Every harvest, hundreds of billions of reais pour through Brazil’s rural credit system, financing everything from soybean seeds to combine harvesters across one of the world’s agricultural superpowers. Yet a sweeping new analysis covering all 5,570 Brazilian municipalities suggests that this river of money behaves less like a rising tide and more like a network of irrigation canals — watering some regions generously while bypassing others almost entirely. The study, published in the open-access journal Discover Agriculture by economists Luis Abel da Silva Filho of the Regional University of Cariri and Pedro Vasconcelos Maia do Amaral of the Federal University of Minas Gerais, applies an advanced spatial econometric toolkit to credit and production data from 2019 to 2023 and reaches a conclusion with profound implications for agricultural policy: the power of a loan to lift farm output depends not only on its size and purpose, but on what is happening just across the municipal border.

Agriculture occupies a strategic position in the Brazilian economy, anchoring the trade surplus, food supply and rural employment, and the authors argue that credit is the hidden infrastructure behind that performance. Access to financing allows producers to acquire the technologies, inputs, machinery and equipment that raise productivity and competitiveness, making credit not merely a financial instrument but a structuring mechanism for modernization across the countryside. Brazil’s National Rural Credit System, created by Law No. 4,829 of 1965 and operationalized through the Central Bank’s Rural Credit Manual, organizes lending into three main categories: operating-cost lines that fund a single production cycle, investment lines that build fixed capital and drive technological upgrading, and marketing lines that support sales and storage. Targeted programs such as Pronaf for family farms and Pronamp for medium-sized producers segment the beneficiary public, while federal public banks, private banks, the national development bank BNDES and the Constitutional Financing Funds for the North, Northeast and Central-West — created by the 1988 Constitution expressly to reduce regional inequality — supply the resources.

The study’s methodological core is the Spatial Durbin Model, an econometric specification built for data in which geography matters. The authors start from an agricultural production function in which output depends on capital — including credit — labor, land and climate, treating rainfall as a quasi-fixed input that behaves like an additional factor of production in a country where most farming is rainfed. Because credit is disbursed through bank branches and spreads across contiguous territorial networks of information, technology and labor, the researchers linked every municipality to all neighbors sharing a border or even a vertex, using a Queen contiguity matrix. Before adopting a preferred specification, they staged a systematic model shootout: ordinary least squares, the spatial lag model, the spatial error model, the spatial Durbin model and the SDEM and SARAR variants were compared using Lagrange multiplier and robust Lagrange multiplier tests, Moran’s I randomization tests and Monte Carlo simulations on the residuals. Ordinary least squares failed the diagnostic — its residuals carried a Moran’s I statistic of 14.005, with the SLX and SAR models close behind at 13.248 and 11.193 — while the Durbin specification drove residual autocorrelation to statistical zero, minimized the Akaike Information Criterion and maximized the log-likelihood.

Formally, the model writes municipal production as y = ρWy + Xβ + WXθ + ε, a compact equation with far-reaching implications. The term Wy means output in one municipality depends directly on output in its neighbors, with ρ measuring the strength of that feedback, while WXθ captures how credit and other drivers in neighboring municipalities spill over administrative boundaries. The authors also estimated a quantile version, Qτ(y|X) = ρτWy + Xβτ + WXθτ + ετ, which allows the effect of credit to differ at the bottom, middle and top of the production distribution — a crucial innovation, because a single average effect can conceal opposite impacts at different points of the spectrum. Following the decomposition framework of LeSage and Pace, every estimated coefficient can be split through the inverse matrix (I − ρW)⁻¹ into a direct effect on the municipality itself, an indirect spillover onto neighbors, and a total effect. To handle the endogeneity inherent in a spatially lagged dependent variable, the study applied the standard instrumental-variable procedure of Kelejian and Prucha, and a test of the common factor hypothesis — which would collapse the Durbin model into a simpler error specification — proved insignificant, leaving the full structure intact.

The exploratory mapping alone tells a striking story. Using Moran’s I and its local variant LISA, the authors show that high-production clusters — neighboring municipalities that all post above-average gross value of agricultural production per capita — are intensely concentrated in the South, Southeast and Midwest. In 2019 these high-high clusters contained 776 significant municipalities, 13.93 percent of the national total; by 2023 they had shrunk to 736, or 13.21 percent. Low-low clusters, concentrated in the North and Northeast, moved in the opposite direction, expanding from 1,356 municipalities to 1,392, or from 24.34 to 24.98 percent of the map. Credit tells a parallel story: clusters of high per capita lending fell from 1,394 municipalities to 1,148 between 2019 and 2023, while low-credit clusters shrank from 1,000 to 703, a measurable redistribution that nonetheless left historically disadvantaged regions confronting the same barriers of collateral, technical assistance and infrastructure. The underlying statistics reveal how skewed the distribution remains. Per capita investment credit in 2019 averaged R$318.51 but carried a median of just R$39.17; operating credit averaged R$847.43 against a median of R$44.13; and medians for BNDES and Constitutional Fund disbursements were often zero, meaning many municipalities received nothing at all in a given year. Extreme outliers — a maximum of R$96,479.92 per capita from BNDES in 2019 — underscore the institutional concentration of resources, as does formal farm employment, which averaged 117 to 122 workers per municipality against a median of only 14.

The headline econometric result is that space amplifies everything. The spatial autocorrelation parameter ρ came in at 0.2570, which implies a spatial multiplier of 1.3459: once feedback through neighboring municipalities is taken into account, the total effect of any credit coefficient is roughly 35 percent larger than its direct effect alone. Agricultural production also proved deeply persistent — the 2019 per capita production value carried a coefficient of 0.872 into the 2023 equation, capturing productive inertia, installed capital and technological learning — while its spatial lag was negative at −0.204, suggesting that booms next door can siphon away inputs, credit and labor in a quiet pattern of intermunicipal competition. Short-term financing credit behaved exactly as theory predicts: disbursements made in 2023 raised output in the same year with a positive, highly significant direct effect of 0.007, reinforcing the literature that identifies cost financing as the workhorse of current production. The same line granted in 2021, however, turned significantly negative, plausibly reflecting the climate shocks and operational distress of that cycle.

Investment credit told a more subtle story. Capital disbursed in 2020 showed a slightly negative direct effect by 2023, consistent with the maturation lag of machinery, irrigation and infrastructure, yet its spatial lag was positive and significant, meaning investment in one municipality generated measurable productive externalities next door. Loans from BNDES were largely statistically insignificant, echoing earlier findings that the development bank’s impacts are localized and uneven. The Constitutional Financing Funds produced the study’s most provocative asymmetry: 2023 financing displayed a negative direct effect of −0.005 but a positive, highly significant indirect effect of 0.020 on neighboring municipalities, while 2020 investment credit combined a strongly negative direct effect of −0.016 with a positive neighborhood effect of 0.020. Recipient municipalities, the authors suggest, may lack the institutional capacity to absorb funds quickly, while their neighbors capture the gains through production linkages, processing and marketing networks — a dynamic that turns regional development funds into engines whose benefits leak across borders.

The quantile estimates sharpen the picture considerably. In the lowest quartile of the production distribution, investment credit granted to neighboring municipalities in 2020 carried a positive elasticity of 0.008, easing to 0.006 at the median — evidence that less productive municipalities benefit disproportionately from the investments made around them, most likely through shared infrastructure and supply chains. By contrast, Constitutional Fund investment credit from 2022 showed negative spillovers of −0.007 at the first quartile and −0.006 at the median, pointing to execution and absorption failures in regions with weak institutional capacity. Geography itself mattered: the spatial lag of the North region dummy reached 0.407 in the first decile and the Northeast’s 0.115 in the first quartile, both statistically significant. At the top of the distribution, most credit variables lost significance altogether — high-productivity municipalities tap private banks and cooperatives that dilute the imprint of public funds — while bottom-tier producers, constrained by liquidity, respond most strongly when credit finally arrives.

The authors frame their findings against the turbulence of the 2019–2023 window, which spans the COVID-19 pandemic, disrupted supply chains and an expanded Plano Safra, and they are candid about the limits of the design. Credit allocation may still respond to municipalities’ past performance, a simultaneity the lagged output term, structural controls and spatial specification mitigate but do not eliminate, so the results should be read as robust, spatially structured association rather than strict causal identification. Municipal-level aggregation, and an agricultural census that has not been updated since 2017, further obscure the microeconomic mechanisms — mechanization, technology diffusion, rural extension — through which credit does its work. Even so, the policy message is unambiguous. Because the effects of credit vary by modality, source, geography and position in the productivity distribution, uniform national lending rules will keep misallocating resources: short-term financing needs speed, investment credit needs complementary infrastructure, technical assistance and patience, and regions with low absorption capacity need institutional support before the money arrives.

What makes the study a genuine first, the authors argue, is the unprecedented application of a quantile spatial Durbin model to rural credit across an entire nation’s municipal grid, with every estimated effect decomposed into direct, indirect and total impacts. The results confirm that Brazil’s credit-and-production landscape is carved into self-reinforcing clusters, that spillovers are real and quantifiable, and that the municipalities that need credit most — those stranded in the low-low pockets of the North and Northeast — are precisely those where public funds currently buy the least on their own. Redirecting even part of the system’s enormous flow toward territory-sensitive design, the study concludes, could transform rural credit from a mechanism that quietly reproduces regional inequality into an instrument that finally begins to dismantle it.