Geography, Institutions, and Global Cropland Dynamics

The Policy Research Working Paper Series disseminates the findings of work in progress to encourage the exchange of ideas about development issues. An objective of the series is to get the findings out quickly, even if the presentations are less than fully polished. The papers carry the names of the authors and should be cited accordingly. The findings, interpretations, and conclusions expressed in this paper are entirely those of the authors. They do not necessarily represent the views of the International Bank for Reconstruction and Development/World Bank and its affiliated organizations


Introduction
Cultivated areas have constantly been increasing since the beginning of agriculture about 10,000 years ago, reflecting the history of human settlements and their economic development (Ramankutty andFoley, 1999, Ellis, 2011).This trend has accelerated over the past three centuries, with a fivefold increase in the global amount of cropland between 1700and 1990(Goldewijk and Ramankutty, 2004).Over the recent decades, megatrends have emerged that will have a significant impact on the amount of land that is cropped globally and its distribution across and within countries.These megatrends include climate change which started half a century ago and is predicted to dramatically accelerate (IPCC, 2019), and the ongoing massive urban population growth in sub-Saharan Africa and Asia, which will likely continue over the next two decades (World Urbanization Prospects, 2018), increased demand for food from emerging economies such as China (van Dijk et al. 2021), and the rising global interest in land FDI since the 2008 commodity price boom (Arezki et al., 2015).
With the availability of satellite imagery products and increasing computational capacities, it has been possible to measure global cropland dynamics from space (Ramankutty and Foley, 1999, Kolb et al., 2013, Klimanova et al., 2018, Li et al., 2018, Liu et al., 2018, Song et al., 2018, Mousivand and Arsanjani, 2019, Vaclavik et al., 2013, Potapov et al., 2021, Winkler et al., 2021).Such studies offered the possibility of a consistent measure of global cropland, circumventing issues of national definitions and incompleteness of reporting when relying only on national statistics.They also make it possible to study spatial patterns of cropland dynamics (Haney and Cohen, 2015).
Satellite imagery assessments of cropland, however, are not exempt from issues ranging from measurement errors associated with spatial and temporal resolution and the use of different classification algorithms to identify cropland.In fact, depending on the satellite imagery product, measures of cropland can vary significantly (Gibbs et al., 2011).Potapov et al. (2021) argue that the MODIS product is the most appropriate given relatively high temporal resolution and ground validation.However, most satellite imagery studies remain descriptive in their nature, with only a few attempts at assessing the drivers of land cover change.The methodologies range from combining different datasets and discussing potential causality (Bren d'Amour, 2017) to modeling land cover changes based on Markov chains (Halmy, 2015, Mousivand and Arsanjani, 2019, Mansour et al., 2020), cellular automata applied to the land cover grid or a combination of the two (Yiulianto et al., 2019;Wang et al., 2019), and simulations in Agent-Based Models (Ligmann Zielinska et al., 2014, Magliocca et al., 2014, Valbuena et al., 2010).These models are computationally intensive and often only applied to specific regions.Even though these approaches identify correlations, none of them establish causality.These studies find that cropland transitions are correlated with important global trends, including climate change (Song et al., 2018), urbanization through encroachment on existing cropland (Bren d'Amour, 2017, Debolini et al., 2018), and demography through increased demand for agricultural production (Wirsenius et al., 2010, Mora et al., 2020).Some papers also discuss the role of institutions (De Beurs and Henebry, 2004) as well as economic globalization in driving land cover changes (Lambin and Meyfroidt, 2010), but these effects have not been studied empirically.
The goal of this paper is to address the above gaps by assessing the recent dynamics of cropland (i.e., measuring movements into and out of cropland) at the global scale and to causally assess the impacts of potential drivers and disentangle their respective contributions.To do that, we consider global cropland changes measured from space using the MODIS data product at a resolution of 500m * 500m over the 2003-2018 period.We empirically estimate a model of cropland changes accounting for the impact of biophysical, socio-economic, and institutional contexts.To our knowledge, our study is the first to embed a high-resolution analysis of global cropland change in an econometric setting.Closest to our approach is the paper by Haney and Cohen (2015), who developed a grid cell level model of cropland transitions explained by global historical trends and global population dynamics.However, our paper significantly departs from their approach given our consideration of pixel and national level determinants-as opposed to assessing how each pixel responds to global demographic dynamics-and given the breadth of the variables included in our analysis.More importantly, we use measures of actual cropland changes rather than modeled data.Our paper is as follows: In sections 1 and 2 below, we present the datasets used in our analysis as well as stylized facts regarding patterns of cropland distributions and transitions into and out of cropland.Section 3 details our empirical approach and discusses how we deal with potential endogeneity concerns.Section 4 presents the results, and Section 5 concludes.

Section 1. Data
We measure land cover at the scale of the world using satellite imagery.This provides a global and consistent measure of cultivated land.Our "direct" approach circumvents the issues that emerge with the aggregation of disparate national-level statistics for the whole world.Above all, it allows us to focus on actual cropland location, which would be impossible from national-level statistics as published in a database such as FAOSTAT.We use the MCD12Q1 product from Moderate Resolution Imaging Spectroradiometer (MODIS) available since 2001 (see Sulla-Menashe and Friedl, 2018), which we chose for its relatively high temporal resolution and high level of accuracy (Potapov et al., 2021).For the analysis, we regroup the 16 land cover classes based on the Annual University of Maryland (UMD) Classification into 8 categories.These categories include: water bodies, forests, shrub and grasslands, wetlands, croplands, cropland/natural vegetation mosaics, urban, and non-vegetated/barren.This gives an annual breakdown of land cover by pixels of approximately 500 meters by 500 meters for the whole world.In order to smooth away biases from potential misclassification, we allocated each pixel to the predominant class over these 3 consecutive years.For the analysis, we use a restricted sample that covers the years 2003-2018.Our main focus is on "cropland" defined as pixels that have at least 60 percent of their area cultivated.The category of "cropland mosaic" (pixels that have between 40 and 60 percent of cropland with natural tree, shrub, or herbaceous vegetation) is a hybrid category that can capture intermediate steps towards cultivation or degradation or mixed uses such as agricultural and residential at the periphery of cities. Figure A1.1 in the Appendix shows the 2018 global land cover by land cover category.
The MODIS dataset provides an estimate of global cropland of 1.2 billion hectares in 2018, which is very close to that of Potapov et al. (2021), who combined MODIS and Landsat to enhance the accuracy of cropland extent measurement.If we include cropland mosaics, we obtain a total mass of cultivated land of slightly more than 1.3 billion hectares.This is smaller than FAO's estimate of global "cropland" of 1.6 billion hectares based on a broader definition, which includes temporary meadows and land temporary fallow.This gives us confidence in focusing on the raw MODIS product and its cropland category for our analysis.Cropland mosaic is a hybrid category that can be considered a transitional category in the dynamic processes whereby land is put into or out of cultivation and thus needs to be treated separately for the purpose of studying cropland dynamics.
To study cropland dynamics, we also mobilize a set of biophysical variables at the same pixel level.As a measure of climate stress, we use the annual number of drought months reconstructed from the Palmer Drought Severity Index (PDSI) developed by TerraClimate (https://www.climatologylab.org/terraclimate.html),where drought months are defined as periods for which the index was below -3.We calculated the distance to rivers and coast using the shapefile from Natural Earth (https://www.naturalearthdata.com/). 1 We also calculated the distance to the nearest city using the UrbanPop dataset (Blankespoor et al., 2017).Land suitability for cultivation was obtained from the FAO Global Agro-Ecological Zones (GAEZ) v4 product (https://gaez.fao.org/).We constructed an aggregate index comprised between 0 and 100 that measures how suitable the pixel is for rainfed cultivation of an umbrella crop of seven major crops under a high input scenario (i.e., taking the maximum individual suitability index over these seven crops). 2e also use measures of institutional quality.This includes measures of land governance from two different datasets.The first is the Quality of Land Administration index (denoted QLA hereafter) from the World Bank Doing Business 2020 database.3QLA is a composite of five subsidiary indexes measuring the reliability of land administration infrastructure, the transparency of information, the geographic coverage of registries, the legal framework for land dispute resolution, and whether men and women have equal access to property rights.Alternatively, we use the Institutional Profile Database (CEPII, Agence Française Développement, and Ministère de l'Économie et des Finances: http://www.cepii.fr/institutions/en/ipd.asp) for 2012.The database provides several country-level measures of land governance, which we summarize as an index measuring "land tenure insecurity of rural populations". 4Finally, we use the 2017 overall indicator in the Enabling Business of Agriculture dataset (https://eba.worldbank.org/en/eba),which captures the ease for farmers to operate their businesses.We use the overall score that aggregates the following sub-indicators supplying seed, registering fertilizer, securing water, registering machinery, sustaining livestock, protecting plant health, trading food, and accessing finance.

Section 2. The global distribution of cropland and its dynamics
The MODIS product allows us to assess patterns of geographic cropland distribution and its evolution over time.Table 1 below provides a breakdown of global cropland by region (using the World Bank definition) at the beginning and the end of our study period, showing that 65 percent of global cropland is located in the Europe and Central Asia region and Asia.North America has over 15 percent of the stock of cropland.
The total land under cultivation and regional shares have been relatively stable over the 2003-2018 period, although concealing large movements in and out of cropland.5Cropland locational characteristics exhibit wide variations within and across regions.Appendix Figure A2.1 shows the regional distributions of cropland pixels in 2003 (the beginning of our study period) with respect to locational characteristics, including distance to the nearest city of 50,000 people or more, the number of drought months between 2003 and 2018, land suitability for agriculture, distance to the coast, and distance to the nearest river.These graphs show similar patterns across regions in terms of distance to rivers and distance to cities, capturing the fact that access to markets, the presence of infrastructures, and the potential for surface water irrigation are quasi-universal conditions for agriculture.For instance, in each region, 80 percent of 2003 cropland is found to be located close to cities (within 31.5 km of a city in the Middle East and North Africa, and within 142.5 km of a city in North America, possibly reflecting better transport networks in North America).There are, however, stark disparities regarding the other contextual variables: Regions are very differently exposed to drought, with North America facing a more favorable climate and the Middle East and North Africa being the most exposed to droughts.In the Middle East and North Africa, the median 2003 cropland cell was exposed to 56 months of droughts over the 2003-2018 period, compared to 11 months for the median cropland cell in North America.The distribution of cropland also varies a lot in terms of distance to coasts, with the mass of cropland in North America being far from coasts (the Great Plains) and the bulk of cropland in the Middle East and North Africa being close to coasts (as the rest of the landmass in the region is mostly barren).
As shown in Table 1 above, the total amount of cropland has been relatively stable in each region over the 15 years of the studied period with the exception of Latin America and the Caribbean which experienced a notable increase in cropland (see Graesser et al., 2015).This relative stability conceals large global and regional variations in both cropland gains and cropland losses, with a total of 116 million hectares of cropland converted to other uses over the period and 119 million hectares of new cropland.These quantities are equivalent to about 10 percent of the existing stock of cropland in 2003.Both movements in and out of cropland are observed for all regions.Latin America lost 10 million hectares of cropland and gained 22 million hectares representing quantities equivalent to 11 percent and 24 percent of 2003 cropland in the region.Sub-Saharan Africa also experienced large movements in and out of cropland with 22 million hectares lost and 20 million gained, representing 21 percent and 19 percent of the total amount of cropland in 2003. 6 These gains and losses, however, are not uniform over space.Regarding movements out of cropland, Figure 1 below represents their spatial distribution according to the same locational characteristics discussed in at the beginning of this section.The first graph in the figure (Panel A) shows that the rate of cropland loss is greater in proximity to the sea and far away from it.This could reflect a variety of factors, from the salinization of coastal land, to the urbanization of the seashore or arid climate and isolation from markets far from ports.Panel B shows that the rate of cropland loss is smaller close to rivers, possibly reflecting better soil quality and the possibility of maintaining irrigated agriculture in proximity to rivers.
Panel C shows that beyond 400 km away from a city, cropland starts to degrade at a much faster rate.This could possibly capture that cropland located outside the "catchment area" of cities (i.e., more than a few hours' drive) is more challenging to maintain.The number of drought months (Panel D) and the cell's cropland suitably (Panel E) play as expected, with the rate of cropland loss steadily increasing with droughts and decreasing with cultivation suitability.Movements into cropland also exhibit an interesting geographic pattern, with new cropland pixels clustering in already cropped locations.Figure 2 shows the count of new cropland pixels (i.e., transitioning to cropland) with distance to the nearest 2003 cropland pixel.Strikingly, 80 percent of movements into cropland occur within only 2.4 km of existing cropland, and this global pattern is almost identical in every region.This could be due to a number of concurring mechanisms, including the expansion of existing cropland, the presence of infrastructure and transport logistics near locations that already have cropland, proximity to markets, or land suitability as revealed by the presence of existing cropland in the same area.

Section 3. Empirical approach Identification strategy
As discussed in the introduction, the empirical literature on cropland dynamics has mostly resorted to descriptive approaches using Markov chains and explored causal mechanisms in calibrated computational models, with very few studies resorting to regression analysis.Our empirical strategy to identify and disentangle the determinants of cropland dynamics revolves around a linear probability model estimated in two stages.This two-stage procedure is separately applied to movements in and out of cropland and accounts for the drivers of cropland dynamics both at a relevant geographical scale (the 500x500 meter pixel) and at the national scale (for country-wide economic and institutional variables).
As regards movement out of cropland, we focus on the sample of 2003 cropland pixels and run the following pixel-level regression to capture the impact of country fixed effects on land degradation while controlling for geographical contexts.
where   is a dummy variable equal to 1 if pixel i (in country c) transitioned out of cropland between 2003 and 2018 and equal to 0 if it remained cropland.  are k explanatory variables for pixel  including Euclidian distance to the nearest city over 50,000 residents and other bio-physical conditions and spatial attributes, namely the Palmer Drought Severity Index, land suitability for our umbrella of major crops according to GAEZ, distance to coast, and distance to the nearest river (see descriptions in the data section above). () are country fixed effects. is a the constant and   are error terms.All explanatory regressions are mean-centered by country.
We voluntarily use a parsimonious specification for this equation where our variables are all likely to be exogenous.This is of course the case of all the biophysical variables since the larger geographical context determines them and is not determined by farmers' individual decisions whether or not to crop the land.Distance to the nearest city is also exogenous since city locations preexisted in the 2003-2018 period.In our parsimonious specification, we voluntarily do not control for the presence of infrastructure, which could be a source of endogeneity given potential reverse causality (for instance, if places that have stable cropland attract roads).Although this omission could be a source of omitted variable bias, the risk of such bias is greatly reduced by controlling for distance to the nearest city, which also proxies for the presence of infrastructure.
Following this first step, we recover the country-fixed effects and then run a country-level regression of those fixed effects on variables that capture the national context.The specification is: where   ̂ is the set of country fixed effects estimated in the first stage,   are the national-level variables for country , and   is the error term.We use four national variables, including GDP per capita in constant dollar terms, the Gini Index for 2018, cropland per capita for 2018, and the Doing Business Quality of Land Administration for 2020. 7We also experimented with two other institutional variables in lieu of the Doing Business variable, including a measure of Rural land tenure insecurity that we constructed using the 2012 Institutional Profile Database and the overall indicator in the 2017 Enabling Business of Agriculture dataset (see the data section above).
For movements into cropland, the approach is similar to assessing the determinants of movements out of cropland.The sample is different though as we need to consider the hundreds of millions of pixels that were not cropland in 2003, raising computational issues.To address the large dimensionality of the problem, we restrict ourselves to the sample of pixels within 3 kilometers of other pixels that were initially classified as cropland in 2003.There are 30.3 million such pixels, of which 4.29 million (covering 107.2 million hectares) transitioned to cropland between 2003 and 2018.This 3-kilometer buffer thus captures 90.1 percent of cropland gains over the period (119 million hectares).We then proceed to a similar 2stage approach for our assessment of the drivers of cropland loss.In the first stage, we run a regression similar to specification (1) where   is now a dummy variable equal to 1 if pixel i (in country c) transitioned into cropland between 2003 and 2018 and is equal to 0 if the pixel remained non-cropped.We consider the same set of explanatory variables   as before.The country-fixed effects are then recovered and regressed on the same national-level characteristics specified in (2).

Samples and descriptive statistics
For the first stage regression of cropland losses, we consider the universe of 48.3 million pixels that were initially under cropland in 2003.Appendix Table A3.1 shows global summary statistics for the five explanatory variables associated with these pixels, including distance to a city, a coast, and a river, the number of drought months over the studied period, and land suitability for agriculture.The figures in the table show the existence of large global variations for these variables.For the first stage regression of cropland gains, we consider the 30.3 million pixels located within 3 kilometers of a cropland pixel in 2003 but that were not cropland at the time.Appendix Table A3.1 also shows descriptive statistics for this sample for the same five explanatory variables.As expected, the latter sample is more "peripheral" and thus further from cities, coasts, and rivers, less suitable for agriculture, but it is only slightly more exposed to drought than the sample of 2003 cropland pixels used for the cropland loss regression.
For the second stage regression for both cropland losses and cropland gains, we use variables from a set of seven national-level explanatory variables presented in Appendix Table A3.2.These institutional and economic variables include the quality of land administration, an index of Rural land tenure insecurity, the indicator Enabling the business of agriculture, GDP per capita, the Gini index, and two measures of Cropland per capita, one for 2003 and one for 2018.Cropland per capita was calculated by dividing the country's aggregate amount of cropland for these years in the MODIS product by the corresponding national population for the same years.Note that some of these variables, however, are not available for all countries: Whereas cropland per capita can be calculated for more than 200 economies, the indicator from Enabling the business of agriculture is only available for 101 countries, potentially limiting the size of the sample in the second stage when introducing this indicator as an explanatory variable.Table A3.2 shows that these variable vary much across countries.

Section 4. Results
We ran the first stage regression of movements out of cropland on pixel-level measures capturing local contexts.Table 2 below reports the OLS estimates for regression (1) focusing on the conversion out of cropland.Given the large dataset (over 48 million observations), all coefficients are significant at the 1 percent level.Estimated coefficients are consistent with theoretical intuitions: Cropland far from cities is likely to disappear at a more intense pace than cropland close to cities.With cropland pixels distant by 60 km from the nearest city on average, an additional one standard deviation in the distance to a city (i.e., considering locations that are an additional 55 km away from a city) increases the likelihood of cropland loss by .8percent over the 15-year period.This could reflect greater investment incentives to keep land in production close to markets.Similarly, an increase by one standard deviation in the number of months of exposure to droughts (i.e., having 63 months of droughts instead of 38 over the period) increases the likelihood of cropland loss by 1.4 percent.A decrease by one standard deviation in cultivation suitability (i.e., a reduction of the suitability index from 84 to 66) increases the likelihood of cropland loss by 2.8 percent.An increase in one standard deviation in the distance to a river increases the probability of cropland loss by .3percent.Proximity to a river also decreases the likelihood of cropland loss, which could capture the role of rivers in keeping soils moist as well as the likelihood of surface water irrigation close to rivers: A decrease by one standard deviation in the distance to a river decreases the probability of cropland loss by .6 percent.Finally, proximity to the coast is harmful to keeping land under cultivation.A one standard deviation in the distance to the coast increases the likelihood of land degradation by .3percent.This could be potentially due to the salinization of coastal areas.We see that the factor which by far has the greatest impact on transitions out of cropland is the quality of soils (as measured by the cultivation suitability index), followed by climate events (as measured by droughts).Interestingly, the role of access to cities in keeping land cultivated is greater than that of proximity to rivers or distance to coasts.As a robustness check, we run the same regression on a sample including both cropland and cropland mosaics (53 million pixels) and consider movements out of either one of these two land use categories.This produces very similar results presented in Table A4.1, which all coefficients having the same sign and same order of magnitude.We then recover the country fixed effects from the above regression on the cropland sample and regress them on national level socio-economic and institutional variables as in specification (2).Given the small number of observations, we aim for a parsimonious specification, including cropland per capita-as a proxy for demographic pressure exerted on land utilization-an institutional variable, and controls for the macro-economic and socio-economic contexts.Table 3 below shows the coefficient estimates, starting with cropland per capita in column (1) and subsequently adding the Quality of Land Administration rate in column (2) as well as two controls in column (3): GDP per capita and the Gini Index.In columns ( 4) and ( 5), we substitute the Quality of Land Administration with other institutional variables, namely our index of Rural land tenure insecurity and the index for Enabling the business of agriculture, respectively.We do not include the institutional variables simultaneously due to both our concern to keep a parsimonious specification and the high correlation between these institutional variables.In Table 2, a negative sign is associated with a lower fixed effect and thus with a slower pace of cropland loss in regression (1).
For each variable, the sign of coefficients and the order of magnitude remain the same across specifications.Our preferred specification is column (3), which has the greater number of observations, although columns (3)-( 5) consistently point towards significant comparable effects for both cropland per capita and the institutional variable.Column (3) shows that in countries that have greater cropland per capita, cropland disappears at a slower pace: A one standard deviation decrease in cropland per capita leads to 5.8 percent increase in cropland losses over the study period.This can be due to increased demographic stress from the relative scarcity of cropland that could provide incentives for excessive cultivation.Interestingly, the Quality of Land Administration consistently plays a role in preventing land degradation: A one standard deviation increase in the Quality of Land Administration (i.e., comparing the quality of land administration in a country such as Poland versus that of a country such as Burkina Faso) results in the likelihood of cropland loss that is 5.2 percent lower over the period.Better land management or more stable investment in land likely follows from an efficient land registration system as measured by the QLA index, reducing the pace of cropland loss.Interestingly, inequality is also associated with faster cropland loss.Although we cannot test for this mechanism, one potential reason could be unsustaible land management and cultivation practices associated with poverty.The effect of Rural land tenure security is comparable to that of poor quality of the land administration, with a one standard deviation increase in Rural land tenure insecurity also leading to a 5.2 percentage point in the pace of cropland loss.Similarly, a one standard deviation increase in Enabling the business of agriculture decreases the pace of land loss by 5.0 percent.The effects of the demographic pressure and institutional variables are of the same order of magnitude as our controls: GDP per capita is consistently associated with a greater pace of cropland loss, although the effect is not significant in column ( 5) and is only significant at the 15 percent level in columns ( 3) and (4).Interestingly, inequality is associated with faster cropland loss at the 10 percent level in columns (3) and ( 5) and at the 5 percent level in column (4): A one standard deviation increases in inequality (i.e., comparing a country such as Morocco to a country such as the Republic of Korea) increases the pace of cropland loss by 3.0 to 4.2 percent, depending on the specification.This is consistent with previous studies that found a correlation between poverty and land degradation (Lambin et al., 2021).Interestingly, all those country-level effects seem to be significantly larger-and in some cases by an order of magnitude-than the effects of biophysical conditions presented in Table 1.We now turn to the assessment of the factors that drive movements into cropland, adopting the same two-stage procedure.For this, we focus on pixels within a 3-km distance of existing cropland in 2003 and estimate a regression similar to (1) but where the left-hand side variable is a dummy equal to 1 if the pixel transitioned to cropland in 2018 and 0 otherwise and where explanatory variables are the same as in Table 1.Results are presented in Table 4 below.All coefficients are highly significant, and the signs of these coefficients for distance to the city, the number of drought months, land suitability, and distance to the coast are opposite to what we reported in Table 1 for transitions out of cropland.Here, a one standard deviation decrease in proximity to a city is associated with an increased probability of 8.7 hundredths of a percent that the pixel would transition to cropland.The effect is in line with theory and previous work that showed that cropland expands in place with better market access (see Berg et al., 2017, for an assessment in Sub-Saharan Africa).Droughts are of course a major deterrent to putting land into cultivation, with a one standard deviation increase in the number of droughts reducing the likelihood of transition to cropland by 1.3 percent.Land suitability is a predictor of conversion to cropland, with a one standard increase in suitability leading to a pixel being 1.9 hundredths of a percent more likely to transition to cropland.Consistent with the result from Table 1, distance to coast increases the probability of transition to cropland with a one standard deviation increase in distance coast leading to an increase in the likelihood of cropland transition by 0.8 hundredth percent.The only surprising result is that distance to a river is associated with a greater likelihood of transition to cropland.This surprising result could be explained by our inability to for groundwater irrigation potential, for which there is no globally available data.We perform a robustness check by redefining the left-hand side variable to measure transitions into cropland or cropland mosaic.Table A4.2 presents the results for this extended definition of the land-use transition.Results are very similar to those in Table 4, with coefficients having the same sign and order of magnitude except for the coefficient on distance to a city which is noticeably greater in absolute value.8As before, we recover the country fixed effects from the first-stage regression and regress them on the same national-level variables (keeping only one institutional variable at a time).Table 5 presents the findings.Model (1) is the full specification where the institutional variable is the Quality of Land Administration.Models (2) and ( 3) respectively substitute Rural land tenure insecurity and Enabling the business of agriculture to the Quality of Land Administration.Model (3) is our preferred specification because it has the greatest fit, although it also has the lowest number of observations due to data limitations for Enabling the business of environment.The signs of the coefficients are consistent across all specifications, with cropland per capita consistently associated with greater movements into agriculture-possibly capturing countries that have greater specialization in agriculture-and with GDP per capita and the Gini index associated with lower movements into agriculture.This mirrors our results for land degradation, where national wealth and inequality were associated with faster cropland loss.As regards institutional variables, only Enabling the business of agriculture is loosely significant (at the 15 percent level) in Model (3): A one increase in the standard deviation of this indicator increases the pace of movement into agriculture by 2.1 percent.

Section 5. Conclusion
This paper, to our knowledge, is the first to try and characterize the dynamics of cropland using a global georeferenced dataset combined with both geographic characteristics and national-level economic and institutional contexts.Our two-stage approach allows us to disentangle the purely geographic effects at the granular local level from the national level contexts.Beyond the expected role played by local geographic characteristics (i.e., droughts driving land degradation) or access to markets (with proximity to cities stimulating conversion to cropland), a key contribution of this study is to show that national economic contexts and national institutions matter.The pressure on land resources as proxied by a low amount of cropland per capita can accelerate cropland loss.Similarly, weak land governance and economic inequality can also accelerate cropland loss, with both channels possibly percolating through agricultural practices that facilitate land degradation.Interestingly, we find that although geographic characteristics explain movements in and out of cropland, national-level variables are less able to explain movements into cropland.This could be due to the specific role of government policies in putting land into cultivation, which could not be captured with globally available data.
Our paper focuses on the extensive margin, i.e., on the expansion or shrinkage of land put into cultivation.
Other adjustments could nevertheless occur through the intensive margin, i.e., with the intensification of cultivation.Although data has become increasingly available to measure "greenness" or "biomass" (e.g., using the Normalized Vegetation Index or the Net Primary Product), such measures do not necessarily capture well cropland productivity and do not differentiate between vegetation and cropland, making it difficult to use them as proxies for cultivation intensity.A study of cultivation intensity would instead require a comprehensive dataset ideally accounting for crop type, the frequency of cultivation, and cultivation schedules that are very likely to vary across locations, which is presently unavailable at the global scale.
Even though this study shows that the global cropland mass is not undergoing drastic increases (a .25 percent increase over 15 years), there are significant movements in both directions that balance one another.With varying degrees of exposure to climate change, conflicts concentrating in some regions of the world, and uneven demographic changes, these movements could result in a reshuffling of global cropland.Identifying the factors driving these changes as we do in this paper is a first step to anticipating the direction of these movements and the role that policies can play in addressing these emerging challenges.

Figure 1 -
Figure 1 -The geographic distribution of cropland losses Panel A

Figure 2 -
Figure 2 -Cropland gains by distance to existing cropland

Table 2 . Drivers of cropland losses (pixel level)
Note: The regression is applied to all 2003 cropland pixels.The dependent variable takes value 1 if the pixel transitioned out of cropland over the 2003-2018 period and 0 otherwise.Standard errors are in parentheses.*** p<0.01, ** p<0.05, * p<0.1

Table 3 -National-level drivers of cropland loss Dependent variable: Fixed effect recovered from the pixel-level cropland loss regression
Note: The explained variable in these regressions are the fixed effects recovered from the pixel-level regression of cropland losses.Cropland per capita is from MODIS 2018; the Quality of Land Administration is from Doing Business 2020; GDP per capita is in current US$, the Gini Index uses the latest value reported in the World Development Indicators.Standard errors in parentheses.*

Table A4 .1. Drivers of cropland losses
The regression is applied to all 2003 cropland and cropland mosaic pixels.The dependent variable takes value 1 if the pixel transitioned out of cropland or cropland mosaic over the 2003-2018 period and 0 otherwise.Standard errors are in parentheses.*** p<0.01, ** p<0.05, * p<0.1 Note:

Table A4 .2. Drivers of cropland gains
Note: The results are for a regression that is restricted to the sample of pixels that were neither cropland nor cropland mosaic in 2003 but are located within 3 kilometers of existing 2003 cropland.The dependent variable is equal to 1 if the pixel became cropland in 2018 and 0 otherwise.Standard errors are in parentheses.*** p<0.01, ** p<0.05, * p<0.1