Impact of small farmers' access to improved seeds and deforestation in DR Congo

The study investigates whether promoting modern, higher-yielding seed varieties to smallholders affects deforestation in the Congo Basin, specifically distinguishing impacts on primary versus secondary forests. The context is Sub-Saharan Africa, where agriculture supports about 75% of livelihoods but yields lag due to limited diffusion of improved varieties and input constraints. The Borlaug hypothesis suggests technological intensification can spare land, yet local market imperfections, elastic demand, and labor supply responses can lead to the opposite (Jevons paradox). In the Democratic Republic of Congo’s Equateur region, input and output markets are weak, fertilizers are scarce, and labor for land clearing is binding during peak seasons. The paper aims to provide causal evidence on how access to modern seed varieties affects land conversion decisions and deforestation types, with implications for biodiversity and climate.

Prior work links modern seed diffusion to substantial gains in yields, GDP per capita, and health outcomes across low- and middle-income countries. The land-sparing Borlaug hypothesis posits that yield growth reduces expansion into forests, while alternative theories highlight conditions where intensification raises profitability and can expand cultivated area, especially with elastic demand and labor (Jevons paradox). Empirical evidence at global and national scales is mixed: some studies credit intensification with land sparing, others find ambiguous or context-dependent relationships. Reviews note a paucity of causal estimates and limited evidence for a strong positive link between technological progress and deforestation, especially in the Congo Basin, where credible causal studies are lacking. Differentiating between primary and secondary forest is emphasized as essential due to their distinct biodiversity and carbon implications.

The study leverages a randomized controlled trial nested in the DRC Ministry of Agriculture’s PARRSA program in the former Equateur province (Nord Ubangi, Sud-Ubangi, and Mongala). Ninety-two villages were stratified by population, remoteness, and extension exposure; 60 were randomly assigned to receive a seed subsidy intervention and 32 served as controls. Within treatment villages, households were randomly offered vouchers redeemable for up to 10 kg of seeds at one of four subsidy levels (30%, 60%, 90%, or 100%) for rice (short cycle), maize (high-yielding), groundnuts (high-yielding), or cassava (disease tolerant). To study spillovers, the share of households receiving vouchers varied randomly by village (approximately 20%, 45%, or 70%). To reduce transport frictions, 35 villages were randomly assigned to on-site truck delivery of seeds shortly after voucher distribution; other treatment villages relied on distant seed multipliers’ offices. Logistical circumstances led trucks to supply legumes more reliably than stores, which stocked relatively more cereals. Timing: Vouchers were distributed in February 2013, too late for primary forest clearing that year. Seeds harvested in 2013 could be recycled, making 2014 the main season to observe land conversion responses. The analysis therefore focuses on 2014, with some satellite-based outcomes aggregated 2013–2016. Data: (1) Remote sensing: Village-level deforestation is computed from Hansen et al. (30 m) as the sum of grid cells with >90% canopy in 2000 that declined to <25% tree cover, aggregated using two approaches to approximate village extent (survey-based polygons and 5 km buffers) and averaged, then normalized per household. FACET data (60 m) distinguishing primary versus secondary forest were used in a robustness exercise by combining with Hansen, acknowledging added noise and coarser temporal resolution. (2) Household survey: In mid and late 2014, detailed plot-level data were collected for 904 households on all plots cultivated in the main season, including prior land cover classification by farmers (primary vs secondary forest, fallow, previously cultivated, savanna/other), land areas, crops planted, and labor allocations (household, hired, and labor-sharing for land preparation). Labor income from agricultural wage work was also collected. Outcomes and estimation: Primary outcomes are deforested area (ha) by forest type at household level (2014) and village-level deforestation per household (2013–2016 and 2014). The inverse hyperbolic sine transformation of areas is used in regressions, though at small values it approximates levels. Ordinary least squares regressions compare treatment to control, distinguishing truck vs no-truck villages, with standard errors clustered at village level and strata fixed effects. Additional analyses compare high vs low subsidy recipients and examine variation by treatment density. Labor allocation outcomes (members working on farm, person-days of land preparation via household labor and labor sharing, and agricultural wage income) test mechanisms. Randomization inference and robustness checks (binary extensive-margin outcomes; non-transformed areas; alternative village boundary definitions; combined Hansen–FACET) support the findings.

• Village-level satellite outcomes: No evidence that seed promotion increased total deforestation. For 2013–2016, joint F-test P=0.60; lottery without truck coefficient −0.006 (SE 0.182, P=0.972); lottery with truck −0.150 (SE 0.166, P=0.368). For 2014, joint F-test P=0.13; lottery without truck −0.092 (SE 0.094, P=0.337); lottery with truck −0.144 (SE 0.086, P=0.099). Point estimates are negative.
• Household-level land conversion (2014): No statistically significant change in total forest cleared (joint F-test P=0.54). However, significant increases in primary forest clearing with treatment (joint F-test P=0.00): • Villages without truck: +0.108 to +0.11 ha per household vs control (coefficient 0.108, SE 0.038, P=0.006; text cites 0.11 ha). • Villages with truck: +0.06 to +0.071 ha (SE 0.031–0.039; P≈0.052–0.074). • Given legumes were primarily delivered by truck, this pattern suggests greater access to improved cereals (no-truck) increased primary forest conversion more strongly.
• Offsetting reductions in secondary forest conversion: Point estimates show decreased reliance on secondary forest, especially among high-subsidy households without truck (−0.110 ha; SE 0.071; P=0.124). Overall totals did not increase because increased primary clearing was partly offset by reduced secondary clearing and shifts away from previously cultivated/fallow land.
• Subsidy intensity: High-subsidy, no-truck households had the largest primary forest increment (+0.126 ha; SE 0.049; P=0.012). Low-subsidy, no-truck households: +0.086 ha (SE 0.039; P=0.028). With truck, high-subsidy: +0.071 ha (SE 0.039; P=0.074); low-subsidy: +0.053 ha (SE 0.031; P=0.090). Total cultivated area did not increase significantly.
• Robustness: Binary indicators of using any primary forest land and regressions with raw area units confirm direction and significance patterns. Combined Hansen–FACET village-level analysis (2013–2016) shows consistent directions (more primary, less secondary deforestation), though not statistically significant.
• Labor allocation: Treatment increased the number of household members working on the farm (joint significance P=0.00) and labor-sharing person-days for land preparation (P=0.06), with a (non-significant) decrease in members’ agricultural wage income (P=0.25). These patterns indicate reallocation toward labor-intensive clearing, consistent with increased primary forest conversion, and help explain why village-level totals did not rise markedly due to labor constraints.
• Crop type: Where access favored legumes (truck villages), the increase in primary forest clearing was smaller, suggesting cereal-focused improved varieties, with higher nitrogen demand, drive stronger shifts toward primary forest.

The RCT provides causal evidence that promoting modern seed varieties in a remote Congo Basin context did not increase overall deforestation but altered its composition: more primary and less secondary forest were cleared. This pattern aligns with a mechanism where improved cereal varieties raise nitrogen demand, and with limited fertilizer availability and weak knowledge or adoption of alternative soil fertility practices, farmers source nutrients by shifting cultivation to nitrogen-richer primary forest. Limited labor market depth and the labor intensity of clearing likely constrained total expansion, resulting in minimal net changes in aggregate deforestation. These findings nuance the Borlaug hypothesis in this setting: intensification via seeds alone did not spare land locally, nor did it trigger broad expansion (Jevons paradox), but it did shift pressure toward ecologically critical primary forests, with severe biodiversity implications. The smaller primary forest impact where legumes were more available supports the mechanism but stems from unplanned variation in crop supply; thus, causal attribution by crop type is suggestive rather than definitive. Policy implications are clear: seed promotion programs should be coupled with soil fertility management (mineral or organic inputs, conservation practices) or emphasize crops less demanding of nitrogen to avoid biodiversity losses.

The study contributes two key advances: (1) rigorous, causal evidence from a randomized trial on how modern seed promotion affects deforestation in the Congo Basin; and (2) demonstration that distinguishing primary from secondary forest is essential, as improved seed access increased primary forest clearing while reducing secondary clearing, leaving total deforestation largely unchanged. The mechanism is consistent with nitrogen demand of improved cereals combined with missing input markets and labor constraints. For policy, promoting improved varieties without simultaneous soil fertility strategies risks biodiversity costs through primary forest loss. Programs should bundle seed access with context-appropriate soil fertility management and consider increasing access to legumes. Future research should experimentally vary access by crop type, rigorously test integrated soil fertility interventions, and better quantify carbon and biodiversity trade-offs across forest types under different intensification pathways.

• Remote sensing limits: Hansen data do not distinguish forest types, requiring household reports for primary vs secondary classification and a combined Hansen–FACET robustness exercise with coarser temporal resolution and added noise.
• Measurement error: Farmer-reported plot areas and land cover categories may be imprecise; however, extensive-margin and labor-based results support main findings. Misclassification would bias results only if correlated with randomized treatment, which is unlikely.
• Labor market scope: Data exclude large commercial farms/plantations; broader market-level effects cannot be assessed.
• Boundary approximation: Village-level deforestation relies on approximate village polygons and buffer methods; results are robust across definitions but measurement error remains.
• External validity: Findings reflect a specific, remote context with missing input/output markets and may not generalize to settings with better market access. Timing and logistics also led to unplanned variation in crop availability (legumes via trucks), complicating crop-specific causal attribution.