The World Bank Economic Review, 38(2), 2024, 229–250 https://doi.org10.1093/wber/lhad042 Article Agricultural Productivity and Land Inequality: Downloaded from https://academic.oup.com/wber/article/38/2/229/7499128 by Sectoral Library Rm MC-C3-220 user on 01 May 2024 Evidence from the Philippines Ludovic Bequet Abstract How do agricultural productivity gains affect the distribution of agricultural land? Exploiting three waves of census data from the Philippines covering 21 years and 17 million plots, this article finds that municipalities endowed with favorable soil and weather conditions for genetically modified (GM) corn cultivation experience a relative increase in landholding inequality. Agricultural land is decreasing during this period and this decrease is driven by a decline in the size of large farms. The introduction of GM corn slows down this process by keeping more land under cultivation, which contributes to the documented relative increase in inequality. JEL classification: O13, Q12, Q14, Q15 Keywords: Land inequality, agricultural technology, land reform 1. Introduction The structure of a country’s agricultural sector is strongly linked to its development level. Low-income countries are characterized by a large number of smallholder farmers, while in high-income countries, farms tend to be larger and fewer.1 This difference can be explained by the process of structural transfor- mation, whereby workers move out of agriculture into the industrial and the service sectors. This implies a substantial reallocation of agricultural land between those who leave and those who stay. How this re- allocation takes place shapes the land distribution, which has implications for the distribution of income and wealth at the national level. Gains in agricultural productivity have been identified as a key driver of this structural transformation, as they reduce the demand for agricultural labor and increase the demand for manufacturing goods. While there has been an extensive literature studying the impact of agricultural productivity on land expansion (see Villoria, Byerlee, and Stevenson (2014) for a review), its effect on land inequality has been relatively less investigated. This is striking given that modern agricultural technologies are often blamed for favoring Ludovic Bequet is a researcher at the University of Namur, Namur, Belgium; his email address is ludovic.bequet@unamur.be. Research on this project was financially supported by the Excellence of Science (EOS) Research project of FNRS O020918F. The author thanks Jean-Marie Baland, Catherine Guirkinger, Benoit Decerf, Matthieu Chemin, Peter Lanjouw, Tanguy Bernard, Marc Sangnier, four anonymous reviewers, and participants at the EEA Conference 2021 and the EOS Annual Workshop 2021 for helpful comments, as well as to Eric Edmonds for his outstanding editorial work. The author also thanks Andres Ignacio and Alberto Marin for their assistance with the data collection and for sharing their work on PSGC codes. A supplementary online appendix is available with this article at The World Bank Economic Review website. 1 Using agricultural census data from 92 countries, Lowder, Skoet, and Raney (2016) find that farms smaller than 2 ha account for 30–40 percent of land in low- and lower-middle-income countries and less than 10 percent in upper-middle- and high-income countries. C The Author(s) 2023. Published by Oxford University Press on behalf of the International Bank for Reconstruction and Development / THE WORLD BANK. All rights reserved. For permissions, please e-mail: journals.permissions@oup.com 230 Bequet large farms, at the expense of smallholder farmers, leading to an increase in land concentration. These claims are especially common for genetically modified (GM) crops but are rarely backed by data, or only based on very loose empirical analysis (Catacora-Vargas et al. 2012; Phélinas and Choumert 2017). Herbicide tolerance and pest resistance—the two main traits in GM crops—are labor saving as they decrease the need for manual weeding and pesticide spraying respectively. As Bustos, Caprettini, and Downloaded from https://academic.oup.com/wber/article/38/2/229/7499128 by Sectoral Library Rm MC-C3-220 user on 01 May 2024 Ponticelli (2016) show, this kind of labor-augmenting technology can drive structural transformation and is therefore likely to lead to a redistribution of agricultural land. Moreover, the higher return on capital is likely to favor better-off farmers and lead to higher levels of inequality. This paper presents an empirical evaluation of the effect of agricultural productivity on land inequality, focusing on the two decades surrounding the introduction of GM corn seeds in the Philippines. Corn is the second most cultivated crop in the country, mostly by smallholder farmers who rank among the poorest categories of the population (Reyes et al. 2012). GM seeds were introduced in 2003, rapidly adopted by the farmers, and can be considered as the most important technical innovation for corn agriculture in the recent decades. The economic literature on land distribution usually studies the impacts of land inequality rather than its drivers. The most compelling argument for a more equal land distribution comes from a series of papers, starting with Alesina and Rodrik (1994), showing a negative correlation between inequality— especially land inequality—and economic growth.2 Historical evidence suggests that this is driven by lower investment in physical and human capital in areas with unequal land distribution.3 Likewise, land redistribution policies have been shown to decrease poverty in India (Besley and Burgess 2000), South Africa (Keswell and Carter 2014), and the Philippines (Reyes 2002; World Bank 2009). This may be due to the fact that a more equal distribution generates more employment per hectare (and per unit of output) as small-sized farms are more labor intensive and access to land provides a safety net which may encourage non-farm business investment (Binswanger-Mhkize, Bourguignon, and van den Brink 2009). Furthermore, as agricultural activity in developing countries exhibits diseconomies of scale—the so-called inverse farm-size productivity—redistributing land to smallholder farmers may lead to efficiency gains. This is supported by Vollrath (2007), who finds a negative relationship between land Gini and agricultural productivity using cross-country data. However, this claim has been challenged by Foster and Rosenzweig (2017) who show with microdata that the relationship between farm productivity and size is in fact U- shaped and that large farms are as efficient as small ones, even in developing countries.4 Land inequality has also been linked with an increased likelihood of conflict (Peters 2004; de Luca and Sekeris 2012; Thomson 2016), environmental degradation (Sant’Anna 2016; Ceddia 2019), unemploy- ment (Ortiz-Becerra 2021), and reduced resilience to natural disasters (Anbarci, Escaleras, and Register 2005).5 Despite this large number of studies on the—mostly negative—effects of land inequality, there exists surprisingly little research on its drivers. Two notable exceptions are Bardhan et al. (2014) who use rich panel data from West Bengal to show that household division is a much larger driver of land distribution than land market transactions or the land reform, and Berg et al. (2023) who find that land ownership inequality responds positively to labor-market integration. At a more aggregate level, Lowder, Skoet, and Raney (2016) and Jayne et al. (2016) also provide a detailed description of agricultural land distribution, respectively for the whole world and in four African countries. 2 See also Easterly (2007), Fort (2007), Neves, Afonso, and Silva (2016), and Cipollina, Cuffaro, and D’Agostino (2018) for meta-analyses. 3 Banerjee and Iyer (2005) , Galor, Moav, and Vollrath (2009), Cinnirella and Hornung (2016), Baten and Hippe (2018). 4 Similarly, Adamopoulos and Restuccia (2019) find that land redistribution during the agrarian reform in the Philippines led to a 17 percent decrease in productivity. 5 See also Guereña and Wegerif (2019) for a recent multi-disciplinary review. The World Bank Economic Review 231 The question of the distributional impacts of agricultural technology is, however, not new in economics and echoes an old literature studying the distributive effects of the Green Revolution, especially in South Asia6 but relying on very limited data sources, usually from a few hundred households. Moreover, these papers only focused on describing the change in inequality and did not rely on causal identification strate- gies. The present work therefore addresses an old question using modern empirical tools. It is also linked to Downloaded from https://academic.oup.com/wber/article/38/2/229/7499128 by Sectoral Library Rm MC-C3-220 user on 01 May 2024 the literature on agricultural productivity and structural transformation, in particular Bustos, Caprettini, and Ponticelli (2016)7 and can be seen as a description of the land redistribution process resulting from a more structural change of the economy. To document the landholding inequality in the Philippines during the decades surrounding the intro- duction of GM corn in 2003, I use three waves of agricultural census data covering 21 years and 17 million plots. As the census only surveys agricultural operations, the inequality measures used in this paper reflect landholding inequality and not land ownership inequality, which would require additional information on non-farming landowners. See the Data and Descriptive Statistics section for a discussion on the limits and the benefits of agricultural census data. At the national level, landholding inequality increased between 2002 and 2012, despite an ongoing land reform aimed at redistributing agricultural land. Theil’s inequality decomposition reveals that within- municipality inequality accounts for 80 percent of total inequality. Changes in national inequality are therefore highly likely to be driven by changes at the local level and the rest of the empirical analysis takes the municipality as the unit of observation. As the census data do not distinguish between GM and non-GM corn, it is not possible to correlate the use of the technology with land-inequality measures. Moreover, such an empirical strategy would be subject to reverse-causality bias as it is not clear whether a positive correlation would mean that higher adoption rates lead to higher land concentration, or simply that the technology is adopted in places where land is less equally distributed. To overcome this identification issue, I take advantage of exogenous variations through space and time. First, I compare data collected in 2002—one year before GM seeds were commercialized—with data from 2012, in a long-difference setting, similar to a municipality fixed effects model. Second, cross-sectional variation in profitability of GM corn adoption is used, following the approach of Bustos, Caprettini, and Ponticelli (2016). More specifically, the Food and Agriculture Organization (FAO)’s Global Agro-Ecological Zones (GAEZ) database computes maximum potential yield of corn by using local soil and weather characteristics. Importantly for this approach, those potential yields are computed using three levels of input use intensity: low, medium, and high. The difference in potential yield between low and high input levels serves as proxy for our treatment as it indicates whether the productivity of corn cultivation is more or less likely to be impacted by the new technology. This allows us to compare the change in land inequality between municipalities that benefited substantially from the technology and those that only benefited marginally. Results show that landholding Gini and the share of land occupied by farms in the top decile increased in more impacted municipalities relative to those less impacted. This effect appears to be driven by a smaller contraction of agricultural land in those municipalities, rather than by land consolidation per se. The introduction of GM corn slowed down the trend in land contraction occurring over the same period, which was characterized by a decline in the size of the largest farms. By allowing more land to remain under cultivation, belonging on average to larger farms, the new technology prevented a decline in land inequality in areas where it was more beneficial. 6 Bardhan (1974) , Raju (1976), Chaudhry (1982), Prahladachar (1983), Otsuka, Cordova, and David (1992), Freebairn (1995). 7 Note that, while the new corn variety described in Bustos, Caprettini, and Ponticelli (2016) is a land-augmenting tech- nology, the introduction of GM corn in the Philippines was likely labor augmenting and is more comparable to that of GM soy in their paper. 232 Bequet On the other hand, change in crop mix, migration, and economic development do not appear to play any particular role. Heterogeneity analysis reveals that the effect is stronger in municipalities that adopted modern inputs later, i.e. where the potential for yield increase was higher, and with stronger credit penetra- tion 10 years before the seeds’ commercialization. Geographical heterogeneity analysis shows a stronger effect on the southern island of Mindanao. Finally, looking at land ownership inequality instead of land- Downloaded from https://academic.oup.com/wber/article/38/2/229/7499128 by Sectoral Library Rm MC-C3-220 user on 01 May 2024 holding inequality reveals that this measure follows a similar pattern, although its measurement is more problematic because of data limitation. To assess the robustness of the results, a series of tests are presented. First, a placebo test comparing data from 1991 and 2002—i.e. before GM corn introduction—fails to find any differential pre-trends in landholding inequality, bringing support to the common trend assumption which underlies the estimation strategy. Pooling all three census waves in a triple-difference setting confirms the absence of results before 2002 and the positive correlation between potential yield and inequality afterwards. Adding this third wave of data also allows to control for municipality-specific time trends, which does not affect our results. Other robustness tests involve adding topographical and geographical controls, taking spatial correlation into account with Conley standard errors and by clustering at the provincial level, running the analysis at the level of the barangay (village), using an alternative definition of the treatment variable, and matching on observables. All of those confirm our main results. 2. Background Corn is the second-most-cultivated crop in the Philippines. It is used for consumption and is also sold to the booming animal feed industry. In 2003, the country approved the commercialization of GM corn seeds. Farmers were fast to adopt this new technology and, by 2014, 62 percent of the hectarage devoted to corn was planted with GM seeds (ISAAA 2017).8 In addition to the patented GM seeds, illegal open- pollinated varieties (OPVs) containing herbicide-tolerant traits have been reported in the south of the country. These varieties, locally known as sige-sige, are the result of cross-breeding between traditional cultivars and GM corn seeds and would represent between 35 and 50 percent of maize farm land in Mindanao (south) and the Visayas (center) (De Jonge, Salazar, and Visser 2021).9 Figure 1 shows the evolution of corn and rice yields (in tons per hectare) between 1990 and 2016, using official data from the Department of Agriculture. In the decade following the introduction of GM corn, corn yield almost doubled. Such a large gain in productivity was not observed in rice, the main crop of the Philippines. Despite sustained economic growth during our study period 1991–2012, poverty incidence remained high in rural areas, as 57 percent of agricultural households were characterized as poor in 2009, three times the proportion of non-agricultural households (Reyes et al. 2012). The Philippines is also characterized by a high level of income, wealth, and land inequality, owing to the legacy of Spanish colonialism which constituted a landed elite class occupying prominent positions in the country’s political and economic apparatus. This high level of inequality is at the root of the civil conflicts that have beset the country in the past decades, among which is the Moro insurgency on the island of Mindanao (McDoom et al. 2019). In an effort to address the issue of land inequality, the country has undergone a series of land reforms since the beginning of the twentieth century. The most recent one, the Comprehensive Agrarian Reform Program (CARP), started in 1988. The scope of this reform was extensive, as it covered all agricultural 8 The first generation of biotech corn included the Bacillus thuringiensis (Bt) trait, which confers the plant pest tolerance. In 2005, new varieties were commercialized exhibiting herbicide tolerance (Ht) as well. By 2012, the overwhelming majority of GM corn planted in the Philippines had both traits (Bt/Ht) (Aldemita, Villena, and James 2014). 9 Very little is known about the exact characteristics, origin, and spread of this sige-sige corn. These figures are in line with those found by Bequet (2020) in a case study in Northern Mindanao. The World Bank Economic Review 233 Figure 1. Temporal Evolution of Corn and Rice Yield. Downloaded from https://academic.oup.com/wber/article/38/2/229/7499128 by Sectoral Library Rm MC-C3-220 user on 01 May 2024 Source: Data from the Bureau of Agricultural Statistics 2005, 2008, and 2013 and the Philippines Statistics Authority 2018. Note: Observed yield of corn and rice in tons per hectare. land with a few exceptions.10 Both tenants and regular farm workers were included as recipients, as long as they were landless or smallholder farmers (with less than 3 ha of land) and the reform put an upper limit on ownership of agricultural land at 5 ha.11 Thirty years after the start of the implementation, the CARP claims to have redistributed 4.8 million ha to 2.8 million households (Ballesteros, Ancheta, and Ramos 2017). These figures, however, appear unrealistically high.12 In addition, several scholars have criticized the reform implementation process for being captured by the landed elite and resulting in little distribution of wealth and power to the landless and smallholder farmers (Lanzona 2019; Borras 2006; Borras, Carranza, and Franco 2007). Such critiques are not specific to the Philippines, as Ghatak and Roy (2007) find that the land reform in India had an insignificant effect on land ownership inequality and even slightly increased operational land inequality in regions with poor implementation. 3. Data and Descriptive Statistics The main source of data for this paper consists of three waves of the Census of Agriculture and Fisheries (CAF), collected in 1991, 2002, and 2012 by the Philippines Statistical Agency (PSA), under the supervi- sion of the FAO’s World Census of Agriculture. An additional geocoded data source and the Census of Population (CP) are also used to create control variables. 10 Exceptions include military reservations, penal colonies, educational and research fields, timberlands, undeveloped hills with 18-degree slopes, and church areas. 11 This limit was increased by 3 ha per heir of minimum 15 years at the time of the reform, provided that they were willing to continue tilling or managing the farm. 12 Indeed, according to the agricultural census, there were 3.76 million farmers in the Philippines in 1991 and when we add up the land area under leasehold and tenancy with the area owned in excess of 5 ha, we only reach 4.1 million ha. If the redistribution numbers are true, we would therefore observe a much larger decrease in land inequality than what is found in the subsequent censuses. 234 Bequet 3.1. Agricultural Census The use of agricultural censuses to investigate land inequality is common in the literature, going back to Deininger and Squire (1998). However, it presents several drawbacks, highlighted by Bauluz, Govind, and Novokmet (2020) who advocate for the use of household surveys instead. They show that, while both data sources give comparable land Gini coefficients, adjusting for the landless population and Downloaded from https://academic.oup.com/wber/article/38/2/229/7499128 by Sectoral Library Rm MC-C3-220 user on 01 May 2024 the land value—both absent from census data—leads to important changes in inequality measures. Moreover, agricultural censuses survey farmers and not landowners, which implies that only landholding, and not land ownership, inequality measures can be computed. As noted by Vollrath (2007), landholding inequality matters for efficiency while land ownership inequality is more relevant from an equity perspec- tive. Both measures are, however, highly correlated, with landholding inequality usually considered as a lower bound for land ownership inequality. While agricultural censuses do have shortcomings, they also offer the extensive coverage needed for the kind of analysis carried out in this paper. Indeed, computing land-inequality indicators at the local level (municipality or even village) using household surveys would be highly imprecise given the low number of households typically surveyed in each location. Moreover, household surveys only take into account household farms, and therefore systematically miss company-owned farms, which tend to be larger. For example, in the CAF 2012, corporations accounted for 0.05 percent of farms but 3.4 percent of agricultural land.13 The analysis presented in this paper therefore includes farms with all types of legal status.14 3.1.1. Data Harmonization CAF data provide plot-level information including size, tenure status, main use, and the crops cultivated over the past year. Harvest and input information are unavailable except for some very coarse measures of input use in 1991. Farms are defined at the level of the household, and those with a total land area below 0.1 ha are removed from the analysis, a cutoff used in the 2002 census.15 Only the 2012 CAF wave provides a complete enumeration of all the farms in the country. In 1991 and 2002, a sample of barangays was drawn from each municipality. All farming households living in the sampled barangays were then enumerated. Sampling weights allow the computation of municipality-level statistics and are used in all of the empirical analysis. Plot location is reported at the barangay level in 1991 and 2012 and only at the larger, municipality level in 2002. As our land-distribution measures are computed based on the physical location of the plot and not on the residence of its operator, all of the analysis is carried out at the municipality level. Furthermore, if a farm is located across two municipalities A and B, it is treated as two different farms, one with the land in A and the other with the land in B and the inequality measure of municipality A only uses the land located in A.16 The CAF also reports the land tenure status of each plot. For the purpose of the analysis, it is divide between ownership (full ownership, owner-like possession, and various forms of community ownership) and tenancy (rental, leasehold, rent-free occupation). When the farmer is a tenant, there is no information regarding the owner of the plot. Indicators of land inequality therefore measure landholding inequality and not land ownership inequality. 13 As an extreme example, Lowder, Skoet, and Raney (2016) show that in Guatemala, the top 2 percent largest farms from the agricultural census, representing 57 percent of total land, are absent from the LSMS household survey. 14 The legal status of agricultural holdings reported in the CAF are the following: individual, partnership, corporation, cooperative, other private institution, government corporation/institution, and other. 15 This ensures that the temporal variations we find in the land distribution are not the result of changing farm definitions and that the households considered devote a significant amount of resources to their farming activity. 16 In 2012, 1.5 percent of farms had land located in two different municipalities while 8 percent of the cultivated land was located in a different municipality from where the farmer lived. The World Bank Economic Review 235 Table 1. Summary Statistics of National Land Distribution 1991 2002 2012 Agricultural area (million ha) 8.57 9.56 7.56 Number of farms (million) 3.76 4.80 4.55 Average farm size (ha) 2.28 1.99 1.64 Downloaded from https://academic.oup.com/wber/article/38/2/229/7499128 by Sectoral Library Rm MC-C3-220 user on 01 May 2024 Landholding Gini 0.590 0.576 0.606 Share top 1%a 18.73% 15.34% 19.68% Share top 10% 46.85% 44.86% 48.02% Share bottom 50% 13.10% 13.74% 12.32% Share tenanted land 34.07% 31.19% 27.80% Share tenanted farms 31.01% 25.30% 25.75% Population (millions)b 60.703 75.698 92.100 Share of rural population 51.3% 48.9% 50.9% Source: Data from the Philippines Census of Agriculture and Fisheries (CAF) and Census of Population (last two rows). Notes:a Shares of top and bottom percentiles represent the share of total agricultural land occupied by farms in the corresponding percentiles. b Computed from the Population Censuses of 1990, 2000, and 2010. Most of the empirical analysis of this paper focuses on the difference in municipality-level inequality between 2002 and 2012. In order to ensure that any difference we find is not driven by the sample compo- sition, the 2012 data are restricted to the barangays enumerated in 2002 when computing municipality- level indicators. In addition, municipalities with less than 50 ha of agricultural land and those located in Metropolitan Manila (National Capital Region) are dropped from the analysis. This restricts the sample to areas where farming is of some importance. This sample restriction alleviates the issue of outliers driv- ing our results and is applied throughout the rest of the empirical analysis. A more detailed description of the data cleaning process, variable definitions, and differences between CAF waves can be found in supplementary online appendix S1. 3.1.2. Land Distribution across Farms The distribution of agricultural landholdings in the Philippines is described in table 1. The total land devoted to agriculture increased between 1991 and 2002 from 8.6 to 9.6 million ha and then strongly decreased to 7.5 million ha in 2012. This pattern is driven by a strong increase in farm numbers between 1991 and 2002 and a steady decrease in average farm size over the whole period, which was probably driven by the land reform. In addition, total population strongly increased over the period, from 60 to 92 million inhabitants, while the share of rural population remained relatively constant, around 50 percent. This strong demographic expansion increased the pressure on land and may also explain part of the decline in farm area and farm size. Land-inequality measures also exhibit a non-linear pattern, decreasing in the first decade and then increasing to levels higher than in 1991. The Gini coefficient—the most commonly used inequality indicator—is 0.606 in 2012, up from 0.590 in 1991 and 0.576 in 2002. Such levels are high for the ASEAN region but remain below those found in Latin American countries by Guereña (2016). The share of land occupied by different fractiles shows a very similar pattern of decreasing inequality between 1991 and 2002, which is reversed between 2002 and 2012. At the end of the period, farms in the top percentile (decile) control almost 20 percent (50 percent) of the land, a share that has increased by more than 4 percentage points (3 percentage points) since 2002. At the other end of the distribution, the 50 percent smallest farms occupy 12 percent, down from 14 percent in 2002.17 17 To compute the share of land held by the top decile, I rank the farms by land size, sum the land occupied by the top 10 percent largest farms and divide that amount by the total farm area. 236 Bequet Finally, the share of tenanted land decreases steadily over the two decades, while the share of tenanted farms declines sharply between 1991 and 2002 and then remains stable around 25 percent. This indicates that land ownership inequality exhibits a different pattern from landholding inequality.18 3.1.3. Inequality Decomposition and Municipality-Level Land Inequality Downloaded from https://academic.oup.com/wber/article/38/2/229/7499128 by Sectoral Library Rm MC-C3-220 user on 01 May 2024 Since land is an immobile asset, it is expected that most of the inequality is to be found at the very local level. Intuitively, farmers need to live close to their farms, either because they work in them or because they need to be able to monitor their workers. It is therefore not possible for large farmers to concentrate in specific areas in the same way that wealthy individuals live in the same neighborhoods. Indeed, Theil’s T decomposition shows that within-municipality inequality accounts for around 84 percent of total land inequality in 2012.19 . See supplementary online appendix S2 for the technical details of the decomposition. Maps of municipality-level land Gini for the three waves of data are reported in fig. S4.1. Spatial correlation appears relatively limited, except for some regions characterized by strong land inequality, such as the island of Negros in 1991 and central Mindanao in 2012. Temporal persistence, on the other hand, is high as unequal regions in 1991 tend to be more unequal in 2002 and 2012. The increase in land inequality over time is reflected by the darker colors in 2012. 3.2. Additional Data Sources Aside from the CAF data, the analysis presented in this paper relies on additional data sources. First, the Census of Population (CP), available for the years 2000 and 2010, gives the municipality population and allows the computation of the share of rural population and the share of farming households. Second, GIS data from various sources are used to complement farm- and household-level data. r Crop suitability measures come from the FAO Global Agro-Ecological Zones (GAEZ) database, which predicts yields for each crop based on soil, climate conditions, and agricultural practices at a resolution of 10 km per pixel. This measure will be further detailed in the section presenting the empirical strategy.20 r Geophysical measures such as altitude and ruggedness are computed thanks to the Space Shuttle Radar Topography Mission (SRTM) digital elevation model, which has a pixel size of 90 m. r Night-light data come from the Defense Meteorological Program Operational Line-Scan System (DMSP- OLS), with a pixel size of 1 km. Each administrative area in the Philippines is uniquely identified by a Philippines Standard Geographic Codes (PSGC). These codes are used to match the different waves of CAF and CP data over time and with GIS data, using administrative boundaries shapefiles, obtained from the UN Office for the Coordination of Humanitarian Affairs (OCHA). As municipality boundaries changed over the period of analysis, manual matching by barangay names was carried out in order to increase the quality of the match.21 18 Sampling weights were used to compute those national statistics. See table S3.1 in the supplementary online appendix for the same statistics computed without sampling weights. The number of farms and farm area are much smaller without the weights (as expected), while the inequality measures are relatively similar. 19 Using Theil’s L decomposition, within-municipality inequality accounts for 81 percent of total inequality. This share is slightly lower in 1991 (79 percent and 78 percent for Theil’s T and L respectively) and slightly larger in 2002 (85 and 82 percent respectively 20 The data used in the analysis come from the third version of the GAEZ. 21 In case of split/merge between municipalities over the course of the study period, I always aggregate barangays to form the largest stable entities. I am grateful to Andres Ignacio from ESSC for providing me his match between the PSGC 2000 and PSGC 2010. The World Bank Economic Review 237 4. Identification Strategy This paper focuses on the period surrounding the introduction of genetically modified corn in the Philip- pines, which took place in 2003. We therefore have a first census conducted 12 years before (CAF 1991), another one conducted 1 year before (CAF 2002), and the last one 10 years later (CAF 2012). The main empirical analysis compares the two latest censuses, while using the first one to control for historical Downloaded from https://academic.oup.com/wber/article/38/2/229/7499128 by Sectoral Library Rm MC-C3-220 user on 01 May 2024 differences that may be correlated with the treatment. Because the data do not distinguish between different corn varieties, GM corn adoption cannot be observed and it is not possible to look at the direct impact of adoption on land use and distribution, regardless of the endogeneity of technology adoption. To overcome this issue, I use the empirical strat- egy developed by Bustos, Caprettini, and Ponticelli (2016) in their paper on structural transformation in Brazil. This strategy exploits the fact that differences in soil and weather characteristics lead to differences in potential gain from adopting the technology, thereby creating exogenous cross-sectional variation in adoption and in treatment intensity. The measure of this exogenous potential gain from GM corn cultiva- tion is obtained from the FAO GAEZ database, which predicts yields for each crop based on soil, climate conditions, and agricultural practices.22 Crucially for our strategy, those agricultural practices include various degrees of input level intensity. The low level of inputs implies that “the farming system is largely subsistence based. Production is based on the use of traditional cultivars..., labor-intensive techniques, and no application of nutrients, no use of chemicals for pest and disease control and minimum conser- vation measures.” The high input level implies that “commercial production is a management objective. Production is based on improved or high yielding varieties, is fully mechanized with low labor intensity and uses optimum applications of nutrients and chemical pest, disease and weed control.” The difference in potential corn yield between high and low levels of inputs therefore serves as a proxy for the profitability gain from improved agricultural technology—in our case, GM seeds adoption. Importantly, this measure is only based on exogenous soil and weather characteristics and not on observed yields, which are endogenous to the technology adoption. Given that most of the corn cultivation in the Philippines is rain fed, data for this water source regime is used throughout the analysis. The variation used to identify the effect is therefore the potential increase in yields, which is assumed to be correlated (although not perfectly) with the actual yield gain.23 Although GM corn introduction is not the only explanation for the increasing corn yields over the period, it is the most important technological change and is therefore likely to have largely contributed to it. For the sake of readability, in the rest of the paper, when talking about the potential gain from GM corn, we are referring to the overall change in corn potential yield, which is largely driven by the new technology. Moving from low to high input level more than triples the potential corn yield on average, from 0.8 to 2.8 tons per hectare, with some areas gaining as much as 8 tons per hectare. This average value is lower than the average actual yields given that they are computed over the entire country, including the areas not suitable for corn cultivation. The geographical distribution of the potential gain in corn yield is presented in the first map of fig. 2. The estimation strategy can be formalized with the following first difference equation: yi = α + β Ai + γ1 Xi + γ2 Zi,1991 + i, (1) 22 The input variables for the GAEZ model include precipitation, number of rainy days, mean minimum, mean maximum temperature, diurnal temperature range, cloudiness, wind speed, and vapor pressure for climatic characteristics; organic carbon, pH, soil water-holding capacity, soil depth, cation exchange capacity of the soil and the clay fraction, total exchangeable nutrients, lime and gypsum content, sodium exchange percentage, salinity, textural class, and granulometry for the soil characteristics; slope gradients, aspects of slopes, and median elevation for terrain characteristics; land cover classes, and protected areas. 23 In a cross-country analysis, Alvarez and Berg (2019) show that potential yield is positively correlated with actual yield, especially so in the East Asia and Pacific region (R2 = 0.46). 238 Bequet Figure 2. Geographical Distribution of Potential Corn Yield Gain, Corn Cultivation, and GM Corn Adoption. Downloaded from https://academic.oup.com/wber/article/38/2/229/7499128 by Sectoral Library Rm MC-C3-220 user on 01 May 2024 Source: Data from the Food and Agriculture Organization Global Agro-Ecological Zones (FAO-GAEZ) Database (left). Philippines Census of Agriculture and Fisheries 2012 and Philippines Statistics Authority (PSA) (right). Note: The left-hand graph shows the geographic distribution of the difference in potential yield of rain-fed corn between high and low input levels. The right-hand graph shows the share of agricultural land devoted to corn in the CAF 2012 (background color) and the share of genetically modified corn, among cultivated corn, in 2014. Data on GM corn obtained from the Philippines Statistics Authority (PSA). where yi is the change in outcome variable between 2002 and 2012 for municipality i and Ai is the change of potential corn yield, and β is our coefficient of interest.24 Estimates of β have a causal explanation provided that changes in potential yields are independently distributed from the outcome variable once we control for all time-invariant characteristics and common shocks. If areas that benefited more from the technology were on different trends from those that benefited less, this assumption would be violated and the estimates would be biased. To alleviate this concern, the model includes time-invariant geographical controls Xi and socioeconomic indicators computed from the CAF 1991, Zi, 1991 . The variables Xi include the log of municipal area and, in some specifications, elevation, ruggedness, longitude, and latitude. Controlling for these last four variables is, however, problematic as they enter the formula used to compute the potential yield Ai . The interpretation of the coefficient β is therefore going to be different when they are included. On the other hand, excluding them may bias the estimates as they are correlated with other determinants of land-inequality trends, such as market access or the occurrence of natural disasters. In a robustness check (table S3.6), I show that the results hold when each variable is added individually. Trends in land inequality and technology adoption are likely to differ depending on baseline land scarcity. In frontier regions where new land can be cleared, we would expect lower agricultural productiv- 24 This equation is equivalent to a fixed effect model, with Ait taking the value under low level of inputs before 2003 and under high level of inputs after that. As the agricultural sector did not change from being completely traditional to being fully mechanized with the introduction of GM corn seeds, robustness tests are carried out using intermediate levels of inputs either in the pre- or in the post-adoption period. The World Bank Economic Review 239 ity and different land market dynamics compared to places where all the land is already under cultivation. For this reason, Zi, 1991 includes the share of total municipal area dedicated to agriculture in 1991. More- over, over the study period, corn prices have experienced a sharp increase, being multiplied by 3 between 2002 and 2012 (IMF 2021). This implies that regions where corn production is more widespread are on a different trend. As these regions are likely to be those with a high suitability, the share of corn in total Downloaded from https://academic.oup.com/wber/article/38/2/229/7499128 by Sectoral Library Rm MC-C3-220 user on 01 May 2024 agricultural area in 1991 is also included as control. Finally, night-light intensity in 1992 controls for a combination of initial population density and economic development.25 While adding these control variables alleviates the risk of a violation of the common trend assump- tion, it does not completely rule out the possibility that municipalities that benefited more from the new technologies were on a different trend from those that benefited less. To address this concern, I carry out robustness tests showing that the change in land inequality between 1991 and 2002 is uncorrelated to the potential yield gain and that the main result holds in a triple-difference setting (pooling the three waves of CAF data). 5. Results 5.1. First-Stage Effect As previously mentioned, the agricultural census does not distinguish between different corn varieties and does not provide output information. It is therefore impossible to test whether adoption and yield gains were higher in more suitable areas.26 To the best of my knowledge, the most disaggregated data on GM corn adoption, from the Department of Agriculture, are at the regional level and are available for the years 2003–2009 and 2014. The second map in fig. 2 presents the share of agricultural area devoted to corn in each region in the 2012 census (background color), along with the share of GM corn in 2014 (shaded bars). Adoption is particularly high in Luzon (north), which coincides with the high potential yield gain documented in the first map. In the Visayas (central archipelago) and Mindanao (south), on the other hand, adoption is almost non-existent.27 However, this is precisely where the illegal sige-sige seeds—allegedly originating from Southern Mindanao—are the most widespread. Adoption is therefore seriously underestimated in those regions. In any case, as GM corn represented almost two-thirds of all corn area by the end of the period, a positive correlation between GM corn adoption and potential yield is actually not needed to identify the effect. Assuming that the adoption rate was the same over the entire country—and therefore uncorrelated with crop suitability—we would still expect more suitable regions to be more impacted by the new technology. Corn cultivation relatively increased in regions that benefited more from GM seeds. Indeed, when we estimate equation (1) using the change in corn area and the share of agricultural land devoted to this crop, the coefficient of potential gain is positive and significant (table 2). The magnitude of the coefficients (with control variables) implies that a 1 standard deviation increase in potential yield leads to a 0.13 standard 25 Gibson, Olivia, and Boe-Gibson (2020) challenge the ability of night-light data to accurately measure economic devel- opment in rural areas. They show that these data are particularly unreliable when aggregated over small areas—due to blurring and overglow—and for temporal comparisons—because of satellite change and sensor adjustment to moon- light. Given that we aggregate the data at the municipality level and only use one cross-section, these concerns are unlikely to bias our results. Moreover, Gibson, Olivia, and Boe-Gibson (2020) show that night lights is more correlated with economic activity in urban areas, while sparsely populated rural areas remain dark even after electrification. Our municipality-level night-light measure therefore captures the development of the urban center and acts as a proxy for the local market. 26 Bustos, Caprettini, and Ponticelli (2016) are able to directly address this question and find that the soy potential yield gain is positively correlated with the change in GM soy area share and negatively correlated with the change in non-GM soy area share ( table 6 in their paper). 27 Luzon is the large island in the north, the Visayas the archipelago in the center, and Mindanao the large island in the south of the country. 240 Bequet Table 2. Productivity Change and Corn Cultivation Corn area (log) Corn share Farm area (log) Farm nb (log) Potential gain from GM corn 0.074∗∗∗ 0.120∗∗∗ 0.012∗∗∗ 0.012∗∗∗ −0.009 0.026∗ −0.031∗∗ −0.011 (0.027) (0.029) (0.003) (0.003) (0.013) (0.013) (0.012) (0.013) Municipality area (log) – −0.028 – −0.017∗∗∗ – 0.076∗∗∗ – 0.025 Downloaded from https://academic.oup.com/wber/article/38/2/229/7499128 by Sectoral Library Rm MC-C3-220 user on 01 May 2024 (0.041) (0.004) (0.019) (0.018) 1991 Agricultural area (share) – −0.402∗∗∗ – −0.073∗∗∗ – 0.133∗∗ – 0.105∗ (0.135) (0.014) (0.062) (0.061) 1991 Corn share – −0.054 – −0.060∗∗∗ – −0.133∗∗ – −0.049 (0.112) (0.013) (0.052) (0.047) 1992 Night lights (log) – −0.089∗∗∗ – −0.005∗∗∗ – −0.067∗∗∗ – −0.051∗∗∗ (0.027) (0.002) (0.012) (0.012) Observations 1,434 1,434 1,434 1,434 1,434 1,434 1,434 1,434 R-squared 0.006 0.020 0.019 0.090 0.000 0.047 0.006 0.029 Source: Data from the Philippines Census of Agriculture and Fisheries (CAF) and Food and Agriculture Organization Global Agro-Ecological Zones (FAO-GAEZ) database. Note: Changes in dependent variables are calculated over the years 2002 and 2012. Potential gain from Genetically Modified (GM) corn is the difference between potential rain-fed corn yield with high and low levels of inputs. The unit of observation is the municipality. Robust standard errors in parentheses. ∗∗∗ p < 0.01, ∗∗ p < 0.05, ∗ p < 0.1 Table 3. Productivity Change and Landholding Inequality Gini Share top decile Potential gain from GM corn 0.491∗∗ 0.459∗∗ 0.317 0.467∗∗ 0.101 0.509∗∗ 0.554∗∗ 0.117 (0.191) (0.210) (0.205) (0.210) (0.192) (0.213) (0.239) (0.209) Municipality area (log) – 1.074∗∗∗ 0.660∗∗ 1.058∗∗∗ 0.462∗ – 1.212∗∗∗ 0.573∗ (0.301) (0.302) (0.303) (0.280) (0.346) (0.310) 1991 Agricultural area (share) – 3.624∗∗∗ 2.891∗∗∗ 3.554∗∗∗ 2.995∗∗∗ – 2.737∗∗ 2.275∗∗ (0.964) (0.912) (0.970) (0.866) (1.104) (0.975) 1991 Corn share – 0.540 1.270 0.574 1.575∗ – 1.436 2.498∗∗ (1.044) (0.982) (1.048) (0.899) (1.191) (1.013) 1992 Night light (log) – 0.362∗∗ 0.728∗∗∗ 0.396∗∗ 0.687∗∗∗ – 0.418∗∗ 0.665∗∗∗ (0.183) (0.180) (0.187) (0.176) (0.203) (0.202) Farm area (log) – – 5.491∗∗∗ – 10.553∗∗∗ – – 12.075∗∗∗ (0.690) (0.889) (1.123) Nb farms (log) – – – 0.671 −7.439∗∗∗ – – −10.978∗∗∗ (0.586) (0.879) (1.142) Observations 1,434 1,434 1,434 1,434 1,434 1,434 1,434 1,434 R-squared 0.005 0.025 0.132 0.026 0.215 0.005 0.023 0.230 Source: Data from the Philippines Census of Agriculture and Fisheries (CAF) and Food and Agriculture Organization Global Agro-Ecological Zones (FAO-GAEZ) database. Note: Changes in dependent variables are calculated over the years 2002 and 2012 and expressed in percentage points (ranging from 0 to 100). Potential gain from Genetically Modified (GM) corn is the difference between potential rain-fed corn yield with high and low levels of inputs from the FAO-GAEZ. The unit of observation is the municipality. Robust standard errors in parentheses. ∗∗∗ p < 0.01, ∗∗ p < 0.05, ∗ p < 0.1 deviation increase in corn share, corresponding to an increase of 72 ha (or 1.5 percentage points) for the average municipality. Farmers therefore reacted differently to the new technology, depending on the soil and weather characteristics of their land. Note that this effect needs to be understood as a relative one as we are comparing municipalities that benefited more with those that benefited less. 5.2. Landholding Distribution Land inequality increased in municipalities that benefited more from GM corn seeds. As table 3 shows, the landholding Gini and the share of land occupied by top decile are positively correlated with the potential The World Bank Economic Review 241 yield gain.28 Results with control variables (columns 2 and 7) imply that a 1 standard deviation increase in potential yield leads to a 0.6 percentage point increase in the Gini index and a 0.7 percentage point increase in the top 10 percent share. These marginal effects are large considering that, between 2002 and 2012, the average municipality Gini increased by 0.6 percentage points and the average share of the top decile by 0.2 percentage points. Downloaded from https://academic.oup.com/wber/article/38/2/229/7499128 by Sectoral Library Rm MC-C3-220 user on 01 May 2024 Looking at the effect along the landholding distribution shows that the potential yield is positively correlated only with the share of land in the top decile and is negatively correlated with all other deciles. This is shown in fig. S4.2, which plots the potential yield coefficient obtained by replicating column 7 of table 3 using the land share of each decile as outcome variable. 5.2.1. Mechanisms: Farm Area and Farm Number A change in the land distribution is the result of three possible changes: (a) a reallocation of the previously farmed land between farmers, (b) an expansion/contraction of the farm area, and (c) a increase/decrease in the number of farms. These mechanisms are not mutually exclusive, as a new farm can encroach on new land, thereby increasing both number of farms and agricultural area. Total agricultural area did increase in more suitable regions, when baseline controls are included. The magnitude of the coefficient in column 6 of table 2 implies that a 1 standard deviation increase in potential productivity lead to a 3.2 percent increase in cultivated area, corresponding to 153 ha for an average municipality. However, as the general trend over the period is a contraction in agricultural land, this does not necessarily imply an expansion of agricultural land, as it is possible that the positive effect corresponds to a smaller decrease in farm area. This relative land expansion is not reflected in the number of farms, which is uncorrelated to potential gain from GM corn once baseline controls are included. Controlling for the change in agricultural area, the correlation between landholding Gini and potential yield gain decreases and becomes insignificant (column 3 of table 3). On the other hand, controlling for the change in number of farms has no effect, which was expected given its lack of correlation with potential yield gain. Adding both controls together further reduces the point estimate, which becomes statistically different from that of column 2 at the 10 percent level. This indicates that land reallocation between existing farmers does not play an important role and that the increase in land inequality is driven by municipalities that experienced a relative increase in agricultural land. In a context of massive agricultural land contraction characterizing the Philippines during the study period, this relative increase typically means a lower contraction in municipalities with higher potential gains. Indeed, as shown by fig. S4.3, only 23 percent of municipalities experienced an increase in agri- cultural area between 2002 and 2012. When re-estimating our main equation controlling for a set of categorical variables corresponding to the quintiles of the change in agricultural land area (table S3.2), it becomes clear that our results are driven by the municipalities with the largest land contraction. In- deed, the potential gain coefficient decreases strongly when controlling only for the first two quintile categories, i.e. when differentiating municipalities that experienced a strong decrease in agricultural land from the rest. In the opposite, controlling for the last two quintiles—those where there was no change or an increase—does not affect the coefficient of interest. To gain a better understanding of the interplay between potential gain, change in agricultural land, and change in land distribution, fig. 3 shows the average change in farm size at each decile of the distribution of the change in agricultural area. For municipalities with high land contraction (change in farm area below median, corresponding to a 24 percent contraction), there is a drastic decline in the size of farms, and this decline increases as we move along the distribution, with average farm size declining by 29 and 42 percent in the 90th and 99th percentiles respectively. When land contraction is lower (change in farm 28 To improve the readability of the table, the dependent variables are expressed in percentage points, i.e. ranging from 0 to 100 instead of 0 to 1. 242 Bequet Figure 3. Relative Change in Farm Size at Each Decile. Downloaded from https://academic.oup.com/wber/article/38/2/229/7499128 by Sectoral Library Rm MC-C3-220 user on 01 May 2024 Source: Data from the Philippines Census of Agriculture and Fisheries (CAF) 2002 and 2012. Note: Each dot shows the average log difference in farm size at a specific decile of the landholding distribution, between 2002 and 2012, averaged across municipalities. High land contraction restricts the sample to municipalities with a change of agricultural land below the median, low land contraction to those above the median but with land contraction, and no land contraction to those with an increase in agricultural land. area above median but still negative), the farm size decreases in every decile, but the decline at the end of the distribution is much smaller, around 12 and 15 percent for the top decile and percentile respectively. Finally, when agricultural land increases, the change in farm size is close to zero for most deciles and is even positive in the right-hand tail of the distribution. The trend of land contraction occurring during the study period was therefore characterized by a decrease in the size of large farms. As shown in table 2, the introduction of GM seeds slowed down this land contraction trend, which resulted in a relative increase in land inequality. 5.2.2. Crops, Population and Economic Development This section explores (and rules out) alternative mechanisms that could explain the positive correlation between agricultural productivity and land inequality, namely a change in crop mix, internal migration patterns, and economic development. First, corn is not a land-intensive crop and is mostly cultivated by smallholder farmers, which makes it unlikely that the increase in corn cultivation following the introduc- tion of GM seeds is driving our results. As shown in table S3.3, land inequality is negatively correlated with the share of land planted in corn and adding this control increases (insignificantly) the potential gain coefficient. Adding the change in land share for other common crops, such as rice, sugarcane, coconut, and banana, does not have a significant impact either. Second, the results are not driven by people migrating from low to high productivity municipalities or by differential trends in rural–urban migration within municipalities, as they are robust to controlling for the change in municipality population and the change in the share of rural population. Finally, the correlation between potential gain and land inequality is not the result of differences in economic develop- ment, as controlling for the change in night-light intensity between 2002 and 2012 leaves our coefficient of interest unchanged. Adding all the additional controls in a single regression leads to similar results. The World Bank Economic Review 243 5.2.3. Modern Inputs and Credit Penetration Turning the potential yield gains from GM corn into actual yield gains requires specific conditions. First, cultivating GM corn is more expensive than traditional varieties due to higher seed prices and larger returns on fertilizer, making the use of financial services more prevalent among its farmers. It is therefore expected that the effect will be larger where credit availability is stronger. To test this hypothesis, the share Downloaded from https://academic.oup.com/wber/article/38/2/229/7499128 by Sectoral Library Rm MC-C3-220 user on 01 May 2024 of farmers in 1991 who contracted a credit (formal or informal) over the preceding year is used as a proxy for financial development. This variable is not available in subsequent CAF rounds, but gives a measure of credit penetration 10 years before the introduction of GM seeds. Second, a stronger effect can also be expected in areas with more traditional agricultural systems, i.e. where the potential for improvement was the largest. The degree of agricultural development is proxied by the share of farmers in 1991 who were using high-yielding varieties (HYVs). Results show that the correlation between potential yield and land inequality is indeed stronger in municipalities with more credit and fewer HYVs in 1991 (table S3.4). Given that these two measures are positively correlated and have opposite effects on inequality, it is important to include them both in the same regression. The magnitude of the interaction coefficient implies that, for a municipality with average potential yield gain, the marginal effect of potential gain decreases by 0.5 percentage points when past HYV use increases by 10 percentage points and increases by 0.5 percentage points when credit availability increases by 10 percentage points. 5.2.4. Geographical Heterogeneity Heterogeneous effects based on the location of municipalities can help us gain a better understanding of the geographical distribution of the main effect. First, the correlation between agricultural productivity and land inequality is only significant for coastal municipalities (table S3.4), which are likely to be different from interior municipalities on many levels (exposure to climate events, transportation, communication, etc.).29 This differential effect disappears when controlling for the change in farm area and in farm number. Second, when different effects by island groups or regions are allowed, the correlation between po- tential yield gain and land inequality is only significant in the island of Mindanao (south) and is absent from the main corn-growing regions of Luzon in the north (fig. S4.4). This is not surprising given that Mindanao is the poorest region of the country, with low penetration of modern agricultural inputs and therefore the largest potential for yield increases. 5.3. Land Ownership Inequality The analysis so far has focused on landholding inequality, i.e. computing the land distribution using oper- ated farm as the basic unit. However, land ownership inequality is also an important measure as it is more closely linked to wealth and poverty. Due to data constraints, it is impossible to repeat the analysis using the same type of inequality measures for land ownership. Instead, we can look at the share of land that is not owned by the household cultivating it. An increase in that measure indicates an increase in land ownership inequality given that land ownership tends to be less equally distributed than landholding. Similarly, an increase in the share of tenanted farms also indicates more ownership inequality (among all farmers). Table 4 presents the results obtained by estimating equation (1) using the two aforementioned land ownership measures as dependent variables. The share of tenanted land decreases in municipalities that benefited more from the technology, although this effect loses some significance once control variables are added (p-value = 0.14). The share of tenanted farms shows similar results, with a positive correlation with the potential gain that becomes insignificant once the control variables are added (p-value = 0.19). Overall, this suggests that the increase in landholding inequality is reflected in the land ownership distri- 29 The interaction between potential gain and coastal is, however, not statistically different from that of potential gain with interior municipalities. 244 Bequet Table 4. Productivity Change and Land Ownership Inequality Share tenanted land Share tenanted farms Potential gain from GM corn −0.492∗ −0.465 0.714∗∗ 0.464 (0.287) (0.315) (0.321) (0.353) Municipality area (log) – −1.383∗∗∗ – −2.408∗∗∗ Downloaded from https://academic.oup.com/wber/article/38/2/229/7499128 by Sectoral Library Rm MC-C3-220 user on 01 May 2024 (0.448) (0.418) 1991 Agricultural area (share) – −5.242∗∗∗ – −8.212∗∗∗ (1.577) (1.533) 1991 Corn share – 0.567 – −3.251∗∗ (1.483) (1.426) 1992 Night light (log) – −0.284 – −0.047 (0.291) (0.286) Observations 1,434 1,434 1,434 1,434 R-squared 0.002 0.016 0.005 0.053 Source: Data from the Philippines Census of Agriculture and Fisheries (CAF) and Food and Agriculture Organization Global Agro- Ecological Zones (FAO-GAEZ) database. Note: Changes in dependent variables are calculated over the years 2002 and 2012 and expressed in percentage points (ranging from 0 to 100). Potential gain from Genetically Modified (GM) corn is the difference between potential rain-fed corn yield with high and low levels of inputs from the FAO-GAEZ. The unit of observation is the municipality. Robust standard errors in parentheses. ∗∗∗ p < 0.01, ∗∗ p < 0.05, ∗ p < 0.1 bution as a smaller proportion of farms own a larger (or similar) share of the land. This may be driven by the relative land expansion in the last decile of the landholding distribution. 6. Robustness Tests This section presents additional robustness tests for the main result. First, it tests for a violation of the common trend assumption using data from 1991 and 2002, before the introduction of GM seeds. Second, it adds topo-geographical controls to the main equation. Third, it shows that the results are robust to correcting standard errors for spatial correlation. Fourth, it replicates the analysis at the barangay (village) level, and fifth it uses alternative measures of potential corn yield and matching on observables. 6.1. Pre-treatment Trends One of the key identifying assumptions in our estimation strategy is that trends in land inequality are uncorrelated with the potential yield gain once we control for municipality area and pre-determined variables. This would be violated if previous productivity growth had already put more profitable areas on different trends. For example, the potential yield might be capturing the returns to land consolidation, and the increase in land inequality would have happened even without the introduction of GM seeds. If this was the case, we should observe a similar trend before the introduction of the technology. However, re-estimating our main equation using data from 1991 and 2002 fails to find any effect of potential yield gain on landholding inequality (table S3.5). On the other hand, the common trend assumption is violated for land ownership inequality as municipalities with higher potential yield gain experienced a decrease in the share of tenanted land and in the share of farmers who do not own their land prior to the introduction of GM seeds. This violation goes against the positive (and marginally significant) effect found between 2002 and 2012. To combine the main results with those from the placebo test, the three CAF waves from 1991, 2002, and 2012 are pooled in a triple-difference regression (table S3.6).30 For landholding inequality, the effect 30 Contrary to what we have done so far, municipality-level measures of land inequality are not computed from the same set of barangays given the sampling method of the CAF 1991 and 2002. The World Bank Economic Review 245 of potential gain is insignificant between 1991 and 2002 and positive and significant between 2002 and 2012. Pooling all the data also allows to control for different trends at the municipality level through municipality fixed effects. When they are added (columns 2 and 4), the effect of potential gain in the second period remains positive and statistically significant. This triple-difference estimation also confirms the signs of an increase in land ownership inequality, as the potential gain in the second period is positively Downloaded from https://academic.oup.com/wber/article/38/2/229/7499128 by Sectoral Library Rm MC-C3-220 user on 01 May 2024 correlated with the share of farms under tenancy and is uncorrelated with the share of tenanted land. The same amount of land is therefore distributed among a larger number of tenants. 6.2. Topo-geographical Characteristics The empirical strategy used in this paper relies on a measure of potential yield gain, which is computed using soil and weather characteristics. However, these characteristics may affect the trend in land inequal- ity through other channels than land productivity. For example, elevation and ruggedness determine the availability of transport infrastructure and therefore input availability and market access. Similarly, ex- treme weather patterns affect the accumulation of physical capital, with consequences for the trend in economic development. On the one hand, omitting these variables from the regression, as has been the case so far, might bias our estimates. On the other hand, if we control for them, the potential gain variable loses part of its substance and it is not clear how to interpret the coefficients. To address this issue, the main regression is re-estimated, adding topo-geographical control variables (table S3.7). In order to keep some informational value in the potential gain variable, the controls are added individually in each regression. When the average elevation and the ruggedness index are included, the point estimate becomes larger and more significant, indicating that, if anything, the omitted variable bias was pushing our coefficient downwards. Likewise, the result is robust to controlling for longitude and latitude, which are strongly correlated with weather patterns, especially extreme weather, since tropical cyclones hit the northern half of the country on a yearly basis while missing almost systematically the southern part. Finally, allowing different trends in land inequality for each province through fixed effects marginally decreases the point estimate and its precision, thereby decreasing its significance (p-value = 0.156). 6.3. Spatial Correlation Given that soil and weather characteristics are not distributed randomly over the country, potential corn yield is likely to exhibit some level of spatial auto-correlation. Not taking this into account leads to an underestimation of standard errors, thereby increasing the probability of excluding the null hypothesis when we should not. For this reason, table S3.8 reports the p-value obtained when re-estimating our main results with alternative clustering techniques. The first row shows the p-values obtained from the robust standard errors that have been reported so far. The second and third rows present p-values after the correction suggested by Conley (1999) using a 25 km and a 50 km radius and the last row when standard errors are clustered at the provincial level. When control variables are not included, the coefficients remain below the 5 percent threshold with the 25 km radius and below the 10 percent with the 50 km radius. When controls are included, p-values are larger, but always remain below 15 percent. Provincial-level clustering yields standard errors somewhere between the two radius values. 6.4. Barangay-Level Analysis Due to the geographical characteristics of the country, the level of within-municipality heterogeneity in the Philippines tends to be high. For example, the median municipality area is equal to 119 km2 and the median elevation range (difference between highest and lowest altitude) is 543 m, reflecting the hilliness of the country. Similarly, the within-municipality (between-village) standard deviation in potential yield is equal to 0.65 on average, which corresponds to half of the standard deviation computed between municipalities. Given this heterogeneity, we cannot be sure that the increase in land inequality is actually 246 Bequet observed in areas that became more productive or is the result of spillover effects coming from nearby areas. To address this issue, the analysis is replicated at the barangay level. Before interpreting the results, it is important to recall the differences between barangay- and municipality-level data. First, the plot phys- ical location is only available at the municipality level in 2002. Barangay land-inequality measures are Downloaded from https://academic.oup.com/wber/article/38/2/229/7499128 by Sectoral Library Rm MC-C3-220 user on 01 May 2024 therefore computed on the total land cultivated by people living in the barangay, not on the land located within its boundaries. While both sets are the same in most cases, large farms straddling administrative boundaries and absentee landlords will create a wedge between them. It is therefore possible to have a value of agricultural area larger than the total barangay area, which was not the case in the municipality data. Second, due to the sampling method used in the successive rounds of the CAF, the number of obser- vations will vary depending on the variables included in the analysis. More specifically, when controlling for 1991 variables, the number of observations will strongly decline, restricted to the balanced sample over the three waves.31 Third, while municipalities with less than 50 ha of agricultural land were excluded from the analysis, this threshold is decreased to 10 ha for barangays. Once again, this avoids taking into account areas where farming is a marginal activity. Finally, while most municipalities comprise both ur- ban and rural areas, barangays usually fall into only one of those categories. Given that agricultural land inequality is not a relevant issue in urban areas, it makes sense to restrict the sample to rural barangays only. Results from table S3.9 are remarkably similar to the main results, especially when we restrict the analysis to rural barangays: in those, a 1 standard deviation increase in potential yield leads to a 1.1 percentage point increase in Gini coefficient and a 0.6 percentage point increase in the top decile land share. 6.5. Alternative Measures of Potential Yield and Specification The measure of potential gain from GM corn used so far was defined as the difference between the potential corn yield with high and low levels of inputs. The high level corresponds to optimal modern agricultural practices, while the low level corresponds to traditional practices with no external inputs. The agricultural sector in the Philippines, however, did not change from being completely traditional to being fully mechanized over the decade 2002–2012 and the introduction of GM seeds can certainly not account for such a drastic change. As an additional robustness test, alternative measures of potential gain from GM corn are used, re- computing it using the potential yield with intermediate levels of inputs either in the pre- or in the post- adoption period. The first four columns of table S3.10 present the results when it is defined as the differ- ence between intermediate and low levels of inputs, and columns 4 to 8 when it is defined as the difference between high and intermediate levels of inputs. Results are in line with those presented in the rest of the paper. When using the difference in yield between intermediate and low levels of inputs, however, the effect is much less precisely estimated and becomes insignificant. This is the result of the lower variation in potential gain with this definition: the standard deviation is 0.35 compared to 1.27 when we take the difference between high and low levels of inputs. This decreases the statistical power of the analysis, leading to a non-rejection of the null hypothesis, although the point estimates are twice as large as in the baseline regressions. Finally, the last two columns show that the results are robust to using matching techniques. Given the non-binary nature of our treatment, the generalized propensity score adjustment for the dose–response function proposed by Bia (2008) is used. 31 See supplementary online appendix S1 for the details regarding the sampling structure and the weights recomputation. The World Bank Economic Review 247 7. Conclusions This paper shows that gains in corn productivity are an important factor explaining the evolution of land inequality in the Philippines during the first decade of the twenty-first century, following the introduction of genetically modified corn seeds. Results show that municipalities that benefited more from this tech- nology experienced a relative increase in landholding Gini and in the share of land occupied by the farms Downloaded from https://academic.oup.com/wber/article/38/2/229/7499128 by Sectoral Library Rm MC-C3-220 user on 01 May 2024 in the top decile. This appears to be driven by a smaller contraction of agricultural land: in a context of overall land contraction driven by a reduction in the size of the largest farms, GM seeds allowed more land to remain under cultivation, thereby slowing down the decline in cultivated area and the resulting decline in land inequality. Several heterogeneous effects are identified and robustness checks are carried out. First, the effect is stronger in places where agricultural credit transactions were widespread and improved seeds were less used in 1991. Second, the effect is heterogeneous across regions, although it is present on all major island groups. Third, it does not appear to be driven by migration between municipalities or by rural–urban migration within municipalities. Fourth, it is not present in the decade preceding the introduction of GM corn. One potentially important aspect that is overlooked in this paper is the implementation of the CARP land reform, which took place over the entire study period. This omission is the result of a complete lack of data regarding the amount of land redistributed at a disaggregated level. Given the high level of redis- tribution reported by the government, this may pose a threat to the validity of the results if landlords’ opposition to the process was stronger in regions that benefited more from the new technology. However, given that the CARP started in the 1990s, landlords would have needed to anticipate the arrival of the technology in order to keep their land until 2002. If this was the case, we would observe a similar effect between 1991 and 2002, which is not the case, as reported in table S3.5. Such a political economy expla- nation is therefore unlikely to be driving the results. Moreover, the actual amount of land redistributed by the policy remains largely uncertain given that official statistics appear unrealistically high. The welfare effect of the relative increase in inequality remains an open question, as our results are driven by a smaller contraction in agricultural land in more affected municipalities rather than by land consolidation per se. As a result, it is not clear whether the negative impacts of land inequality docu- mented in the literature apply to this context. Assessing the overall welfare effect of the new technology is beyond the scope of this paper and would require taking into account many additional outcomes, such as agricultural production, farmer incomes, and environmental side effects resulting from the change in pesticide use. Finally, while the empirical analysis uses an exogenous variation in profitability to identify the effect of the new technology, agricultural data needed to go beyond the reduced-form equations are lacking. The change in agricultural land is identified as an important mechanism and, while there exists a large literature on how agricultural productivity affects the size of the cultivated area, how land distribution is affected by land expansion/contraction has received less attention. Detailed information on productivity and profitability over a long enough period would be needed to better understand this connection. Data availability statement The data and code used for the analysis can be accessed at https://www.dropbox.com/scl/fo/ mv5lks2vhd2v62z10pep9/h?rlkey=oydkpolc845tixec736humvn2&dl=0 Household-level data used to construct tables 1, S2.1 and S3.1 are proprietary and cannot be made public. REFERENCES Adamopoulos, T., and D. Restuccia, 2020. “Land Reform and Productivity: A Quantitative Analyisis with Micro Data.” American Economic Journal: Macroeconomics 12(3): 1–39. 248 Bequet Aldemita, R. R., M. M. C. A. Villena, and C. James, 2014. “Biotech Corn in the Philippines: A Country Profile.” Los Baños, Laguna: International Service for the Acquisition of Agri-biotech Applications (ISAAA) and Southeast Asian Regional Center for Graduate Study and Research in Agriculture - Biotechnology Information Center (SEARCA BIC). Alesina, A., and D. Rodrik, 1994. “Distributive Politics and Economic Growth.” Quarterly Journal of Economics Downloaded from https://academic.oup.com/wber/article/38/2/229/7499128 by Sectoral Library Rm MC-C3-220 user on 01 May 2024 109(2): 465–90. Alvarez, J., and C. Berg, 2019. “Crop Selection and International Differences in Aggregate Agricultural Productivity.” IMF Working Papers 19(179). Anbarci, N., M. Escaleras, and C. A. Register, 2005. “Earthquake Fatalities: The Interaction of Nature and Political Economy.” Journal of Public Economics 89(9–10): 1907–33. Ballesteros, M. M., J. Ancheta, and T. Ramos, 2017. “The Comprehensive Agrarian Reform Program after 30 Years: Accomplishments and Forward Options.” PIDS Discussion Paper Series (No. 2017-34). 2017(34). Banerjee, A., and L. Iyer, 2005. “History, Institutions, and Economic Performance: The Legacy of Colonial Land Tenure Systems in India.” American Economic Review 95(4): 1190–213. Bardhan, P. K., 1974. “Inequality of Farm Incomes: A Study of Four Districts.” Economic and Political Weekly 9(6/8): 301–3. Bardhan, P., M. Luca, D. Mookherjee, and F. Pino, 2014. “Evolution of Land Distribution in West Bengal 1967-2004: Role of Land Reform and Demographic Changes.” Journal of Development Economics 110: 171–190. Baten, J., and R. Hippe, 2018. “Geography, Land Inequality and Regional Numeracy in Europe in Historical Perspec- tive.” Journal of Economic Growth 23(1): 79–109. Bauluz, L., Y. Govind, and F. Novokmet, 2020. “Global Land Inequality.” halshs-03022360. Bequet, L., 2020. “Biotech Crops, Input Use and Landslides Case Study of Herbicide Tolerant Corn in the Philippine Highlands.” Ecological Economics 177: 106773. Berg, C., B. Blankespoor, M. S. Emram, and F. Shilp, 2023. “Does Market Integration Increase Rural Land Inequality?” World Bank Policy Research Working Paper 10286. Besley, T., and R. Burgess, 2000. “Land Reform, Poverty Reduction, and Growth: Evidence from India.” Quarterly Journal of Economics 115(2): 389–430. Bia, M., 2008. “A Stata Package for the Estimation of the Dose-Response Function through Adjustment for the Gen- eralized Propensity Score.” Stata Journal 8(3): 354–73(20). Binswanger-Mhkize, H. P., C. Bourguignon, and R. van den Brink, 2009. Agricultural land redistribution: toward greater consensus. World Bank Publications. Borras, S. M. J., 2006. “The Philippine Land Reform in Comparative Perspective: Some Conceptual.” Journal of Agrarian Change 6(1): 69–101. Borras, S. M. J., D. Carranza, and J. C. Franco, 2007. “Anti-poverty or Anti-poor? The World Bank’s Market-Led Agrarian Reform Experiment in the Philippines.” Third World Quarterly 28(8): 1557–76. Bustos, P., B. Caprettini, and J. Ponticelli, 2016. “Agricultural Productivity and Structural Transformation: Evidence from Brazil.” American Economic Review 106(6): 1320–65. Catacora-Vargas, G., P. Galeano, S. Zanon Agapito-Tenfen, D. Aranda, A. T. Palau, and R. O. Nodari, 2012. “Soy- bean Production in the Southern Cone of the Americas: Update on Land and Pesticide Use.” Cochabamba: GenØk/UFSC/REDES-AT/BASE-Is. Ceddia, M. G., 2019. “The Impact of Income, Land, and Wealth Inequality on Agricultural Expansion in Latin Amer- ica.” Proceedings of the National Academy of Sciences of the United States of America 116(7): 2527–32. Chaudhry, M. G., 1982. “Green Revolution and Redistribution of Rural Incomes: Pakistan’s Experience.” Pakistani Development Review XXI(3): 117–123. Cinnirella, F., and E. Hornung, 2016. “Landownership Concentration and the Expansion of Education.” Journal of Development Economics 121: 135–52. Cipollina, M., N. Cuffaro, and G. D’Agostino, 2018. “Land Inequality and Economic Growth: A Meta-Analysis.” Sustainability 10(12): 4655. Conley, T. G., 1999. “GMM Estimation with Cross Sectional Dependence.” Journal of Econometrics 92(1): 1–45. De Jonge, B., R. Salazar, and B. Visser, 2021. “How Regulatory Issues Surrounding New Breeding Technologies Can Impact Smallholder Farmer Breeding: A Case Study from the Philippines.” Plants People Planet 4(1): 96–105. de Luca, G., and P. G. Sekeris, 2012. “Land Inequality and Conflict Intensity.” Public Choice 150(1–2): 119–35. The World Bank Economic Review 249 Deininger, K., and L. Squire, 1998. “New Ways of Looking at Old Issues: Inequality and Growth.” Journal of Devel- opment Econ 57(2): 259–87. Easterly, W., 2007. “Inequality Does Cause Underdevelopment: Insights from a New Instrument.” Journal of Devel- opment Economics 84(2): 755–76. Fort, R., 2007. “Land Inequality and Economic Growth: A Dynamic Panel Data Approach.” Agricultural Economics Downloaded from https://academic.oup.com/wber/article/38/2/229/7499128 by Sectoral Library Rm MC-C3-220 user on 01 May 2024 37(2–3): 159–65. Foster, A. D., and M. R. Rosenzweig, 2022. “Are There Too Many Farms in the World? Labor-Market Transaction Costs, Machine Capacity and Optimal Farm Size.” Journal of Political Economy 130(3): 636–80. Freebairn, D. K., 1995. “Did the Green Revolution Concentrate Incomes? A Quantitative Study of Research Reports.” World Development 23(2): 265–79. Galor, O., O. Moav, and D. Vollrath, 2009. “Inequality in Landownership, the Emergence of Human-Capital Promot- ing Institutions, and the Great Divergence.” The Review of economic studies 76(1): 143–79. Ghatak, M., and S. Roy, 2007. “Land Reform and Agricultural Productivity in India: A Review of the Evidence.” Oxford Review of Economic Policy 23(2): 251–69. Gibson, J., S. Olivia, and G. Boe-Gibson, 2020. “Night Lights in Economics: Sources and Uses.” Journal of Economic Surveys 34(5): 955–80. Guereña, A., 2016. “Unearthed: Land, Power and Inequality in Latin America.” Oxfam International. Guereña, A., and M. Wegerif, 2020. Land Inequality Trends and Drivers Land 9(4): 101 IMF, 2021. “Primary Commodity Price System.”https://data.imf.org/?sk=471dddf8- d8a7- 499a- 81ba- 5b332c01f8b9. ISAAA, 2017. “Global Status of Commercialized Biotech/GM Crops in 2017: Biotech Crop Adoption Surges as Eco- nomic Benefits Accumulate in 22 Years.” ISAAA brief 53: 25–26. Jayne, T. S., J. Chamberlin, L. Traub, N. Sitko, M. Muyanga, F. K. Yeboah, W. Anseeuw, et al., 2016. “Africa’s Changing Farm Size Distribution Patterns: The Rise of Medium-Scale Farms.” Agricultural Economics 47(S1): 197–214. Keswell, M., and M. R. Carter, 2014. “Poverty and Land Redistribution.” Journal of Development Economics 110: 250–61. Lanzona, L. A., 2019. “Agrarian Reform and Democracy: Lessons from the Philippine Experience.” Millennial Asia 10(3): 272–98. Lowder, S. K., J. Skoet, and T. Raney, 2016. “The Number, Size, and Distribution of Farms, Smallholder Farms, and Family Farms Worldwide.” World Development 87: 16–29. McDoom, O. S., C. Reyes, C. Mina, and R. Asis, 2019. “Inequality between Whom? Patterns, Trends, and Implications of Horizontal Inequality in the Philippines.” Social Indicators Research 145(3): 923–42. Neves, P. C., Ó. Afonso, and S. T. Silva, 2016. “A Meta-Analytic Reassessment of the Effects of Inequality on Growth.” World Development 78: 386–400. Ortiz-Becerra, K., 2021. Land Consolidation and Rural Labor Markets: Theory and Evidence from Colombia, PhD thesis. Davis, CA, USA: University of California Davis. Otsuka, K., V. Cordova, and C. C. David, 1992. “Green Revolution, Land Reform, and Household Income Distribution in the Philippines.” Economic Development and Cultural Change 40(4): 719–41. Peters, P. E., 2004. “Inequality and Social Conflict over Land in Africa.” Journal of Agrarian Change 4(3): 269–314. Phélinas, P., and J. Choumert, 2017. “Is GM Soybean Cultivation in Argentina Sustainable?” World Development 99: 452–62. Prahladachar, M., 1983. “Income Distribution Effects of the Green Revolution in India: A Review of Empirical Evi- dence.” World Development 11(11): 927–44. Raju, V. T., 1976. “Impact of New Agricultural Technology on Farm Income Distribution in West Godavari District, India.” American Journal of Agricultural Economics 58(2): 346–350. Reyes, C. M., 2002. “Impact of Agrarian Reform on Poverty.” PIDS Discussion Paper Series. 2002(2) Reyes, C. M., A. D. Tabuga, R. D. Asis, and M. B. G. Datu, 2012. “Poverty and Agriculture in the Philippines: Trends in Income Poverty and Distribution.” PIDS Discussion Paper Series 2012(09). Sant’Anna, A. A., 2016. “Land Inequality and Deforestation in the Brazilian Amazon.” Environment and Development Economics 22(1): 1–25. Thomson, H., 2016. “Rural Grievances, Landholding Inequality, and Civil Conflict.” International Studies Quarterly 60(3): 511–19. 250 Bequet Villoria, N. B., D. Byerlee, and J. Stevenson, 2014. “The Effects of Agricultural Technological Progress on Deforesta- tion: What Do We Really Know?” Applied Economic Perspectives and Policy 36(2): 211–37. Vollrath, D., 2007. “Land Distribution and International Agricultural Productivity.” American Journal of Agricultural Economics 89(1): 202–16. World Bank, 2009. “Land Reform, Rural Development, and Poverty in the Philippines: Revisiting the Agenda - Tech- Downloaded from https://academic.oup.com/wber/article/38/2/229/7499128 by Sectoral Library Rm MC-C3-220 user on 01 May 2024 nical Working Paper.” Washington DC: World Bank Group.