Internet Use and Technical Efficiency of Grain Production in China: A Bias-Corrected Stochastic Frontier Model

The study addresses whether and how Internet use affects the technical efficiency of smallholders engaged in grain production in China. Agricultural production in many developing countries, including China, often follows an extensive, high-input model with low technical efficiency, which threatens both economic competitiveness and environmental sustainability. Prior improvements have focused on institutional reforms and promoting new technologies, but farmers’ adoption decisions are constrained by information access, capital, and skills, particularly among smallholders. As a general-purpose technology, the Internet can reduce information costs, speed up technology diffusion, and improve coordination, potentially increasing production efficiency. However, evidence for staple grains is limited and may differ from cash crops due to lower market incentives and different management requirements. This paper investigates the impact of Internet use on grain producers’ technical efficiency and explores mechanisms through which the Internet may operate, providing policy-relevant insights for enhancing efficient and sustainable agricultural production.

Existing work emphasizes the Internet’s role in improving market efficiency (e.g., electronic markets, information in supply chains) and in production through reduced information frictions and faster technology diffusion. Studies have linked ICT and Internet adoption to improved output and technology adoption in various contexts (Aker et al., 2016; Jensen, 2007; Khan et al., 2022; Zheng et al., 2022). Evidence on technical efficiency is emerging: Mwalupaso et al. (2019) examined mobile communication and maize efficiency without accounting for self-selection; Zhu et al. (2021) and Zheng et al. (2021) found positive effects for apples and bananas in China, respectively. Differences between cash crops and grains (technical requirements, management needs, and lower price incentives) raise questions about external validity to grain production. The study contributes by explicitly incorporating Internet use into the grain production efficiency framework, correcting for selection biases, and testing mechanisms via financial constraints, social capital, and mechanization.

Data: 2018 China Labor Dynamics Survey (CLDS) by Sun Yat-sen University, employing multi-stage, multi-level probability sampling. The sample includes 1699 rural households engaged in grain production across 25 provinces. Datasets from family, household head, and village questionnaires were merged; observations with zero cultivated area or missing key variables were excluded. Internet use is a binary indicator equal to 1 if the household accesses the Internet via mobile phone, tablet, or computer (59.4% users). Control variables span individual (head’s age, gender, education, ethnicity, health), household (size, log total income), and village (topography—plain vs. other, village size, distance to county) characteristics. Models: (1) Selection-corrected stochastic production frontier (SPF) following Greene (2010), with a sample selection (Internet use) equation and a translog production frontier for grain output using inputs: labor (number of agricultural laborers >3 months/year), capital (expenditures on cultivation, fertilizer, pesticide, machinery, etc.), and land (grain planting area). Technical efficiency (TE) is derived as the ratio of expected output to the frontier in the absence of inefficiency. A likelihood ratio test favored the translog over C-D (LR=59.22). The selection term (ρ) tests for selectivity due to unobservables. Conventional SPF and bias-corrected SPF were estimated for pooled, users, and non-users subsamples. (2) Propensity score matching (PSM) to estimate the average treatment effect on the treated (ATT) of Internet use on TE using logit-based propensity scores and nearest-neighbor, radius, and kernel matching. Common support, balance tests (Pseudo-R², LR Chi², mean/median bias, Rubin’s B) were conducted. Sensitivity to hidden bias was assessed via Rosenbaum bounds (Gamma). (3) Endogenous switching regression (ESR) with a first-stage Internet use decision and separate outcome equations for TE among users and non-users. The proportion of Internet users in the same village served as an instrumental variable; strength confirmed by Cragg-Donald Wald F=129.670. Mechanism tests: Using PSM, the study assessed three channels: financial constraints (whether the household borrowed), social capital (gift expenditures), and mechanized farming (use of machinery in grain production).

- Determinants of Internet use (logit marginal effects): - Head’s age: −1.1 percentage points per additional year (p<0.001). - Head’s education: +4.2 pp per education level (p<0.001). - Head’s health: +2.6 pp per health level (p<0.01). - Household size: +1.7 pp per person (p<0.01). - Log household income: +9.0 pp per unit (p<0.001). - Village topography (plain): +6.0 pp (p<0.05). Gender, ethnicity, village size, and distance to county not significant. - Production frontier and selection: Translog preferred to Cobb-Douglas (LR=59.22). In bias-corrected SPF, the selection term ρ is significant (p<0.05), indicating selectivity bias; labor and land inputs significantly contribute to output, with high labor elasticity. - Internet use and technical efficiency (TE): - PSM ATT on TE is positive and statistically significant across matching methods; using bias-corrected SPF TE measures yields larger effects than conventional SPF, indicating conventional SPF underestimates the impact. - Average ATT (bias-corrected SPF): approximately +0.011 in TE for users versus non-users; radius matching ATT=0.012 (p<0.10), nearest neighbor ATT=0.010 (p<0.10), kernel ATT=0.010 (p<0.10). - ESR confirms positive treatment effects: ATT for users ≈ +0.004 (p<0.01); ATU for non-users if they adopted ≈ +0.003 (p<0.01). TE levels around 0.777 (users) and 0.771 (non-users) in observed states. - Robustness: PSM balance achieved (post-match Pseudo-R² ≤ 0.006; LR Chi² < 15; mean/median bias ≈ 4%/3–4%; Rubin’s B < 25%). Rosenbaum bounds indicate conclusions remain robust to hidden bias up to Gamma ≈ 1.5–1.8 depending on matcher. - Mechanisms (PSM ATT, means across matchers): - Financial constraints: probability of borrowing +0.084 (significant). - Social capital: gift expenditures +1.688 units (significant at 1%). - Mechanized farming: probability of using machinery +0.105 (significant).

The findings directly address the research question by showing that Internet use significantly improves technical efficiency in grain production among Chinese smallholders, even after correcting for both observable and unobservable selection biases. The significance of the selection term and the larger ATT under bias-corrected SPF indicate that studies ignoring selection may underestimate ICT’s productivity effects. The determinants analysis highlights which populations are more or less likely to adopt, informing targeted interventions. Mechanism tests substantiate plausible pathways: improved access to finance, enhanced social networks that facilitate information exchange and technology diffusion, and greater mechanization through lower search and transaction costs for service outsourcing. Together, these results suggest that Internet use alleviates key constraints (information, capital, labor) that traditionally impede smallholders’ movement from extensive to efficient production, reinforcing the Internet’s role as enabling infrastructure for agricultural modernization.

The study integrates Internet use into the technical efficiency framework for grain production and, using bias-corrected SPF, PSM, and ESR on nationally representative household data, shows a robust positive effect of Internet adoption on smallholders’ technical efficiency. It identifies who is more likely to adopt (younger, healthier, better educated heads; larger and higher-income households; villages with plain terrain) and validates three channels—alleviated financial constraints, expanded social capital, and increased mechanization—through which the Internet boosts efficiency. Policy implications include: investing in rural Internet infrastructure and lowering access costs; tailoring outreach and training to heterogeneous smallholder groups; establishing technical training centers and leveraging diverse information channels (including social media/apps); and coordinating with financial institutions, cooperatives, research bodies, and agribusinesses to provide integrated financial, technological, and service support. These measures can help transition grain production toward higher efficiency and sustainability.

- Measurement of Internet use is limited to basic access (mobile/PC/tablet) and does not capture intensity, purpose, or quality of use; hence, the study cannot precisely measure how Internet-derived information supports production decisions. - The analysis uses cross-sectional data (2018 CLDS), limiting the ability to assess dynamic effects or causal trajectories over time; future work with panel data is needed. - As with any PSM/ESR application, residual unobserved heterogeneity may remain despite robustness checks, and external validity beyond the sampled regions and period should be considered.