in the MNS sample comes from the Indian Agricultural Census in 2005 and is the level reported for 2005. Data on combines comes from the Indian Agricultural Input Survey in 2005. For the MNS sample I use the number of tractors reported to the Ministry of Statistics of Andhra Pradesh as a proxy for the number of combines. This likely overestimates the number of combines in Andhra Pradesh. Data on the share of holdings >4 HA comes from the Agricultural Census in 2005 for both samples. The share of crop coverage reports the average share of the subdistrict area that is classified as agricultural on October 31st over the years 2003-2005 in the SEDAC Indian Winter Cropping dataset (Jain et al., 2017; NASA, 2017). There are two main takeaways from Figure A23. The first is that the frequency of fires is lower in Andhra Pradesh than in the states of India with the most fires. This confirms the results in Figure 2 that show Andhra Pradesh is near the median state in the number of pre-MNREGA fires. The second is that the subdistricts included in the MNS sample are, broadly, not the same subdistricts in which fires most frequently occur in Andhra Pradesh but in the majority of the MNS sample districts fires are present. Because the assignment of subdistricts to treated and control in the MNS RCT did not result in perfect balance in the frequency of pre-treatment fires across the treated and control subdistricts I use the same difference-in-differences approach in the main text when analyzing the RCT sample. I modify the estimating equation from 1 to be: ( ) log µ(Ximy ) = βTimy + ωi Wimy + δmy + ψi (2) where in the RCT sample I replace Nimy with Timy – an indicator for treatment having occurred in subdistricts i in month m in year y where treatment occurs in the treated subdistricts after the baseline survey in 2010 as in Muralidharan et al. (2016). Further, in the RCT sample ψi becomes a subdistrict, rather than district, fixed effect. Othwerise terms are the same as in the main estimating equation in the text. 93 Figure A24: Average crop coverage on October 31st in Andhra Pradesh Notes: This Figure shows the average share of pixels in a subdistrict across Andhra Pradesh is available that have crop coverage on October 31st over the years 2003-2005. Crop coverage is measured by reflectivity detected by satellite as described in Jain et al. (2017). Data comes from the Center for International Earth Science at Columbia University NASA (2017). 94 Appendix 6.2 Mean impacts using the RCT sample The results in Table A9 show that the improvement in the implementation of MNREGA did not have a large average impact on agricultural fires in the subdistricts of Andhra Pradesh where the experiment was conducted. While the estimates in Table A9 are imprecisely estimated - I cannot rule out an increase in fires of roughly 28% with 95% confidence - the point estimates for both specifications (controlling for weather and not) are close to zero. Table A9: Impact of randomized improvements in NREGA implementation in Andhra Pradesh on monthly fires Cropland Fires [Treated x Post] 0.027 0.018 (0.137) (0.143) Subdistricts 145 145 Months 104 104 N 15,080 15,080 Avg. monthly fires pre-NREGA .2 .2 Subdistrict FE X X Year × Month FE X X Weather Controls X Notes: Each column represents separate regressions. In all columns the outcome is monthly cropland fires. In all columns the coefficient can be interpreted as the approximate percentage change in fires after treatment in the RCT in Muralidharan et al. (2016) (MNS) occurs in a subdistrict. In all columns the base regression is a Poisson of fires in subdistrict i in month m in year y that includes subdistrict fixed effects and year by month fixed effects. In column 2 I include controls for the monthly average cloud cover, precipitation and minimum and maximum temperature in subdistrict i in month, year my.N refers to the number of subdistricts × months included in each regression. The sample is a balanced, monthly panel of subdistricts in the MNS sample from 2003 to 2012. Heteroskedasticity robust standard errors clustered at the district level are in parentheses. (* p<.10 ** p<.05 *** p<.01). There are several explanations for the difference in results. The first is that fires do not appear to have been widely used in Andhra Pradesh prior to the implementation of MNREGA. As Figure 2 showed, the average number of monthly fires prior to MNREGA in Andhra Pradesh is far below the average in states known for using fires. To the extent that there is learning by doing (i.e. I learn how to use fires from my neighbor who uses fires) or protection from legal ramifications of burning (i.e. I can blame sparks from my neighbors’ field for the fire on mine if the authorities attempt to fine me) we may expect areas with higher levels of fires prior to MNREGA to have seen larger effects. Second, as Table 1 shows, the agricultural fires are associated with certain crops and certain agricultural practices. In particular, places that practice coupled rice-wheat production are more likely to use fires and areas that have higher cropping levels in October have more fires (Jain et al., 2014; Bhuvaneshwari et al., 2019; Shyamsundar et al., 2019). Andhra Pradesh has relatively little area in agriculture in October and, on average, plants less area in coupled rice- wheat production than states in which fires are widespread (see Table A8). Further, the level of combines, an important determinant of the share of residue in a given field that is burned (Yang et al., 2008), is lower in Andhra Pradesh than in states with more fires. However, there are some subdistricts in Andhra Pradesh that used fires at rates comparable to 95 the states with the most frequent use of fires. In particular, dividing Andhra Pradesh subdistricts into quartiles based on pre-MNREGA fire use, I observe that subdistricts in the fourth quartile use fires at roughly the same rates as Haryana (Figure A25). Figure A25: Mean monthly agricultural fires by state from 2003-2005 Notes: The mean of the average number of fires within each subdistrict in a state. The mean for Andhra Pradesh, the location of the RCT, are shown with a lighter color bar and highlighted in red. The mean number of fires is calculated by finding the average number of fires by subdistrict across the pre-MNREGA years and taking the mean of these by state. The overall average for Andhra Pradesh, the location of the MNS RCT, is shown as well as the mean for the subdisricts in each quartile of the distribution of the number of pre-treatment fires. These are highlighted in red. The following are omitted from the Figure because their levels are so high including them would make it difficult to see variation in the remaining states: Punjab (Mean: 18.88), Haryana (Mean:3.37), and Andhra Pradesh, Q4 (Mean: 2.61). The similarity in the use of fires at the top of the distribution suggests that there may be some places in Andhra Pradesh where conditions are such that the implementation of MNREGA may have had similar effects as in the national sample. To test this I construct a similar mechanization index for Andhra Pradesh subdistricts as I calculate for the whole country. I cannot construct exactly the same mechanization index for the RCT sample because of a lack of data on inputs by subdistrict. Specifically, I do not have access to data that measures the number of combines in Andhra Pradesh prior to the implementation of MNREGA. This is an important part of the mechanization index because the ability to rent time from a combine operator, a primary method of mechanizing in this context, is facilitated when there are more combines. Despite this, I can construct a measure of mechanization index that uses the number of tractors to proxy for the number of combines. This is likely an overestimate of the number of combines in a given district. Appendix 6.3 Heterogeneity by mechanization index in the MNS sample Examining heterogeneity across my Andhra Pradesh specific mechanization index using the RCT sample confirms the results in the main text. While the results are very noisily estimated, I see the same pattern where the only quartile in which fires increase is that in which mechanization is predicted to be easiest. Despite the noise in the estimates, this is also the only quartile in which 96 the estimate is statistically different from zero at conventional (10%) levels (Table A10 and Figure A26). While the setting in Andhra Pradesh is different from the rest of the country in the use of fires in the baseline – likely due to differences in the use of the coupled rice-wheat production process – I observe similar results in the heterogeneity analysis when I look at how fires changed after the improvement of MNREGA in subdistricts with a high probability of mechanization. These results highlight two features of the primary results. First, the effects are driven by places that use fires as a management tool in a coupled rice-wheat production process. I see no effect in Andhra Pradesh on average, likely due to the fact that coupled rice-wheat production is much less prevalent compared to the areas in which is most common. Second, in those areas of Andhra Pradesh that score highly on the mechanization index, calculated in part by considering how much land is in a coupled rice-wheat production process, I observe large changes in fire use after MNREGA was improved experimentally in those areas that had a higher probability of mechanizing the harvest. On the whole these results are consistent with the results on the impacts of MNREGA, and the proposed mechanism, reported in the main text. 97 Table A10: Heterogeneity of treatment impact by ease of mechanization index Andhra Pradesh (A)Quartile 1 of Ease of Mechanization Index Post-NREGA -0.305 (0.231) Subdistricts T:17 C:8 Months 79 N 2,844 Avg. monthly fires pre-NREGA .35 (B)Quartile 2 of Ease of Mechanization Index Post-NREGA -0.050 (0.217) Subdistricts T:20 C:9 Months 88 N 3,432 Avg. monthly fires pre-NREGA .17 (C)Quartile 3 of Ease of Mechanization Index Post-NREGA -0.637 (0.476) Subdistricts T:22 C:8 Months 86 N 3,010 Avg. monthly fires pre-NREGA .16 (D)Quartile 4 of Ease of Mechanization Index Post-NREGA 0.452∗ (0.264) Subdistricts T:10 C:8 Months 64 N 2,240 Avg. monthly fires pre-NREGA .11 District FE X Month × Year FE X Weather Controls X Notes: The outcome is monthly cropland fires. The coefficient can be interpreted as the approximate percentage change in fires after NREGA was statutorily implemented in a district. The ] subdistricts [ sample is the ( ∑ in[ Andhra Pradesh included in the ] MNS RCT. ) The specification is a fixed effects Poisson of the form E Fimy |Ximy = exp β 4 z =1 P ostimy × T reatedi × M echiz + γi + δmy where Fimy is the outcome in district i in month m in year y . Postimy is a dummy variable equal to one after MNS treatment and Treatedi is a dummy indicating that subdistrict i was among the treated subdistricts. Mechiz is an indicator for where subdistrict i falls in the distribution of the ease of mechanization index within Andhra Pradesh. The ease of mechanization index is the sum of a district’s Z score across measures of land concentration, tractor presence and crop types. Note this is not directly comparable to the Z score in the main text because it lacks a measure of combine presence. The mechanization index is calculated based on levels of each component variable in the subdistrict prior to 2006. γi are district fixed effects while δmy is a year by month fixed effect. γi are subdistrict fixed effects while δt is a year by month fixed effect. Each panel is a different quartile of the mechanization index with quartile 4 corresponding to the places mechanization is predicted to be easiest. N refers to the number of subdistrict × months included in each regression. Subdistricts reports the treated and control subdistricts in each sample. The average number of monthly fires (the outcome) in the pre-treatment period in each quartile are reported. T The samples are a balanced, monthly panel of subdistricts in Andhra Pradesh from 2003 to 2012. All columns include controls for weather in the month the outcome number of fires is measured. Heteroskedasticity robust standard errors clustered at the subdistrict level are in parentheses. (* p<.10 ** p<.05 *** p<.01). 98 Figure A26: Estimated impact of MNREGA in MNS sample by mechanization index Notes: Bars indicate the approximate percentage change in fires after the MNREGA was im- provement in a subdistrict. The sample is the subdistricts in Andhra Pradesh included [ the in ] MNS RCT. Estimates come from a fixed effects Poisson specification of the form E Fimy |Ximy = ( ∑ [ ] ) exp β 4 z =1 P ostimy × T reatedi × M echiz + γi + δmy where Fimy is the outcome in district i in month m in year y . Postimy is a dummy variable equal to one after MNS treatment and Treatedi is a dummy indicating that subdistrict i was among the treated subdistricts. Mechiz is an indicator for where subdistrict i falls in the distribution of the ease of mechanization index within Andhra Pradesh. The ease of mechanization index is the sum of a district’s Z score across measures of land concentration, tractor presence and crop types. Note this is not directly comparable to the Z score in the main text because it lacks a measure of combine presence. The mechanization index is calculated based on levels of each component variable in the subdistrict prior to 2006. γi are district fixed effects while δmy is a year by month fixed effect. γi are subdistrict fixed effects while δt is a year by month fixed effect. 99