Digital Technology Adoption and the Optimal Allocation of Rural Household Labor, Land, and Capital: Evidence from the Yellow River Basin

Abstract

Digital technologies have rapidly penetrated rural China, transforming household economic structures and reshaping agricultural factor markets. Drawing on 3930 household-level observations from six provinces in the Yellow River Basin (2020–2023), this research aims to examine how digital technology adoption influences the optimal allocation of household labor, land, and capital. To address self-selection and endogeneity, we employed an Endogenous Switching Probit (ESP) model and conducted counterfactual analysis, supplemented by propensity score matching (PSM), instrumental variable probit (IV-Probit), replacement of the core explanatory variable, and exclusion of special samples as four robustness checks. The Average Treatment Effects on the Treated show that digital technology adopters would have reduced the probabilities of non-farm employment, farmland transfer-out, and productive loan access by 24.5, 19.3, and 16.7 percentage points, if they had not adopted digital technology. Similarly, digital technology non-adopters would have improved 27.4, 22.2, and 18.9 percentage points if they adopted digital tools. These impacts are stronger over time, in the upper reaches of the Yellow River Basin, among households with larger landholdings, and among younger farmers. Mechanism analysis further indicates that digital technologies expand information access, strengthen social networks, and ease credit constraints, thereby jointly promoting more efficient labor, land, and capital allocation. The policy implications of these findings are as follows: the importance of improving rural digital infrastructure, tailoring regional policies, and enhancing farmers’ digital skills to narrow the digital divide and support inclusive rural revitalization.

1. Introduction

Achieving common prosperity in rural areas is central to China’s rural revitalization strategy and requires efficient flows of labor, land, and capital [1,2,3,4]. Recent policy documents—including the 2020 Opinions on Improving the Market-Based Allocation of Factors of Production and the 2025 No. 1 Central Document—highlight the need to strengthen factor security and promote the orderly, market-based circulation of production factors across urban and rural areas [1,4,5]. Improving factor allocation is thus essential for enhancing farmers’ self-development capacity and supporting high-quality rural revitalization.

Since 2020, China has promoted the construction of ‘digital villages,’ integrating digital technology into agricultural production, rural governance, and household livelihoods [6,7]. Digital applications reduce information delays, lower transaction costs, and expand access to employment, land transfer, and credit services, thereby helping households make more informed production decisions and facilitating the allocation of labor, land, and capital [3]. Rapid improvements in rural connectivity—including high broadband penetration, accelerated 5G deployment, and the growth of e-commerce—indicate that rural households are increasingly benefiting from digital dividends [3,8]. Accordingly, digital technologies have become an important driver of optimized factor allocation, household wealth creation, and the integration of agricultural digitization with rural economic development [8].

The Yellow River Basin—often regarded as the cradle of Chinese civilization—holds strategic importance in national agricultural development. With intertwined ecological, energy, and economic corridor functions, it exemplifies how revitalizing local resources can stimulate rural economies and promote high-quality agricultural development alongside cultural continuity [9]. As a key ecological zone and economic belt, the Basin plays an indispensable role in ecological conservation, cultural heritage, and rural revitalization, thereby possessing irreplaceable strategic value for China’s sustainable development [10]. Since the 18th National Congress of the CPC, President Xi Jinping has conducted multiple field inspections in the region, emphasizing the coordination of ecological protection and socioeconomic advancement [11].

Despite these advantages, the Basin continues to face significant factor misallocation due to complex geography and underdeveloped factor market mechanisms [12]. Such inefficiencies waste resources, hinder agricultural transformation, constrain household wealth accumulation, and reduce overall socioeconomic efficiency [1,13,14], specifically in the following areas: (i) labor—the dual urban–rural structure and limited access to employment information restrict labor transfer and non-farm opportunities [15]; (ii) land—hierarchical social norms and the “differential mode of association” impede farmland transfer, and widespread informal transactions weaken the market’s price mechanism [16]; and (iii) capital—asymmetric credit information [17], limited financing demand [18,19], and risk-averse behavior among farmers [20,21] restrict the conversion of assets into productive loans [22]. These persistent distortions define the core problem of this study: whether and how digital technology adoption can alleviate rural household factor misallocation and support more efficient use of labor, land, and capital in the Yellow River Basin.

Existing research has explored digital technology’s role in agricultural production, non-farm employment, land transfer, and rural finance, but most studies examine single outcomes in isolation and rely on empirical strategies that only partially address selection bias and unobserved heterogeneity [23,24]. In particular, there is limited micro-level evidence on how digital technology jointly affects the allocation of labor, land, and capital within rural households, and very few studies focus systematically on the Yellow River Basin as an integrated ecological–economic region [25,26,27]. To fill these gaps, this paper pursues the following three objectives: (i) to estimate the impact of household digital technology adoption on the optimal allocation of labor, land, and capital; (ii) to identify the mechanisms through which digital technology operates, namely information access, social networks, and credit constraints; and (iii) to analyze heterogeneity in these effects across regions and household types. By jointly modeling multiple factor allocation decisions within an Endogenous Switching Probit framework and incorporating mechanism and heterogeneity analysis, the study contributes to the literature on digital transformation and rural factor market reform and provides empirical evidence to inform targeted, digital inclusive policies for the Yellow River Basin.

Using micro-survey data from rural households in six provinces—Gansu, Ningxia, Inner Mongolia, Shaanxi, Shanxi, and Henan—collected in 2020, 2022, and 2023, this study empirically examines how digital technology adoption affects the optimal allocation of rural household production factors in the Yellow River Basin, a region central to China’s high-quality development strategy [11]. The ESP model is applied to correct for self-selection and endogeneity biases, and the analysis is supplemented by robustness checks and heterogeneity tests by year, region, land scale, and age demographic group. Mechanism models further trace the pathways through which digital adoption affects resource allocation, focusing on information access, social networks, and credit constraints. Previewing the main empirical results, digital technology adoption exerts the strongest effect on labor allocation (non-farm employment), followed by moderate impacts on land transfer-out and a smaller but significant effect on access to productive loans. The findings provide theoretical and empirical support for rural factor market reform, the transformation of agricultural development models, and the construction of digital villages, offering practical insights for promoting inclusive rural economic growth.

2. Literature Review

2.1. Labor Reallocation and Non-Farm Employment

Labor, land, and capital are the three fundamental production factors in rural household economies, and their allocation decisions are shaped by both market conditions and household characteristics [5,28,29]. With the deepening of industrialization and urbanization, non-farm employment has become a key channel for absorbing surplus rural labor and raising household income [15,27,30,31]. A large body of empirical work shows that access to non-farm jobs reshapes household labor allocation, reduces underemployment in agriculture, and promotes livelihood diversification [27,30]. At the same time, the degree of part-time farming and the extent of labor reallocation depend on wage gaps, migration barriers, and information frictions in labor markets [14,17,32,33,34,35]. Resource endowments such as education, skills, and social capital also influence the likelihood of taking up non-farm work [36,37,38,39,40]. Overall, this literature suggests that better access to labor market information and lower transaction costs are associated with higher probabilities of off-farm employment and more efficient labor allocation [5,18,28,38,41,42,43,44]. Accordingly, we expect that digital technology adoption will increase households’ participation in non-farm employment by easing information and coordination frictions in labor markets.

2.2. Land Transfer Institutions and Transactions

Land is the core productive asset for rural households, and the functioning of land transfer institutions has important implications for factor allocation and structural transformation [15,27,30,31]. Incomplete property rights, hierarchical social norms, and the “differential mode of association” can impede formal land transfers and encourage informal or short-term arrangements, which weaken price signals and limit scale operations [15,27,30,31]. Empirical studies show that secure tenure, transparent registration, and contract enforcement promote the transfer of land from less efficient to more efficient producers, thereby improving land-use efficiency and encouraging specialization [14,17,32,33,34,35]. Transaction costs—such as search, negotiation, and monitoring—also play an important role: high costs discourage transfer-out and keep land fragmented, while reductions in transaction costs facilitate matching between landlords and tenants [36,37,38,39,40]. In sum, this literature indicates that institutional security and lower transaction costs are associated with higher probabilities of farmland transfer-out and more efficient land allocation. Building on this work, we expect that digital technology adoption can support land transfer-out by improving access to land market information, strengthening trust and monitoring, and reducing the costs of searching for and contracting with trading partners.

2.3. Rural Credit, Digital Finance, and Capital Allocation

Capital allocation in rural households is heavily influenced by credit constraints and the structure of local financial markets. Traditional studies document that asymmetric information, lack of collateral, and risk-averse behavior among farmers restrict their ability to convert assets into productive loans and undertake profitable investments [17,30,31,32,33]. Households therefore rely on a mix of formal, semi-formal, and informal credit channels, each with different costs and access conditions [18,19]. Recent work on inclusive and digital finance shows that new financial technologies can ease information frictions, broaden access to credit products, and reduce transaction and monitoring costs, thereby relaxing liquidity constraints and enabling more productive capital allocation [45,46,47,48,49,50,51,52]. At the same time, heterogeneity in financial literacy, digital skills, and local financial infrastructure shapes who can actually benefit from these innovations [43,44,45,46]. Overall, the evidence suggests that improved access to credit—especially through digital or inclusive finance—raises the likelihood that households obtain productive loans and invest in higher-return activities. We thus expect that digital technology adoption, by lowering information and access barriers in credit markets, will increase households’ access to productive loans and support more efficient capital allocation.

2.4. Household and Village Endowments

Household- and village-level endowments condition both digital technology adoption and factor allocation decisions. At the household-level, education, age structure, health, asset holdings, land scale, and social capital shape the capacity and incentives to reallocate labor, land, and capital [5,18,28,38,41,42,43,44]. At the village-level, infrastructure such as roads, broadband connectivity, public services, and local market development influences transaction costs and the availability of non-farm jobs, land transfer opportunities, and credit services [41,42]. Empirical studies find that households with better resource endowments and those located in more developed villages are more likely to participate in labor migration, land transfers, and formal credit markets, and are also more likely to adopt new technologies [5,18,28,38,41,42,43,44]. These findings imply that endowments can amplify or dampen the effects of digital technology adoption on factor allocation, leading to heterogeneous impacts across regions and household types. We therefore expect that stronger household and village endowments will be associated with both higher digital technology adoption and more active engagement in labor, land, and credit markets.

2.5. Digital Technology, Factor Misallocation, and Mechanism Channels

Empirical studies on factor misallocation emphasize that distortions in labor, land, and capital markets reduce aggregate productivity and slow structural transformation [14,17,32,33,34,35]. Existing work on digital technology in rural China shows that internet use, digital platforms, and digital finance can affect agricultural production, non-farm employment, land transfers, and credit access, but often focuses on single outcomes and relies on estimation strategies that only partially address selection bias [43,44,45,46,47,48,49,50,51,52,53,54,55,56]. Few studies jointly consider labor, land, and capital in households’ optimal allocation decisions, and even fewer explicitly link digital technology adoption to the combined configuration of these three factors [28,38,41,42,43,44]. Most existing research focuses on a single margin of adjustment—non-farm employment [13,22,23,24,25,26], farmland transfer [4,14,15,20,21,27,28,29,30,31,32], or agricultural credit [5,16,18,33,35,36,55]—and typically relies on OLS-, SUR-, or PSM-type estimators that only partially address selection bias and unobserved heterogeneity in factor allocation [11,28,37,48,49,50,51,52]. Studies on digital technology in rural China mainly examine its impact on agricultural production or specific market outcomes [49,50,51,52,53,54,55,57], with limited attention to how digital adoption reshapes the joint allocation of labor, land, and capital in key strategic regions such as the Yellow River Basin.

Against this backdrop, this study contributes in four ways. First, it adopts a joint factor allocation perspective, examining how digital technology simultaneously affects household participation in labor, land, and credit markets, thereby providing a more comprehensive picture of production factor optimization than single-outcome studies [5,11,17,37,38,39,40,41]. Second, it employs an Endogenous Switching Probit framework combined with counterfactual analysis to address both observable and unobservable selection in digital technology adoption, complementing earlier work that relies on conventional regression approaches [11,28,37,48,49,50,51,52]. Third, it uses multi-period household microdata from six provinces in the Yellow River Basin, enriching the evidence base on digital transformation and rural factor market reform in a less-developed but nationally strategic region. Fourth, it links digital adoption to concrete policy levers by clarifying the information, social network, and credit channels through which digital tools restructure household production factors, thereby offering targeted implications for rural revitalization policies in the Basin and comparable regions. Table 1 also specifies how each strand of literature relates to the optimal allocation of rural household labor, land, and capital.

Table 1. Representative literature on digital technology, rural transformation, and the optimal allocation of labor, land, and capital (2018–2025).

Author(s), Year	Journal/Region and Data	Focal Variables/Setting	Identification or Methods	Relation to Optimal Allocation of Labor, Land, and Capital
Li et al., 2024 [53]	Technological Forecasting and Social Change; China; quasi-natural experiment using “Broadband China” rollout	Digital infrastructure (broadband), urban–rural inequality	Policy quasi-experiment; spatial models	Broadband infrastructure reduces urban–rural income inequality by improving information access and enabling the reallocation of rural labor and capital from low-productivity agriculture to higher-return non-agricultural activities, thereby enhancing overall factor allocation efficiency.
Zheng et al., 2024 [54]	Scientific Reports; China; national microdata (household-level)	Digital economy index; land transfer-out; entrepreneurship; online social interaction	IV/mechanism analysis	Digital platforms and online social interaction foster rural entrepreneurship and non-farm jobs, promoting a more efficient allocation of household labor and capital between agricultural and entrepreneurial activities.
Guo et al., 2025 [55]	Information Technology for Development; China; policy exposure to Broadband China	Digital infrastructure; household entrepreneurship	Econometric evaluation of infrastructure shocks	Shocks to digital infrastructure significantly raise rural entrepreneurial activity, encouraging households to shift labor and capital toward higher productivity uses and improving the efficiency of factor allocation.
Chen et al., 2025 [23]	Social Indicators Research; China; national microdata	Internet usage; non-farm wage and local non-farm employment; wages/benefits	Microeconometric estimation	Household internet use increases the probability and quality of non-farm employment, suggesting a more optimal allocation of labor away from low-return farm work and better utilization of human capital.
Yu et al., 2024 [56]	SSM—Population Health; China; CGSS 2021 (rural women)	Internet embeddedness; non-farm work; rights perception (mediator)	Mediation analysis	Digital technology promotes non-farm employment partly by enhancing rights perception, helping households reallocate labor from subsistence agriculture toward better-protected, higher-quality jobs, thereby improving labor allocation.
Xu et al., 2024 [58]	Sustainability; China; provincial panel (2011–2022 as context)	Digital Inclusive Finance (DIF); land transfer	Panel regression; chain mediation (livelihood capital; digital information use)	Digital village construction strengthens livelihood capital and digital information use, which in turn supports more coordinated adjustment of labor and capital portfolios, moving households closer to an optimal factor mix.
He et al., 2025 [59]	Humanities and Social Sciences Communications (Nature portfolio); China	Broadband China policy; urban–rural income disparity	Policy evaluation; econometric decomposition	Digital inclusive finance reduces income disparity by relaxing credit constraints and guiding capital toward more productive households and sectors, thereby improving the efficiency of capital allocation.
Yan et al., 2025 [60]	Humanities and Social Sciences Communications; China; provincial panel 2011–2022	DIF; agricultural mechanization services	Fixed-effects model	Digital inclusive finance promotes investment in agricultural mechanization services, facilitating substitution of capital for labor and contributing to a more efficient labor–capital combination in agricultural production.
Liu et al., 2021 [57]	China Agricultural Economic Review (Elsevier page index); CFD/CFPS-type data	Digital inclusive finance; agricultural production and investment	IV/PSM-style causal checks	Digital inclusive finance stimulates agricultural investment and technology adoption, enabling households to upgrade capital and reorganize labor and land-use, thus improving the optimal allocation of production factors.
Deng et al., 2023 [61]	Plos one; China	Broadband China; urban–rural consumption gap; DIF	Policy evaluation; mediation	Broadband rollout and digital inclusive finance narrow the urban–rural consumption gap by expanding access to non-farm income and financial services, indirectly driving a more efficient reallocation of rural labor and capital.
Song et al., 2025 [62]	China Agricultural Economic Review; 285 prefecture-level cities	Digital economy index; high-quality agricultural development; labor migration	Panel models; robustness	The digital economy accelerates industrial integration and labor migration, promoting cross-regional and cross-sector reallocation of labor and capital that supports more efficient rural development.
Yuan et al., 2025 [63]	Frontiers in Sustainable Food Systems; China	Internet use; employment transformation of rural labor	Panel/logit variants (per PDF)	Internet use significantly accelerates the shift in rural labor from agriculture to non-farm sectors, indicating a more efficient labor allocation consistent with higher returns to human capital.
Nie et al., 2025 [24]	Agricultural and Food Economics (SpringerOpen); China	Internet access; female non-farm employment; income	Econometric models	Internet access increases women’s non-farm employment and income, improving intra-household labor allocation and making the use of family labor resources more efficient and gender-inclusive.

3. Theoretical Analysis and Research Hypotheses

3.1. Digital Technology and the Optimal Allocation of Production Factors

Digital technology has become a key driver of rural production efficiency due to the rapid expansion of digital infrastructure and network platforms in China [45,49,64,65,66,67]. From the perspective of household behavior theory, farmers are rational agents who adjust labor, land, and capital decisions according to expected returns and resource endowments [28,38,41,42,43,44,68]. Consistent with Coasean theory, digital tools reduce information, coordination, and transaction costs in employment, land transfer, and credit markets, thereby reshaping households’ production efficiency frontier and moving them closer to a Pareto-optimal allocation of production factors [52,67].

3.1.1. Labor Allocation

Digital adoption enhances non-farm employment opportunities by lowering labor market search costs and improving information flow [27,30,31,68]. Digital communication tools also strengthen social networks, reduce psychological barriers to migration, and facilitate stable, higher-quality employment. These improvements encourage households to reallocate labor from low-productivity agricultural work to higher-return non-farm sectors.

3.1.2. Land Allocation

Digital platforms improve farmland transfer-out by broadening access to land market information and reducing information asymmetry between transacting parties [17,34,69,70,71]. Digital social networks expand farmers’ social circles, create trustworthy negotiation channels, and improve contract enforcement. These functions reduce transaction risks and support more rational, transparent land transactions.

3.1.3. Capital Allocation

Digital finance tools improve access to productive loans by easing information frictions and expanding formal and quasi-formal credit sources [20,21,36,68]. Online credit-rating technologies lower borrowing costs, increase approval efficiency, and reduce spatial restrictions in traditional banking. As a result, digital adoption relaxes capital constraints and enables households to invest in production, mechanization, or entrepreneurship.

H1a. 

Digital technology adoption increases the probability of non-farm employment.

H1b. 

Digital technology adoption increases the probability of farmland transfer-out.

H1c. 

Digital technology adoption increases the probability of accessing productive loans.

H1. 

Digital technology adoption promotes the optimal allocation of labor, land, and capital in farm households.

3.2. Mechanisms Linking Digital Technology to the Optimal Allocation

3.2.1. Information Access Channel

Digital tools generate substantial information dividends by providing timely, reliable, and low-cost access to market information. This reduces search and transaction costs in labor markets, land transfers-out, and productive lending [5,22,27,30,49,69]. Enhanced job matching, greater land price transparency, and better credit market information help align household decisions with market signals and facilitate more efficient allocation of production factors (Figure 1).

H2. 

Digital technology adoption promotes optimal factor allocation by broadening information access.

3.2.2. Social Networks Channel

Digital technology transforms rural social relations by expanding and densifying interpersonal networks. Online communication reduces the cost of maintaining social ties, overcomes spatial and temporal barriers, and alleviates information disadvantages associated with traditional “differential mode of association” [15,49,72]. Stronger digital networks foster trust, improve contract reliability in land transfers [34], expand informal credit channels, mitigate information asymmetry, and reduce irrational borrowing behavior, all of which support more efficient factor allocation (Figure 1).

H3. 

Digital technology adoption promotes optimal factor allocation by strengthening social networks.

3.2.3. Credit Constraints Channel

Digital finance integrates big data analytics and cloud-based assessment into credit evaluation, enabling lenders to assess borrower creditworthiness using transaction histories and digital footprints [5,39]. As a result, farmers gain access to faster, lower-cost, often unsecured loans through platforms such as Alipay and WeChat Microloan. Improved credit access supports labor mobility [27,30], facilitates farmland transactions [35,71], alleviates precautionary motives, and enhances productive investment, further optimizing the joint allocation of labor, land, and capital (Figure 1).

H4. 

Digital technology adoption promotes optimal factor allocation by alleviating credit constraints.

Figure 1. Conceptual framework illustrating the mechanism through which digital technology adoption optimizes the allocation of production factors in rural households.

Figure 1. Conceptual framework illustrating the mechanism through which digital technology adoption optimizes the allocation of production factors in rural households.

4. Materials and Methods

4.1. Data Sources and Sampling Design

The empirical analysis draws on data from two large-scale survey programs: (1) Sustainable Development of Regional Characteristic Agricultural Policies, organized by Gansu Agricultural University, and (2) Ecological Protection and High-Quality Agricultural and Rural Development in the Yellow River Basin, conducted by Northwest A&F University. These surveys targeted the natural, geographic, economic development, and socioeconomic characteristics of the middle and upper reaches of the Yellow River Basin, encompassing six provinces—Gansu, Ningxia, Inner Mongolia, Shaanxi, Shanxi, and Henan.

To ensure representativeness and data reliability, a stratified random sampling method was employed. The sampling frame comprised registered rural households in agricultural administrative villages located within Yellow River Basin counties of the six provinces. Within each province, counties and townships were first grouped by basin location and level of agricultural development, then administrative villages were randomly selected within each stratum, and finally rural households were randomly drawn from village rosters provided by local committees. Primary data were collected through structured questionnaires and face-to-face interviews conducted during June–July 2020, July–August 2022, and June–August 2023. The three rounds captured household-level information for the reference years 2019, 2021, and 2022, respectively. The survey instrument covered a wide range of variables, including demographic characteristics, household assets, income and expenditure, industrial engagement, digital technology adoption, non-farm employment, farmland transfers, and access to credit. In total, the surveys encompassed 75 townships and 185 administrative villages across the six provinces. Urban residents, non-agricultural enterprises, collective units, and permanently migrated households were not eligible for interview, so the sample is representative of resident rural households in the surveyed Yellow River Basin counties rather than of all residents in these provinces.

This study systematically processed the data, and the specific procedures are as follows: Firstly, village-level data were matched with household-level data to construct a complete basic dataset. Secondly, samples were organized by household, retaining the information of the household head as the representative of the family and deleting duplicate samples to ensure the uniqueness of each observation. Subsequently, missing values of key variables were handled, and individual samples with missing information were deleted. The final dataset comprised 1280 valid households in 2020, 1340 in 2022, and 1310 in 2023, yielding a total of 3930 valid observations and an effective response rate of 89% (Table 2). All observations were merged into a pooled cross-sectional dataset, and subsequent empirical analyses were performed using Stata 18.

Table 2. Sample distribution.

Provinces	Year	Number of Administrative Villages	Number of Samples	Total Sample and Proportion by Province
Gansu	2020	25	180	857 (21.8%)
	2022	49	413
	2023	38	264
Ningxia	2020	21	282	738 (18.8%)
	2022	19	338
	2023	13	118
Inner Mongolia	2020	13	165	511 (13.0%)
	2022	10	113
	2023	15	233
Shaanxi	2020	39	150	536 (13.6%)
	2022	25	131
	2023	50	255
Shanxi	2020	64	372	852 (21.7%)
	2022	25	159
	2023	65	321
Henan	2020	21	131	436 (11.1%)
	2022	27	186
	2023	21	119

4.2. Variable Selection

4.2.1. Dependent Variables: Optimal Allocation of Household Production Factors

The dependent variable captures what we term the ‘optimal allocation of household production factors’, defined as the extent to which rural households participate in market-oriented reallocations of labor, land, and capital that move resources toward higher productivity uses. Following prior studies on rural transformation and factor market participation [15,20,21,27,30,36], we operationalize this construct through three complementary dimensions rather than a single composite index, because each dimension reflects a distinct but interrelated decision margin. Three binary indicators were used as follows:

Non-farm employment (1 = at least one household member engaged in off-farm work; 0 otherwise), representing labor factor allocation.

Farmland transfer-out (1 = household leased out its contracted farmland; 0 otherwise), representing land factor allocation.

Productive loan access (1 = household obtained loans for productive or agricultural activities; 0 otherwise), representing capital factor allocation.

4.2.2. Core Explanatory Variable: Digital Technology Adoption

The core explanatory variable is digital technology adoption, defined as the integration of internet-based applications with agricultural information systems and extension services [7,8,22,46,47,67]. Digital technology allows farmers to use online platforms and electronic devices to acquire information and services, apply them in production, lower information search costs, and improve household utility. As a new-generation technology, it merges internet applications with traditional agricultural information and extension systems, reconstructing information dissemination mechanisms. Consequently, farm households can obtain customized information and services for all stages of production and management, thus reducing search costs and enhancing utility maximization.

Adoption was measured through three dimensions:

Online learning: use of devices to search, learn, and apply agricultural- or input-related knowledge.

Online commerce: use of devices to obtain agricultural market information, policy updates, digital payments, online transactions, and e-lending services.

Social networking: use of platforms such as WeChat, Weibo, Douyin (TikTok), or Kuaishou for communication and social interaction.

Households adopting any of these functions were coded 1; otherwise, 0. This measure is binary and does not capture the intensity or partial degree of digital technology adoption.

4.2.3. Mechanism Variables

To capture potential transmission mechanisms [44], three variables were included as follows:

Information access: measured by the ease of obtaining production-, sales-, and livelihood-related information online (five-point scale; higher values = stronger capability).

Social networks: a composite index derived via the entropy method, incorporating network size, interaction strength, and trust.

Credit access: measured by perceived ease of obtaining loans (five-point scale; higher values = greater credit availability).

4.2.4. Control Variables

To mitigate omitted variable bias, eighteen control variables were selected across four dimensions:

Household head characteristics: gender, age, education, political affiliation, marital status, cadre identity, and health status.

Household characteristics: household size, child dependency ratio, and labor force share.

Land characteristics: number of plots, certification, and landholding size.

Village characteristics: presence of new agricultural business entities, village social relations, transport conditions, rural infrastructure, and financial service availability. A year dummy was also included to control for temporal effects.

4.2.5. Instrumental Variable

Following behavior imitation theory [34], the instrumental variable was defined as the average level of digital technology adoption among other households within the same village (excluding the target household) (Table 3). Peer adoption satisfies the relevance condition, as behavior in close rural communities influences individual decisions, and the exogeneity condition, as one household’s adoption does not directly determine another’s factor allocation outcome.

Table 3. Variable definitions and descriptive statistics.

Variable Type	Variable Name	Symbol	Meaning and Assignment of Variables		Mean	Std. Dev.
Dependent Variables	Allocation of Labor Factors	LF	Non-farm employment: Whether at least one rural household member engages in non-farm employment (Yes = 1, No = 0).		0.598	0.520
	Allocation of Agricultural Land Factors	ALF	Farmland transfer-out: Whether the rural household has leased out its farmland (Yes = 1, No = 0).		0.465	0.499
	Allocation of Capital Factors	CF	Productive loan access: Whether the household has obtained a productive loan for agricultural activities (Yes = 1, No = 0).		0.588	0.492
Core Explanatory Variable	Digital Technology Adoption	DTA	Whether the rural household uses digital devices (mobile phone, computer, etc.) for any of the following: online learning, e-commerce, or social networking (Yes = 1, No = 0).		0.673	0.583
Mechanism Variables	Information Access Channels	IAC	Ease of obtaining production/livelihood information online (Relatively Difficult = 1, Slightly Difficult = 2, Moderate = 3, Relatively Easy = 4, and Very Easy = 5).		3.697	1.140
	Social Network	SN	Social network—size	No. of online groups and offline community activities (count),	137.562	401.844
			Social network—strength	Gift-expenditure-to-total-expenditure ratio (%)	0.089	0.151
			Social network—trust	Level of trust in strangers (Extremely distrustful = 1, Slightly distrustful = 2, Neutral trust = 3, Fairly trusting = 4, and Extremely trusting = 5).	3.952	0.904
	Credit Constraint	CC	Credit access	Ease of obtaining loans (Relatively Difficult = 1, Slightly Difficult = 2, Moderate = 3, Relatively Easy = 4, and Very Easy = 5).	2.731	1.199
Control Variables	Household Head Characteristics	HHC	Gender	Household head’s gender (Male = 1, Female = 0).	0.953	0.241
			Age	Household head’s age (years).	55.736	11.187
			Education	Years of schooling of household head.	7.525	3.656
			Political affiliation	Household head is a Party member (Yes = 1, No = 0).	0.088	0.283
			Marital status	Household head is married (Yes = 1, No = 0).	0.782	0.413
			Cadre identity	Whether any household member is a cadre (Yes = 1, No = 0).	0.066	0.249
			Health status	Household head’s health (1 = unhealthy, 2 = average, and 3 = healthy).	2.829	0.488
	Family Characteristics	FC	Household size	Total number of household members (persons).	3.985	1.677
			Child dependency ratio	Proportion of dependents in the household: ratio of dependents under age 16 to household labor force.	0.284	0.215
			Labor force share	Share of labor force in total household members (%).	0.665	0.359
	Characteristics of Agricultural Land	CAL	Number of plots	Number of farmland plots owned (pieces).	6.105	13.403
			Land certification	Whether household has a land-use rights certificate (Yes = 1, No = 0).	0.890	0.313
			Landholding size	Per capita farmland area (mu/person).	5.596	12.455
	Village Characteristics	VC	New agricultural business entities	Whether the village has family farms, cooperatives, or agricultural enterprises (Yes = 1, No = 0).	0.075	0.264
			Village social relations	Whether mutual help among neighbors exists in agricultural production (Yes = 1, No = 0).	0.927	0.261
			Transport conditions	Distance from village to township/county (km).	11.736	11.841
			Rural infrastructure	Whether the village has 5G base stations or fiber-optic networks (Yes = 1, No = 0).	0.105	0.307
			Financial service availability	Whether the village has a convenience financial service station (Yes = 1, No = 0).	0.770	0.321
Instrumental Variable	Peer Adoption of Digital Technology	PADC	Average level of digital technology adoption among other households in the same village (excluding respondent household).		0.694	0.512

Note: The mechanism variable ‘social network’ is a composite index constructed using the entropy method that integrates three fundamental dimensions: social network size, social network strength, and social network trust.

The reliability of the instrumental variable was first verified. Based on the first-stage regression results, the instrument exhibits a statistically significant positive impact on an individual household’s adoption probability, thus satisfying the relevance condition. Regarding the validity of the instrumental variable, this paper regresses the instrumental variable and control variable on the explained variable. The results show that the estimated coefficients of the instrumental variable on non-agricultural employment, farmland transfer, and agricultural production loans are 0.058, 0.077, and 0.078, respectively, with corresponding p-values of 0.343, 0.245, and 0.159. This indicates that the instrumental variable has no significant impact on non-agricultural employment, farmland transfer, and agricultural production loans, satisfying the exogeneity condition of the instrumental variable. Additionally, the estimation results of the instrumental variable model (IV-2SLS) for digital technology usage and the three explained variables show that the F value in the first stage is 10.764, which is greater than the maximum instrumental variable critical value of 8.96 at the 15% level of the Stock–Yogo test, indicating that the instrumental variable is not a weak instrumental variable. Therefore, “the average level of digital technology adoption among other households within the same village (excluding the target household)” is an effective instrumental variable.

4.3. Modeling Specification

(1)

Endogenous Switching Probit (ESP) Model

Households’ decisions to adopt digital technology and their factor allocation outcomes are endogenous: both reflect self-selection and may depend on unobservables, leading to biased estimates if not modeled jointly [72]. Moreover, each household is observed in only one state (adopter or non-adopter), so counterfactual outcomes are missing [73]. To address selection on observables and unobservables, we employ the Endogenous Switching Probit (ESP) framework as the baseline specification. Compared with the strong reliance of the propensity score matching method on observable variables and the limitation of the instrumental variable Probit model that can only provide local average treatment effects, this method not only effectively corrects selection bias but also reconstructs the outcome distribution of the sample in the counterfactual state, thereby obtaining more generally meaningful treatment effect estimates. In this model, the ‘selection equation’ refers specifically to the first-stage digital technology adoption equation, constructed to identify the factors influencing rural households’ adoption of digital technology, which estimates the probability that a household adopts digital tools and generates the correction terms used in the outcome equations.

The ESP estimation proceeds in two stages.

Stage 1 (selection equation). Let D_i denote the latent propensity to adopt digital technology and D_i the observed adoption status [72]:

$$D_i^{\ast } ={ \gamma Z}_i + {\mu }_i$$

(1)

$$D_i =\left\{\begin{array}{c}1, D_i^{\ast } > 0\\ 0,{ D}_i^{\ast } \le 0\end{array}\right.$$

(2)

where $D_i^{\ast }$ = 1 indicates that the rural household adopts digital technology, while $D_i^{\ast }$ = 0 indicates that the rural household does not adopt digital technology. $Z_i$ represents exogenous variables affecting adoption, including individual, household, land, and village characteristics. $\gamma$ is the parameter vector to be estimated, and μ_i is the error term.

Stage 2 (regime-specific outcome equations). Under two scenarios—adoption and non-adoption of digital technology—we construct the outcome equations to examine the impact of digital technology use on the optimal allocation of production factors in rural households as follows:

$$R_{1i{D}_{\mathit{ij}}}^{\ast }=X_{1i}{\beta }_{1D_{\mathit{ij}}}+{\epsilon }_{1i{D}_{\mathit{ij}}},if D_i=1$$

(3)

$$R_{0i{D}_{\mathit{ij}}}^{\ast }=X_{0i}{\beta }_{0D_{\mathit{ij}}}+{\epsilon }_{0i{D}_{\mathit{ij}}},if D_i=0$$

(4)

here, D_i denotes rural households’ digital technology adoption status: D_i = 1 indicates that the rural household adopts digital technology, otherwise D_i = 0. j indexes outcome categories: non-farm employment (j = 1), farmland transfer (j = 2), and productive loans (j = 3). R_1i^*_Dij and R_0i^*_Dij represent the latent and observed outcomes, respectively. X_i denotes household- and village-level covariates; β_Dij are parameters to be estimated; and ε_iDij are error terms.

Due to the self-selection problem in digital technology adoption, the correlation coefficient between ${\mathrm{\epsilon }}_{{\mathrm{i}0}_{\mathrm{j}}}$ and ${\mathrm{\epsilon }}_{{\mathrm{i}1}_{\mathrm{j}}}$, namely ρ₂, is not zero. Direct estimation of Equations (3) and (4) will yield inconsistent results. It is assumed that the three random error terms (${\mu }_i{,\epsilon }_{{i1}_j},{\epsilon }_{{i0}_j}$) follow a trivariate joint normal distribution with a mean of 0, and their variance–covariance structure is as follows:

$$\Omega =\left[\begin{array}{ccc}{\sigma }_{\mu }^2& {\sigma }_{\mu 1}& {\sigma }_{\mu 0}\\ {\sigma }_{1\mu }& {\sigma }_1^2& {\sigma }_{10}\\ {\sigma }_{0\mu }& {\sigma }_{01}& {\sigma }_0^2\end{array}\right]$$

(5)

Among them, ${\sigma }_{\mu }^2$, ${\sigma }_1^2$, and ${\sigma }_0^2$ denote the variances of the error terms in Equations (1), (3), and (4), respectively; ${\sigma }_{\mu 1}$ and ${\sigma }_{1\mu }$ represent the covariance between ${\mu }_i$ and ${\epsilon }_{{i1}_j}$; ${\sigma }_{\mu 0}$ and ${\sigma }_{0\mu }$ stand for the covariance between ${\mu }_i$ and ${\epsilon }_{{i0}_j}$; and since R_1i^*_Dij and R_0i^*_Dij cannot be observed simultaneously, ${\sigma }_{10}$ and ${\sigma }_{01}$ are not defined. By setting the correlation coefficients between the error terms of the selection equation and the outcome equations as ${\rho }_1$ = corr(${\epsilon }_{{i1}_j}$,${\mu }_i$) and ${\rho }_0$ = corr(${\epsilon }_{{i0}_j}$,${\mu }_i$), the conditional expectations of ε_i1j and ε_i0j can be expressed as follows:

$$E[{\epsilon }_{{i1}_j}|D_i=1]={\sigma }_{1\mu }{\displaystyle \frac{\phi (\gamma Z_i/\sigma )}{\Phi (\gamma Z_i/\sigma )}}={\sigma }_{1\mu }{\lambda }_{1i}$$

(6)

$$E[{\epsilon }_{{i0}_j}|D_i=0]={\sigma }_{0\mu }{\displaystyle \frac{-\phi (\gamma Z_i/\sigma )}{1-\Phi (\gamma Z_i/\sigma )}}={\sigma }_{0\mu }{\lambda }_{0i}$$

(7)

In Equations (6) and (7), ${\lambda }_i$ denotes the inverse Mills ratio. Following Lokshin and Sajaia [73], the estimation process of the ESP model is as follows: in the first stage, Probit regression is performed using the Full Information Maximum Likelihood (FIML) method to estimate the probability of rural households using digital technology; and in the second stage, the inverse Mills ratio calculated from the first stage is incorporated into the outcome equations for bias correction, so as to ensure the consistency of parameter estimates.

(2)

Counterfactual Analysis

Based on the ESP estimates, we apply counterfactual analysis to estimate the Average Treatment Effect of digital technology use. This is achieved by comparing the behaviors of rural households—including non-farm employment, farmland transfer, and agricultural productive loan access—between those who used or did not adopt digital technology under both actual and counterfactual scenarios:

The expected value of the optimal allocation of household production factors for rural households that have adopted digital technology is as follows:

$$E[R_{1i{D}_{\mathit{ij}}}^{\ast }|D_i=1]=X_{1i}{\beta }_{1D_{\mathit{ij}}}+{\sigma }_{1\mu }{\lambda }_{1i}$$

(8)

The expected value of the optimal allocation of household production factors for rural households that have not adopted digital technology is as follows:

$$E[R_{0i{D}_{\mathit{ij}}}^{\ast }|D_i=0]=X_{0i}{\beta }_{0D_{\mathit{ij}}}+{\sigma }_{0\mu }{\lambda }_{0i}$$

(9)

In the counterfactual scenario, the expected value of the optimal allocation of household production factors for rural households that have adopted digital technology (under the condition of not adopting it) is as follows:

$$E[R_{0i{D}_{ij}}^{\ast }|D_i=1]=X_{1i}{\beta }_{0D_{ij}}+{\sigma }_{0\mu }{\lambda }_{1i}$$

(10)

The expected value of the optimal allocation of household production factors for rural households that have not adopted digital technology (under the condition of using it) is as follows:

$$E[R_{1i{D}_{\mathit{ij}}}^{\ast }|D_i=0]=X_{0i}{\beta }_{1D_{\mathit{ij}}}+{\sigma }_{1\mu }{\lambda }_{0i}$$

(11)

From Equations (8) and (10), the treatment effect of digital technology adoption on the optimal allocation of household production factors for rural households that have adopted digital technology can be derived as follows:

$$ATT=E[R_{1i{D}_{\mathit{ij}}}^{\ast }|D_i=1] - \mathrm{E}[R_{0i{D}_{\mathit{ij}}}^{\ast }|D_i=1]=X_{1i} ({\beta }_{1D_{\mathit{ij}}}-{\beta }_{0D_{\mathit{ij}}}) +{\lambda }_{1i}({\sigma }_{1\mu }-{\sigma }_{0\mu })$$

(12)

From Equations (9) and (11), the treatment effect of digital technology adoption on the optimal allocation of household production factors for rural households that have not adopted digital technology can be derived as follows:

$$ATU=E[R_{1i{D}_{ij}}^{\ast }|D_i=0] - E[R_{0i{D}_{ij}}^{\ast }|D_i=0]=X_{0i} ({\beta }_{1D_{ij}}-{\beta }_{0D_{ij}})+{\lambda }_{0i}({\sigma }_{1\mu }-{\sigma }_{0\mu })$$

(13)

(3)

Mechanism Testing Model

Following Ting [74], when the causal relationship between mediating variables and outcomes is theoretically clear, it is sufficient to test whether the explanatory variable influences the mediators. Accordingly, the following mechanism test examines whether digital technology adoption significantly affects information access, social networks, and credit constraints:

$$M_{\mathit{ni}}={\alpha }_0+{\alpha }_1{D}_i+\beta {\mathit{controls}}_i+{\epsilon }_i$$

(14)

where M_ni represents the mediating variable (n = 1, 2, and 3 for information access, social networks, and credit constraints, respectively); α₀ is the constant term; D_i indicates digital technology adoption; ${\mathit{controls}}_i$ are control variables; α₁, β are parameters to be estimated; and ε_i is the error term.

5. Results and Discussion

5.1. Instrument Validity and Model Diagnostics

After the verification in the previous text, the instrumental variables in this article meet the conditions of relevance, exogeneity, and non-weak instrumental variables. On this basis, this paper uses the above-mentioned valid instrumental variables to directly test the existence of endogeneity through the Conditional Mixed Process (CMP) model (Table 4). The results show that the correlation coefficient atanh (ρ₁₂) is significant at the 5% level, confirming that unobservable characteristics affect both the adoption decision and production factor allocation, and clarifying the necessity of dealing with endogeneity.

Table 4. Conditional Mixed Process (CMP) model estimates.

Variable	Non-Farm Employment The Second Stage of the CMP Model (1)	Adopt Digital Technology The First Stage of the CMP Model (2)	Farmland Transfer-Out The Second Stage of the CMP Model (3)	Adopt Digital Technology The First Stage of the CMP Model (4)	Productive Loan Access The Second Stage of the CMP Model (5)	Adopt Digital Technology The First Stage of the CMP Model (6)
adopt digital technology	0.328 ** (0.127)		0.553 ** (0.255)		0.659 *** (0.208)
instrumental variable		0.479 *** (0.165)		0.411 ** (0.208)		0.486 *** (0.173)
control variables	Controlled	Controlled	Controlled	Controlled	Controlled	Controlled
atanh (ρ12)	0.384 ** (0.161)		−0.415 ** (0.192)		−0.326 ** (0.175)
N	3930	3930	3930	3930	3930	3930

Notes: Standard errors are in parentheses. ** and *** indicate significance at the 5% and 1% levels, respectively.

Consequently, the Endogenous Switching Probit (ESP) model was employed to correct for selection bias. As shown in Table 5, model-fit statistics are highly significant (p < 0.01), and at least one of the correlation coefficients ρ₀ or ρ₁ is non-zero, rejecting the null of equation independence. These results validate the ESP framework and ensure consistent estimation of treatment effects.

Table 5. Baseline regression results: Endogenous Switching Probit (ESP) model estimates.

Variable	Selection Equation (Whether to Adopt Digital Technology)	Non-Farm Employment Result Equation		Selection Equation (Whether to Adopt Digital Technology)	Farmland Transfer-Out Result Equation		Selection Equation (Whether to Adopt Digital Technology)	Productive Loan Access Result Equation
	(1)	Users (2)	Non-Users (3)	(4)	Users (5)	Non-Users (6)	(7)	Users (8)	Non-Users (9)
Instrumental variable	0.316 *** (0.126)			0.474 *** (0.154)			0.492 *** (0.137)
Gender	0.223 ** (0.093)	0.309 *** (0.112)	−0.073 (0.109)	0.094 *** (0.002)	0.085 (0.118)	0.059 *** (0.021)	0.267 *** (0.091)	0.284 *** (0.102)	−0.2308 (0.824)
Age	−0.007 *** (0.002)	−0.005 *** (0.002)	−0.001 (0.023)	−0.007 *** (0.002)	−0.008 *** (0.002)	−0.004 *** (0.001)	−0.006 *** (0.002)	−0.012 (0.021)	−0.048 (0.038)
Education	0.012 ** (0.006)	0.005 (0.006)	0.085 (0.065)	0.012 ** (0.006)	0.019 ** (0.009)	0.015 (0.013)	0.019 *** (0.006)	0.164 ** (0.083)	0.022 ** (0.011)
Political affiliation	0.132 *** (0.043)	0.103 (0.083)	0.097 (0.084)	0.126 * (0.073)	0.143 (0.103)	−0.033 (0.088)	0.107 *** (0.042)	0.014 (0.066)	0.043 (0.134)
Marital status	0.202 *** (0.054)	0.368 *** (0.058)	0.490 *** (0.062)	0.203 *** (0.055)	−0.058 (0.130)	0.040 ** (0.017)	0.213 *** (0.054)	0.173 *** (0.061)	0.143 ** (0.069)
Cadre identity	0.046 (0.081)	0.049 (0.095)	−0.008 (0.096)	0.048 (0.087)	0.138 (0.107)	0.049 (0.105)	0.048 (0.084)	−0.106 (0.915)	−0.067 (0.153)
Health status	0.007 (0.042)	0.125 *** (0.044)	0.005 (0.049)	−0.018 (0.043)	0.017 ** (0.008)	0.021 (0.066)	−0.277 (0.420)	−0.195 (0.465)	0.130 (0.741)
Household size	−0.019 * (0.010)	−0.039 *** (0.015)	−0.032 ** (0.016)	−0.017 ** (0.008)	0.009 (0.017)	−0.008 (0.017)	−0.013 *** (0.004)	−0.010 (0.004)	0.022 (0.025)
Child dependency ratio	0.194 ** (0.088)	−0.018 (0.120)	−0.170 (0.104)	0.190 ** (0.087)	0.131 (0.140)	−0.181 (0.165)	0.233 *** (0.085)	0.136 (0.198)	0.156 * (0.087)
Labor force share	0.027 (0.063)	0.756 *** (0.068)	0.036 (0.053)	0.018 (0.063)	0.021 (0.083)	−0.012 (0.072)	0.055 (0.067)	0.089 (0.172)	−0.029 (0.103)
Number of plots	0.009 (0.017)	0.025 (0.021)	0.001 (0.002)	0.001 (0.002)	0.004 ** (0.002)	0.016 (0.024)	0.025 (0.029)	0.018 (0.019)	0.094 (0.096)
Land certification	0.010 (0.063)	−0.080 (0.068)	0.134 (0.093)	0.089 (0.076)	0.216 ** (0.090)	0.108 (0.097)	0.185 (0.645)	−0.069 (0.704)	−0.297 (0.841)
Land size	0.106 *** (0.037)	0.094 *** (0.031)	−0.083 (0.075)	0.092 *** (0.031)	0.006 ** (0.003)	0.008 (0.021)	0.096 *** (0.028)	−0.007 (0.026)	−0.011 (0.028)
New agricultural business entities	0.125 (0.081)	−0.128 (0.086)	0.037 (0.097)	0.143 (0.181)	0.078 (0.176)	−0.104 (0.132)	0.094 (0.118)	0.087 *** (0.031)	0.178 ** (0.076)
Village social relations	0.137 *** (0.043)	−0.233 (0.198)	0.265 (0.198)	0.165 ** (0.083)	−0.021 (0.134)	0.246 ** (0.100)	0.084 ** (0.041)	0.330 ** (0.143)	0.101 (0.192)
Traffic conditions in villages	0.095 *** (0.037)	0.101 *** (0.032)	−0.068 (0.078)	0.086 *** (0.025)	0.005 *** (0.002)	0.009 (0.022)	0.091 *** (0.032)	0.024 (0.201)	0.020 (0.103)
Rural infrastructure	0.042 ** (0.017)	0.152 ** (0.077)	−0.068 (0.078)	0.026 *** (0.008)	−0.088 (0.087)	−0.022 (0.080)	0.028 ** (0.014)	0.075 ** (0.035)	−0.142 (0.172)
Situation of convenient financial services in villages	−0.088 (0.074)	0.033 (0.051)	−0.005 (0.058)	−0.102 (0.653)	0.095 (0.713)	−0.075 (0.657)	−0.959 (0.997)	1.251 *** (0.056)	0.736 (0.785)
Constant term	−0.014 (0.247)	−0.659 *** (0.248)	−0.140 (0.267)	−0.050 (0.262)	−0.302 (0.6230)	0.436 (0.395)	0.087 (0.245)	0.305 (0.245)	−1.891 *** (0.554)
ρ1		−11.672 *** (3.509)			−0.181 (0.863)			−0.994 *** (0.010)
ρ0			7.346 *** (2.255)			−3.393 *** (1.248)			−0.271 (0.476)
Goodness-of-fit test for models	90.07 ***			89.83 ***			96.47 ***
Log pseudo-likelihood value	−483.80			−520.990			−453.350
Independence test of equations	37.880 ***			24.526 ***			12.290 ***
Sample size	3930			3930			3930

Note: Robust standard errors are reported in parentheses. *** p < 0.01, ** p < 0.05, * p < 0.10. ρ₁ and ρ₀ denote the estimated correlations between the selection equation and the outcome equations for adopters and non-adopters, respectively; statistical significance of either parameter indicates selection on unobservables. The bottom row reports the Wald “independence test of equations” (H₀: ρ₁ = ρ₀ = 0); rejection of H₀ supports the use of the ESP framework. Selection equation: Whether to adopt digital technology represents the first-stage adoption equation of the ESP model. This equation estimates the probability of digital technology adoption and generates the correction terms used in the outcome equations for labor, land, and capital allocation.

5.2. Determinants of Digital Technology Adoption

Table 5 presents the determinants of adoption behavior. The following individual characteristics play a decisive role: gender, education, political affiliation, and marital status positively influence adoption, while age exerts a strong negative effect. These results imply that younger, better-educated, and socially connected male farmers are more likely to engage with digital tools. Similar findings are reported by Yang et al. [75] in China Economic Review, who show that digital finance diffusion is driven by younger, educated households capable of processing information efficiently.

At the household-level, a higher child dependency ratio significantly raises the likelihood of adoption, suggesting that families with dependents are more inclined to explore new income-enhancing technologies. In contrast, larger household size reduces adoption probability, likely because larger households face greater resource allocation challenges and decision-making frictions, leading to coordination difficulties in technology uptake.

In terms of land characteristics, the positive association between land scale and adoption indicates that larger farms, being more market-oriented, seek digital platforms for price discovery and input procurement.

Village characteristics also exert a pronounced influence. Strong social relations, improved transportation, and modern infrastructure (5G and fiber networks) all increase adoption. A possible explanation for this is that stronger trust networks encourage information exchange and digital communication, while better infrastructure and transportation lower barriers to technology adoption. These factors lower information and transaction costs and promote information diffusion. This result echoes Zou and Mishra [76] in Computers and Electronics in Agriculture, who found that better digital and transport infrastructure accelerates land market participation through enhanced communication efficiency. Collectively, these findings confirm that both individual capacity and institutional context jointly shape digital technology adoption in rural China.

5.3. Factor Allocation Behavior Among Adopters and Non-Adopters

The outcome equations in Table 5 (columns 2–3, 5–6, and 8–9) reveal distinct patterns in the determinants of labor, land, and capital allocation between adopters and non-adopters.

Labor allocation. For both adopters and non-adopters, younger age, higher education, better health, and shorter distance to town significantly increase the probability of non-farm employment, indicating that human capital and local infrastructure are key drivers of labor reallocation. Conditional on these determinants, digital technology adoption is associated with a substantially higher likelihood of non-farm work, consistent with the largest ATT effects in Table 4 and with evidence that digital inclusion enhances labor mobility and accelerates structural transformation. These results are consistent with Yang, Wu and Huang [75], who emphasized that digital inclusion enhances labor mobility and accelerates structural transformation.

Land allocation. Among adopters, education, health, land certification, larger contracted area, and better transportation conditions significantly raise the probability of transferring out land, while older age reduces it, suggesting that secure tenure and improved infrastructure facilitate formal land transactions. Relative to non-adopters, digital adopters are more likely to transfer out land, in line with the ATT estimates and the view that internet use formalizes land exchange mechanisms and reduces uncertainty. This reinforces the argument of Zou and Mishra [76] that internet use increases the probability of farmland transfer-out by formalizing transactions and reducing uncertainty.

Capital allocation. For capital, education, marital status, village financial services, and social relations are significant determinants of access to productive loans, highlighting the combined role of human capital, social capital, and local financial infrastructure. Conditional on these factors, digital adopters exhibit a higher probability of obtaining productive loans than comparable non-adopters, which is consistent with the positive but smaller ATT effects for capital and with findings that fintech penetration expands rural credit supply. Digital platforms reduce credit information asymmetry and link farmers with formal lenders, aligning with Zhang et al. [77], who observed that fintech penetration expands rural credit supply.

5.4. Treatment Effects of Digital Technology Adoption

Counterfactual analysis based on the ESP estimates quantifies the magnitude of adoption effects (Table 6). The ATT of digital technology adoption on non-farm employment, farmland transfer-out, and productive loan access is 0.245 (95% CI [0.210, 0.280]), 0.193 (95% CI [0.160, 0.226]), and 0.167 (95% CI [0.136, 0.198]), respectively, implying that, had adopters not engaged with digital technologies, their likelihood of reallocating labor, land, and capital in these ways would have been 24.5, 19.3, and 16.7 percentage points lower. The corresponding ATU estimates—0.274, 0.222, and 0.189—summarize the potential gains for non-adopters if they were to adopt.

Table 6. Estimates of Average Treatment Effects (ATT and ATU) of digital technology adoption on factor allocation.

Farmer Category	Adoption of Digital Technology Adoption	Non-Adoption of Digital Technology	ATT			ATU
			Non-Farm Employment	Farmland Transfer-Out	Productive Loan Access	Non-Farm Employment	Farmland Transfer-Out	Productive Loan Access
adopters	0.598 (0.010)	0.353 (0.009)	0.245 *** (0.013)	/	/	/
	0.587 (0.012)	0.394 (0.006)	/	0.193 *** (0.009)	/
	0.562 (0.007)	0.395 (0.010)	/	/	0.167 *** (0.011)
Non-adopters	0.661 (0.006)	0.387 (0.005)	/			0.274 *** (0.008)	/	/
	0.633 (0.005)	0.411 (0.007)				/	0.222 *** (0.009)	/
	0.598 (0.008)	0.409 (0.005)				/	/	0.189 *** (0.004)

Note: ATT = Average Treatment Effect on the Treated; ATU = Average Treatment Effect on the Untreated. Standard errors in parentheses; *** p < 0.01. Estimates are derived from the Endogenous Switching Probit (ESP) model using 3930 household-level observations.

Economically, these estimates correspond to approximately 961 additional non-farm employment arrangements, 759 extra farmland transfer-out contracts, and 656 extra productive loans across the 3930-household sample. The magnitude hierarchy (labor > land > capital) indicates that digital adoption first and foremost promotes labor reallocation, while also meaningfully strengthening land and credit market participation, which together substantiate Hypothesis H1 and corroborate earlier evidence from Yang, Wu, and Huang [75], who reported that digital technology development reallocates production factors toward more efficient sectors.

5.5. Mechanism Analysis

Building on the ATT estimates in Table 6, the effects of digital technology adoption are largest for non-farm employment, moderate for farmland transfer-out, and smaller but still statistically significant for access to productive loans. This hierarchy is consistent with the notion that information effects materialize fastest in labor markets, that transaction formalization and trust-building operate more gradually in land markets, and that digital finance relaxes capital constraints more slowly but meaningfully. The mechanism analysis below maps these three channels—information access, strengthened social networks, and alleviated credit constraints—onto the observed pattern of treatment effects on labor, land, and capital allocation. The mechanism tests (Table 7) identify the channels through which digital technology promotes optimal factor allocation.

Table 7. Mechanism test results.

Variable	Expanding Channels for Information Acquisition	Strengthening Social Networks	Alleviating Credit Constraints
	Model 1	Model 2	Model 3
Digital Technology Adoption	0.202 *** (0.041)	0.127 *** (0.011)	0.124 *** (0.038)
Controls	Controlled	Controlled	Controlled
Dummies	Controlled	Controlled	Controlled
R2	0.039	0.070	0.108
Observations	3930	3930	3930

Note: Robust standard errors in parentheses. *** denotes significance at the 1% level.

5.5.1. Mechanism 1: Expanding Information Access

First, information access is higher among adopters. Table 7 (Model 1) shows that digital technology adoption is positively and significantly associated with the information access index after controlling for household and village characteristics. Within the ESP framework, this suggests that adopters obtain more timely and accurate information on employment opportunities, land prices, and credit products, which is consistent with the larger ATT estimates observed for non-farm employment relative to land transfer-out and productive loans. Hypothesis H2 is supported by lowering search and negotiation costs and reducing information asymmetry [50,66].

5.5.2. Mechanism 2: Strengthening Social Networks

Second, social networks appear stronger for adopters. Table 7 (Model 2) indicates that digital technology adoption is positively related to the social network index, and the coefficient is statistically significant. This pattern is consistent with the idea that online communication tools reduce the cost of maintaining ties and facilitate reputation-building and cooperation in both land and labor markets, helping to lower relational and transaction barriers to farmland transfer-out and off-farm work. The result supports Hypothesis H3 and aligns with [1], who noted that online interaction enhances cooperative behavior in land and labor markets.

5.5.3. Mechanism 3: Easing Credit Constraints

Third, credit constraints are less severe among adopters. Table 7 (Model 3) shows that digital technology adoption is significantly and negatively associated with the credit constraint indicator at the 1% level. This association is in line with the view that digital finance and online credit services expand access to productive loans, especially where financial infrastructure is better developed, even though the ATT estimates for capital are smaller than those for labor and land [21,37,38,39]. It helps them improve their credit evaluation levels and encourages them to use new digital inclusive financial products and wealth management software, effectively alleviating their credit constraints consistent with Hypothesis H4.

5.6. Heterogeneity Across Time, Region, Land Scale, and Age

While the baseline and mechanism results show that digital adoption primarily operates through information, social networks, and credit channels, these channels are unlikely to be equally strong in all contexts. To examine where they ‘bite’ most strongly, we estimate subgroup-specific treatment effects by year, region, farmland scale, and age group. The heterogeneity patterns reported in Table 8 thus provide evidence on how the information, transaction formalization, and digital finance channels vary over time and across different types of households and regions.

Table 8. Heterogeneous effects of digital technology adoption on the optimal allocation of household production factors.

Category	Group	Non-Farm Employment	Farmland Transfer-Out	Productive Loan Access
By Year	2019	0.237 *** (0.052)	0.169 *** (0.037)	0.126 *** (0.040)
	2021	0.252 *** (0.054)	0.192 *** (0.041)	0.151 *** (0.043)
	2022	0.276 *** (0.055)	0.195 *** (0.043)	0.183 *** (0.048)
By Region	Upper Yellow River	0.297 *** (0.059)	0.201 *** (0.031)	0.217 *** (0.043)
	Middle Yellow River	0.266 *** (0.057)	0.194 *** (0.029)	0.208 *** (0.041)
By Farmland Scale	Small-scale	0.186 *** (0.042)	0.195 *** (0.031)	0.196 *** (0.049)
	Medium-scale	0.205 *** (0.046)	0.217 *** (0.032)	0.209 *** (0.052)
	Large-scale	0.227 *** (0.052)	0.231 *** (0.034)	0.218 *** (0.054)
By Age Group	Younger Generation (≤55 years)	0.236 *** (0.054)	0.197 *** (0.038)	0.225 *** (0.051)
	Older Generation (>55 years)	0.199 *** (0.048)	0.172 *** (0.034)	0.193 *** (0.043)

Note: Robust standard errors in parentheses. *** p < 0.01. ATT values are derived from the Endogenous Switching Probit model to capture subgroup-specific treatment effects.

5.6.1. By Year

The heterogeneity results (Table 8) reveal that the effects of digital adoption intensify over time (Figure 2 and Figure 3). Between 2019 and 2022, the contribution to non-farm employment increased from 23.7 percent to 27.6 percent, reflecting the accelerating spread of digital infrastructure and rural e-commerce policies under China’s 14th Five-Year Plan, it has provided farmers with more non-agricultural employment opportunities. The impact on farmland transfer-out peaked in 2021, aligning with new national programs promoting “Smart Agriculture + Land Circulation,” which makes digital technology play a pivotal role in driving the transfer-out of agricultural land. While digital technology progressively expanded credit inclusion, raising productive loan access by 18.3 percent by 2022 because it enhances farmers’ access to finance and the diversity of production factors. As time goes by, it can alleviate the difficulties faced by rural households in obtaining loans and financing, and the facilitating role of digital technology becomes more pronounced.

Figure 2. Average Treatment Effects on the Treated (ATT) of digital technology adoption on the optimal allocation of household production factors by year. The figure shows ATT values for labor (non-farm employment), land (farmland transfer-out), and capital (productive loan access) derived from the Endogenous Switching Probit (ESP) model for 2019, 2021, and 2022.

Figure 3. Temporal trend of the Average Treatment Effect on the Treated (ATT) of digital technology adoption on productive loan access from 2019 to 2022. The upward trend reflects the increasing contribution of digital adoption to alleviating credit constraints and improving rural households’ access to productive finance, consistent with the broader expansion of digital financial services during the study period.

5.6.2. By Region

Regional differences show stronger effects in the upper reaches of the Yellow River Basin. The heterogeneity results (Table 8) reveal that compared with the counterfactual, adopters in the upper reaches were 29.7% more likely to engage in non-farm work, 20.1% more likely to transfer farmland, and 21.7% more likely to obtain productive loans. In contrast, the corresponding probabilities for adopters in the middle reaches were 26.6%, 19.4%, and 20.8%. This suggests that digital technology generates larger benefits in economically less-developed areas. In the upper reaches, when digital technologies are widely used by members of farmers’ families, the dividend effect brought about by digital technologies can compensate for weaker infrastructure and lower human capital endowment [77]. It not only helps farmers find matching non-agricultural jobs but also alleviates their worries about participating in farmland transfer and obtaining productive funds for investment in the upper reaches of the Yellow River, thereby promoting the development of the rural land market and capital market. This implies that digital adoption yields the highest marginal benefits in less-developed regions by reducing information inequality and spatial transaction costs. However, the differences across regions also imply that while the dividend effect of digital technology application promotes the optimal allocation of production factors, it also gives rise to a “digital divide” [65], which in turn widens the “income gap” between rural households in the upper reaches and the middle reaches of the Yellow River [10,13,52,65].

5.6.3. By Farmland Scale

Farmland scale is a critical determinant of production decisions. Accordingly, households were grouped into small-, medium-, and large-scale categories. The heterogeneity results (Table 8) reveal that compared with non-adopters, adopters with large holdings benefit most. They were 22.7% more likely to engage in non-farm employment, 23.1% more likely to perform land transfer-out, and 21.8% more likely to obtain productive loans. These effects weaken progressively as land scale decreases. The likely explanation is that from the perspective of household decision-making, the information dividends and production technology promotion services brought by digital technology can help large-scale farmers reduce the costs of off-farm work and on-farm operations, promote the transfer of surplus agricultural labor to non-agricultural sectors, increase the availability of production credit for farmers, and boost their disposable income [44,71]. Furthermore, farmers with larger farmland scales have a stronger economic base and demonstration effects, which also encourage more efficient reallocation of resources across households of varying scales [14]. The goal is to achieve the optimal plan for maximizing the utilization of household labor endowment, farmland endowment, and capital endowment [14].

5.6.4. By Age Group

Given the severity of rural aging, households were divided into “new-generation” (≤55 years) and “older-generation” (>55 years) groups based on the mean age of household heads. The heterogeneity results (Table 8) reveal that adopters aged ≤55 experienced increases of 23.6% in non-farm employment, 19.7% in farmland transfer-out, and 22.5% in productive loans, all higher than the effects for the older generation. A possible reason is that the new generation of farmers often has a higher level of education, stronger ability to accept new things, and better digital skills than the older generation [3]. Under the same conditions, they are more likely to obtain non-agricultural employment opportunities, access information on farmland transfer, and secure more credit resources through digital technology channels. This finding echoes Gao et al. [3], who demonstrated that generational renewal accelerates rural digital transformation. However, the differences between farmers of different age groups also provide a new perspective for the subsequent implementation of differentiated digital technology support policies targeting different farmer groups [4,5].

5.7. Robustness Tests

Robustness was assessed using two complementary estimators: propensity score matching (PSM), an instrumental variable Probit (IV-Probit) model, replacing the core explanatory variable and excluding special samples.

The PSM results (Appendix A Table A1) indicate that the estimated treatment effects of digital technology adoption on non-farm employment, farmland transfer-out, and productive loan access remain positive and statistically significant across nearest-neighbor-, radius-, and kernel matching specifications.

The IV-Probit estimates (Appendix A Table A2) similarly confirm that the digital adoption coefficient remains significant and of comparable magnitude once endogeneity is addressed. These PSM and IV-Probit specifications therefore serve as robustness and sensitivity analyses of the ESP-based results.

Replacing the core explanatory variable (Appendix A Table A3) refers to selecting the degree of farmers’ participation in the digital economy to measure the actual situation of digital technology adoption. The degree of digital economy participation is represented by a 1–5 variable, where a higher value indicates a higher level of participation. The replaced core explanatory variable is then re-substituted into the model for re-estimation. The estimation results show that the degree of farmers’ participation in the digital economy has a significantly positive impact on rural households’ non-farm employment, farmland transfer-out, and access to agricultural productive loans.

Excluding special samples (Appendix A Table A4) refers to the following consideration: as a typical major agricultural province among all sample provinces, Henan Province exhibits significant differences from other provinces in terms of agricultural economic scale, factor endowments, per capita GDP level, and digital infrastructure development. These differences may exert a leverage effect on the estimation results, thereby affecting the empirical findings. Therefore, this paper conducts a re-test after excluding Henan Province as a special sample. The estimation results indicate that after removing the sample of Henan Province, digital technology adoption still has a significantly positive impact on rural households’ non-farm employment, farmland transfer-out, and access to agricultural productive loans.

Overall, the results of the aforementioned robustness tests all pass the significance test, indicating that the benchmark regression results of this paper are highly reliable.

6. Conclusions and Policy Implications

6.1. Conclusions

This study shows that adopting digital technology is associated with higher probabilities of non-farm employment, farmland transfer-out, and access to productive loans, with the largest ATT effects observed for labor allocation, followed by land and capital. These gains are concentrated among households in the upper reaches of the Yellow River Basin, those with larger landholdings, and those headed by younger and better-educated farmers.

6.1.1. Determinants of Adoption

Farmers’ adoption decisions are jointly shaped by individual, household, land, and village characteristics. Beyond personal preferences and capabilities, adoption is further constrained by structural factors such as uneven digital infrastructure, limited network coverage, high device and data costs, and imperfect local service provision, particularly in more remote areas. Male household heads with higher educational attainment, Party membership, and stable marital status demonstrate a greater propensity to adopt digital tools. Younger farmers, those with lighter dependency burdens, stronger economic resources, larger farmland holdings, and access to well-developed transport and communication infrastructure are also more likely to participate. These findings highlight that both human capital and infrastructural endowment are critical preconditions for digital integration in agriculture, consistent with previous empirical evidence.

6.1.2. Overall Effect on Factor Allocation

Digital technology adoption significantly enhances the efficiency of labor, land, and capital allocation. However, the dependent variables in this study are dichotomous and capture only whether households engaged in these activities, not the extent or intensity of participation. Counterfactual analysis shows that, without digital engagement, adopters’ probabilities of non-farm employment, farmland transfer, and productive loan access would decline by 24.5%, 19.3%, and 16.7%, respectively. Conversely, non-adopters could raise these probabilities by 27.4%, 22.2%, and 18.9% if they embraced digital tools. These effects demonstrate that digital participation facilitates labor mobility toward higher productivity sectors, promotes market-based land transactions, and improves credit accessibility through digital finance, thereby reducing transaction costs and information asymmetries in rural factor markets.

6.1.3. Heterogeneity of Impacts

The benefits of digital adoption exhibit marked heterogeneity across time, regions, landholding scales, and age demographic characteristics. The magnitude of effects increased between 2019 and 2022, reflecting policy momentum from the 14th Five-Year Plan for Digital Economy Development. Stronger effects were observed in the upper reaches of the Yellow River, where baseline infrastructure is weaker and information gains are larger, and among households with greater land scale or younger decision-makers. This pattern underscores that digital transformation generates both efficiency gains and potential divides, necessitating inclusive strategies to avoid widening inequality.

6.1.4. Mechanisms of Influence

The mechanism analysis provides evidence that is consistent with three reinforcing pathways through which digital technology adoption may support more efficient allocation of labor, land, and capital. First, adopters exhibit higher levels of information access, which aligns with the stronger effects on non-farm employment. Second, they report denser and more active social networks, which is consistent with reduced relational barriers and somewhat larger effects on farmland transfer-out. Third, adopters face weaker credit constraints, in line with the positive but smaller effects on productive loan access. These patterns, together with the ESP-based counterfactual estimates and the observed heterogeneity by year, region, land scale, and age group, suggest that information, network, and credit channels jointly shape how digital technology interacts with local infrastructure and institutional conditions to influence rural factor allocation.

Despite these contributions, several limitations remain. First, the analysis relies on pooled cross-sectional data, which limits our ability to observe household dynamics over time. Second, the instrumental variable reflects cross-sectional social interaction patterns and may not fully capture longer-term diffusion or network spillovers. Third, digital technology adoption is measured as a binary indicator, which cannot reflect variation in adoption depth or usage intensity. Future research could address these limitations by using panel data, staggered policy rollouts, or detailed digital usage measures to more precisely identify how digital technology shapes factor allocation over time.

6.2. Policy Implications

Based on these empirical findings—namely, that digital technology adoption significantly improves the allocation of labor, land, and capital but with stronger effects for younger, better-educated, larger-scale households and in regions and villages with better infrastructure—we propose three concrete policy levers. These levers target the specific channels and covariates that significantly shape adoption and outcomes in our ESP estimates (Table 5) and ATT (Table 6)/heterogeneity results (Table 8): digital skills support for disadvantaged households, land market digitization in key regions, and integrated financial and digital credit services in well-connected villages.

6.2.1. Strengthen Rural Digital Infrastructure and Integrate Technology with Factor Markets

Accelerating the construction of ‘digital villages’ under national and provincial programs is essential to translate the positive effects of digital adoption on labor, land, and capital allocation into broad-based gains. Upgrading broadband networks, 5G coverage, and digital service stations at the village-level can reduce transaction costs in employment, land transfer, and credit markets and make it easier for households to access non-farm jobs, formal land contracts, and productive loans. In relatively remote areas, especially in the upper reaches of the Yellow River Basin where our results show both strong potential gains and greater infrastructure gaps, priority should be given to building integrated digital platforms that link labor information, land transfer listings, and rural financial services. Such platforms can facilitate real-time matching of labor, land, and finance resources and help narrow regional disparities in factor allocation efficiency.

6.2.2. Develop Differentiated Regional Policies to Address External Barriers to Adoption

Given substantial spatial and socioeconomic heterogeneity within the Yellow River Basin, policy design should reflect the differentiated impacts of digital adoption identified in this study. In regions and households where adoption is still low—such as smaller-scale farms, older farmers, and areas with weaker infrastructure—targeted subsidies for devices and data, support for village-level service intermediaries, and measures to reduce perceived risks of online transactions can help overcome external barriers beyond individual preferences and capabilities. These external constraints underscore that adoption gaps are not solely driven by farmers’ individual willingness or skills, but also by the institutional and infrastructural environment in which they operate.

Strengthening coordination between the Ministry of Agriculture and Rural Affairs, financial regulators, and local governments is also crucial for integrating digital infrastructure planning with rural land and labor policies. Region-specific pilot programs that bundle digital infrastructure investment, land transfer platform development, and inclusive digital finance can help identify effective combinations of interventions that promote both economic efficiency and social equity across the Basin.

6.2.3. Enhance Farmers’ Digital Literacy and Skills Through Targeted Capacity-Building

Human capital development is central to sustaining digital transformation and ensuring that its benefits are not confined to a small group of early adopters. Our findings that younger and better-educated households are more likely to adopt and benefit from digital technology suggest the need for differentiated capacity-building programs tailored to different demographic groups. For younger and more market-oriented farmers, training should emphasize advanced digital applications—such as e-commerce, precision agriculture services, and online financial management—to deepen the intensity of use. For older, less-educated, or more vulnerable households, programs should focus on basic smartphone operation, safe use of digital platforms, and simple online transactions related to employment, land transfer, and credit. Embedding such training within existing agricultural extension and rural vitalization programs can help ensure that digital dividends are widely shared.

6.3. Limitations and Future Research Directions

Although the analysis provides robust empirical evidence, several limitations remain. First, the use of cross-sectional micro-survey data restricts the ability to capture long-term behavioral adaptation to digital transformation. Second, potential measurement bias may arise from self-reported indicators of technology use. Third, digital technology adoption is measured only as a binary variable, so we cannot distinguish partial or intensity-based adoption, which future studies should address. Fourth, our dependent variables are also dichotomous and thus do not reflect the extent of factor allocation (e.g., hours worked, land area transferred, or loan amounts), which may limit the precision of the estimated effects. Future research could integrate panel data, spatial econometric models, and digital trace datasets to examine dynamic effects and inter-village spillovers. Moreover, exploring the interaction between digitalization, environmental sustainability, and gender equity would further deepen understanding of inclusive digital transformation in rural China.