August 14, Abstract

Transcription

1 THE IMPACT OF INDIA S GREEN REVOLUTION: AN EMPIRICAL INVESTIGATION OF MODERN AGRICULTURAL DEVELOPMENT JACOB MOSCONA August 14, 2017 Abstract This study exploits rapid technological development during the Green Revolution to estimate the causal effect of agricultural productivity growth during the second half of the 20th century. I use variation in the maximum potential impact of the Green Revolution to construct a measure of predicted agricultural productivity; using predicted productivity as an instrument, I investigate the impact of actual productivity growth across districts in India between I find that agricultural productivity growth increased several proxies of rural well-being and spurred employment and land use in the agricultural sector, albeit a less equal distribution of land ownership. It also reduced urbanization and the size of the manufacturing and service-sector labor force. Additionally, the Green Revolution increased the electoral success of agrarian opposition parties and district-level violent conflict. Using an analogous identification strategy across countries, I find qualitatively identical effects on the sectoral distribution of land and labor, but I find no evidence that the Green Revolution increased national incomes. If anything, the relationship for low-income countries is negative. Greater entrenchment in the agricultural sector is a robust effect of the Green Revolution both at the sub-national and national level. Broadly, these results suggest that a productivity shock may have limited or negative consequences in equilibrium by shifting individuals, resources, or political capital away from potentially more productive activity. JEL Codes: J43, O13, O14, O33, O47, Q10, Q16. Keywords: Agricultural productivity, Green Revolution, economic development, structural change. I especially thank Nathan Nunn for his support and guidance from the project s outset. I am grateful to Daron Acemoglu, Abhijit Banerjee, Esther Duflo, Claudia Goldin, Eduardo Montero, Ben Olken, Jonas Poulsen, Carl Riskin, James A. Robinson, Andrei Shleifer, and members of the Harvard Economic History Workshop and M.I.T. Political Economy Lunch for extremely helpful feedback. Francesca R. Jensenius generously shared her data on state assembly elections in India. Department of Economics, M.I.T., 50 Memorial Drive, Cambridge, M.A , U.S.A. ( moscona@mit.edu; website: 1

2 1 Introduction In recent decades, the battle against hunger was radically transformed by an assault of technological breakthroughs. The second half of the 20th century was a period of rapid and unprecedented growth in agricultural productivity and food production. Global food grain production increased from one to two billion tons between 1960 and 2000 (Khush 2001). The increase in output was driven almost entirely by increases in technological productivity and not by an expansion of land devoted to agriculture (Ball et al. 1997). This period of rapid 20th century agricultural development has been named the Green Revolution. The Green Revolution has been credited with drastically reducing hunger in the developing world and countries ability to support rapidly growing populations (eg. Davis et al. 2010, Pingali 2012), with improved overall well-being and poverty reduction (eg. Rao 1985, Thirtle et al. 2003, Evenson & Gollin 2003), and with overall great...economic benefits (Frankel 1971). At the same time, it has been associated with higher poverty levels (eg. Griffin 1974, Harriss 1977), with lower local industrial development (eg. Foster & Rosenzweig 2004), and with higher levels of political discord and violent conflict (eg. Ladejinsky 1970, Frankel 1971, Shiva 1991). As Clifton Wharton Jr. observed in Foreign Affairs in 1969, just as modern variety crop adoption was first accelerating: On the one hand, some observers now believe that the race between food and population is over, that the new agricultural technology constitutes a cornucopia for the developing world, and that victory is in sight in the War on Hunger. Others see this development as opening a Pandora s box; its very success will produce a number of new problems which are far more subtle and difficult than those faced during the development of the new technology. (Wharton Jr. 1969) Polarizing debates on the impact of the Green Revolution are far from conclusive it is difficult to think of many development projects that have been simultaneously so praised and decried. Yet regardless of the divergent stances of onlookers, the Green Revolution has transformed global agricultural production and this fact is nowhere more obvious than in India. In the Indian state of Punjab, for example, by % of rice cultivation relied on modern varieties (or high yield varieties, henceforth HYVs) developed since the 1960s. Between 1961 and 1981 in Punjab, overall agricultural productivity had grown by 1340 hectograms per hectare (or 138%). 1 If the 1 This is based on my calculation from Sahghi et al. (1998) described below. 1

3 Green Revolution did have major economic and social impacts, we would be most likely to observe them in a region like Punjab. Yet while some regions in India were at the forefront of HYV adoption and leaders in agricultural productivity growth, in other regions HYV adoption was almost impossible (Foster & Rosenzweig 1996, 2004). 2 This combination of factors regions that were ground zero for HYV-led agricultural productivity combined with diverse responses to new agricultural technologies across districts makes India an ideal setting to study the impacts of the Green Revolution. The main contribution of this study is to provide a rigorous empirical analysis of the effects of modern agricultural productivity growth the Green Revolution on sectoral change and economic growth. I propose causal estimates of the impact of the Green Revolution in India and use the Green Revolution as a tool in order to better understand the impact of changes in recent agricultural productivity more broadly. Few existing studies claim to isolate the causal impact of the Green Revolution and causal evidence is especially scarce in contexts where it may be possible to identify equilibrium effects. 3 This lack of causal evidence is troubling, not only because agricultural productivity change is a context in which statistical correlations and anecdotal evidence are likely to be biased and difficult to interpret, 4 but also because the Green Revolution was and has been a major component of global food policy. 2 Mean agricultural productivity growth was 374 hectograms per hectare. 3 Some studies focus on the local level, whereas the main focus of this study is to evaluate effects that emerge in equilibrium. For example, Foster & Rosenzweig (2004) use the village as their unit of observation and their analysis is focused on the rural sector. This paper, however, focuses on districts within India and studies the movement of labor between the rural and urban sectors I also supplement sub-national analysis with cross-country analysis. While Foster & Rosenzweig (2004) use a HYV productivity index to derive reduced form estimates of the impact of the Green Revolution, this paper incorporates measures of actual productivity changes. Foster & Rosenzweig (2004) also focuses on a slightly different time period. Their analysis begins in 1971, which is after the release of several HYVs including HYV rice, wheat, and maize that play a major role in my empirical results (see Figure 1 of this paper). Last, their HYV index is calculated from information on yield potential from informed sources in the village based on past and existing crop yields, whereas the variation in my measure of predicted agricultural productivity is from exclusively baseline geographic and ecological characteristics. Dasgupta (2014) uses irrigation coverage in 1966 as an instrument for HYV adoption in India it seems unlikely, however, that this instrument would satisfy the appropriate exclusion restriction since the adoption and use of irrigation is endogenous and may be correlated with subsequent economic dynamics. Thirtle et al. use cross country regressions but their identification strategy consists only of running a causal chain model and treating their lagged measure of agricultural research and development as weakly exogenous. Most existing studies are qualitative or rely purely on correlations in cross-sectional or panel data. 4 First, it seems likely that wealthier areas are more likely to invest in land and crop productivity, which would bias the result due to reverse causality agricultural productivity is highly endogenous. Second, there may be some additional feature like institutional quality that is correlated both with agricultural productivity and wealth. Strong and enforceable property laws, for example, might increase the likelihood that individuals and firms invest in the productivity of their land and in increased yield and have an independent effect on economic growth. Resource endowment has been shown to be highly correlated with measures of institutional quality (eg. Acemoglu, Johnson & Robinson 2001, 2002). This potential omitted variable bias makes OLS estimates unreliable. 2

4 In order to investigate the causal link between changes in agricultural productivity and development in India, I exploit rapid increases in crop-level potential productivity that resulted from the staggered release of HYVs during the Green Revolution. I calculate a measure of predicted change in agricultural productivity at the district level that estimates the maximum potential impact of HYV releases based on time-invariant geographic, ecological, and climatic conditions. I then use this plausibly exogenous change in predicted productivity to instrument overall growth in agricultural productivity during the second half of the 20th century. Predicted productivity is calculated from crop-level models of maximum potential yield from the Food and Agriculture Organization (FAO) crucial for this study, the FAO model can be calculated either assuming the use of traditional agricultural technologies or the systematic use of modern varieties. Section 3 of this paper is devoted entirely to developing and testing the instrument. Using this identification strategy, I find that agricultural productivity growth associated with the Green Revolution increased several proxies for rural well-being in India, including access to education, healthcare, and high quality roads. 5 Agricultural productivity growth also increased agricultural wages. Furthermore, I find that agricultural productivity growth spurred growth of the district-level agricultural economy, and that this growth came at the expense of development in the urban and manufacturing sectors. The Green Revolution led to greater employment in farming and a larger portion of district land devoted to agriculture. It also led to a reduction in district-level manufacturing and service sector labor across a variety of occupations. This effect is explained both by lower levels of urbanization and lower relative employment in the rural non-farm sector. The effect on non-farm labor dynamics is attributable not just to movement of labor across sectors within districts, but also to movement between districts lower net migration from other districts explains part, but not all, of the negative relationship between the Green Revolution and nonfarm labor. Highlighting the magnitude of these effects, in the most conservative specifications, a 1000 hectogram per hectare increase in productivity below the average productivity change in Punjab increased farm labor by 27% and decreased manufacturing labor by 33% holding population constant. 5 These are commonly used proxies for district level wealth since, during the early part of the period under analysis, no district-level measures of income per capita exist. See, for example, Banerjee & Iyer (2005, p. 1199). Similar quantities for the urban sector do not exist at the district level. 3

5 At first glance, this finding may seem at odds with the conventional wisdom that agricultural productivity growth increases industrial development. However, significant theoretical and empirical work suggests that this assumption is not always borne out in reality. Matsuyama (1991 p. 643), for example, notes that, [A] takeoff is possible in an economy with less productive agriculture while an economy with productive agriculture will be trapped into the state of preindustrialization. This result, once stated, is intuitive. A low productivity in agriculture implies an abundant supply of cheap labor that the manufacturing sector can rely on. He later notes, and formalizes in Matusyama (1992), that the conventional law that there are positive links between agricultural productivity and industrialization is based almost entirely on the history of England during the Industrial revolution and may not be applicable to the present day, especially low income countries. Field (1978), who argues that early industrialization in Massachusetts resulted in part from low agricultural productivity, supports this hypothesis. The employment effects that I document are perhaps also a result of the fact that, more than earlier breakthroughs in agricultural technology, advancements during the Green Revolution technologies were often labor-complementing (Ladejinsky 1970, Hossain 1998, Lal et al. 2016). John Mellor, for example observes an increase in the demand for labor owing to the higher labor intensity of the modern rice varieties (Hossain 1998 p. 9). Bustos et al. (2016) develop a model in which the impact of new agricultural technologies on labor movement depends on whether the new technology, on average, was labor-saving or land-augmenting. In this framework, the fact that Green Revolution technologies require high labor intensity may lead to a prediction in line with my empirical finding. Greater entrenchment in rural areas and the agricultural sector was, moreover, not limited to patterns of employment. Coinciding with the escalation of the Green Revolution in agriculture was the development and increasing electoral success of political parties explicitly aligned with agrarian interests (Varshnay, 1995 for a full discussion). Earlier work has posited a link between the rise of a commercial agrarian class as a result of agricultural productivity growth and the proliferation of agrarian political parties, competing with the Indian National Congress (INC) for representation (Varshnay 1995, Rudolph & Rudolph 1980, Brass 1980; Dasgupta 2014, for a recent empirical analysis). Rudolph & Rudolph (1980 p. 417) argue that during the Green Revolution there emerged a new rural class we call bullock capitalists. Their political strategy was to 4

6 challenge Congress alleged urban and industrial bias by generalizing the interests of independent agrarian producers to the agricultural sector and rural society as a whole. New party manifestoes called for agricultural price regulation and greater government investment in the rural sector, and were highly critical of the INC (Brass 1980 p ). Using data from Jensenius (2016) on Indian state assembly elections since 1961, I find that agricultural productivity growth indeed led to better electoral outcomes both in terms of votes cast and actual assembly seats won for rural opposition parties. It also led to worse electoral outcomes for the INC. These results contribute to a narrative in which agricultural productivity growth far from spurring urbanization and industrial development increased the economic and political strength of the rural sector, which redoubled its commitment to continued investment in and development of the farm economy. Further probing case-study accounts of political upheaval associated with the Green Revolution, I test hypothesis that the Green Revolution led to an increase in violent conflict. Several earlier studies have suggested this relationship (Cleaver Jr. 1972, Griffin 1990, Shiva 1993). Griffin (1990), for example, avers that the Green Revolution has increased the power of landowners, especially the larger ones, and this in turn has been associated with a greater polarization of classes and intensification of conflict (p. xxi). I find that agricultural productivity growth led to increased district-level conflict and deaths from conflict events. Consistent with several qualitative accounts and the mechanism suggested by Griffin (1990), I also find that the Green Revolution led to a greater concentration of land wealth. Thus, Green Revolution contributed to the redistribution of resources not only between urban and rural areas but also within the farm sector. A major question that remains is how the results of this sub-national analysis aggregate to the country level. It is unclear whether cross-country patterns should mirror those uncovered in the sub-national analysis or whether they should follow a different set of dynamics. First, it may be that agricultural productivity growth had the opposite impact on labor movement at the national level. If countries more closely resemble a closed economy setting than districts within a single country, which can more easily trade amongst themselves, we may expect the countrylevel results to be starkly different from results from within India. In Murphy, Shleifer & Vishny (1989), for example, greater agricultural wealth spurs industrialization by generating demand for manufactures, but local demand is only a pre-requisite for manufacturing development if trade is 5

7 costly. Indeed, the contraction of non-farm activity and growth of the farming sector in districts exposed to the Green Revolution may have been an optimal outcome since it allowed districts to concentrate on their comparative advantage. This set of effects may lead to an overall positive effect on income. Alternatively if, following sub-national trends, the Green Revolution led to greater focus on the agricultural sector at the national level, it may have had country-level effects that dominate the local positive impact of agricultural productivity growth. Lower manufacturing and service sector growth at the national level may lead to an overall negative effect on national income, especially if, for example, there are spillover or agglomeration effects in manufacturing (see Patibandla & Petersen 2002, Lall & Chakravorty 2004, Lall & Mengistae 2005 on agglomeration effects in Indian cities). Gollin, Lagakos & Waugh (2014) show that, especially in developing countries, value added per worker is substantially higher in the non-agricultural sector, so reducing movement from the agricultural to non-agricultural sector might hurt overall growth. These effects would not be captured by the sub-national analysis. I examine the impact of the Green Revolution across countries by exploiting an analogous identification strategy at the country-level. 6 I find, paralleling results from the local level, that at the national level the Green Revolution led to greater fractions of land and labor devoted to agriculture and lower levels of urbanization. Moreover, I find no evidence that the Green Revolution had a positive impact on national total or per-capita income. If anything, the effect is negative for low-income countries. Greater entrenchment in the rural sector is a robust effect of the Green Revolution both at the sub-national and national level. More broadly, these results may suggest that a productivity shock in the less productive of multiple sectors may have limited or even negative impacts by pulling individuals, resources, and political capital from more productive sectors. 7 While agricultural productivity growth associated with the Green Revolution had a positive direct effect on rural income in Indian districts, indirect and equilibrium effects may, depending on the context and 6 The only difference is that I use a broader set of crops to construct the instrument in order to accommodate the broader geographic diversity globally. This is discussed in detail in Section 5. 7 Relatedly, Hsieh & Olken (2014) conclude that [T]he problem of economics development...is how to relieve the differential constraints faced by large firms. Indeed, this view suggests that programs such as microcredit or simplified tax regimes that benefit only small firms may worsen the development problem by further increasing the incentive to stay small (p. 107). While this logic applies to more or less-productive firms, an analogous logic may apply across sectors, where particular policies or innovations increase the incentive to stay in the less productive sector. 6

8 scale, counteract those gains. This study builds on a vast existing literature on the role of agricultural productivity in growth and structural change. Early works by Nurkse (1953), Schultz (1953), Rostow (1960), and A.H. John (1965), and more recently Irz et al. (2001) and Kogel & Prskawetz (2001), argue that agricultural productivity growth is highly associated with and perhaps a precondition for industrial development. Others, however, suggest that the impact of agricultural productivity growth is heterogeneous and contingent (Field 1978, Matsuyama 1992). Hornbeck & Keskin (2015) find that in the U.S. Plains region, for example, access to the Ogallala aquifer, despite increasing agricultural revenue and production, did not generate positive local spillovers to the nonagricultural sector. This paper also contributes to a deeper understanding of the impact Green Revolution in particular. In addition to the work cited above, Evenson & Gollin (2003) [2], Dethier & Effenberger (2012), and Pingali (2012) provide helpful reviews of recent work on the Green Revolution and HYV crop introduction. This paper is also linked to a large body of work that studies the impact of resource abundance and resource based specialization (eg. Cordon & Neary 1982, Auty 1993, Sachs & Warner 2005, and Michaels 2007). A main finding of this paper is that the Green Revolution increased specialization in the agricultural sector. Michaels (2007) is of particular relevance studying the impact of oil abundance in U.S. counties during the 20th century, he finds that the oil abundance increased county-level per-capita income and in the long run reduced the manufacturing employment share. 8 He shows, however, that oil abundance did not reduce the absolute size of the manufacturing sector, and argues that this is because it led to population in-flows. This is distinct from the experience of the Green Revolution in India, where I show that that agricultural productivity growth reduced net inward migration in Indian districts, contributing to the contraction of the non-farm sector. These findings also contribute to a large literature in development economics on determinants of migration and rural-urban migration in particular and variation in wages, within and across sectors. Several existing studies have investigated sources of variation in migration (eg. Banerjee & Newman, 1998, Young, 2013, Bryan, Chowdhury & Mobarack, 2014, Munshi & Rosenzweig, 2016, Imbert & Papp, 2016) and some argue that reducing barriers to migration would have large 8 Although, interestingly, he finds that in the short run oil abundance primarily crowded out agricultural production. 7

9 positive effects. Unlike some existing work that has focused on short-term labor migration, this study analyzes longer run population movement. Foster & Rosenzweig (2004), of particular relevance, find that within the rural sector exposure to the Green Revolution hindered village-level non-farm growth. Whereas their focus is within India s rural sector, this paper studies district and country-level effects in order to investigate equilibrium outcomes. Related prior work has also analyzed sources of variation in rural wealth and wages (eg. Foster & Rosenzweig, 1993, Jayachandran 2006, Goldberg 2016). This paper contributes to a deeper understanding of the forces that drive urbanization and labor force composition in developing countries, and how changes to both effect both local and aggregate income. The instrument used in this paper is motivated by Acemoglu & Johnson (2007). The measure of predicted productivity is in many ways analogous to Acemoglu & Johnson (2007) s predicted mortality instrument My identification strategy also builds on a growing body of work in economic development and economic history that uses variation in land characteristics as a source of plausibly exogenous variation impacting technology adoption (eg. Nunn & Qian 2011, Alesina et al. 2013, Bustos et al. 2016). Bustos et al. (2016), in particular, use the difference in local soybean potential yield at high and low input levels to estimate the impact of the introduction of genetically engineered soybeans in Brazil during the 1990s. The paper proceeds as follows. In the next section, I describe the Green Revolution in more depth along with the data used in the subsequent analysis. Section 3 details the construction of the instrument and presents the first stage results along with a series of robustness checks. Section 4 presents the main empirical results on the impact of the Green Revolution across districts in India. Section 5 presents the cross-country results. Section 6 concludes. 2 Background and Data 2.1 The Green Revolution The 20th century s Green Revolution was a period of steep increases in agricultural productivity that resulted from coordinated global research into the development of high yield varieties of a set of staple crops. Before 1960, few major advances in crop productivity had taken place and the high-yield technologies that did exist were largely limited to conditions in the developed world 8

10 (Evenson & Gollin 2003). While individuals began experimenting with improved agricultural technologies long before the mid-20th century (eg. van Zanden 1991, Sonnenfeld 1992), during the 1960s a process of rapid and coordinated discovery led to drastic increases in potential crop yield (eg. Evenson & Gollin 2003, Foster & Rosenzweig 1996, Sasaki et al. 2002). Several scholars cite 1960 especially the release of much higher yielding rice and wheat varieties during the mid- 1960s as the start of this period of rapid change. Khush (2001 p. 815) notes, It took almost 10,000 years for food grain production to reach 1 billion tons, in 1960, and only 40 years to reach 2 billion tons, in A series of institutional changes occurred during the 1960s that encouraged investment in agricultural research and development and that may have resulted in the rapid production of higher yielding crops. Before the 1960s, there was no effective intellectual property of crop varieties...[but] in the 1960s, Plant Breeders Rights were developed in order to provide incentives for private breeding programs (Evenson & Gollin 2003 p. 2). This institutional adjustment increased potential profits from private sector investment. At the same time, international organizations, backed by an international set of donors, formed to coordinate the development of high yielding crops especially crops suited to low-income countries. This resulted in the establishment of several international agricultural research centers (IARCs) which ultimately coalesced to form the Consultative Group for International Agricultural Research (CGIAR). The combination of private and public sector institutional changes increased crop breeding research and boosted the release of high yield varieties. The canonical example of targeted development of high yield technology has become the release of genetically engineered HYV rice in Genetic engineering allowed wet and dryland rice plants to produce significantly greater quantities than had previously been possible. The first, and arguably most game-changing breakthrough in rice technology was the the IR8 variety so called miracle rice developed during the early 1960s by Dr. Peter Jennings at the International Rice Research Institute (IRRI) in the Philippines. IRRI, one of the IARCs, was established by the Ford and Rockefeller foundations in 1960 to promote crop variety research. The striking difference between IR8 and pre-existing rice varieties was first noted in 1966 when S. K. De Datta found that while pre-existing varieties produced on average one ton of rice per hectare, the IR8 variety could produce 5 tons per hectare even without the use of fertilizer. Moreover, IR8 took fewer days 9

11 to mature than other rice varieties. Distinct from earlier periods of agricultural development, agricultural technologies developed during the Green Revolution had several complementarities with labor. The labor intensive work of modern variety crops has been noted by several scholars (Ladejinsky 1970, Hossain 1998). More recently, researchers have sought to address several shortcomings of the Green Revolution and, in particular, the fact that Green Revolution technologies required high amounts of labor, preventing the movement of labor towards non-farm industries. For example, Lal et al. (2016) propose a new Evergreen Revolution that would introduce technologies that would be...labor saving and that would tend to free labor for nonfarm uses (p. 438). They hope that the next round of agricultural innovations will substitute for, rather than complement, resources under stress and avoid labor retention in agriculture (p. 438). While the Green Revolution indeed made agriculture more productive, significant amounts of labor were required to implement new techniques and utilize new technologies. 2.2 Coding Crops To estimate the impact of changes in agricultural productivity, I exploit plausibly exogenous variation in exposure to the impact of the Green Revolution to calculate a measure of predicted productivity and use predicted productivity as an instrument. The predicted productivity measure relies on theoretical models of maximum potential crop yield from the Food and Agriculture Organization Global Agro-Ecological Zones (FAO GAEZ). 9 These data:...reflect yield potentials with regard to temperature, radiation and moisture regimes prevailing in the respective grid-cells. The model requires the following crop characteristics: Length of growth cycle (days from emergence to full maturity); length of yield formation period; maximum rate of photosynthesis at prevailing temperatures, leaf area index at maximum growth rate; harvest index; crop adaptability group; sensitivity of crop growth cycle length to heat provision; development stage specific crop water requirements, and coefficients of crop yield response to water stress (FAO GAEZ). Crucially, the FAO potential yield model is constructed using parameters derived from controlled experiments, and not from data on actual agricultural inputs and output (see eg. Costinot 9 As noted in the Introduction, FAO data have been used in several recent works to estimate the variation in suitability for agricultural technologies or individual crops, including Qian (2008), Nunn & Qian (2011), Alesina et al. (2013), Fiszbein (2015), Bustos et al. (2016), Costinot & Donaldson (2016). 10

12 & Donaldson, 2016 p. 18). 10 The data are reported by the FAO as a 9.25km 9.25km raster grid, with each grid cell containing the maximum attainable yield for a given crop in that area. Moreover, the data are constructed at both low and high input levels. The low-input level version assumes the general use of traditional cultivars in agriculture while the high-input level assumes the use of high-yield variety crops (FAO GAEZ). While the low input level data indeed account for the fact that there might be some unsystematic use of early high-yield technologies as indeed there was before 1960 it makes the assumption that they are treated in the same way as local cultivars and that the use of these technologies was not coordinated (FAO GAEZ). I also collect information about timing of high yield variety release, primarily from a series of crop reports compiled in Evenson & Gollin (2003) and from Dalrymple (1986). Since this paper s main results focus on long-difference specifications, they do not hinge on the exact years that were identified; however, high yield varieties seem to have been adopted rapidly and broadly following the release of high yield technologies. Evenson & Gollin (2003) find that globally the adoption of modern varieties aggregated across all crops had reached 46% by The figure was 63% by The adoption rate was greater in many parts of East and South East Asia for example, by 1981, the proportion of rice growing area devoted to high-yielding varieties was 74.5% in Indonesia (p. 50). 11 The fraction of India s rice, wheat, and maize cropland devoted to high yield varieties between , averaged across the 271 districts in Sanghi et al. (1998), is presented in Figure 1. The vertical dotted line in each graph indicates the year used for each crop s HYV release year in the construction of the instrument. In all cases, the fraction of cropland devoted to the HYV began to increase in the indicated year. By 1981, on average, the fraction of a district s land devoted to rice cultivation in which high yield varieties were used was over 40%. The proportion was even greater for wheat and maize. Thus, in India high yield variety adoption was widespread and quickly followed technological innovation. 10 The FAO produces two basic measures of potential crop growth: maximum potential yield (described in the quote above) and a crop suitability index. I calculate instrument values using the potential yield measure because it seems less likely to be influenced by endogenous district-level characteristics or growth patterns. The suitability index is calculated using edaphic rating[s] for each soil/slope combination along with soil type. It seems likely that soil quality and local terrain may have been influenced by human behavior and technology. Soil quality does not factor into the FAO calculation of potential yield. 11 The proportion in 1981 was 45.4% in then-burma (Dalrymple 1986, p. 41), 43.5% in Pakistan (p. 56), 78.3% in the Philippines (p. 60), and 70.8% in Sri Lanka (p. 62). 11

13 2.3 Actual Productivity and Outcome Variables For the analysis of Indian districts, both the actual productivity measure and several outcome variables are from the India Agriculture and Climate Dataset, compiled by Sanghi et. al. (1998), which calculates yield measures for 5 major and 15 minor crops at the district level for 271 districts from These data have been used extensively in the economics and political science literature on agriculture in India. 12 Crop-level agricultural output in Sanghi et. al. (1998) is measured in hectograms per hectare. In order to calculate a measure of agricultural productivity that incorporates nutritional content, I matched each crop in the data set to United States Department of Agriculture (USDA) data on calorie content to estimate the caloric load of a hectogram of each agricultural product. Then the district level productivity measure is the sum of the output of all crops (either in hectograms or kilocalories) divided by the area under cultivation for all crops. 13 Since the focus of this paper and of the Green Revolution itself is the impact of productivity growth in crops intended for consumption, I exclude cash crops (eg. tobacco) from my calculation of district level productivity. 14 However, I show that all results are robust to the inclusion of cash crops in the calculation of productivity and, if anything, results using consumption-crop productivity that I present in the body of the paper are more conservative (see Table A3). Outcome variables derive from a variety of sources. Data on population composition, employment, migration, and literacy rates are from India s 1961, 1971, and 1981 censuses. 15 For data on public goods provision (schools, education, and roads), I use measures that were computed from India s censuses in Iyer (2010). Data on state assembly elections are from Jensenius (2016). Each assembly constituency (AC) fits entirely within a single district ACs were matched to districts using the 1961, 1967, and 1976 Indian Delimitation Reports of electoral boundaries. Major opposition and rural opposition parties were identified based on the classification in the Appendix of Dasgupta (2014). Data on conflict are from the Integrated Network for Societal Conflict Research (INSCR) publication Center for Systemic Peace, India Sub-National Problem Set, which 12 For example, Duflo & Pande (2007), Edmonds, Pavcnik & Topalova (2010), Iyer (2011), Dasgupta (2014), Taraz (2014). 13 If N crops c n, n (1, N), are produced in a district, where crop n is k n kilocalories per hectogram, h n hectares of land are used, and H n hectograms are produced, then productivity becomes ( 1 N H nk n )/( 1 N h n). 14 In the main results, I do not factor tobacco, sunflower, cotton, sesame, sugar, jute, or rapeseed production into the district level productivity calculation. Crops included are: Bajra (pearl millet), jowar (sorghum), maize, rice, wheat, barley, groundnut, gram, other pulses, potato, ragi, tur, and soybean. 15 These were compiled for analysis by Reeve Vanneman at the India District Database 12

14 compiles events data drawn from representative news accounts of violent conflicts in India to identify and delineate spatial, temporal, and intensity parameters of societal conflict processes. This data set also reports the local location of the conflict event (sometimes the location given is a district, sometimes it is a smaller geographic unit) I individually matched conflict locations to districts in the 1961 census. Descriptive statistics, including all variables at both the start and end of the period under investigation, are reported in Table 1. 3 Predicted Productivity & First Stage Results 3.1 Calculating Predicted Productivity Using the data described in section 2.2, I construct the predicted productivity instrument as: P it = c[(1 I ct )P L ci + I ctp H ci ] N it (1) P it is predicted agricultural productivity in district i at time t. I ct is an indicator variable that takes on the value 1 in the year that high-yield variety of crop c is released and in all subsequent years. P L ci and PH ci are the maximum potential yield measures for crop c in district i at low and high input levels respectively. For the analysis of India, I use the FAO theoretical model of (i) rice, (ii) wheat, and (iii) maize cultivation since in India, a large portion of HYV-induced productivity growth is attributable to these three crops (Evenson & Gollin 2003, Pingali 2012, Hazell & Ramasamy 1991). 16 N it is the number of crops used in the construction of the instrument that it is possible to grow in district i. In words, the district level value for the instrument is the sum of the district-level predicted productivity measures for all Green Revolution crops divided by the number of crops that could theoretically be grown at the relevant input level. The predicted productivity measure for each crop uses the low-input level version of the FAO s theoretical data for the years before the release of its high yield variety and switches to the high-input level version for all years after the release of the high yield variety. The first stage is presented graphically as a partial correlation plot in 16 Clearly, however, predicted productivity can be constructed with any set of crops for which FAO potential yield models exist. In particular, I show in Section 3.3 that the first stage relationship does not hinge on choosing exclusively rice, wheat, and maize by constructing a version of the instrument that also includes sorghum and barley. 13

15 Figure 2a with the change in the instrument on the x-axis and the change in actual productivity on the y-axis the t-statistic is 7.51, a first indication of a strong first stage relationship. 17 The validity of the instrument relies on the assumption that Cov(P it ɛ it ) = 0, where ɛ it is the error term in the second stage regression (equation (3), below). Since the indicator variable I ct changes when the HYV was released, in the same year for all districts, potentially endogenous rates of HYV adoption do not bias the instrument. As (1) shows, variation in the instrument is determined entirely by baseline district level characteristics that influence the potential responsiveness to modern variety versions of a set of important crops, combined with the timing of global technological developments. A majority of the analysis relies on long-difference specifications, which do not incorporate variation in the timing of HYV releases and rely exclusively on baseline characteristics. It seems unlikely that this instrument is correlated with any regional changes in economic, institutional, or geographic conditions. 3.2 Zeroth Stage Part of the logic behind the first stage is that in districts where there was a larger potential productivity gain from the adoption of HYVs, HYVs were more broadly adopted leading to increases in overall agricultural productivity. Table A1 documents a strong, positive correlation between predicted productivity the instrument and district-level HYV adoption. This relationship is robust to the inclusion of a broad set of controls and remains strong when rice, wheat, or maize is dropped from the construction of predicted productivity. 3.3 Baseline First Stage Estimates The first stage relationship between predicted and actual productivity is modeled as: X it = βp it + η i + ξ t + Z γ + u it (2) X it is actual agricultural productivity in district i at time t. All specifications include both time and district fixed effects. For the purpose of the first stage, the coefficient of interest is β, which 17 Controls are latitude, longitude, initial log agricultural wages, initial population density, and the initial male literacy rate. 14

16 measures correlation between the predicted productivity measure and actual district-level output per hectare. Z is a vector of controls that changes depending on the specification and u it is an error term. Table 2 presents the baseline first stage results. In Columns 1-4, both predicted and actual productivity are measured in hectograms per hectare. In the remaining columns, both are measured in kilocalories per hectare. All regressions are long difference specifications, including just the years 1961 and These years were chosen since data in Sanghi et al. (1998) covers and, within that period, only 1961 and 1981 are years in which the census was conducted. Data from the census are used extensively in the second stage so, in order for the first stage regressions to correspond appropriately to the second stage, in most of the analysis only 1961 and 1981 are considered. The first column includes exclusively the instrument and the fixed effects on the right hand side and suggests a strong first stage relationship between predicted and actual productivity (F-statistic = 20.01, β = 0.219). The second column adds a set of initial district-level characteristics latitude, longitude, a coastline indicator, log of average agricultural wages, and tractor ownership normalized by population interacted with an indicator variable that equals 1 in 1981 ( postyear indicator interaction). This allows for differential trends based on a range of initial geographic and agricultural characteristics. Despite the inclusion of these controls, the point estimate is slightly larger (β = 0.265) and the first stage relationship increases in statistical significance (F-statistic = 29.47). The third column adds to these the postyear indicator interacted with the 1961 adult male literacy rate and population density, potential proxies for development at the start of the period. The coefficient magnitude and significance are similar (F-statistic = 27.81, β = 0.249). Despite the inclusion of these controls, designed to capture the role of initial geographic and economic characteristics on trends in agricultural productivity, the first stage relationship remains robust. Since this identification strategy builds on insight from Acemoglu & Johnson (AJ) (2007), I next address the main critique of the AJ approach. Bloom, Canning & Fink (BCF) (2009) argue that the predicted mortality instrument does not adequately account for differences in initial mortality rates. They argue that the AJ (2007) result depends crucially on their assumption that initial health and income do not affect subsequent economic growth. Similarly here, one might argue that initial agricultural productivity (analogous to initial mortality) impacts subsequent 15

17 economic growth. In their response to BCF (2009), AJ (2014) show several additional specifications that they favor over those used by BCF (2009). However, to preempt a critique in the spirit of BCF (2009), I explain here why the present analysis is robust to their concerns. First, I do not use actual productivity data in the construction of the predicted productivity instrument. Preintervention values, like post-intervention values, were taken from the FAO s theoretical models of potential yield that are based on topological and climatic characteristics. As a result, values of the predicted agricultural productivity measure are not determined by potentially endogenous pre-intervention agricultural productivity or any real world data. Second, I control directly for initial agricultural productivity in the first-stage regression model to show that the strength of the instrument does not hinge on district-level differences in initial agricultural productivity. Column 4 of Table 2 adds to the existing controls an interaction term between actual agricultural productivity (hg/ha) in 1961 and a post-year indicator. 18 If variation in the instrument were in large part determined by differences in agricultural productivity at the start of the period, we might expect the coefficient on predicted productivity to lose significance. However, the magnitude and significance of β remain virtually unchanged despite the inclusion of this control. Reassuringly also, the magnitudes of the coefficients, β, are logical. Whereas the instrument is a measure of maximum attainable yield based on geographic and ecological characteristics, the outcome X it is the yield that was actually attained in district i at time t. Therefore, it makes sense that β (0, 1) in all first stage specifications. Columns 5-8 repeat the same specifications as Columns 1-4 except the outcome variable and the instrument are both measured in kilocalories per hectare. The story is largely the same as in the first four columns except that the F-statistic is slightly lower in all cases I find no indication, however, that the instrument is weak. 18 Ideally, I might want to control for changes in agricultural productivity or in income over a period before the Green Revolution to account for historical trends in the data. In their response to BCF (2009), AJ argue that controlling for actual mortality several decades before the period that they study is more appropriate than controlling for mortality at the start of the period of study, as Bloom et. al. propose. Unfortunately, systematic data on agricultural yield do not exist for years prior to 1957 in the case of India or prior to 1960 at the country level, so I am unable to control for earlier agricultural productivity or earlier changes in agricultural productivity. However, when I control for the actual value of our instrumented variable at the start of the period, the first stage relationship seems to remain strong. See AJ (2014) for a full discussion. 16

18 3.4 Robustness & Falsification Tests The first stage results in Table 2 do not allow for the inclusion of lead or lagged variables on the right hand side. Moreover, they assume that contemporaneous predicted productivity impacts actual productivity. If, however, future predicted productivity is correlated with current actual productivity, this would cast doubt on the validity of the first stage. Table 3 presents a series of results using a panel that includes a single observation every five years in order to allow for the inclusion of lagged and lead variables. Again, in the first set of columns (1-5) agricultural productivity is measured in hectograms per hectare and the second set (6-10) it is measured in kilocalories per hectare. All specifications include district and year fixed effects along with a full set of time indicator interactions with latitude, longitude, log agricultural wages in 1961, population density in 1961, and the adult male literacy rate in For reference, the first column includes just the instrument on the right hand side (along with the controls) to show the baseline first stage when a panel, rather than long difference is used. The specifications in Table 2 do not allow for mean reversion in agricultural productivity. Column 2 of Table 3 includes a five-year lag of actual agricultural productivity (hg/ha), allowing past productivity to affect current productivity. There is evidence for mean reversion: the coefficient on lagged productivity is statistically significant. Nevertheless, the first stage relationship remains strong β is significant at below the 1% level (F-statistic = 20.12). In column 3, current and lagged (five-year) log agricultural wages and adult male literacy are included on the right hand side. β is virtually unchanged from the first column. One shortcoming of including these lagged variables on the right hand side is that, since all specifications include fixed effects, the estimates suffer from Nickell bias. 19 As a result, the baseline specifications (Table 2) do not include any lags and the estimates in Table 3 should be treated with a degree of skepticism; nevertheless, it is reassuring that in all specifications the first stage relationship remains highly significant. Column 4 uses an instrument constructed using two additional crops whose modern varieties 19 One solution to this problem is to remove the district fixed effects in the regressions that include the lagged dependent variable. The coefficient on the instrument remains significant. Reestimating column 1 without district fixed effects, β = (se = 0.009) in Panel A and β = (se = 0.009) in Panel B. An alternative solution would be to use the second lag of the dependent variable to instrument the first lag of the dependent variable. However, using five year lags, this approach would require dropping observations before 1970 which eliminates the variation in the instrument. HYVs for rice, wheat, and maize were all released prior to 1970, so the impact of these releases on variation in the instrument would be excluded. 17