When does higher credit lower forest area? Amazon contexts change credit’s impacts

Significance

Agricultural and pastoral activities (crops and cattle) drive forest loss in the Brazilian Amazon and globally. Factors which support them—such as rural credit—can drive each to expand over forests. Yet, credit can instead raise productivity and spare native vegetation. Examining cattle and crops separately, and distinguishing land use and organizational contexts at the municipal level, we show relevant variation in the magnitude—and even direction—of credit’s forest impacts across a vast and diverse Brazilian Amazon. Our typology of contexts can help policymakers align rural credit with development and conservation.

Abstract

Economic activities have transformed over half of the Earth’s surface and the large-scale conversion of rural forests to agropastoral production is ongoing. Raising access to credit may lead producers to expand at the expense of the forest or, instead, invest to raise yields, reducing forest loss per unit output. Looking across Brazil’s enormous Amazon, we show that extremely different local contexts differ in terms of which effect of credit dominates. We distinguish those contexts using factors that affect cattle and crop production. For each of these two outputs, we find distinct credit impacts by context. Cattle credit raises deforestation where pasture expands over native forest on far frontiers. Crop credit raises deforestation where farms expand over pasture and forest remnants. Understanding impacts heterogeneity can help to reduce losses of rural forests, even while credit is supporting local economies.

Agriculture and livestock compete with forests for land. Thus, support for agropastoral activities can raise forest loss. Yet, the literature’s estimates of rural credit’s effect on vegetation loss are mixed, in magnitude and even direction. Rural credit can raise forest loss (1–3) when it supports the expansion of lands in agropastoral activities. Yet, credit can have zero impact on or reduce forest loss when it raises agropastoral yields per unit land (4). We offer a simple typology of distinct contexts to help explain when each of these heterogeneous impacts arises.

Forest impacts of credit (5, 6) are shaped by land governance and by markets, which affect land expansion (7, 8). We go beyond existing work showing a single credit impact on forest in demonstrating how that impact varies by sector (crops vs. cattle) and by context (less vs. more isolated). We do so for Brazil’s Amazonia, a region with the size and geophysical and socioeconomic variation of a continent. This informs optimal rural-credit policies to balance development and conservation.

We describe contexts using their distinct shifts in production and forest before changes in credit access. Across Brazil’s enormous Amazon, municipalities featuring greater occupation, with contested property rights—often smaller, older frontiers with fragmented forest—tend to be more “governance dense” (9). Moving out from those centers, organized rural production is both considerable and still expanding. Even further out, native forest is prevalent in larger, less accessible municipalities that have featured lower formal land governance (10). These areas, with limited legal land titles and more undesignated public lands, present opportunities for land invasions as well as for deforestation as a means to claim land ownership (11–13).

These preexisting conditions influence how rural credit affects deforestation. In part, that is due to past credit supply. It was easier in the well-established centers but challenging for remote frontiers—suggesting more impact there for current credit shocks. Public monitoring and enforcement also differ considerably across contexts, again given challenges on remote frontiers. In sum, raising access to rural credit is less likely to reduce forest within governed contexts that long had credit access. There, given scarce and costly lands (8), substantial credit even may encourage intensification (higher yields, smaller footprints, “forest sparing”). Conversely, raising rural credit is more likely to reduce forest, via expansions of production areas, in less-regulated remote regions where land is abundant (14, 15).

These hypotheses concerning large landscapes align with a pattern noted for the Brazilian Amazon and other parts of South America: Pasture replaces forest; then crops occupy “cattle-opened” areas (16, 17) where early roads and frontier institutions reduced transport and transaction costs (18). In light of these dynamics, we examine cattle and crop dynamics separately. For each sector, we estimate the impacts of rural credit on forest loss for each context. We study temporal changes to deal with time-invariant factors and we use a shift-share instrument for credit supply to address endogeneity, such as reverse causality when deforestation-based growth raises demand for rural credit. The locally exogenous credit “shift” part of our shift-share instrument is computed using the temporal variation in total credit supply from each financial institution (for all municipalities excluding the one in question). The baseline, preshock credit fraction of each institution in the municipality in question provides the “share” in our instrument. We construct local-credit-shift instruments separately for cattle and crops to allow for sector-specific dynamics. Thus, we use the variations in local credit which are exogenous to isolate the effects upon forest from credit-supply shocks (vs. effects in the other direction).

We combine annual deforestation with rural credit data from 2006 to 2017 for 452 municipalities with at least 50% of their area in the Amazon biome (excluding coasts, the Cerrado biome, and state capitals). We use land-cover data from MapBiomas’ 6th Collection, based upon Landsat images at a 30 m spatial resolution for 1985–2019, focusing on the land use transition data that distinguish cattle from crops. For formal occupation, we use registered farms from the National System of Rural Register (Sistema Nacional de Cadastro Rural, SNCR), now integrated into the new Land Management System (Sistema de Gestão Fundiária, SIGEF) (19–21). We gather the number of certified private lands in SIGEF serves as our measure for the agribusiness large farms linked to global markets. Rural credit data are from Municipal Bank Statistics (Estatística Bancária Mensal por Município), on a monthly basis, accessed via the Brazilian Central Bank website. These data differentiate between credit for agriculture and credit for cattle, again keeping separate these rural production sectors. We use Central Bank data to compute total lending for each financial institution, which is used to calculate the baseline shares of each financial institution’s credit allocation within our shift-share instrument for local shocks to credit supply.

We distinguish our precredit-shock contexts using land cover, expansion of agropastoral production, municipal area, and formal land occupation. To estimate the effects on forest of raising rural credit access, we employ panel regressions using logged credit (again with a shift-share instrument). As anticipated for each productive sector (crops vs. cattle), in looking across our contexts, from older to more recent frontiers, the forest impact of credit differs significantly in magnitude, statistical significance, and even direction.

In the older, smaller, more formally governed areas with fragmented native vegetation (our 1st and 2nd contexts), credit was available previously and neither cattle nor crop credit shocks raised forest loss. If anything, increased access to substantial credit may have reduced forest losses. In the larger, more formalized crop production (our 3rd context), for crops but not cattle greater access to rural credit raised forest losses. Finally, within the more remote, less densely populated and less occupied municipalities characterized by larger and less formally governed areas (in 4th and especially 5th context), for cattle but not crops greater credit access raised forest losses. Separating contexts allows policy to better balance agropastoral production and forest conservation.

Theoretical Framework

Widely applied in geography and economics, the von Thünen model explains spatial patterns of agricultural production using costs of transporting goods to consumer centers (22). Its focus on distance to market implies spatial gradients: Goods with high transport costs, relative to the good’s market value, become economically infeasible as market distance increases. Lands near centers produce goods with high transport costs, while far from center hosts goods with low transport costs.

This idea has been applied to explain the dynamics of land-use change in tropical frontiers (15, 23), including for deforestation, our outcome (5, 24). Active forest frontiers feature rapid exploitation of land and natural resources, at times linked to global commodity markets (18). Some drivers are local, including supply chain agglomerations (25), while others are external and exogenous such as global product demand, technological innovation, federal infrastructure (15)—e.g., new roads that raise economic returns and crop expansion (26, 27)—and other policies such as taxes and subsidies, including for rural credit (28). When an abundance of natural resources is confronted with favorable drivers, rapid and even nonlinear agropastoral expansion can occur (15).

A common trajectory on such frontiers, at least in Latin America, begins with land appropriation (that may displace indigenous and other forest dwellers who lack land titles). Cattle expansion follows on low-priced lands which are sold later at higher prices to larger-scale land speculators who formalize titles and make land available for conversion to production of commodities (18). As all this advances, states and corporations promote surveying, monitoring, taxation, and governance (15), often raising the cost of deforestation including through fines (29, 30). When the resources are exhausted, or regulation is established, rapid transformation ceases and a frontier may be considered closed (18).

We focus upon the heterogeneity of credit’s forest impacts across these contexts, for crops and separately for cattle (Fig. 1). We describe and document a spatial gradient of land formalization and governance [or “territorialization” per “value creation” (15, 18)] that moves outward from closed frontiers with settled parcels and little native vegetation all the way to little-occupied native forests that lack legal registrations but feature abundant space for clearing and land grabbing. Between those extremes lie active frontiers for natural resources and commodities which are pushed forward by both exogenous and internal driving processes (Fig. 1).

Fig. 1.

Theoretical Framework: rural credit’s impact on deforestation under varying conditions of land dynamics, organization, and governance.

Our empirical typology describes each context using three observed factors (Fig. 2). First, precredit-shock land use. If it is steady, vs. rising, combined with if native vegetation remains abundant, distinguishes postfrontier, active-frontier, and prefrontier (15). Second, municipal area: In the Amazon (and Brazil as a whole), municipalities are divided into smaller units as a consequence of decision-making decentralization to new local governments, stimulated by fiscal gains (31) and influenced by urbanization and increased organization of economic sectors (32). Thus, we assume that small municipal areas indicate the presence of a local organization process. Third, formal land registration. It is a key feature of land governance in Brazil and a strong proxy of the presence of agri-businesses whose production is linked to international markets and intensely employs capital and specialized labor.

Fig. 2.

Classification tree, with spatial representation of each step. Maps (A–E) show the criteria used to compose conditions on land organization and governance (F). First, municipalities were classified according to land use and stability (C), which is based on vegetation cover (A) and rural expansion rate (B) data. Then, municipal area (D) is used to distinguish small and big municipalities, and finally, formal land occupation (E) is used to identify high and low occupation by large private farms. The numbers in parentheses refer to the number of municipalities that fulfill that criterion at that point of the tree.

Using those three factors, we distinguish five contexts, Conditions #1 to #5 (Fig. 2). Condition#1 has small municipalities with dense populations and high shares of formal land registration, indicating “old” (“postfrontier”) areas (18, 24). They have lost most native forest, featuring steady native cover at or below the 20% required by law (for native grasslands). Condition#2 also features small municipalities with high population density, yet less formally registered private land and more native vegetation, consistent with extraction instead of agri-business. Condition#3 has larger municipalities that are expanding in crop production, with intermediate vegetation cover and registered farms. In this commodity frontier, we expect intermediate but rising roads, other infrastructure, and governance. Finally, Conditions #4 and #5 feature big municipalities with low land tenure. Condition#5’s prefrontiers have the largest forest cover. In these remote areas, which are more difficult to access, small populations in villages (not cities) have low infrastructure and development. A few, often longstanding small producers are disconnected from markets, especially within #5, and large indigenous populations exist. Government is frequently absent. Pressure to occupy land and to change land-use cover is increasing or such pressures are in the neighborhood.

Finally, crops and cattle differ in their spatial dynamics; thus, we examine credit and deforestation by sector. Cattle notably relates to illegal land occupation in forested frontiers. Often, for instance, some heads are sold to intermediaries, not directly to slaughterhouses (33). Large-scale agricultural commodities link to international markets and may promote deforestation through processes considered legal (34).

The above suggests hypotheses by Condition (Fig. 1). In #1 and #2, we expect rural credit will not increase the loss of forest for crops or cattle. There is no cheap (“no man’s”) land for expansion. There is land governance and environmental monitoring that restricts rural land expansion. More credit might even support intensification that could reduce the area. In #3, however, agriculture expands over cheaper lands that include old pastures but also large forest remnants. Thus, agricultural credit might support crop expansions that clear remaining forests. In #4, we instead would expect cattle to expand rural frontiers, with no or few constraints. Credit to cattle can raise deforestation, in light of currently low formal occupation of the land and less economic activity, which imply opportunities to expand and deforest. Finally, #5 is both figuratively and somewhat literally right around the corner on remote forest frontiers. We would, then, again expect credit to cattle to increase deforestation.

Results

To investigate how land contexts (our Conditions #1 to #5) shape the impacts of rural credit on forests, we run panel regression analyses for each sector (cattle and crops) for each context. We use the log difference in credit between years to capture credit variation (credit) in Eq. 4, employing an instrumental variable (see more below) to address the potential endogeneity between rural credit and deforestation. Our coefficients represent the increase in deforestation due to an increase in credit (and not the opposite effect of an increase in credit loans as a consequence of raised deforestation).

Relationship Between Credit Supply and Demand.

Our first stage investigates if credit ($C$) is strongly driven by our instrumental-variable measure of credit-supply shocks ($Z$), as described in Material and Methods. First-stage results are presented in Table 1, with columns (1) to (5) for crops and (6) to (10) for cattle (one column for each sector for each Condition). Those results show that the shift-share instrument, when interacted with the relevant Condition, is positively correlated with rural credit for both crops and cattle, for all of our Conditions. This means credit-supply shocks raised rural credit across Brazil’s Amazon. The F-statistics for the excluded instruments show strong correlations, ranging from 8.4 to 39.4 for agriculture and from 3.9 to 42.6 for cattle.

Table 1.

Impact of credit supply shock on the rural credit growth (agriculture and cattle raising) under five different conditions of land organization in the Amazon Biome

	Credit growth
	Agriculture					Cattle
	C1 (1)	C2 (2)	C3 (3)	C4 (4)	C5 (5)	C1 (6)	C2 (7)	C3 (8)	C4 (9)	C5 (10)
Z $\times$ C1	2.474∗∗∗ (0.935)	0.028 (0.052)	−0.053 (0.084)	−0.019 (0.082)	−0.026∗∗ (0.013)	0.086 (0.062)	−0.006 (0.004)	−0.043∗∗	−0.010 (0.008)	0.001 (0.001)
Z $\times$ C2	$-$1.932 (1.198)	0.848 (0.732)	−0.168∗∗ (0.083)	−0.035 (0.059)	−0.020 (0.023)	1.384∗ (0.832)	1.722∗∗ (0.827)	0.028 (0.082)	−0.163∗ (0.089)	0.010 (0.016)
Z $\times$ C3	−0.339 (1.071)	−0.013 (0.030)	2.206∗∗∗ (0.535)	0.032 (0.053)	0.007 (0.007)	0.543∗∗∗ (0.208)	−0.017 (0.012)	0.795∗∗∗ (0.201)	−0.070∗∗ (0.028)	0.003 (0.003)
Z $\times$ C4	−0.979 (1.052)	0.007 (0.030)	−0.054 (0.063)	1.635∗∗∗ (0.480)	0.004 (0.014)	0.582∗ (0.322)	−0.041∗ (0.024)	−0.014 (0.051)	0.819∗∗∗ (0.3162)	0.001 (0.006)
Z $\times$ C5	−0.604 (1.090)	−0.018 (0.042)	0.008 (0.077)	0.278∗∗ (0.140)	1.689∗∗∗ (0.559)	0.699∗∗∗ (0.246)	−0.0002 (0.001)	0.047 (0.056)	−0.311∗∗ (0.152)	1.152∗∗∗ (0.186)
F-test of excl. instr.	20.8∗∗∗	8.4∗∗∗	21.6∗∗∗	22.0∗∗∗	39.4∗∗∗	3.9∗∗∗	15.4∗∗∗	4.3∗∗∗	4.9∗∗∗	42.6∗∗∗

Notes: Credit growth refers to the difference in the log of the rural credit. Z are the instrumental variables (for agriculture and for cattle) and refer to our measure of credit supply shock, explained in Eq. 2. All regressions have 3,164 observations and include control variables (precipitation and squared population density), municipality fixed effects, state-time trend fixed effects, and clustered SEs by municipalities (in parentheses). As for conditions, C1 up to C5 indicate conditions 1 to 5, respectively. P-values: ***P$<$ 0.01, **P$<$ 0.05, and *P$<$ 0.1.

Credit Impact on Deforestation.

Table 2 presents our estimates of rural credit’s impact on deforestation (OLS in columns 1 and 3 and IV in Columns 2 and 4). IV estimates deal with endogeneity biases, downward (Cattle C#5) or upward (Agri C#2), and at times shift significance (Agri C#1). Upward bias may indicate that decisions to deforest led to more credit, e.g., with urbanization in Condition #2. In Condition #5, the negative bias likely reflects the link, across these huge landscapes, between low credit access and high deforestation feasibility.

Table 2.

Impact of rural credit on deforestation under five conditions of land organization, Amazon Biome, OLS, and IV estimates (for unweighted regressions)

Dependent variable:	Log diff native		Log diff native
	Conversion to agriculture		Conversion to pasture
	(1)	(2)	(3)	(4)
Model:	OLS	IV	OLS	IV
Agri Credit	$-$0.0519∗ (0.0292)	−0.0323 (0.0776)
Agri Credit $\times$ C2	0.0984∗∗∗ (0.0356)	0.0880 (0.2383)
Agri Credit $\times$ C3	0.0474 (0.0315)	0.1814∗∗ (0.0892)
Agri Credit $\times$ C4	0.0509 (0.0312)	0.1770 (0.1123)
Agri Credit $\times$ C5	0.0706∗∗ (0.0322)	−0.0104 (0.0984)
Cattle Credit (C1)			0.0112 (0.0084)	−0.1222∗ (0.0739)
Cattle Credit $\times$ C2			0.0016 (0.0109)	0.0546 (0.0982)
Cattle Credit $\times$ C3			−0.0119 (0.0097)	0.0409 (0.0740)
Cattle Credit $\times$ C4			$-$0.0001 (0.0094)	0.2993∗∗∗ (0.1070)
Cattle Credit $\times$ C5			0.0529∗∗∗ (0.0181)	0.4922∗∗∗ (0.1268)
Fit statistics
Observations	3,164	3,164	3,164	3,164
R2	0.02375	$-$0.05081	0.04819	$-$0.25072
Within R2	0.00269	$-$0.07348	0.02341	$-$0.28329

Notes: Deforestation refers to the growth rate or the difference of the logarithm of the deforestation rate over time. Credit growth refers to the difference in the logarithm of the rural credit. All regressions include control variables (precipitation and squared population density), municipality fixed effects, and clustered SEs by municipalities (in parentheses). IV regressions also include state-time trend effects. As for conditions, C1 up to C5 indicate conditions 1 to 5, respectively. P-values: ***P$<$ 0.01, **P$<$ 0.05, and *P$<$ 0.1.

Results of credit shocks (Columns 2 and 4) indicate that rural credit raises deforestation for crops in Condition #3, and for cattle in Conditions #4 and #5, yet for neither in Conditions #1 or #2. If anything for Condition #1—small municipalities occupied by large farms and largely devoid of forested areas (SI Appendix, Table S1)—we find a suggestion of lower forest loss (credit driving intensification). SI Appendix, Tables S2–S5 show results when weighting municipalities by baseline forest area. They are qualitatively the same—lacking only the negative effect of credit on deforestation in Condition #1 when we weight by native area (in hectares) in the baseline (SI Appendix, Table S3).

For where credit raises deforestation, Condition #3’s statistically significant effect of 0.1814, for crops, means that for credit growth of 1%, deforestation would rise by 0.18%. As we move to Conditions #4 and #5, cattle credit raises deforestation. Conditions #4 and #5 share features, yet differ, consistent with cattle credit’s impact rising in significance and magnitude as we move outward across contexts from Condition #3 (even while crop credit’s impact falls in that movement outward). Credit for cattle has its greatest impact on forest loss in Condition #5, at about 0.49 (a rounded half percent rise for each percent raise in credit). We note Conditions #4 and #5’s SEs are relatively higher too, suggesting variations also within each of those contexts.

Sensitivity Analysis.

To explore what drives these results, we break each of our contexts (Conditions #1 to #5) into its components (land use stability, municipal area, degree of occupation by large formally registered farms). We then test for effects of each of these underpinning factors, using the same specification. More sensitivity analyses—weighting regressions by baseline forest area, varying samples and variables’ limits, and aggregating conditions—are in SI Appendix.

Table 3 shows that upward local shocks to rural credit for crops raise deforestation when agropastoral production has been expanding and with a higher degree of occupation by large farms capable of such expansion. Local credit supply shocks for cattle if anything have forest-sparing effects in small municipalities, far from distant frontiers, yet raise deforestation when agropastoral output has been rising (though, unlike for crops, this is with few large registered farms). Cattle credit also raises forest loss on distant frontiers, where native vegetation was stable. Consistent with that, cattle credit raises deforestation in larger municipalities.

Table 3.

Sensitivity analysis for the impact of rural credit on deforestation in the Amazon Biome, unweighted IV estimates

Dependent variable:	Log diff native
	Conversion to agriculture	Conversion to pasture
Model:	(1)	(2)
Panel A: Heterogeneity by Land Use
Credit (Agro stable)	−0.0331 (0.0772)	−0.0992 (0.0626)
Credit $\times$ Agro dynamic	0.1713∗∗ (0.0863)	0.1044∗ (0.0608)
Credit $\times$ Native dynamic	−0.0145 (0.1027)	0.3838∗∗∗ (0.1012)
Panel B: Heterogeneity by Municipal Area
Credit (Small municipalities)	0.0401 (0.1380)	−0.0823∗ (0.0455)
Credit $\times$ Big	0.0723 (0.1362)	0.1702∗∗∗ (0.0534)
Panel C: Heterogeneity by Large Farms Occupation
Credit (High occupation)	0.1178∗∗ (0.0461)	−0.1105 (0.0693)
Credit $\times$ Low	−0.0326 (0.0756)	0.2063∗∗∗ (0.0762)
Total observations	3,164	3,164

Notes: Deforestation refers to the growth rate, or the difference of the logarithm of the deforestation rate over time. “Credit” refers to the difference in the logarithm of the rural credit. All regressions include control variables (precipitation and squared population density), municipality fixed effects, state-time trend effects, and clustered SEs by municipalities (in parentheses), and the credit is instrumented by our measure of credit supply shock. Each panel and column shows the results for a unique regression: Panel (A) tests for credit effect when only land use and stability is considered (1–agricultural credit affecting conversion to agriculture; 2–cattle credit affecting conversion to pasture); Panel (B) when only municipal area is considered (column 1 for agricultural and 2 for cattle); and Panel (C) when only large farms occupation is considered for agriculture (column 1) and pasture (column 2). P-values: ***P$<$ 0.01, **P$<$ 0.05, and *P$<$ 0.1.

Discussion

We showed empirically that easily distinguishable contexts differ in how rural credit affects loss of important forests, for the Brazilian Amazon, in each of the two dominant rural sectors (cattle and crops). The significance, magnitude, and even direction of credit’s impacts vary across our contexts.

Examined separately or aggregated into our Conditions, the factors in our context typology across an enormous landscape—land use, municipality size, and formal land registration—influence the forest impacts of this important development policy. This heterogeneity runs contrary to prior claims to find the universal impact of rural credit. So too does our distinguishing crops from cattle. This heterogeneity includes that opposing results when credit raises productivity—diminishing deforestation (Borlaug’s hypothesis) or raising it (Jevons paradox or Rebound hypothesis)—can coexist in the same region, each for distinct contexts.

Specifically, for distant frontiers with low population and land registration but high forest [our Conditions #4 and #5 that correspond to resource or primary frontiers (15, 18)], cattle credit raised deforestation. Cattle historically signal land tenure and can be associated with land disputes (12), removal of native vegetation on Amazon frontiers (24) and opening for crops (16–18). This dynamic is more powerful when constraints such as surveillance by locals, governments, civil society, or markets are absent. For such Amazonian contexts, making rural credit available accelerates forest loss.

That contrasted greatly with our Condition #1. There, with very different characteristics, in which frontier dynamics have ended and formal institutions have been established, if anything more cattle credit lowered the rate of forest loss. In between those Amazonian realities, i.e., for our Conditions #2 and #3, we found no impacts from greater cattle credit.

Turning to crops, within sites previously cleared for cattle, conditions can be favorable for large private farms, such as for soy (35). As pasture productivity falls (36), degradation results in cheap land attractive for agri-business expansion (16, 18), including over postcattle forest remnants. Further, such agri-business is more demanding of financial capital and thus credit. While the infrastructure it uses for transport and enhanced formal institutions can make governance easier (15, 24), still credit directed to crops raised deforestation for this context (our Condition #3), as is typical of commodity or secondary frontiers (18, 28). This contrasts strongly, though, with more developed Conditions #1 and #2 where increased access to rural credit for crops had no significant effect on deforestation (similar to cattle credit). No crop credit effect was also true for the more distant Conditions #4 and #5 in resource and new frontiers—where cattle dominates.

Thus, for the majority of the massive Brazilian Amazon, some rural credit increases deforestation: crop credit in our Condition #3; and cattle credit in #4 and #5 (even on isolated (pre-)frontiers that might seem beyond the reach of such credit shocks). Finding that raising credit access tended to raise frontier forest loss supports an interpretation of low governance (37, 38) allowing exogenous drivers to accelerate frontier activation (15, 39), especially for cattle. Still, we also showed credit may not harm the forest when alongside effective constraints (Conditions #1 and #2) (40). It even can help forests, in the most constrained and highly capitalized situations (for cattle credit in Condition #1’s old-frontiers).

In sum, deforestation could rise or fall due to increased credit—even in a single region—while crop and cattle credit’s impacts vary by context. Whether credit-supply shocks lower, raise, or have no impact on native vegetation areas depends on economic and socioinstitutional (historical and policy) factors that are relevant for land use and easily observed.

As credit has such heterogeneous impacts, policymakers should consider credit’s contexts. Credit supply in Brazil reflects a combination of national policy decisions, macroeconomic conditions (such as commodity prices and income), as well as local demand and land-use dynamics (local population increase and the development of agglomeration economies). For sites where credit will raise deforestation, restricting credit access can lower it, but also land governance before credit may be sensible (15, 37). Credit restrictions linked to environmental compliance have been implemented in the past (Resolution 3,545) and proved effective in curbing deforestation (41), avoiding competition with existing environmental policies when fostering development, for example, via the National Development Bank (in Portuguese BNDES) (42). However, Resolution 3,545 was withdrawn in 2021 [along with other environmental setbacks (43, 44)].

Other complementary policies can help forests too. Policies that enhance surveillance in municipalities with high deforestation rates such as the Priority List policy can reduce resource exploitation (30, 45), as can improved land tenure for indigenous people as well as protected areas on forested frontiers (24, 46), and also fostering local sociobiodiversity production with low-cost state investments such as expanding access to information and technical training (47). Another complementary approach would be to integrate environmental value (of carbon but also biodiversity, water, and soil fertility) into lending decisions. Financial institutions could align credit with Brazil’s environmental and climate goals, while incentivizing borrowers to adopt sustainable land uses. In a balanced, sustainable development trajectory, credit could promote rural development alongside forest conservation.

Materials and Methods

Data.

We combine municipal-level annual deforestation data with rural credit information from 2006 to 2017 for all (504) municipalities within the Amazon Biome considering the ones with at least 50% of their area within the Amazon Biome’s geographic boundaries as defined by MapBiomas. Municipalities from the Legal Amazon but within the Cerrado Biome, as well as state capitals and coastal municipalities, were excluded from our sample.

Land uses and deforestation.

We use land-cover data from the MapBiomas Project 6th Collection (48), which offers annual, cloud-free, and automatically classified land-cover data based on Landsat images with a 30 m spatial resolution, spanning from 1985 to 2019. Our analysis focuses on the period from 2006 to 2017. We used the land use transition data matrix from MapBiomas, which distinguishes areas that transitioned from forested to deforested, at a pixel scale. We distinguished deforestation due to the conversion of native vegetation to agricultural land from deforestation for pasture expansion. We represent the annual deforestation rate as the percentage loss of previously forested area from the previous year, per municipality. MapBiomas land-cover data were also used to calculate the rate of expansion for rural areas (crops and pasture) and the percentage of municipal land covered by native vegetation in 2010. This information helped classify baseline land conditions for the period between 2006 and 2010.

Land occupation.

To distinguish between levels of legal land organization, we use data on formally registered farms in Brazil from the National System of Rural Register (Sistema Nacional de Cadastro Rural, SNCR) now incorporated into the new Land Management System (Sistema de Gestão Fundiária, SIGEF) (19–21). We gather the number of registered private lands per municipality. The SIGEF land register is an incomplete database of private lands, mainly due to high costs related to georeferencing and registering farm boundaries. These processes need to follow specific methodologies and be verified by technical officers from the National Institute of Colonization and Agrarian Reform (INCRA, in Portuguese Instituto Nacional de Colonização e Reforma Agrária). High costs related to such methodology and the obligation directed exclusively to large parcels hinder a more extensive registration in SIGEF. Thus, the georeferenced database encompasses part of the existing occupied land, mainly large farms.

Rural credit.

Rural credit data were obtained from the Municipal Bank Statistics database (Estatística Bancária Mensal por Município) available on a monthly basis on the Brazilian Central Bank’s website (49). The credit data, adjusted to US dollars,^* are detailed regarding targeted activity, i.e., agriculture or cattle raising. In addition, credit data from Municipal Bank Statistics allowed us to use credit type intended for implementation (costs $+$ investments), excluding credit for commercialization and industrialization; however, the data do not allow for further details, such as separating costs from investments or identifying specific credit lines.^† Data from the Central Bank were also used to compute credit distribution per financial institution, enabling the calculation of each institution’s market share in each municipality by year. Credit information is aggregated at the municipality level annually.

Other variables.

We also collect additional variables to control for relevant factors that affect deforestation. Population growth significantly affects deforestation rates (50). We obtain data on population by municipality from the Instituto Brasileiro de Geografia e Estatística (IBGE) (National Institute of Geography and Statistics) via its online database SIDRA (51) for the study period. We calculate average population density per municipality by dividing the annual population by the total municipal area. Squared population density was included as an independent variable to capture potential nonlinear effects on deforestation.

We also include annual average precipitation data per municipality from the CHIRPS, accessible through Columbia University’s climate library database (52), to control for potential residual variation in Landsat mapping between dry and rainy years. Human population and precipitation were included in all our models as control variables. Variables commonly linked to deforestation, such as agricultural production, were excluded as they would be bad controls, potentially creating endogeneity by influencing both credit loans and deforestation rates.

Empirical Approach.

First, we explain our typology of land contexts to distinguish samples used to estimate potentially heterogeneous credit impacts. This entails a detailed exploration of various factors, including the historical dynamics of land occupation, usage, and proxies of land organization such as municipal area and extension of land formally registered. Second, we describe our empirical strategy for estimation, including the specifics of our instrumental variable.

Definition of land conditions.

Building on the discussion above, we define five distinct land conditions based on a deforestation-relevant gradient that exists prior to the rural credit intervention. These conditions are characterized using three main dimensions: recent land use and its stability, municipal area (which summarizes significant land-use patterns and organization), and the extent of preexisting formal land registration under current land governance frameworks.

Land use & stability.

Land use, for our analysis, is defined by vegetation cover, which summarizes the historical implications of land use forest cover. Land-use stability is measured by the expansion rate of rural production activities, indicating trends before the rural credit intervention. Vegetation cover is defined as the percentage of municipal area covered with native vegetation in 2010 (Fig. 2, map A), the final year of our baseline period (2006–2010). Rural activity expansion is captured as the average rate of increase in agricultural and pasture areas combined during the baseline period (2006–2010) (Fig. 2, map B). Thus, the “Agropastoral & Stable” group then encompasses municipalities with less than 20% native vegetation cover and low or no expansion of rural areas. The “Native Stable” group consists of municipalities with more than 90% native vegetation cover, indicating minimal past production expansion. All other municipalities that do not fall into these two extremes are categorized as “Agropastoral & Dynamic.” The spatial distribution of these groups is presented in Fig. 2 (map C).

Municipal area.

A historical process of negotiations over land governance often led to an increase in the number of administrative units, thereby reducing their average size. Thus, we use municipal area as a proxy for social processes, including land governance, suggesting that land in smaller municipalities has been more intensively governed than in larger ones. Municipalities falling within the smallest 35% quantile (less than 175,000 ha) are classified as “Small” (Fig. 2, map D).

Formal land registration.

We assess the extent of formal land registration using data on legally registered private land up to 2010, i.e., prior to the rural credit provision, expressed as the proportion of municipal area with registered land. These data are obtained from the Land Management System register data (SIGEF) (with previous data from Sistema Nacional de Cadastro Rural, SNCR). These records reflect the degree of legal land organization. Due to legal requirements, only large private farms are required to register, and even then, only a fraction comply. Up to quantile 55% of municipalities have 0% occupation by large farms, 1% occupation represents around quantile 65% and 5% represents quantile 80%. Fig. 2 (map E) illustrates the spatial distribution of land organization. Given this distribution, we classify municipalities with at least 1% of their area registered as “More Formal,” indicating a higher presence of large farms. It is important to note that our SIGEF-based measure is not intended as a proxy for total land area—metrics available in the Agricultural Census serve that role—but rather as an indicator of formally certified land that can be mobilized in regulated markets, traded, or used as collateral.

Land conditions.

Finally, Fig. 2 also summarizes the criteria and classification scheme that define the five different land organization contexts considered in this analysis. The spatial distribution of municipalities according to observed factors and the identified land conditions across the Amazon Biome are shown in Fig. 2 (map F).

Identification Strategy.

We argue that the effect of credit on deforestation varies across different land contexts. To test it, we evaluate the impact of rural credit, specifically for agriculture and cattle raising, on deforestation by interacting credit variables with baseline land conditions. We address potential endogeneity in the relationship between credit and deforestation by employing an instrumental variable that can be interpreted as a credit supply shock.

Instrumental variable: Credit supply shifts.

Our main independent variables, rural credit for cattle and agriculture, might be endogenous due to two main concerns. First, there is potential reverse causality: Not only can credit affect deforestation, but higher or lower levels of deforestation could change the need for credit. Anecdotal evidence suggests that increased deforestation often leads to more cattle, which are commonly used as collateral for loans. This potential dynamic provides an explanation for why deforestation might increase access to credit, as banks tend to view cattle as desirable collateral. Consequently, deforestation itself may drive demand for agricultural and cattle credit, potentially leading to upward bias in Ordinary Least Squares (OLS) estimates. Second, endogeneity might arise from omitted variables correlated with both rural credit and deforestation, such as local economic growth and fluctuations in agricultural and cattle prices. These variables, associated with the demand for credit, could also impact deforestation. In the case of cattle prices, for example, not controlling for it might generate an upward bias for cattle-driven deforestation, as higher prices increase the pressure for deforestation, while boosting the demand for cattle credit. However, cattle prices might correlate negatively with the demand for agricultural products, also producing a downward bias.

To address these potential endogeneity concerns, we propose an instrumental variable that separates the demand component from the supply component. The demand component is specific to each municipality and might be correlated with local time-varying characteristics (omitted from the analysis). The supply component, on the other side, is more related to external sources, such as the banking sector’s overall supply of credit and macroeconomic conditions. Therefore, our instrument, inspired by Assunção et al. (41), is based on a shift-share approach, focusing on supply-side variations. In this sense, we can isolate an exogenous variation in the effect of credit on deforestation. We develop two separate instruments: one for agricultural credit and another for cattle credit.

To construct our instrument, we use credit lending data by financial institution ($f$), municipality ($m$) for the baseline period (2006–2010) to calculate the market share of each financial institution in each municipality, $s_{f,m}$. This is the “share” component of our instrument. We also calculate the market participation of each financial institution by agricultural and cattle credit ($j=agriculture,cattle$) within Amazon Biome municipalities ($s_{f,m}^j$), based on data from the baseline. We use initial-period bank market shares to construct the shift-share measure of credit supply, which deliberately fixes exposure to avoid endogeneity from banks’ strategic responses to deforestation (e.g., Banco do Brasil’s alignment with the Soy Moratorium or Priority List of municipalities). The instrument captures shocks under constant initial exposure rather than actual evolving credit flows. For the observed period from 2011 to 2017, we use total agricultural and cattle credit, $C_{f,m,t}^j$, to estimate the average supply of credit by financial institutions during the period (${\delta }_{f,t}^j$), while controlling for the municipality’s demand for credit per year (${\lambda }_{m,t}^j$). From Eq. 1, we get the estimated bank-year supply shock:

$$\begin{array}{c}\hfill \mathrm{\Delta }ln(C_{f,m,t}^j)={\lambda }_{m,t}^j+{\delta }_{f,t}^j+{\epsilon }_{f,m,t}^j\end{array}$$ [1]

Finally, we calculate our instruments, $Z^j=(Z^{\mathit{agric}},Z^{\mathit{cattle}}$), using the following equation:

$$\begin{array}{c}\hfill {\widehat{Z}}_{m,t}^j={\displaystyle \sum _f}({\widehat{\delta }}_{\mathit{ft}}^j{\widehat{s}}_m^j)\end{array}$$ [2]

Eq. 2 shows that the instrument represents the weighted average of the financial institution’s estimated credit supply in period $t$ (${\delta }_{f,t}^j$), where the weights are based on the institution’s market share in the municipality, as determined using data from a baseline period. We have estimated different instruments using data for cattle and agriculture.

To address concerns regarding the exogeneity of the shift-share instrument, we leverage the fact that Brazil’s banking system is highly oligopolistic, with a dominant public player, Banco do Brasil, responsible for most rural credit. This structure implies that deforestation patterns are unlikely to significantly influence bank profitability or expansion decisions in rural areas. Since rural credit is almost exclusively supplied by a public bank, this sector is less vulnerable to shifts in credit supply linked to local deforestation, supporting the exogeneity of our instrument. Furthermore, the possibility of endogenous bank penetration is limited due to the external factors influencing the presence of national and international banks, especially in areas like the Amazon. Bank penetration in these regions is often driven more by regulatory policies and broader economic priorities than by local deforestation patterns, further bolstering the validity of our IV approach.

Empirical model.

We have two endogenous credit variables, one for each sector ($j$): agriculture and cattle credit, as well as two exogenous variables capturing the supply shock for both credit types. As a result, our model is exactly identified. We employ a two-stage model estimated using a first-difference procedure. As our main independent variable is the interaction between credit and baseline land conditions, we have a separate first-stage equation for each land condition and each type of credit—cattle and agricultural. The basic first stage equation is defined as

$$\begin{array}{cc}\hfill ln\left(\frac{C_{m,t}^j}{C_{m,t-1}^j}\right)=& {\lambda }_m+{\theta }_{s}t+{\alpha }_j{\widehat{Z}}_{m,t}^j+\eta X_{m,t}+{\epsilon }_{m,t}\hfill \end{array}$$ [3]

The second stage equation is as follows:

$$\begin{array}{c}\hfill ln\left(\frac{y_{m,t}}{y_{m,t-1}}\right)={\alpha }_m+{\delta }_{s}t+{\beta }_{1j}ln\left(\frac{C_{m,t}^j}{C_{m,t-1}^j}\right)+\\ \hfill {\beta }_{2j}ln\left(\frac{C_{m,t}^j}{C_{m,t-1}^j}\right)L_m+{\beta }_3{X}_{m,t}+{ϵ}_{m,t}\end{array}$$ [4]

Here, $y$ represents the deforested area (expressed as the logarithm difference between two consecutive years) in each municipality $m$ for year $t$. $C$ denotes the credit utilized (also expressed as the log difference between two consecutive years) for $j$ activity in municipality $m$ for year $t$. $L$ refers to the land condition in municipality $m$. $X$ is the covariate matrix, including precipitation and population density, and $ϵ$ is the error term. We control for municipality fixed effects ${\alpha }_m$ to account for time-invariant differences among municipalities (e.g. variables related to land profitability, such as distance to roads, rivers, and slope and quality of the soil) and state-time fixed effects ${\delta }_{\mathit{st}}$, allowing each state to have unique temporal patterns (to control for the potential confounding influences of state-level policies implemented over the analyzed period). We run separate first-stage regressions for each credit activity (agriculture and cattle). All procedures were conducted using R (Version 3.6.3.) (53), with the Fixest package (Version 0.10.4) used for mean conditional panel regressions (54).

Supplementary Material

Appendix 01 (PDF)

Data, Materials, and Software Availability

The data that support the findings of this study are available in public platforms and websites: brasil.mapbiomas.org (48); atlasagropecuario.imaflora.org (20); www.bcb.gov.br/estatisticas/estatisticabancariamunicipios (49); sidra.ibge.gov.br (51); and iridl.ldeo.columbia.edu/ (52). The compiled dataset and R-code used for analyses are available at: github.com/pruggiero2/Credit_Deforestation_Research_data_code (55).