Abstract
This paper studies the relationship between geographic heterogeneity and technology adoption in the context of the direct planting system (DPS) in Brazil. The DPS is a no-till farming technique that increases productivity and decreases soil degradation. However, it requires adaptation to local conditions to be profitably used. Combining detailed geographic and agricultural data, we show that geographic heterogeneity reduces DPS adoption. This effect is robust to the inclusion of controls and not observed for technologies that do not require local adaptation. These findings are consistent with models in which geographic heterogeneity increases the cost of adapting technologies to local conditions. (JEL O13, Q55)
1. Introduction
Differences in the adoption of modern technologies account for a substantial share of the cross-country differences in productivity (Comin and Hobijn 2010). Understanding the determinants of these differences in technology use is therefore essential for understanding the existing differences in productivity. Because most of the cross-country differences in productivity are due to productivity differences in agriculture, this topic is of special importance in this sector (Gollin, Lagakos, and Waugh 2014). This motivated a large literature investigating the determinants of the adoption of modern agricultural technologies. This literature has associated adoption rates with factors like hidden costs, market failures, or even behavioral biases.1
This paper provides evidence that geographic heterogeneity is another important determinant of the adoption of modern technologies in agriculture. We combine fine-level geographic data with information on the adoption of the direct planting system (DPS) in Brazil to document that increases in geographic heterogeneity are connected to lower technology adoption. The DPS is a no-till technique developed in southern Brazil in the 1970s in areas prone to land degradation. It later evolved into a full farming method with higher revenues and lower costs than the traditional methods used in the country (Inoue 2003). The limited use of tillage also prevents soil degradation and the loss of nutrients, which boosts productivity in the long run.2 Despite these benefits, just 10% of farmers had adopted the technique in 2006. Underadoption is often connected to difficulties in the diffusion of information about this technology, since farmers must adapt the DPS to local conditions (Landers 2005; Derpsch et al. 2010).
Our empirical exercise uses municipality-level information on DPS adoption and geographic heterogeneity in Brazil. Because the use of this technology is influenced primarily by soil characteristics, we use soil heterogeneity as a proxy of geographic heterogeneity. Soil heterogeneity is defined as the inverse of the Herfindahl index of soil concentration. To ensure that technology adoption does not influence this measure of heterogeneity, we construct it using a physicochemical classification of soil types that is invariant to land use.
We then examine whether this measure of heterogeneity in soils influences DPS adoption by comparing municipalities in the same state that have similar geographic endowments but different levels of heterogeneity. We begin by using soil types, gradient, and altitude as controls. We subsequently include the agronomic potential to cultivate different crops, temperature, rainfall, latitude, longitude, and state fixed effects as controls. Regardless of the choice of controls, we find that DPS adoption is significantly lower in municipalities with more heterogeneous soils. The coefficients imply that a one standard deviation increase in soil dissimilarities reduces DPS adoption by 1.49 to 1.62 percentage points.
While exhaustive, the controls do not rule out the possibility that other (unobserved) geographic determinants of technology adoption drive the relationship we document in the data. Using the method proposed by Oster (2017), we provide evidence that these unobserved geographic determinants would have to be at least 2.9 times more important than observed geographic determinants for the effect of soil heterogeneity on DPS adoption to become zero. Other robustness tests indicate that the results do not change when we consider different definitions of soil heterogeneity, use different samples, or weight observations.
We perform placebo tests to examine whether more heterogeneous municipalities also have lower adoption rates of technologies that do not require adaptation to local conditions. The results suggest that soil heterogeneity neither affects the use of general-purpose technologies such as electric power nor the use of agricultural technologies such as harvesters.
Because geographic heterogeneity does not influence the costs and benefits of adoption directly, it is essential to discuss the mechanisms that might explain the documented relationship between geographic heterogeneity and technology adoption. We rule out explanations based on the connection between heterogeneity and land distribution, human capital, mechanization, and the provision of public goods (Alesina, Baqir, and Easterly 1999; Michalopoulos 2012). However, the absence of effects of geographic heterogeneity on technologies that do not require adaptation to local conditions indicates that the results might be tied to difficulties in the diffusion of information about modern technologies in heterogeneous areas (Ellison and Fudenberg 1993; Diamond 1997; Munshi 2004).
Our findings support Diamond’s (1997) assertion that the diffusion of agricultural technologies is slower in areas with large geographic dissimilarities. There is a burgeoning empirical literature linking geographic dissimilarities to determinants of economic development, like ethnic fractionalization or state centralization (Michalopoulos 2012; Fenske 2014). However, there is no empirical evidence that geographic heterogeneity directly influences technology diffusion. Moreover, we provide evidence that geographic heterogeneity influences only the diffusion of technologies that require adaptation to local conditions to be successfully adopted.
The results from our paper are also related to the extensive literature on the determinants of technology adoption. Griliches (1957) documents and discusses the differences in the origins, slopes, and ceilings in the diffusion of hybrid corn in the United States.3 His influential contribution inspired numerous empirical studies documenting determinants of the diffusion process.4 Some studies in this literature (e.g., Rahm and Huffman 1984; Abdulai and Huffman 2014) document the importance of geographic characteristics of the plots a farmer cultivates to his adoption decisions. Nevertheless, there is no direct evidence of the relationship between the dissimilarities in geographic characteristics in the farmer’s region and his adoption decisions.
Within the literature on the determinants of technology adoption, our work is closest to studies discussing the role of social learning in technology adoption. (Ellison and Fudenberg 1993; Foster and Rosenzweig 1995; Munshi 2004; Abdulai and Huffman 2005; Bandiera and Rasul 2006; Conley and Udry 2010). Munshi’s (2004) study is particularly important to our discussion. This author uses a theoretical model to argue that farmers should respond more (less) to their neighbors’ choices when the population is homogeneous (heterogeneous). Munshi (2004) then uses data from India to provide evidence that rice growers (more heterogeneous) respond less to neighbors’ choices regarding the adoption of high-yielding varieties than wheat growers (more homogeneous). However, it is not possible to rule out that other differences between technologies (e.g., spatial and serial correlation in profits) or farmers (e.g., credit constraints) drive the results he documents.
We believe our paper complements Munshi’s (2004) in three dimensions. First, by focusing on the diffusion of one technology in environments with comparable geographic endowments but different levels of geographic heterogeneity, our empirical design mitigates the concern that differences between technologies or farmers drive the relationship between heterogeneity and the diffusion of new technologies. Second, by using direct geographic information, we demonstrate that it is possible to test whether heterogeneity reduces the diffusion of new technologies without panel data. Third, by focusing on a different context, our paper provides evidence that heterogeneity delays technology diffusion even in the context of large and mechanized farms.
2. Background
The DPS
The DPS is a no-till technique developed in Brazil at the beginning of the 1970s. No-till techniques can be described as agricultural practices in which tillage is absent and crop residue is left on the surface. These practices reduce the loss of nutrients and erosion, but their use in large-scale agriculture requires the presence of effective herbicides, because tillage is an important tool for weed control. Indeed, the substantial benefits of tillage in terms of weed control implied that it was worth using despite its cost in terms of land degradation and erosion, until the development of more effective herbicides after the 1940s. Technological developments in weed control triggered more research about no-till techniques in the following decades. These techniques have become popular among proponents of sustainable agricultural practices because of their positive environmental consequences (Baker et al. 2007).
The DPS is a distinct no-till method in which the absence of tillage is permanent and the use of green manure crops to cover soils is widespread (Derpsch et al. 2010). These features resulted from efforts to adapt no-till techniques to the specific geographic conditions that farmers face in southern Brazil, where the technique was developed in the 1970s. The DPS has since became popular in other South American regions such as the Pampas in Argentina and the Cerrado in Brazil.
The DPS is novel production technology, as adopters continue to use similar inputs and reach a higher output. Its adoption may thus be characterized as the adoption of a different production process that uses the same inputs. This feature is important for the empirical investigation because it is often difficult to separate technology underadoption from input underuse (Foster and Rosenzweig 2010).5
DPS adoption is associated with environmental and economic gains. Its environmental benefits are derived from both lower carbon emissions and higher carbon sequestration. Both changes have positive externalities to the environment and mitigate climate change (West and Post 2002; Metay et al. 2007). Other gains have been reported in terms of reduced environmental contamination and increased biodiversity (Derpsch et al. 2010). The DPS’s economic benefits are derived from a combination of higher revenues and lower costs. The DPS reduces soil degradation and erosion and improves soil properties, resulting in higher revenues. The expenditures associated with herbicides increase, but the reduction in the use of machines more than compensates for this increase, resulting in lower costs (Derpsch et al. 2010; Baker et al. 2007). Appendix Table A1 summarizes the public and private benefits from DPS adoption.
A number of studies have quantified the benefits farmers obtain from DPS adoption. Inoue (2003) presents evidence that DPS adoption increases productivity in soy cultivation by 17% in Brazil. The author also presents evidence that average costs decrease by 9% with DPS adoption. Sorrenson and Portillo (1997) report that DPS increases net income by 33% in the first year of adoption among farmers in Paraguay. The authors also argue that the benefits gained from the DPS increase over time. Trigo et al. (2009) estimates substantial economic benefits from DPS adoption in Argentina, derived from both increases in production and decreases in costs. Simulations reported by Ringler et al. (2014) indicate that widespread adoption of no-till techniques would increase yields of staple crops, even considering the effects of climate change on temperature and rainfall. The authors therefore argue that no-till is a promising farming technique for mitigating climate change’s impact on agriculture.
An important feature of the DPS is that neither credit constraints nor incomplete insurance appear to be barriers to its adoption, because the technology does not have relevant upfront costs and does not increase risk exposure. DPS adoption also does not require additional infrastructure. However, information appears to be an important barrier to DPS adoption. Derpsch (1999) argues that adjustments required to make the DPS suitable are unique to the type of soil considered (although there are some general practices that must be implemented in all contexts). The author suggests that a lack of site-specific knowledge has been a significant barrier to the diffusion of DPS in Latin America. Indeed, his first two recommendations for farmers willing to adopt such techniques are as follows:
“Improve your knowledge about all aspects of the system but especially in weed control.”
“Analyze your soil.”
There are at least three dimensions through which soil type affects DPS use. First, it influences the type of residue to be left on the surface. Whether the residue will come from maize, millet, rye, sorghum, or other crops will heavily depend on both the type of soil and the climate (Sisti et al. 2004; Bolliger et al. 2006). Second, soils influence the thickness of the residue layer covering the soil. The thickness of this layer changes exposure to sunlight, heat, and humidity, which, in turn, affects the prevalence of weeds. Thus, soils with high temperature will typically require a thicker layer of residue, whereas humid soils will usually require a thinner layer of residue (Derpsch 1999; Inoue 2003). Third, the type of soil influences the types of herbicides used for weed control. Since no-till techniques increase weed prevalence, the use of the right herbicides is fundamental for cultivation using DPS. However, both soils and crops influence the likelihood of different weeds. This, in turn, determines not only the most effective herbicides but also how they should be applied (Nascente and Crusciol 2012; Bolliger et al. 2006).
DPS Diffusion
The DPS was first implemented at the beginning of the 1970s in southern Brazil among farmers cultivating sorghum and wheat. The technique was developed to fight rain erosion affecting soils in the region of Ponta Grossa in the state of Paraná. Pioneer farmers studied no-till techniques abroad and incurred the risk of importing equipment and knowledge and testing the method on a large scale. These farmers exerted significant influence over the development of the DPS in Brazil, inducing research about the DPS among seed producers, machine manufacturers, and research institutes.
Anecdotal evidence suggests that large farmers in the Ponta Grossa region started adopting the DPS in 1976. Technological diffusion was concentrated in southern Brazil in the first decade after the technique’s implementation. Diffusion to other regions began later. The DPS became particularly successful in the Cerrado biome, where its use expanded among crop producers. Diffusion has accelerated in all regions since the 1990s.
An important feature of the DPS’s diffusion was the creation of an association called Clube da Minhoca (“Earthworm Club”) in 1979. This organization is located in Ponta Grossa and is a diffusion center whose goal is to spread knowledge about the technology. The organization promoted meetings in which farmers (adopters and nonadopters) would discuss issues related to farming. It also organized national meetings and sponsored the publication of technical material about the DPS. It inspired the creation of private associations with similar objectives in other areas.
These associations are called Clube Amigos da Terra (“Friends-of-the-Earth Club”). The first such organization was established in 1982 in the state of Rio Grande do Sul. These associations have been an essential tool for the diffusion of the DPS throughout Brazil. They coordinate learning efforts and information exchange among farmers and, along with other organizations that promote the technique, have been essential for spreading information and knowledge about the method. Indeed, these private organizations appear to have been more important to DPS diffusion than public extension services.
However, the presence of these associations was not sufficient to induce most Brazilian farmers to adopt the DPS. Almost 90% of farmers did not adopt the technique despite the substantial increase in adoption since the 1990s. Adoption rates are similar across farm sizes and are higher in the crop-intensive areas in southern and central Brazil.
3. Data
Soil Heterogeneity
We build the soil heterogeneity measure using detailed GIS information from a Brazilian soil map developed by Embrapa, the Brazilian Agricultural Research Corporation (Embrapa Solos 2011). The data are based on an international soil classification system. This classification uses a hierarchical taxonomy: a hypothetical soil Aa1 belongs to order A, suborder a, and group 1. “Order” is the first and most general classification level, and the following levels are subdivisions. Although the classification system allows for finer levels, the map does not report information beyond the third level. Information is presented at a scale of 1:5.000.000 for each level.
The classification system is based on soils’ physicochemical composition. This composition is a major determinant of the physical properties of each type of soil. Physical properties, in turn, define the suitability for different agricultural methods. For example, in the case of the DPS, high soil temperature often calls for a thick layer of residue on the surface to decrease exposure to sunlight and avoid excessive heat. Hence, a different type of soil calls for some adjustment (a micro-innovation) in the use of the DPS. This classification system is smooth in the sense that, for agricultural purposes, the difference in physical properties between two soils is roughly the same for each soil pair.
The baseline empirical specification considers the most general level (order) to construct the soil heterogeneity measure. This level has the advantage of being invariant to land use. Although different practices may either enrich or impoverish the soil by affecting the levels of several nutrients, this process cannot go as far as to change its basic chemical structure. Therefore, this measure is independent of the adoption of the DPS or other agricultural practices. We provide evidence in the robustness section that results remain unchanged when one considers orders, suborders, and groups to construct the soil heterogeneity measure.
There are 35 different orders across Brazil. We merge the soil map with the map of municipalities to build a measure of the share of each municipality covered by each soil order. We use the same procedure to build measures for the other levels. These shares are used to construct a Herfindahl index of soil types for each municipality. High values indicate homogeneity and low values indicate heterogeneity.
Soil heterogeneity (S) is defined as the inverse of a municipality’s Herfindahl index of soil concentration. In line with the industrial organization literature, we interpret this index as a measure of the effective number of soils. S is chosen because it is simple to interpret. The variation in the data also makes it difficult to use other measures from the existing literature. It is important to highlight that S is an artificial variable with no direct impact on agricultural production. Panel A in nTable 1 reports descriptive statistics for the main soil heterogeneity measure. We trim the upper 1% tail of S’s distribution to remove some extreme outliers. The average value of S is 1.69 with variance 0.69.
Descriptive Statistics
It is important to note that our approach for measuring heterogeneity of growing conditions stands in contrast to that used by Munshi (2004). The author classifies regions as more or less heterogeneous using information on the crops cultivated in these regions. Our approach uses direct geographic information, being more similar to the approach of Michalopoulos (2012) and Fenske (2014), who correlate similar heterogeneity measures to determinants of economic development.
Agricultural Data
The outcome used in our empirical exercises is the DPS adoption rate. This measure is defined as the percentage of farms that use the DPS, and it is constructed using data from the 2006 Brazilian Agricultural Census.6 We restrict the baseline estimates to municipalities with adoption levels above 5% to ensure that we are investigating adoption in municipalities in which the DPS is viable. The results are robust to including municipalities below this threshold. We also exclude municipalities without information of one or more control variables. The results are also robust to including these municipalities in the specification without the full set of controls. The final sample has 1,681 municipalities.
Figure 1 displays the spatial distribution of adoption in Brazil. Adoption rates are higher in the areas where crop cultivation is concentrated, mostly in the states of Rio Grande do Sul, Santa Catarina, and Paraná in the South region and in the state of Mato Grosso in the Center-West region. However, there are municipalities with adoption rates above the 5% distributed throughout the whole country. This is important because our empirical design uses only variation within states to identify the effect of geographic heterogeneity on technology adoption.7
Distribution of Direct Planting System (DPS) Adoption across Brazilian Municipalities in 2006
Our empirical exercises further use the Brazilian Agricultural Census to construct controls used in the regressions. We build the following municipality-level controls: number of farms, average farm revenues, share of farms with tractors, share of farmers with more than 11 years of schooling, share of farmers with access to technical assistance, and share of farmers associated with cooperatives.
Moreover, we use administrative data on diffusion centers (Clube Amigos da Terra locations) to calculate the distance (in kilometers) to the nearest diffusion center, and administrative data on the location of bank branches to calculate the number of bank branches in each municipality. We calculate the number of Banco do Brasil and non–Banco do Brasil bank branches, because this financial institution is the principal supplier of rural credit in Brazil. Table 1, Panels B and C reports descriptive statistics of these variables.
Geographic Data
The empirical specifications include several other geographic and socioeconomic variables as controls. Average temporary rainfall and temperature for the period 1970 to 2010 are calculated for each observation using gridded data on rainfall and temperature obtained from the Terrestrial Air Temperature and Precipitation Version 3.01.8 Rainfall is measured in millimeters of precipitation, while temperature is measured in degrees Celsius. Land gradient is calculated using raster data collected from the 90-meter Shuttle Radar Topography Mission (SRTM) radar.9 We combine these data with a municipalities map to construct the average land gradient at the municipality-level. The gradient is measured in degrees. Agronomic potential is measured using raster data from the FAO Global Agro-Ecological Zones (GAEZ) dataset.10 We combine these data with a municipalities map to calculate the average agronomic potential to cultivate at the municipality level for each of the six main crops cultivated in Brazil (soy, maize, sugarcane, rice, beans, and cotton). Our measure of agronomic potential is based on potential yield for rain-fed cultivation in the high-input regime. The results are robust to using the intermediate input level instead. Latitude, longitude, and altitude were obtained through the Ipeadata website.11 Table 1, Panels D–F reports descriptive statistics of these variables.
4. Empirical Framework
The literature on technology adoption highlights that—conditional on the availability of the technologies—regional differences in the adoption of modern technologies might come from regional differences in the speed of the diffusion process (“slope”) and in its equilibrium level (“ceiling”) (Griliches 1957; Rogers 2003; Jackson 2010). Hence, we model the effects of geographic heterogeneity on DPS by letting it influence these two elements.
Let Ait denote the DPS adoption in municipality i and period t, and Si denote the level of soil heterogeneity in municipality i. Also let εit be an error term of municipality i and period t. We depict the relationship between DPS adoption and soil heterogeneity using the following empirical model:
[1]
This empirical model explicitly considers the dynamic nature of the diffusion process emphasized in the literature on the S-curve (e.g., Young 2009; Jackson 2010; Foster and Rosenzweig 2010). Moreover, it enables heterogeneity to influence technology use by affecting both its “ceiling” (β1) and its “slope” (β2).
The empirical relationship depicted in equation [1] might be generated by different theoretical mechanisms. For instance, Ellison and Fudenberg (1993) and Munshi (2004) show that heterogeneity between potential users of a technology might inhibit learning from peers. This increases the costs of adopting the technology and, therefore, might influence both the level and the rate of growth for the DPS adoption rate. The authors also argue that this mechanism will be particularly important for the diffusion of technologies, like the DPS, that require adaptation to local conditions. Other theoretical mechanisms might also connect soil heterogeneity and technology adoption. For instance, Alesina, Baqir, and Easterly (1999) discuss how heterogeneity in a population reduces the provision of public goods. To the extent that these public goods influence the technology’s returns, this might reduce the level (equilibrium) rate of adoption.
Irrespective of the theoretical mechanism, separately identifying the parameters from equation [1] requires data on technology adoption for multiple periods. This is a problem in our setting because data on DPS adoption are available for only one period. Nevertheless, it is possible to show that the relationship between DPS adoption and geographic heterogeneity in one period is informative about the magnitudes of these structural parameters.
Because equation [1] is valid for all periods, we can write technology adoption as a function of initial adoption and soil heterogeneity:
[2]
Suppose we can approximate the municipality fixed effect αi using observable geographic characteristics Xi. Moreover, notice that the term ρtAi0 converges to zero, since initial adoption is zero. Then, it is possible to rewrite equation [2] as
[3]
in which
and
Equation [3] can be estimated using cross-sectional information on DPS adoption and soil heterogeneity. The parameter β is the cumulative effect of soil heterogeneity on DPS adoption. Since technology adoption is persistent (ρ > 0), the sign of this parameter is informative about the relative magnitudes of the effects of geographic heterogeneity on the level and the growth of technology adoption across the municipalities in our sample. A negative β indicates either that both effects are negative or that the negative effect dominates the positive one, while a positive β indicates either that both effects are positive or that the positive effect dominates the negative one.12
The parameter β will be identified under the assumption that (conditional on the vector of covariates) soil heterogeneity is not correlated with the error term. This assumption will not be satisfied if there are geographic characteristics that are correlated with both soil heterogeneity and DPS adoption. The vector Xi includes numerous geographic characteristics to mitigate this concern. We include soil types, temperature, rainfall, altitude, land gradient, latitude, longitude, and agronomic potential for the six main crops cultivated in the country in this vector. We also include state fixed effects to control for unobservable characteristics of the states that might correlated with both soil heterogeneity and DPS adoption.
It is important to notice that because soil heterogeneity in neighboring municipalities might be correlated, the error term from equation [3] might be spatially correlated. This creates problems for estimating standard errors using methods (e.g., standard errors robust to heteroskedasticity) that assume the error term is independent across observations. To account for this issue, we cluster the standard errors at the microregion level. This allows for arbitrary correlation in the errors of municipalities located in the same microregion. There are 360 microregions in our main sample. In robustness tests, we use Conley’s (1999) method to estimate standard errors robust to the presence of spatial correlation using three different distance cutoffs (50 km, 100 km, and 150 km).
Measurement error is another important issue for estimation of equation [3] because our index of heterogeneity probably measures the relevant heterogeneity with noise. This will bias the estimates of the parameter β. Hence, we perform robustness tests using other indexes of soil heterogeneity to test whether our estimates are robust to the choice of heterogeneity index.
5. Results
The Effect of Soil Heterogeneity on Technology Adoption
Figure 213 graphically displays the relationship of interest in the paper. It shows that there is a negative relationship between soil heterogeneity and DPS adoption. This correlation provides prima facie evidence for the hypothesis that the adoption of modern agricultural technologies is obstructed by geographic heterogeneity.
Soil Heterogeneity versus Direct Planting System (DPS) Adoption
To test whether this correlation is only the result of the correlation between geographic heterogeneity and mean geographic characteristics, Table 2 reports estimates of equation [3] using different sets of geographic controls. Column (1) reports the coefficient from the regression of DPS adoption on soil heterogeneity, controlling for soil types, altitude, and gradient. The coefficient on soil heterogeneity is negative and significant at the 5% level. Its magnitude suggests that an increase of one standard deviation in soil heterogeneity (0.69) reduces the DPS adoption rate by 1.62 percentage points.
Soil Heterogeneity and Technology Adoption
Column (2) includes measures of agronomic potential as additional controls. These controls mitigate the concern that the relationship between soil heterogeneity and DPS adoption is driven by differences in DPS profitability for different crops. The estimates indicate that soil heterogeneity continues to exert a negative and significant effect on DPS adoption.
Column (3) includes temperature and rainfall in the vector of covariates. These variables help to ensure that differences in the climate not captured in the measures of agronomic potential drive the relationship between soil heterogeneity and DPS adoption. The coefficient on soil heterogeneity decreases but continues to be negative and significant at the 5% level.
Column (4) further includes latitude and longitude as controls. These coordinates help to mitigate the concern that our results are driven by some geographic characteristic not included in the controls from the previous columns. The coefficient on soil heterogeneity of this specification is quite close to the coefficient estimated in column (2).
Column (5) further includes state fixed effects as controls. The fixed effects control for state-specific unobserved geographic characteristics that might influence both soil heterogeneity and DPS adoption. This is our preferred specification. It indicates that a one standard deviation increase in soil heterogeneity reduces the DPS adoption rate by 1.49 percentage points.
While exhaustive, the controls do not rule out the possibility that unobserved geographic determinants of technology adoption drive the relationship we document in the data.14 However, Altonji, Elder, and Taber (2005) show that if the coefficients of interest change little with the inclusion of controls with a lot of explanatory power, the unobserved controls must be much more important than the observed ones to be driving the results. In the specifications from Table 2, the coefficient on soil heterogeneity decreases less than 10% (from 2.39 to 2.17), while the R-squared increases 60% (from 0.40 to 0.64) with the inclusion of state fixed effects and controls for agronomic potential, temperature, rainfall, latitude, and longitude. This indicates that these (unobserved) geographic determinants are unlikely to be driving this relationship.
In the Appendix, we use the method proposed by Oster (2017) to compute numerically the relative size of selection on unobservables compared to selection on observables required to zero the coefficients estimated in column (5) of Table 2. Appendix Table A3 reports that selection on unobservables would need to be between 2.87 to 4.97 times the size of the selection on observables to drive the effect of heterogeneity to zero. We also use this method to compute bounds on the effects of the soil heterogeneity under the hypothesis of proportional selection (selection on unobservables equal to selection on observables). Appendix Table A4 reports that the lower bound of the effects of soil heterogeneity on DPS adoption ranges between 81% and 94% of the effects reported in Table 2.
Together, the evidence provides support to the hypothesis that geographic heterogeneity influences technology adoption. While we are unable to separate whether the effect of geographic heterogeneity comes from a level or a growth effect, our results strongly indicate its net effect is negative. Because the results are robust to the inclusion of numerous geographic controls, this does not appear to be a consequence of correlation between our heterogeneity measure and other geographic determinants of technology adoption.
Mechanisms Linking Soil Heterogeneity and Technology Adoption
Let Mi be a mechanism that potentially mediates the relationship between soil heterogeneity and technology use. To examine the importance of this mediator, it is useful to compute the conditional direct effect, that is, the effect of soil heterogeneity not mediated by the mechanism Mi. A common procedure used in empirical work is to interpret the coefficient from a regression further controlling for the mediator Mi as this effect. However, because the mediator is endogenous to the treatment, this method will not identify the conditional direct effect, but a combination of this effect and a selection effect induced by the control (Rosenbaum 1984; Angrist and Pischke 2008). For this reason, we use the sequential g-estimator proposed by Acharya, Blackwell, and Sen (2016) to examine the role of different mechanisms in explaining the relationship between soil heterogeneity and technology use.
Let Wi be a vector of determinants of technology use that might be correlated with Mi.15 In the first stage of the procedure, we regress DPS adoption on soil heterogeneity, controlling for Mi, Xi, and Wi:
[4]
Then, we use the coefficient on the mediator obtained in equation [4] to demediate the outcome:
Finally, in the second stage of the procedure, we regress this demediated outcome on soil heterogeneity, controlling for Xi:
[5]
Acharya, Blackwell, and Sen (2016) prove the coefficient on soil heterogeneity obtained in equation [5] identifies the effect of soil heterogeneity not coming from its effect on the mediator under the hypothesis of “sequential unconfoundedness.” This hypothesis implies that, conditional on the geographic controls Xi, soil heterogeneity must not be related unobserved geographic determinants of either the DPS use or the mediator. Moreover, it implies that, conditional on soil heterogeneity, the geographic controls Xi and the economic controls Wi, the mediator must not be related to other unobserved geographic and economic determinants of the DPS use. While restrictive, it is possible to examine how the conditional direct effects identified using the sequential g-estimator are sensitive to violations in sequential unconfoundedness.
Figure 316 uses this method to examine whether the connection between soil heterogeneity and DPS adoption is mediated by the following variables: the log of the number of farms, the log of farm size, the share of farms that own a tractor, years of schooling, the share of farms with access to technical assistance, the number of Banco do Brasil branches, the number of non–Banco do Brasil branches, the share of farms that belongs to a cooperative, and the proximity to diffusion centers.17 The points report the effects of soil heterogeneity not coming from each of these mechanisms, while the lines report 95% confidence intervals. The effects of soil heterogeneity not coming from these mediators are close to the total effects estimated in Table 2. The size of the effects diverges by more than 5% for only two mediators (the number of farms and the share of farms that own a tractor), by more than 10% for only one mediator (the number of farms), and by more than 15% for no mediators. Appendix Figures A1–A3 report the results from sensitivity analyses that show the results are robust to violations in the identification hypotheses.
Effects of Soil Heterogeneity on Direct Planting System Adoption Conditional on Nine Different Mechanisms
These findings indicate that the impact of soil heterogeneity on DPS adoption is not mediated by economic determinants of adoption like farm size, education, credit, or technical assistance. A possibility is that soil heterogeneity reduces technology adoption by creating barriers to the diffusion of information about modern technologies (Ellison and Fudenberg 1993; Diamond 1997; Munshi 2004). To the extent that farmers learn from their peers how to use modern technologies, this might explain the relationship between soil heterogeneity and technology adoption. This mechanism might be of particular importance in the case of technologies that require adaptation to local conditions, such as the DPS. While we are not able to test this hypothesis directly, we can test it indirectly by examining whether the connection between soil heterogeneity and DPS adoption exists for technologies that do not require local adaptation.
Table 3 reports estimates of the relationship between soil heterogeneity index and the use of electric power and harvesters. Columns (1)–(3) present the results for the former variable, while columns (4)–(6) present the results for the latter. Columns (1) and (4) control for geographic characteristics. Columns (2) and (5) add state fixed effects. Columns (3) and (6) further add economic controls.
Soil Heterogeneity and Other Technologies
Columns (1)–(3) indicate that the use of electricity is not influenced by soil heterogeneity. The estimates indicate that soil heterogeneity does not influence the adoption of this technology. The coefficients are economically and statistically insignificant in all specifications. The evidence presented in columns (4)–(6) is consistent with the evidence from columns (1)–(3). The effects of soil heterogeneity on harvester use are statistically and economically insignificant regardless of the specification. This provides support for the hypothesis that geographic heterogeneity influences technology use through its effect on adaptation costs.18
Robustness
In Table 4, columns (1) and (2) examine the robustness of the results to different definitions of soil heterogeneity. First, we build a measure of heterogeneity using more detailed information from the soil map. While the original measure uses information of the most general soil class available in the map (order), this measure uses information of the more detailed classes available (order, suborder, and group). Because soil groups might be influenced by agricultural practices, this different measure is not as invariant to land use as the original one. However, it uses more variation to construct the soil heterogeneity index than the baseline measure. Second, we build a measure of heterogeneity using information on soils from neighboring municipalities. The measure is the weighted average of the soil heterogeneity from the microregion in which the municipality is located. The weights are the areas of each spatial unit. This alternative measure considers that farmers can learn from farmers located in other municipalities as well.
Robustness
Column (1) reports the results obtained using the soil heterogeneity classification built using information on soils orders, suborders, and groups. The specification is the same as the one used in Table 2, column (5). The coefficient on soil heterogeneity is negative and statistically significant at the 1% level. It implies that an increase in one standard deviation in this index of soil heterogeneity (about 0.92) reduces DPS adoption about 1.8 percentage points. This is about 20% more than the results obtained using the original soil heterogeneity classification. Column (2) reports the results obtained using the soil heterogeneity of the whole microregion in which the municipality is located. The coefficient on soil heterogeneity continues to be negative and becomes significant at the 1% level. A one standard deviation increase in soil heterogeneity (about 0.89) reduces DPS adoption by 2.2 percentage points, roughly 50% more than the effect obtained using the original measure. Together, these results indicate the effect of soil heterogeneity documented in our preferred specification might be the lower bound of the effect of soil heterogeneity on DPS adoption.
Table 4, columns (3) and (4) examine the robustness of the results to the inclusion of municipalities in which the share of farms using the DPS is below 5% in the sample. Our preferred specification excludes these municipalities to ensure we are focusing on municipalities in which the technology is known to be viable. However, it is useful to consider other thresholds, to ensure the findings are not entirely driven by the threshold used in our preferred specification.
Column (3) includes all municipalities with information on the geographic controls and positive DPS adoption. The sample increases to 4,480 municipalities. The coefficient on soil heterogeneity becomes significant at the 1% level. However, its absolute value is smaller. Because learning from peers is not relevant in municipalities with few adopters, this result is consistent with hypotheses in which soil heterogeneity influences DPS use through its effect on the costs of learning. Column (4) includes all municipalities with information on the geographic controls. The sample increases to 5,331 municipalities. The coefficient on soil heterogeneity continues to be negative and significant at the 1% level. Its magnitude is roughly 10% smaller than the coefficient from the previous column.
In the Appendix, we additionally test whether the results are robust to different hypothesis on the structure of the error term and to the use of weights. Appendix Table A6 provides evidence that the results are robust to estimating standard errors using the method proposed by Conley (1999), while Appendix Table A7 shows that the results are robust to using either the municipality area or the municipality farmland as a weight.
6. Conclusion
Low adoption of modern technologies is the object of extensive research in economics because of its impact on economic development. This paper provides evidence indicating that low adoption of modern technologies in agriculture is deterred by geographic heterogeneity, using data from the adoption of the DPS in Brazil.
Conditional on geographic characteristics and socioeconomic characteristics, we estimate a negative relationship between an index of soil heterogeneity and DPS adoption. The results also indicate no relationship between soil heterogeneity and technologies in which adaptation costs are not relevant. Robustness checks indicate that these findings are unlikely to be driven by unobserved determinants of technology adoption.
Our findings illustrate that the process of learning can be deterred by dissimilarities across adopters and supports the theoretical mechanism highlighted by Ellison and Fudenberg (1993) within a context where large-scale agriculture is prevalent and for a production technique with important private and public benefits. The results have important implications for policies aimed at promoting technology adoption. Provisional and focused training can induce substantial adoption in homogeneous areas because social learning will induce technological diffusion in these areas. Nevertheless, lasting and widespread investments in training will be needed to induce adoption in more heterogeneous areas, because adaptation will be more difficult in these areas.
Acknowledgments
We are grateful to Gustavo Gonzaga, Bernardo Mueller, Leonardo Rezende, Romero Rocha, Rodrigo Soares, Alban Thomas, and seminar participants at CPI, FUCAPE, USP, the European Meeting of the Econometric Society, the North American Meeting of the Econometric Society, the Latin American Meeting of the Econometric Society, and the Brazilian Meeting of Econometrics for comments and suggestions. We thank Ana Ribeiro and Fabio Magrani for excellent research assistance.
Footnotes
↵1 Foster and Rosenzweig (2017) argue that transaction costs in labor markets reduce the size of farms and, as a consequence, the profitability of modern inputs in Indian agriculture. Suri (2011) discusses the role of differences in comparative advantage in explaining adoption of modern technologies in agriculture in Kenya. Foster and Rosenzweig (1995) and Conley and Udry (2010) document the role of social learning in technology adoption in India and Ghana. Karlan et al. (2014) investigates the role of market failures in investments and technology adoption in Ghana. Duflo, Kremer, and Robinson (2011) show the role of behavioral biases in explaining adoption decisions among Kenyan farmers.
↵2 DPS adoption is also associated with an increase in carbon sequestration and a reduction in greenhouse gas emissions in agriculture. See West and Post (2002) and Metay et al. (2007) for evidence on these gains in different settings.
↵3 A related contribution is by Griliches (1960), who documents how geographic differences in the use of hybrid corn may be explained by distinct profitability.
↵4 See Foster and Rosenzweig (2010) for a recent review of this literature.
↵5 The optimal choice of inputs often changes under a new technology. However, this is different from change of inputs given a technology.
↵6 See https://sidra.ibge.gov.br/pesquisa/censo-agropecuario/censo-agropecuario-2006/segunda-apuracao.
↵7 Appendix Table A2 reports the average adoption and the number of municipalities in each state of the country. There is more than one municipality with adoption above 5% in all states but the Federal District (which is not divided in municipalities as the other states). In the sample used in the empirical analysis, average adoption is 30%. In all municipalities, it is close to 10%.
↵8 See http://climate.geog.udel.edu/~climate/html_pages/archive.html.
↵9 See http://srtm.csi.cgiar.org/.
↵10 See http://gaez.fao.org.
↵11 See http://www.ipeadata.gov.br.
↵12 It is possible to formalize the discussion from the main text using the following expressions:
↵13 This figure reports the bivariate relationship between soil heterogeneity on DPS adoption for different adoption levels. The black dots plot the average DPS adoption rate in 40 evenly spaced bins of soil heterogeneity. The blue line plots the fit of a nonparametric regression (LOESS) of DPS adoption on soil heterogeneity.
↵14 It is important to distinguish between unobserved geographic determinants of technology adoption and other unobserved determinants of technology adoption. The former are exogenous to heterogeneity, while the latter are endogenous to it. For this reason, the unobserved geographic determinants threaten identification, while the other unobserved determinants are mechanisms that might explain the relationship between heterogeneity and technology adoption.
↵15 It is important to highlight that these controls differ from the geographic determinants Xi discussed before, because they might be influenced by soil heterogeneity.
↵16 The points represent the point estimates, while the lines represent 95% confidence intervals. All coefficients are computed using the method proposed by Acharya, Blackwell, and Sen (2016). For comparison, the vertical line reports the total effect of soil heterogeneity on DPS adoption.
↵17 We do not observe a continuous measure of schooling. We construct it using information on the share of farmers with different levels of schooling. We attribute 0 years of schooling to farmers with no education, 4 years to farmers with incomplete elementary education, 8 years to farmers with elementary education, 11 years to farmers with high school education, and 15 years to farmers with college education. The results are robust to using the share of farmers with 11+ years of schooling as our measure of human capital.
↵18 In the Appendix, we formalize these ideas in a theoretical model that shows how information barriers might generate a relationship between geographic heterogeneity and technology adoption. In the model, profit-maximizing farmers must decide whether to adopt a new technology that requires adaptation to be locally adapted. Nonadopters must adapt the new technology from the current adopters, and adaptation entails a cost that increases as the difference between farmers’ geographic characteristics increases. These assumptions capture the intuition that farmers learn from their peers (e.g., Foster and Rosenzweig 1995; Conley and Udry 2010) and that it is easier to learn from similar peers than from different ones (e.g., Ellison and Fudenberg 1993; Munshi 2004). This model predicts that adaptation costs will be greater in more heterogeneous regions than in more homogeneous ones, because farmers are located farther from each other in the former regions than in the latter ones. As a consequence, fewer farmers will find it profitable to adopt the DPS in regions where geographic heterogeneity is high. Our results are consistent with this prediction.
