Geography and Agricultural Productivity: Cross-Country Evidence from Micro Plot-Level Data

Transcription

1 Geography and Agricultural Productivity: Cross-Country Evidence from Micro Plot-Level Data Tasso Adamopoulos York University Diego Restuccia University of Toronto February 2015 Abstract The disparity in agricultural output per hectare (land productivity or yield) between the richest and poorest 10 percent of countries is a factor of more than 3-fold. What explains low land productivity in agriculture in poor countries? Is it natural disadvantage such as low land quality and unfortunate geography in poor countries or misguided economic choices given their geography? We study these questions using high-resolution micro-geography data from the Global Agro-Ecological Zones project (GAEZ). The data provides not only actual yields for the crops produced in each plot of land covering the entire globe, but also potential yields for all the main crops in each plot. Potential yields account for soil quality, climate conditions, and terrain topography. We develop a simple spatial-accounting framework that allows us to aggregate up from the plot-crop level resolution to the country level. Our main finding is that low agricultural productivity in poor countries is not due to poor land quality and geography. If both rich and poor countries produced according to their potential yield given their internal dispersion of land quality, the rich-poor agricultural yield gap would virtually disappear, from more than 200 percent to less than 5 percent. In addition, we find evidence that the yield gap is mostly due to low efficiency in operating each plot in poor countries and less so to distorted crop-mix choices and crop-location choices within a country. JEL classification: O11, O14, O4. Keywords: agriculture, land quality, productivity, cross-country. Contact: Tasso Adamopoulos, tasso@econ.yorku.ca; Diego Restuccia, diego.restuccia@utoronto.ca 1

2 1 Introduction Understanding the large differences in output per worker across rich and poor countries is a fundamental issue on the research agenda in economics. It is now well understood that the differences in agricultural output per worker are considerably larger than those at the aggregate level, particularly when comparing the very rich to the very poor countries. Why are poor countries so much more unproductive in agriculture than in the rest of the economy? There is mounting evidence that the agricultural productivity problem in poor countries can be traced to a large extent to economic decisions that affect either the level of inputs used or the allocation of factors across heterogeneous farmers within agriculture. 1 Nevertheless, an alternative explanation for the agricultural productivity gap is that poor countries have a natural disadvantage in agriculture when compared to rich countries. The natural disadvantage would stem from poor countries endowment of geography and the quality of land available for agriculture relative to rich countries. In this paper, we study the role of geography and land quality for agricultural productivity gaps by documenting differences in the internal distribution of land quality across countries and assessing the quantitative role that land quality plays in accounting for differences in agricultural productivity across countries. To accomplish these objectives, we use a high resolution micro-geography data from the Global Agro-Ecological Zones project (GAEZ) of the Food and Agricultural Organization (FAO). The data provides not only actual yields for the crops produced in each plot of land covering the entire globe, but also potential yields in each plot for all the main crops. The data combines established agronomic models with detailed micro-geography information at the plot level. We develop a simple spatial-accounting framework that allows us to aggregate up from the plot-crop level resolution to the country level. The analysis imposes minimal 1 The explanations include: low intermediate input use Restuccia et al. (2008); poor transport infrastructure Adamopoulos (2011); misallocation of labor between agriculture and non-agriculture Restuccia et al. (2008), selection Lagakos and Waugh (2013), misallocation of factors across farms within agriculture Adamopoulos and Restuccia (2014), international transport frictions Tombe (2014), and idiosyncratic risk Donovan (2014). Gollin et al. (2014b) and Herrendorf and Schoellman (2013) study the role of measurement. 2

3 structure, using the detailed data through a series of relevant counterfactuals to determine to what extent low land productivity in poor countries is the result of unfortunate geography or the result of unfortunate economic choices given their geography, in terms of what crops are planted and where they are planted within the country. Our main finding is that low agricultural productivity in poor countries is not due to poor land quality and geography. If both rich and poor countries produced according to their potential yields given their internal dispersion of land quality, the rich-poor agricultural yield gap would virtually disappear, from more than 200 percent to less than 5 percent. We also find evidence that the yield gap is mostly due to low efficiency at the plot level, with crop choices in each location and location of crops within a country playing secondary roles. By nature, agriculture is an activity that takes place across space, using location-specific inputs such as soil quality, climate conditions, and terrain topography. 2 These attributes could matter not only for what yield a farmer gets for any crop harvested, but also for what crops are planted in each location and what locations are used in agriculture. Traditional measures of land quality in crosscountry comparisons, such as the percentage of agricultural land classified as arable or cropland, or the percentage of arable land or cropland that is irrigated face two main problems. First, they can be affected by economic decisions and hence are not exogenous. Second, they are average measures for the whole country and may confound important within-country variation in land quality. To the best of our knowledge, our paper is the first systematic quantitative assessment of the role that differences in geography and land quality play in understanding agricultural productivity differences across countries using micro-geography data. We use micro data from the Global Agro-Ecological Zones (GAEZ) project of the Food and Agriculture Organization of the United Nations (FAO) in collaboration with the International Institute for Applied Systems Analysis (IIASA). GAEZ provides high-resolution spatial-grid data on soil quality, climate conditions, as well as topography and terrain, covering the entire globe. The unit 2 In contrast to agriculture the quality of land is less of a factor for manufacturing plants or service-sector establishments. 3

4 of analysis in the GAEZ data is a plot, a cell in the grid, of roughly 10 by 10 kilometers. GAEZ combines detailed geographic information and local land quality attributes with state-of-the-art agronomic models to estimate a crop-specific potential yield for each of 23 main crops in every plot of land. Importantly, potential yields for each crop are provided in all plots, even for plots of land not engaged in agricultural production. In addition, GAEZ provides information by plot on actual production of each crop, actual harvested area by crop, actual yield by crop, as well as land utilization. Using this data we can measure not only the overall land quality of a country but also the entire distribution of crop-specific land quality attributes within a country at a very fine spatial resolution. To aggregate up any variable from the plot-level to the country-level we combine the GAEZ data with political boundaries data in ArcGIS. Given the data, the measure of agricultural productivity that we focus on is real agricultural output per unit of cultivated land, also known as land productivity or yield. We consider two types of yield measures. First, the aggregate yield over groups of crops, measured as total output in each country valued at the common set of international crop prices from the FAO for all plots and all countries. Second, physical output in tonnes per unit of land for individual crops. The upside of our first measure is that it allows us to construct summary measures of overall agricultural productivity across countries. The upside of our second measure is that it allows us to compare the same good across countries ( apples for apples ) without the need of aggregation using international prices. We develop a simple quantitative spatial-accounting framework to decompose aggregate agricultural productivity, measured as output per hectare (yield), into the contributions of: (i) within-location efficiency, and (ii) across-location efficiency. In particular, we consider a world in which each country consists of a given number of plots (roughly km). Each plot can produce any of a number of crops. However, plots are heterogeneous with respect to their suitability in producing each crop. A plot s suitability in producing a given crop is captured by its potential yield, which measures how much output per hectare could be produced given the plot s particular geographic attributes. For 4

5 each plot we also have what part of it is used for crop production, what crops are actually planted there, and what the actual yields are for the planted crops. A country s aggregate yield can be expressed as the ratio of the real value of its output to the total harvested land. We use this spatial accounting framework to measure four sources of inefficiency, conducting alternative counterfactual experiments: 1. Production inefficiency at the crop-plot-level: we hold the set of crops produced in each plot and the amount of land used in their production constant to the actual ones, but crops are produced according to their potential yields rather than their actual. 2. Production inefficiency in the crop mix within plots: we hold the total amount of land used in production in each plot fixed to its actual, but the highest yielding crops (given the plot s land quality attributes) are produced rather than the actual ones. 3. Production inefficiency across cultivated plots: we hold the total amount of a country s land allocated to a given crop fixed but allow it to be produced in the country s cultivated plots with the highest yield for each crop. 4. Production inefficiency across cultivable plots (all plots not only those were production takes place): we hold the total amount of a country s land allocated to a given crop fixed but allow it to be produced in any of the country s plots (except urban or forest) with the highest yield for each crop. We find that if countries produced according to their potential yields rather than their actual, keeping everything else constant, the aggregate yield disparity between the richest and poorest 10% of countries would drop from 3.14-fold to 1.05-fold. In other words, the productivity disparity would virtually disappear. In addition, we find that if countries produced in each plot the highest yielding crops then the rich-poor aggregate yield disparity would drop further to 0.77-fold. In other 5

6 words, poor countries produce systematically lower yielding crops given their land quality. We also find that reallocating crop production across plots within countries to take advantage of the local crop-specific land quality characteristics would lead to efficiency gains in poor relative to rich countries in the range of 2-4-fold depending on the individual crop. These findings suggest that poor land quality is not the source of low agricultural productivity in poor countries. Instead, it is the economic choices made by poor countries in terms of what is produced, how it is produced, and where it is produced given their geography, that lies behind their low productivity problem. Our paper contributes to a growing literature that studies agricultural productivity gaps across countries. One branch of this literature, including Restuccia et al. (2008), Caselli (2005), Adamopoulos (2011), Lagakos and Waugh (2013), Adamopoulos and Restuccia (2014), Tombe (2014), Donovan (2014) among others, has used quantitative models to assess the contribution of particular factors that can affect either the level of inputs used in agriculture or their allocation. Another branch of this literature has focused on measuring sectoral agriculture-nonagriculture productivity gaps across countries (Herrendorf and Schoellman 2013, Gollin et al. 2014b) or cross-country agricultural productivity disparities (Prasada Rao 1993, Restuccia et al. 2008, Gollin et al. 2014a). While our paper is also about measurement, we differ from the above literature in two important ways. First, we focus on measuring the role of land quality and geography for agricultural productivity gaps. Second, we use spatially explicit micro-geography data to measure actual and potential agricultural yield gaps across countries. Our paper is also related to the literature that studies cross-country differences in aggregate land quality indices and their effect on agricultural productivity. While traditional measures of land quality, such as the Peterson Index, were subject to the criticism that they included components that were not exogenous, more recent efforts have tried to focus on measures that are affected less by human decisions. For example, Wiebe (2003) constructs a land quality index which utilizes soil and climate properties to classify a country s cropland (a much more narrow concept of land, already 6

7 used for agricultural production). The index ranges from 3 to 11, where a higher number reflects a higher land quality class. While this index varies across the countries in our sample, it does not vary systematically with income. Wiebe and Breneman (2000), use the same underlying data to construct a complementary measure of land quality: the share of country s cropland that is not limited by major soil or climate constraints to agricultural production (cropland in the highest three land quality classes). This measure is a fraction and ranges from 0 to 1. They include this measure in a regression to study the effect of land quality on agricultural labor productivity. They find that good soils and climate are associated with a 13% increase in agricultural productivity relative to poor soil and climate. We differ from these approaches in that: (a) we exploit the explicit spatial nature of the micro-geography data in GAEZ, and (b) we use an accounting framework which allows us to assess the contributions of land quality vs. a set of economic choices in accounting for agricultural productivity. Our findings are consistent with the above literature on land quality, in that land quality does not appear to be a key determinant of agricultural productivity differences between rich and poor countries. We are not the first economists to use the GAEZ data. Nunn and Qian (2011) use the GAEZ data to assess the suitability of Old World regions for the cultivation of the potato, in order to estimate the effect of the potato on population and urbanization. Costinot et al. (2014) use the GAEZ data to study the effects of projected climate change on aggregate output through the evolution of comparative advantage and the accompanying adjustments in production and trade. Costinot and Donaldson (2014) use the GAEZ data to study the gains from economic integration within the United States. We are however the first to use the GAEZ data to study their macro-level implications for cross-country differences in agricultural productivity, an issue that is paramount for understanding the foundation of poverty across the world. The paper proceeds as follows. The next section describes the GAEZ data in some detail and provides some observations on land quality dispersion across countries based on the GAEZ data. 7

8 In Section 3 the spatial accounting framework and describe our counterfactuals. In Section 4 we present our main findings. We conclude in Section 5. 2 GAEZ Data The Global Agro-Ecological Zones (GAEZ) project, developed by the Food and Agricultural Organization (FAO) in collaboration with the International Institute for Applied Systems Analysis (IIASA), is a standardized framework for the characterization of climate, soil and terrain conditions relevant to agricultural production. GAEZ combines state-of-the-art established agronomic models with high resolution spatial data on geographic attributes to, (a) classify land according to its suitability for the production of specific crops, and (b) calculate the potential yield that could be attained for each crop in each spatial unit, including those where the crop is not actually produced. GAEZ also provides spatially detailed land utilization as well as actual crop-specific production information. The GAEZ information is available at the 5 arc-minute resolution. To picture it, imagine superimposing a grid of about 9 million cells or pixels covering the entire world. While the size of each cell is constant at 5 arc-minutes, it is not constant in terms of squared kilometers or hectares, as the mapping from arc-minutes to square km depends on the distance from the equator (latitude). As a rough approximation the size of each cell can be described as kilometers. For each cell in the grid GAEZ has micro data on the following location-specific geographic attributes that are important for agricultural production: (a) soil quality, which includes depth, fertility, drainage, texture, chemical composition; (b) climate conditions, which includes temperature, and moisture environment; (c) terrain and topography, which include elevation, slope. This micro-geography data constitute the main input into agronomic models that combine crop-specific 8

9 biophysical requirements, along with a level of water supply (irrigated or rain-fed), and cultivation practices (input intensity use and management). 3 Then for each of 49 crop varieties the models classify each cell according to the extent of its soil, terrain, climate constraints for the production of that crop. The same information is used to construct a crop-specific suitability index, and most importantly to calculate a potential yield for that crop in that cell, measured as the maximum potential output (in tonnes) per hectare that can be attained in the cell given its production requirements and the cell s characteristics. Potential yields are calculated not only for the crops actually produced in the cell but for any of the set of 49 crops. The GAEZ potential yields are calculated for both historical climate conditions (with the baseline reference period being ) as well as projected future climate conditions. The GAEZ database also provides at the 5 arc-minute resolution data on crop choice, actual production, actual area harvested, and actual yield, i.e., tons of production per hectare of the crop actually planted. In addition, the database contains land cover data that classify land in cell in terms urban, cultivated, forest, grassland and woodland, water bodies etc. GAEZ provides the information for each variable in raster (grid cell) files, which we work with in ArcGIS. To aggregate cell-level information to administrative units, such as regions, countries, provinces etc. we use world borders data shape files from Thematic Mapping. 3 Low level of inputs (traditional management) assumes subsistence based farming, labor intensive techniques, no application of nutrients, chemicals, pesticides. Intermediate level of inputs (improved management) assumes partly market oriented farming, improved varieties with hand tools and/or animal traction, medium labor intensity, use of some fertilizer, chemicals, pesticides. High level of inputs (advanced management) assumes mainly market oriented farming, high yield variety seeds, fully mechanized with low labor intensity, optimum application of nutrients, chemicals, pesticides as well as disease and weed control. 9

10 3 Accounting Framework Consider a world that consists of a fixed number of administrative units, indexed by u U {1, 2,..., U}. These units could be countries, or lower administrative units, such as provinces, states or counties, within countries. Each administrative unit u consists of a finite number G u of grid cells (or pixels) of fixed size - in the GAEZ data the grid cell resolution is 5 arc-minutes. Note that while the size of a cell in arc-minutes is constant, it is not constant in terms of hectares as the mapping from arc-minutes to hectares depends on the distance from the equator. We index grid cells by g G u. Cells can be aggregated up to various levels of administrative units (regions, countries, states, counties etc.) using a mapping of cells to administrative boundaries in ArcGIS. Each grid cell can produce any of C crops, indexed by c C {1, 2,..., C}. Cells are heterogeneous with respect to their productivity across crops. In particular, grid cells differ with respect to their potential yield or land productivity (tons per hectare) if they were to produce crop c. We denote the potential yield from producing crop c in grid cell g in unit u by ẑgu. c Note that for each cell g in unit u there are C such numbers, each of which reflects the inherent productivity of that cell in producing crop c. In reality cells are either used for crop production or some other activity (could be agricultural such as raising livestock or non-agricultural, or some other land cover category). If a cell is used for crop production it would produce specific crops which may differ from the ones in which the cell has the highest potential yield. We use an indicator function Igu c to denote whether production of crop c takes place in grid cell g in unit u, with the convention that Igu c = 1 if production takes place and 0 otherwise. If Igu c = 0 for all c then no crop production takes place in that cell. When Igu c = 1, i.e., when crop c is produced in cell g in unit u, we denote by zgu c the actual yield, by ygu c the real output (in tons), and by l c gu the amount of land in hectares. Note then that ygu/l c c gu = zgu. c Note that the size of each vector is C 1, that is the total number of crops in the GAEZ project (18 10

11 crops). In each cell g all the vectors have non-zero elements only for the crops actually produced there. The only vectors that have non-zero elements for every crop are the potential yield vector and the common international price vector. The potential yield vector is specific to each cell g but has non-zero elements because it provides the yield that would be possible had you produced any of the crops. The price vector is common to all administrative units and tells you the common international price for any crop. Land Allocation Variables Land devoted to the production of crop c in cell g: l gc u. Total land (in hectares) devoted to production (cultivated land) in cell g, l gu = c C l c gu Total land (in hectares) in unit u devoted to the production of crop c, L c u = g G u l c gu Total land (in hectares) in unit u devoted to production of any crop (cultivated land), L u = c C L c u g G u l gu Amount of land in cell g not used for crop production, l 0 gu = 1 l gu 11

12 Total amount of land not used for crop production in unit u, L 0 u = g G u l 0 gu Real Output Variables Given that we consider different crops, in order to calculate aggregate output by aggregating up from crops we need to use crop prices. Given that we are interested in calculating real measures of aggregate output we consider a common set of international prices {p c} c C. Actual output of crop c in cell g: y c gu = z c gul c gu Total actual real output in cell g in unit u, y gu = c C p cy c gu Total actual output (in tons) of crop c produced in unit u, Y c u = g G u y c gu g G u z c gul c gu Aggregate actual real output in unit u, Yu = p cyu c ygu c C g G u Measure of Productivity Given that we lack information on labor or other inputs the measure of productivity for each administrative unit u that we focus on is the aggregate yield (land 12

13 productivity), i.e., aggregate real output per unit of land devoted to agricultural production: r u = Y u L u When making comparisons across administrative units i and j we look at the ratio of their aggregate yields, r i /r j. Counterfactuals We construct a set of counterfactual or aggregate potential yields for each administrative unit u by exploiting the set of potential yields by crop at the cell level g and the spatial distribution of production by crop across cells. We are interested in two types of comparisons. First, how does each administrative unit s aggregate actual yield r u compare to its potential aggregate yield r cf u. Second, how do the actual yield differences across administrative units r i /r j compare to the potential yield differences r cf i /r cf j. Note that rcf u is the potential yield if either the aggregate real output or aggregate labor input were equal to their counterfactual values, Y,cf u and L cf u respectively. r cf u = Y u,cf L cf u There are four margins that contribute to the determination of the actual aggregate yield: (a) Whether each crop is produced in each plot according to its potential? (b) Whether each plot produces the crop in which it possess an absolute advantage? (c) Whether crops are produced within the country in the highest yielding locations across plots of cultivated land? Whether crops are produced within the country in the highest yielding locations across plots of potentially cultivable land? To examine the importance of these margins we consider the following set of counterfactual experiments. 13

14 Counterfactual 1: Production efficiency at the crop-plot-level The question that we ask here is the following: how would the aggregate yield change if in each plot g in administrative unit u each crop was produced according to its potential yield rather than its actual, holding the set of crops and the allocation of land across crops fixed to their actual values. Counterfactual 2: Production efficiency in the crop mix within plots The question that we ask here is the following: how would the aggregate yield change if in each plot g the crop mix was changed towards the production of the highest yielding crops, keeping the amount of land allocated to production within each plot fixed. Counterfactual 3: Production efficiency across cultivated plots The question we ask here is the following: how would the aggregate yield change if we reallocated the production of crops across cultivated plots according to their comparative advantage holding the total land allocated to each crop in the administrative unit at its actual level. Counterfactual 4: Production efficiency across cultivable plots The question we ask here is the following: how would the aggregate yield change if we reallocated the production of crops across potentially cultivable plots according to their comparative advantage holding the total land allocated to each crop in the administrative unit at its actual level. Note that in counterfactuals 1 and 2, the aggregate potential yield changes only because of changes in the numerator Yu,cf, while in counterfactuals 3 and 4, the aggregate potential yield changes only because of changes in the denominator L 1,cf u. 14

15 4 Data Analysis and Findings 4.1 Land Quality Characteristics In Figures 1-4 we present maps we have constructed in ArcGIS using the micro-geography data from GAEZ to illustrate the diversity of some important land quality and geography characteristics across the world. Figure 1 classify the soil at the 5 arc-minute resolution according to its nutrient content, an important indicator of soil fertility particularly in low intermediate-input intensive environments. The classification determines how constrained the soil in each cell is in terms of its nutrient content, ranging from no/slightly constrained to severely constrained. Figure 2 shows median altitude within each pixel for the whole world. Altitude is an important indicator of terrain suitability for agricultural production. Soil Fertility Constraints No or slight Moderate Severe Very severe Mainly non-soil Premafrost zone Water bodies 0 3,200 6,400 12,800 Kilometers Source: Soil Resources, Land Resources, GAEZ. Figure 1: Soil Fertility 15

16 Median Altitude High : 6563 Low : ,200 6,400 12,800 Kilometers Source: Terrain Resources, Land Resources, GAEZ. Figure 2: Median Altitide In Figures 3 and 4 we present pixel-level indicators thermal regimes (mean temperature) and moisture regimes (annual precipitation) respectively. Both measures are important measures of agroclimatic conditions and serve as key inputs into the GAEZ methodology in constructing crop-specific potential yields by cell. What is important to note is that while there seems to be considerable variation in each one of the above land quality attributes (as well as others not presented here) across the world, we are ultimately interested to study whether all these measures together contribute to systematically different growing conditions across rich and poor countries. In order to determine this we consider we work with potential yields, which summarize crop-specific land quality and geographic attributes, in the following set of counterfactuals. 16

17 Mean Annual Temperature High : ,200 6,400 12,800 Kilometers Low : Source: Thermal Regimes, Agro-Climatic Resources, GAEZ. Figure 3: Mean Temperature 4.2 Results Using plot-level production and land use data within countries we calculate for each country aggregate output per hectare (actual yield) using FAO international prices. In Figure (5) we plot for 162 countries the aggregate yield against real GDP per capita, both in logs. The aggregate yield varies systematically with the level of development, with the correlation in logs being The 10% richest countries have an average yield that is 3.1 times higher than the average in the 10% poorest countries in our sample. To what extent are these yield differences due to differences in land quality and geography across rich and poor countries? We address this questions through a set of counterfactual experiments outlined in Section 3. Counterfactual 1: Production Efficiency at the crop-plot-level. The question that we ask 17

18 Annual Precipitation High : 5440 Low : 0 0 3,200 6,400 12,800 Kilometers Source: Moisture Regimes, Agro-Climatic Resources, GAEZ. Figure 4: Annual Precipitation here is the following: how would the aggregate yield within each country change if in each plot of every country each crop was produced according to its potential yield rather than its actual, holding the set of crops and the allocation of land across crops fixed to their actual values? This experiment allows us to construct for each country a counterfactual or potential aggregate yield. In the first panel of Table 1 we present the results of this counterfactual for the richest and poorest countries. The first column shows the actual aggregate yield for the 10% richest and poorest countries, as well as their ratio. The second column, presents the potential aggregate yield, computed under counterfactual 1, for the richest and poorest countries and their ratio. In the third column we compute the yield gap for each group of countries, as the ratio between the aggregate potential and actual yields. The results are striking. If poor countries produced the crops they are actually producing but according the their potential yields in each plot, then the aggregate yield disparity between rich and poor countries would drop from 3.14 to 1.05, that is it would virtually disappear. 18

19 Actual Yield by Country Log of Actual Yield (Int.$ per Hectare) ZAR AFG BEL ARE KWT EGY GRC IRL LUX SLB UZB JAM MYSKOR GBR JPNCHE CHN COL CRI NZL GER FJI TKM MLT ISR FRA DNK GUY AUT IDN LBN TWN CHL ITA PNG PRI PER SUR CZE SVN OMN SWESGP TJKVNM USA SWZ BGD COG LKA MNE DOM BLZ BRA CYP CIV CUB ESP NOR GHALAO PAK PRY KGZJOR THA HRV VENHUN GTM VUT POL URY SVK ARG SAU FIN QAT BEN SYR CMRALB PHL AZE ECUTUR PANGAB BHS BRN BIH LTU MDG LVA BLR IND AUS CAN KHM MKD PRT UGA BOL SLV UKR ROM ARM BGR ZAFMEX LBR HND IRN EST MWI NPL BDI CAF GMB MDA NGA NIC TGO ZMB AGO GIN GNQ LBY TZA GEO ZWE RUS GNB RWA MLI SLE MOZ SEN BTN DZA HTI KEN ETH KAZ BFA YEMTCD TUN LSO MAR ATG SDNMNG IRQ NAM SOM ERI MRT NER BWA Log of Real GDP Per Capita (2000) NLD ρ = , N = 162 Figure 5: Actual Yield across Countries The yield gap (third column) is very telling. While it is true that both rich and poor countries are not producing at their full potential, the distance from the full potential is much larger for poor countries than it is for rich countries. In Figure 6 we present the yield gap (potential to actual yield) against real GDP per capita in logs for the entire set of countries in our sample. As is clear the yield gap is indeed systematically negatively correlated with the level of development, with a correlation coefficient of In Figure 7 we plot the standard deviation of plot-level yields gaps within each country against the level of income. This figure suggests that not only do poor countries have on average larger gaps from their potential yields but the dispersion of their yield gaps also appears to be slightly higher than in rich countries. In the second panel of Table 1 we present the results of the first counterfactual experiment for the sub-group of cereal crops, which covers a set of crops produced in most countries of the world such 19

20 Yield Gap by Country Log of Yield Gap (Potential Over Actual Yield) AFG ZAR BWA ATG LSO NER RWA SDN HTI MAR IRQ TUN SOM GNQ GEO TZA CAFBFA MOZ ARM BDI ZMB UGA KEN BTN LBR GNB ERI MLI YEM SLETCD ZWE AZE ETH TGO MRT DZA BLR MDG MWI TJKMDA BEN GIN KHM NPL MNG AGO ALB UKR BIH ROM GAB LBY PRT BOL MNE KGZ MKD BGR RUS NAM CIV BRN HND CUB LVA LTU HRV NGA TKMPOL CMR COG DOM EST GMB NIC PNG UZB SEN GHA IND VUT SWZ ECU KAZ ZAF PRI PHL PER SLV IRN SVK TUR PAN CHL MLT LAO HUN BGD JOR LKA MEX URY SYR PAK GTMTHA SUR SVN PRYFJI LBN VENARG BLZ ESP COL CZECYP ISR SGP VNM SLB GUY BRA CHN IDN CRI MYS USA BHS EGY SAU GRC NZL OMN QAT TWNITA SWE AUS CAN AUT CHE BEL DNK KOR GBR FRA JPN GERARE NOR LUX JAM FINNLD IRLKWT ρ = , N = Log of Real GDP Per Capita (2000) Figure 6: Yield Gap across Countries as wheat, rice, maize, sorghum, millet, and others. The rich-poor disparity in the actual yield for cereals at 4.61-fold is larger than in the total crop group (3.14). If all countries were to produce their cereal crops in each location according to their potential yields then the rich-poor country disparity would drop down to 1.24, echoing a similar message to the total crop category. Again, it is clear from the third column that the yield gap is 3.7 times higher in poor than in rich countries. In Table 2 we repeat counterfactual 1 for each of the three most representative crops produced across the world: wheat, rice, and maize. In this case the yield is measured as output in tonnes per unit of land, as we do not have to use international prices to aggregate real output. The rich-poor disparity in the actual yield differs across crops with 6.53-fold and 5.11-fold disparities for maize but only 2.53-fold disparity for wheat. Producing these crops according to potential yields would reduce the rich-poor gap to 1.60-fold in the case of maize and 1.22-fold in the case of rice. In the case of wheat the rich-poor disparity drops to Despite these differences across individual 20

21 Dispersion of Yield Gaps within Countries Std. of Yield Gap (Potential Over Actual Yield) BWA LSO MAR LBY TJK SDN TUN TKM BDI HTI IRQ MOZ MRT SWZ ZMB TCD KEN AGO ECU NAM TZAHND SOMAZE ZWE ATG BTN AFG ERI NER MWI PNGPER ZAR BFA CMR BEN BOL DOM ETH RWA UGA MLI YEM IND GTM LBR MDG SLE UZB SEN SYRMNE ZAF NGA GHA NPL CAF NIC GIN ARM BIH GNQURY SAU PRT BGD ALB COGCOLCRI GEO GNB DZAIRN KOR KGZ HRV ISR PRI TGO GMB GUY CHN ARG KHM LAO PAK SLV MKD MEX MNG PHLTHA VNM PAN CIVEGY BGR BLZ BRA MDA IDN JOR PRY LKACUB KAZEST BLR CHL MLT CZE CYP BHS ESP FJI ROMRUS TUR SURVEN LBN JAM LVA GAB LTUPOL HUN MYS GRC NZL OMN FIN GBR ITA JPN SLB VUT UKR SVK SVN TWN SWE FRA GER BEL AUS DNK IRL CANCHE AUT NLD SGP KWTARE USA NOR ρ = , N = 162 BRN Real GDP Per Capita (2000) QAT LUX Figure 7: Within-Country Dispersion of Yield Gaps crops, the main message remains the same: poor countries are producing much further away from their potential than rich countries (third column of each panel), and if they were able to produce according to their potential the rich-poor gap would virtually go away. Counterfactual 2: Production Efficiency in the crop mix at the plot-level. The question that we ask here is the following: how would the aggregate yield within each country change if in each plot the highest yielding crop was produced, holding the amount of land used in each plot for agricultural production at its actual level? This experiment allows us to construct for each country a counterfactual or potential aggregate yield, that reflects the extent to which countries are producing crops in which they have an absolute advantage. In Table 3 we present the results of the counterfactual 2 for the total crop category. If all countries were to shift their crop mix to the highest yielding crops, plot-by-plot, then the aggregate yield disparity between rich countries would drop from 3.14 that it actually is to In other words 21

22 poor countries would be about 30% more productive than rich countries. Counterfactual 3: Production efficiency across cultivated plots The question we ask here is the following: how would the aggregate yield change if we reallocated the production of individual crops across cultivated plots according to where they exhibit the highest yield in the country holding the total land allocated to each crop in the country at its actual level. In Table 4 we show the results of counterfactual 3 for the individual crops of wheat, rice, and maize. We find that reallocating crop production within countries to its highest yielding locations would have substantial gains in the aggregate yield and would reduce the rich-poor gap by a factor of about 2 in the case of wheat and about a factor of 4 for rice and maize. 5 Conclusions Using detailed micro-geography data we study the macro-level consequences of land quality for agricultural productivity, measured as output per hectare (yield). In particular, we examine to what extent differences in agricultural yields across countries are the result of natural advantages/disadvantages or the result of economic choices. We find that land quality differences cannot justify the agricultural productivity gaps between rich and poor countries. Instead we trace the problem to what crops are produced, where they are produced within the country, and what level of inputs they use. We see more need for future research on cross-country agricultural productivity gaps to try to understand what factors lead poor countries to not better exploit their land endowments. 22

23 References Adamopoulos, T. (2011). Transportation costs, agricultural productivity, and cross-country income differences. International Economic Review, 52(2): Adamopoulos, T. and Restuccia, D. (2014). The size distribution of farms and international productivity differences. American Economic Review, 104(6): Caselli, F. (2005). Accounting for cross-country income differences. Handbook of Economic Growth, 1: Costinot, A. and Donaldson, D. (2014). How large are the gains from economic integration? theory and evidence from us agriculture, Technical report. Costinot, A., Donaldson, D., and Smith, C. B. (2014). Evolving comparative advantage and the impact of climate change in agricultural markets: Evidence from 1.7 million fields around the world. Journal of Political Economy, Forthcoming. Donovan, K. (2014). Agricultural risk, intermediate inputs, and cross-country productivity differences. Unpublished manuscript, University of Notre Dame. Gollin, D., Lagakos, D., and Waugh, M. E. (2014a). Agricultural productivity differences across countries. The American Economic Review: Papers and Proceedings, 104(5): Gollin, D., Lagakos, D., and Waugh, M. E. (2014b). The agricultural productivity gap. The Quarterly Journal of Economics, 129(2): Herrendorf, B. and Schoellman, T. (2013). Why is measured productivity so low in agriculture? Unpublished manuscript, Arizona State University. Lagakos, D. and Waugh, M. E. (2013). Selection, agriculture, and cross-country productivity differences. The American Economic Review, 103(2): Nunn, N. and Qian, N. (2011). The potato s contribution to population and urbanization: Evidence from a historical experiment. The Quarterly Journal of Economics, 126(2):

24 Prasada Rao, D. (1993). Intercountry comparisons of agricultural output and productivity. Technical report. Restuccia, D., Yang, D. T., and Zhu, X. (2008). Agriculture and aggregate productivity: A quantitative cross-country analysis. Journal of Monetary Economics, 55(2): Tombe, T. (2014). The missing food problem: Trade, agriculture, and international productivity differences. American Economic Journal: Macroeconomics, Forthcoming. Wiebe, K. D. (2003). Land quality, agricultural productivity, and food security: Biophysical processes and economic choices at local, regional, and global levels. Edward Elgar Publishing. Wiebe, K., S. M. N. C. and Breneman, V. (2000). Resource quality and agricultural productivity: A multi-country comparison. USDA. 24

25 Table 1: Production by Potential Yield (Counterfactual 1) All Crops (country obs. = 162) Actual Yield Potential Yield Yield Gap Rich 10% , Poor 10% , Ratio /2.99 Cereal Crops (country obs. = 160) Actual Yield Potential Yield Yield Gap Rich 10% , Poor 10% Ratio /3.72 Note: Rich and Poor refer to the highest and lowest decile of the world income distribution, where income is Real GDP per capita in 2000 (PWT 6.3). Actual and potential yields are measured as total real gross output per hectare in international prices (GK $/ha). Each country-level yield is constructed by aggregating up from the GAEZ pixel-level information at the 5 arc-minute resolution (roughly 10-by-10 km). Potential yield assumes high level of input use for every pixel. The yield gap refers to the ratio of potential to actual yield. 25

26 Table 2: Production by Potential Yield (Counterfactual 1) Wheat (country obs. = 110) Actual Yield Potential Yield Yield Gap Rich 10% Poor 10% Ratio /2.96 Rice (country obs. = 104) Actual Yield Potential Yield Yield Gap Rich 10% Poor 10% Ratio /4.18 Maize (country obs. = 142) Actual Yield Potential Yield Yield Gap Rich 10% Poor 10% Ratio /4.09 Note: Rich and Poor refer to the highest and lowest decile of the world income distribution, where income is Real GDP per capita in 2000 (PWT 6.3). Actual and potential yields are measured as tons per hectare. Each country-level yield is constructed by aggregating up from the GAEZ pixel-level information at the 5 arc-minute resolution (roughly 10-by-10 km). Potential yield assumes high level of input use for every pixel. The yield gap refers to the ratio of potential to actual yield. 26

27 Table 3: Production by Maximum Potential Yield (Counterfactual 2) All Crops (country obs. = 164) Actual Yield Counterfactual Yield Rich 10% ,498.3 Poor 10% ,254.7 Ratio Note: Rich and Poor refer to the highest and lowest decile of the world income distribution, where income is Real GDP per capita in 2000 (PWT 6.3). Actual and potential yields are measured as total real gross output per hectare in international prices (GK $/ha). Each country-level yield is constructed by aggregating up from the GAEZ pixel-level information at the 5 arc-minute resolution (roughly 10-by-10 km). Potential yield assumes high level of input use for every pixel. The yield gap refers to the ratio of potential to actual yield. 27

28 Table 4: Yield Gains from Spatial Reallocation (Counterfactual 3) Wheat (country obs. = 105) Actual Yield Counterf. Yield Yield Gap Rich 10% Poor 10% Ratio /2.15 Rice (country obs. = 96) Actual Yield Counterf. Yield Yield Gap Rich 10% Poor 10% Ratio /3.82 Maize (country obs. = 144) Actual Yield Counterf. Yield Yield Gap Rich 10% Poor 10% Ratio /4.38 Note: Rich and Poor refer to the highest and lowest decile of the world income distribution, where income is Real GDP per capita in 2000 (PWT 6.3). Actual and potential yields are measured as tons per hectare. Each country-level yield is constructed by aggregating up from the GAEZ pixel-level information at the 5 arc-minute resolution (roughly 10-by-10 km). Potential yield assumes high level of input use for every pixel. The yield gap refers to the ratio of potential to actual yield. 28