China’s digital village policy boosts agricultural entrepreneurship, with regional trade-offs

China’s national Digital Village Pilot Policy increased the rate at which counties launched digitally oriented agricultural businesses after the program began in 2020, according to a county-level quasi-experimental study of 1,743 jurisdictions from 2014 through 2023. The researchers estimated that the program produced an additional 0.0875 to 0.0989 newly registered agricultural ventures per 10,000 residents, depending on the statistical specification, but also found that entrepreneurial activity declined in neighboring counties relative to the modeled counterfactual. The analysis suggests that institutional digitalization can create new opportunities for smart agriculture, precision farming, agricultural e-commerce, drones, big data and Internet of Things services, while also redistributing talent, capital and markets toward places selected for early policy support.

The study, published in Frontiers in Sustainable Food Systems on October 21, 2026, treats the first batch of national digital village pilot areas as a quasi-natural experiment. China initially designated 117 counties, cities and districts in 2020, but the researchers excluded 42 pilot jurisdictions because they were urban districts, administrative units with incompatible fiscal or land systems, areas affected by boundary changes or places with severe data gaps. That left 75 treated counties and 1,668 comparison counties, producing a panel with 17,430 county-year observations. The outcome was not the digital upgrading of established farms and cooperatives. It was the entry of new enterprises whose registration information and business scopes indicated a principal focus on digitally driven agricultural production, operations or services.

Enterprise data were assembled with a Python program that searched Tianyancha records for seven categories: smart agriculture, precision agriculture, smart agricultural management systems, agricultural e-commerce, agricultural plant-protection drones, agricultural big data and agricultural Internet of Things. Researchers first screened firms using China’s national industry classifications, including agriculture, forestry, animal husbandry and fishery, agricultural and sideline food processing, and internet-related services. They then examined the semantic content of company business descriptions, excluding firms whose principal activity was general information technology, software development, consulting or e-commerce unrelated to agriculture. New firms were normalized by county population to measure entries per 10,000 residents. A robustness analysis also counted enterprises that remained active after two years, providing a partial check on whether the result simply reflected short-lived registrations rather than more persistent venture creation.

The study’s main model was an interactive double machine-learning model. Unlike a conventional linear difference-in-differences regression, it allows the policy effect to vary nonlinearly with a large set of county characteristics while using cross-fitting and county-clustered standard errors. The model controlled for economic and agricultural development, industrial structure, credit and fiscal conditions, cultivated land, population density, education, roads, existing informatization, and county and year fixed effects. A partially linear double machine-learning model, conventional difference-in-differences, propensity-score matching combined with difference-in-differences, and a synthetic-control version of the method all yielded positive policy estimates. Event-study results found no significant differences between pilot and comparison counties before implementation, while the gap widened after 2020.

The authors also tested whether the result could be explained by pilot selection, omitted variables or reverse causality. An instrumental variable combined the post-2020 policy period with each county’s distance from the node cities of China’s optical-fiber backbone network, a communications geography established before the sample period. Its first stage was strongly significant, and the estimated policy effect remained positive at 0.0655 in the second stage, with county-clustered standard error 0.0140. A one-period lag of the policy indicator produced a similar result. These checks strengthen the causal interpretation, but they do not make the pilot assignment random: the authors describe the design as a quasi-experiment, and the findings apply most directly to institutional settings comparable to the first Chinese pilot program.

Three channels appeared to transmit the policy effect. The program increased agricultural patent applications, an indicator of technological innovation; the coefficient on innovation was positive, and including it reduced the estimated direct policy effect from 0.0878 to 0.0848, while innovation itself was associated with more new agricultural businesses. Digital infrastructure also carried a mediation role. Adding the infrastructure measure lowered the policy coefficient from 0.0878 to 0.0621, and the infrastructure coefficient was strongly positive. A third pathway involved digital finance: its inclusion reduced the policy coefficient to 0.0436 and was positively associated with entrepreneurial entry. These results identify statistical transmission channels rather than directly observed changes in an individual entrepreneur’s behavior. Because digital infrastructure was measured mainly through the frequency of infrastructure-related terms in county government work reports, it primarily captures policy attention rather than the physical density of broadband, base stations or connected devices. A proxy based on household broadband access produced broadly consistent results.

The benefits depended heavily on local absorptive capacity. The estimated effect was 0.1117 in counties with above-median pre-policy rural internet access, compared with 0.0653 in counties with lower access. It was 0.1487 where transport infrastructure was stronger, versus 0.0776 where it was weaker, and 0.1517 in counties with higher digital-finance development, versus 0.0942 in lower-digital-finance counties. All three regional estimates were positive, but central China showed the largest coefficient at 0.1487, followed by western China at 0.1023 and eastern China at 0.0800. Counties designated as advantageous areas for characteristic agricultural products also had a larger effect, 0.1387, than non-advantageous counties, 0.0958. The authors interpret this pattern as evidence that digital opportunities become ventures most readily where connectivity, logistics, finance and established agricultural industries are complementary.

Spatial econometric analysis revealed an important trade-off. Entrepreneurial activity was positively spatially autocorrelated across adjacent counties, and pre-policy event-study coefficients for neighboring exposure were statistically indistinguishable from zero. In a spatial Durbin difference-in-differences model, the direct effect of digital village status was 0.0610, the estimated indirect effect on neighbors was negative at −0.0206, and the combined total effect remained positive at 0.0404. The negative neighboring effect was stronger in counties experiencing greater agricultural labor outflow and stronger market competition. The authors identify factor siphoning and product-market substitution as plausible explanations: pilot counties may attract labor, capital and technology while gaining first-mover advantages in digital agricultural markets. The data cannot cleanly separate those mechanisms, however, and the study does not establish that any individual entrepreneur transferred from a neighboring county to a pilot area.

Several limitations constrain the conclusions. The panel ends in 2023, only about three years after the first pilots began, so persistence, strengthening or attenuation over a longer horizon remains unknown. Patent counts do not measure innovation quality or practical adoption, the infrastructure proxy is imperfect, and county-level data cannot fully distinguish labor mobility, market crowding out and differences in digital capabilities. The entrepreneurial measure chiefly captures entry into new enterprises, not revenue, growth, innovation performance or the digital transformation of existing family farms and cooperatives. It is therefore a measure of new-business formation in selected agricultural domains, not a complete index of rural digital entrepreneurship. Taken together, the study indicates that digital village programs can accelerate agricultural venture creation where local systems are ready to absorb new technology, but their wider effects depend on whether digital resources expand regional opportunity or merely redistribute it toward designated pilot counties.

Subject of Research: Effects of institutional digitalization on new agricultural business formation in Chinese counties.

Article Title: How institutional digitalization reshapes entrepreneurial activity in new agricultural business forms: quasi-natural experimental evidence from China's Digital Village Pilot Policy

Article References: Li, S., & Yang, C. (2026). How institutional digitalization reshapes entrepreneurial activity in new agricultural business forms: quasi-natural experimental evidence from China's Digital Village Pilot Policy. Frontiers in Sustainable Food Systems, 10, Article 1946061. https://doi.org/10.3389/fsufs.2026.1946061

Image Credits: AI Generated

DOI: 10.3389/fsufs.2026.1946061

Keywords: digital agriculture, rural entrepreneurship, Digital Village Pilot Policy, quasi-experiment, double machine learning, China, spatial spillovers, digital finance, agricultural innovation, county-level analysis, institutional, digitalization